machine-bridge-mcp 2.0.0 → 3.0.0-beta.102
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +1142 -0
- package/CONTRIBUTING.md +23 -25
- package/GOVERNANCE.md +2 -2
- package/README.md +58 -20
- package/SECURITY.md +161 -120
- package/SUPPORT.md +1 -1
- package/browser-extension/broker-auth.js +83 -0
- package/browser-extension/broker-liveness.js +35 -0
- package/browser-extension/browser-error-boundary.js +96 -0
- package/browser-extension/browser-operations.js +1456 -101
- package/browser-extension/devtools-input.js +195 -34
- package/browser-extension/devtools-observation.js +429 -0
- package/browser-extension/devtools-session.js +48 -0
- package/browser-extension/manifest.json +4 -3
- package/browser-extension/page-automation.js +578 -118
- package/browser-extension/pairing-bootstrap.js +42 -0
- package/browser-extension/pairing.js +28 -12
- package/browser-extension/service-worker.js +126 -106
- package/docs/AGENT_CONTEXT.md +13 -10
- package/docs/ARCHITECTURE.md +159 -67
- package/docs/AUDIT.md +1302 -4
- package/docs/CLIENTS.md +19 -6
- package/docs/COMPUTER_USE.md +297 -0
- package/docs/ENGINEERING.md +72 -13
- package/docs/GETTING_STARTED.md +18 -20
- package/docs/LOCAL_AUTHORIZATION.md +115 -72
- package/docs/LOCAL_AUTOMATION.md +27 -21
- package/docs/LOGGING.md +40 -19
- package/docs/MANAGED_JOBS.md +37 -34
- package/docs/MULTI_ACCOUNT.md +98 -37
- package/docs/OPERATIONS.md +145 -51
- package/docs/OVERVIEW.md +14 -11
- package/docs/POLICY_REFERENCE.md +71 -3
- package/docs/PRIVACY.md +41 -2
- package/docs/PROJECT_STANDARDS.md +29 -6
- package/docs/RELEASING.md +238 -64
- package/docs/TESTING.md +150 -39
- package/docs/THREAT_MODEL.md +164 -67
- package/docs/TOOL_REFERENCE.md +653 -60
- package/docs/UPGRADING.md +78 -18
- package/native/macos/MachineBridgeBackgroundInput.swift +1051 -0
- package/native/macos/MachineBridgeBackgroundInputSmokeFixture.swift +295 -0
- package/native/macos/MachineBridgeTrustBroker.swift +233 -0
- package/package.json +94 -16
- package/scripts/accepted-candidate-tarball.mjs +77 -0
- package/scripts/candidate-runtime-store.mjs +93 -0
- package/scripts/check-plan.mjs +73 -3
- package/scripts/check-runner.mjs +129 -17
- package/scripts/commit-message-check.mjs +5 -2
- package/scripts/consumer-package-security.mjs +293 -0
- package/scripts/coverage-check.mjs +251 -15
- package/scripts/coverage-generation.mjs +41 -0
- package/scripts/coverage-range-merge.mjs +41 -0
- package/scripts/foreground-daemon-recovery.mjs +88 -0
- package/scripts/generate-policy-reference.mjs +28 -5
- package/scripts/generate-tool-reference.mjs +1 -1
- package/scripts/generate-worker-types.mjs +21 -11
- package/scripts/github-backlog.mjs +11 -10
- package/scripts/github-push.mjs +62 -16
- package/scripts/github-release-asset.mjs +58 -0
- package/scripts/github-release.mjs +241 -132
- package/scripts/global-package-installation.mjs +68 -0
- package/scripts/hardened-npm-session.mjs +42 -0
- package/scripts/install-published-prerelease.mjs +198 -0
- package/scripts/local-release-acceptance.mjs +103 -45
- package/scripts/macos-background-input-smoke.mjs +515 -0
- package/scripts/network-retry.mjs +8 -1
- package/scripts/npm-global-prefix.mjs +27 -0
- package/scripts/npm-publication-policy.mjs +17 -0
- package/scripts/official-mcp-conformance.mjs +267 -0
- package/scripts/persistent-activation-process.mjs +101 -0
- package/scripts/prepare-pinned-npm.mjs +12 -69
- package/scripts/prerelease-activation.mjs +160 -0
- package/scripts/privacy-check.mjs +14 -6
- package/scripts/promotion-digest.mjs +128 -0
- package/scripts/publish-npm.mjs +260 -0
- package/scripts/published-release.mjs +132 -0
- package/scripts/release-acceptance.mjs +62 -54
- package/scripts/release-browser-extension-store.mjs +172 -0
- package/scripts/release-candidate-manifest.mjs +52 -0
- package/scripts/release-channel.mjs +106 -0
- package/scripts/release-diagnostic.mjs +41 -0
- package/scripts/release-impact-check.mjs +12 -22
- package/scripts/release-oauth-canary-core.mjs +325 -0
- package/scripts/release-oauth-canary-evidence.mjs +100 -0
- package/scripts/release-oauth-canary.mjs +127 -0
- package/scripts/release-publication-guard.mjs +67 -0
- package/scripts/release-soak.mjs +284 -0
- package/scripts/release-state.mjs +1 -1
- package/scripts/run-checks.mjs +44 -7
- package/scripts/run-worker-dry-run.mjs +21 -0
- package/scripts/sarif-security-gate.mjs +2 -1
- package/scripts/sbom-check.mjs +106 -0
- package/scripts/start-release-candidate.mjs +267 -44
- package/scripts/syntax-check.mjs +36 -14
- package/scripts/verification-generation-guard.mjs +15 -0
- package/scripts/verification-idle-sleep-guard.mjs +37 -0
- package/scripts/verification-state.mjs +108 -0
- package/scripts/wrangler-command-lifecycle.mjs +115 -0
- package/src/local/account-access.mjs +13 -1
- package/src/local/account-admin-response.mjs +92 -0
- package/src/local/account-admin.mjs +52 -19
- package/src/local/agent-context-projection.mjs +26 -7
- package/src/local/agent-context.mjs +88 -49
- package/src/local/agent-skill-discovery.mjs +16 -5
- package/src/local/app-automation-macos-jxa.mjs +350 -0
- package/src/local/app-automation.mjs +873 -163
- package/src/local/application-capability-projection.mjs +24 -0
- package/src/local/authority-context.mjs +106 -0
- package/src/local/autostart-log-maintenance.mjs +36 -0
- package/src/local/browser-bridge-http.mjs +5 -3
- package/src/local/browser-bridge.mjs +109 -31
- package/src/local/browser-broker-auth-http.mjs +67 -0
- package/src/local/browser-broker-auth.mjs +98 -0
- package/src/local/browser-broker-routes.mjs +63 -20
- package/src/local/browser-broker-server.mjs +17 -7
- package/src/local/browser-command.mjs +79 -30
- package/src/local/browser-computer-observation-service.mjs +266 -0
- package/src/local/browser-extension-identity.mjs +50 -0
- package/src/local/browser-extension-path.mjs +22 -0
- package/src/local/browser-extension-protocol.mjs +50 -11
- package/src/local/browser-operation-service.mjs +147 -117
- package/src/local/browser-pairing-grant.mjs +109 -0
- package/src/local/browser-pairing-http.mjs +36 -0
- package/src/local/browser-pairing-launch.mjs +58 -0
- package/src/local/browser-pairing-store.mjs +85 -48
- package/src/local/browser-request-registry.mjs +40 -14
- package/src/local/browser-request-settlement.mjs +78 -0
- package/src/local/browser-resource-input.mjs +74 -0
- package/src/local/browser-trusted-input-health.mjs +63 -0
- package/src/local/call-authority.mjs +21 -0
- package/src/local/call-capacity.mjs +45 -0
- package/src/local/call-registry-drain.mjs +54 -0
- package/src/local/call-registry.mjs +30 -15
- package/src/local/capability-observer.mjs +5 -0
- package/src/local/capability-ranking.mjs +3 -2
- package/src/local/child-process-settlement.mjs +132 -0
- package/src/local/cli-account-admin.mjs +40 -6
- package/src/local/cli-activate.mjs +162 -0
- package/src/local/cli-local-admin.mjs +20 -45
- package/src/local/cli-options.mjs +12 -13
- package/src/local/cli-ready-output.mjs +52 -0
- package/src/local/cli-service.mjs +91 -11
- package/src/local/cli.mjs +232 -150
- package/src/local/computer-use-application-observation.mjs +224 -0
- package/src/local/computer-use-arguments.mjs +270 -0
- package/src/local/computer-use-deadline.mjs +50 -0
- package/src/local/computer-use-dispatch-settlement.mjs +40 -0
- package/src/local/computer-use-expectation.mjs +171 -0
- package/src/local/computer-use-observation.mjs +769 -0
- package/src/local/computer-use-recovery.mjs +82 -0
- package/src/local/computer-use-result-budget.mjs +145 -0
- package/src/local/computer-use-snapshot-store.mjs +67 -0
- package/src/local/computer-use.mjs +2255 -0
- package/src/local/daemon-http-relay-auth.mjs +26 -0
- package/src/local/daemon-http-relay-connection.mjs +298 -0
- package/src/local/daemon-http-relay-request.mjs +64 -0
- package/src/local/daemon-http-relay-sequence.mjs +49 -0
- package/src/local/daemon-process.mjs +40 -7
- package/src/local/delegated-process-sandbox.mjs +148 -0
- package/src/local/device-identity.mjs +156 -49
- package/src/local/device-root-provider.mjs +81 -0
- package/src/local/directory-metadata.mjs +47 -0
- package/src/local/doctor-reporting.mjs +23 -0
- package/src/local/durable-process-spec.mjs +80 -0
- package/src/local/errors.mjs +25 -15
- package/src/local/exclusive-file.mjs +74 -26
- package/src/local/execution-limits.mjs +9 -0
- package/src/local/execution-routing.mjs +276 -0
- package/src/local/execution-surface.mjs +23 -0
- package/src/local/file-mutation-coordinator.mjs +47 -0
- package/src/local/file-snapshot-preservation.mjs +24 -0
- package/src/local/filesystem-identity.mjs +54 -0
- package/src/local/fixed-process-environment.mjs +21 -0
- package/src/local/full-access-test.mjs +31 -13
- package/src/local/git-commit.mjs +69 -0
- package/src/local/git-config-safety.mjs +55 -0
- package/src/local/git-log-parser.mjs +41 -0
- package/src/local/git-metadata-boundary.mjs +77 -0
- package/src/local/git-metadata-tree-safety.mjs +59 -0
- package/src/local/git-operation-state.mjs +19 -0
- package/src/local/git-service.mjs +165 -30
- package/src/local/hardened-npm-download-timeout.mjs +46 -0
- package/src/local/hardened-npm-download.mjs +109 -0
- package/src/local/hardened-npm-extract.mjs +20 -0
- package/src/local/hardened-npm-verification.mjs +104 -0
- package/src/local/hardened-npm.mjs +199 -0
- package/src/local/job-runner.mjs +244 -89
- package/src/local/lifecycle.mjs +6 -0
- package/src/local/log.mjs +19 -45
- package/src/local/macos-background-input.mjs +398 -0
- package/src/local/macos-idle-sleep-assertion.mjs +65 -0
- package/src/local/macos-trust-broker.mjs +404 -0
- package/src/local/managed-job-cancellation.mjs +42 -0
- package/src/local/managed-job-capacity.mjs +16 -0
- package/src/local/managed-job-directory-generation.mjs +60 -0
- package/src/local/managed-job-directory.mjs +75 -0
- package/src/local/managed-job-durable-process.mjs +49 -0
- package/src/local/managed-job-lock.mjs +46 -14
- package/src/local/managed-job-plan.mjs +38 -14
- package/src/local/managed-job-projection.mjs +23 -15
- package/src/local/managed-job-retention.mjs +128 -0
- package/src/local/managed-job-runner-claim.mjs +70 -0
- package/src/local/managed-job-runner.mjs +35 -10
- package/src/local/managed-job-storage.mjs +22 -6
- package/src/local/managed-job-terminal-maintenance.mjs +40 -0
- package/src/local/managed-job-terminal.mjs +150 -0
- package/src/local/managed-jobs.mjs +424 -287
- package/src/local/npm-cli.mjs +42 -0
- package/src/local/npm-environment.mjs +37 -0
- package/src/local/operation-authorization.mjs +68 -324
- package/src/local/operation-risk.mjs +51 -4
- package/src/local/owner-state-lock.mjs +118 -0
- package/src/local/package-identity.mjs +9 -0
- package/src/local/patch.mjs +20 -18
- package/src/local/path-inspection.mjs +23 -0
- package/src/local/policy.mjs +19 -0
- package/src/local/private-toolchain-integrity.mjs +54 -0
- package/src/local/process-execution.mjs +87 -75
- package/src/local/process-foreground-timeout.mjs +50 -0
- package/src/local/process-identity.mjs +41 -7
- package/src/local/process-nonreplayable-settlement.mjs +50 -0
- package/src/local/process-result-projection.mjs +22 -0
- package/src/local/process-session-events.mjs +47 -0
- package/src/local/process-session-remote-activity.mjs +10 -0
- package/src/local/process-session-termination.mjs +73 -0
- package/src/local/process-sessions.mjs +144 -112
- package/src/local/process-tracker.mjs +140 -13
- package/src/local/process-tree-ownership-types.d.ts +37 -0
- package/src/local/process-tree-ownership.mjs +57 -0
- package/src/local/process-tree-signal.mjs +34 -0
- package/src/local/process-tree-snapshot.mjs +77 -0
- package/src/local/process-tree-supervisor.mjs +69 -0
- package/src/local/process-tree.mjs +11 -53
- package/src/local/project-metadata.mjs +19 -6
- package/src/local/relay-call-recovery.mjs +58 -38
- package/src/local/relay-connection-classification.mjs +131 -0
- package/src/local/relay-connection-support.mjs +91 -0
- package/src/local/relay-connection.mjs +179 -272
- package/src/local/relay-diagnostics.mjs +91 -0
- package/src/local/relay-heartbeat.mjs +119 -0
- package/src/local/relay-liveness.mjs +109 -0
- package/src/local/relay-peer-diagnostics.mjs +29 -0
- package/src/local/release-runtime-lock.mjs +18 -0
- package/src/local/remote-activity-idle-sleep-guard.mjs +77 -0
- package/src/local/remote-configuration.mjs +48 -0
- package/src/local/resilient-relay-connection.mjs +178 -0
- package/src/local/resource-admission-diagnostics.mjs +52 -0
- package/src/local/resource-admission-policy.mjs +115 -0
- package/src/local/resource-admission.mjs +339 -0
- package/src/local/resource-build-root.mjs +83 -0
- package/src/local/resource-cargo-concurrency.mjs +38 -0
- package/src/local/resource-cmake-concurrency.mjs +42 -0
- package/src/local/resource-command-concurrency.mjs +37 -0
- package/src/local/resource-command-profile.mjs +176 -0
- package/src/local/resource-coordinator-accounting.mjs +21 -0
- package/src/local/resource-cpu-window.mjs +20 -0
- package/src/local/resource-disk-headroom.mjs +13 -0
- package/src/local/resource-elastic-memory.mjs +19 -0
- package/src/local/resource-elastic-request.mjs +36 -0
- package/src/local/resource-foreground-wait.mjs +29 -0
- package/src/local/resource-go-concurrency.mjs +43 -0
- package/src/local/resource-gradle-concurrency.mjs +57 -0
- package/src/local/resource-host-cache.mjs +39 -0
- package/src/local/resource-host-darwin.mjs +54 -0
- package/src/local/resource-host-linux.mjs +35 -0
- package/src/local/resource-host-sample-file.mjs +23 -0
- package/src/local/resource-host-snapshot.mjs +99 -0
- package/src/local/resource-lease-accounting.mjs +119 -0
- package/src/local/resource-light-command.mjs +6 -0
- package/src/local/resource-make-concurrency.mjs +42 -0
- package/src/local/resource-maven-concurrency.mjs +38 -0
- package/src/local/resource-ninja-command-concurrency.mjs +20 -0
- package/src/local/resource-ninja-concurrency.mjs +28 -0
- package/src/local/resource-operations.mjs +26 -10
- package/src/local/resource-pressure.mjs +67 -0
- package/src/local/resource-probe-command.mjs +35 -0
- package/src/local/resource-process-admission.mjs +59 -0
- package/src/local/resource-process-ancestry-cache.mjs +43 -0
- package/src/local/resource-process-ancestry.mjs +32 -0
- package/src/local/resource-process-priority.mjs +15 -0
- package/src/local/resource-project-key.mjs +53 -0
- package/src/local/resource-pytest-concurrency.mjs +33 -0
- package/src/local/resource-release-control-classification.mjs +14 -0
- package/src/local/resource-release-control-executable.mjs +70 -0
- package/src/local/resource-release-control-workspace.mjs +41 -0
- package/src/local/resource-request-contract.mjs +30 -0
- package/src/local/resource-script-classification.mjs +44 -0
- package/src/local/resource-shell-analysis.mjs +58 -0
- package/src/local/resource-staging-recovery.mjs +97 -0
- package/src/local/resource-swift-concurrency.mjs +31 -0
- package/src/local/resource-transaction-lock.mjs +189 -0
- package/src/local/resource-wait.mjs +48 -0
- package/src/local/resource-waiters.mjs +139 -0
- package/src/local/resource-xcode-command.mjs +7 -0
- package/src/local/resource-xcode-concurrency.mjs +34 -0
- package/src/local/resource-xcode-non-build.mjs +19 -0
- package/src/local/runtime-activation.mjs +700 -0
- package/src/local/runtime-activity-projection.mjs +48 -0
- package/src/local/runtime-capabilities.mjs +48 -15
- package/src/local/runtime-diagnostic-state.mjs +14 -0
- package/src/local/runtime-diagnostics.mjs +40 -13
- package/src/local/runtime-info-projection.mjs +60 -0
- package/src/local/runtime-info-relay-projection.mjs +21 -0
- package/src/local/runtime-paths.mjs +2 -1
- package/src/local/runtime-process-routing.mjs +33 -0
- package/src/local/runtime-relay-connection-options.mjs +44 -0
- package/src/local/runtime-relay-control.mjs +98 -0
- package/src/local/runtime-relay.mjs +26 -50
- package/src/local/runtime-reporting.mjs +37 -21
- package/src/local/runtime-resource-service.mjs +77 -0
- package/src/local/runtime-tool-handlers.mjs +13 -10
- package/src/local/runtime.mjs +244 -208
- package/src/local/secure-file.mjs +73 -20
- package/src/local/security-audit-dispatch.mjs +35 -0
- package/src/local/security-audit-log.mjs +215 -0
- package/src/local/security-audit-state.mjs +139 -0
- package/src/local/security-audit-storage.mjs +135 -0
- package/src/local/security-audit-warning.mjs +31 -0
- package/src/local/security-audit-worker.mjs +87 -0
- package/src/local/service-convergence.mjs +42 -1
- package/src/local/service-definition.mjs +24 -0
- package/src/local/service-environment.mjs +18 -10
- package/src/local/service-owner.mjs +146 -0
- package/src/local/service-ownership.mjs +17 -0
- package/src/local/service-restart-handoff.mjs +64 -0
- package/src/local/service-restart-scheduler.mjs +49 -0
- package/src/local/service-runtime-convergence.mjs +52 -0
- package/src/local/service-runtime.mjs +105 -0
- package/src/local/service-status.mjs +60 -0
- package/src/local/service.mjs +385 -92
- package/src/local/shell.mjs +39 -19
- package/src/local/ssh-key.mjs +158 -42
- package/src/local/state-inventory.mjs +50 -24
- package/src/local/state-owner-lock-inventory.mjs +23 -0
- package/src/local/state-root-owned-namespaces.mjs +83 -0
- package/src/local/state-root-retirement.mjs +89 -0
- package/src/local/state.mjs +386 -96
- package/src/local/stdio.mjs +96 -66
- package/src/local/support-state-projection.mjs +30 -0
- package/src/local/system-network-route.mjs +76 -0
- package/src/local/systemd-removal.mjs +20 -0
- package/src/local/tool-executor.mjs +85 -20
- package/src/local/tool-result-boundary.mjs +51 -0
- package/src/local/toolchain-operation-lock.mjs +41 -0
- package/src/local/tools.mjs +7 -5
- package/src/local/trusted-executable.mjs +53 -0
- package/src/local/trusted-git-executable.mjs +34 -0
- package/src/local/trusted-github-cli.mjs +24 -0
- package/src/local/windows-service-convergence.mjs +49 -0
- package/src/local/windows-service.mjs +52 -45
- package/src/local/worker-deployment-fingerprint.mjs +102 -0
- package/src/local/worker-deployment.mjs +48 -60
- package/src/local/worker-health.mjs +12 -4
- package/src/local/worker-secret-file.mjs +37 -17
- package/src/local/workspace-file-service.mjs +156 -239
- package/src/local/workspace-file-transaction.mjs +189 -0
- package/src/local/workspace-search.mjs +55 -0
- package/src/local/wrangler-toolchain/package-lock.json +1505 -0
- package/src/local/wrangler-toolchain/package.json +23 -0
- package/src/local/wrangler-toolchain-verification.mjs +135 -0
- package/src/local/wrangler-toolchain.mjs +188 -0
- package/src/shared/access-contract.json +7 -1
- package/src/shared/activation-recovery.mjs +49 -0
- package/src/shared/admin-auth.d.mts +2 -1
- package/src/shared/admin-auth.mjs +3 -3
- package/src/shared/authority-revocation.d.mts +9 -0
- package/src/shared/authority-revocation.mjs +28 -0
- package/src/shared/daemon-auth.d.mts +11 -0
- package/src/shared/daemon-auth.mjs +17 -0
- package/src/shared/device-session-auth.d.mts +20 -0
- package/src/shared/device-session-auth.mjs +51 -0
- package/src/shared/foreground-timeout.d.mts +15 -0
- package/src/shared/foreground-timeout.mjs +70 -0
- package/src/shared/log-redaction.d.mts +9 -0
- package/src/shared/log-redaction.mjs +51 -0
- package/src/shared/mcp-protocol.d.mts +27 -0
- package/src/shared/mcp-protocol.mjs +275 -0
- package/src/shared/project-overview-projection.d.mts +4 -0
- package/src/shared/project-overview-projection.mjs +96 -0
- package/src/shared/relay-contract.json +28 -0
- package/src/shared/result-projection.d.mts +2 -1
- package/src/shared/result-projection.mjs +13 -2
- package/src/shared/server-metadata.json +15 -9
- package/src/shared/tool-argument-validation.d.mts +17 -0
- package/src/shared/tool-argument-validation.mjs +325 -0
- package/src/shared/tool-call-capacity.d.mts +41 -0
- package/src/shared/tool-call-capacity.mjs +59 -0
- package/src/shared/tool-catalog.json +626 -57
- package/src/worker/access.ts +12 -0
- package/src/worker/account-admin.ts +55 -33
- package/src/worker/authority-revocations.ts +174 -0
- package/src/worker/authority.ts +52 -6
- package/src/worker/daemon-auth.ts +49 -68
- package/src/worker/daemon-channel.ts +36 -0
- package/src/worker/daemon-http-auth.ts +63 -0
- package/src/worker/daemon-http-channel.ts +120 -0
- package/src/worker/daemon-http-controller.ts +163 -0
- package/src/worker/daemon-http-protocol.ts +62 -0
- package/src/worker/daemon-http-registry.ts +82 -0
- package/src/worker/daemon-last-observation.ts +30 -0
- package/src/worker/daemon-liveness.ts +8 -6
- package/src/worker/daemon-ready-dispatch.ts +17 -0
- package/src/worker/daemon-ready-messages.ts +73 -0
- package/src/worker/daemon-ready-waiters.ts +79 -0
- package/src/worker/daemon-recovery-budget.ts +13 -0
- package/src/worker/daemon-registry.ts +58 -0
- package/src/worker/daemon-relay-diagnostics.ts +127 -0
- package/src/worker/daemon-socket-attachment.ts +55 -0
- package/src/worker/daemon-sockets.ts +72 -62
- package/src/worker/daemon-status.ts +38 -0
- package/src/worker/device-session-verifier.ts +129 -0
- package/src/worker/dpop.ts +157 -0
- package/src/worker/durable-process-timeout.ts +29 -0
- package/src/worker/errors.ts +57 -0
- package/src/worker/http.ts +112 -26
- package/src/worker/index.ts +446 -440
- package/src/worker/mcp-access.ts +57 -0
- package/src/worker/mcp-controller.ts +188 -0
- package/src/worker/mcp-http-accept.ts +11 -0
- package/src/worker/mcp-http-contract.ts +322 -0
- package/src/worker/mcp-initialization-compat.ts +141 -0
- package/src/worker/mcp-jsonrpc.ts +11 -29
- package/src/worker/mcp-removed-protocol.ts +17 -0
- package/src/worker/mcp-response-proxy.ts +135 -0
- package/src/worker/mcp-response-stream.ts +65 -0
- package/src/worker/mcp-stale-schema-compat.ts +57 -0
- package/src/worker/mcp-stream-proxy-contract.ts +48 -0
- package/src/worker/mcp-tool-call-input.ts +23 -0
- package/src/worker/nonce-store.ts +33 -35
- package/src/worker/oauth-authorization-page.ts +21 -1
- package/src/worker/oauth-client-admin.ts +52 -0
- package/src/worker/oauth-client-contract.ts +38 -0
- package/src/worker/oauth-controller.ts +107 -47
- package/src/worker/oauth-field-contract.ts +31 -0
- package/src/worker/oauth-record-contract.ts +7 -0
- package/src/worker/oauth-refresh-authority.ts +27 -0
- package/src/worker/oauth-refresh-exchange.ts +162 -0
- package/src/worker/oauth-refresh-families.ts +130 -27
- package/src/worker/oauth-refresh-persistence.ts +116 -0
- package/src/worker/oauth-state.ts +33 -18
- package/src/worker/oauth-store-validation.ts +131 -0
- package/src/worker/oauth-token-derivation.ts +33 -0
- package/src/worker/oauth-token-issuance.ts +119 -0
- package/src/worker/oauth-tokens.ts +56 -190
- package/src/worker/observability.ts +38 -5
- package/src/worker/pending-admission.ts +15 -0
- package/src/worker/pending-call-capacity.ts +108 -0
- package/src/worker/pending-call-contract.ts +42 -5
- package/src/worker/pending-call-deadlines.ts +75 -0
- package/src/worker/pending-calls.ts +131 -125
- package/src/worker/runtime-alarm-storage.ts +37 -0
- package/src/worker/runtime-alarm.ts +136 -0
- package/src/worker/server-info-activity.ts +36 -0
- package/src/worker/server-info-tool-delivery.ts +15 -0
- package/src/worker/server-info.ts +168 -0
- package/src/worker/tool-catalog.ts +47 -1
- package/src/worker/tool-timeout.ts +75 -14
- package/src/worker/websocket-protocol.ts +26 -2
- package/src/worker/worker-edge-guard.ts +90 -0
- package/src/worker/worker-edge-log.ts +65 -0
- package/src/worker/worker-entry.ts +52 -0
- package/src/worker/worker-mcp-config.ts +23 -0
- package/src/worker/worker-metadata.ts +48 -0
- package/src/worker/worker-rate-limit-key.ts +30 -0
- package/src/worker/worker-runtime-config.ts +20 -0
- package/src/worker/worker-static-routes.ts +54 -0
- package/src/worker/worker-task-lifetime.ts +18 -0
- package/tsconfig.local.json +17 -1
- package/wrangler.jsonc +13 -0
- package/src/local/cli-approval.mjs +0 -117
- package/src/local/operation-state-lock.mjs +0 -92
- package/src/worker/mcp-session.ts +0 -72
package/docs/AUDIT.md
CHANGED
|
@@ -1,5 +1,1293 @@
|
|
|
1
1
|
# Security and privacy audit notes
|
|
2
2
|
|
|
3
|
+
## 2026-08-18 beta.102 end-to-end interruption re-audit
|
|
4
|
+
|
|
5
|
+
**Live beta.101 falsified “fast detection equals recovery.”** A normal daemon-backed read reproduced the user-visible interruption while Worker-local `server_info` remained available. The Worker had no authenticated, ready, probing, or candidate daemon channel and the call exhausted the bounded recovery wait. After the daemon returned, retained evidence showed a 323,839 ms outage with 13 reconnect attempts ending in `relay_connect_timeout` and 9,830 ms of inbound silence before the previous ready WSS closed. launchd still reported the same PID, `runs=1`, and no termination signal; the matching interval contained no macOS Sleep/Wake event. A second live beta.101 interruption later recovered after about 10.6 seconds without a daemon restart. This proves the remaining product defect was not Computer Use, launchd crash, or idle sleep: beta.101 detected a broken WSS promptly but still depended on successful WSS reconnection as its only remote-daemon recovery transport. The host default route was carried by a tunnel-like interface during investigation, but available evidence does not identify a particular VPN/TUN product, node, Wi-Fi component, Cloudflare edge, ISP, or upstream device as the physical trigger.
|
|
6
|
+
|
|
7
|
+
**Independent signed HTTPS transport removes WSS as a single recovery point.** Beta.102 keeps WSS preferred but adds one root-certified ephemeral P-256 daemon HTTP channel using the same daemon instance and authority. Every fallback request signs the fixed Worker origin/route/server/version, a bounded one-time timestamped nonce, and the exact body SHA-256. Requests, responses, queues, sequence numbers, and retry cadence are bounded; request timeout is seven seconds, verified-channel liveness is twelve seconds, ordinary poll cadence is one second, minimum request-start interval is 750 ms, and the shared new-call recovery ceiling is fifteen seconds. Bidirectional contiguous sequence/ACK state makes a lost HTTP response retransmit the same transport envelope; duplicates are discarded before business handling and gaps fail closed. A response from an old/reset HTTP session cannot commit its sequence into the replacement session. The first authenticated request enters probing directly; no obsolete challenge state or public MCP recovery/session model is introduced.
|
|
8
|
+
|
|
9
|
+
**Asymmetric black holes require explicit same-instance takeover.** A local transport watchdog can know WSS is dead while the Worker has not received the close and still considers that socket ready. The first fallback draft therefore failed precisely where it was most needed: Worker HTTP admission returned standby whenever any ready WSS existed. The corrected protocol marks takeover only after an established active WSS fails locally. The Worker verifies the signed request, new-session preconditions, and exact daemon instance before retiring a Worker-side zombie WSS; malformed, stale, or different-instance requests cannot preempt the incumbent. In-flight pending ownership is rebound without replaying the business call. A real Wrangler regression holds an in-flight call on WSS, performs same-instance HTTPS takeover while the old WSS is still Worker-ready, settles the original call through HTTPS, then verifies a newly ready WSS reclaims primary ownership. A separate regression proves an invalid signed takeover receives a bounded rejection while the healthy WSS remains ready.
|
|
10
|
+
|
|
11
|
+
**Daemon-proven non-delivery now recovers transparently without weakening mutation safety.** Beta.101 added `resume_calls_ack.missing_ids`, but used that proof only to return retryable `unavailable`. Beta.102 tightens handshake order first: the daemon reconciles Worker `resume_calls`, processes `ready_ack`, commits replacement readiness, emits sequenced `https_ready` where applicable, and only then sends `resume_calls_ack`. A missing ID is generated from the union of local active calls and retained unacknowledged terminal results, so it proves the same daemon never acquired execution ownership for that call and that the replacement channel is ready. Only that state can trigger transport redelivery. The Worker retains the original pending record and retransmits the same call ID, tool arguments, and authority with a timeout reduced to the original remaining execution budget; it refuses the optimization when less than one second remains or the ready channel cannot accept the envelope. Request cancellation still removes pending ownership. Active calls, retained results, a different daemon instance, post-dispatch unknown outcomes, and any mutation that may have executed are never automatically replayed. Repeated transport recovery therefore cannot convert uncertainty into duplicate side effects.
|
|
12
|
+
|
|
13
|
+
**Additional race and observability closure.** HTTP `tool_call` handling now reaches the local ownership path before an asynchronous control-handler yield can let the transport ACK outrun `activeRelayCalls`. Inbound sequence commit also revalidates the same authenticated HTTP session after handler settlement so an old response cannot poison a freshly reset session. HTTPS global readiness clears the bridge-level outage projection while retaining WSS background reconnect diagnostics separately; the Worker reports the correct twelve-second fallback liveness rather than the ninety-second WSS fallback. Browser broker/extension liveness independently negotiates matching Ping/Pong sequencing and reconnects a half-open local extension socket; ambiguous trusted input quarantines that extension generation and never authorizes mutation replay. Process-session retention, Computer Use end-to-end deadlines, remote idle-sleep ownership, managed-job restart/finally behavior, resource admission, launchd handoff, release activation, and privacy/logging paths were re-reviewed independently and retained their existing safety boundaries. Finally, a live all-channel outage no longer erases the previous transport evidence: `server_info.daemon.previous_connection` retains only one in-memory last-verified transport/timestamp/sanitized-relay snapshot, excluding policy, tool membership, account/client identity, daemon instance/connection identity, call IDs, arguments, and results, and never using that snapshot as authority or routing state.
|
|
14
|
+
|
|
15
|
+
**Release boundary.** These are packaged beta.102 bytes. No beta.101 candidate, activation, OAuth-canary, or acceptance evidence authorizes them. The final tree must pass fresh architecture/lint/type checks, browser and Computer Use regressions, security/privacy checks, repeated relay fault loops, frozen `check:fast`, frozen `check:full`, exact candidate creation, and install-only verification. Persistent beta.102 activation still requires a new explicit owner authorization; only the activated package's OAuth canary and live WSS/HTTPS/browser/Computer Use acceptance can support final release acceptance.
|
|
16
|
+
|
|
17
|
+
## 2026-08-18 beta.101 interruption root-cause and independent hardening review
|
|
18
|
+
|
|
19
|
+
**Two interruption mechanisms, not one.** Live beta.100 observability had already separated `computer_act` failures from successful low-level application operations: the compound tool accumulated timeout/conflict failures while direct application mutation completed. Deterministic timing instrumentation then showed Chrome/application capture and mutation stages completing individually inside their own local budgets, exposing the composition defect instead of an Accessibility `set_value` hang. The hosted projection defaulted configurable foreground work to twenty seconds while local Computer Use defaulted to thirty, and the orchestration could hand nearly that full timeout to preflight, dispatch, and post-capture sequentially. Beta.101 gives `computer_act` one monotonic deadline across preflight, its one mutation attempt, verification, and post-observation; `computer_observe` likewise shares one deadline across screenshot, Accessibility inspection, and fallback window revalidation. Verification is bounded by the smaller of its own window and the remaining operation budget. A preflight deadline error is computed outside broad stale/unavailable catches so a definite no-side-effect timeout cannot be misclassified. The review additionally caught two edge conditions after the first repair: a final mutation-budget check originally happened after snapshot claim, so a definite pre-dispatch timeout could consume one-shot mutation authority without side effects; that check now precedes claim. The millisecond remainder also used `ceil` when projected into integer child seconds, which could grant an internal stage nearly one second beyond the actual end-to-end remainder; projection now uses `floor`. A proposed additional 500 ms local settlement reserve was removed during the same review because hosted calls already have a separate five-second Worker reply-settlement margin outside daemon execution, and double-charging that headroom unnecessarily shortened valid first-stage and verification budgets. Worker defaults for only the two compound Computer Use tools move to thirty seconds; the shared explicit remote maximum remains forty-five seconds and ordinary configurable foreground defaults remain twenty.
|
|
20
|
+
|
|
21
|
+
**Whole-daemon outage falsified as a crash.** When the local control plane later became unreachable even for a read-only file call, Worker `server_info` still answered but showed no ready daemon. After the service returned, launchd evidence showed the same daemon PID, `runs=1`, and no process restart. `pmset -g log` supplied the missing causal timeline: the Mac entered Sleep at 06:50:08 for about 509 seconds, Maintenance Sleep at 06:58:42 for about 980 seconds, and Sleep again at 07:15:04 for about 741 seconds, returning to full wake near 07:27:30. Daemon records line up with those intervals: an event-loop stall of about 505 seconds on the first wake, a relay outage recorded as 979/981 seconds across the second sleep, and an event-loop stall of about 718 seconds on the final wake. The apparent sixteen-minute `relay_heartbeat_timeout` therefore occurred while the host was suspended, not while the daemon was awake and network-stalled. This evidence does not support a crash, launchd restart, or an awake sixteen-minute Cloudflare outage.
|
|
22
|
+
|
|
23
|
+
**Awake interruption root cause and amplification.** A later incident supplied the missing non-sleep case. Around 09:12:30 local time macOS unified network state reported route/reachability recovery and primary connectivity up while targeted `pmset` evidence showed no Sleep/Wake transition. The launchd service still had the same PID, one lifetime run, and no exit. The ready relay did not produce its close `1006` / `connection_interrupted` transition until about 09:13:39 and was ready again at about 09:13:41. The roughly two-second close-to-ready outage therefore understated the user-visible interruption: beta.100 had no fast transport-only probe, so its twenty-five-second application heartbeat and seventy-five-second local silence decision (with a ninety-second Worker fallback) could leave an OPEN-but-black-holed WebSocket eligible for dispatch far longer than the ordinary twenty-second and compound Computer Use thirty-second hosted budgets. A deterministic regression first failed because no protocol probe existed at five seconds. Beta.101 adds `RelayLiveness`: protocol-level WebSocket Ping every five seconds with ten seconds without inbound proof before `relay_transport_timeout`, while preserving the separate twenty-five-second JSON application heartbeat and its seventy-five-second application-silence decision as `relay_heartbeat_timeout`; the Worker retains its wider ninety-second fallback. Protocol Pong therefore cannot mask a Worker application path that has stopped replying. For a scheduling-responsive daemon, the executable contract guards the full transport-detection horizon—one probe interval plus the silence threshold, fifteen seconds—below the ordinary twenty-second hosted execution default, leaving reconnect/result-settlement headroom instead of comparing only the timeout value. A diagnosed local event-loop stall follows the separate recovery-grace path and can intentionally extend that interval without being misclassified as remote transport silence. Diagnostics now preserve `previous_ready_inbound_silence_ms` across reconnect so operators can distinguish the silent pre-close interval from `outage_duration_ms`, which measures only close through verified readiness. This evidence establishes the Machine Bridge amplification defect; it does not identify which VPN/TUN, Wi-Fi, intermediary, edge, or upstream component caused the underlying network-path transition.
|
|
24
|
+
|
|
25
|
+
**Post-reconnect lost-dispatch settlement.** The transport watchdog repair exposed one more delivery-state hole that is independent of reconnect speed. The Worker registers a pending call before writing `tool_call` to the current ready socket. On a half-open transport, `WebSocket.send()` can return successfully even though the frame never reaches the local dispatcher. Before beta.101, a same-instance reconnect rebound that Worker pending record and sent the one-way `resume_calls` list, but the daemon only cancelled local work that the Worker no longer wanted; it did not tell the Worker which requested resume IDs it had never owned. Such a call could therefore remain pending until its original deadline even after the relay itself had recovered. The safety boundary for repairing this without replay is explicit: every relay call that reaches local dispatch enters `activeRelayCalls` before execution, and a completed result enters the bounded `pendingResults` ledger before delivery and remains there until Worker acknowledgement. Reconciliation now snapshots the union of those two ownership sets before cancelling or discarding anything. A Worker-resumed ID absent from both is therefore daemon-proven non-delivery, not an ambiguous mutation outcome. The daemon returns only those IDs in pre-ready `resume_calls_ack.missing_ids` on the same authenticated relay generation; failure to send that acknowledgement interrupts the generation instead of claiming successful reconciliation. The Worker strictly validates at most thirty-two unique bounded call IDs, rejects only pending records currently owned by that replacement socket, and returns retryable `unavailable` with `side_effects_started=false` and `reason=daemon_call_not_received_after_reconnect`. Active calls and retained terminal results continue under their existing ownership and are never replayed automatically. Local and Worker logs expose only aggregate missing-call counts; raw resume IDs, tool names, arguments, account identity, and results remain excluded.
|
|
26
|
+
|
|
27
|
+
**Bounded runtime mitigation.** Verification already had a whole-run macOS `caffeinate -i` wrapper, but normal remote daemon workflows had no equivalent. Permanently caffeinating a login daemon would be an unacceptable power-management change, so beta.101 separates the fixed macOS primitive from its owners. `macos-idle-sleep-assertion.mjs` is the only runtime adapter that starts `/usr/bin/caffeinate -i -w <owner-pid>` with `shell: false`, ignored stdio, and process-lifetime binding. The daemon-side `remote-activity-idle-sleep-guard.mjs` composes it only after policy/account/operation authorization and argument validation, so rejected or malformed relay traffic cannot keep the host awake. Authorized handlers hold reference-counted activity for their full execution lifetime: concurrent calls share the assertion, no inactivity timer runs while any handler is active, the fixed five-minute grace begins only after the last daemon-side activity settles, a new activity invalidates stale release timers, and surplus end calls are idempotent rather than extending the grace. A remote process session acquires an additional logical hold only after resource admission succeeds and releases it on child settlement or startup failure, so a child that outlives `start_process` is not reduced to the handler grace. Remote account managed-job runners cannot depend on daemon ownership because they intentionally survive relay/service replacement; after confirming the runner claim and validating persisted account ownership, each remote runner acquires its own runner-PID-bound assertion and retains it across recovery handoff, resource admission, steps, `finally_steps`, and terminal persistence before top-level `finally` releases it. Local managed jobs do not acquire this remote-continuity assertion. Runtime shutdown terminates process sessions before releasing the daemon assertion. The five-minute relay grace is fixed rather than exposed through a shell environment variable because the launchd service definition persists only its controlled `PATH` and offers no durable configuration channel for arbitrary shell overrides. Assertion/spawn/timer/logging failures remain fail-open for tool/job settlement and expose only a coarse error class; owner `diagnose_runtime` projects only daemon-side supported/enabled/active/grace/error-class fields and does not enumerate sessions or jobs. These mechanisms prevent ordinary Idle Sleep during owned remote work but intentionally do not promise to override explicit sleep or lid-close behavior.
|
|
28
|
+
|
|
29
|
+
**Independent architecture/documentation/privacy review.** A fresh current-state pass found two stale pieces of guidance in addition to the timeout implementation defect: README still summarized remote foreground process/shell/browser/application work as a blanket sixty-second daemon envelope even though remote process one-shots have been durable 10-second-acceptance/1–600-second-execution jobs since beta.93, and `docs/COMPUTER_USE.md` still claimed the first application verification capture could use the normal full action timeout. Both are corrected and now guarded. The same review challenged the first sleep-guard draft rather than accepting a green unit test: lease renewal was moved after authorization/schema validation to prevent rejected traffic from becoming a power-resource DoS; the inactivity timer was moved from handler start to the last concurrent handler's settlement; surplus end calls became idempotent; logging failures became fail-open; unexpected assertion-child exit gained coarse observability; an implicit `spawn` shell default became explicit `shell: false`; credential/path-shaped synthetic test strings were replaced with neutral markers; and a proposed environment grace override was removed after service tracing showed launchd does not persist arbitrary shell environment as configuration. The review then found that handler-only ownership still left long process sessions and detached jobs vulnerable after the five-minute grace, producing the separate session/job lifetime ownership described above. Computer Use/tool metadata explicitly describes one end-to-end deadline, architecture/operations/testing distinguish the two thirty-second compound defaults from ordinary twenty-second foreground defaults, and logging/diagnostics document power-management state without recording commands, tool names, paths, PIDs, arguments, results, session identities, or job identities. Every runtime assertion launch uses the same fixed absolute executable and flags with a numeric owner PID and never invokes a shell. Existing privacy/history checks remain the authoritative repository-wide sensitive-data gates.
|
|
30
|
+
|
|
31
|
+
**Verification and release boundary.** Focused Computer Use and Worker timeout tests pass after the deadline repair; runtime infrastructure and runtime self-tests pass after the sleep-guard integration. The remaining architecture, privacy, generated-doc, fast/full, candidate, install-only, live-activation/canary, browser-extension, and acceptance gates must be regenerated from the final frozen beta.101 tree. Beta.100 was already live-activated and its activated-package canary had passed, but browser extension acceptance remained incomplete; packaged beta.101 changes make all beta.100 candidate/acceptance evidence stale for release authorization.
|
|
32
|
+
|
|
33
|
+
## 2026-08-17 beta.100 independent architecture/privacy review
|
|
34
|
+
|
|
35
|
+
**Starting evidence and release boundary.** The review began from the beta.99 working tree after its local targeted checks, 103-task fast plan, 130-task full plan, candidate preparation, and disposable install-only preflight had succeeded. Beta.99 had never been live-activated or accepted. The owner runtime remained the previously activated beta.98 service. The review therefore treated beta.99's green verification as a baseline to challenge, not as evidence that the current design or documentation was complete. Any packaged change found here invalidates the beta.99 receipt/candidate and requires a new prerelease rather than silently reusing prior evidence.
|
|
36
|
+
|
|
37
|
+
**Obsolete internal process-call shape.** `runtime-tool-handlers.mjs` dispatches public `exec_command` as one validated argument record, but `runtime-process-routing.mjs` and `LocalRuntime.execCommand()` still supported an older `(commandString, timeout, context)` overload. Repository search found no production consumer of that shape: only runtime-self/full-access fixtures called it. The overload duplicated timeout/context interpretation and maintained a second branch beneath an otherwise record-shaped MCP contract. Beta.100 removes it, updates the fixtures to `{ command, timeout_seconds }`, and adds an architecture guard rejecting `legacyCall`/`timeoutOrContext` reintroduction. This is internal compatibility cleanup; the public MCP schema, shell limits, resource admission, and durable relay-origin process routing are unchanged.
|
|
38
|
+
|
|
39
|
+
**Audit shutdown and catch-hygiene edge cases.** `SecurityAuditLog.close()` previously computed its deadline with `Math.max(1, Number(timeoutMs) || CLOSE_TIMEOUT_MS)`. A direct control experiment passed `Infinity` through a non-settling synthetic Worker; Node emitted `TimeoutOverflowWarning: Infinity does not fit into a 32-bit signed integer` and reported that the timer duration was set to 1. The fake Worker did not keep the event loop alive and the deliberately unref'ed timer therefore did not provide a terminal close result, so this audit does not claim a measured one-millisecond production settlement. The causal defect is nevertheless explicit: an invalid internal delay could reach `setTimeout` and be inverted into Node's overflow behavior; the same Node timer boundary also rejects oversized finite delays. Beta.100 normalizes non-finite/uncoercible/non-positive delays to the existing five-second close default and caps every positive override at that same five-second shutdown window before timer creation. Regression coverage checks both `Infinity` and `Number.MAX_SAFE_INTEGER` and keeps the close pending until a synthetic worker acknowledgement instead of allowing timer overflow to decide settlement. The same review found that repository hygiene rejected empty `.catch(() => {})` but not optional-chained `.catch?.(() => {})`; `security-audit-log.mjs` contained exactly one such unexplained best-effort termination catch. The production catch now documents why termination cannot change the already-failed audit settlement, and the hygiene detector plus self-checks cover both syntaxes.
|
|
40
|
+
|
|
41
|
+
**Package-impact semantics and historical wording.** Direct inspection of the beta.99 tarball showed no `package/tests/**` entries. The current `package.json.files` manifest likewise excludes `tests`, and `release-impact-check.mjs` computes package relevance from that manifest plus package/lock metadata. This falsifies the current beta.98 shorthand that described its regression-test change itself as packaged bytes. The release policy was conservative rather than unsafe—beta.98 also changed real packaged source/version/documentation—but the wording could mislead future maintainers into treating every test-only change as a package mutation. An attempted check of older beta.74–86 package manifests could not establish their historical tarball policy because those version tags are not retained in the current local repository, so older incident wording is deliberately left untouched rather than inferred from today's manifest. Beta.100 corrects the current beta.98 statement, documents that tests are verification evidence rather than tarball entries, adds an explicit release-impact regression proving a `tests/**`-only change does not require an npm bump, and adds architecture guards so silently adding `tests` to `package.json.files` requires a deliberate policy/documentation update.
|
|
42
|
+
|
|
43
|
+
**Verification concurrency documentation.** `docs/TESTING.md` said the parallel-safe fast worker pool used four workers by default. Source tracing shows the actual default is `Math.max(1, Math.min(4, availableParallelism()))`; the explicit environment override remains 1–16. This matches beta.99 runs that reported two or three available workers and means the prose had promoted a cap into a fixed default. Beta.100 documents the dynamic `availableParallelism()` bound and protects the source/doc relation in the release-contract suite.
|
|
44
|
+
|
|
45
|
+
**Redundancy and maintainability.** Agent-context, full-access, and runtime-self tests carried byte-for-byte identical synthetic healthy-host functions. They now import one test fixture so deterministic resource assumptions have one source. The production-size review also confirmed several intentionally broad modules remain close to executable ceilings (`runtime.mjs`, `relay-connection.mjs`, `computer-use.mjs`, browser operations/page automation, and app automation). No ceiling was raised and no Strategy/archive refactor was introduced: there is no demonstrated failure that justifies a wide rewrite. The practical maintenance constraint is that the next feature on those surfaces should extract a coherent owned lifecycle before adding more responsibility rather than increasing the cap.
|
|
46
|
+
|
|
47
|
+
**Release/JSON and privacy review with falsified branches.** Because beta.96 failed after an inner persistent activation produced non-JSON stdout, the review re-traced the current persistent `activate --json` path instead of assuming the beta.97 Wrangler fix closed every equivalent channel. The installed CLI invocation remains captured by the outer wrapper; JSON activation routes service logging through quiet/stderr-only loggers; managed-job authentication cannot start interactive Wrangler login; inner JSON Worker deployment is non-interactive. The remaining inherited-stdio execution is the separate non-persistent foreground-candidate path and does not feed the parsed JSON child. No second stdout-contamination mechanism was found, so no speculative release rewrite was retained. A targeted sensitive-metadata scan across source/scripts/extension/docs also found no concrete owner home path, account identifier, real `workers.dev` endpoint, private-key header, bearer header, or `client_secret`; the repository privacy/history gates still remain authoritative and are rerun on the final frozen tree.
|
|
48
|
+
|
|
49
|
+
**Recurrent message-timeout symptom and resource-admission evidence.** During this review the owner again observed the client message “send timed out; retry.” The live beta.98 control plane did not support a blanket relay/tool-timeout explanation: `server_info` reported a ready authenticated daemon with no pending calls, Worker observability reported zero timed-out tool calls, and the recent close-1006 relay interruption had recovered after roughly 4.5 seconds. `diagnose_runtime` did, however, show a distinct machine-user scheduling condition: sampled resource pressure was Green while one heavy lease remained active, nine roots were queued, six had aged into protected rank, and the oldest waiter was already about twenty-one minutes old. Reading a durable `pgrep` job later provided stronger terminal evidence: it failed after the complete thirty-minute resource-admission ceiling with `resource_error` and an empty `steps` result, proving the child never spawned. A later process query was still queued. These facts establish severe pre-spawn resource waiting during the same broad period; they do not establish that the ChatGPT client message error itself was caused by that queue rather than a host/UI/network layer that Machine Bridge cannot observe.
|
|
50
|
+
|
|
51
|
+
The scheduler's protected-drain behavior is intentional, not reclassified as a pressure-policy defect. Existing tests require an aged feasible rank-zero waiter blocked by coordinator-owned reservations to stop later fitting backfill until those reservations drain; consequently `pressure.state=green` can coexist with fairness blocking because the pressure state and queue fairness answer different questions. Beta.100 leaves admission limits, priorities, aging, and backfill semantics unchanged and repairs observability instead. `ResourceCoordinator.snapshot()` now exposes only the Boolean `waiters.drain_active`, calculated by the exact same selection state machine used for admission, and owner diagnostics project that Boolean without command text, path, PID, waiter identity, or reservation details. Managed-job status switches to the fixed low-cardinality phase `resource_admission` before acquisition and restores `steps`/`finally_steps`/`recovery-cleanup` after successful admission, so `read_job` can distinguish a process that has not spawned from a slow child. The runner retains numeric `resource_admission_ms` while preserving `duration_ms` as the total orchestration duration. Local/owner `read_job` projection exposes that pre-spawn portion for delayed-success postmortems, but delegated non-owner reads strip it because isolating machine-user admission delay would otherwise sharpen a cross-workload contention side channel beyond the existing total-duration signal. A status-write failure after lease acquisition but before spawn triggers an explicit lease-release attempt before the status error propagates; if that release also fails, the runner surfaces both failures together instead of silently suppressing the cleanup fault. The thirty-minute ceiling is now one shared durable-delivery contract value used by the runner, Worker tool descriptions, and `server_info.tool_delivery`; the separate 1–600-second process timeout remains child execution time after admission.
|
|
52
|
+
|
|
53
|
+
**Additional race/hygiene closure.** The bounded daemon-fixture exit helper still had a narrow event-registration race: it checked `exitCode`/`signalCode` before attaching the `exit` listener, so a child exiting between those operations could lose the event and turn into a false thirty-second timeout. The helper now attaches the listener, installs single-settlement timeout handling, and immediately rechecks child state; the architecture contract binds that post-registration check. Repository catch-hygiene was also widened without broad pattern matching: an empty Promise rejection handler is rejected across direct or optional-chained `.catch`, harmless whitespace, `async`, and parameterized arrow forms, while a non-empty explanatory comment or handling body remains accepted.
|
|
54
|
+
|
|
55
|
+
**Verification and release consequence.** The first targeted pass exposed only an extra final blank line left after fixture extraction; repository hygiene caught it and it was fixed. Focused resource-admission, security-audit, release-impact, agent-context, and full-access regressions passed. The correct `tests/local-self-test.mjs` entry completed every named phase in about thirty seconds, including the listener/recheck teardown boundary. Managed-job integration then passed twice: the second run exercised the final projection in which delegated operator reads omit `resource_admission_ms` while local/owner reads retain it. Worker runtime-infrastructure tests passed with the shared 1,800,000 ms admission contract and tool-delivery projection, and architecture, lint, typecheck, package, privacy, version, tool-reference generation, and whitespace gates have passed during the review; gates rerun before the final freeze remain authoritative. Attempts to start the broader Worker integration while unrelated machine-user work held capacity were rejected before spawn as `host_pressure_red` or `fairness_wait`; those responses are resource-admission evidence, not test failures, so Worker integration still requires an admitted run in the final full plan. Because beta.100 changes shipped local runtime and documentation bytes, beta.99's local full receipt and candidate are stale. Beta.100 requires a fresh frozen fast/full run and candidate/install-only cycle, and a later explicit owner authorization before live activation; no beta.100 acceptance, commit, push, npm publication, tag, or GitHub Release exists at this point.
|
|
56
|
+
|
|
57
|
+
## 2026-08-17 beta.99 pre-acceptance shipped-testing-contract drift
|
|
58
|
+
|
|
59
|
+
**Observed beta.98 live evidence.** The exact beta.98 candidate was owner-authorized and activated through the guarded persistent-candidate command. The durable activation job completed successfully after re-verifying the exact candidate tarball, and post-handoff `server_info` reported `3.0.0-beta.98` with one authenticated/ready daemon socket, readiness verified, and no pending calls. The canary executed from the package root derived from the activation record `runtime_entry`; authorization-code exchange, authenticated MCP, refresh rotation, refreshed MCP, and synthetic client/account cleanup all succeeded. `diagnose_runtime` then reported an operational runtime, authenticated/ready relay, zero active/draining processes, green resource admission, and a healthy security-audit worker with a verified chain, `content_logged=false`, and zero dropped records. The browser store had also converged on the beta.98 stable extension bytes, but the currently loaded browser extension still required reload; browser live verification therefore was not complete.
|
|
60
|
+
|
|
61
|
+
**Blocking pre-acceptance defect.** Before recording acceptance, a source/runtime/documentation cross-check found that shipped `docs/TESTING.md` still claimed `full-access:test`, `local-self-test`, and `agent-context-test` used five-minute real-host-pressure resource-admission waits. Current source contradicts that text: full-access and agent-context use isolated coordinators, synthetic healthy-host samplers, and ten-second process-admission budgets; beta.98 additionally changed runtime self-test to the same synthetic healthy-host isolation and ten-second one-shot/process-session admission, added a thirty-second daemon child-exit bound, and emitted per-phase progress markers. The five-minute managed-job CLI/stdio background-settlement observation windows are separate and remain valid. Because `docs/` is explicitly included in `package.json.files`, the stale prose is part of the candidate bytes rather than an out-of-band note.
|
|
62
|
+
|
|
63
|
+
**Release consequence and guard.** Beta.98 is not accepted or publishable despite successful activation and OAuth canary. No acceptance record was written. The documentation is corrected in beta.99, and the architecture release-contract suite now both rejects the obsolete five-minute fixture wording and requires the current isolated/synthetic ten-second contract. This is a packaged-byte change, so beta.98 live evidence cannot be reused as beta.99 acceptance evidence; beta.99 must repeat frozen verification, candidate preparation, install-only preflight, explicit owner activation, activated-package canary, observed live verification, and acceptance. The live beta.98 runtime remains a valid operational baseline while beta.99 is prepared; no rollback is justified by this documentation-only defect.
|
|
64
|
+
|
|
65
|
+
## 2026-08-17 beta.98 exact-head CodeQL API-shape drift
|
|
66
|
+
|
|
67
|
+
**Exact-head hosted finding.** Beta.97 was locally frozen, owner-activated, canary/live-verified, accepted, committed as `e7bf465790961db203ad751e8d876a60db880f98`, and guarded-pushed to PR #85 before hosted provider validation. Dependency Review, Governance, Workflow Policy, CodeQL's Actions analysis, package-audit, Ubuntu full verification, and Windows platform/install all passed on that exact head. JavaScript/TypeScript CodeQL completed analysis successfully but the repository's zero-unaccepted-SARIF gate rejected five `js/superfluous-trailing-arguments` results at the shared timeout helper and its Worker/test callers. The existing allowlisted direct-process shell-sink result remained accepted and is unrelated to this failure.
|
|
68
|
+
|
|
69
|
+
**Root cause, not waiver.** `src/shared/foreground-timeout.d.mts` still declared `remoteForegroundMaximumSeconds(name: string): number`, and all call sites still passed the tool name, but beta.97's JavaScript implementation had been simplified to `remoteForegroundMaximumSeconds()` because the maximum is currently uniform across configurable foreground tools. JavaScript tolerates the extra argument, so behavior tests, ESLint, and TypeScript did not expose the runtime/declaration arity mismatch; CodeQL correctly identified the stale contract shape. The repair restores the formal `_name` parameter while leaving the function body and timeout value unchanged. No SARIF exception is appropriate because the declaration/implementation inconsistency is real and trivially repairable.
|
|
70
|
+
|
|
71
|
+
**CodeQL regression.** A new local assertion binds `remoteForegroundMaximumSeconds.length === 1`. It failed against the beta.97 implementation before the source repair and passed afterward; worker runtime infrastructure, lint, typecheck, architecture, and `git diff --check` also pass.
|
|
72
|
+
|
|
73
|
+
**Independent macOS hosted timeout.** The same beta.97 CI run exposed a second provider-only failure that must not be normalized into the CodeQL result. The macOS platform job completed setup and tasks 1–117 of the 122-task platform plan by `2026-08-17T09:48:01Z`, then started task 118 (`self-test`). No self-test terminal output followed. At `10:06:42Z` GitHub cancelled that step at the job's 20-minute workflow ceiling, and the subsequent documented global-install step was skipped. Ubuntu full verification and Windows platform/install on the same PR succeeded, so this is not evidence of a general package or cross-platform failure. It is also not evidence that increasing the macOS job timeout is the correct repair: the same self-test normally completes in tens of seconds.
|
|
74
|
+
|
|
75
|
+
**Harness hardening with bounded causal claims.** The beta.97 self-test did not emit phase progress, so the provider log cannot prove which internal phase was active during the long silence. Source review found two independent harness defects that can make this non-resource test inherit unbounded or multi-minute machine conditions. `runtimeSelfTest()` used a private resource-coordinator root but still sampled the real shared host and set `processResourceWaitMs` to five minutes for four `LocalRuntime` fixtures; multiple process assertions could therefore wait on provider host pressure even though resource-pressure policy is independently covered elsewhere. Separately, `local-self-test.mjs` waited for daemon-fixture child `exit` events without any deadline. Beta.98 follows the already-established agent-context/full-access isolation pattern: runtime fixtures keep their private coordinator but inject the existing synthetic healthy host sample and reduce their harness-only resource wait to ten seconds; daemon child exit is bounded by the existing 30-second fixture budget; and each self-test phase logs start/completion so a future hosted stall identifies its phase. This does not weaken production resource admission or its dedicated pressure/fairness tests. On the owner Mac the hardened self-test completed every phase successfully in 29.3 seconds.
|
|
76
|
+
|
|
77
|
+
**Coverage ownership after isolation.** The first beta.98 full run then failed one local gate after the self-test fix: `coverage:test` reported `src/local/resource-host-cache.mjs` at 75% function coverage rather than its required 100%. This was a useful consequence of removing accidental real-host execution from the broad runtime self-test, not a reason to restore shared-host sampling. The resource-admission suite already owned explicit cache freshness/cross-scope semantics but did not execute the `withCachedIo()` quick-refresh path. Beta.98 adds deterministic cases for both a non-I/O CPU refresh that retains a valid scope-local I/O hint and an I/O request within the five-second fresh-I/O window. The dedicated coverage gate then passes with `resource-host-cache.mjs` at 100% function coverage (4/4), keeping policy/cache behavior and its coverage in the specialized suite.
|
|
78
|
+
|
|
79
|
+
**Release consequence.** These repairs change packaged shared runtime, version metadata, and documentation bytes after beta.97 acceptance; the regression-test changes are repository verification inputs rather than npm tarball entries. Beta.97's live acceptance remains valid evidence for the exact beta.97 candidate but cannot authorize beta.98. The live beta.97 service remains unchanged while beta.98 is rebuilt locally and must complete a fresh exact-tree verification/candidate/explicit-activation/canary/live-acceptance cycle before another guarded push.
|
|
80
|
+
|
|
81
|
+
## 2026-08-17 beta.97 managed-job activation authentication/JSON settlement
|
|
82
|
+
|
|
83
|
+
**Observed live release-path defect.** The beta.96 source passed frozen fast/full verification, produced an exact candidate, and the candidate later completed owner-terminal activation, activated-package OAuth canary, live Worker/daemon/browser verification, and local acceptance. After the local control plane recovered, however, the earlier activation that the coding agent had launched through the repository-required detached managed-job carrier had a definitive failed terminal record: it ran for about 793.6 seconds and exited with `persistent candidate activation did not return valid JSON`. It had verified and installed the exact beta.96 tarball and entered persistent activation before failing. The later owner-terminal invocation of the same guarded activation command succeeded. A targeted `pmset` review found no Sleep/Wake event during the failed 15:45–15:59 local interval and showed the machine kept a user-idle-sleep prevention assertion, so the beta.95 verification-sleep failure mechanism does not explain this activation incident.
|
|
84
|
+
|
|
85
|
+
**Source invariant violation and bounded inference.** The failed activation did not record the private Wrangler stage that was active, so this audit does not claim as an observed fact that the first run entered `wrangler login`. Source tracing nevertheless exposed a release-blocking branch whose timing and final parse error match the live symptom. `activatePersistentRuntime()` stopped the verified login provider before `prepareRemoteState()` called `ensureWorkerDeployment()`. That deployment function performed a captured `wrangler whoami`, but on authentication failure it invoked `wrangler login` without capture; the shared executable runner therefore inherited stdout/stderr and allowed login up to ten minutes. The outer release wrapper simultaneously ran the inner candidate as `machine-mcp activate ... --json` and treated the complete child stdout as one JSON document. Interactive Wrangler text could therefore both prolong the service-offline interval and make a successfully converged inner activation look like a JSON parse failure. Independently of whether that exact branch caused the beta.96 run, allowing an interactive login after service stop and inside a machine-readable stdout channel violates the documented managed-job activation contract.
|
|
86
|
+
|
|
87
|
+
**Repair.** Beta.97 moves Wrangler authentication to a release-wrapper preflight before the release-runtime/service handoff. The preflight uses captured `whoami`. On the `managed_job` execution surface, missing authentication returns typed `worker_authentication_required` with `sideEffectsStarted=false` and never starts login; the operator authenticates Wrangler in an ordinary owner terminal before retrying. On an ordinary local terminal, the wrapper may run interactive login while the existing service is still untouched, then requires a second captured `whoami`. The inner Worker deployment layer independently rejects interactive login whenever `args.json` is true, closing the preflight-to-handoff authentication-expiry race; ordinary non-JSON startup still supports login but also re-verifies authentication before deploy. Tests fault-inject all four branches, and architecture checks require the outer preflight to precede the persistent release-runtime lock and the JSON guard to precede login.
|
|
88
|
+
|
|
89
|
+
**Release consequence.** The defect was discovered after beta.96 acceptance, but the repair changes packaged release tooling and local runtime behavior. Under the prerelease invariant, beta.96 acceptance cannot authorize these bytes. Beta.97 therefore requires a fresh exact-tree fast/full receipt, candidate, install-only preflight, new explicit live-activation authorization, managed-job activation proof, activated-package OAuth canary, observed live verification, and acceptance before any commit/push/publication decision.
|
|
90
|
+
|
|
91
|
+
## 2026-08-17 beta.96 independent review-report adjudication and second-pass hardening
|
|
92
|
+
|
|
93
|
+
**Report adjudication, not checklist acceptance.** A supplied review was written against beta.94 and mixed accurate observations with stale or unsupported conclusions, so each claim was re-tested against the beta.95 working tree before modification. The claimed relay timer leak/double-resolution mechanism was not present: `RelayConnection` already clears connect/handshake/readiness/reconnect/outage timers on the relevant stop/close/fatal paths, heartbeat start is stop-before-rearm, socket handlers bind the captured socket generation, and the first-readiness resolver is nulled immediately after settlement. The claimed unbounded pending/waiter memory path was also not present: the Worker pending map and readiness waiters consume the same thirty-ordinary-plus-two-reserved admission algebra, terminal/timed-out records are removed, detached calls remain capacity-accounted, and Durable Object alarm/event-boundary sweeps expire due records. `outageCount` is a scalar diagnostic counter, not a retained queue. The claimed missing 1002/1008/1009/1012 WebSocket mapping was contradicted by the existing explicit close-category/protocol tests. Changelog size and the large Computer Use/application modules remain maintenance observations rather than demonstrated defects; existing executable line ceilings are retained instead of introducing an unmotivated archive or Strategy-pattern rewrite.
|
|
94
|
+
|
|
95
|
+
**Actual relay lifecycle defect found while falsifying the timer claim.** A minimal current-source reproduction started a relay with a socket that never reached readiness, called `stop()`, and raced the returned `start()` Promise against a short timer; the Promise remained `pending`. Timer cleanup therefore worked, but the first-start settlement contract did not. Beta.96 captures the pending resolver before shutdown, clears both resolver references, performs the existing timer/socket cleanup, and settles that start as no-readiness. The runtime layer additionally checks lifecycle ownership after the await and in the rejection path, so a concurrent `stop()` that has moved the runtime to `stopping`/`stopped` cannot be overwritten by a late `markRunning()` or `markFailed()`. Regression tests prove stop-before-ready settles and leaves no scheduler tasks, while a late relay-start rejection cannot replace a stopped runtime lifecycle.
|
|
96
|
+
|
|
97
|
+
**Deadline arithmetic: public paths were already bounded; one internal fail-closed gap was real.** Public foreground timeout parsing, Worker configurable-tool validation, `read_process.wait_ms`, daemon-ready grace, and relay-contract budgets already reject or clamp ordinary malformed user input. Relay JSON cannot represent `NaN`/`Infinity`. However, direct `PendingCallRegistry`/`PendingCallDeadlines` use still normalized with `Math.floor(Number(value) || 1)`, so an internal `Infinity` produced both an infinite timer delay and an infinite next-alarm delay. Beta.96 centralizes strict pending-delay validation: operation deadlines are bounded by `maximumRelayToolTimeoutMs`, reconnect grace by the shared reconnect contract, and the timer layer independently rejects non-finite/non-integer/non-positive/overflow values. Invalid timing is rejected before a call is detached or registered, preserving pending state. This is defense in depth against an internal contract regression, not evidence that a remote client could previously encode JSON `Infinity`.
|
|
98
|
+
|
|
99
|
+
**Relay envelope validation tightened where the report was directionally right.** The local relay parser already bounded message size, required a plain argument record, validated account/client/family identifier shapes, normalized account roles, and capped timeout values. It was nevertheless too permissive about shape: an arbitrary short string could be a call ID, a syntactically arbitrary tool string could reach dispatch, numeric authorization versions could be coerced from strings, and missing/string/negative/over-limit `timeout_ms` values could be defaulted or clamped. Beta.96 aligns call IDs with the existing `call_...` resume grammar, requires a lowercase bounded tool-name grammar, keeps authorization version/role strictly typed, and requires an explicit integer timeout inside the shared 1–50-second relay envelope. Unknown but syntactically valid tools still fail at the ordinary tool authorization/dispatch boundary; catalog membership is deliberately not duplicated in the transport parser.
|
|
100
|
+
|
|
101
|
+
**Logging/privacy claim narrowed to the real boundary.** JWT redaction was already present in `sanitizePortableLogText`, synthetic JWT coverage already exercised free-form local/Worker logging, and repository privacy scanning already rejected JWT-shaped bearer values. Compact DPoP proofs use the same JWS/JWT shape and therefore traverse that existing redaction. No production path was found logging `cf-connecting-ip` or a remote socket address; the rate-limiter hashes that address before using it as a state key. A generic IPv6 regex was therefore rejected: it would add false-positive/corruption risk without closing an observed leak. Privacy documentation now makes the design explicit—compact JWT/JWS/DPoP-shaped values are sanitized, while client/network identity should be omitted or coarsened at the source rather than entrusted to a universal IP scrubber.
|
|
102
|
+
|
|
103
|
+
**Silent-error review found a different maintenance gap.** The named `daemon-sockets.ts` and relay transport paths already classify meaningful transport failures; the report's example location was inaccurate. Repository-wide review did find uncommented best-effort/fail-closed catches in browser DevTools cleanup, snapshot preflight, extension reconnect/UI decoration, application cleanup/probing, release temp cleanup, diagnostics, audit shutdown, and macOS helper cleanup. Those paths now carry explicit comments explaining why the primary settlement remains authoritative instead of emitting noisy or privacy-risking raw exception text. Repository hygiene rejects newly introduced empty `catch {}` and empty Promise `.catch(() => {})` bodies under `src/`, `scripts/`, and `browser-extension/`. The browser extension's intentionally empty WebSocket `onerror` now documents that `onclose` owns handshake failure/reconnect settlement.
|
|
104
|
+
|
|
105
|
+
**Independent second-pass drift finding.** The previous beta.95 review corrected README/architecture/operations timing, but a fresh cross-surface search found two current-state leftovers that its architecture gate did not cover. `docs/TESTING.md` still said ordinary daemon tools used up to 30 seconds, and `src/shared/server-metadata.json`—which is delivered as server guidance rather than being passive prose—still described ordinary tools and remote process one-shots with the old 30-second foreground model. Beta.96 updates both and expands the release-contract guard. Shared guidance now distinguishes ordinary one-step durable process tools, which already return a `job_id` and should be observed with `read_job`, from owner-only `start_job` for multi-step/resource/finally workflows. This is a concrete contract correction; the historical beta.91/beta.89 audit entries remain untouched as historical evidence.
|
|
106
|
+
|
|
107
|
+
**Release consequence.** These fixes change packaged local runtime, Worker timing validation, browser extension/script bytes, shared server metadata, tests, and documentation after a beta.95 exact-tree candidate had already been prepared locally. That beta.95 receipt/candidate is therefore stale by construction. The package is advanced to beta.96 and requires a fresh targeted/fast/privacy/full/candidate cycle. No live beta.96 activation, Worker deployment, acceptance, Git commit, push, npm publication, or GitHub release is implied by this source review.
|
|
108
|
+
|
|
109
|
+
**Verification-carrier boundary found by the fresh full gate.** After the report-driven fixes, one frozen beta.96 `check:full` passed all 130 tasks in 596.0 seconds—only about four seconds inside the 600-second detached execution ceiling of remote `run_process`. The dominant variance was `install:test` at 229.7 seconds; the run itself did not classify why that install was slower than earlier observations, so no network or registry cause is asserted. This exposed an orchestration hazard rather than a product-test failure: a legitimate full suite can outlive the ordinary one-step remote process budget. Guidance now keeps ordinary one-step work on the durable process tools only when it fits 600 seconds and explicitly routes owner-authorized remote full verification (or another legitimate >600-second step) through detached `start_job` with a larger step timeout. The final beta.96 full receipt must therefore be regenerated after this guidance change using the wider carrier; the suite's own contents and pass criteria remain unchanged.
|
|
110
|
+
|
|
111
|
+
## 2026-08-17 beta.95 independent reply-safety and privacy review
|
|
112
|
+
|
|
113
|
+
**Independent contract review.** The beta.94 runtime behavior remained consistent with its incident repair: relay-origin `start_process` performs zero-queue cooperative admission, ordinary hosted tools default to 20 seconds plus a distinct five-second settlement margin, durable process one-shots separate ten-second acceptance from up-to-600-second execution, public HTTP/SSE delivery remains request-scoped, and same-daemon relay reconnect consumes the original absolute pending-call deadline rather than extending it. The review found no new SSE replay/session requirement and no evidence that the removed 30-second remote-process foreground model still governed a production relay route.
|
|
114
|
+
|
|
115
|
+
**Drift removed.** Two implementation/documentation leftovers could have reintroduced misleading behavior later. `daemon-ready-waiters.ts` still carried a private ten-second fallback even though the authoritative relay contract had moved new-call recovery to five seconds, and the shared contract plus `server_info` still exposed a 30-second process-foreground ceiling even though all remote `exec_command`/`run_process`/`run_local_command` calls route through durable job acceptance. Beta.95 derives the waiter fallback and request-owned `start_process` budget from `relay-contract.json`, removes the obsolete process-foreground field from the shared/public contract, and reports the explicit process-session startup envelope instead. The lower process service intentionally retains a private 30-second defense ceiling so an accidental internal bypass of durable relay routing still fails closed; that ceiling is not advertised as a supported remote delivery mode. README/architecture/testing text that still described the beta.93 timings, plus an unrelated stale 1.5-second host-snapshot freshness claim where code uses 500 ms, is corrected and guarded by architecture checks.
|
|
116
|
+
|
|
117
|
+
**Independent resource-capacity defect.** Live review found one interactive resource waiter that had remained pre-spawn for several minutes even though host pressure was green and there were no leases. Its fixed request was `pytest -n 8` on an eight-core host. Pure policy reproduction proved the request still returned `cpu_pressure_window` with zero observed CPU because the interactive launch policy intentionally retains one core of headroom: the best-case launch window is therefore below eight cores. That condition can never recover by waiting, and the old `selectedResourceWaiter` starvation exception could also treat an empty-lease `cpu_pressure_window` as potentially feasible. Beta.95 classifies fixed CPU requests above the best-case priority-specific launch window as `cpu_request_exceeds_launch_window`, fails them before the retry sleep loop with `retryable=false`, preserves explicit argv/concurrency unchanged, and leaves elastic/unbounded fitting plus genuinely transient CPU-window pressure unchanged. Regression coverage proves an eight-worker fixed request is rejected on an idle eight-core host, a seven-worker interactive request remains admissible, the impossible waiter cannot reserve protected drain, and the public process boundary advises reducing explicit parallelism instead of advertising a temporary retry.
|
|
118
|
+
|
|
119
|
+
**Verification idle-sleep defect.** A later frozen-tree `check:fast` began at `2026-08-17 12:27:33 +0800`; `pmset` records Idle Sleep at `12:27:39` for 594 seconds and DarkWake at `12:37:33`. The fast runner then reported `worker-types-generator:test` after about 593.7 seconds because a fixture that normally exits with code 7 instead reached its two-second Wrangler hard timeout immediately after wake. The failure duration therefore matched host suspension rather than active test execution. Beta.95 does not relax those short lifecycle timeouts. Instead the macOS verification entrypoint re-executes the complete fast/platform/full run under `/usr/bin/caffeinate -i`, preventing ordinary Idle Sleep for the verification lifetime while preserving the same task semantics and exit status. The guard is verification-only, recursion-bounded, absent on other platforms, and does not treat explicit/lid-close sleep as a supported verification environment.
|
|
120
|
+
|
|
121
|
+
**Relay/sleep evidence correction.** During this review the live beta.94 relay reported close `1006` at `2026-08-17T03:56:46.879Z` and was ready again at `03:56:51.232Z`. The launchd daemon stayed on the same verified PID with one lifetime run and no termination signal; event-loop lag remained bounded. A targeted `pmset` check then showed `Wake from Deep Idle` at `2026-08-17 11:56:46 +0800`, followed by normal wake activity. This event is therefore a sleep/wake transport reconstruction, not evidence of a daemon crash or hosted-deadline regression. A later outage began at `2026-08-17T04:21:16.022Z`, was classified as `relay_transport_error` / `network_error` with close `1006`, and recovered in about 5.5 seconds; `pmset` recorded another Deep Idle wake at `12:21:07 +0800`, roughly ten seconds before that transport failure. That second event is strongly wake-adjacent but is recorded more conservatively as a post-wake network-recovery interruption rather than a same-instant wake disconnect. Neither later event invalidates the distinct earlier beta.94 incident evidence that host-pressure reply queueing and a separately observed awake relay interruption were independent mechanisms.
|
|
122
|
+
|
|
123
|
+
**Privacy and observability review.** The beta.94 acceptance record contains only fixed release metadata and cryptographic digests; it contains no machine path, account identifier, endpoint, command/output content, or credential. Operational logs continue to omit arguments/results. To preserve the incident discriminator after pressure has recovered, resource-admission failures may add only an allowlisted reason and coarse pressure state to the existing debug-only tool failure event. Unknown or future reason text collapses to fixed fallback values rather than becoming free-form log data. No new default-level event, path, process name, sampled command output, account identity, or resource identifier is recorded.
|
|
124
|
+
|
|
125
|
+
## 2026-08-17 beta.94 hosted reply headroom under host pressure
|
|
126
|
+
|
|
127
|
+
**Observed incident and falsified alternatives.** A live hosted-call investigation reproduced local resource-admission failures while a heavy Xcode/Simulator workload kept the macOS host awake: the coordinator reported red load backlog and returned both `host_pressure_red` and `coordinator_busy`. The Machine Bridge daemon itself was not CPU-bound. The relay close-1006 episode examined during that beta.94 incident also occurred while the host was awake and recovered in a few seconds, so sleep/wake did not explain that incident-window transport event. The installed service was a verified launchd daemon, the private Wrangler toolchain was healthy, and the Worker's call counters did not show a Worker execution-timeout defect; the foreground-candidate/Wrangler hypothesis was therefore rejected rather than patched. Later sleep/wake-correlated relay interruptions are recorded separately rather than rewriting this historical evidence.
|
|
128
|
+
|
|
129
|
+
**Causal repair.** Beta.93 made remote process one-shots durable, but `start_process` intentionally remained a request-scoped daemon-lifetime session. Its local process-session manager still inherited the owner-local ten-second cooperative resource wait, so a hosted request could queue under known host pressure before spawn even though the caller could safely receive a retryable pre-spawn error instead. Relay-origin `start_process` now uses a zero-queue admission policy: it performs the normal host/capacity check once and, if not admitted, returns the existing retryable `unavailable` classification with `side_effects_started=false` at the Worker boundary. Local/stdio callers keep the ten-second wait, and explicit configured waits remain authoritative. The Worker also reduces `start_process` to a ten-second execution plus five-second settlement envelope, ordinary daemon-backed defaults to 20+5 seconds, and configurable browser/application defaults to 20 seconds without reducing their explicit 45-second maximum. New calls wait only five seconds for daemon readiness; in-flight reconnect still consumes the call's original absolute deadline and can rebind to the same daemon instance.
|
|
130
|
+
|
|
131
|
+
**Transport boundary retained.** The relay already retains completed daemon results until Worker acknowledgement and replays those results after a same-instance reconnect; the Worker retains detached pending calls through the bounded reconnect interval and sends `resume_calls`. The close-1006 path therefore did not justify a second MCP terminal-result store, replay identifier, or cross-request response session. Public HTTP/SSE closure remains cancellation of transient request-owned work, while operations that must outlive delivery use managed jobs. This repair reduces the probability that Machine Bridge itself consumes the Host's reply window; it cannot make a disconnected Host render a terminal frame, and `host_terminal_receipt_observable=false` remains the explicit external boundary.
|
|
132
|
+
|
|
133
|
+
## 2026-08-16 beta.93 durable process settlement and disconnect recovery
|
|
134
|
+
|
|
135
|
+
**Remaining lifecycle defect.** Beta.91 bounded hosted foreground calls and made reconnect consume the original absolute deadline; beta.92 hardened activation rollback. Those changes reduced how long a broken call could occupy the host, but they did not change the ownership model of ordinary remote process one-shots: `exec_command`, `run_process`, and `run_local_command` still spawned work inside the current MCP call. A relay interruption, daemon replacement, host deadline, or cancellation could therefore terminate the child or discard its final result even when the underlying task was otherwise valid. During this repair's targeted verification the ready daemon actually observed a close-1006 relay interruption at `2026-08-16T05:44:26.304Z` and returned ready at `05:44:29.076Z`; the detached managed verification job continued through that transport loss. The event is direct evidence that task correctness cannot depend on an uninterrupted response channel.
|
|
136
|
+
|
|
137
|
+
**Durable-first process contract.** Remote `exec_command`, `run_process`, and `run_local_command` now validate their original tool/cwd/argv/registered-command authority first, then commit the operation as a principal-bound one-step managed job before execution. The MCP request owns only a 10-second acceptance window plus the Worker's separate five-second settlement margin; the detached step has an independent 600-second maximum and returns a `job_id` with `read_job`/`list_jobs` recovery metadata. Remote calls require a caller-held `idempotency_key` before daemon dispatch, so even loss of the acceptance response leaves the caller with a credential that can be retried against the same retained job instead of duplicating execution. Cached hosts that omit the new field receive a normal no-side-effect tool error requesting a schema refresh rather than a protocol-only validation failure. The durable step uses interactive resource-admission priority so persistence does not silently demote an interactive tool call behind unrelated background builds. `start_process` remains the daemon-lifetime session path for interactive stdin; explicit `start_job` remains the owner-only multi-step/finally/resource API.
|
|
138
|
+
|
|
139
|
+
**Authority and isolation preservation.** Automatic persistence does not turn operator process authority into owner managed-job authority. The internal durable path authorizes the original source tool, persists the existing principal binding, remains visible/cancellable through the existing account ownership rules, and never exposes `start_job` to operators. Non-owner process execution also keeps the verified macOS delegated workspace sandbox. The runner persists only a boolean delegated-execution requirement and reconstructs the sandbox wrapper at the actual step spawn using the job's own runtime directory, so daemon replacement cannot make the sandbox depend on an expired daemon temp path. Missing or failed sandbox verification still fails closed.
|
|
140
|
+
|
|
141
|
+
**Diagnostic boundary repair.** `diagnose_runtime` previously used resource-admitted public process paths for its fixed local spawn/shell probes. A long resource-admission wait could therefore make the diagnostic report `local-process-spawn=false` even though the OS spawn boundary was healthy. The fixed spawn and shell probes now use the implementation-owned fixed-process boundary and bypass cooperative resource admission; resource pressure/waiters remain reported independently. This prevents queueing from being mislabeled as an OS/process failure.
|
|
142
|
+
|
|
143
|
+
**Activation ownership transaction.** Beta.89's execution-surface guard remains intact: managed-job runner launch still overwrites the environment with the `managed_job` marker, so persistent activation may run only from the existing durable release surface or an ordinary local terminal. Beta.92's delayed launchd stop observation and previous-service rollback retries remain. Beta.93 additionally moves `service_owner` commit out of definition installation: activation keeps the candidate owner pending through provider start and exact daemon/Worker convergence, then commits under the still-held machine-service lock. Recovery carries the same pending identity; failed recovery stops an uncommitted candidate, while owner-commit failure stops the verified provider and leaves the pending record fail-closed. This avoids both a false committed-but-inactive owner and an unsafe owner-only rollback after the provider definition may already have changed. No live `release:candidate:activate` is required to validate this lifecycle repair.
|
|
144
|
+
|
|
145
|
+
## 2026-08-16 beta.92 activation stop-settlement rollback
|
|
146
|
+
|
|
147
|
+
**Live incident.** The exact beta.91 candidate passed 130/130 local full verification and disposable installation, then was activated through the required durable managed-job surface. The managed activation step failed after about 9.1 seconds with `the existing autostart service could not be stopped before activation (launchd)`. The Worker remained beta.90, but the previously ready local daemon was no longer connected and did not return until approximately 17 minutes later. This proved that the stop operation's reported failure was not equivalent to “no side effect”: a launchd `bootout` can still settle after the helper's short observation window, while the immediate one-shot rollback start can race that in-progress removal and fail to restore persistence. The durable job prevented duplicate activation and retained the terminal error, but availability recovery was too weak.
|
|
148
|
+
|
|
149
|
+
**Causal repair.** `stopLaunchdService()` now observes safe unload for up to about ten seconds rather than the generic two-second service-convergence window. More importantly, once a bootout is dispatched against a service that was active at preflight, the result carries `restore_required=true` even if post-mutation status is unavailable or still loaded. Persistent activation already records the verified previous runtime before mutation; beta.92 uses that evidence to retry the idempotent previous-provider start/bootstrap operation and inspect the exact previous service-mode daemon after each attempt. Recovery is bounded to 30 attempts. An activation stop failure receives stable code `autostart_stop_failed`; if rollback converges, that original failure remains the candidate result, and if rollback cannot converge its error is aggregated rather than hidden.
|
|
150
|
+
|
|
151
|
+
**Related reply reliability.** The live incident also reproduced beta.90's old pending-call behavior: calls detached by the daemon shutdown remained pending far beyond their original foreground budget before clearing. Beta.91 already fixes that separate mechanism by preserving the original absolute pending-call deadline across detach/rebind, caps remote synchronous execution and process polling, and emits five-second SSE heartbeats. Beta.92 retains those changes. Historical relay logs also contain long outage durations, but code review confirms reconnect backoff itself tops out at roughly 15.5 seconds; long gaps can include OS/event-loop suspension or network/VPN unavailability, so increasing reconnect frequency was rejected as an unsupported fix. Current heartbeat recovery already forces a stale relay rebuild after a sufficiently long event-loop stall.
|
|
152
|
+
|
|
153
|
+
## 2026-08-16 beta.91 hosted-reply settlement closure
|
|
154
|
+
|
|
155
|
+
**Remaining blank-response mechanism.** Beta.90 fixed daemon-replacing activation ownership and narrowed process one-shots, but an independent request-lifecycle review found two remaining ways a valid hosted tool call could monopolize the assistant reply window. First, tools without a configurable `timeout_seconds` field still received a 60-second daemon execution budget, and `read_process` could block a single remote poll for 30 seconds. Second, when the daemon disconnected during an in-flight call, `PendingCallRegistry` paused the operation timer and armed the separate 120-second daemon reconnect grace; a successful same-instance rebind then restarted the saved remaining operation budget. The reconnect design preserved execution opportunity but silently enlarged the host-visible call lifetime, so a short call could remain unresolved for minutes and the external host could time out the whole assistant turn without displaying a final message.
|
|
156
|
+
|
|
157
|
+
**Repair and boundary.** Hosted synchronous execution now has a separate reply-safe projection instead of reusing the local administrator envelope. Ordinary daemon-backed tools receive at most 30 seconds; configurable browser/application foreground tools are capped at 45 seconds; process one-shots default to 20 seconds and are capped at 30 seconds; `read_process` waits at most five seconds per hosted poll and receives only a 5–10-second execution budget; `start_process` receives 20 seconds to establish the retained session. Local stdio/CLI behavior is unchanged. The public SSE proxy heartbeat moves from ten to five seconds. More importantly, detach/rebind preserves the original absolute pending-call deadline: reconnect grace is capped by the remaining operation budget and reconnect time is no longer paused, so daemon recovery cannot extend a current HTTP/SSE request beyond the deadline it had before the outage. Durable work still belongs in managed jobs, which are intentionally independent of one request or daemon lifetime.
|
|
158
|
+
|
|
159
|
+
**Rolling-upgrade and residual limitation.** Hosts may cache old tool schemas. Maximum-only version skew for both `timeout_seconds` and `read_process.wait_ms` therefore returns a normal MCP `isError` result with `invalid_request`, `side_effects_started=false`, and schema-refresh/session/job guidance; malformed types and unrelated schema failures remain protocol errors. `server_info.tool_delivery` exposes the ordinary, configurable-foreground, process, poll, and settlement ceilings for diagnosis. Machine Bridge still cannot observe a host UI's terminal-frame consumption (`host_terminal_receipt_observable=false`), so no server-side timeout can mathematically guarantee that a particular client renders the final assistant message. The repair instead removes known server-side paths that could unnecessarily consume or enlarge that external budget and makes long work recoverable through sessions/jobs with short polling.
|
|
160
|
+
|
|
161
|
+
## 2026-08-16 beta.90 stale-host timeout-schema compatibility
|
|
162
|
+
|
|
163
|
+
**Live finding.** After beta.89 was activated successfully through a durable managed job and its activated-package OAuth canary passed, a deliberate live probe exercised the new process ceiling through the same ChatGPT connector that had opened the conversation before beta.89 deployment. The connector's local action snapshot still described `run_process.timeout_seconds` as maximum 60, so it accepted a request for 46 seconds. The beta.89 Worker had already published the correct 45-second schema and rejected the call before daemon dispatch, but returned the normal JSON-RPC `-32602` used for input-schema violations. The connector translated that protocol error into a generic `UNKNOWN: ExceptionGroup` rather than preserving the useful validation data. No child process ran, but the user-facing settlement was materially worse than the intended explicit long-task routing guidance.
|
|
164
|
+
|
|
165
|
+
**Causal boundary.** Tool schemas are cacheable host metadata; a Worker deployment cannot make every already-open host/tool snapshot refresh atomically. Therefore lowering a server-side maximum creates a temporary compatibility interval in which a stale client can construct an argument the current Worker knows is invalid. Treating that expected version-skew condition identically to a malformed request delegated presentation quality to every connector's JSON-RPC error mapping. The problem was not the 45-second limit itself and not a daemon failure; it was an incomplete rolling-upgrade contract between cached tool metadata and current server validation.
|
|
166
|
+
|
|
167
|
+
**Repair.** `McpController` now recognizes the narrow case where all validation issues are `/timeout_seconds` with keyword `maximum`. It returns a successful JSON-RPC envelope containing a standard MCP tool result with `isError=true`, structured `invalid_request`, `side_effects_started=false`, `schema_refresh_recommended=true`, the bounded validation issues, and instructions to refresh `tools/list` or move longer work to process sessions/managed jobs. The daemon is never called. All other input-schema failures still use `-32602`; tests explicitly keep malformed timeout types, unknown fields, missing names, and unknown tools on the protocol-error path. This gives new clients the strict schema and old clients a useful fail-closed settlement without widening execution authority or accepting the obsolete timeout.
|
|
168
|
+
|
|
169
|
+
**Release consequence.** beta.89 proved the durable daemon-replacement fix and canary path, but this live connector compatibility defect changes packaged Worker behavior after beta.89 activation. beta.89 is therefore superseded without acceptance; beta.90 requires fresh full verification, exact candidate preparation, durable activation, activated-package OAuth canary, and live stale-schema verification before acceptance.
|
|
170
|
+
|
|
171
|
+
## 2026-08-16 beta.89 daemon-handoff and long-task settlement repair
|
|
172
|
+
|
|
173
|
+
**Live activation failure and causal mechanism.** beta.88 passed its local frozen-tree verification and candidate install preflight, then persistent activation was launched through a Machine Bridge process session. The Worker updated to beta.88, but the old login daemon disappeared and `machine-mcp service status` subsequently reported a committed beta.88 service owner with launchd `loaded=false`, `active=false`. Code review established the causal chain rather than treating this as a generic launchd flake: process-session children are registered in the current daemon's `ProcessTracker`; activation intentionally calls `stopAutostart()` during foreground-to-service handoff; `LocalRuntime.stop()` drains tracked process groups with `SIGKILL`; on POSIX the process-tree implementation targets the child's process group. The old daemon could therefore terminate the very activation transaction that was replacing it, after service-definition/owner mutation but before `startAutostart()` completed. Manual `machine-mcp service start` recovered the already-committed beta.88 owner and restored a ready beta.88 daemon, confirming that the broken state was the handoff boundary rather than candidate package identity.
|
|
174
|
+
|
|
175
|
+
**Repair.** Foreground one-shot processes, daemon-lifetime process sessions, and detached managed-job runners now receive an internal execution-surface marker after environment construction. Persistent release activation inspects that marker before candidate installation, Worker deployment, or service mutation. `foreground_process`, `process_session`, and any unknown non-empty surface fail closed with a no-side-effects error; an ordinary local terminal and the detached `managed_job` surface remain valid. The managed runner overwrites any inherited/spoofed marker even when full environment inheritance is requested. This deliberately reuses the existing durable runner, whose detached/unreferenced process and persisted job state already survive MCP loss and daemon restart, rather than creating a second release-specific detachment mechanism.
|
|
176
|
+
|
|
177
|
+
**Long foreground calls and blank-response risk.** A separate review found that remote process tools could request the complete 60-second interactive execution window while Worker settlement was only budgeted after that execution deadline. In practice, resource admission, transport delay, process termination settlement, or host response overhead near that boundary can consume the remaining UI/tool-call envelope; the user then experiences a terminated task and, in the worst case, no visible assistant reply. Remote `exec_command`, `run_process`, and `run_local_command` now default to 30 seconds and reject requests above 45 seconds before spawn. The general 60-second interactive ceiling remains for browser/application operations that have their own bounded settlement logic. Work expected to exceed the process foreground budget must use `start_process`/`read_process`; work that must survive daemon replacement or MCP disconnect must use `start_job`/`read_job`. The design intentionally does not auto-detach an already-started one-shot command at timeout, because silently changing one-shot ownership/cleanup semantics after side effects begin would create a different ambiguity class.
|
|
178
|
+
|
|
179
|
+
**Additional boundary repair.** Execution-surface regression coverage instantiated `ProcessSessionManager` without a resource coordinator and exposed a latent contract defect: `acquireProcessResources(null, ...)` returned environment metadata but dropped the original command/argv, producing an undefined executable at spawn. The no-coordinator path now preserves command/argv exactly. Dedicated resource-admission tests remain responsible for real admission behavior; standalone process-session composition no longer accidentally depends on a coordinator being present.
|
|
180
|
+
|
|
181
|
+
**Nested resource-wait finding.** A subsequent full run reached `full-access:test` and then waited 331.9 seconds before its internal managed job timed out in `running`. Runtime diagnostics showed the outer `check:full` holding a verification-plan heavy lease while multiple older background waiters occupied the machine-user queue. The full-access fixture already supplied a private coordinator root and synthetic host sampler to its `LocalRuntime`, but that root was used only by foreground/process-session execution; `ManagedJobManager` launched its detached runner from the ambient environment, so the inner job rejoined the global queue. `LocalRuntime(resourceCoordinatorRoot=...)` now propagates that root as a runner-environment override for both normal managed-job launch and recovery relaunch. Production runtimes that do not override the coordinator retain the default machine-user namespace. Re-running `npm run full-access:test` under the same outer Machine Bridge process carrier completed immediately, proving the nested queue dependency was removed rather than masked with a longer timeout.
|
|
182
|
+
|
|
183
|
+
**Release consequence.** beta.88 cannot be accepted or reused: live mutation occurred, the activation did not complete normally, and this repair changes packaged source, release tooling, Worker timeout metadata, tests, documentation, and synchronized version surfaces. beta.89 therefore requires a fresh full receipt, exact candidate, install-only verification, durable-job activation, activated-package OAuth canary, live service/Worker/process-settlement verification, and a new acceptance record before any guarded source push.
|
|
184
|
+
|
|
185
|
+
## 2026-08-15 beta.88 workflow-surface terminology closure
|
|
186
|
+
|
|
187
|
+
**Authorization-layer attribution review.** A nested SSH maintenance probe returned an explicit remote forced-command allowlist after Machine Bridge had successfully spawned the local SSH client. Treating that result as a Machine Bridge policy denial would have widened the wrong boundary. Current shared instructions, runtime diagnostics, README, and operations guidance now make the evidence rule explicit: pre-response blocks are host-side unknowns, true local spawn failures belong to the local execution boundary, and a child exit code with bounded output proves spawn occurred so any remote allowlist/forced-command refusal must be diagnosed downstream. The change adds no new Machine Bridge authority and avoids using a broader local policy as compensation for a narrower target-side restriction.
|
|
188
|
+
|
|
189
|
+
**Module-responsibility review.** `computer-use.mjs` had accumulated expectation-schema normalization, intrinsic application-state expectations, and Accessibility check/uncheck target validation alongside orchestration, dispatch, post-observation, and settlement. Those pure validation responsibilities now live in `computer-use-expectation.mjs`; the architecture inventory treats it as a first-class boundary and lowers the parent module line ceiling. This is a structural extraction only: existing Computer Use behavior tests remain authoritative for browser/application edge cases and ambiguous mutation settlement.
|
|
190
|
+
|
|
191
|
+
**CLI responsibility review.** `cli.mjs` was at 947 lines under a 950-line architecture ceiling even though its ready/start presentation was a pure formatting concern. Start JSON, the human/JSON MCP-ready banner, initial-owner one-time credential display, and policy-summary formatting now live in `cli-ready-output.mjs`. CLI entrypoint/service regressions preserve the observable behavior, the parent falls below 900 lines, and the architecture contract lowers its ceiling to 900 while budgeting the new presentation module separately. `runtime.mjs` remains near its 700-line ceiling, but its remaining concentration is stateful runtime composition and lifecycle authority; it was not split merely to improve a line-count metric without a clearer responsibility seam.
|
|
192
|
+
|
|
193
|
+
**Resource host-sample generation race.** The full verification gate exposed a cross-process cache race rather than a policy failure: `host-sample.json` is deliberately replaced atomically, while a concurrent reader opens the old generation and then verifies the pathname identity. If another coordinator publishes between those two observations, `secure-file` correctly raises `MBM_IDENTITY_CHANGED`; `freshHostSnapshot()` previously propagated that expected publication race as an `execution_failed`. Host-sample reads now retry only this identity-change class for four fresh opens. The path-identity check remains enabled on every attempt, hard links and symlinks remain rejected, malformed JSON is not retried, and persistent generation churn still fails closed. Deterministic tests cover both recovery and exhaustion.
|
|
194
|
+
|
|
195
|
+
**Structured file mutation generation race.** A separate review of local file deletion found a narrower time-of-check/time-of-use defect in `commitPatchTransaction()`: source content was hashed before commit, then the pathname was renamed to an internal backup, but the moved generation was not revalidated before replacement publication or backup cleanup. An external process could therefore replace or mutate the source in the interval after the preflight hash and before the rename; a delete could then remove that newer generation, and an update/move could treat it as the observed source. The transaction now reads and hashes the exact backup generation immediately after rename. A mismatch enters the existing rollback path before any target publication; if rollback itself cannot restore the pathname because another actor has reoccupied it, the existing non-retryable `patch_recovery_incomplete` settlement preserves the ambiguity instead of deleting competing state.
|
|
196
|
+
|
|
197
|
+
**Optimistic write convergence.** `edit_file` always writes against the SHA-256 of the text it just read, and `write_file` may accept an explicit `expected_sha256`; both previously used a second check followed by overwrite-capable rename, leaving the same final check-to-rename window. Expected-hash atomic writes now delegate to the generation-safe patch transaction, so patch updates/deletes and optimistic single-file writes share one quarantine, post-rename hash verification, no-overwrite publication, rollback, and cleanup-warning implementation. Unconditional `write_file` without a precondition remains intentionally last-writer-wins. Deterministic fault injection mutates the source inside the injected rename boundary and requires both delete and expected-hash write to reject while preserving the concurrent generation.
|
|
198
|
+
|
|
199
|
+
**Release verification reuse.** The prerelease handoff previously ran the complete 129-task local gate and then `release:candidate` immediately invoked the complete gate again before packing. That second run added several minutes but did not create an independent authority or environment: both executions observed the same owner machine and same source tree. The workflow now persists a private `.project-local/full-verification.json` receipt only after a stable `check:full` completes. The full-run guard still watches generated `.release-candidate` state so concurrent candidate regeneration invalidates the run, while the reusable receipt binds the exact verified source/configuration generation and deliberately excludes that next-stage generated candidate directory. It also binds package name/version, Node version, platform/architecture, and a six-hour freshness window. Starting any new full run deletes the old receipt before executing tests, so an observed failure or interrupted rerun cannot fall back to prior success. Candidate preparation recomputes the current source generation and rejects missing, stale, runtime-mismatched, or source-mismatched evidence before packing; candidate output or a retry of the pack stage does not by itself force another full run. The optimization therefore removes duplicate execution, not a distinct proof boundary.
|
|
200
|
+
|
|
201
|
+
**Exact-main finding.** beta.87 was fully live-verified and accepted, then its exact PR head passed Ubuntu/full, macOS/platform+install, Windows/platform+install, package audit, Governance, Workflow Policy, Dependency Review, and both CodeQL analyses. The repository permits squash merges only; PR #81 merged cleanly and the resulting `main` tree hash was byte-for-byte identical to the accepted beta.87 PR-head tree. The newly triggered exact-main CI nevertheless displayed `Verify interactive candidate acceptance` for its portable acceptance steps. Repository search found the same retired concept in the portable verifier's success output, the workflow-policy contract's descriptive label, TESTING documentation, and one architecture-test error message.
|
|
202
|
+
|
|
203
|
+
**Correction and regression.** Current workflow step names, verifier output, workflow-policy description, TESTING guidance, and test diagnostics now uniformly call this evidence `local candidate acceptance`. Architecture verification reads the actual CI YAML, portable verifier, workflow policy contract, and TESTING guide, rejects `interactive candidate acceptance` case-insensitively, and positively requires the new terminology. This expands the beta.87 regression from selected release scripts/docs to the complete current CI presentation surface. Historical records that describe the former policy remain unchanged.
|
|
204
|
+
|
|
205
|
+
**Release consequence.** The beta.87 runtime behavior and exact candidate remain valid evidence for the underlying authorization/discovery fixes, but current workflow/docs/tests and synchronized version metadata change in beta.88. The beta.87 acceptance therefore cannot authorize beta.88 publication; the complete candidate/live acceptance cycle must run again.
|
|
206
|
+
|
|
207
|
+
## 2026-08-15 beta.87 current-contract terminology closure
|
|
208
|
+
|
|
209
|
+
**Post-push finding.** beta.86 completed its fresh fast/full/frozen-candidate gates, direct-Node install-only verification, owner-authorized activation, exact packaged OAuth canary, live Worker/service/relay/application/browser verification, acceptance, commit, and guarded push. The guarded-push success line itself still said `Verified interactive local candidate acceptance`. A repository-wide follow-up then found the same obsolete adjective in the portable CI acceptance verifier plus live operational guidance that still described an owner-terminal ceremony, real TTY streams, an agent stopping to hand activation back to the owner, and a failed `owner-terminal` activation command. These are not historical entries; they are current execution/operations surfaces and therefore contradict the policy now implemented by the release guard.
|
|
210
|
+
|
|
211
|
+
**Correction.** Current push/CI diagnostics now call the evidence `local candidate acceptance`. ENGINEERING states the actual publication invariant: explicit current-task owner authorization, the `--owner-confirm` acknowledgement, exact release-integrity gates, and one publication lock holder; TTY presence is optional. RELEASING requires the non-live direct-Node preflight and then permits the explicitly authorized owner or agent to run the exact activation command, while prohibiting an agent from racing a command the owner says they will run. OPERATIONS describes failure by activation outcome rather than terminal shape. Historical changelog/audit passages that accurately describe older owner-terminal releases remain unchanged.
|
|
212
|
+
|
|
213
|
+
**Regression and review lesson.** Architecture/documentation tests now forbid the obsolete current-contract phrases in `scripts/github-push.mjs`, `.github/scripts/verify-release-acceptance.mjs`, ENGINEERING, RELEASING, and OPERATIONS. This corrects a verification blind spot: beta.85's audit recorded that no current operational TTY contract remained, but the post-push observable disproved that claim. The new gate checks both required positive language and explicit stale phrases rather than relying on a partial search. Since current packaged scripts/docs/tests changed after beta.86 acceptance, a new beta.87 candidate and live acceptance are required.
|
|
214
|
+
|
|
215
|
+
## 2026-08-15 beta.86 live effective-tool projection repair
|
|
216
|
+
|
|
217
|
+
**Live finding.** beta.85 completed its final fast/full/frozen-candidate gates, direct-Node install-only check, owner-authorized activation, exact activated-package OAuth canary, and live Worker/service/relay diagnostics. `server_info` correctly reported an authenticated owner with full effective policy and 54 effective tools. A live `list_local_applications` call nevertheless returned `discovery=true` while projecting `open=false`, `accessibility_inspection=false`, `structured_accessibility_actions=false`, and `window_screenshot=false`. The mismatch was therefore inside the local runtime's capability/routing projection rather than the account/Worker authorization layer or the actual ToolExecutor gate.
|
|
218
|
+
|
|
219
|
+
**Causal mechanism.** `buildAuthorityContext()` correctly intersects the daemon and account policies, but that intersection deliberately carries the synthetic profile name `custom` even when its capability bits are identical to canonical full. `runtime.effectiveToolNames()` then re-ran `toolNamesForPolicy()` over that synthetic policy. The `full` availability class intentionally requires the literal canonical `full` profile as well as full capability bits, so the owner projection collapsed from 54 to 37 tools. Actual execution did not use this derived set for authorization: ToolExecutor first applies the daemon `PolicyGate` and then the account-role gate, so the defect primarily corrupted capability/routing metadata and any consumers that trust `effectiveToolNames()`.
|
|
220
|
+
|
|
221
|
+
**Repair.** `runtime.effectiveToolNames()` now derives its result from the same two authoritative executable gates: the daemon PolicyGate's allowed names intersected with the authenticated account role's allowed names. Local calls receive the daemon PolicyGate set directly. This avoids inferring tool availability from a synthetic policy label, preserves every custom-daemon restriction, and makes owner/full projection agree with the 54-tool authorization contract. The application integration test now exercises both relay reviewer and relay owner inventories: reviewer keeps read-only discovery, projected paths, and no launch/UI capabilities; owner must retain launch, Accessibility, and window-observation capabilities.
|
|
222
|
+
|
|
223
|
+
**Release consequence.** beta.85 remains useful live evidence that activation/canary/service/relay paths were healthy, but it was intentionally not accepted after the projection mismatch. The packaged runtime/test/documentation bytes change for beta.86, so all beta.85 candidate evidence is non-transferable to beta.86.
|
|
224
|
+
|
|
225
|
+
## 2026-08-15 beta.85 authorization and read-only discovery review
|
|
226
|
+
|
|
227
|
+
**Owner-requested policy correction.** The beta.84 source had completed live acceptance, merge, and exact-head hosted validation, but no Git tag/GitHub Release/npm publication had occurred when the owner rejected two workflow restrictions as too strict. The first required GitHub publication to have TTY-backed stdin/stdout/stderr even when the owner had explicitly authorized the operation in conversation. That rule was an anti-accident ceremony, not a meaningful same-user authentication boundary: code already executing as the owner's OS user can emulate terminal presence. The new contract therefore treats an explicit current-task owner request as sufficient user authorization and requires the repository command's explicit `--owner-confirm` acknowledgement instead of a TTY. This does not turn version changes into standing publication authority; npm publication, global/registry installation, Worker/service mutation, credential mutation, and source publication still require a specific owner request, and an agent must not race an operation the owner says they will execute themselves.
|
|
228
|
+
|
|
229
|
+
**Publication guard analysis.** Removing TTY presence does not remove the release-integrity state machine. `github-release.mjs` still requires a clean checkout on `main`, fetches the remote, verifies the version channel and changelog, creates an integrity-pinned hardened npm session, runs the complete verification plan and synchronized-version check, verifies exact local candidate acceptance (and soak for stable), requires `HEAD === origin/main`, requires successful CI for that exact commit, stages only the accepted candidate tarball, rejects conflicting local/remote tags, publishes through trusted absolute `git`/`gh` executables, verifies the GitHub asset SHA-256 through REST, waits for release metadata convergence, and serializes publication through the owner-state lock rooted in the common Git directory. Missing `--owner-confirm` fails before fetch or remote mutation. Concurrency, digest, exact-commit, and remote-state guards therefore remain the security/integrity controls; terminal shape is now observational metadata only.
|
|
230
|
+
|
|
231
|
+
**Application-discovery finding.** `resolve_task_capabilities` previously returned `application_discovery.available=false` with `reason=effective_authority` to delegated accounts because `list_local_applications` was classified as `full`. That bundled a low-impact inventory read with high-impact desktop control. The tool only enumerates installed application/launcher metadata and does not read documents or UI contents; restricted policy also projects every external absolute path through `<external-path:…>`. The catalog now classifies it as `always`, and account-access revision 4 records the resulting reviewer/editor/operator capability change. The outer runtime authorization pipeline remains intact: daemon policy is asserted first, relay calls then pass `accountAccessGate.assert(role, tool)` before handler invocation, and operation/audit middleware still runs normally.
|
|
232
|
+
|
|
233
|
+
**Delegated path-projection defect.** The first fast-plan run after widening discovery found a stale integration assertion, and strengthening that assertion exposed a real implementation bug rather than merely an outdated test. `AppAutomationManager.listApplications(args, context)` used `context` for cancellation and scanning but projected both application paths and discovery-warning paths through `this.displayPath(path)` without forwarding the request context. The runtime injects `displayPath` as an authority-sensitive function; omitting `context` therefore fell back to local owner/full policy on a full daemon and could expose an absolute installed-application path to a delegated account. All three application result projections now pass the originating context, including `openApplication`'s resolved path for consistency. The integration fixture executes `list_local_applications` as an authenticated reviewer against both a valid application root and an invalid discovery root, then requires application and warning paths to omit the private temporary root while still discovering the named app.
|
|
234
|
+
|
|
235
|
+
**Capability and routing separation.** Simply widening the catalog exposed a second architectural problem: the old bootstrap returned the raw application backend capability object whenever discovery was permitted, which would have told a reviewer that launch, Accessibility inspection/actions, screenshots, and background visual input were available. A dedicated `application-capability-projection.mjs` now intersects each advertised backend capability with the effective tool set. A further direct-result review found the same raw backend capability object was still returned by `list_local_applications` itself; the runtime now injects the same authority-aware projector into `AppAutomationManager`, so both bootstrap metadata and the direct inventory result agree with the account's effective tool set. `execution-routing.mjs` likewise separates the one-tool `application-discovery` route from the actionable `application` route. Direct regressions prove a discovery-only role receives inventory without UI authority, while full authority preserves the complete backend. A later governance pass deliberately invalidated an in-progress full check after noticing that the new module was only transitively packaged and had no explicit architecture budget. It is now an explicit package-inventory item, architecture boundary module with a 50-line ceiling, strict-JS typecheck root, and 100% function/branch critical-coverage target. Its extraction also preserved the existing 100-line ceiling on `runtime-capabilities.mjs` instead of raising that budget.
|
|
236
|
+
|
|
237
|
+
**Privacy and scope decision.** The wider inventory intentionally reveals installed application names and bundle/launcher identifiers to authenticated delegated roles; those values are now documented as device metadata rather than hidden capability state. Raw external paths remain projected for restricted authority. The review did **not** generalize this change to every read-only tool: `inspect_local_application`, browser tab/source/semantic inspection, and similar reads can expose window text, URLs, authenticated DOM, or other private content, so retaining `full` authority is a privacy boundary rather than unnecessary ceremony. `MULTI_ACCOUNT.md`, application/context/privacy guidance, generated policy reference, and generated 54-tool reference now describe that distinction.
|
|
238
|
+
|
|
239
|
+
**Independent review evidence.** Account authorization, agent-context routing, policy, runtime-handler inventory, runtime-boundary behavior, full local self-test, release-publication guard behavior, release-state diagnostics, architecture/source graph, lint, strict local/Worker typecheck, structured logging, security audit, property-based security tests, current-tree privacy, reachable-history privacy, npm package modes, consumer-package security, and CycloneDX validation were rerun while reviewing this change. Current privacy scanning covers 659 tracked/unignored files; reachable-history scanning covers 3733 blobs and 145 commit messages; package inspection contains 462 files; the consumer artifact has 11 production dependencies/8 SBOM components; the workspace CycloneDX graph has 117 components/118 dependencies. The new source file initially had mode `0600`, which package review identified and corrected to the existing `0644` source contract. No current operational documentation retains the removed `--owner-terminal-confirm`/mandatory-TTY rule; historical changelog/audit entries remain untouched as then-true records.
|
|
240
|
+
|
|
241
|
+
**Release consequence.** These are packaged policy/runtime/tooling/documentation changes, so beta.84's acceptance cannot authorize beta.85. A new beta.85 candidate must run fresh fast/full/frozen-candidate/install-only verification, then be activated and observed with the widened application-discovery contract before acceptance, review, merge, and any explicitly authorized source publication.
|
|
242
|
+
|
|
243
|
+
## 2026-08-15 beta.84 ChatGPT host-routing and raw-browser-source review
|
|
244
|
+
|
|
245
|
+
**Host-routing finding.** The repeated statements that Machine Bridge had been "directly disabled by the platform" were tested rather than accepted as root-cause evidence. The running beta.82 server returned owner/full authority with 54 effective tools, a ready daemon and relay, and healthy fixed filesystem/process/shell/resource probes; it also correctly reported that final host tool exposure is not known to the server. ChatGPT workspace administration showed the `machine-mcp` custom app enabled. One fresh Work conversation launched through the app surface successfully invoked `server_info`; a second fresh Work conversation without an initially visible app pill also successfully routed an explicit `machine-mcp server_info` request. Both returned beta.82 with daemon readiness verified. The two quoted failure statements were independently found in the same older conversation. This disproves a blanket current server/platform disable but does not reveal the host's private pre-dispatch decision. Conversation/surface app routing, a stale per-host action/tool snapshot, host filtering, connector-gateway routing, client routing, or platform policy remain possible host-owned causes until host-side evidence distinguishes them.
|
|
246
|
+
|
|
247
|
+
**Diagnostic repair.** Runtime diagnostics and server instructions now name conversation/surface app-routing state and stale host action/tool snapshots as host-side possibilities for a call that never returns a structured Machine Bridge result. They retain the stronger invariant introduced in beta.79: such a failure is not observable by Machine Bridge and must not be attributed to a platform without evidence. Client guidance adds a differential test: if a fresh supported conversation can invoke the same app/tool while an older conversation fails before dispatch, that is evidence against a blanket Machine Bridge outage and directs investigation to the host conversation/surface/snapshot layer. Workspace app enablement and the live server catalog are documented as distinct from host-cached action definitions; app refresh/review is required after action/schema changes where the host provides that workflow.
|
|
248
|
+
|
|
249
|
+
**Raw browser-source privacy finding.** A raw-source read against an authenticated application during this investigation demonstrated that serialized DOM can carry hidden bootstrap/session/account/authentication values not visible in rendered text. No live value from that observation is copied into this audit, tests, logs, or release evidence. The runtime continues to return raw site-controlled markup by design rather than applying an incomplete token scrubber that would both break source semantics and miss arbitrary application formats. The safer contract is explicit minimization: `browser_get_source` now warns that the entire result is sensitive and is for markup-required tasks only; `browser_inspect_page` or `computer_observe` is the default for semantic/actionability work. Routing tests pin the warning, and privacy/security/local-automation documentation records the same boundary.
|
|
250
|
+
|
|
251
|
+
**Independent review.** The follow-up sweep rechecked empty catches, exception-to-success transitions, structured logging, current/history privacy boundaries, destructive path containment, the beta.83 Windows state-root fix, platform-specific tail fixtures, TODO/FIXME/HACK markers, source/documentation drift, and browser-extension growth ceilings. The reviewed empty catches remain cleanup or fail-closed ambiguity paths rather than swallowed success. No second confirmed mixed-Windows-path destructive containment defect or new Windows tail portability defect was found statically. The large packaged browser-operation/page-automation modules remain a maintainability hotspot but are already bounded by executable architecture ceilings; this release does not refactor them without a concrete behavior defect.
|
|
252
|
+
|
|
253
|
+
**Release consequence.** beta.83 was never owner-activated or accepted. Because these diagnostic, tool-metadata, and documentation bytes are shipped in the package, beta.83 evidence is non-transferable and beta.84 must repeat fast/full/candidate/install-only validation before any owner activation.
|
|
254
|
+
|
|
255
|
+
## 2026-08-15 beta.83 Windows state-root current-CLI containment
|
|
256
|
+
|
|
257
|
+
**Hosted evidence.** The exact beta.82 candidate completed local focused review, the 102-task fast plan, two independent 129-task full/candidate gates, direct-Node install-only verification, owner activation, the activated-package OAuth canary, live Worker/service/relay/browser/privacy verification, exact acceptance, commit, guarded push, and PR update. Exact-head hosted Windows then directly vindicated the beta.81 launchd repair: `service-platform:test` completed successfully in 0.5 seconds. The same run also passed tasks 112-116 before the first failure at task 117 `self-test`. The state phase expected `validateStateRootForRemoval()` to reject a synthetic state root after `process.argv[1]` was moved underneath its release-channel runtime tree, but the current-CLI containment guard returned false.
|
|
258
|
+
|
|
259
|
+
**Causal mechanism.** State-root removal first resolves the existing target with ordinary `realpathSync`. The current-entrypoint guard was added later and canonicalized only the entrypoint through `canonicalizePotentialPath()`, whose existing ancestor is resolved with `realpathSync.native`. On Windows, equivalent drive paths can occupy ordinary and extended/native namespaces. A control experiment using `node:path.win32` proves that `relative("D:\\root", "\\\\?\\D:\\root\\child")` is absolute, while comparing the same two paths after both use the native namespace yields the expected relative child. Feeding the mixed pair into the existing containment predicate therefore skips the live-runtime protection. This is a production destructive-operation safety defect exposed by the hosted fixture, not merely a portability expectation error.
|
|
260
|
+
|
|
261
|
+
**Repair and invariant.** `currentEntrypointInsideStateRoot()` now canonicalizes both `root` and `process.argv[1]` through the same potential-path canonicalizer in one transaction. If that canonicalization fails, both operands fall back together to `path.resolve()`; the guard never mixes canonical path families. The existing synthetic state-root fixture remains unchanged in purpose and gains platform-independent `path.win32` assertions that pin the namespace mechanism. Architecture regression requires root and entrypoint canonicalization plus `path.relative(canonicalRoot, resolved)`. No state marker, reserved namespace, active lock/job/daemon inventory, workspace exclusion, generation-bound quarantine, or recursive deletion rule is relaxed.
|
|
262
|
+
|
|
263
|
+
**Release consequence.** beta.82 remains the currently live accepted runtime, but its exact hosted Windows validation failed and the newly fixed packaged runtime byte invalidates beta.82 acceptance for any subsequent commit. beta.83 must repeat the complete candidate/live evidence lifecycle before another guarded push. The still-unreached Windows tail after `self-test` remains hosted evidence to obtain rather than an assumed success.
|
|
264
|
+
|
|
265
|
+
## 2026-08-15 beta.82 independent recovery-evidence and verification-boundary review
|
|
266
|
+
|
|
267
|
+
**Candidate invalidation and review stance.** The beta.81 launchd stop repair had already passed its targeted regression, the 102-task fast plan, a 129-task full gate, a second frozen 129-task candidate gate, and direct-Node install-only verification. It had not been owner-activated, accepted, committed, or pushed. This review deliberately treated that prepared tarball as an untrusted baseline rather than release evidence. New packaged findings therefore invalidate beta.81 instead of being layered onto an already-approved candidate; its launchd change remains unchanged and is carried into beta.82.
|
|
268
|
+
|
|
269
|
+
**Recovered-success privacy defect.** `runtime-activation.mjs` previously derived successful recovery detail from the lower-layer exception message, bounded only by whitespace removal and a character ceiling. That detail propagated through CLI JSON and owner output, and both local-candidate and published-prerelease writers could persist it in schema-2 activation evidence. A filesystem path, endpoint, account-shaped string, or credential-shaped value emitted by a dependency could therefore cross from local causal state into durable release evidence even though the recovered outcome itself was already classified by one of only three stable reasons. The repair introduces one shared `activation-recovery.mjs` contract mapping `relay_authentication_failed`, `autostart_install_failed`, and `autostart_start_failed` to fixed canonical detail. Runtime settlement, persistent-child validation, activation-record validation, and current writers use this source of truth. Historical schema-2 records still fail closed on malformed recovery metadata but normalize any otherwise bounded legacy detail to the canonical reason-owned text before returning it; current writers therefore cannot re-persist the legacy raw string. Tests inject `/private/tmp/operator-secret`, a reserved `.invalid` URL, and publication-safe credential placeholders at runtime, IPC, writer, and disk-read boundaries and require those values to disappear from public/persisted recovery evidence.
|
|
270
|
+
|
|
271
|
+
**Process-settlement defense in depth.** A second review of successful/error projection found a narrower but real weakness below the UI wrappers. Generic non-replayable pre-spawn settlement converted an unclassified spawn exception into an exposed `BridgeError` whose message was the lower-layer exception text. Current application/browser callers usually reclassified that reason into their own fixed message, but privacy depended on every future caller remembering the second wrapper. The low-level boundary now emits only `process failed before spawn`, retains the original exception as local `cause`, and keeps the stable `process_failed_before_spawn` reason. Already-classified `BridgeError` values remain unchanged, and once process start has occurred the existing non-replayable unknown settlement is still authoritative. Regression tests deliberately place a private-looking path in the lower-layer exception and verify it is absent from the public message/details.
|
|
272
|
+
|
|
273
|
+
**Verification architecture gap.** The review then inspected whether the new shared security boundary was actually protected by repository-wide gates. It exposed a pre-existing blind spot: the package lint command and the correctness ESLint file set omitted `src/shared`, while `scripts/syntax-check.mjs` also omitted that root. Shared modules received complexity limits but could contain an undefined binding or unused import without the advertised production correctness gate seeing it. `src/shared` is now part of both the dynamic syntax scanner and correctness lint CLI/configuration; `lint-gate-test` proves undefined-binding and unused-import rejection on a shared-module path. The activation-recovery module is additionally a strict `tsconfig.local.json` root, an explicit npm package inventory item, a 70-line architecture-budget target, and a critical-coverage target. The fresh coverage run measures 4/4 functions and 28/28 branches (100%/100%). Architecture source guards prevent raw recovery exception text, loss of these gates, or restoration of the removed pre-spawn raw-message helper.
|
|
274
|
+
|
|
275
|
+
**Package-mode and documentation findings.** Explicit package testing of the new module initially failed because the newly created source file had owner-only mode `0600`. The existing package contract correctly rejected it; the file was normalized to `0644` and the 461-file package manifest test then passed. No package-mode rule was relaxed. Separately, `RELEASING.md` and `ARCHITECTURE.md` still claimed schema-1 prerelease activation records using the legacy `previous` field were accepted and normalized. Git history shows that beta.33 once provided that compatibility, while the beta.62 hardening deliberately made current activation reads schema-2-only and installed executable architecture guards against restoring the legacy reader. The current documentation now matches the executable contract: schema-1 files are historical evidence, not supported operational inputs. Historical changelog/audit entries retain their then-true description rather than being rewritten.
|
|
276
|
+
|
|
277
|
+
**Platform, logging, redundancy, and privacy sweep.** The repeated hosted-Windows failures were generalized into a review rule rather than another single fixture patch: a test that injects logical `darwin`/`win32` behavior must not silently use incompatible UID, POSIX mode-bit, signal-delivery, path, or service-manager semantics from the physical host. The remaining Windows platform tail after `service-platform:test` was reviewed for the same pattern; macOS-only trust/native-build fixtures already skip Windows, and process-tree tests assert final settlement through the Windows `taskkill /T` path rather than requiring JavaScript SIGTERM delivery. No additional confirmed portability defect was identified statically. Structured logging and security-audit regressions pass; operational events remain low-cardinality and do not record tool arguments or process output. Empty catches in production were reviewed as cleanup, best-effort detach/release, or preflight probes whose failure later becomes an explicit ambiguous/identity failure; no branch was found that turns an unclassified operational failure into success. A source-import reachability pass initially identified only intentional dynamic/entry modules such as CLI, job runner, audit worker, service restart handoff, trusted GitHub CLI, and Worker index; no orphan production module was found.
|
|
278
|
+
|
|
279
|
+
**Security/privacy evidence.** Current-tree privacy checking passes over 657 tracked/unignored files with the local denylist enabled. Reachable-history privacy checking passes over 3681 blobs and 142 commit messages. Runtime privacy tests, structured logging tests, security-audit tests, consumer-package security, package manifest validation, generated policy/tool reference checks, project metadata checks, and CycloneDX validation (117 components / 118 dependencies) pass. The review did not find an additional confirmed tracked credential, private key, stale current beta/channel snapshot, TODO/FIXME/HACK debt, or active schema-1 release path. The largest remaining maintainability hotspot is still the packaged browser extension: `browser-operations.js` and `page-automation.js` remain large even though executable line ceilings now prevent silent growth. A future staged decomposition around tab/screenshot lifecycle, mutation settlement, and page semantic/action execution would reduce review density, but this release-hardening pass does not refactor those high-risk files without a concrete behavioral defect.
|
|
280
|
+
|
|
281
|
+
**Release consequence.** All beta.81 candidate, full-gate, and install-only evidence predates these packaged changes and is intentionally non-transferable. beta.82 must repeat the complete frozen candidate lifecycle from the reviewed bytes; no owner activation, live canary, acceptance, commit, push, tag, GitHub Release, Worker publication, or npm publication is implied by this source review.
|
|
282
|
+
|
|
283
|
+
## 2026-08-15 beta.81 launchd stop target-resolution portability
|
|
284
|
+
|
|
285
|
+
**Hosted evidence.** The exact beta.80 candidate completed local fast/full/candidate validation, direct-Node disposable install, owner activation, the activated-package OAuth canary, live Worker/service/relay/browser/privacy verification, acceptance, commit, and guarded push. Exact-head Windows then directly proved both prior portability repairs: `resource-admission:test` completed in 3.5 seconds and `macos-background-input:test` completed in 0.3 seconds. The platform plan continued to `service-platform:test`, where its first synthetic `stopLaunchdService()` call failed with `launchd requires a numeric user id` before the mocked status reader could classify the service as unavailable.
|
|
286
|
+
|
|
287
|
+
**Causal mechanism.** `stopLaunchdService()` computed both `launchdDomainTarget()` and `launchdServiceTarget()` before reading current service status. That ordering was stronger than the operational contract requires: the status-unavailable and safely-unloaded branches must not mutate launchd at all, so they do not need a launchd target or user id. On macOS `process.getuid()` masks the unnecessary dependency; hosted Windows has no `process.getuid()`, so a platform-independent synthetic launchd contract exposed the eager resolution. Loaded-service synthetic cases have the same host dependency unless the test provides a deterministic UID.
|
|
288
|
+
|
|
289
|
+
**Repair and invariant.** Target resolution now occurs only after status confirms the service is loaded and a `bootout` mutation is actually required. `stopLaunchdService()` accepts an internal explicit UID dependency for deterministic synthetic contracts; production callers omit it and retain the same `process.getuid()` default. The regression deliberately supplies an invalid UID to the no-mutation unavailable/absent cases, proving they never resolve targets, while loaded cases inject UID 501 and assert the exact `launchctl bootout gui/501/dev.machine-bridge-mcp.daemon` command. Provider selection, launchd label/domain semantics, stop observation, plist fallback, and production identity remain unchanged. Because packaged source changed after beta.80 acceptance, beta.81 requires fresh exact candidate and live evidence.
|
|
290
|
+
|
|
291
|
+
## 2026-08-15 beta.80 Windows synthetic macOS-helper cache portability
|
|
292
|
+
|
|
293
|
+
**Hosted evidence.** The exact beta.79 candidate completed local fast/full/candidate verification, direct-Node install-only validation, owner activation, activated-package OAuth canary, live Worker/service/relay/browser/privacy verification, acceptance, commit, and guarded push. Exact-head Windows then ran through `resource-admission:test` successfully in 3.8 seconds, directly validating the beta.78 host-native cwd fixture repair. The platform run failed later in `macos-background-input:test` at `helper compiled more than once for the same source digest` with two observed synthetic compiler invocations instead of one.
|
|
294
|
+
|
|
295
|
+
**Causal mechanism.** The production helper cache is macOS-only and deliberately trusts a cached native binary only when `executableRegularFile()` proves a regular non-symlink file owned by the current/root user, with at least one POSIX execute bit and no group/other write bits. The deterministic unit fixture forces `platform: "darwin"` so the macOS state machine can run on every CI host, but it still uses the real host filesystem for temporary cache files. Node's Windows filesystem permission model does not let `chmod(0o700)` establish the POSIX execute-bit evidence consumed by that trust predicate. A second `ensureBuilt()` therefore correctly refuses to reuse the synthetic Windows file and invokes the fake compiler again. This is a fixture/evidence mismatch, not evidence that the production macOS cache lost digest or permission validation.
|
|
296
|
+
|
|
297
|
+
**Repair and invariant.** Production `MacosBackgroundInputService` and `executableRegularFile()` remain unchanged. The cross-platform fixture always requires at least one compilation and continues to validate the hashed owner-local source copy, temporary-source cleanup, helper payload/result projection, coercion rejection, probe semantics, and mutation-settlement behavior. The exact one-compile cache-reuse assertion runs only when the host can supply POSIX mode evidence; hosted macOS still exercises that invariant against the real trust predicate. Because the packaged regression changed after beta.79 acceptance, beta.80 requires a fresh exact candidate and live-evidence lifecycle.
|
|
298
|
+
|
|
299
|
+
## 2026-08-15 beta.79 diagnostic attribution, Computer Use privacy, and architecture review
|
|
300
|
+
|
|
301
|
+
**Platform-disable diagnosis.** The reported statement that a retried local execution channel was "directly disabled by the platform" was not supported by Machine Bridge evidence. In the same session, authenticated `server_info` reached a ready beta.77 daemon through the relay with the owner effective catalog, and `diagnose_runtime` reached the local runtime and passed its fixed filesystem/process/shell/resource probes. The server also explicitly cannot observe the host's final post-relay tool filtering. The defect was therefore an attribution contract: `runtime-diagnostics.mjs` described a call blocked before any Machine Bridge response as `host/platform or connector gateway`, which made an unobservable pre-runtime failure look diagnosed. The repaired projection separates fact from hypothesis: successful delivery is positively stated; a pre-response refusal is `not observable by Machine Bridge` and lists host filtering, connector gateway, client routing, and platform policy only as possible causes requiring host-side evidence.
|
|
302
|
+
|
|
303
|
+
**Successful-result privacy boundary.** Independent review found that Computer Use could copy lower-layer `error.message` text into successful result fields such as `dispatch_error`, `post_observation_error`, verification detail, or exposed preflight details. This path bypasses ordinary tool-error sanitization because the overall MCP call itself succeeds. The public projection now emits stable settlement/reason text plus `error_class` only. Raw causes stay attached only as local exception causes where useful. Tests deliberately place `/private/tmp/operator-secret` in ambiguous trusted-input, post-dispatch wait, cancellation, and document-state errors and assert that no public Computer Use result/details contain it. Browser-extension direct tools already have their independent `browser-error-boundary.js` allowlist/canonicalization, so this closes the composition-layer gap rather than replacing that boundary.
|
|
304
|
+
|
|
305
|
+
**Responsibility boundaries.** The review identified two orchestration files whose previous line ceilings had become ineffective: `computer-use.mjs` was within only a few lines of its 2670-line ceiling, while `app-automation.mjs` mixed Node authorization/settlement logic with roughly 350 lines of fixed JXA implementation. Request validation moved to `computer-use-arguments.mjs`; fixed Accessibility JXA moved to `app-automation-macos-jxa.mjs`. The main modules now have materially lower ceilings, the extracted modules have their own budgets, the browser/security guard reads the JXA module directly for selector-index and secure-field suppression, and the largest extension execution modules now have explicit growth ceilings. Behavior tests remain the authority; the extraction does not add caller-selected script/CDP surfaces or weaken mutation settlement.
|
|
306
|
+
|
|
307
|
+
**Documentation, secrets, and history.** The live upgrade guide contradicted its own "current path only" rule by retaining a `Current beta.61 transition` heading and an old npm-channel snapshot. That text is now explicitly historical migration rationale with no current-channel claim. Architecture/local-automation/Computer Use/client documentation was updated to match the new module and diagnostic boundaries. Static secret-pattern review found no obvious private-key/API-token material, the tracked-tree privacy check passed, and the history privacy scan passed across 3650 reachable blobs and 140 commit messages. Existing changelog/audit history is retained because it is release evidence rather than stale operational guidance.
|
|
308
|
+
|
|
309
|
+
**Release consequence.** The exact beta.78 candidate had passed local fast/full/candidate/install-only validation but had not been owner-activated or accepted when this review changed packaged source. That candidate is invalidated rather than retroactively accepted. beta.79 must repeat the complete candidate and live-evidence lifecycle from frozen source.
|
|
310
|
+
|
|
311
|
+
## 2026-08-15 beta.78 resource-admission cwd portability
|
|
312
|
+
|
|
313
|
+
**Observed hosted defect.** The exact accepted beta.77 candidate passed local fast/full verification, owner activation, activated-package OAuth canary, live Worker/service/relay/browser checks, acceptance, and guarded push. Exact-head Windows no longer failed at the beta.76 Wrangler cleanup-race fixture; the platform plan continued until `resource-admission:test` failed at its literal shell cwd assertion. Production returned `D:\\tmp\\project-a`, while the fixture expected the POSIX source literal `/tmp/project-a`.
|
|
314
|
+
|
|
315
|
+
**Causal mechanism.** `resourceCommandEffectiveCwd()` intentionally canonicalizes the request cwd with Node `path.resolve()` and applies the same host-native canonicalization to trusted literal shell `cd` targets. That invariant is required so resource contention keys refer to the actual host path rather than preserving syntax from the shell payload. On POSIX, `resolve("/tmp/project-a")` remains `/tmp/project-a`; on Windows the same rooted input resolves to the current drive-qualified absolute path. The fixture compared the raw input token instead of the function's canonical-path contract. Its adjacent dynamic-cwd fallback assertion had the same latent assumption for `/tmp/shared` and would have failed next.
|
|
316
|
+
|
|
317
|
+
**Repair and invariant.** The regression now imports `resolve` from `node:path` and compares both results against `resolve("/tmp/project-a")` and `resolve("/tmp/shared")`. No resource-admission production code, contention scoping rule, shell parser, path trust boundary, or resource quota behavior changes. Because the packaged regression changed after beta.77 acceptance, beta.78 requires fresh candidate and live evidence.
|
|
318
|
+
|
|
319
|
+
## 2026-08-15 beta.77 Wrangler cleanup-race fixture portability
|
|
320
|
+
|
|
321
|
+
**Observed hosted defect.** The exact beta.76 candidate passed local fast/full verification, owner activation, the activated-package OAuth canary, live Worker/service/relay/browser checks, acceptance, and guarded push. Exact-head Windows then directly validated both earlier portability repairs: `process-output:test` completed in 1.3 seconds and `managed-job-boundary:test` completed in 0.2 seconds. The platform run failed later in `worker-types-generator:test` with `Wrangler cleanup-race fixture did not cross the completion grace`. The assertion was not observing Wrangler lifecycle state directly; it expected the fixture process's `process.on("SIGTERM")` handler to run and write a stderr marker after the parent requested cleanup.
|
|
322
|
+
|
|
323
|
+
**Causal mechanism.** That marker is not a portable cleanup-request witness. Node documents `SIGTERM` as unsupported for ordinary signal delivery on Windows even though a listener can be registered; `subprocess.kill("SIGTERM")` uses Windows termination emulation and unconditionally terminates the target rather than reproducing POSIX signal-handler delivery. The production lifecycle's actual invariant is different: after a valid completion marker, failure to exit within the bounded completion grace must request cleanup; if the child races that request and still closes successfully before forced cleanup, the completed command remains successful. Binding that invariant to child-side SIGTERM output made the fixture host-dependent and tested the operating system instead of the lifecycle transition.
|
|
324
|
+
|
|
325
|
+
**Repair and invariant.** `runCompletedWranglerCommand()` now resolves one internal `killChild` dependency, defaulting exactly to `target.kill(signal)`, and uses it for the existing graceful and forced cleanup calls. The cleanup-race regression injects only this dependency: it records the requested signal, treats the graceful request as delivered without sending an OS signal, and allows the real fixture child to take its pre-existing 250 ms normal-exit path. The assertion therefore proves that the 20 ms completion grace caused exactly one `SIGTERM` cleanup request while the command still settles successfully from the child's real zero exit. Hanging fixtures continue to use the default OS kill path, preserving behavior coverage for bounded cleanup. No default production timeout, signal, success classification, or forced-cleanup policy changes. Because packaged bytes changed after beta.76 acceptance, beta.77 requires fresh candidate and live evidence.
|
|
326
|
+
|
|
327
|
+
## 2026-08-15 beta.76 managed-job Windows boundary fixture portability
|
|
328
|
+
|
|
329
|
+
**Observed hosted defect.** The exact accepted beta.75 candidate passed the local fast/full candidate gates, owner activation, activated-package OAuth canary, exact protocol-3 browser handshake, live background-tab Computer Use click/hover checks, consumed-snapshot rejection, runtime/privacy audit, acceptance, and guarded push. On the beta.75 exact-head Windows platform run, the preceding beta.75 repair was directly vindicated: `process-output:test` completed successfully in 1.2 seconds. The job then failed in `managed-job-boundary:test` at its explicit-directory-mode assertion with `actual: null`, `expected: 448 (0o700)`.
|
|
330
|
+
|
|
331
|
+
**Causal mechanism.** `withPinnedManagedJobDirectory()` has an intentional platform split. POSIX opens an already-existing directory through `O_RDONLY | O_NOFOLLOW | O_DIRECTORY` with mode `0o700`, pins the descriptor, and cross-checks descriptor/path/canonical filesystem identity. The Windows branch does not invoke `openSync`; it rejects symbolic-link/non-directory state through `lstat` and applies the existing canonical/path device-and-inode identity checks. `docs/TESTING.md` already describes the explicit descriptor mode requirement specifically as **POSIX directory pinning**. The fixture nevertheless installed an `openSync` observer and asserted that it saw `0o700` on every host, so a correct Windows execution left the observer at `null` and failed. This is a test-portability mismatch, not evidence that beta.75 weakened managed-job directory permissions or identity validation.
|
|
332
|
+
|
|
333
|
+
**Repair and invariant.** The real-host descriptor-mode assertion now runs only when `process.platform !== "win32"`. A second platform-independent regression explicitly calls the resolver with `platform: "win32"` and an `openSync` stub that records-and-throws if invoked; the resolver must return the same canonical job root while proving that Windows did not enter the POSIX descriptor path. Production managed-job code remains unchanged. Because this packaged test change follows beta.75 acceptance and push, beta.76 requires a fresh candidate, activation, canary, live verification, acceptance, guarded push, and hosted exact-head validation.
|
|
334
|
+
|
|
335
|
+
## 2026-08-15 beta.75 Windows process-session fixture portability
|
|
336
|
+
|
|
337
|
+
**Observed hosted defect.** The exact accepted beta.74 candidate was owner-activated and passed its packaged deployed OAuth canary, stable-path protocol-3 extension handshake, live background-tab Computer Use observation/click/hover checks, one-shot replay rejection, and runtime/privacy audit before guarded source push. Exact-head Ubuntu/full and macOS/platform+install passed; package-audit, governance, dependency review, workflow policy, JavaScript/TypeScript CodeQL, Actions CodeQL, and the zero-unaccepted-CodeQL gates also passed. Windows alone failed `check:platform`: `process-output:test` ran for about 31 seconds and `testSessionAuthorityRevocation` finally surfaced `process-session shutdown was incomplete; runtime ownership must be retained` from its cleanup path.
|
|
338
|
+
|
|
339
|
+
**Causal mechanism.** The failure was produced by the fixture's synthetic termination handles, not by a demonstrated slow real Windows child shutdown. On Windows, `terminateProcessTree()` requests tree termination by spawning `taskkill.exe /PID ... /T /F` and returns true when that helper is successfully spawned; whether a synthetic target PID exists is settled asynchronously. The fixture instead used fake children whose `kill()` returned false and assumed that this would prove delivery failure synchronously on every platform. As a result, the Windows forced-kill assertion received a successful request result, threw before deleting the fake session, and entered the outer `finally`. Final `clearAndWait()` then waited on that impossible fake session and threw its own teardown error, masking the original assertion. The earlier beta.72 test-only 15-second manager budget made the same platform mismatch slower rather than fixing it; the approximately 31-second hosted duration is consistent with two 15-second synthetic settlement waits plus fixture overhead.
|
|
340
|
+
|
|
341
|
+
**Repair and invariant.** The failed-delivery branches now inject `manager.terminateTree = () => false` only for the duration of the synthetic assertion, so they test `ProcessSessionManager`'s definite-delivery-failure contract independently of the host tree-kill mechanism. Each fake session is removed in its own `finally` block and the original terminator is always restored, preventing a failed assertion from contaminating final cleanup or replacing the actual failure. The separate delivered-but-unsettled case keeps an explicit 100 ms synthetic deadline and also cleans itself in `finally`. The authority fixture no longer overrides `terminationSettlementWaitMs`, so its real surviving child is again cleaned with the unchanged production-default five-second settlement contract. No runtime termination semantics or retryability contract is relaxed. Because this packaged test change follows beta.74 acceptance and push, beta.75 requires a fresh candidate, activation, canary, live verification, acceptance, guarded push, and hosted exact-head validation.
|
|
342
|
+
|
|
343
|
+
## 2026-08-14 beta.74 bounded Chromium DevTools session settlement
|
|
344
|
+
|
|
345
|
+
**Observed live defect.** The exact beta.73 candidate was owner-activated, its candidate-bound deployed OAuth canary passed, activation published the extension into the new stable release-channel directory, and Chrome successfully loaded that directory and completed an exact protocol-3 `3.0.0-beta.73` handshake with Computer Use/CDP/trusted-input capabilities present. Browser acceptance still failed on a real background synthetic tab. `browser_inspect_page` repeatedly proved the tab document was complete and readable while `computer_observe` intermittently exceeded its broker deadline. A snapshot-bound click later hit the browser preflight five-second timeout, so that action was not replayed. The live evidence therefore separated extension installation/pairing and ordinary page scripting from the Computer Use DevTools path.
|
|
346
|
+
|
|
347
|
+
**Causal mechanism and repair.** `devtools-session.js` serialized one debugger session per tab but awaited raw `chrome.debugger.attach`, `sendCommand`, and `detach` promises with no independent settlement deadline. The broker can request cancellation after its own timeout, but a pending `sendCommand` cannot reach the next `throwIfCancelled()` check and keeps the per-tab queue unresolved; later CDP observations then wait behind the same unresolved session. Beta.74 bounds attach and each command at five seconds and detach cleanup at 1.5 seconds. A late successful attach after an attach timeout schedules best-effort detach. Observation code treats the internal DevTools-timeout marker as fatal for that CDP capture instead of swallowing it as an optional component failure, allowing the existing `observe_computer` semantic fallback to settle and release the queue.
|
|
348
|
+
|
|
349
|
+
**Mutation safety and regression.** The shared session is also used by trusted browser input, so the deadline is not treated as replay authorization. `devtools-input` already sets `dispatchStarted=true` before invoking each Input command; a command timeout therefore remains `safeToFallback=false` and flows through the existing unknown-mutation settlement. Regression fixtures make a DevTools command promise never resolve and verify prompt timeout, detach, and successful reuse of the same tab queue; a separate trusted-hover fixture proves an `Input.dispatchMouseEvent` timeout remains post-dispatch/unknown rather than fallback-safe. Focused trusted-input, DevTools-observation, and browser Computer-observation tests pass. Because the fix changes packaged extension bytes after beta.73 activation/canary, beta.74 requires fresh candidate preparation, owner activation, canary, live browser verification, hosted checks, and acceptance.
|
|
350
|
+
|
|
351
|
+
## 2026-08-14 beta.73 stable unpacked-extension release path
|
|
352
|
+
|
|
353
|
+
**Observed release defect.** The exact beta.72 candidate was owner-activated and its packaged deployed OAuth canary passed, but browser acceptance remained blocked after the documented extension Reload. The extension itself was still installed under the pinned Machine Bridge extension ID; the broker's `extension_reload_required` state proved that an extension WebSocket candidate had reached the broker but had not completed a valid protocol-3 hello. Source review then exposed a release-state contradiction independent of Chrome UI visibility: local-candidate `extension_path` was the current package's `browser-extension` directory inside `release-channels/runtimes/v<version>-<sha>-<nonce>/`, while successful activation deliberately pruned every inactive versioned runtime. Chrome retains the source directory selected by **Load unpacked**, so a later Reload cannot move that installation to a newly created candidate directory. This violates the documented load-once/reload-on-upgrade contract and makes beta.72 unfit for acceptance even though its Worker/daemon/OAuth path is healthy.
|
|
354
|
+
|
|
355
|
+
**Repair and safety ordering.** Versioned local-candidate runtimes now resolve browser setup/status to one owner-only `release-channels/browser-extension` physical directory. Ordinary checkout/global runtimes remain package-local so development and globally installed package behavior are not silently redirected through release state. Candidate activation first proves exact Worker and background-daemon convergence, then copies the installed candidate extension to the stable directory. Source reads reject symbolic links and use opened regular-file descriptors; destination files are replaced atomically, stale files are removed, and `manifest.json` is replaced last so its `version_name` is the commit marker only after the rest of the candidate bytes are present. Inactive versioned runtimes are pruned only after that publication succeeds, and activation evidence is written afterward. State-root removal recognizes this bounded real-file namespace and still rejects symbolic links or other unexpected entry types.
|
|
356
|
+
|
|
357
|
+
**Regression and migration.** `candidate-runtime-store:test` proves that two candidate versions publish to exactly the same path, observes the old manifest while the new non-manifest bytes are already staged, verifies stale-file removal and final exact bytes, rejects source symlinks, and proves only versioned candidate packages are redirected to release state. Architecture guards bind live-convergence → stable-extension publication → inactive-runtime prune → activation-record ordering and require browser bridge/CLI routing through the shared path helper. Existing installations loaded from beta.72-or-earlier candidate-runtime paths require one manual **Load unpacked** migration to beta.73's reported stable path; browser-internal extension management remains outside Machine Bridge automation. Subsequent candidate upgrades use Reload on the same path, while explicit extension-icon confirmation remains required when replacing different browser-local pairing state.
|
|
358
|
+
|
|
359
|
+
## 2026-08-14 beta.71 browser action timing settlement follow-up
|
|
360
|
+
|
|
361
|
+
**Observed live defect.** The exact beta.70 candidate was owner-activated, matched its candidate provenance, passed the packaged deployed OAuth canary, and paired the packaged Chromium extension at protocol 3 with Computer Use/CDP capabilities present. A controlled loopback page then separated observation from mutation behavior. `computer_observe` returned a navigation-coherent DOM/Chromium Accessibility snapshot and ordinary read-only browser inspection remained healthy. A DOM `focus` action on the same control settled successfully, while DOM `click` and `hover` requests repeatedly reached the extension but exceeded their broker deadlines. The click effect was independently observed afterward on the page, proving that mutation could complete while its response remained unavailable. Machine Bridge's existing settlement boundary behaved correctly: the request became non-retryable unknown rather than a false success or a replay opportunity. That safety property prevented duplicate input, but the inability to obtain an ordinary confirmed browser mutation is a release blocker, so beta.70 was not accepted.
|
|
362
|
+
|
|
363
|
+
**Causal boundary and repair.** Both failing actions share deterministic scroll, stable-box, and pointer-hit preparation; `focus` does not. That preparation waited with page-world `setTimeout(50/100)`. Chromium is permitted to batch, intensively throttle, or freeze timers in background/hidden pages, so a short geometry poll could outlive Machine Bridge's foreground request deadline even though the page action later resumed and executed. The repair does not delete or weaken geometry/actionability checks. When the page module is running inside the packaged extension, short polling waits now cross `chrome.runtime.sendMessage` to a service-worker-only `machine_bridge_internal_delay` endpoint. The endpoint accepts only messages whose sender ID equals the running extension ID and only exact integer delays from 1 through 250 ms. A timing-service failure before mutation remains definite; after scroll or another side-effecting preparation step it flows through the existing unknown-mutation settlement. A renderer-timer fallback exists only when the extension runtime API is absent, which keeps the pure VM fixture usable without reintroducing the packaged background-tab dependency.
|
|
364
|
+
|
|
365
|
+
**Regression and release impact.** A test-first fixture disables renderer `setTimeout` while providing the internal extension delay service and requires DOM click to retain deterministic scrolling/stability and settle successfully; the pre-fix source fails that fixture at the exact renderer timer. Service-worker coverage independently rejects cross-extension and over-250-ms delay messages. Page automation, service worker, trusted input, browser request settlement, Computer Use, browser security/source guards, and module responsibility limits pass after the repair, followed by the 102-task fast plan. The repair changes packaged browser-extension bytes, so beta.70 candidate/canary evidence cannot authorize it. The unused beta.71 namespace was checked in npm, local/remote Git tags, and GitHub Releases before the version bump; beta.71 must repeat full candidate preparation, owner activation, deployed canary, live observation, and acceptance.
|
|
366
|
+
|
|
367
|
+
## 2026-08-14 beta.70 Computer Use integration and independent review
|
|
368
|
+
|
|
369
|
+
**Integration boundary.** Computer Use vNext was reviewed as a reference implementation rather than merged as a branch. The independently developed source had diverged into a parallel Machine Bridge trunk and still carried the older stateful MCP initialization/session/resumption design. Beta.70 instead keeps the current beta.69 codebase authoritative and ports the Computer Use capability onto the native request-scoped `2026-07-28` runtime. Remote HTTP initialization compatibility remains the bounded stateless adapter for `2025-06-18` and `2025-11-25`; it does not create or accept `Mcp-Session-Id`, recovery GET, `Last-Event-ID`, persisted replay/delivery rows, or initialization-owned authorization state. The beta.69 structured Git, metadata-boundary, support-projection, daemon-readiness, release, and runtime protections are retained rather than overwritten by the older parallel trunk.
|
|
370
|
+
|
|
371
|
+
**Computer Use contract.** The resulting product adds `computer_observe` and `computer_act` over the existing browser and application backends. Browser snapshots combine bounded executable DOM evidence with Chromium Accessibility/DOMSnapshot evidence, frame/document epochs, trusted-input bindings, and optional native MCP screenshots. macOS application snapshots combine Accessibility state with process-birth and owner-window evidence and optionally a window screenshot. Snapshot IDs are process-local opaque handles, not authorization; every call still passes normal policy/operation authorization. Mutation authority is one-shot: read-only preflight precedes an atomic snapshot claim, mutation handoff is attempted at most once, ambiguous lower-layer settlement is never automatically replayed, and post-state/continuation evidence is returned separately from dispatch settlement.
|
|
372
|
+
|
|
373
|
+
**Independent seam review.** Reference-source parity was not treated as sufficient proof because beta.70 also had to compose vNext with newer request settlement, process settlement, and 7 MiB result-budget boundaries. The review found that `post_screenshot_included` reported policy intent rather than whether native image content actually reached the result; screenshot capture failure could therefore claim an image that was absent. It now reflects actual returned image content. A second defect existed when an oversized post-action screenshot was removed by the post-capture budget before outer action-result fitting: the action remained safe and continuable, but the structured budget-omission reason could disappear. Both capture-level and outer action-level omission now converge on `tool_result_budget`, preserve dispatch/effect settlement and the bounded `post_snapshot_id`, and keep same-action replay disabled. The compact-result fallback was also found to manufacture that screenshot reason for a hypothetical semantic-only overflow; it now preserves the reason only when an image was actually omitted for the budget.
|
|
374
|
+
|
|
375
|
+
**Documentation and regression boundaries.** The page automation module emits snapshot version 3 with per-document epoch semantics, while local-automation and architecture documentation still named version 2; both are corrected and an architecture test now binds source and documentation versions. README incorrectly stated that all clients had no `initialize` handshake even though the current Worker intentionally supports stateless remote initialization compatibility, and the system overview simultaneously counted four authority questions as three and referred to nonexistent current MCP session state. Repository-hygiene tests now reject those stale positive claims while requiring the current native/stateless-compatibility model. Focused behavior fixtures were added for Computer Use result-budget downgrade, browser request settlement, BrowserOperationService including Computer Use adapter entrypoints, and ordinary/non-replayable fixed-process settlement. These fixtures participate in the normal fast plan and critical coverage instead of living only as ad hoc review probes.
|
|
376
|
+
|
|
377
|
+
**Candidate-gate resource race.** A later candidate rerun under unrelated concurrent machine jobs passed the first 119 full-plan tasks, then `full-access:test` failed while resource admission inspected a coordinator staging inode still owned by a live publisher. That state is legitimate for the short interval between complete-before-visible staging publication and cleanup; deleting it would violate ownership, but surfacing it immediately as an internal error made ordinary unrelated work spuriously fail. Live lease/waiter staging now carries the stable `MBM_RESOURCE_STAGING_BUSY` classification. `ResourceCoordinator.withLock()` releases its transaction generation and permits only four five-millisecond retries; a publisher that remains live still fails closed, while an ordinary short publication may settle. Exhausted staging or transaction-lock contention is projected through the existing retryable resource-admission unavailable result rather than an internal exception. The pre-existing dead-publisher identity/link-count recovery is unchanged. Focused resource-admission and process-admission regressions pass, the existing responsibility cap is preserved without increase, and the exact `full-access` real-machine fixture that exposed the race passes after the repair.
|
|
378
|
+
|
|
379
|
+
**Second independent post-closure review.** After the resource-staging repair passed 129/129 and an install-only closure candidate was prepared, a fresh review deliberately treated that candidate as a baseline rather than proof. It found that Computer Use snapshot TTL and application verification duration still used an injectable wall clock; the snapshot authority state is now extracted to `computer-use-snapshot-store.mjs`, defaults to `performance.now()`, clamps injected clock rollback, and the verification loop uses the shared monotonic deadline helper. AppAutomation already had a monotonic manager clock, but its resource-backed exact-value verification handle incorrectly used `Date.now()` for the 60-second secret-bearing fallback TTL; retention and pruning now use the same monotonic clock. The review also found that the extension already classified an oversized/unserializable successful mutation result as unknown, but the local broker did not attach structured settlement for that fixed message, and direct application wrappers discarded lower-level post-spawn settlement details. Both now preserve the non-replayable result together with bounded `side_effects_started`, `termination_requested`, and `effect_settlement` evidence. Architecture governance now includes every Computer Use core/observation/recovery/snapshot/native/browser adapter module plus extension size caps, and ties the duplicated local/extension 7 MiB byte ceiling to one executable drift guard. Critical coverage now measures the Computer Use kernel itself; snapshot-store coverage reaches 100% functions / 95.2% branches, browser Computer Use observation 92.0% / 75.9%, and the existing core observation/recovery/native paths remain above their new floors. Five cheap GUI security fixtures move from platform-only to fast, expanding the fast plan from 97 to 102 without changing the 121-task platform or 129-task full set. No credential, private-key, absolute user-home path, Computer Use content log, TODO/FIXME/HACK residue, or current positive session/resumption contract was found; one documentation-only internal `Codex/daily/ops` scratch-path assumption was generalized.
|
|
380
|
+
|
|
381
|
+
**Verification.** Before this independent pass, the beta.70 integration completed the then-current 94-task fast plan. After the review changes, direct Computer Use, browser observation/bridge, process-settlement, architecture, Markdown, and typecheck checks pass. The frozen critical-coverage run passes without lowering thresholds: `process-nonreplayable-settlement.mjs` measures 100% functions / 96.4% branches, `computer-use-result-budget.mjs` 100% / 90.7%, `browser-request-settlement.mjs` 100% / 100%, and `browser-operation-service.mjs` 83.3% / 79.4%, above its existing 80% function threshold. The expanded frozen fast plan passes 97/97 twice, including after the audit/README/overview corrections. The first 129-task full run passed tasks 1-123 and then correctly stopped at `package:test`: 17 files brought over with the Computer Use work still had owner-only `0600` checkout modes. Those non-executable source/documentation/native files were normalized to `0644` without relaxing the package gate; the targeted 455-file package manifest test passed, and the clean rerun then passed all 129/129 tasks including package, SBOM, install, stdio, Worker integration, and OAuth-browser verification. Privacy history scanning also passes over 651 tracked/unignored files, 3462 reachable history blobs, and 135 commit messages; both complete and production-only high-severity npm audits report zero vulnerabilities, while registry verification reports 117 packages with valid signatures and 38 with attestations. A closure-baseline beta.70 candidate was prepared after a 129/129 full gate and passed direct-Node install-only verification, but the second independent review changed packaged runtime bytes and therefore explicitly invalidated that tarball as release evidence. Candidate generation for the reviewed bytes occurs only after the new frozen gates; `.release-candidate/manifest.json` is the authoritative final digest rather than this audit prose. No owner activation, Worker/service replacement, acceptance, commit, push, tag, release, or npm publication is implied by these local review results.
|
|
382
|
+
|
|
383
|
+
## 2026-08-13 beta.69 bounded Git mutation and metadata-boundary follow-up
|
|
384
|
+
|
|
385
|
+
**Observed facts.** The beta.68 candidate completed the full owner-machine live verification sequence and was locally accepted. The subsequent local commit step, however, was stopped before reaching the Machine Bridge daemon when expressed through the generic process tool. Repeated diagnostics continued to reach the daemon with owner/full effective authority, and `OperationAuthorizer` classifies owner shell execution as allowed. The public catalog still described `run_process` and `exec_command` accurately as non-read-only, destructive, non-idempotent, open-world execution surfaces. During this repair an ordinary `npm version` direct-process call and several source edits containing Git-mutation terminology were also stopped before daemon execution, while adjacent structured edits and fixed tests were accepted. These observations falsify the hypothesis that Machine Bridge's owner role or internal operation authorizer had newly denied `git commit`; the unstable decision is upstream of daemon authorization and can vary with the exposed tool/request shape.
|
|
386
|
+
|
|
387
|
+
**Causal design defect.** A routine, repository-local, additive history operation had no structured surface of its own. Read-only Git already had `git_status`, `git_diff`, `git_log`, and `git_show`, but commit required the same arbitrary-process capability that can launch unrelated executables, shells, helpers, and external side effects. Changing `run_process.destructiveHint` to false would misrepresent that much larger capability and weaken host/user risk signaling. Beta.69 instead adds `git_commit`, a fixed implementation-owned write operation that consumes only an existing staged index. The trusted Git executable is selected by Machine Bridge; message text crosses stdin rather than argv; repository hooks and signing are disabled; no editor is started; staging, amend, push, tag, and release operations are absent from the schema and implementation. The surface is available at the existing direct-exec policy level rather than the edit-only file-mutation level. Commit creation is plumbing-only (`write-tree` → `commit-tree` → compare-and-swap `update-ref`) so a concurrent ref change cannot be overwritten silently. Generic process tools retain their existing destructive/open-world annotations.
|
|
388
|
+
|
|
389
|
+
**Adjacent boundary found during independent challenge.** The first `git_commit` draft authorized only the detected worktree root. That is insufficient for linked worktrees, submodules, and gitdir indirection because `.git` can be a pointer to metadata and a common object/ref store outside the selected workspace. The same issue pre-dated beta.69 in the fixed read-only Git tools: asking Git itself to discover the repository could read that external metadata before Machine Bridge had enforced its path boundary. Repository discovery is now filesystem-first. It walks authority-permitted ancestors, rejects a symbolic `.git` marker, reads gitdir and commondir through bounded no-follow single-link regular-file reads, authorizes each canonical metadata directory, and only then permits a Git subprocess. A real linked-worktree fixture confines the resolver to the linked worktree and proves the external metadata pointer is rejected before the injected Git runner is called. The same discovery now canonicalizes the primary object store, recursively resolves bounded `objects/info/alternates`, and authorizes HEAD/refs/logs/packed-refs/reftable/object-info paths before Git runs; a confined alternate-object fixture is rejected before subprocess execution while an owner/unrestricted fixture remains readable. Direct `.git` reads and writes are also classified sensitive, preventing delegated file tools from bypassing the structured Git boundary through control-plane metadata. Owner/full unrestricted-path behavior remains available by design.
|
|
390
|
+
|
|
391
|
+
**Failure and privacy properties.** `git_commit` does not return the message, Git stdout, author identity, object path, or hook path. A nonzero Git result becomes a bounded `git_commit_failed` classification and leaves callers to inspect ordinary staged/status state; this avoids returning repository configuration or helper diagnostics across a mutation boundary. Empty staged state produces no commit, an oversized UTF-8 message is rejected before process execution, and a failed retry cannot silently use `--allow-empty` because that option is never present. The random nonexistent hooks path is per invocation and is not persisted. The message-size schema bound plus byte-level backend bound prevents multi-byte Unicode from expanding beyond the internal stdin budget.
|
|
392
|
+
|
|
393
|
+
**Second independent review before freeze.** A fresh defect-first pass challenged the repaired Git boundary rather than assuming the first review was complete. A real loose-object control proved that Git follows a symbolic-link object stored below an otherwise authorized `.git/objects` directory; restricted structured Git therefore now scans the metadata trees it may traverse and rejects symbolic-link or special-file descendants before process launch. A separate real repository proved that a literal directory named like `:(top,glob)**` was interpreted as pathspec magic and could return another file's diff; all structured Git commands now force literal pathspec semantics. `project_overview` was also found to retain a second raw `rev-parse --show-toplevel` path and now reuses `GitService` discovery. Ref validation delegates to `git check-ref-format`, fixing a local regex that rejected Git-valid names such as `refs/heads/-topic`. Repository `diff.orderFile` and partial-clone/promisor configuration are rejected so the non-open-world contract cannot read an unapproved local file or lazily fetch missing objects.
|
|
394
|
+
|
|
395
|
+
The same pass tightened adjacent contracts instead of broadening trust. Internal fixed-process environment overrides are exact-value allowlisted to the five structured-Git safety toggles, so loader/startup variables cannot be introduced through that seam. Git path and revision schemas now reject oversized strings before path/object parsing. The new Git/privacy regressions participate in critical coverage, selected pure boundary modules opt into strict checked JavaScript, generated tool/policy references are regenerated from the catalog, and every new shipped/test file is normalized to package-safe `0644`. The support path no longer asks users to manually sanitize an exact Worker endpoint from `doctor`: successful health reports only reachability and the state section is a bounded support projection without endpoint/name/workspace/key/resource identifiers. Owner-local `status` remains the richer diagnostic surface.
|
|
396
|
+
|
|
397
|
+
The first frozen full verification pass surfaced one additional availability defect outside the Git changes. The Worker started recovery timing before awaiting `waitForReadyDaemon()`, so even its immediate-ready branch could cross a busy event-loop scheduling boundary and lose a millisecond from a one-second execution timeout. That overhead is not daemon recovery and must not reduce the daemon's execution budget. `daemon-ready-dispatch.ts` now checks for an already-ready socket first and reports zero recovery delay without sampling the recovery clock; only a real wait is timed and passed to `daemon-recovery-budget.ts`. Deterministic unit coverage proves both branches, and the same real Wrangler integration assertion that exposed the defect passes without loosening its exact execution-timeout expectation.
|
|
398
|
+
|
|
399
|
+
**Third independent review (2026-08-14).** A real merge repository falsified an implicit assumption in the first plumbing design: `write-tree` → single-parent `commit-tree` → `update-ref` can succeed while `MERGE_HEAD` exists, flattening a merge into ordinary history and leaving the repository in an inconsistent in-progress state. `git_commit` now rejects merge, rebase, cherry-pick, revert, bisect, and sequencer markers before staged-tree plumbing; exhaustive marker tests and the real merge fixture prove HEAD/state remain unchanged. The review also found that restricted Git metadata-tree inspection honored an entry/depth ceiling but not the runtime cancellation signal, allowing a cancelled or timed-out call to continue a large pre-process traversal; the scan now preserves and throws the call's abort reason during root, directory, and entry traversal. The five fixed Git safety environment values were duplicated between the service and its validator, so they now come from one exact-value source. Finally, `SECURITY.md` still described the removed resumable/session-backed MCP transport even though the implementation, architecture, and integration tests are request-scoped and reject `Mcp-Session-Id`, recovery GET, and `Last-Event-ID`; the stale section is replaced with the current non-resumable model and repository-hygiene tests pin that documentation contract.
|
|
400
|
+
|
|
401
|
+
**Verification status.** Real temporary-repository tests pass staged-only commit creation with a deliberately failing pre-commit hook and `commit.gpgSign=true`, proving both are suppressed; the committed tree excludes an untracked file and exact multi-line message content survives stdin. Catalog, schema validation, the 52-handler matrix, operation authorization, capability routing, lint, and local TypeScript checks also pass. The strongest host-compatibility proof cannot occur until beta.69 is owner-activated: after packaged canary/live verification and acceptance, this same conversation must create the repository commit through the newly exposed `git_commit` MCP tool. If that structured call is still stopped before daemon execution, the remaining incompatibility belongs to the host safety/control plane rather than to Machine Bridge's generic-process annotations and must be reported as such instead of broadening this tool.
|
|
402
|
+
|
|
403
|
+
## 2026-08-13 beta.68 hosted-provider follow-up
|
|
404
|
+
|
|
405
|
+
Beta.67 completed owner activation, candidate-bound deployed OAuth canary, observed live verification, local acceptance, commit, and guarded branch push before the exact pushed head reached hosted provider gates. The local candidate evidence remains valid for beta.67 itself, but the pushed head is not publishable: Ubuntu full CI, Windows platform CI, macOS platform CI, and JavaScript/TypeScript CodeQL each produced a blocking result. Because every repair below changes packaged bytes after beta.67 activation/acceptance, the beta.67 acceptance record is removed from the new source generation and beta.68 must repeat the complete candidate/activation/canary/live-observation/acceptance sequence.
|
|
406
|
+
|
|
407
|
+
Ubuntu exposed a real managed-job destructive-cleanup ABA. `removeManagedJobDirectoryIfCurrent()` compared the complete pre-rename filesystem identity, but after quarantine rename it checked only device and inode. On Linux an immediately removed/recreated directory can reuse the same numeric inode, allowing a replacement pathname generation to look like the original and reach recursive deletion. Comparing ctime after rename is not a correct repair because rename itself changes directory ctime; this was independently reproduced on macOS while the inode remained constant. The repair instead holds an `O_NOFOLLOW|O_DIRECTORY` descriptor across the POSIX rename and moved-path check, keeping the original inode referenced so reuse cannot occur inside that window. The regression now preserves the replacement generation, and the architecture line budget remains unchanged.
|
|
408
|
+
|
|
409
|
+
Windows exposed a liveness defect in the process-session authority-revocation settlement wait. The shared `waitForSessionChange()` timer was always unreferenced, including when revocation/shutdown correctness depended on that timer to complete a bounded top-level await. In an otherwise-idle Node process, the event loop could therefore terminate with the revocation Promise still unsettled. Ordinary observation waits remain unreferenced, while the security-critical termination path now explicitly keeps its deadline timer referenced. The process-output/session regression completes locally with the production five-second revocation contract; hosted Windows remains the required platform-fidelity proof for the beta.68 head.
|
|
410
|
+
|
|
411
|
+
Hosted macOS exposed a test-harness ordering defect rather than a failed POSIX group signal. The self-test deliberately times out one command, whose public error correctly reports pending effect settlement while tree cleanup may continue, and then immediately started a separate termination test after a fixed delay. A busy runner could therefore start the second case while the first process still owned tracker/resource state; a fast termination could also reject before the later assertion attached a rejection handler. A direct detached-process probe confirmed that process-group SIGTERM works on the host. The self-test now waits within the production escalation budget for the prior tracker lifecycle to settle, observes a new tracker ownership increment before terminating the next process, and attaches its rejection observer at Promise creation time. It no longer assumes a 50 ms machine-speed boundary or bypasses resource admission.
|
|
412
|
+
|
|
413
|
+
CodeQL reported `js/insecure-temporary-file` at two descriptor-open sites reached by temporary-directory test data. Both production operations open existing objects read-only with no-follow semantics: bounded project metadata reads and state-root directory pinning. Rather than accept or suppress the findings, beta.68 makes the intended private mode explicit at the open boundary (`0600` for metadata, `0700` for the directory pin) and adds a metadata regression for the mode. The repository's zero-unaccepted-CodeQL-findings policy remains unchanged; only a hosted beta.68 CodeQL run can close those provider findings.
|
|
414
|
+
|
|
415
|
+
Targeted local validation after the repairs passes managed-job boundary, process-output/session, project-metadata, state-root-retirement, self-test, architecture, lint, structured logging, and privacy gates. This does not substitute for the exact-head hosted Ubuntu/Windows/macOS/CodeQL rerun. The beta.68 namespace was verified unused in npm, local/remote Git tags, and GitHub Releases through the local GitHub CLI before the version bump.
|
|
416
|
+
|
|
417
|
+
## 2026-08-13 beta.67 relay-disconnect incident and independent post-fix review
|
|
418
|
+
|
|
419
|
+
Observed facts were separated from inference before patching. The running login daemon was a single stable service process rather than a crash/restart loop, while historical daemon logs contained repeated relay WebSocket `1006`, `network_error`, and connect-timeout episodes. One 2026-08-13 outage lasted about fifteen and a half minutes and required fifty reconnect attempts; after the shared 120-second same-daemon reconnect grace expired, the runtime reported that one already-completed queued result had been discarded. That is direct server-side evidence for a class of incidents where local execution can complete but its terminal delivery cannot be recovered. The exact host/UI behavior after such a terminal-delivery failure remains outside Machine Bridge observability, so the review does not claim to prove which ChatGPT rendering path produced every visually empty conversation.
|
|
420
|
+
|
|
421
|
+
Large `runtime.event_loop.stall` records were checked against macOS power events instead of being assumed to be CPU starvation. Several multi-minute stalls overlap Idle Sleep/DarkWake/FullWake intervals, confirming local suspend/resume as one fault class. A separate long relay outage had no matching sleep event, falsifying the hypothesis that sleep alone explains the instability. The current system route is carried through a VPN/TUN, which is relevant transport context but not sufficient evidence to name the VPN product or any specific upstream node as the cause. The supported conclusion is therefore two independent external disruption classes: local scheduling suspension during sleep/wake and network-path interruption while the daemon remains scheduled. Machine Bridge can harden recovery semantics around both; it cannot make the external network continuously available.
|
|
422
|
+
|
|
423
|
+
The relay heartbeat policy preserves grace for a short scheduling stall. After a pause longer than heartbeat timeout plus recovery grace, it reconnects immediately only when the relay is also still silent beyond heartbeat timeout; fresh inbound traffic processed after resume proves the socket live and keeps it on the ordinary recovery path. A deterministic manual-scheduler regression proves both the fresh-inbound race and the stale-plus-silent termination boundary, including the structured `runtime.event_loop.stall` transition from `relay_disconnect_deferred=true` to `false` before reconnect.
|
|
424
|
+
|
|
425
|
+
The terminal-loss path was also made directly observable. Reconnect-grace expiry now emits `relay.calls.reconnect_expired` at warning level with only three aggregate fields: cancelled call count, discarded completed-result count, and grace duration. No argument, result, account identity, request key, or call identifier is added to this warning. Runtime regression coverage proves the aggregate values. Operations and testing documentation now distinguishes short-stall heartbeat recovery, long-pause-plus-silence stale-socket reconnect, and reconnect-grace expiry instead of requiring operators to parse a free-form warning.
|
|
426
|
+
|
|
427
|
+
The requested independent review was then performed from a fresh skeptical pass over architecture, module responsibilities, Worker pending/waiter capacity, timeout arithmetic, relay/result ownership, release-runtime provenance, resource admission, state-root retirement, logging, privacy, tests, and documentation. The 30 ordinary + 2 reserved Worker capacity remains shared between already-dispatched calls and pre-dispatch recovery waiters; delegated accounts still receive hidden activity projections; recovery waiting is deducted from the original execution/settlement envelope; release-control canary provenance remains bound to the activated package/runtime Node/service identity; state retirement and bounded-file gates remain fail-closed. Repository privacy scanning, structured-logging checks, architecture/documentation contracts, lint, typecheck, MCP/Worker/security/resource tests, and the 88-task fast verification plan pass on the reviewed source. No tracked credential/key artifact, TODO/FIXME/HACK residue, or new cross-principal pending-activity disclosure was found.
|
|
428
|
+
|
|
429
|
+
The review itself found two follow-up consistency defects and repaired them rather than weakening gates. First, the new structured relay warning temporarily pushed `relay-call-recovery.mjs` past its existing 170-line responsibility limit; the implementation was compacted back under the existing boundary and the architecture gate passed without increasing the threshold. Second, beta.66 had been successfully activated on the owner machine before this incident repair, even though its earlier failed-activation audit paragraph still used an unqualified present-tense statement that no beta.66 activation record existed. This section explicitly scopes that historical statement to the earlier failure point. Because repository policy requires every blocking post-activation defect to use a new prerelease number, the repaired source is beta.67; beta.66 activation evidence is retained as historical evidence and is not reused as acceptance for the new package bytes.
|
|
430
|
+
|
|
431
|
+
Residual boundary: the current MCP delivery model is request-scoped and intentionally has no durable terminal replay log. Short same-daemon outages are recoverable, but an outage beyond the bounded reconnect/request lifetime must eventually terminate rather than preserve side-effecting foreground work indefinitely. If the product requirement changes to guarantee recovery of user-visible terminal delivery across arbitrarily long disconnects, that is not another timeout tweak: it requires a durable request/result protocol with explicit idempotency and host receipt semantics, especially for side-effecting calls. Managed jobs remain the existing durable execution mechanism for long work.
|
|
432
|
+
|
|
433
|
+
## 2026-08-13 beta.66 failed-activation and post-candidate availability follow-up
|
|
434
|
+
|
|
435
|
+
The first prepared beta.66 candidate was not activated. The repository owner ran the required interactive-terminal activation command and it failed during the initial current-source/candidate promotion check with `npm pack dry-run returned an invalid file mode: 384`. Decimal `384` is mode `0600`; an independent npm dry-run isolated exactly two post-candidate packaged source files with that mode: `src/local/resource-foreground-wait.mjs` and `src/worker/daemon-ready-waiters.ts`. The promotion digest intentionally accepts only `0644` and `0755`, so the failure is fail-closed package hygiene rather than a validator-format bug. `start-release-candidate.mjs` computes that digest before creating the hardened npm session, installing a candidate runtime, deploying a Worker, handing off the service, or writing activation evidence. No beta.66 activation record exists. Both files were normalized to `0644`, and the subsequently extracted `daemon-recovery-budget.ts` was likewise normalized after creation.
|
|
436
|
+
|
|
437
|
+
That mode failure also exposed a more important generation problem: comparing the frozen tarball with a fresh dry-run package showed that the checkout had gained packaged runtime changes after the candidate was prepared. The delta included process execution/session admission, relay contract changes, Worker reconnect handling, and two new modules. The old tarball therefore cannot be made valid merely by changing file modes; it is stale evidence and must be replaced by a fresh candidate after the final tree passes the complete frozen gates. This incident also exposed a workflow ownership defect: the owner activation command was being used as the first post-preparation freshness check. The repository now requires direct Node argv `node scripts/start-release-candidate.mjs --install-only` immediately before owner handoff. That path reuses the activation helper for the same source/promotion/tarball preflight and disposable install without Worker/service activation or activation evidence, but resolves its preflight npm CLI from the running Node installation with lifecycle/fallback discovery disabled; project/user `script-shell` therefore cannot interpose before the agent freshness proof.
|
|
438
|
+
|
|
439
|
+
Independent review of those post-candidate changes found three additional availability defects before refreeze. First, the new ten-second new-call daemon reconnect grace ran before the Worker created the tool timeout budget and pending deadline. A nominal 60-second foreground call could therefore consume about ten seconds of recovery plus sixty seconds of execution plus five seconds of settlement, contradicting the beta.31 hosted-delivery contract that admission/transport latency consumes the existing interval rather than extending it. Short calls were worse: `timeout_seconds=1` could first wait ten seconds. The Worker now creates the original budget before waiting, caps reconnect grace by the execution budget, measures elapsed recovery with monotonic `performance.now()`, subtracts it before dispatch while preserving the five-second settlement margin, and refuses dispatch with retryable `unavailable` plus `side_effects_started=false` if recovery consumed the whole execution window.
|
|
440
|
+
|
|
441
|
+
Second, `resource-foreground-wait.mjs` introduced a proportional two-to-ten-second foreground admission default, but production `LocalRuntime` still converted an omitted `processResourceWaitMs` into the old explicit `2000` value before constructing either process service. The helper therefore passed isolated unit tests while being dead in normal runtime composition. The runtime now preserves an omitted override as `undefined`, rejects malformed explicit overrides, and retains the existing thirty-minute cap for deliberate test/diagnostic values. Production one-shot admission consequently uses 20% of the execution timeout with a two-second floor, ten-second ceiling, and never more than the whole execution timeout; process-session startup defaults to ten seconds. A runtime-composition regression requires both services to receive no explicit override in the ordinary constructor path.
|
|
442
|
+
|
|
443
|
+
Third, the first daemon-ready waiter implementation had its own 64-waiter limit, outside the Worker's established 30 ordinary + 2 reserved-control pending-call algebra. During a disconnect, ordinary new calls could therefore consume a parallel capacity pool and undermine the reserved recovery slots. Waiter admission now folds the current pending-call snapshot and all already-waiting tool names into the same shared admission function before adding a waiter. A deterministic regression starts with 29 detached ordinary calls, allows exactly one additional ordinary waiter, rejects the next ordinary waiter, still permits `diagnose_runtime` and `list_roots` into the two reserved slots, and rejects the next call at the combined 32-call ceiling. Ready notification, timeout, and cancellation still remove waiter state exactly once. A final observability pass also makes owner `server_info` distinguish already-dispatched calls from pre-dispatch daemon waiters and report their combined ordinary/reserved capacity usage; delegated activity remains hidden. Architecture budgets were kept intact by extracting the post-recovery timeout arithmetic instead of increasing `tool-timeout.ts`, and the three new availability modules are explicit critical-coverage targets rather than relying only on broad Worker/resource fixtures.
|
|
444
|
+
|
|
445
|
+
## 2026-08-13 beta.66 pre-commit npm-lifecycle follow-up
|
|
446
|
+
|
|
447
|
+
Beta.65 completed the intended live release-validation sequence before Git publication: the owner activated the exact candidate, activation evidence matched its SHA-1/SRI/promotion digest, the authenticated daemon and Worker both reported `3.0.0-beta.65`, the deployed OAuth canary passed authorization-code exchange/authenticated MCP/refresh rotation/refreshed MCP/cleanup, post-canary runtime and audit health remained clean, repository privacy checks passed, and local acceptance was recorded. No beta.65 commit or push had occurred.
|
|
448
|
+
|
|
449
|
+
The subsequent staged-diff review challenged a different layer of the same liveness exception. Proving which outer `npm` executable PATH selects is insufficient to prove what `npm run` executes next: npm honors `script-shell` configuration from project/user/environment configuration. An isolated temporary npm project with a project `.npmrc` pointing `script-shell` at a wrapper proved the wrapper runs before an otherwise ordinary package script while `npm run` still exits successfully. The beta.65 checkout itself had no such setting (`npm config get script-shell` returned `null`), so this does not invalidate the observed OAuth result; it invalidates the general zero-resource accounting guarantee claimed by the source. Beta.65 acceptance was therefore removed before commit, and beta.65 remains activated evidence only. It was never committed, pushed, tagged, released, or published.
|
|
450
|
+
|
|
451
|
+
Beta.66 removes npm lifecycle execution from the release-control exception rather than trying to enumerate every npm configuration source. A further independent pre-candidate review then rejected the first direct-Node variant because it still executed `scripts/release-oauth-canary.mjs` from the mutable checkout while admission compared only that entry file with the packaged copy. The entry could remain byte-identical while one of its local imported modules changed, allowing same-version checkout code to run under zero-resource accounting. No beta.66 candidate had been prepared or activated at that point, so the fix stays within the same beta.66 source generation. The only light form is now exact direct Node argv `node <activated-runtime-package>/scripts/release-oauth-canary.mjs --allow-live-oauth-canary`: the script operand must be absolute and canonicalize to the running package's own canary, while the checkout is only the command cwd and candidate/evidence data root. `resource-release-control-executable.mjs` independently resolves the actually invoked Node and requires its canonical identity to equal the daemon's own `process.execPath`; PATH shadowing therefore fails closed. The same proof rejects environment keys that can inject code, expose sensitive transport debugging, alter TLS trust/output, inflate runtime instrumentation, or materially alter Node/native startup before the entrypoint, including `NODE_OPTIONS`, `NODE_PATH`, `NODE_DEBUG`, `NODE_PRESERVE_SYMLINKS`, V8 coverage/compile-cache inputs, `NODE_TLS_REJECT_UNAUTHORIZED`, extra CA/certificate/key-log/OpenSSL settings, `UV_THREADPOOL_SIZE`, `LD_PRELOAD`/`LD_AUDIT`/library-path injection, and the corresponding DYLD injection paths. The canary runner applies this same policy itself before candidate validation or any live OAuth mutation, imports all local modules from the activated package, and uses `process.cwd()` for the current source/candidate/evidence data. For prereleases it additionally canonicalizes the executing package root and the package root containing activation `runtime_entry` and requires exact equality; workspace/developer copies therefore cannot generate release evidence even if their entry bytes match the candidate. For every release channel it also reuses the hardened daemon-owner verifier to require a current startup-ready service daemon with the candidate version, an `entryScript` canonically equal to the canary package's `bin/machine-mcp.mjs`, and a canonical Node executable equal to the canary's `process.execPath`, and daemon-lock Node runtime metadata equal to `process.versions.node`; this gives stable releases the same code-origin proof without adding a parallel activation-record format. Because canary evidence does not record alternate local-state identity or Node startup flags, the runner itself now requires exact single-flag argv, requires `process.execArgv` to be empty, and refuses `--state-dir`, `--workspace`, or any other extra argument before state access. This makes ordinary invocations with Node preload/loader/debugger/profiler/runtime CLI options fail even though they never received the release-control light profile. Workspace-relative Node, copied canary, shell wrappers, npm lifecycle invocation, alternate package managers, lookalikes, extra argv, and mismatched cwd/runtime entry bytes retain ordinary accounting. This is still cooperative resource accounting rather than a hostile same-UID execution-authentication boundary.
|
|
452
|
+
|
|
453
|
+
The canary still needs npm for its explicit `--ignore-scripts` pack dry-run used by the promotion-content digest, but it no longer trusts lifecycle-provided `npm_execpath`. `npm-cli.mjs` centralizes npm CLI discovery and, for the release canary, restricts candidates to layouts derived from the running Node installation. The first owner-machine probe exposed a Homebrew layout distinction: the Cellar Node binary did not contain npm under its versioned tree, while npm lived under the same package-manager prefix. The resolver now derives that prefix from a Cellar Node path and validates the resulting regular file/mode; a synthetic regression reproduces that layout. The beta.66 registry version, Git tag, and GitHub Release namespaces were all confirmed absent before the version bump. Fresh frozen verification, candidate creation, owner-terminal activation, direct-Node deployed canary, live observation, and acceptance are required for beta.66.
|
|
454
|
+
|
|
455
|
+
## 2026-08-13 beta.65 pre-commit executable-identity follow-up
|
|
456
|
+
|
|
457
|
+
Beta.64 completed the repaired live sequence successfully before Git publication: the owner activated the exact candidate; the activation record matched candidate SHA-1, SRI, and promotion digest; the authenticated daemon and recorded Worker deployment both reported `3.0.0-beta.64`; the canonical deployed OAuth canary started immediately rather than waiting on machine resource admission; and its evidence bound the exact beta.64 hashes while authorization-code exchange, authenticated MCP, refresh rotation, refreshed MCP, and cleanup all succeeded. Post-canary relay readiness, security-audit chain health, `content_logged=false`, and repository privacy checks were clean, and local acceptance was recorded. No commit or push had occurred yet.
|
|
458
|
+
|
|
459
|
+
The staged-diff review then found a remaining consistency defect in the resource-control liveness exception. Generic arbitrary-process light admission had already stopped trusting PATH-resolved basenames, but the beta.64 release-control path still accepted lexical `npm run release:oauth-canary` plus cwd/runtime package identity without proving which executable the daemon PATH would actually resolve for the outer `npm`. That is not an OAuth authorization bypass and resource admission remains cooperative rather than a hostile same-user security boundary, but it can still under-account a wrapper/lookalike and contradicts the executable-identity rule adopted for the adjacent light-command family. Because this was found before commit, beta.64 acceptance was deleted and beta.64 remains activated evidence only; it was never committed, pushed, tagged, released, or published.
|
|
460
|
+
|
|
461
|
+
Beta.65 adds `resource-release-control-executable.mjs` at the async process-admission boundary. For a bare npm command it walks the actual execution search order and stops at the first executable match; an untrusted first match is rejected rather than skipped in favor of a later safe-looking npm. The selected target is canonicalized, constrained to npm/npm-cli target names, rejected when it resolves inside the command cwd or HOME, and on POSIX rejected when group/world writable. Explicit absolute npm paths go through the same canonical target and ownership-surface checks. Only after this proof, the exact npm `run release:oauth-canary` lexical shape, and the existing runtime-bound package/version/entrypoint proof all pass does the command receive `release-control/light`; every failure falls back to ordinary accounting. Regression fixtures cover a trusted PATH npm, a HOME-shadowing first match, writable and missing executables, a cwd-local executable, explicit absolute invocation, and a non-npm target. The current owner machine's real npm resolves through this proof successfully.
|
|
462
|
+
|
|
463
|
+
The `3.0.0-beta.65` npm version, Git tag, and GitHub Release were confirmed unused before the bump. Since executable-resolution logic is packaged runtime code and beta.64 acceptance is no longer valid for these bytes, beta.65 must repeat frozen verification, exact candidate preparation, owner-terminal activation, deployed canary, observed live verification, and acceptance before Git publication.
|
|
464
|
+
|
|
465
|
+
## 2026-08-12 beta.64 live release-admission follow-up
|
|
466
|
+
|
|
467
|
+
The owner activated the exact beta.63 candidate and repository activation evidence matched its SHA-1, SRI, and promotion digest. The live daemon and Worker both reported `3.0.0-beta.63`, relay readiness was verified, and doctor/Worker health were clean. The next required deployed OAuth canary, however, revealed a separate resource-scheduling defect before any OAuth request was sent. Its managed runner stayed at step zero for roughly twenty minutes while unrelated long-running workflows held the machine-user I/O reservation. OS process inspection showed the canary runner had no spawned step, and cancellation settled with `steps=[]`; `.release-candidate/oauth-canary.json` remained absent. This establishes that no release-canary account/client mutation began, so there is no ambiguous remote cleanup state to carry into the repaired candidate.
|
|
468
|
+
|
|
469
|
+
The cause is the command-profile lexical heuristic rather than host pressure or OAuth behavior. `packageManagerTokensHeavy()` intentionally treats package scripts containing `release` as orchestration because release candidates commonly fan out into build/test/package work. The documented `npm run release:oauth-canary -- --allow-live-oauth-canary` therefore inherited the generic `js-build` reservation of 2 CPU, 0.5 I/O, and 2048 MiB even though its implementation performs bounded candidate/activation evidence reads, an `--ignore-scripts` npm pack dry-run for promotion identity, and network control-plane requests. On this host two unrelated low-runtime-pressure workflows legitimately held enough conservative I/O reservation to keep that false build request queued indefinitely. `resource-release-control-classification.mjs` now owns one narrow exception for direct npm argv targeting the exact `release:oauth-canary` script; shell-wrapped forms, alternate package managers, `release:candidate`, and lookalikes remain heavy. The release architecture contract also pins the canary script body and forbids `prerelease:oauth-canary`/`postrelease:oauth-canary` hooks, preventing npm lifecycle expansion from silently adding unclassified local work. The exception is runtime-bound rather than globally name-bound: `resource-release-control-workspace.mjs` requires the command cwd package name/version to equal the running daemon package, rejects workspaces and lifecycle hooks, and compares the cwd canary entrypoint with the runtime-packaged entrypoint using bounded descriptor-verified reads. Thus an unrelated same-name script remains heavy, and beta.63 cannot grant beta.64's liveness exception before beta.64 is actually activated. This is a cooperative local resource-accounting correction, not a relaxation of authorization, release evidence, OAuth mutation, candidate identity checks, or a hostile-same-user code-authentication boundary.
|
|
470
|
+
|
|
471
|
+
Adversarially composing that first repair exposed a broader pre-existing shell classifier weakness before any gate was run. The old `isLightCommand()` accepted shell text by matching a known-light prefix and rejecting only a small set of explicitly named heavy commands after control operators. A payload such as `echo ready; chmod ...` could therefore inherit a zero-resource bypass even though later shell work had never been classified. Trying to prove a growing set of shell syntax safe was itself the wrong abstraction: startup profiles, environment-controlled executable resolution, wrappers, and PATH lookalikes can change execution before or beneath the apparent payload. Beta.64 therefore removes the generic arbitrary-shell light class entirely. Shell commands remain adaptive or are promoted by the existing heavy shell/build classifiers, and the release-control exception is direct npm only. This eliminates the shell laundering family rather than extending a denylist of syntax cases.
|
|
472
|
+
|
|
473
|
+
Reviewing the positive direct-command allowlist found the same executable-identity issue. A bare `ps`, `uptime`, `sleep`, or other apparently cheap basename is still resolved through the daemon environment PATH and therefore does not prove which executable will run; read-oriented heads such as `find`, `awk`, `osascript`, `open`, ordinary Git, filesystem metadata/lookup tools, search, and file-stream processors can additionally launch helpers or perform work proportional to caller-selected input. Generic zero-resource admission is now identity-bound in `resource-light-command.mjs`: only a small set of standard absolute executables qualify, while the same basenames reached through PATH remain adaptive. This mirrors the existing disk-reclaim rule that already refused PATH-resolved deletion. Implementation-owned Git/diagnostic liveness is unchanged because those operations execute through fixed code-constructed argv with isolated settings and bypass arbitrary-process admission as an internal control-plane boundary.
|
|
474
|
+
|
|
475
|
+
Re-reviewing the beta.63 Linux state-root fix found a narrower post-rename check/use interval that the original provider regression did not exercise. The POSIX descriptor pin already prevents the original directory inode from being recycled, but the full moved-root safety traversal occurs after the first retired-path device/inode check. A concurrent replacement installed during that traversal could therefore survive verification and become the pathname passed to recursive removal. Beta.64 keeps the descriptor open and rechecks the retired pathname against that pinned device/inode again after the complete safety traversal; a deterministic fixture replaces the quarantine from inside the verifier and requires the replacement to remain as non-reclaimable evidence. Portable Node.js still cannot express recursive `openat`/`unlinkat` deletion bound entirely to a directory capability, so a determined same-user process can race the final identity check and pathname `rm`; the architecture now states that residual explicitly instead of claiming full hostile-same-user race elimination.
|
|
476
|
+
|
|
477
|
+
Because these changes are packaged local runtime source, beta.63 activation cannot authorize beta.64 publication. Beta.63 remains activated but unaccepted. The repaired source is versioned `3.0.0-beta.64` and must pass a new frozen gate, exact candidate preparation, owner-terminal activation, and the deployed canary. The strongest live regression is the canary itself: after beta.64 activation it must start without waiting for unrelated build reservations and still produce candidate-bound authorization-code/MCP/refresh/cleanup evidence before acceptance.
|
|
478
|
+
|
|
479
|
+
## 2026-08-12 beta.63 exact-head provider follow-up
|
|
480
|
+
|
|
481
|
+
Beta.62 completed every local and live release stage before hosted verification: the exact candidate was activated from the owner TTY, its candidate/activation identities matched, the deployed OAuth canary completed authorization-code exchange, authenticated MCP, refresh rotation, refreshed MCP, and both cleanup operations, the verified launchd service/Worker reported `3.0.0-beta.62`, privacy-bounded live log inspection found no credential/raw-home patterns, acceptance was recorded, and guarded push produced exact PR head `7e3bcfe8a6b4fe85f867117ada3a965c7bc2e1a7`. That evidence remains valid for the bytes it observed, but it does not authorize publication after provider failures. The exact head passed Dependency Review, Governance, workflow policy, package audit, and CodeQL Actions, while Ubuntu full, Windows platform, and JavaScript/TypeScript CodeQL failed. Per the release contract, beta.62 acceptance is therefore invalidated rather than grandfathered into a repaired source generation.
|
|
482
|
+
|
|
483
|
+
Ubuntu failed `state-root-retirement:test` at the replacement-generation fault injection. The pre-rename guard compared full lossless filesystem identity, including change-time generation, but the post-rename guard intentionally compared only device/inode because a legitimate rename may change directory ctime. On Linux the injected remove/recreate could immediately recycle the just-freed inode before the replacement was renamed into quarantine, so the replacement satisfied that numeric post-rename pair and incorrectly reached `verifyMovedRoot`. The repair keeps the original POSIX directory open through an `O_NOFOLLOW|O_DIRECTORY` descriptor after matching its full pre-rename generation. The descriptor pins the original inode allocation across the rename window; the moved pathname must then match that descriptor's device/inode before verification or recursive deletion. A changed generation returns false and remains quarantined. This closes the ABA mechanism without treating rename-mutated ctime as stable evidence.
|
|
484
|
+
|
|
485
|
+
Windows failed later for a separate fixture boundary. `testSessionAuthorityRevocation` intentionally set `terminationSettlementWaitMs=100` so a synthetic child that reports successful termination delivery but never closes can prove revocation remains retryable. That test-level override also governed the real spawned session left for final `clearAndWait`; Windows process-tree termination and libuv close can legitimately exceed 100 ms, so the fixture manufactured `process-session shutdown was incomplete` even though production uses a five-second settlement budget. The short deadline is now installed only around the artificial never-close record and is restored in `finally`; real session revocation/cleanup continues under the production budget. The focused process-output lifecycle regression passes with that separation.
|
|
486
|
+
|
|
487
|
+
CodeQL independently rejected ten findings rather than sharing either platform cause. Five were obsolete third arguments to the two-argument resource aggregation API plus the corresponding production coordinator call; one modeled the Promise-valued host-snapshot single-flight entry as a missing-await flow; and three were check/use file-race reports in verification/test code. The dead arguments were removed. The host-snapshot map now stores a stable `{ promise }` flight record and awaits only its Promise, preserving single-flight semantics while making lifecycle identity explicit. Verification-generation hashing no longer performs an ordinary path read after type inspection: regular-file bytes/mode come from the existing bounded descriptor read with post-open path-identity verification. Secure-file and Windows launcher tests likewise use identity-coherent read evidence or separate temporary pathname generations instead of a check followed by later reuse. No CodeQL threshold or acceptance entry was added for these new findings.
|
|
488
|
+
|
|
489
|
+
An independent clean detached worktree of the same beta.62 head ran `npm ci` and `CI=true npm run check:full` on macOS and passed all 117 tasks, which distinguishes the Linux/Windows provider-specific failures from a general exact-head regression. At this point in the sequence the repaired source was `3.0.0-beta.63`; its npm version, remote tag, and GitHub Release had been confirmed unused before the bump, and beta.62 remained the live-verified runtime pending beta.63 owner-terminal activation. The later beta.64 section above records the subsequent beta.63 activation and canary-admission rejection.
|
|
490
|
+
|
|
491
|
+
## 2026-08-12 beta.62 rejected-candidate and release-gate concurrency follow-up
|
|
492
|
+
|
|
493
|
+
The first owner-terminal activation of beta.61 was deliberately not converted into acceptance after the deployed OAuth canary exposed a cross-layer success-status mismatch. `/admin/clients` deletion commits client/code/token cleanup and returns JSON `200`, whereas the activated beta.61 local client required `204` for every DELETE. The repaired local client now distinguishes OAuth-client deletion (`200`) from account deletion (`204`), account creation (`201`), and the remaining `200` operations, with reverse-status regressions. Because that repair changes packaged bytes after activation, beta.61 remains an activated but unaccepted/unpublished candidate and the repaired source is versioned beta.62; none of beta.61's activation evidence can substitute for a fresh beta.62 candidate activation and deployed canary.
|
|
494
|
+
|
|
495
|
+
That cleanup incident also exposed a broader settlement invariant at the local admin-client boundary. A mutation can commit remotely before the client discovers that an otherwise successful response used the wrong 2xx status or an empty, oversized, malformed, or non-object body. Treating those cases as a bare `protocol_error` omitted the fact needed to prevent unsafe automatic retry. Successful-response protocol failures are therefore split by method: reads keep ordinary protocol-error semantics, while mutations are non-retryable and record only `request_delivery=sent` plus `effect_settlement=unknown`. The response body and remote error content remain excluded from these settlement details. Reverse-status DELETE tests and a malformed successful PATCH response pin this behavior independently of the deployed canary.
|
|
496
|
+
|
|
497
|
+
The same credential-boundary review found that signed account-administration fetches still inherited the platform default redirect-following policy. Those requests can carry device-session proof headers and mutation bodies, so a Worker or intermediary redirect must not cause a second request to a Location target. Account administration now sets `redirect=error` on every request. The release OAuth canary applies the same fail-closed default to dynamic registration, token exchange, and authenticated MCP calls; its authorization request alone uses `redirect=manual` because the expected `303` callback is itself validated as data. Focused fixtures assert both policies. This is defense in depth against unexpected redirect behavior and does not rely on assumptions about which headers a particular Fetch implementation strips across origins.
|
|
498
|
+
|
|
499
|
+
The beta.62 freeze then surfaced a separate verification-lifecycle race only because two safe verification jobs intentionally overlapped. The complete 117-task gate passed on one job while the concurrent fast gate failed `worker-types-generator:test` with a normal Wrangler dry-run reported as a completed-but-resident CLI. The fixture itself is process-isolated, so the divergent result was traced to the shared completion wrapper rather than shared test files: after observing the completion marker, its short grace timer could run while a naturally exiting child still had `exitCode === null`; `requestTermination()` set the cleanup flag, and a later `close(code=0)` was classified by that flag before the actual zero exit code. The wrapper now gives a real zero exit precedence over the raced cleanup request and accepts the bounded-hang path only when the completed child actually closes from `SIGTERM`. A deterministic synthetic mode installs a no-op SIGTERM handler, emits the completion marker, crosses the short grace, and then exits zero; the old ordering necessarily mislabeled it, while the corrected ordering does not. The existing resident-after-completion fixture still closes by SIGTERM and must retain the cleanup diagnostic. This is release-gate correctness rather than a runtime feature change, but it invalidates the preceding beta.62 full-gate result as final evidence and requires a new frozen run.
|
|
500
|
+
|
|
501
|
+
That incident was followed by a second, distinct evidence-integrity failure: independent work on the same checkout changed coverage policy and audit files while a nominally frozen full gate was still running. The resulting output could no longer establish anything about one exact source generation, even if every individual task were green. Coverage already had a narrower source-generation check, but it protected only its own fixture interval and input subset. The fast/full/platform entrypoint now computes a cryptographic generation over the complete verification input surface before starting the plan and again after it settles, including source, scripts, tests, browser extension, GitHub workflows/scripts, documentation, native/bin content, release acceptance/candidate evidence, top-level build/release/privacy configuration, and POSIX mode or symlink topology. Any drift invalidates the whole run with an explicit “discard this run” error. If a task also failed while inputs changed, generation drift wins because the mixed-generation task failure is not reliable causal evidence; if inputs are stable, the original failure object is preserved. Unit and architecture tests pin both cases. The Wrangler lifecycle regression file was separately found at accidental `0600` mode and restored to ordinary `0644`; the mode itself participates in the new generation identity.
|
|
502
|
+
|
|
503
|
+
The next architecture pass caught a simpler structural regression before that rerun: the account-administration client had grown to 250 lines against its 240-line responsibility limit as response-size, JSON-shape, remote-error, mutation-settlement, and endpoint-status hardening accumulated. The budget was not raised and formatting was not compressed to hide the growth. The HTTP response boundary now lives in `account-admin-response.mjs`, which owns strict endpoint success status, the one-MiB streamed response ceiling, fatal UTF-8/JSON-object validation, bounded terminal-safe remote error text, response cancellation, and delivered-but-unsettled mutation protocol errors. `account-admin.mjs` returns to account command construction, device-session signing, request dispatch, and account identifier normalization. Architecture caps the new helper at 120 lines; the stabilized source measures 162 lines for the client and 92 for the response helper, with the new packaged source normalized to `0644`. The existing account-admin regression remains end-to-end through the client and passes after extraction. An independent coverage review then caught that the extraction had moved security-relevant response semantics outside the old `account-admin.mjs` threshold even though the fixture itself still exercised them. The critical coverage table now names `account-admin-response.mjs` explicitly at 100% functions / 75% branches; the first gated run measured 100% (9/9) functions and 85.7% (36/42) branches while the caller remained 100% functions / 75% branches. No threshold was lowered to obtain that result.
|
|
504
|
+
|
|
505
|
+
## 2026-08-11 beta.61 hosted OAuth persistence incident and independent re-review
|
|
506
|
+
|
|
507
|
+
A ChatGPT hosted-connector failure exposed both a production persistence defect and a diagnosis-process defect. The user-visible connector message was a generic account-connection failure and was deliberately not accepted as root-cause evidence. An exact beta.38 rollback control still failed, proving that the current beta.61 feature set alone was insufficient to explain the outage. Fresh dynamic registration and account authorization succeeded, and the resulting client became trusted to the owner account, but admin inventory showed zero active access tokens and zero active refresh tokens. Production Worker tail then showed `/oauth/token` returning an unexpected 500.
|
|
508
|
+
|
|
509
|
+
The investigation initially spent too much time on plausible branches before adding sufficiently discriminating instrumentation. DPoP interoperability, a cross-request Promise mutex, and refresh-state growth were each investigated. A privacy-bounded request-shape check proved that the failing ChatGPT token requests were form encoded and carried no DPoP header, eliminating DPoP from this incident. Replacing the old instance-level `oauthQueue` Promise chain with a `DurableObjectState.blockConcurrencyWhile()` boundary did not restore connectivity. Compaction of expired refresh-retry payload materially reduces retained refresh-state size and remains useful capacity hardening, but it also did not move the failing stage. These changes must therefore never be cited as the causal repair for this outage; any retained change requires its own invariant and regression evidence.
|
|
510
|
+
|
|
511
|
+
A fixed-enum authorization-code stage discriminator then located the failure after OAuth-store load, refresh-store load, PKCE verification, and token-pair issuance but during persistence: the deployed Worker reported `oauth_token_stage_persist`. The persistence path used the supported Durable Object multi-entry convenience form to update the main OAuth and refresh stores together. Local Wrangler/workerd integration had already exercised authorization-code exchange, refresh rotation, and concurrent refresh retry successfully, so the production failure demonstrated a real parity gap between that local evidence and the deployed edge behavior. The project does **not** claim that Cloudflare multi-key `put()` is unsupported; current API/runtime sources support it, and the exact provider-internal reason this deployed path failed remains unresolved.
|
|
512
|
+
|
|
513
|
+
The production repair replaced that security-critical convenience write with one explicit Durable Object storage transaction containing two named single-key writes. It preserves atomicity while avoiding the deployed failing path and attaches only bounded stage plus error-class diagnostics (`oauth`, `refresh`, or `commit`) without request values or raw exception messages. After owner activation of the exact rebuilt beta.61 candidate, the ChatGPT connector completed authorization and an authenticated `server_info(full)` call returned through Worker, relay, and the local daemon. A subsequent fixed `diagnose_runtime` call also reached the local runtime. Worker observability reported `oauth_refresh.rotated=1`, proving that at least one refresh-token rotation had also committed through the repaired persistence path; the result is stronger than a UI-only connection indicator.
|
|
514
|
+
|
|
515
|
+
The independent follow-up found the same risky convenience-write pattern still reachable through the authority-revocation helper when a protected multi-record write carried no revocation record. That branch is now routed through the same explicit transaction/per-key persistence discipline, and its regression rejects any return to an outer bulk-object `put()` while requiring all protected records to commit. The helper also refuses writes that try to replace its reserved `authority-revocations` key. The test storage transaction was upgraded from callback-on-self behavior to copy-on-commit semantics so rollback/fault-injection tests do not accidentally validate partial state that a real transaction would abort. A worst-shape 1024-entry authority queue is now regression-budgeted below 512 KiB of JSON envelope, providing explicit headroom rather than relying on the count cap alone.
|
|
516
|
+
|
|
517
|
+
A later state-shape pass found that the same authority queue budget was only true for state produced by current code. Its schema-1 validator checked known field values but accepted and spread arbitrary additional queue/record fields back into memory; damaged or unexpected persisted state could therefore carry an unbudgeted payload through the next transaction despite the 1024-record cap. Schema-1 validation now rejects unknown top-level and record fields, and deterministic tests cover both forms. This intentionally prefers fail-closed repair over silent forward-field preservation for a versioned security-control queue.
|
|
518
|
+
|
|
519
|
+
The audit then followed that mechanism rather than stopping at the first instance. OAuth's deep validators bounded every known value but likewise treated unknown object fields as harmless. Because current store objects are subsequently mutated and persisted again, that behavior could preserve an unbudgeted or sensitive payload that no current schema owns. The fix centralizes exact field sets in a dedicated OAuth structural contract and applies them to the schema-1 main OAuth store, schema-3 refresh store, refresh tokens/replay markers/revoked-family records, and schema-1 consumed-token shards; identifier/value grammar remains in the separate record contract. The accepted field sets include the beta.61 optional `registration_revision` and `access_scope` additions; the beta.55 persisted shape therefore remains representable, while unsupported same-version fields now fail closed with the existing repair-required path. Regression fixtures cover unknown main-store, client, refresh-envelope, replay-marker, shard-envelope, and shard-marker fields.
|
|
520
|
+
|
|
521
|
+
The same bounded-state review found a smaller version of the issue in the generic Worker nonce store used by daemon preflight, account administration, and DPoP replay protection. Callers supplied maximum cardinalities, but those maxima governed only new insertion after a full persisted map had already been accepted and copied; the module also exposed an unused read helper with no cardinality parameter at all. Persisted nonce maps are now validated against the caller's maximum before normalization, oversized state fails closed rather than being silently repaired through an unbounded read, and the dead helper is removed so replay state has one validation/mutation path.
|
|
522
|
+
|
|
523
|
+
That replay-state boundary also needed a temporal cap. The persisted validator accepted any positive safe-integer expiry, so damaged state could stay within the entry-count limit while setting every marker arbitrarily far into the future and permanently denying new proofs. The three callers have explicit proof-time windows: daemon preflight and admin authorization accept timestamps within their TTL, and DPoP within its replay window; in each case a valid proof's farthest expiration is at most twice that window ahead of the current clock. The shared nonce store now requires a caller-supplied maximum future window and rejects persisted expirations beyond it before copying or pruning the map.
|
|
524
|
+
|
|
525
|
+
The local error-contract pass found one misleading branch family in the account-admin client: malformed or oversized successful Worker responses constructed `BridgeError("invalid_response")`, even though `invalid_response` is not a supported bridge error code and the constructor therefore normalized it to `execution_failed`. No caller depended on the nonexistent code. The response-contract failures now explicitly use the existing `protocol_error` classification, and tests assert the code as well as the bounded message so source intent and wire behavior cannot drift silently again.
|
|
526
|
+
|
|
527
|
+
That pass also distinguished MCP result sanitization from direct local CLI behavior. `AccountAdminClient` is used by the terminal account command directly, so a non-success Worker JSON `message` or `error` could previously become a large control-character-bearing local exception before any MCP public-error projection existed. Remote admin error text is now normalized and capped at 2,000 characters at the client boundary itself, including removal of terminal controls and Unicode bidirectional/invisible controls. A hostile synthetic Worker response verifies the direct exception is bounded and control-free.
|
|
528
|
+
|
|
529
|
+
The same branch review found that every admin HTTP status other than 401/404/409 was classified as `invalid_request`. The Worker implementation makes its ordinary domain failures explicit 400/404/409 responses and converts uncaught route failures to 500, so a 5xx is not evidence that the operator supplied invalid input. The client now classifies 5xx as `unavailable`. Read-only admin calls may retry; mutation calls do not, and carry delivered-but-unsettled details because an HTTP 5xx proves the request reached the Worker but does not prove whether a storage mutation committed before the failure.
|
|
530
|
+
|
|
531
|
+
The response-state review then found that all 204 responses were accepted before method-specific validation. That meant an impossible GET/PATCH/POST 204 could be normalized to `{removed:true}`, and other unexpected 2xx statuses were likewise treated as valid JSON successes. The client now enforces the Worker contract explicitly: DELETE expects 204, account creation expects 201, and current read/update/rotate operations expect 200. An unexpected 2xx is a protocol error and its body is cancelled without being trusted.
|
|
532
|
+
|
|
533
|
+
The redundancy pass also removed `remoteBridgeError()`. It had no production consumer anywhere under `src/`; the only reference outside its definition was a unit assertion that instantiated it directly. With the old stateful remote/session result paths already deleted, preserving a separately tested error-deserialization contract that no executable route could reach increased maintenance surface and created a misleading signal that remote structured errors were reconstructed through that helper. Current transport-specific error paths keep their own explicit settlement and exposure boundaries instead.
|
|
534
|
+
|
|
535
|
+
The storage-capacity review also showed that refresh replay compaction by itself did not establish a durable single-value bound. With 500 active refresh records, 4096 consumed markers, 1024 revoked-family tombstones, and a legitimate short retry-window burst, the old aggregate `oauth-refresh` value could still approach or exceed the provider value ceiling even after old markers compacted. The final persistence format therefore keeps active refresh tokens and revoked-family tombstones in the main `oauth-refresh` value but moves consumed-token replay markers into 8 deterministic hash shards, each capped at 1024 records. A worst-shape full retry-window shard is regression-budgeted below 1.1 MiB, while the global replay cap remains 4096, cutting refresh-state write fan-out roughly in half compared with the initial 16-shard draft without approaching the provider single-value ceiling. Existing schema-2/schema-3 main values remain readable; load reconstructs one in-memory `consumed` map, validates shard ownership/duplicates, compacts/prunes as before, and atomically migrates legacy main-value markers into the fixed shard set. Every token/admin/authority save path goes through the same persistence codec, so no branch can silently re-aggregate consumed replay state into the main value. Tests cover legacy migration, empty main `consumed`, shard cardinality/headroom, over-cap fail-closed behavior, refresh retry/replay semantics, and the absence of direct `oauth-refresh` write sites outside the codec.
|
|
536
|
+
|
|
537
|
+
The incident also changed the project evidence rules. Production diagnosis now records **observed fact**, **inference**, **falsified hypothesis**, and **unknown** separately; privacy-safe stage evidence or a minimal control experiment precedes semantic patching when the cause is ambiguous; a disproved patch is removed before finalization unless an independent invariant justifies it; credentials, accounts, proof-of-possession, state, and remote resource identity are not generic diagnostic reset knobs; and hosted-runtime behavior requires a deployed canary when local emulation cannot prove equivalent semantics. Verification evidence must also come from a frozen source tree. During this independent follow-up, one full local gate was mistakenly allowed to overlap a Worker source edit; Wrangler correctly hot-reloaded and a WebSocket integration assertion failed with close code 1006. That run was discarded as invalid evidence rather than retried or reported as a product failure, and the rule now explicitly requires completing all writes before the final full gate.
|
|
538
|
+
|
|
539
|
+
The independent documentation pass also found a protocol-contract contradiction introduced during the ChatGPT compatibility repair. An earlier beta.61 audit correctly removed the **stateful** `2025-11-25` session/recovery implementation, but its wording generalized that result into “only `2026-07-28` is executable”. The final Worker now intentionally has a much narrower remote-HTTP interoperability layer for `2025-06-18` and `2025-11-25`: only `initialize`, `notifications/initialized`, `ping`, `tools/list`, and `tools/call`; tools are dispatched through the current request-scoped controller; `Mcp-Session-Id`, recovery GET, `Last-Event-ID`, replay state, and the removed prepare/subscribe/session architecture stay absent. stdio remains `2026-07-28` only. Current engineering, upgrading, testing, changelog, and release-contract documentation now separates this stateless host compatibility from native protocol/state support so a future “legacy cleanup” cannot accidentally delete a proven hosted-client boundary.
|
|
540
|
+
|
|
541
|
+
The follow-up current-doc scan found one last copy of the earlier over-generalization in `CLIENTS.md`: it still said the product “supports MCP `2026-07-28` only” even though the same beta.61 source deliberately serves the bounded remote-HTTP initialization compatibility dates. The client guidance now distinguishes native/current-only stdio and `2026-07-28` request semantics from the narrow stateless `2025-06-18`/`2025-11-25` HTTP method set, without implying any restored session or replay model.
|
|
542
|
+
|
|
543
|
+
The compatibility path itself was then checked as a potential policy bypass rather than merely a protocol adapter. It already shares Origin, OAuth/DPoP, bounded body/media validation, request-ID, role-visible tool and input-schema gates, and forces compatibility tool responses through the current JSON-only controller path. One ambiguity remained: because old clients are allowed to omit the beta.61 mirrored routing headers, the adapter ignored those headers even when a caller supplied contradictory values. Compatibility now treats `Mcp-Method`, applicable `Mcp-Name`, and schema-declared `Mcp-Param-*` headers as optional-but-binding: absence remains compatible, but presence must match the body before dispatch. Unit regressions cover mismatched method/name and a matching mirrored tool call.
|
|
544
|
+
|
|
545
|
+
That routing review found a second ambiguity shared by native and compatibility requests: validation checked every mirror the selected method/tool expected, but silently ignored reserved mirror headers that were not applicable. A direct same-origin or non-browser caller could therefore attach `Mcp-Name` to a method with no name mirror, or attach undeclared/other-tool `Mcp-Param-*` headers, creating intermediary-visible semantics that the Worker did not own. The mirrored namespace is now closed per request: inapplicable names and any parameter header outside the selected tool's declared bindings fail with `-32020` before authorization-dependent dispatch. Compatibility still permits the current mirrors to be absent; it no longer permits them to be semantically extraneous.
|
|
546
|
+
|
|
547
|
+
The same independent pass found two local lifecycle defects that were not exposed by the green fast gate. Retained isolated POSIX resource leases correctly survive leader exit while descendants remain, but the group-liveness check ran before process-generation validation; after the original group disappeared, PID/PGID reuse by unrelated work could therefore make a stale lease look permanently live. Reclamation now gives an explicit `pid_reused` identity mismatch precedence over numeric group liveness while preserving the legitimate descendant case. Separately, `ProcessSessionManager.start()` checked cancellation again after spawn/bind but previously threw without terminating the just-created child or returning its session handle. That post-spawn cancellation branch now removes the unreachable session and force-terminates its tree, leaving close handling to release the already-bound lease exactly once. Deterministic regressions cover both mechanisms.
|
|
548
|
+
|
|
549
|
+
The same PID-generation mechanism also applied to crash-only resource staging recovery. Internal exclusive-file staging names encode a publisher PID but not its start identity; recovery therefore treated any currently live reuse of that number as proof that an uncommitted `.tmp` still had a live owner and could block every subsequent lease/waiter scan indefinitely. The staging file's own write time is now used as an ownership lower bound with the existing process-instance inspector: a genuine publisher necessarily started before it wrote the staging file, while a process generation that started materially later is reclaimable. A regression backdates a staging artifact beneath the current numeric PID and proves it is reclaimed, while the existing live-publisher case remains protected.
|
|
550
|
+
|
|
551
|
+
Documentation was also rechecked against executable protocol behavior rather than copied audit prose. `TESTING.md` and `THREAT_MODEL.md` still described `subscriptions/listen` as a zero-capacity rejection even though the current controller and integration test deliberately accept a bounded valid filter, acknowledge an empty supported subset, emit correlated completion, and close without persistent subscription state. Those stale claims are corrected to match the actual finite stateless contract.
|
|
552
|
+
|
|
553
|
+
Relay control handling had one analogous generation-gating omission. `tool_call` and terminal-result acknowledgement already require the authenticated ready relay generation, while `cancel_call` was accepted before that gate and validated only as a string. Worker handover does not require this exception: after a reconnect probe succeeds, the Worker rebinds pending calls and sends `resume_calls`, queued revocations, then `ready_ack` synchronously on the same ordered WebSocket before later cancellation can target that rebound socket. `cancel_call` is therefore now accepted only in the ready generation and only for the canonical call-ID grammar; pre-ready or malformed frames are protocol violations and cannot mutate local call state.
|
|
554
|
+
|
|
555
|
+
The revocation-delivery path was separately challenged for a possible "mutation succeeded but local authority stayed live forever" failure. The Worker side does preserve the core durability invariant: a revocation-producing account/OAuth save validates and updates the durable revocation queue in the same storage transaction as the authority mutation, so invalid queue state aborts the mutation rather than committing without a cleanup record; transient post-commit queue-read or relay-send failure also leaves the record durable. A later independent local-runtime pass found the missing half of that proof. `ManagedJobManager.revokeAuthority()` caught per-job inspect/cancel failures, logged them, and returned a partial count, so the daemon could still send `authority_revoke_ack` and let the Worker delete the durable queue record even when a matching managed job had not been cancelled. That was a genuine fail-open acknowledgement path. Revocation application now attempts every job but throws if any matching/possibly matching durable state cannot be safely inspected or cancelled; the control handler withholds acknowledgement and interrupts the current relay generation, causing the still-persisted queue to replay promptly during the next verified reconnect. Partial successful cancellations are idempotent on replay. The follow-up also proved that this internal safety path must not reuse the public `cancel_job` policy gate: a durable job accepted while writes were enabled could otherwise become impossible to revoke after the daemon policy narrowed. Revocation now calls a module-internal cancellation state machine that keeps maintenance, ownership, transition-lock, and persistence checks but intentionally bypasses only public tool availability. Tests prove public cancellation is denied under the narrowed policy while authority revocation still cancels the retained job. Tests also hold a matching job transition lock to force the failure, prove no acknowledgement is emitted, prove the relay generation is interrupted, then release the conflict and prove the retained revocation cancels the job on retry. `MULTI_ACCOUNT.md` now documents this complete request-scoped, durable acknowledgement contract rather than the deleted MCP session model.
|
|
556
|
+
|
|
557
|
+
A later managed-job deletion review found that terminal evidence validation still did not by itself bind the final recursive directory removal to the object that had been inspected. Retention could prove one job generation complete and then remove the same pathname after another generation replaced it. Whole-directory retirement now revalidates the lossless observed filesystem generation, atomically renames that exact directory to an internal name carrying device/inode identity, revalidates the moved object, and only then recursively deletes it. The review then challenged the crash path rather than stopping at the normal path: a process can die after rename but before deletion, so retired generations are recognizable persistent state. A later maintenance pass may reclaim only a generation whose encoded identity still matches; wrong-type, mismatched, or unreadable retired state is surfaced through active-state inventory as `unreadable` and blocks uninstall. Full uninstall attempts valid retired cleanup only while already holding its state-maintenance and machine-service locks, then repeats the fail-closed inventory check. Deterministic tests cover same-name replacement, valid crash residue, corrupt encoded identity, wrong-type residue, state-inventory projection, and uninstall behavior. The new retirement module is also in the critical coverage contract and currently measures 100% function / 72.7% branch coverage against a 90/70 gate.
|
|
558
|
+
|
|
559
|
+
After a frozen 115/115 local gate passed, the review deliberately continued and found that the first retirement namespace draft (`job_retired_*`) still matched the public managed-job ID grammar (`job_` plus at least 24 canonical ID characters). That made one internal crash residue eligible for a second ordinary-job/list/lock scan; a valid retired generation could therefore be projected once as cleanup-pending and again as an unreadable job, while an abnormal entry could leak its encoded device/inode through generic job diagnostics. The existing tests used `.some(...)` and therefore missed the duplicate projection even though the full gate was green. The internal namespace is now `retired_job_*`, which is outside the public grammar. Active-state inventory reports only a privacy-bounded `state_kind=retired_managed_job` with cleanup-pending/unreadable status and no internal filename/job ID, uninstall emits only a coarse `retired-managed-job:<status>` blocker, and public job/lock scans cannot reinterpret the directory. Recognized retired entries still count against the hard 50-item retained-state capacity until safely reclaimed, so namespace separation cannot become a capacity bypass. Regressions now require exact retired projection cardinality, no public-job-ID overlap, no filesystem-identity text in the uninstall error, and capacity refusal with 49 staged jobs plus one abnormal retired record. That source edit invalidates the earlier 115/115 run as final evidence and requires another frozen full gate.
|
|
560
|
+
|
|
561
|
+
The same final verification exposed two stale `local-self` assertions rather than product failures. Forced verified daemon reclamation still set its structured `forced` result and emitted an actionable warning, but the fixture required the old internal phrase `forcing process` after routine PID-oriented logging had been removed; the assertion now binds the stable semantics (`ignored graceful termination` plus forcing) instead of restoring identifier-heavy wording. Likewise, full uninstall correctly refused an active or unreadable managed-job state, but its test still matched the older narrower phrase `managed jobs are active`; that check now requires the current fail-closed uninstall meaning and inspection guidance. Removing the obsolete `job approve` parser table also briefly caused unknown job actions to be intercepted by a generic positional-count error. Unknown job actions now bypass the known-action positional table and reach the single command handler, preserving the explicit `Unknown job action` result without restoring any approval execution path.
|
|
562
|
+
|
|
563
|
+
The first frozen full gate then failed at `full-access:test` after 105 earlier tasks had passed. The failure was not a sandbox or execution defect: the diagnostic's first direct-process probe inherited the ordinary two-second foreground resource-admission window, while the full gate's own concurrent work temporarily produced Yellow `cpu_pressure_window`; the probe therefore returned the expected retryable resource-admission error after about two seconds. This was the same test-versus-cooperative-load mismatch already corrected in `local-self-test` and `agent-context-test`, but only the full-access managed-job observation window had previously been lengthened. `runFullAccessTest` now gives its internal `LocalRuntime` a five-minute cooperative admission budget while leaving command timeouts, detached-job step timeouts, and ordinary production foreground admission unchanged. A release architecture contract requires that override, and the focused real-machine full-access test passes with it.
|
|
564
|
+
|
|
565
|
+
The review then followed process ownership rather than treating ChildProcess `error` as equivalent to process death. Three local execution surfaces had the same causal defect in different forms: one-shot execution released tracker/resource ownership directly from an `error` event; detached managed-job steps cleared `activeChild` and released their lease from that event; and fixed internal `runExecutable()` commands settled before `close`. All three now latch the asynchronous error and wait for close/exit settlement before releasing ownership. Process-session startup had a related post-spawn gap: bind failure or cancellation deleted the session/untracked/released the lease after only a best-effort kill request. Failed starts now retain an internal hidden session, tracker entry, and resource lease until real close, and return a non-retryable ambiguous-side-effect record stating whether termination was requested. A direct forced `kill_process` or authority-revocation SIGKILL/tree request that returns false no longer reports success or deletes the process-session handle. Deterministic fault injection proves error-before-close ownership, failed bind/cancel termination retention, exactly-once lease release after eventual close, and revocation retry. Runtime authority revocation also attempts pending calls, process sessions, and managed jobs independently before aggregating any failure, minimizing revoked-authority lifetime without allowing acknowledgement of a partial result.
|
|
566
|
+
|
|
567
|
+
A separate retry review found that idempotent `start_job` replay could inspect an old `queued` status and relaunch a dead runner before reconciling a terminal `result.json` that had already committed in the result-first/status-second settlement window. That could repeat main effects under the same idempotency key after a status-write failure. Replay now performs full reconciliation before any queued-runner launch decision, and its prior-status presence check uses the bounded no-follow state reader so only real `ENOENT` is absence. Fault injection writes a queued status plus durable terminal result and requires replay to converge to that result without executing the main marker again. The detached runner fatal handler was tightened at the same boundary: it may write a fatal terminal record only after confirming this process owns the runner claim and after re-reading a valid same-job active status; unreadable, terminal, or ownership-mismatched state is retained rather than overwritten and scrubbed.
|
|
568
|
+
|
|
569
|
+
The destructive-state pass found that whole-state uninstall still validated one state-root pathname generation and then recursively removed whatever generation occupied that name. The documented same-user race attacker could therefore replace the root between validation and deletion. `state-root-retirement.mjs` now binds deletion to the exact lossless inspected generation: rename to a reserved sibling retirement name, recheck the moved device/inode, rerun the complete marker/allowed-content/workspace safety contract, and only then recursively delete. A crash-left valid retired root is recognizable and reclaimable; malformed reserved names, wrong types, encoded-generation mismatch, unreadable state, or a replacement generation are retained and block destructive cleanup. State/profile/job inventory was tightened in parallel so only `ENOENT` means absent and symbolic-link profile/job roots cannot be traversed as destructive control evidence. The dedicated regression covers normal removal, crash recovery, rename-window replacement, verification-failure quarantine retention and later safe recovery, malformed/wrong-type/mismatched residue, and symbolic-link profile refusal. The new critical module measures 100% function / 81.1% branch coverage against a 90/70 gate, and the complete critical-module coverage gate passes.
|
|
570
|
+
|
|
571
|
+
The follow-up reserved-namespace review found two additional fail-open deletion/capacity cases. First, `retired_job_*` cleanup inventory recognized only well-formed names, so malformed reserved crash residue could be silently skipped; second, the public `job_...` namespace was counted and inventoried only when the directory entry was actually a directory. A wrong-type job file/symlink could therefore escape the 50-item capacity, disappear from destructive inventory, and later be removed with the whole state root. Both namespaces are now reserved by name first and validated by type/grammar second: malformed/wrong-type entries remain `integrity_error` blockers, public wrong-type job state counts toward capacity, owner/local diagnostics expose only coarse unreadable state, and non-owner listing remains isolated. Deterministic target admission also securely inspects an existing idempotent job pathname before pruning, preventing a dangling/invalid deterministic target from evicting terminal history before the request fails. The same principle was applied one directory higher: `profiles/` destructive inventory no longer filters unknown/wrong-type children out of evidence; every child must be a real canonical 24-hex profile directory or uninstall is blocked for inspection.
|
|
572
|
+
|
|
573
|
+
The generation-race review also revisited the beta.60-compatible resource transaction mutex. Ordinary managed-job and state-root retirement now deliberately never roll a failed quarantine back onto the public pathname, because a check-then-rename restore can overwrite a new same-user generation. The resource lock is a constrained rolling-compatibility exception: beta.60 still requires the canonical `transaction.lock` directory to remain the mutex. The highest-risk incomplete-owner case no longer uses quarantine at all; after the bounded grace it rechecks the exact empty directory generation and missing `owner.json`, then uses `rmdir`, so a late beta.60 owner publication makes the directory non-empty and atomically defeats reclamation without ever moving the live mutex. Token-bearing quarantine restore remains only for release/stale-owner compatibility, now rechecks for a replacement before rename, refuses to overwrite one, and has deterministic replacement-preservation coverage. The remaining final check-to-rename window is documented as a temporary rolling-boundary residual rather than generalized as a safe deletion primitive.
|
|
574
|
+
|
|
575
|
+
A later critical-coverage run exposed a separate interruption mechanism that ordinary managed-job runs had only hit intermittently. The detached runner correctly gave background resource admission up to thirty minutes, but every coordinator transaction nested inside that wait still inherited a fixed five-second lock deadline. A transient transaction owner lasting slightly more than five seconds therefore bypassed the outer liveness contract and surfaced as `cleanup_error_class=resource_error`; recovery then permanently committed `recovery_failed` before its finally step had even produced a result. The test now reproduces that failure deterministically by holding the private coordinator transaction lock for 5.5 seconds during dead-runner recovery. Coordinator acquisition uses the caller's remaining admission budget capped at 30 seconds, while post-spawn lease bind/release use a 30-second ownership-settlement budget; the ordinary production two-second foreground admission ceiling therefore remains unchanged. The same investigation found a second bootstrap defect: a recovery runner marked its runner claim terminal-authoritative before recovery-lock handoff completed, so handoff failure could scrub a still-retryable plan. Terminal authority now begins only after the token-bound recovery lock has been released by the handed-off runner; a wrong-type lock fixture proves bootstrap failure retains `interrupted` state and plan with no terminal result. The deterministic lock-contention recovery passes both normally and with `NODE_V8_COVERAGE` enabled.
|
|
576
|
+
|
|
577
|
+
The next frozen full gate reached 113/116 before `install:test` exposed another real interruption boundary: the private hardened-npm bootstrap made one 60-second registry request and failed permanently on `ETIMEDOUT`, even though the requested artifacts are immutable exact URLs whose bytes are independently pinned by SHA-512. The downloader now makes at most three attempts with referenced per-attempt timeout timers and bounded 750/1500 ms backoff, but only for explicit transient network codes, HTTP 429, and HTTP 5xx. Redirects, proxy-configuration errors, certificate/policy failures, byte-limit violations, and other non-transient failures are never retried. This both removes another awaited-work `unref()` lifetime dependency and gives install/package flows bounded resilience without weakening origin/integrity checks. Deterministic tests drive ETIMEDOUT then HTTP 503 then successful bytes, and separately require certificate rejection and HTTP 302 to stop after one request. After a fixture-only timing correction separated the 30-second setup-acquisition budget from the measured 5.5-second resource-lock hold, the same source generation subsequently passed fast 87/87 and full 116/116 in 406.9 seconds, including canonical coverage, real install, stdio, Worker integration, and OAuth browser. A later status-enum review then found that the production `full-access-test` duplicated an incomplete managed-job terminal list and could wait five minutes on an already-terminal `recovered`, `cancelled_before_start`, or `expired_before_start` job. Full-access polling now consumes the shared terminal classifier, and active lifecycle classification is likewise shared by retention, manager reconciliation, and detached-runner fatal settlement. Because that is a source edit after the 116/116 run, that run remains strong evidence but is no longer the final frozen-source acceptance run; the complete gate must be repeated once more.
|
|
578
|
+
|
|
579
|
+
The continued independent review then challenged the **meaning** of that 60-second retry rather than merely its retry count. On the current network path a 256 KiB range from the exact npm 12.0.1 registry artifact completed at roughly 20 KiB/s, which is healthy forward progress but implies well over sixty seconds for the full multi-megabyte tarball. A fixed per-attempt wall deadline therefore restarted the same immutable transfer from byte zero forever under a sufficiently slow but non-stalled connection. Hardened downloads now separate liveness from duration: a referenced 60-second idle timer is reset only by response/body progress, while an independent five-minute per-attempt ceiling still bounds a peer that dribbles forever. Exact HTTPS, proxy validation, redirect rejection, byte ceilings, fixed SHA-512 verification, and the three-attempt transient classification are unchanged. The real `install:test` subsequently crossed the former sixty-second failure point, completed hardened npm construction, package packing, isolated global installation, and zero-argument startup in about seven minutes, and exited zero. The critical coverage gate initially rejected the new modules at 68.4%/38.2% and 100%/77.8%; instead of lowering the thresholds, fault injection was expanded across exact-HTTPS, default proxy-agent, malformed proxy/NO_PROXY, setup/stream failure, declared/streaming oversize, retry exhaustion, progress/idle/absolute timeout, and successful slow-body paths. The final measured coverage is 94.7% function / 78.1% branch for the downloader and 100% / 92.3% for the timeout helper.
|
|
580
|
+
|
|
581
|
+
The same fail-closed status review found a destructive launchd ambiguity and a cross-platform provider-state projection gap. `launchdStatusSummary` previously collapsed every nonzero `launchctl print` into `loaded=false`, and provider stop accepted `active=false` even if the service remained registered. A provider/permission failure could therefore look like absence, while a failed `bootout` followed by an independently exited process could authorize later plist deletion. The local platform probe confirms exit 113 is launchctl's explicit missing-service result. Current launchd handling accepts only exit 0 as loaded evidence and 113 as unloaded absence; every other nonzero result is `status_unavailable`, failed-query stdout is ignored, and stop/removal succeeds only after observing the service both inactive and unloaded. The same review found that systemd `is-active` transition/unknown output and launchd query failure were still projected as boolean `active=false`, which let owned-service mutation and `service status` erase the difference between inactivity and unavailable evidence. Provider activity is now tri-state: recognized active/inactive/failed/confirmed-missing states are boolean, while transition, installed-but-unknown, and unrecognized query state are `null`; launchd/systemd direct start/restart refuse unavailable state, and `effective_active` remains `null` unless a workspace daemon is positively alive. Deterministic tests cover ambiguous pre-stop status, explicit absence, normal unload, still-loaded/inactive state, post-bootout status loss, systemd transition/unknown state, and CLI tri-state projection. These source edits likewise require a fresh final frozen-source gate.
|
|
582
|
+
|
|
583
|
+
One live-runtime observation during this review reinforces the source/deployment distinction. The deployed daemon currently reports exactly 50 retained managed jobs, all listed records in the observed window are terminal, and an attempted new detached coverage job returned a generic failure without creating a new job. That is consistent with the already-identified exact-50 admission defect in the older live beta.61 bytes; the current working tree contains the protected-capacity/terminal-eviction repair, but those source edits have not been activated and are not claimed as live behavior.
|
|
584
|
+
|
|
585
|
+
A further independent verification pass found a source-tooling liveness defect rather than a TypeScript defect. The fast gate first stopped on two ordinary lint errors in the Wrangler lifecycle regression; after those were corrected, `typecheck` repeatedly reached Wrangler's final `Types written` and rerun-reminder output but never advanced to `tsc`. The same behavior reproduced with direct CLI execution, stdin disconnected from the terminal, and both retained 4.115.0 and current 4.120.0 toolchains under Node 26.7.0, so the evidence does not isolate the behavior to the beta.61 Wrangler upgrade. The cross-platform Worker-type generator now removes its target before launch, has a hard pre-completion timeout, and recognizes successful generation only after Wrangler's final command-completion notice and a newly created declaration file. It then allows a bounded natural-exit grace and may send graceful termination to a completed-but-resident CLI; inability to terminate gracefully is still a verification failure. A synthetic lifecycle regression covers clean exit, post-completion residency, nonzero exit, and pre-completion stall, while the real full typecheck chain now completes normally. The exact upstream Node/Wrangler lifecycle cause remains unknown.
|
|
586
|
+
|
|
587
|
+
Final release-readiness review then reproduced the same provider-CLI lifecycle defect in the separate release-only `worker:dry-run` check. The real Wrangler 4.120.0 process printed its complete upload/binding summary and the terminal `--dry-run: exiting now.` notice, but remained resident for more than ninety seconds; this was OS process evidence, not the older live daemon's known process-session close-settlement defect. The raw package script therefore could never finish one of the documented pre-candidate release gates on this host. Worker type generation and deployment dry-run now use one `wrangler-command-lifecycle.mjs` state machine with command-specific completion markers, a hard pre-completion deadline, bounded natural-exit grace, SIGTERM cleanup, and a bounded SIGKILL failure path. Generated types additionally require the newly written declaration target before a completion hang can be classified as successful. Synthetic fixtures drive normal exit, completed-but-resident, nonzero failure, and pre-completion stall for both commands; the exact real dry-run that previously remained resident now exits zero after bounded cleanup, and no matching Wrangler dry-run process remains. This repair changes only source verification lifecycle; it does not authorize or perform a Worker deployment.
|
|
588
|
+
|
|
589
|
+
The subsequent frozen full gate exposed a deeper resource-mutex crash-recovery defect through the managed-job coverage fixture. The first failure arrived as a bare `MBM_RESOURCE_TRANSACTION_BUSY`; the fixture itself also had an unhandled-rejection escape hatch because it awaited only a separate readiness Promise while its hold-lock task could reject first. After that control flow was corrected and the setup failure was labeled, repeated full coverage proved the private coordinator really remained busy for the full 30-second acquisition budget. Privacy-bounded diagnostics showed no readable `owner.json`. The remaining state transition was then mechanically identifiable: `tryAcquireDirectoryLock()` creates the final `transaction.lock/` directory before `replaceFileAtomicallySync()` publishes `owner.json`; a process killed after the atomic writer creates `.owner.json.<pid>.<nonce>.tmp` but before its final rename leaves an ownerless yet non-empty directory. The beta.60 compatibility path deliberately reclaims ownerless directories only with empty-directory `rmdir`, so that crash residue could never be removed and permanently pinned the transaction mutex.
|
|
590
|
+
|
|
591
|
+
The recovery boundary now recognizes only that exact atomic-owner staging grammar after the existing incomplete-owner grace. It inspects the encoded publisher PID against the staging file's write-time generation bound, preserves a still-current publisher, requires a single-link regular file, rechecks the staging filesystem identity, and only then unlinks a dead publisher's uncommitted staging file. Unknown names, multiple entries, wrong types, and hard links remain integrity failures rather than generic cleanup candidates. The caller still performs canonical-directory `rmdir`, so a beta.60 publisher that legitimately completes `owner.json` in the final race window makes the directory non-empty and atomically defeats reclamation exactly as before; no quarantine/split-brain shortcut was introduced. Deterministic resource-admission tests cover dead staging recovery, live-publisher preservation, and unknown-content refusal. `resource-staging-recovery.mjs` retains 100% function coverage and 61.4% branch coverage against its 100/55 gate, and the complete critical coverage suite passes after the production repair. The managed-job fixture keeps the immediate rejection observer too, so future setup-lock failures surface through its own test boundary instead of terminating Node asynchronously.
|
|
592
|
+
|
|
593
|
+
The first owner-terminal activation of that candidate then exposed a separate end-to-end release blocker that unit tests had masked. The live daemon came up cleanly as `3.0.0-beta.61`: relay readiness was verified, process/drain/termination ownership was empty, the resource coordinator had no active leases or waiters, and the security-audit chain was healthy. The deployed OAuth canary completed authorization-code exchange, authenticated MCP, refresh rotation, and refreshed MCP against the activated Worker, but exited with `release OAuth canary temporary state cleanup was incomplete`. A read-only live inventory immediately afterwards found zero synthetic `release-canary-*` reviewer accounts and zero `Machine Bridge release OAuth canary` clients, proving the remote cleanup had in fact settled. Source comparison then found the exact contract mismatch: the Worker `/admin/clients` DELETE path deletes the client, codes, access tokens, and refresh tokens and returns JSON `200`, while local `AccountAdminClient.request()` previously accepted only `204` for every DELETE. The local fixture had reproduced that assumption by returning `204` for both account and client deletion, so it could not detect the cross-layer divergence. The local client now enforces endpoint-specific success statuses: OAuth-client deletion requires `200`, account deletion requires `204`, account creation requires `201`, and all other admin operations require `200`; reverse-status negative tests ensure the two DELETE contracts cannot collapse back together. Because this changes packaged bytes after activation, beta.61 is explicitly rejected from acceptance/publication. The repaired source is assigned to `3.0.0-beta.62` and requires a fresh frozen gate, exact candidate, owner activation, deployed OAuth canary, and observed live verification before acceptance.
|
|
594
|
+
|
|
595
|
+
The same pass found a client-visible error-semantic defect in the current MCP response proxy. If the internal Durable Object SSE reader failed, `pumpUpstream()` called the generic cancellation path first; that path closed the public stream, so the subsequent attempt to error the public controller was necessarily too late. The existing test explicitly tolerated either behavior and therefore allowed a transport failure to become normal EOF. A red regression made the next public read require the synthetic upstream error, reproduced the clean-EOF bug, and then drove the fix: upstream failure still aborts the internal reader and schedules the credential-free cancel control but does not close the public controller before `target.error()`. Ordinary upstream completion, public reader cancellation, request-signal cancellation, random internal stream capability replacement, and credential stripping remain covered separately.
|
|
596
|
+
|
|
597
|
+
Managed-job launch ownership was challenged separately from normal runner execution. `launchRunner()` necessarily spawns before the child pid is known, while the parent then publishes `runner.pid`. The old publication path could create a readable matching claim and still throw during the final owner-only permission operation; its caller would try a best-effort `SIGKILL`, classify the launch as failed, and scrub launch artifacts, but a failed kill could leave the child able to confirm the claim it had already received and execute. Claim publication is now explicitly two-phase: the parent first writes a token-bound `committed:false` claim, hardens it, and only then atomically replaces it with `committed:true`. The child accepts only the committed form and merely waits on a matching uncommitted claim until its startup deadline. The full managed-job/resources integration test passes with this contract, including a valid pid/token uncommitted fault fixture that must remain unchanged and non-executable.
|
|
598
|
+
|
|
599
|
+
The same ownership principle exposed a cross-platform authority-revocation gap. On Windows, `terminateProcessTree()` returns after `taskkill.exe` is spawned while its eventual exit and fallback `ChildProcess.kill()` are asynchronous. A synchronous `true` therefore meant only "request accepted", yet process-session authority revocation previously deleted its session immediately and let the relay acknowledge the durable Worker revocation. Revocation is now asynchronous: all execution categories are attempted, the session termination is requested without serially blocking managed-job cancellation, and the durable revocation is acknowledged only after every targeted process session reports actual `close`; an undelivered or un-settled termination retains the handle and returns retryable `unavailable`. Fault injection now covers both outright kill-request failure and the subtler case where the request reports success but the process never closes.
|
|
600
|
+
|
|
601
|
+
Runtime teardown was aligned with that same standard rather than leaving a second escape hatch. The shared process tracker now has a terminal drain mode: existing children receive forced termination, children tracked after draining begins are immediately terminated, and shutdown waits for the tracker to become empty. Process-session teardown uses `clearAndWait()` and the previous fire-and-forget `clear()` API has been removed. A failed drain/clear moves the lifecycle from `stopping` to a dedicated non-operational `stop_failed` state, which rejects restart but permits another cleanup attempt; it is never reported as `stopped`. Relay supersession/fatal callbacks, failed activation cleanup, stdio shutdown, and CLI signal handling all await the same full runtime teardown before lock release or process exit. The runtime-infrastructure and process-output continuation regressions pass with the close-settled contract, including late-child drain and delivered-but-never-settled termination cases.
|
|
602
|
+
|
|
603
|
+
A later retry-path review found that `stop_failed` was not fully retryable even though its lifecycle API said it was. `ProcessTracker.requestDrainTermination()` marks a child as already requested before sending the forced termination, and the first drain implementation retained that marker when its settlement deadline expired. A second `runtime.stop()` could therefore retain the same child but skip the only termination request that might let it finish. Failed drains now clear the request marker only for children still active at the end of that attempt, preserving within-attempt de-duplication while allowing the next cleanup attempt to reissue termination. The regression forces a first drain to time out, requires the second drain to send a second request, then proves it can settle without losing ownership.
|
|
604
|
+
|
|
605
|
+
The same review compared the architecture's documented `exit` fallback against actual adapters and found that only the managed-job runner was using `child-process-settlement.mjs`; one-shot execution and process sessions still released tracker/session/resource ownership only from `close`. That made the documentation broader than the implementation and left both surfaces vulnerable to an indefinitely missing libuv `close` after an already-observed process `exit`. `attachChildProcessSettlement()` now owns one shared adapter policy for all three surfaces: prefer real `close`, retain the observed exit state, and after the existing one-second grace destroy only residual stdio, detach the already-exited child, and settle exactly once. Deterministic one-shot and process-session fixtures emit `exit` but never `close` and require successful result/session settlement plus tracker/resource cleanup; the managed-job suite passes after its duplicate listener/stdio-fallback wiring was removed in favor of the same helper.
|
|
606
|
+
|
|
607
|
+
One narrower shutdown race remained after that first repair and was found by challenging the ordering rather than the individual termination primitives. `CallRegistry.cancelAll()` aborted each call and immediately removed it from the registry, while the actual `ToolExecutor` handler could still be unwinding toward its lifecycle `finally`. That left a window where `ProcessTracker.drain()` observed no children and returned, after which the still-running cancelled handler could register a new child; the tracker's persistent drain flag would request termination, but runtime teardown was no longer waiting for that late child's close. The call registry now has its own terminal drain controller. Once shutdown starts it rejects new `open()` calls, aborts existing calls without synthetically finishing them, and waits until the ordinary handler lifecycle calls `finish()` for every retained record. Timeout is monotonic and fails retryably while preserving the stuck call. Runtime ordering is now call drain -> process drain -> process-session drain, so no handler remains capable of spawning after the final process-settlement barrier begins. Architecture tests pin that ordering, and critical coverage now explicitly includes both call-drain and process-session termination helpers rather than allowing those extracted safety modules to fall outside threshold accounting.
|
|
608
|
+
|
|
609
|
+
The first frozen full gate also found a packaging-only hygiene defect at task 112/117: six newly added `src/local/*.mjs` files were present with owner-only `0600` modes, which the npm package manifest test rejects for ordinary source. Two were created during this review and four were already present among the pre-existing uncommitted modules. All six were normalized to `0644` without content changes or executable bits, and the package regression reports the expected modes across its 417-file packed manifest. This is deliberately recorded as a release-contract defect even though it has no runtime semantics; it would otherwise survive lint, typecheck, architecture, privacy, and coverage gates until packaging.
|
|
610
|
+
|
|
611
|
+
Logging/privacy was rechecked after those fixes rather than inferred from the earlier privacy gate. Worker observability and the throttled outer-edge logger both bound field counts/names, redact credential-shaped keys, and apply the shared portable value redactor to surviving strings; current call sites use coarse route/event/error classes rather than request values. The stateful rate-limiter's binding-failure fail-open behavior was also challenged but is not an accidental security bypass introduced by this pass: the beta.18 audit and regression suite already record it as an explicit availability trade-off. No new raw credential/token/path logging path was established by this review. Best-effort local cleanup diagnostics that intentionally cannot replace the primary failure now carry an explicit catch rationale rather than unexplained empty catches.
|
|
612
|
+
|
|
613
|
+
This record does not accept the current source or candidate. The live repaired candidate is strong operational evidence, but the independent review continues across architecture, module boundaries, state cardinality, error/log projection, tests, documentation, privacy, and release behavior. Any packaged edit after the successful live candidate invalidates that candidate for final acceptance and requires a new exact candidate plus owner activation and repeated deployed OAuth/MCP verification. Formal acceptance remains withheld until the final source is frozen, all gates pass, and the final deployed candidate demonstrates authorization-code exchange, persisted access/refresh credentials, authenticated MCP operation, refresh rotation, and the applicable continuity/soak evidence.
|
|
614
|
+
|
|
615
|
+
## 2026-08-10 beta.61 stateful-session removal and conservative disk cleanup follow-up (intermediate protocol state)
|
|
616
|
+
|
|
617
|
+
At this intermediate point in beta.61, the review made MCP `2026-07-28` the only executable protocol while deleting the **stateful** `2025-11-25` initialize/session adapter, signed MCP sessions, `Last-Event-ID` replay, persisted MCP delivery/call rows, prepare/subscribe descriptors, era-specific controllers, and their DPoP retry binding rather than hiding them behind aliases. The browser-visible CORS contract also stopped advertising `Mcp-Session-Id` and `Last-Event-ID`; current request ownership remained transient and request-scoped, while same-daemon reconnect could rebind an already-dispatched call beneath the still-owned request without creating client replay semantics. The later 2026-08-11 hosted-client review supersedes only the “sole executable wire date” part of this paragraph by adding bounded **stateless remote-HTTP initialization compatibility** for `2025-06-18` and `2025-11-25`. It does not restore any of the deleted session/replay/delivery state described here.
|
|
618
|
+
|
|
619
|
+
Cancellation evidence was tightened rather than inferred from the local emulator. Wrangler/workerd does not reliably convert a raw local TCP close into a Worker `Request.signal` abort callback, so the integration gate no longer treats that observation as proof. A deterministic response-proxy regression now drives both response-body `cancel()` and request-signal abort, requires a fresh random internal stream capability instead of trusting caller-supplied control headers, binds exactly one credential-free cancel control to the direct request, and verifies that Authorization/DPoP are never copied into the private cancellation path. Real deployed-edge stream-close cancellation remains a candidate-activation observation.
|
|
620
|
+
|
|
621
|
+
State migration remains deliberately narrower than runtime compatibility. A read-only registry check at this source cutoff found public npm channels `latest=2.0.0` and `beta=3.0.0-beta.38`; the owner machine separately has the explicitly recorded live beta.60 candidate. Source comparison proves that 2.0.0 browser pairing stores a single `{token, port}` shape, beta.38 stores schema-2 dual tokens without the current pairing-auth version, and 2.0.0 OAuth refresh state uses schema 2. Those one-way readers therefore remain necessary to avoid destructive re-pairing or reauthorization for real published users. The beta.60 resource-transaction lock bridge likewise remains only for the exact live handoff. Successful migration writes current state and does not make the producing runtime or protocol executable again.
|
|
622
|
+
|
|
623
|
+
Disk cleanup followed the same evidence rule. The host recovered from the earlier roughly 67 GiB free-space low to above 100 GiB without broad deletion. The repository's `.project-local`, `.wrangler`, and dependency tree were small or actively useful; the remaining roughly 2.37 GiB shared `AgentBuilds.noindex` Swift cache did not contain this repository's canonical project partition and was therefore left untouched. Under the system temporary root, only `mbm-candidate-runtime-*` and `mbm-candidate-outside-*` directories older than twelve hours were deleted after source inspection proved they are test-only `mkdtemp` fixtures, no candidate-runtime test process was active, and successful current tests already clean them in `finally`. An active coverage process was detected separately, so every `machine-bridge-coverage-*` directory was preserved. The reclaimed candidate-test residue was only tens of KiB; no larger unknown or cross-project cache was deleted merely to increase the cleanup number.
|
|
624
|
+
|
|
625
|
+
## 2026-08-09 beta.61 scheduler frontier and CognactApp follow-up
|
|
626
|
+
|
|
627
|
+
A second resource-scheduler review compared the beta.60 design against current Bazel `ResourceManager`, Ninja jobserver integration, Buck2 daemon/isolation/local-resource behavior, Linux PSI/cgroup documentation, DRF/Tetris-style multi-resource fairness, production backfill scheduling, and the local CognactApp process runtime. The review deliberately did not replace the coordinator with one of those systems. Bazel and Ninja support the existing work-conserving/shared-token direction; Buck2 reinforces that project state/cache identity must be scoped to the real project rather than a common parent; DRF is a multi-tenant fair-share objective rather than a direct fit for prioritized one-shot local roots; and Linux PSI provides a better contention signal than inventing another load-average threshold.
|
|
628
|
+
|
|
629
|
+
The review found two concrete scheduling defects. First, a safe shell `cd` could still fall back to the daemon workspace when the inner command family was unknown, so sibling repositories under one parent could receive the same contention key. Literal, provable directory changes now refine the project scope even for a generic serialized root; an uncertain directory transition makes subsequent relative transitions uncertain until a new absolute/home path re-anchors the parser. Second, shell-heavy detection looked at arbitrary command text, allowing filenames containing `test`, `build`, or `release` to turn metadata reads into heavy work. Script-name classification is now isolated from shell tokenization and considers only the actually executed shell/Node/Python script operand or package-manager script name. This also closes the opposite blind spot found during the final resource review: direct orchestration roots such as `node scripts/coverage-check.mjs`, `python -u verify-*.py`, and `npm run check` no longer begin with a 0.5-CPU adaptive lease before spawning expensive descendants. Negative tests keep module/eval modes and harmless scripts with heavy-looking later arguments unpromoted.
|
|
630
|
+
|
|
631
|
+
Fit-aware queue scanning remains the default because it avoids head-of-line underutilization. Aging alone, however, could not prevent a large feasible waiter from being bypassed indefinitely by a stream of smaller arrivals. Beta.61 therefore adds a conservative protected-drain phase: a rank-zero waiter must reach its protection age, be blocked by an existing coordinator lease for project/CPU/I/O/memory reservation, and remain structurally feasible after those leases disappear. Only then does it stop fresh backfill; impossible requests still cannot drain the queue. Reservation blockers are classified before the instantaneous CPU window so the scheduler can distinguish coordinator-owned capacity from unrelated host pressure without changing the admission result.
|
|
632
|
+
|
|
633
|
+
Platform verification exposed a separate event-loop liveness defect. The default admission sleep used `setTimeout(...).unref()`: when a full-access path had no other live handles, Node exited with code 13 while the top-level `await coordinator.acquire(...)` was still pending. The same helper also accumulated one abort listener per retry because successful timer completion did not remove it. `resource-wait.mjs` now owns a referenced timer and symmetric listener cleanup. A fast regression performs a real bounded wait instead of injecting a fake sleep, and the previously failing full-access machine sandbox succeeds under the repaired implementation.
|
|
634
|
+
|
|
635
|
+
A subsequent canonical-coverage run exposed a different pressure interaction: `local-self-test` could hit the production 2-second foreground admission window while an unrelated cooperating workload held I/O/CPU reservations, causing `resource_admission` even though the self-test was healthy. The self-test now constructs its four runtimes with a five-minute internal resource-wait override, and it asserts that the same value reaches both `ProcessExecutionService` and `ProcessSessionManager`. This does not weaken or bypass machine-user coordination, and it does not alter the public foreground contract: ordinary runtime calls still default to two seconds, while the long test waits for capacity and starts its command timeout only after admission.
|
|
636
|
+
|
|
637
|
+
The review also found that the critical V8 coverage gate did not explicitly include the new resource scheduler. Functional tests were present in the fast plan, but `coverage-check.mjs` neither executed the direct resource fixtures nor failed if scheduler modules lost coverage. Beta.61 adds both resource fixtures to the canonical coverage run and module-specific function/block thresholds. Deterministic tests now cover Linux `/proc` reader absence, PSI parsing, request-contract failures, protected-queue diagnostics, referenced wait timeout, in-flight abort, and already-aborted signals. Cross-platform host orchestration uses a function-only threshold because one local platform cannot honestly claim branch coverage for all platform samplers. A later canonical coverage run exposed the same test-vs-production admission mismatch in `agent-context-test`: its registered-command runtime could fail after the production two-second window while an unrelated cooperating workload drove host pressure Red. Only that resource-executing test runtime now opts into a five-minute wait, and the release architecture contract requires the override so the coverage fixture continues to wait for shared capacity rather than bypassing coordination or becoming load-sensitive. Production foreground admission remains unchanged. Architecture release contracts also require the direct resource fixtures and core scheduler threshold keys.
|
|
638
|
+
|
|
639
|
+
A full workflow-bundle development gate then exposed a distinct nested-admission deadlock that focused tests had not exercised. The gate process held a bounded `script-heavy` lease while pytest, running inside that process tree, attempted to acquire another bounded lease before starting a heavy child. Naive summation made the parent reservation consume the capacity needed by its own nested command; under Yellow pressure the nested waiter could age until its five-minute admission budget expired. Beta.61 keeps every child lease complete and independently crash-recoverable, but computes live reservations from an ephemeral process-ancestry forest: a parent lease is an orchestration envelope and the effective vector is the component-wise maximum of that envelope and the sum of direct lease children. Pending nested requests are charged only for the delta they add. Same-key contention is ignored only along the requester's actual ancestor lease chain, so siblings still serialize. If ancestry sampling fails, accounting reverts to full summation. A follow-up performance review found that retrying waiters would otherwise execute a whole-process-table `ps` or PowerShell/CIM query roughly every half second, so parent snapshots now have a one-second completion-time cache, including failed/timeout samples; stale or unavailable ancestry can only make accounting more conservative. The previously failing full workflow gate subsequently crossed the old 52-53% failure point with no nested waiter and completed 682 tests successfully.
|
|
640
|
+
|
|
641
|
+
A later live-host diagnosis reproduced the user-visible responsiveness failure more directly. On an eight-core host, load1 reached about 109 while the coordinator reported only roughly two reserved CPU cores. The process table showed a `swift-test` root already reparented to launchd plus about seven concurrent `swift-frontend` descendants, along with a separate Python test process; after those roots ended, load fell to roughly 11 and the resource queue drained. Inspection of the persisted leases showed no lease for the surviving Swift process group. The stale-pruning logic was already process-group aware, but explicit `releaseLease()` deleted a bound lease unconditionally. Beta.61 now interprets caller release as relinquishing the handle, not proof that an isolated POSIX group is gone: a still-live group keeps its persisted reservation until ordinary pruning observes group death. Regression tests release a real detached group while it is running and require the lease to survive, then require pruning after termination.
|
|
642
|
+
|
|
643
|
+
The same diagnosis identified a separate control-plane stall mechanism. The live admission path synchronously executed `ps`, `memory_pressure`, `vm_stat`, one-second `iostat`, `pmset`, whole-table ancestry sampling, and child start-time lookup on the daemon's Node event loop. Under load, the mechanism intended to measure pressure could itself delay relay heartbeats, browser bridging, and MCP responses. A local A/B timer probe around the full Darwin sample measured about 1.15 seconds elapsed with approximately 1.1 seconds of maximum 10 ms timer lag on the old synchronous path; the asynchronous path took about the same total observation time but reduced maximum timer lag to about 3 ms. Host/process probes are now async on the live path, independent Darwin probes run concurrently, concurrent ancestry requests share one in-flight sample, and child start identity is collected before entering the coordinator lock. Successful I/O evidence is reused for five seconds while cheaper signals refresh, and all freshness timestamps are completion-time timestamps.
|
|
644
|
+
|
|
645
|
+
The live disk snapshot also fell from roughly 125 GiB free earlier in the review to about 67 GiB. The shared `~/Library/Caches/AgentBuilds.noindex` root measured about 21.6 GiB, including several multi-gigabyte Cargo partitions. The cache is therefore a material contributor but not the whole disk change. Automatic deletion is intentionally not added without a cross-language maintenance lock, active-lease exclusion, and crash-safe cleanup semantics. Instead beta.61 tightens admission non-destructively: free space below `max(80 GiB, 15%)` becomes Yellow, while the existing post-reservation hard floor stays `max(50 GiB, 10%)`; build partitions now derive from the same canonical anonymous project hash as resource identity so aliases cannot create duplicate caches.
|
|
646
|
+
|
|
647
|
+
Cross-language probing during that repair found a second protocol defect: the Node and Python contention-key formulas were textually identical but canonicalized project paths differently. Node used lexical `path.resolve`, while Python resolved existing filesystem aliases; on macOS this made `/var/...` and `/private/var/...` hash to different project mutexes. The shared v1 key now canonicalizes the existing filesystem path before hashing on both sides; Windows additionally normalizes separator form, case, and extended `\\?\\`/UNC prefixes. Fixed golden vectors bind the exact v1 hash material in both repositories, while real symlink-alias tests exercise filesystem canonicalization. The privacy-preserving lease `project_hash` now derives from the same canonical identity, so diagnostics correlate aliases consistently without making that hash admission authority. A real interoperability probe also wrote a live Node lease and required Python to reject the same canonical project/family as `project_resource_busy`. Persistent lease schema version 1 is unchanged.
|
|
648
|
+
|
|
649
|
+
Linux sampling now reads `MemAvailable/MemTotal` plus `/proc/pressure/{cpu,memory,io}` avg10 values. PSI is optional; missing files remain missing evidence. A 10% recent stall is used only to tighten to Yellow, while memory `full` avg10 at 60% is treated as Red, matching the order of magnitude used by systemd-oomd's memory-pressure action boundary rather than pretending that the coordinator enforces memory. macOS retains its memory-pressure/pageout/swapout/disk/thermal probes.
|
|
650
|
+
|
|
651
|
+
CognactApp supplied an important negative design result. Its `RealProcessRunner` owns `posix_spawn`, `wait4`, process-group quiescence, and exact `rusage`, so it can retain settlement-time wall/user/system CPU, peak RSS, I/O blocks, and descendant evidence before releasing a process lease. That is a credible substrate for later profile calibration. Machine Bridge's Node `ChildProcess` boundary has no equivalent per-child `rusage`; beta.61 therefore does not add a wall-time EWMA or polling-derived pseudo-learning model. A future adaptive-cost implementation should first introduce a small trusted spawn/wait supervisor that emits bounded settlement usage, then calibrate profiles with explicit caps/decay and regression evidence.
|
|
652
|
+
|
|
653
|
+
Beta.60 has no local acceptance record, but it was activated from a local candidate at `2026-08-08T23:23:27.866Z`; the live Worker/daemon therefore remains beta.60 while beta.61 source work proceeds. Its pending candidate directory was subsequently removed before the beta.61 version bump. Beta.61 must regenerate all release evidence and must not inherit acceptance from the activated-but-unaccepted beta.60 runtime.
|
|
654
|
+
|
|
655
|
+
## 2026-08-08 beta.60 cooperative resource-admission review
|
|
656
|
+
|
|
657
|
+
A multi-agent local execution review found that existing per-runtime tool-call and process-session limits could not coordinate independent Machine Bridge daemons, detached managed jobs, or workflow-bundle processes. Static low concurrency would protect the host but waste idle capacity, while load-average-only gating reacts after fan-out has already started. Beta.60 therefore introduces a per-user durable, work-conserving coordinator: multi-resource reservations are combined with current host pressure and a five-second startup reservation window; priority aging prevents permanent background starvation; fit-aware waiter selection avoids head-of-line blocking; and same-project build-cache families serialize without imposing machine-wide serial execution.
|
|
658
|
+
|
|
659
|
+
The review also found two transaction defects while integrating the coordinator. First, Node and Python process helpers could acquire a lease and then fail while preparing an external build root, leaving phantom capacity until stale-lease recovery. Both implementations now release the lease on every post-admission preparation failure and have deterministic fault-injection coverage. Second, the Node build-root helper could ignore an explicit effective `AGENT_BUILD_ROOT` supplied to a managed step because it consulted only daemon process environment. The process helper now propagates only that explicit control override into build-root selection rather than reusing the isolated child `XDG_CACHE_HOME`, which would fragment the intended shared cache.
|
|
660
|
+
|
|
661
|
+
The persistent protocol deliberately excludes argv, shell text, and raw project paths. Lease/waiter ownership is PID/start-time bound and process-group aware; malformed state fails closed. `diagnose_runtime` exposes only coarse pressure/reservation/queue summaries. Internal fixed diagnostics bypass heavy admission, and interactive calls use short retryable waits so resource pressure cannot consume every ordinary tool-call slot. This remains cooperative availability control, not a security sandbox or cgroup-style CPU/RSS quota; an uncooperative same-user process remains outside the guarantee.
|
|
662
|
+
|
|
663
|
+
Coverage fault injection then exposed a staging publication crash window. `createExclusiveFileSync()` can be killed after linking the final lease but before unlinking its internal staging alias; `replaceFileAtomicallySync()` can likewise be killed with an uncommitted single-link staging file beside an older final lease. The coordinator's strict directory scanner originally treated either as permanent corruption, so a later recovery runner could fail closed instead of cleaning up. Recovery now recognizes only the exact internal staging filename grammar, verifies target/staging inode and link-count relationships twice, and removes either a committed two-link alias or a dead-publisher single-link uncommitted staging generation. It still rejects mismatched identities, extra links, live uncommitted publishers, symlinks, directories, and arbitrary dotfiles. Python interoperability implements the same recovery states so either controller can recover a Node crash artifact.
|
|
664
|
+
|
|
665
|
+
The real-world workflow suite then exposed a scheduling-policy defect rather than a correctness failure: Python classified any gate-like shell script as `unbounded`, and admission translated `unbounded` to the entire current CPU budget. The startup-window check therefore required the host to be almost idle before such a script could begin, producing a background waiter even though no lease was active. The profile is now a bounded mixed process-tree reservation shared with the Node implementation. This preserves same-project contention protection without turning an unknown script body into a near-idle-host requirement.
|
|
666
|
+
|
|
667
|
+
## 2026-08-08 beta.59 live-candidate browser pairing correction
|
|
668
|
+
|
|
669
|
+
The owner activated the exact beta.58 candidate, but independent live verification deliberately withheld acceptance. Worker and launchd daemon identity/readiness were healthy at beta.58; browser verification was not. `machine-mcp browser status` reported the broker running with no extension connection, and source tracing showed a release-blocking split between the two explicit pairing entrypoints: MCP `pair_browser_extension` used the new ephemeral fragment launcher, while CLI `browser setup` still invoked `openExternal()` on the sanitized long-lived broker `/pair` URL. Because that HTTP page is intentionally token/grant-free, the packaged content script receives no bootstrap and returns before registering pairing material. The documented first-pair CLI flow therefore could not work. beta.58 remains activated only as the current operational baseline; it has no acceptance record and is superseded by beta.59 before any push/publication.
|
|
670
|
+
|
|
671
|
+
The correction removes grant construction from both callers and makes `browser-pairing-launch.mjs` the single composition boundary: broker port + current extension credential become a 30-second grant only inside the one-shot listener helper. CLI pairing reads only current schema-2/auth-v2 state through the existing bounded/no-follow/single-link parser, opens the helper's random 127.0.0.1 listener, and still exposes only the sanitized broker URL in terminal/JSON output. Failure to spawn the browser closes the launch immediately. A new fast-plan regression exercises the CLI command with real temporary state and a real ephemeral listener, while architecture source guards prevent a return to direct fixed-URL opening.
|
|
672
|
+
|
|
673
|
+
## 2026-08-08 beta.58 independent follow-up: ownership, runner launch, and browser bootstrap
|
|
674
|
+
|
|
675
|
+
### Additional beta.58 independent review findings
|
|
676
|
+
|
|
677
|
+
- **Pairing bootstrap redirect inheritance:** fragment secrecy alone was insufficient when the long-lived broker port itself might be occupied by another local process: an HTTP redirect can inherit the original fragment before the content script runs. Explicit pairing now binds a separate OS-random loopback listener first, opens only that already-owned listener, serves one token-free page, and closes after the first delivery or 30 seconds. The fragment carries the target broker port and grant; the extension strips it at `document_start` before page scripts and authenticates the real broker separately.
|
|
678
|
+
- **Authentication-registry churn:** both already-paired HMAC auth and first-pair bootstrap originally allocated bounded pending state after seeing only a syntactically valid challenge/grant id. A local process without the secret could therefore fill the registry as a denial of service. Both flows now require a secret-derived init HMAC before allocation. Normal auth is keyed by client challenge and repeated identical requests return the same server nonce/proof; pairing auth is keyed by grant id and is likewise idempotent for the same challenge.
|
|
679
|
+
- **Rollback-compatible browser state:** a provisional schema-3 design would have made beta.55 rollback reject browser state. The released design intentionally preserves `schemaVersion: 2`, adds `pairingAuthVersion: 2`, rotates the historically page-exposed extension key, retains the runtime key for migration identity, and requires boolean `migrationPending`. A beta.55-compatible parser fixture prevents future envelope drift.
|
|
680
|
+
- **Authorized admin-body amplification:** device-signed `/admin/*` requests previously used `request.clone().arrayBuffer()` and only then checked 64 KiB. Valid signed chunked bodies could therefore force unbounded materialization. The path now bounded-reads the original stream once, hashes exact bounded bytes, and rebuilds the request from that buffer.
|
|
681
|
+
- **Internal MCP response materialization:** stream-descriptor and terminal HTTP fallback parsing no longer uses bare `response.json()`; both reads are bounded at the consumer using the descriptor or resumable-message contract.
|
|
682
|
+
- **Managed-job normalization:** explicit invalid approval fields are rejected instead of coerced to defaults. Over-limit or malformed persisted resource registries fail closed rather than silently truncating or becoming empty.
|
|
683
|
+
- **Cloudflare Computer assessment:** `cloudflare/computer@8758b51c8891c211dddd1903d2ee2d12a75ac7ff` was reviewed as a potential remote execution layer. Its Durable Object VFS, real rows/statement benchmark, default-closed worker egress, and explicit host capabilities are useful patterns; the package remains preview-only and its container transport still has authentication/frame-limit/process-user open questions. It is therefore not added to the production dependency or owner-machine authority path. A future integration should be default-off, non-authoritative cloud scratch using a separate DO namespace and explicit rows budgets, with worker-shell/worker-JS preferred before container transport.
|
|
684
|
+
|
|
685
|
+
Beta.57 remained unaccepted and unactivated while the filesystem transaction review continued. Two additional ownership gaps were found before a candidate was frozen.
|
|
686
|
+
|
|
687
|
+
- **SSH staging cleanup was not fully descriptor-bound.** The beta.57 installer verified the source/target hard link after publication, but the source unlink and outer `finally` cleanup still addressed the random staging pathname. A replacement appearing after verification could therefore be deleted by cleanup, and a source-path replacement between link creation and post-link validation could leave the just-created target behind. Beta.58 opens and verifies the staging file before publication, keeps that descriptor alive across `link`, refreshes the descriptor identity after the legitimate link-count/ctime change, and accepts the target only when it matches that pinned object. If the source path changes while the target still matches the descriptor, the target is rolled back by its exact generation; source/final staging cleanup likewise deletes only the known object and preserves replacements. A deterministic fault-injection replaces the staging source after the hard link is created and requires the replacement to survive while the installed target is removed. The residual final compare-to-filesystem-syscall race remains the documented same-OS-user limitation of Node pathname APIs.
|
|
688
|
+
- **Provisional runner claims trusted PID without launch identity.** `publishProvisionalRunnerClaim()` accepted an existing `runner.pid` when its PID matched, even if that claim belonged to a different launch attempt. Beta.58 validates the 128-bit hex launch token before publication and requires both PID and launch token to match before treating an existing claim as idempotent. This closes PID reuse and same-parent multi-launch inheritance without weakening the later process-start-time confirmation.
|
|
689
|
+
- **Browser broker discovery and first pairing trusted too much of the loopback transport.** The old runtime/extension WebSocket subprotocols exposed long-lived shared tokens to whichever process owned the candidate port before that process proved broker identity. The first beta.58 repair replaced those bearers with role-separated HMAC server/client proofs, but the initial pairing grant still let the HTTP listener supply the token used to validate its own grant: a fresh extension could therefore be convinced by a malicious local listener that chose a self-consistent token/grant pair. The final beta.58 protocol separates the two problems. Already-paired clients send a random challenge only through an internal-header loopback endpoint, verify the broker's HMAC server proof under the owner-only persisted token, and only then open WebSocket with a one-time client proof; pending proofs expire against a monotonic five-second deadline and are consumed once. First pairing no longer receives a token from HTTP at all. `pair(open=true)` places a 30-second bootstrap only in the URL fragment, so it is absent from the HTTP request; the packaged content script runs at `document_start`, removes the fragment before page scripts, and retains the proof only in the extension isolated world. The extension sends the non-secret grant id, a random challenge, and an init HMAC under the fragment secret to `/pair-auth`; the broker verifies that init proof before allocating any pending exchange state, so fabricated grant IDs cannot exhaust the bounded registry. The real broker then derives the same temporary secret from its owner-only extension token, proves possession, consumes the matching client proof once, and only then returns the extension token. The public `/pair` document is always token-free, manual repair obtains bootstrap material from the isolated content script rather than page DOM, and the pairing envelope keeps schema version 2 for beta.55 rollback readability while `pairingAuthVersion: 2` marks the hardened protocol and rotates the historically exposed extension token. A persistent `migrationPending` marker prevents an old broker on the original port from causing adjacent-port split-brain. The fixed custom request header remains a non-secret browser-origin/challenge-churn guard, not local-process authentication. Same-OS-user malicious code remains outside the repository-only isolation guarantee.
|
|
690
|
+
|
|
691
|
+
The Cloudflare Computer decision is unchanged after rechecking the pinned current upstream head: it remains useful as a future default-off scratch runtime, but its preview status and unresolved container authentication/process-ownership contracts still exclude it from Machine Bridge's production authority core.
|
|
692
|
+
|
|
693
|
+
## 2026-08-08 beta.57 independent re-review and Cloudflare Computer assessment
|
|
694
|
+
|
|
695
|
+
The beta.56 candidate was deliberately treated as untrusted before owner activation. A new source-level review re-walked filesystem mutation/rollback order, state-lock authority, managed-job ownership, service-definition removal, Worker HTTP/OAuth parsing, logs/error projection, dead modules, documentation claims, and external cloud-runtime options. Beta.56 had no activation or acceptance record; its candidate manifest/tarball were removed before beta.57 packaged changes.
|
|
696
|
+
|
|
697
|
+
### New local findings
|
|
698
|
+
|
|
699
|
+
- **SSH validation could mutate an invalid existing path.** Existing private/public paths were chmod-normalized before the pair was cryptographically inspected. A mismatched or non-key public path at `0600` could therefore be widened to `0644` and only then rejected. Beta.57 validates the complete pair first. Permission normalization is subsequently bound to the inspected filesystem generation and performed on an opened descriptor, so replacement between validation and chmod is rejected without touching the replacement.
|
|
700
|
+
- **SSH compensation was pathname-based.** Partial public-key installation and later resource-state persistence failures could remove a private/public pathname even if another process had replaced that entry after generation. Generated-key inspection now retains non-enumerable internal generation identities. Rollback deletes only the same generated objects, attempts both files, preserves replacements, and retains primary-before-cleanup `AggregateError` causality. Same-directory installation also dropped an unreachable `EXDEV` copy fallback. A successful hard-link install verifies source and target are still the same inode/generation before staging cleanup; if source unlink fails, the target is rolled back by that installation identity instead of being left behind.
|
|
701
|
+
- **Exclusive hard-link publication had an internal contention false positive.** The fast parallel plan reproduced a security-audit record failure when one owner-state lock contender observed the successful publisher between `link(staging, final)` and `unlink(staging)`: the complete lock inode legitimately had two links for that tiny interval and the strict reader mistook it for persistent hard-link corruption. Secure-file now gives multiple-link rejection a stable code and a four-attempt/1 ms bounded retry helper. Only consumers of the exact exclusive-publication protocol use it; persistent external hard links still fail after the budget. The same audit found runner-claim reads were still using the generic bounded reader, so they now require no-follow path identity and single-link evidence too.
|
|
702
|
+
- **Managed-job runner corruption could become absence.** `runner.pid` ENOENT is still absence, but successfully read malformed JSON now fails closed. Pruning retains the directory and emits only a bounded error class; active-job inventory conservatively reports unreadable ownership as potentially live so uninstall cannot proceed from damaged runner evidence.
|
|
703
|
+
- **Service definitions were deleted by pathname after provider shutdown.** Launchd/systemd removal now snapshots the existing no-follow, single-link definition generation before stop/disable and removes only that same object afterward. A replacement or a definition that appears after an initially absent snapshot is retained and returns `definition_changed`.
|
|
704
|
+
- **Several workspace mutations read authority before waiting for the startup lock.** SSH resource registration, ordinary start, secret rotation, and persistent activation now use the early state only to address the same lock namespace, then reload the authoritative workspace state after the startup lock is held. A deterministic SSH regression simulates another writer committing a resource while the operation waits and requires the concurrent entry to survive. Daemon-only persisted network environment is also loaded only after this boundary, preventing a wait from pinning obsolete proxy/CA values into `process.env`.
|
|
705
|
+
- **The beta.53 state-location split was not real.** `state-locations.mjs` had zero importers. Its Windows root, environment handling, and workspace hashing had already diverged from the live `state.mjs` contract, while actual direct `state.mjs` fan-in is 22, not the previously reported 17. The module is removed and the earlier architecture/changelog claim is corrected. `package-identity.mjs` remains a real independent boundary.
|
|
706
|
+
|
|
707
|
+
The HTTP/OAuth pass found no new release blocker. Worker production code has no direct `request.json()`, `request.text()`, `formData()`, `arrayBuffer()`, or `blob()` body consumption; JSON/form and unauthorized-body handling use the shared bounded reader. Dynamic OAuth registration caps redirect count/length, canonicalizes only HTTPS or loopback HTTP redirects without credentials/fragments, and the authorization page reflects only whitelisted fields through HTML escaping under a restrictive form-action CSP. Error/log scans found no new public details carrying raw commands, stdout/stderr, credentials, account identity, or local paths.
|
|
708
|
+
|
|
709
|
+
### Cloudflare Computer
|
|
710
|
+
|
|
711
|
+
The review also examined `cloudflare/computer@8758b51c8891c211dddd1903d2ee2d12a75ac7ff` (552 tracked files in the pinned shallow checkout). The package describes itself as preview/alpha (`@cloudflare/computer` `0.1.0-alpha.1`) and explicitly warns that the API is unstable and not suitable for production. Its useful design patterns are real: DOFS benchmarks actual Durable Object SQL statements/rows read/rows written; path resolution uses cache plus recursive SQL/index tuning; Worker-shell starts with outbound networking disabled; Worker-JavaScript exposes only host-granted capabilities and applies source/input/result/output/concurrency budgets; runtime handles require explicit disposal rather than relying on remote-object GC.
|
|
712
|
+
|
|
713
|
+
Those mechanisms are worth borrowing, but the project is not an appropriate production authority dependency yet. The container design documentation still lists connection authentication, RPC frame/in-flight limits, and process-user ownership as unresolved boundaries. The current container callback implementation routes a token-like UUID through the proxy path/query but `CloudflareContainerBackend.handleFetch()` only verifies `/ws` plus WebSocket upgrade and does not itself authenticate that token, relying on Cloudflare container reachability. That is weaker and less explicit than Machine Bridge's root-certified daemon/session, challenge, OAuth, and relay identity boundaries. Its DO-backed VFS would also move ordinary workspace file I/O into the same Durable Object rows-read/write billing domain that beta.54 had to optimize, and upstream performance data shows metadata-heavy operations can be good while large sequential reads and full package installs remain materially slower than native filesystems.
|
|
714
|
+
|
|
715
|
+
**Decision:** do not add Cloudflare Computer to the beta.57 dependency graph or current owner-machine authority path. The pinned public API is a workspace/runtime/backend library and does not provide a production MCP peer surface, so a future integration would be a Machine Bridge-owned backend adapter rather than MCP-to-MCP chaining. It can be valuable as a separately named, default-off cloud scratch/execution backend, preferably Worker-JavaScript or Worker-shell first. It must use an independent Durable Object namespace and explicit row/read/write budgets, host-owned backend allowlists, no inherited local credentials/state, closed-by-default outbound capabilities, bounded results/calls, and explicit disposal. It must not silently mirror or replace the authoritative local workspace. The container backend remains out of scope until upstream authentication, frame/concurrency limits, and unprivileged-process contracts are explicit and testable.
|
|
716
|
+
|
|
717
|
+
## 2026-08-08 beta.56 exact-head CodeQL follow-up
|
|
718
|
+
|
|
719
|
+
The accepted beta.55 branch was not merged after provider checks. Exact-head CodeQL raised one high-severity `js/insecure-temporary-file` alert at the POSIX managed-job directory `openSync` call. The new directory resolver opens an already-existing directory with `O_RDONLY | O_NOFOLLOW | O_DIRECTORY`, pins the descriptor while canonical/path identity is checked, and never uses `O_CREAT`; therefore the runtime operation is not temporary-file creation. The alert was nevertheless actionable as a release gate because CodeQL treats a file open reached from an operating-system temporary-directory test path as an insecure temporary-file sink when no secure mode argument is specified.
|
|
720
|
+
|
|
721
|
+
Beta.56 keeps the descriptor-pinned algorithm and passes an explicit owner-only `0o700` mode argument. Without `O_CREAT`, Node does not create or chmod the directory from that argument; the security behavior remains the existing no-follow, directory-only, descriptor-held identity check. The explicit mode makes the private ownership contract machine-readable to the static-analysis model. A behavior test intercepts the real `openSync` call and requires the third argument to be `0o700`; the architecture contract also requires that literal beside `O_NOFOLLOW` and `O_DIRECTORY`. The beta.55 candidate and acceptance were removed before this packaged-source change entered a new beta.56 release cycle.
|
|
722
|
+
|
|
723
|
+
## 2026-08-08 beta.55 filesystem-generation ABA correction after beta.54 activation
|
|
724
|
+
|
|
725
|
+
The exact beta.54 candidate was owner-activated successfully and independently verified as the live Worker, verified login daemon, committed service owner, and accepted package identity. Cloudflare per-minute analytics then validated the Durable Objects repair under higher request load: the complete beta.50 comparison window from 05:20 through 05:32 UTC produced 24,801 rows read for 546 invocations (45.42 rows/request), while the first three fully ingested beta.54 minutes from 05:34 through 05:36 produced 3,321 rows for 280 invocations (11.86 rows/request). The reduction is therefore real operational evidence, not a low-traffic artifact.
|
|
726
|
+
|
|
727
|
+
Exact-head Ubuntu CI subsequently failed `secure-file:test` before the full suite could proceed. The failing fixture intentionally read a file snapshot, unlinked the source, recreated the pathname, and required `preserveFileSnapshotSync` to reject removal of the replacement. On Linux the allocator reused the just-freed inode, so the replacement had the same device/inode pair and the production identity predicate accepted it. Because `unlinkRegularFileIfIdentitySync`, shared owned-JSON removal, Worker-secret cleanup, lock reclamation, state recovery, audit identity, and SSH-key inspection all rely on the shared identity primitive, this is a real ABA class: an inode number is unique only while that inode is allocated, not a permanent file generation. Merely changing the test to force a different inode would conceal the defect.
|
|
728
|
+
|
|
729
|
+
Beta.55 keeps exact BigInt device/inode identity and adds filesystem change-time generation when available. Real Node 26 BigInt stats expose `ctimeNs`; that exact value is carried into the shared identity. If either side of a comparison contains generation evidence, the other side must contain the same generation. Legacy/synthetic stat doubles without a change-time remain supported only when neither side claims generation evidence. The existing BigInt descriptor/path observations therefore become generation-aware without routing identifiers through lossy Number conversion. The secure-file regression now rejects same-device/same-inode identities whose `ctimeNs` differs, which deterministically models the Ubuntu reuse case.
|
|
730
|
+
|
|
731
|
+
The stronger identity exposed one legitimate self-change in Worker-secret setup: the file identity had been captured before an explicit chmod and reused for cleanup after that chmod. Beta.55 retains the creation identity for failure cleanup, then refreshes it immediately after successful permission enforcement and before the deployment callback. A deterministic test supplies one generation before chmod and a second stable generation afterward; cleanup must accept only the refreshed generation. A later candidate rerun exposed the corresponding directory distinction: managed-job directory ctime legitimately changes while runner children create and remove files, so ctime cannot be its stable object generation. POSIX managed-job resolution now opens the directory with `O_DIRECTORY|O_NOFOLLOW`, holds that descriptor across canonical/path rechecks, and compares exact device/inode to the pinned descriptor. The open descriptor prevents the old inode from being recycled during the check while allowing child-entry ctime churn. Windows retains exact device/inode path rechecks because Node does not provide equivalent directory-descriptor semantics. File snapshots and destructive file cleanup remain generation-aware.
|
|
732
|
+
|
|
733
|
+
Because `filesystem-identity.mjs` and `worker-secret-file.mjs` are packaged production files, beta.54 acceptance is invalid for publication even though its DO quota behavior is proven live. The running beta.54 service remains useful operational evidence. Beta.55 must repeat frozen-source verification, exact candidate preparation, owner activation, independent live verification, acceptance, guarded push, and exact-head provider checks.
|
|
734
|
+
|
|
735
|
+
## 2026-08-08 beta.54 Durable Object quota incident and independent read-amplification review
|
|
736
|
+
|
|
737
|
+
The owner reported a Cloudflare Durable Objects Free-plan warning that the account had consumed 90% of its 5,000,000 daily `rows_read` allowance. This review did not assume the warning was historical or caused by unrelated workloads. After the 00:00 UTC reset, authenticated Cloudflare analytics for the current account showed 1,344,623 rows read and 13,918 rows written by approximately 03:31 UTC. A second read-only query at 04:05 UTC showed 1,417,559 rows read (28.35% of the 5,000,000 daily allowance) and 15,231 rows written, proving that the old live runtime continued accumulating the same cost while the repair was developed. The usage grouped to one active `BridgeRoom` Durable Object rather than multiple namespaces. During the same window the object handled roughly 7,515 successful hibernation events, 3,822 successful HTTP events, 3,407 `responseStreamDisconnected` HTTP events, and 354 alarms. At that average read rate the daily rows-read allowance would be exceeded again, while the write rate was already close enough to the free-tier write budget that a fix could not safely trade reads for a new globally rewritten index.
|
|
738
|
+
|
|
739
|
+
The live beta.50 `server_info` also proved that this was not an unused legacy compatibility path. The current host was actively completing legacy resumable durable calls and opening the internal hibernatable terminal subscribers. Source tracing then identified a systematic O(events x retained-streams) read pattern. `BridgeRoom.fetch()` and every hibernatable WebSocket message called `expireOverdueCalls()`, which unconditionally delegated to durable `expireDue()`. That operation enumerated every `mcp-stream:*` row. A normal legacy call separately enumerated the same rows for request-key deduplication, stream admission, call activation, alarm deadline discovery, and call-ID lookup. Beta.45 intentionally removed beta.44's globally rewritten `mcp-stream-index` to cure rows-written amplification, but making per-stream rows authoritative left too many unrelated event paths using full prefix enumeration. The current quota incident is the read-side consequence of that earlier write optimization.
|
|
740
|
+
|
|
741
|
+
Beta.54 keeps the per-stream authority and does not recreate a global index. Legacy stream admission and durable-call activation are instead one transaction over one bounded prefix-list snapshot: request identity, expiry pruning, capacity, call admission, and the initial active record all use that observation. Production no longer exposes a separate request-key full-scan helper. A lost `prepare` response remains safe because a retry runs the same transaction and returns the existing stream as `resume`; a different argument fingerprint returns `conflict` and never sends a second daemon call. Sessionless legacy requests intentionally retain independent generations. The hot production path also drops one stream mutation because the transient pending stream and active durable call are now published together.
|
|
742
|
+
|
|
743
|
+
New random streams and daemon call IDs share one random 43-character generation suffix while retaining distinct `stream_` and `call_` domains. The suffix is generated from the stream identity, not from the reusable JSON-RPC request ID, so acknowledging one request and intentionally reusing its request ID still creates an unrelated generation. Terminal call lookup derives the stream key and performs one point read. If that row exists but is already terminal, duplicate daemon results stop there. Only when the derived row is absent may a pre-beta.54 independent random call ID use one bounded migration prefix scan; those records age out under the existing pending/reconnect limits.
|
|
744
|
+
|
|
745
|
+
Deadline recovery is now alarm-led rather than event-scan-led. A future persisted runtime alarm is sufficient evidence that all already-existing durable deadlines are protected, so ordinary HTTP and WebSocket entry performs no durable prefix scan. New durable admission commits the active stream row and advances the persisted alarm to no later than the exact operation deadline inside the same Durable Object storage transaction. The event cannot therefore crash between durable ownership publication and alarm coverage; the later scheduler only coalesces daemon/transient deadlines and does not enumerate durable rows. Detach/rebind preserves or extends existing deadlines and still schedules after the state change; settlement/cancellation/expiry only remove deadlines. If alarm inspection fails, or if no alarm exists or the alarm is already due, ordinary entry performs the old bounded durable sweep as a conservative recovery path. Actual alarm delivery remains failure-propagating and enumerates durable records to expire due ownership and schedule the next deadline. This preserves cross-event deadline safety without making every high-frequency hibernation event pay O(retained streams).
|
|
746
|
+
|
|
747
|
+
The regression suite now measures reads as well as writes. With 24 retained background stream rows, a new combined legacy durable call must perform exactly one prefix list at admission. New call-ID recovery must use one point get and zero lists; terminal settlement with the already-known call view and a later duplicate terminal lookup must use zero lists. The combined production lifecycle uses three stream-row mutations through terminal acknowledgement. A synthetic pre-beta.54 call ID must show exactly one derived point miss plus one bounded migration list. Runtime-alarm tests prove that a future alarm suppresses durable deadline reads entirely and a missing alarm retains the recovery scan. Combined-admission tests separately require the active stream row and operation-deadline alarm coverage to be committed by the same transaction. Production `startEventDrivenStreamCall(createStream)` tests also prove first dispatch, identical-request resume without a second socket send, and changed-argument conflict.
|
|
748
|
+
|
|
749
|
+
Operational visibility now includes isolate-local aggregate stream-read counters (`stream_rows_read_estimate`, point gets, prefix lists, and list-returned rows) beside the existing mutation/alarm counters. These metrics never contain storage keys, stream/request IDs, OAuth/account identity, or response content. They are diagnostic estimates rather than Cloudflare billing authority; post-activation validation must compare them with Cloudflare account analytics. The accepted live runtime remains beta.50 until the owner activates a new candidate, so no reduction is claimed yet. Beta.53 never reached candidate preparation, owner activation, or acceptance and is superseded by this release-blocking beta.54 repair. Acceptance of beta.54 requires an exact candidate plus live post-activation evidence that the Cloudflare rows-read slope has materially fallen.
|
|
750
|
+
|
|
751
|
+
## 2026-08-08 beta.53 independent architecture, state-integrity, privacy, and lifecycle re-review
|
|
752
|
+
|
|
753
|
+
This review was performed independently after the beta.52 Pi/LoopX comparison and intentionally invalidated that unactivated candidate when new packaged fixes became necessary. The review re-walked architecture boundaries, dependency fan-in, persistent-state reads and deletion authorization, lock/recovery snapshots, service-provider removal, OAuth Durable Object state, browser/resource flows, public error projections, logging, subprocess settlement, test coverage, documentation claims, ignored local evidence, and dead/duplicate adapters. Beta.52 had no owner activation or acceptance record and its candidate files were deleted before beta.53 preparation.
|
|
754
|
+
|
|
755
|
+
### Findings and changes
|
|
756
|
+
|
|
757
|
+
- **Ownership evidence was not uniformly descriptor-coherent.** Process/startup locks, managed-job locks, and shared owned-JSON removal could combine a pre-open path stat with bytes from a later open. The secure-file boundary now exposes the exact BigInt descriptor stat used for identity; ownership snapshots derive content, device/inode, size, mtime, and link count from that one descriptor, reject hard links, and perform identity/single-link checks again at destructive removal. A real regression discovered during this hardening—ordinary `Stats.mtimeMs` fractional precision versus BigInt-stat integer milliseconds—was fixed rather than hidden by weakening CAS checks.
|
|
758
|
+
- **Corrupt-state backup had a read→rename race.** State JSON can no longer be classified as corrupt from one file and then path-renamed after replacement. `file-snapshot-preservation.mjs` creates a complete backup from the exact read bytes, removes the source only if its identity still matches, removes a misleading backup on mismatch, and preserves primary-before-cleanup error causality. Direct fault injection covers success, mismatch, source failure, cleanup-ENOENT, and AggregateError ordering.
|
|
759
|
+
- **Destructive control evidence had inconsistent file hardening.** Worker deletion inventory, state/recovery/root markers, internal managed-job control JSON, service ownership, and the autostart log-schema marker now use the same no-follow/lossless-identity/single-link rules appropriate to Machine Bridge-owned state. Optional project discovery and explicit user files remain deliberately less restrictive where omission or hard links cannot grant authority or justify deletion.
|
|
760
|
+
- **Remote/provider prose was influencing destructive continuation.** Worker deletion no longer treats `not found`-like Wrangler stderr as success. Systemd unit removal now uses a pure structured decision over definition presence, parsed active state, and disable exit status; nonzero disable cannot authorize removal of an installed definition, and a missing file does not imply an active in-memory unit is stopped. A dead generic service-result message adapter was removed.
|
|
761
|
+
- **Main OAuth state validation lagged the refresh-store boundary.** Schema-1 main OAuth state previously checked only that five top-level maps existed. A dedicated deep validator and shared persisted-ID contract now validate production-shaped records before authorization. Invalid-role quarantine remains separate from structural corruption. Targeted coverage is 100% functions with substantial branch coverage and the module is included in the critical coverage gate.
|
|
762
|
+
- **File validation and later use diverged in browser resources.** Resource permissions, size, canonical path, and injected bytes now come from one validated descriptor snapshot. CLI browser pairing likewise no longer maintains a second credential parser: only the pairing store parses/validates tokens and the CLI consumes a port-only projection.
|
|
763
|
+
- **Persistence-domain fan-in review removed duplicate package identity parsing.** `package-identity.mjs` became the real shared source for static package root/name/version. This beta.53 note originally also claimed that `state-locations.mjs` had become an active path/hash boundary and reduced `state.mjs` fan-in to 17. Beta.57 later proved that part incorrect: `state-locations.mjs` had zero importers and divergent semantics, so it was removed rather than counted as an architectural improvement.
|
|
764
|
+
- **Audit/log maps had prototype-key and determinism gaps.** Structured log maps use null-prototype records, and operation-risk target redaction now both retains prototype-shaped own keys and sorts object keys before hashing so semantically identical objects do not create cardinality from insertion order.
|
|
765
|
+
- **Error control flow contained one remaining message coupling.** `search_text` now classifies skippable files by stable `BridgeError.details.reason` rather than English error strings. Worker health also rejects unbounded or non-token version strings before they can enter status/doctor diagnostics.
|
|
766
|
+
- **Ignored review evidence had regrown unnecessarily.** Reconstructible external clones, 10k-file benchmark corpora, raw coverage scratch, and obsolete endpoint-bearing validation files were removed. The ignored tree fell from about 135 MiB to about 12 MiB before new final-verification scratch was generated; `.workers.dev` endpoint hits in retained ignored files fell to zero at that cleanup point.
|
|
767
|
+
|
|
768
|
+
### Explicit non-findings / decisions
|
|
769
|
+
|
|
770
|
+
- The review did **not** replace real process/service timeout tests with virtual clocks: process-tree identity and OS-provider behavior need real lifecycle coverage.
|
|
771
|
+
- It did **not** impose single-link policy on user-selected external plans/resources. Internal ownership state and user files have different trust semantics; over-hardening this boundary was detected during the review and corrected.
|
|
772
|
+
- It did **not** add a custom MCP JSON serializer. The measured extra mirror stringify cost was about 2 ms at 1 MiB and about 12 ms at 7 MiB, while the largest process results already use the dedicated `$mcpText` projection; the complexity was not justified.
|
|
773
|
+
- It did **not** widen near-full module line caps merely to silence architecture gates. A proposed snapshot helper temporarily pushed `exclusive-file.mjs` over its cap and was instead extracted to a cohesive 20-line domain module.
|
|
774
|
+
- Same-OS-user malicious replacement in the final instant between identity observation and a filesystem syscall remains an explicit residual risk; Node path APIs do not provide an atomic compare-and-unlink/rename primitive. The project now minimizes that window and fails closed on persistent ambiguity, but OS/container/VM isolation remains the boundary for a hostile peer with the same user identity.
|
|
775
|
+
|
|
776
|
+
### Verification intent
|
|
777
|
+
|
|
778
|
+
Beta.53 is not accepted by this document. The source must pass targeted regression tests, architecture/type/lint/documentation gates, the 79-task fast plan, critical coverage thresholds, the full 108-task plan, privacy/history and supply-chain checks, then `release:candidate`. The repository owner must activate that exact candidate in a real terminal before the coding agent may independently verify the live Worker/daemon and record acceptance.
|
|
779
|
+
|
|
780
|
+
## 2026-08-08 beta.52 Pi/LoopX comparative source review
|
|
781
|
+
|
|
782
|
+
Before owner activation of beta.51, the external comparison was repeated against exact upstream heads `earendil-works/pi@e47b8e37a6211ebd0b2942fa87059d64f81eec02` and `huangruiteng/loopx@29b086a1752a5329cb46de220225f2902353af3e`. Shallow local review copies were placed only under ignored `.project-local/` and their HEADs were compared with GitHub metadata. The review enumerated and line-scanned each complete text/source tree (Pi: 1,347 text files / about 314k lines; LoopX: 2,230 text files / about 807k lines), searched repository-wide concurrency/state/retry/cache/lease/output/privacy patterns, and then read the relevant implementations and tests in full. This is a source-level comparative audit, not a claim that every generated fixture or vendored/lockfile line received equal manual attention.
|
|
783
|
+
|
|
784
|
+
Pi's directly transferable mechanism was bounded concurrency around independent metadata work. Machine Bridge already had stronger write semantics than Pi's per-file mutation queue: multi-path canonical reservations, case-alias handling, flushed staging, expected-hash CAS, no-overwrite target publication, rollback, and lossless filesystem identity. It also already had analogues of Pi's stale-generation/single-flight patterns in relay generations, browser `startPromise`/generation ownership, and per-socket daemon cleanup. Pi's provider-level deferred tools are valuable inside model APIs that support `additional_tools`, but they cannot be reinterpreted as task-dependent MCP tool visibility without changing the protocol/client contract. Pi's abortable provider retries similarly do not replace Machine Bridge's existing live relay state machine; the remaining synchronous retry helper is release-only rather than an interactive hot path.
|
|
785
|
+
|
|
786
|
+
The comparison did expose two serial read fan-outs. `list_dir` could inspect up to 10,000 entries but awaited each path metadata read before starting the next. A 10,000-file benchmark measured a stable serial median near 159.7 ms versus 78.7 ms with a conservative batch width of 16; earlier cold/noisy runs showed about 2.6-2.9x improvement. `directory-metadata.mjs` now prefetches only one bounded batch with `Promise.allSettled`, consumes results in directory enumeration order, checks cancellation before scheduling and before each yield, and throws an error only when the caller actually consumes the corresponding entry. An error in a later prefetched entry therefore cannot escape after the caller has already truncated, preserving the observable serial boundary.
|
|
787
|
+
|
|
788
|
+
`search_text` had the larger read amplification: each of as many as 10,000 default / 100,000 maximum files performed secure open/fstat/read/close/UTF-8 work serially. A no-match 10,000-small-file benchmark measured about 649.4 ms serial versus 287.4 ms with ordered batch-16 reads; the initial prototype measured roughly 978.5 ms versus 287.6 ms. `workspace-search.mjs` separates scheduled files from externally visible visited files, settles a bounded batch concurrently, then consumes in original walk order. Old and new algorithms produced byte-identical results for a full no-match scan, a `max_matches=50` early stop, and a `max_files=37` stop. Tests require exact peak concurrency, ordering, limit accounting, cancellation, suppression of a later same-batch failure after an earlier match has already completed the request, and propagation of that same failure when it is actually consumed.
|
|
789
|
+
|
|
790
|
+
LoopX's strongest transferable idea was not its lease/CAS machinery—Machine Bridge's filesystem and release mutations already fail closed at stronger identity/commit boundaries—but its explicit distinction between hot read models and cold diagnostic facts. Its CLI output-budget tests also verify semantic keys and growth slopes, not only one snapshot length. Machine Bridge now applies that discipline to both compact `server_info` and the new `project_overview detail=summary`: absolute output ceilings, exact summary shape, and high-cardinality fixtures make summary size independent of exact tool membership/per-tool pending maps. A coverage expansion caught and corrected an initial projector bug where a non-record summary input could be mistaken for an empty record; the fixed projector returns full/non-record inputs unchanged and reaches 100% function / over 90% block coverage in focused measurement.
|
|
791
|
+
|
|
792
|
+
`project_overview` previously repeated exact tool membership across local tools, daemon tools, and remote authorization, retained account identity/routing fingerprints, and returned path/size for every displayed top-level entry. The summary form removes those cold fields while preserving policies and counts. Critically, remote summary is not sent below the authorization boundary: normal and legacy resumable Worker paths send an empty/default request to the daemon, apply the account-role/daemon-policy intersection, and only then project the compact result. Durable pending-call metadata stores only the validated `summary` selector in addition to the pre-existing account transform fields. Full remains the default and historical persisted records remain valid. On the owner workspace a local full projection measured about 7,484 bytes versus 2,732 bytes summary; applying the Worker owner decoration measured about 10,503 bytes full versus 3,537 bytes summary.
|
|
793
|
+
|
|
794
|
+
Several attractive external patterns were deliberately rejected. LoopX's changed-surface canary planner can return no profile for an unclassified path and would create a second dependency map beside Machine Bridge's already parallel 79-task fast gate, so it is not a release or development correctness boundary here. Pi's database writer leases are unnecessary for local locks that are never reclaimed merely because a live owner exceeded a TTL. LoopX local revision packets do not supersede no-follow descriptor reads, BigInt identity, CAS, atomic/no-overwrite publication, and rollback. Skill/context discovery was measured before considering more concurrency: warm calls were tens of milliseconds, while their large payloads are explicit cold-path instruction/skill/command content with existing omit/fingerprint controls. No second fan-out was added there.
|
|
795
|
+
|
|
796
|
+
Because the directory and search changes affect packaged runtime bytes, the prepared beta.51 tarball and manifest were deleted before any owner activation. Beta.51 has no activation or acceptance record. Beta.52 is the next eligible candidate.
|
|
797
|
+
|
|
798
|
+
## 2026-08-07 beta.51 exact-head provider correction
|
|
799
|
+
|
|
800
|
+
The first guarded push of the owner-activated and locally accepted beta.50 candidate passed package audit, dependency review, governance, workflow policy, macOS platform/install, Ubuntu full, and CodeQL Actions checks, but exact-head provider evidence exposed two independent blockers. JavaScript/TypeScript CodeQL reported `js/useless-assignment-to-local` for `ownsTemporary = false` immediately before the successful return in `replaceFileAtomicallySync()` and `staged = false` immediately before the successful return in `atomicWriteText()`. Both assignments are semantically dead: the surrounding function returns immediately and no `finally` observes either flag. Beta.51 removes only those assignments; failure-path ownership and staging cleanup state remain unchanged.
|
|
801
|
+
|
|
802
|
+
Windows `check:platform` independently reached `security-properties:test` and failed before the intended managed-job canonical-target mismatch assertion. The test injected `/tmp/jobs` as though it were already canonical and compared the second `lstatSync` target against that literal. `resolveManagedJobRootIfPresent()` correctly resolves the requested root first, so Windows produced a drive-qualified path and the fixture returned its synthetic outside identity too early. The same fixture also asserted POSIX output strings for the BigInt identity success cases. Beta.51 makes only the fixture portable by deriving the root/outside paths with `path.resolve()` and the expected job directory with `path.join()`. The production managed-job directory resolver and its lossless identity checks are unchanged.
|
|
803
|
+
|
|
804
|
+
Because the CodeQL repair changes packaged production source, beta.50 cannot be published despite successful owner activation and local acceptance. Its tracked acceptance record is removed in accordance with the release contract. The running beta.50 service remains valid operational evidence but not publication authorization. Beta.51 requires a new full local gate, exact candidate, owner activation, Machine Bridge live verification, acceptance, guarded push, and exact-head provider checks.
|
|
805
|
+
|
|
806
|
+
## 2026-08-07 beta.50 independent adversarial re-review
|
|
807
|
+
|
|
808
|
+
Beta.49 was not activated. A second independent pass treated the prepared beta.49 candidate as untrusted and reconstructed architecture, file-transaction semantics, stream admission, cleanup causality, filesystem identity, logging fields, ignored-local retention, tests, package boundaries, and release documentation from source. That review found a release-blocking patch-transaction defect: the beta.49 mutation coordinator used one case-insensitive resolved-path identity on macOS/Windows, but patch collision preflight case-folded only Windows and the patch transaction used overwrite-capable rename for staged targets. On a case-insensitive macOS volume, distinct patch strings such as `Foo` and `foo` can denote one directory entry; a single multi-file patch could therefore pass preflight and let a later staged target replace an earlier one. The beta.49 private candidate was deleted before any owner activation.
|
|
809
|
+
|
|
810
|
+
Beta.50 makes conflict identity a shared primitive and changes staged patch target publication to no-overwrite hard links. `EEXIST` after preflight becomes `conflict/target_appeared`; the transaction never treats a late alias as permission to overwrite. Transaction mechanics now live in `workspace-file-transaction.mjs`, separate from `WorkspaceFileService` authorization/path/content orchestration. Whole-file and patch staging failures aggregate the primary failure with cleanup failures without exposing private paths or content. Once a create-only target has actually committed, a later staging unlink failure is reported as a fixed warning rather than converting a completed side effect into a false failure. Patch rollback preserves the primary commit error followed by every rollback/artifact cleanup cause.
|
|
811
|
+
|
|
812
|
+
The same cleanup audit found an unused but unsafe branch in the shared synchronous exclusive-file helper. Its former `cleanupTargetOnFailure` option unlinked the target whenever the exclusive hard-link step failed before `linked=true`; on `EEXIST`, that target belongs to the pre-existing winner and must never be removed. No production caller used the option, but beta.50 removes the behavior and adds a regression that supplies the old flag and requires the existing target to survive. Pre-commit descriptor/staging cleanup failures now aggregate with the primary error. Post-commit staging cleanup remains a successful commit with fixed warning metadata and non-enumerable internal cleanup evidence. Worker-secret setup consumes that evidence immediately, performs one identity-checked second deletion attempt, and refuses to invoke Wrangler if a secret-bearing hidden staging file cannot be proven removed.
|
|
813
|
+
|
|
814
|
+
A portability sweep generalized the beta.47 Windows finding. `secure-file.mjs` previously converted `dev` and `ino` to Number and, if either value was not a safe integer, returned `true` from its identity comparator. That was an explicit fail-open path at a primitive used by Worker fingerprinting, owner state, service configuration, pairing, operation authorization, hardened toolchains, security audit, and other sensitive reads. Beta.50 introduces a shared lossless filesystem-identity boundary. Default descriptor/path identity observations use BigInt `fstat`/`lstat`; safe injected Numbers remain supported, while unsafe Numbers fail closed. Process/startup locks, managed-job locks/directories, exclusive owner-file removal, Worker-secret cleanup, audit caching, and SSH-key snapshots use the same exact device/inode representation. Fault injection uses adjacent inode values `2^53` and `2^53+1`, which collapse under IEEE-754 Number, and requires them to remain distinct.
|
|
815
|
+
|
|
816
|
+
The Worker review found over-wide coordination rather than a missing lock. Legacy `McpStreamChannel` serialized every stream subscription through one Promise and held that queue across the post-registration storage read, even though subscription capacity is per stream and the count-to-`acceptWebSocket` step contains no await. Beta.50 removes the global queue. A deterministic test blocks stream A's second storage read while stream B must finish subscription, preserving the required storage-recheck race defense without unrelated-stream head-of-line blocking.
|
|
817
|
+
|
|
818
|
+
Logging and local-retention privacy were reviewed independently of publication scanning. Unexpected Worker HTTP failures now log only a low-cardinality route class, not request path text. Routine Worker deployment logs no longer retain Worker name or endpoint; the explicit operator handoff remains unchanged. The ignored `.project-local` tree contained no detected AWS/GitHub/npm/JWT token shapes, but it had accumulated reconstructible toolchain experiments, old live-endpoint probes, and machine paths. Removing only artifacts with no tracked dependency reduced it from about 406 MiB to 12 MiB and eliminated remaining `workers.dev` endpoint matches. Ignored storage remains machine-private operational state, not an exemption from retention/privacy discipline.
|
|
819
|
+
|
|
820
|
+
A second fail-open search distinguished malformed metadata from storage failure. Startup/daemon process-lock reads, managed-job recovery/transition locks, owner-state locks, and `runner.pid` formerly had catch paths that could collapse a bounded-read or storage error into malformed/null/fallback owner state. In the destructive cases, sufficiently old oversized metadata could then become reclaimable even though Machine Bridge had never reliably read the owner. Beta.50 separates the phases: ENOENT is absence, a successful bounded read followed by malformed UTF-8/JSON is content corruption with the existing bounded policy, and size/permission/I/O/descriptor/identity failures propagate. Recovery/pruning already retain unreadable managed-job state, so removing the deep fallback makes the existing fail-closed outer behavior effective. Oversized lock/runner fixtures verify the existing files survive.
|
|
821
|
+
|
|
822
|
+
The same pattern existed in autostart log rotation. `readLogSchema()` previously converted every exception to an empty string; `trimAutostartLogs()` interpreted that as a schema mismatch and truncated both daemon logs before rewriting the marker. An unreadable or oversized schema marker could therefore erase the very evidence needed to diagnose the storage failure. Beta.50 treats only ENOENT as first-run absence. A readable old marker still performs the intentional schema reset; any bounded-read, path-safety, permission, or I/O error stops before opening/truncating the log streams. The regression uses an oversized marker and requires both daemon logs and the marker to remain unchanged.
|
|
823
|
+
|
|
824
|
+
The source import graph remained acyclic. Near-ceiling modules were reviewed individually rather than split for line-count aesthetics: state-machine-heavy OAuth, daemon-socket, and diagnostic modules remain intact where extraction would scatter invariants. Same-OS-user malicious filesystem replacement remains an explicit residual risk outside the Node process's sandbox guarantees; mutually untrusted users still require OS/container/VM isolation. Beta.50 requires a new exact candidate and owner activation after the final frozen-source gates; no beta.49 activation or acceptance evidence exists.
|
|
825
|
+
|
|
826
|
+
## 2026-08-07 beta.49 comparative runtime review
|
|
827
|
+
|
|
828
|
+
A read-only comparative review of two actively maintained agent/control-plane projects was used as an adversarial design check rather than as a source-code dependency. LoopX's strongest transferable idea was the separation of canonical decisions from compact hot-path projections and explicit write-intent/conflict semantics. Pi's strongest directly applicable implementation pattern was per-canonical-file mutation coordination, including the requirement that cancellation not release a queue while an underlying filesystem operation is still in flight. Machine Bridge already had stronger local commit primitives than either pattern required: SHA-256 preconditions, no-follow and hard-link defenses, flushed staging files, atomic link/rename, and rollback-capable multi-file patches. The review therefore did not replace those primitives or adopt another agent runtime.
|
|
829
|
+
|
|
830
|
+
The concrete beta.49 defect was narrower: `LocalRuntime` serialized every `write_file`, `edit_file`, and `apply_patch` behind one process-wide Promise even when the operations touched disjoint files. The new `file-mutation-coordinator.mjs` accepts already-resolved absolute paths, normalizes case on Windows/macOS for conservative alias safety, registers all unique paths of a transaction before awaiting any predecessor, then releases them only after the callback settles. `WorkspaceFileService` resolves a single write/edit target before reservation and resolves all patch sources/destinations before reserving the complete set; mutable file contents, target existence, hashes, and patch hunks are re-read/evaluated only inside the reservation. Same-file concurrent edits therefore observe fresh state, multi-path patches block overlapping single-file mutations, and unrelated files can proceed independently.
|
|
831
|
+
|
|
832
|
+
Direct tests cover same-path ordering, independent-path concurrency, overlapping multi-path reservations, duplicate-path deduplication, callback failure cleanup, and real concurrent edits through `LocalRuntime`. Architecture checks prevent the global runtime queue from returning and require registration-before-wait/settlement semantics; critical coverage gates the coordinator independently. The same comparative review found a separate context-amplification issue: authenticated `server_info` always returned its full authority/tool/OAuth/observability graph even when an agent only needed version/readiness/relay health. Beta.49 therefore adds an explicit `summary` projection while preserving `full` as the default. Summary is a post-decision projection, not a separate authorization computation; it intentionally omits raw account ID, OAuth metadata, exact effective/advertised tool arrays, per-tool Worker observability, and local security/trust detail. Tests require materially smaller local and remote fixtures and preserve the detail argument through both ordinary and legacy resumable MCP paths. The final projection review also moved daemon ready-socket status formatting out of `BridgeRoom` into a pure `daemon-status.ts` module; stale-socket reclamation remains in the Durable Object, while connected/disconnected status shape and readiness/liveness constants are tested independently. This reduces the Worker composition root from 843 to 820 lines and tightens its enforced ceiling to 830. A third interruption audit found that `ToolExecutor` checked cancellation again after the handler returned even though its inner observability/audit middleware had already committed a completed outcome. For a filesystem or process operation whose final await had crossed a non-cancellable commit boundary, that could expose `cancelled` to the caller after the side effect actually completed. Beta.49 removes only that retroactive veto: pre-handler cancellation and every manager/service cancellation checkpoint remain, remote cancelled owners still suppress terminal delivery, and the local handler result becomes the single settlement fact once it returns. A targeted regression cancels while a synthetic commit is in flight and requires successful completion, one completed metric, zero cancelled/failed metrics, and no registry leak. No credential, account, network, host-UI, release-activation, or authorization boundary changes. The LoopX lock comparison also exposed a local cleanup-causality gap: `withOwnerStateLock` previously threw a release/ownership error before rethrowing an already-captured callback failure, so a secondary cleanup fault could erase the primary state-operation cause. Beta.49 now aggregates primary and release errors in that order; a fault-injection regression corrupts the owned lock during a failing callback and requires both causes to survive.
|
|
833
|
+
|
|
834
|
+
The Pi supply-chain comparison confirmed two repository defaults worth evaluating separately. Machine Bridge already rejects non-exact root dependency ranges in its release architecture contract, so adding `save-exact=true` simply makes ordinary future npm save behavior agree with that invariant. Pi also uses `min-release-age=2`; Machine Bridge intentionally does not adopt that setting. npm 12 defines it as a tree-resolution filter that makes versions published inside the age window ineligible and documents that it may block `npm audit fix` from installing a newly published patched version. This project instead relies on explicit exact upgrades, committed lockfiles, hardened npm, zero-vulnerability audits, registry signatures/attestations, lifecycle-script allowlists, private toolchain verification, and isolated consumer-package installation. A blanket age filter would reduce emergency patch agility without adding equivalent evidence about a version that has aged past the threshold.
|
|
835
|
+
|
|
836
|
+
Beta.48 acceptance remains evidence only for the immutable beta.48 artifact. Beta.49 was later blocked before activation by the beta.50 independent review above and must not be activated or accepted.
|
|
837
|
+
|
|
838
|
+
## 2026-08-07 performance, stability, architecture, and privacy re-review
|
|
839
|
+
|
|
840
|
+
A separate review after the beta.47 Windows/process remediation reconstructed the slow-path and interruption mechanisms instead of treating individual timeout symptoms as transport failures. The original fast plan ran 79 gates serially and took 53.4 seconds in the review environment, leaving almost no margin below the remote 60-second foreground execution ceiling. The runner now preserves every gate while directly executing simple single-Node scripts, retaining npm for lifecycle/compound/non-Node scripts, and using bounded concurrency with explicit process-heavy serial barriers. The syntax gate was independently responsible for hundreds of process launches; it now parses the dynamically discovered JavaScript surface in one non-evaluating VM-module parser subprocess plus the existing shell-wrapper check. The final local fast-plan measurement after these changes was 17.8 seconds, about one third of the original wall time; elapsed time is diagnostic evidence rather than a release guarantee on other hardware. The release-level review also found wall-clock-only fault-injection delays: the direct managed-job suite spent 189.2 seconds waiting for a deliberately long descendant timeout even though the assertion concerns the timeout/termination state transition, not the absolute duration. The same real process-tree fixture now uses a 15-second deadline and completed locally in 23.9 seconds while retaining the ignore-SIGTERM descendant and terminal-state checks. The shell process-tree fixture remains real but uses a ten-second timeout. Coverage no longer launches `runtime-self-test.mjs` separately after `local-self-test.mjs`, because the latter already invokes `runtimeSelfTest`; this removes duplicate execution without removing a covered path.
|
|
841
|
+
|
|
842
|
+
The same review reduced request/context amplification. `project_overview` now runs independent inventory/Git-root probes concurrently and returns at most 40 top-level entries with total/truncation metadata. Server and bootstrap guidance no longer requires `resolve_task_capabilities` before straightforward direct file, Git, or shell work; the resolver remains the explicit refresh/routing path for project instructions, skills/commands, and application/browser tasks. Documentation now distinguishes modern static `server/discover` from the legacy initialization-time bootstrap path.
|
|
843
|
+
|
|
844
|
+
Relay architecture was re-reviewed rather than only retested. The connection lifecycle module had reached its 780-line responsibility ceiling; close/error classification, handshake/readiness validation, user-facing relay causes, and reconnect backoff now live in a dedicated boundary module while the original exports remain compatible. The new boundary is explicitly included in the critical-module coverage threshold rather than inheriting coverage only indirectly through the former file. Architecture, relay lifecycle, Markdown, lint, type checking, and the complete fast plan cover the extraction and the new verification runner. Package inspection also caught and corrected the newly created module's accidental `0600` source mode before candidate preparation; publishable source remains `0644`.
|
|
845
|
+
|
|
846
|
+
Privacy and supply-chain checks were repeated independently. Both production-only and complete `npm audit` runs reported zero vulnerabilities. `privacy:history` scanned the tracked/unignored surface plus 2,864 reachable history blobs and 128 commit messages without a finding. `npm pack --dry-run` ran the package prepack gates and showed no `.project-local` review artifacts, `.env` files, or private-key/container candidates in the publish list; source modules whose names describe secret handling are not secret artifacts. Host-rendered “tool called” UI decorations are outside the MCP server/logging boundary: Machine Bridge keeps routine per-tool logs debug-only and can reduce unnecessary calls, but only the host can change its own invocation indicators. Likewise, a host may cache an older input schema; the connected runtime's 60-second foreground ceiling remains authoritative and an already cached host schema cannot be invalidated server-side.
|
|
847
|
+
|
|
848
|
+
## 2026-08-07 version 3.0.0-beta.48 live candidate activation correction
|
|
849
|
+
|
|
850
|
+
The beta.47 exact candidate passed the local release gates and the repository owner then ran the required persistent activation command. Remote preparation advanced the same-name Worker to beta.47, but activation exited nonzero with `unauthorized`; the failure settlement reported that a compatible candidate service had been installed, started, and verified ready for automatic recovery. Independent post-failure inspection confirmed the live Worker health, launchd service owner, verified service daemon, and authenticated relay had all converged on beta.47. No beta.47 activation record exists. Under the release contract this is recovery evidence only, not successful candidate activation or acceptance evidence.
|
|
851
|
+
|
|
852
|
+
The failure mechanism is a phase-boundary dependency rather than a reason to relax recovered-success rules. General startup calls `ensureInitialOwnerAccount` after remote convergence so a first deployment can create its first owner. Candidate activation reused that same remote-preparation path even though activation requires an existing deployment. The redundant account-inventory request crosses the account-admin device-auth/OAuth boundary before the foreground candidate has proved relay readiness; an HTTP 401 is represented locally as `BridgeError("authentication_failed", "unauthorized")`. A failure there correctly remains nonzero even when post-deployment forward recovery later starts a compatible service, because foreground readiness was never observed.
|
|
853
|
+
|
|
854
|
+
Beta.48 separates those responsibilities. Candidate activation and its one bounded same-identity repair call remote preparation with `provisionInitialOwner=false`; ordinary startup retains the default and still provisions an initial owner when needed. Relay proof, service handoff, recovery, and account-administration authorization are otherwise unchanged. The existing post-deployment-preparation fault test now reproduces the live `authentication_failed`/`unauthorized` shape and still requires nonzero pre-readiness settlement with verified forward service recovery. Architecture checks bind both activation preparation call sites to the opt-out and bind ordinary preparation to the default-on behavior. Beta.47 must not be accepted or published; beta.48 requires a fresh exact candidate and owner activation.
|
|
855
|
+
|
|
856
|
+
The first beta.48 candidate-preparation attempt then exposed a separate coverage-only timing flaw in the process-tree regression: after a real shell timeout the test slept exactly 2.5 seconds before sampling a SIGTERM-ignoring descendant, while production escalation permits a graceful-termination interval plus bounded pre-escalation and final current-ownership observations. Under instrumentation the sample could occur before a valid SIGKILL. The regression now polls monotonically only through that complete production-derived bound and still fails closed if the descendant remains alive; the production supervisor and its deadlines are unchanged.
|
|
857
|
+
|
|
858
|
+
## 2026-08-07 version 3.0.0-beta.47 release-blocking remediation
|
|
859
|
+
|
|
860
|
+
Provider Windows CI exposed that managed-job directory identity used JavaScript Number values for `dev` and `ino`; large platform identifiers can therefore become lossy before equality checks. Beta.47 changes the default observations to BigInt without relaxing containment, symlink, type, canonical-target, or identity-change rejection. Fault-injected unsafe numeric and changing identities remain rejected.
|
|
861
|
+
|
|
862
|
+
A later local full verification caught an independent one-shot process cleanup race under coverage instrumentation: ownership capture and refresh could each consume a full process-snapshot budget before escalation. The supervisor now allocates one shared monotonic budget across those pre-escalation observations and still requires an independently current captured identity before SIGKILL. Failure to prove ownership remains fail closed.
|
|
863
|
+
|
|
864
|
+
The beta.46 acceptance is invalid because both fixes affect packaged production bytes. No beta.47 release claim is valid until a new exact candidate is owner-activated, live-verified, accepted, guarded-pushed, and provider-side Windows and security checks pass on that exact head.
|
|
865
|
+
|
|
866
|
+
## 2026-08-06 version 3.0.0-beta.44 Workflow Bundle independent review
|
|
867
|
+
|
|
868
|
+
A fresh high-risk change was created through the local Universal AI Development Workflow Bundle (`workflow 0.5.0.dev0`) after archiving an obsolete authority epoch through the bundle's own protocol. The review independently reconstructed architecture/import graphs, state and cleanup branches, filesystem/process/network boundaries, release diagnostics, tests, documentation, package inventory, and privacy evidence rather than accepting earlier green suites as proof. No live Worker, daemon, service, credential, GitHub, or npm-registry mutation is part of this local authority scope.
|
|
869
|
+
|
|
870
|
+
The review found three blocking source defects. Managed-job cancellation and job-directory lookup used `existsSync`, allowing storage failure to look like no cancellation and symlinked job directories to escape the job root. Temporary hardened-npm session settlement was duplicated across two owner activation entrypoints. GitHub workflow semantic checks could be forged by required strings in comments or unrelated values and were named Workflow Bundle, colliding with the actual local Workflow Bundle control plane before this review corrected the terminology. Beta.44 adds fail-closed managed-job boundaries, one session-settlement helper, structurally bound workflow policy checks, bounded release diagnostics, and behavior-level fault injection.
|
|
871
|
+
|
|
872
|
+
One lower-priority architectural issue remains explicit: the npm tarball still ships the complete `scripts/` directory, including source-only release and CI utilities, because the installed `package.json` also advertises those commands. Safely reducing that surface requires an explicit source-package/runtime-package split that transforms package scripts and proves package, install, consumer-security, and release workflows together; deleting files alone would create a misleading broken installed package.
|
|
873
|
+
|
|
874
|
+
## 2026-08-06 version 3.0.0-beta.43 workflow, state, and privacy review
|
|
875
|
+
|
|
876
|
+
A separate review after the beta.43 activation/publication corrections examined GitHub workflow governance, security-state discovery, development trust material, release diagnostics, module ownership, tests, documentation, and publication privacy. The repository previously mentioned a `workflow-bundle` check in historical notes but had no implementation. Beta.43 now includes a standalone Workflow Policy Gate workflow plus three independent modules for filesystem confinement, per-workflow policy, and six-workflow semantic contracts. It rejects unpinned or dynamic Actions, unreviewed actions and writes, privileged triggers, checkout credentials, missing timeouts/concurrency, direct event-data shell interpolation, malformed UTF-8, symlinks, hard links, and missing release/security gates. Local fast/full plans and exact-commit release verification require the same policy gate.
|
|
877
|
+
|
|
878
|
+
The review also found repeated `existsSync` prechecks in security-sensitive state. Depending on the platform, access or I/O failures could be normalized to absence before the secure reader ran. Browser pairing could then attempt new credentials; service environment or owner records could be treated as empty; activation/soak evidence and the official conformance checkout could be reported as missing. One shared inspection primitive now treats only `ENOENT` as absence. Pairing, service environment, service ownership, global config, legacy approval state, activation, soak, and conformance paths additionally require expected real file types and bounded no-follow/single-link reads. Fault injection covers EIO/EACCES, symbolic links, hard links, malformed content, and concurrent replacement boundaries.
|
|
879
|
+
|
|
880
|
+
The development macOS trust-broker cache had a second-order integrity gap: its marker covered only Swift source bytes, so a modified cached executable could be reused while the source digest still matched. The marker now binds source and compiled-binary SHA-256 under a versioned JSON schema. Cache reuse checks file type, link count, owner-only executable mode, and the exact binary digest. Compiler output is safely read before atomic installation, and primary plus cleanup failures remain distinguishable.
|
|
881
|
+
|
|
882
|
+
Release and CI helpers also used inconsistent raw `stderr` projection. A shared release diagnostic adapter now reuses the portable log-redaction policy, limits command identity to executable plus first subcommand, removes credential-bearing URLs, tokens, bearer material, email-shaped identities, home paths, control characters, and excessive output, and is itself package-tested. GitHub control scripts are part of syntax and ESLint coverage. A redundant npm-environment forwarding module was deleted so release and runtime callers use one authoritative sanitizer.
|
|
883
|
+
|
|
884
|
+
Beta.42 remains blocked and has no valid activation or soak evidence. These beta.43 changes modify packaged and Worker-relevant bytes. A new exact beta.43 candidate, owner-terminal activation, observed Worker/daemon/service convergence, registry publication, published-package activation, and a fresh complete seven-day soak are required.
|
|
885
|
+
|
|
886
|
+
## 2026-08-05 version 3.0.0-beta.42 independent release-path review
|
|
887
|
+
|
|
888
|
+
A fresh review after beta.41 candidate preparation found multiple blocking gaps that the existing 104-check suite did not cover. First, recovered activation success was phase-aware but not error-class-aware: any exception raised after foreground `ready_ack` could become success when the exact candidate service later converged. Second, npm execution-mode cleanup was duplicated, case-sensitive, and absent from the real owner candidate, registry-install, publish, registry-query, SBOM, and check-runner paths. A live probe proved that uppercase `NPM_CONFIG_WORKSPACES=true` makes npm fail with `ENOWORKSPACES`. Third, the owner install and publication paths still used the ambient npm 12.0.2 bundle, whose local tree contains undici 6.27.0 and brace-expansion 5.0.7. Fourth, broad verification catches treated permission, I/O, and timeout failures as corrupt state and could recursively delete private toolchains. Fifth, the consumer gate omitted optional dependencies and SBOM validation did not require dependency entries for every component.
|
|
889
|
+
|
|
890
|
+
Beta.42 narrows recovered success to explicit operational codes and preserves the original class, bounded detail, wrapper warning, and activation-record metadata. Candidate installation, published global installation, and npm publication create an ephemeral hardened npm session from the exact pinned artifacts; the ambient CLI only launches the repository script. One authoritative case-insensitive environment sanitizer is used across internal npm calls, while critical commands also pass explicit non-dry-run/non-workspace flags. Global package discovery requires a canonical real directory, bounded no-follow manifest, exact package identity, and a contained non-symlink CLI; read or parse errors are not normalized to absence.
|
|
891
|
+
|
|
892
|
+
Private hardened npm and Wrangler verification now distinguish integrity corruption from operational failure. Missing or mismatched private content may be reconstructed under the owner lock, but access denial, I/O failure, storage exhaustion, read-only filesystems, descriptor exhaustion, and timeout are propagated without removal. Critical hardened npm directories and its CLI must be real non-symlink paths. The consumer verifier installs optional production dependencies, and CycloneDX validation requires a unique dependency record for the root and every component with all edges resolving inside the document. Regression coverage includes uppercase/mixed-case npm modes, operational verification timeouts without reinstall, unknown post-ready `TypeError` failure, recovery-record consistency, global-package path identity, missing SBOM entries, dangling edges, and duplicate dependency records.
|
|
893
|
+
|
|
894
|
+
The full installation gate then exposed a macOS path-identity edge in the consumer fixture. npm 12.0.1 saved an external tarball dependency using a `/var/...` spelling while the child process reported its canonical `/private/var/...` cwd, causing `npm ls` to mark the otherwise correct package edge invalid. The gate does not ignore that result. It now verifies the source tarball, copies the exact bytes into a canonicalized fixture root, and runs install/audit/signature/tree/SBOM commands against that canonical root. A symlinked tarball itself remains rejected; only ancestor-directory aliases are normalized.
|
|
895
|
+
|
|
896
|
+
A second pass over the publication state machine found that both npm publication and GitHub Release creation still regenerated tarballs from the source directory after owner acceptance. That left accepted bytes, the GitHub asset, and the npm registry object only indirectly related. Beta.42 now reads the accepted candidate through bounded no-follow/single-link verification, copies those bytes into an owner-private staging directory, and uses that exact staged file for both channels. npm first runs the explicit prepublication gate and a tarball dry-run whose name, version, SHA-1, and SRI must match acceptance; actual upload disables lifecycle scripts. GitHub upload is followed by a REST asset query whose `sha256:` digest must match the current accepted artifact. Registry installation and formal soak repeat the GitHub digest check and npm registry hash comparison before changing global/service state.
|
|
897
|
+
|
|
898
|
+
The same pass found PATH-derived `git`, `gh`, and npm execution inside owner release, guarded push, backlog, portable acceptance, publication-lock, and soak-tag paths. npm lifecycle PATH prefixes could therefore select a repository shim instead of the intended control-plane executable. Those paths now resolve trusted absolute git/gh binaries outside workspace/state/home roots, while internal npm tasks use explicit CLI paths and hardened sessions. The published installer also resolves the owner's actual global npm prefix before invoking hardened npm, preventing a second unrelated global installation. Finally, cleanup after an independently verified irreversible result is classified separately: temporary staging, backfill notes, inactive candidate runtimes, and hardened publication-session cleanup may emit bounded warnings, but no longer convert a verified activation/upload into a false failure. Unknown exceptions and failed evidence writes remain blocking.
|
|
899
|
+
|
|
900
|
+
Beta.41 was never activated and is blocked. The live beta.40 service remains availability-only evidence; beta.42 requires a new exact candidate, successful owner-terminal activation, observed Worker/daemon/service convergence, acceptance, registry publication, published-package activation, and a new seven-day soak.
|
|
901
|
+
|
|
902
|
+
## 2026-08-05 version 3.0.0-beta.41 recovered-activation settlement correction
|
|
903
|
+
|
|
904
|
+
The exact beta.40 owner activation exposed a final state-machine defect. The isolated runtime was installed at 19:57:34 +08:00, the beta.40 Worker deployment evidence was updated at 19:57:53.739 +08:00, and an exact beta.40 launchd daemon reached verified post-`ready_ack` readiness at 19:58:39 +08:00. Authenticated server metadata then reported one ready socket, zero pending calls, matching Worker/daemon version, and no active relay outage. The service owner was committed to beta.40. Nevertheless, the command returned `unauthorized; a compatible candidate service was installed, started, and verified ready for automatic recovery`, exited nonzero, and wrote no activation or acceptance record. Launchd reported four runs and retained one prior exit code 1 while the fourth process remained ready.
|
|
905
|
+
|
|
906
|
+
The timing is important. Worker deployment evidence to final ready service took about 45 seconds, shorter than the ten-attempt candidate authentication exhaustion budget. Together with the verified-recovery message, this establishes that the foreground candidate had already completed end-to-end readiness and a later asynchronous authentication or handoff-stage failure triggered compensation. Beta.40 correctly recovered the exact runtime but incorrectly represented the verified final state as a failed transaction. It remains blocked because the owner command was nonzero and no activation record exists.
|
|
907
|
+
|
|
908
|
+
Beta.41 makes recovery settlement phase-sensitive. `candidateRelayVerified` is now an explicit security boundary. If false, compatible-service recovery remains diagnostic only and the original failure is returned. If true, and only if compensation independently verifies the exact candidate entrypoint/version, service-mode daemon identity, post-`ready_ack` checkpoint, and Worker version, activation returns a recovered success result. The result carries a bounded recovery class, the CLI emits a warning, and JSON output exposes the recovery fields. Any cleanup error or incomplete convergence still overrides recovery and fails with the primary and compensation errors preserved.
|
|
909
|
+
|
|
910
|
+
Fault-directed tests cover failed installation followed by verified reinstallation, an asynchronous `relay_authentication_failed` after foreground readiness, strict service-start failure followed by raw-provider convergence, authentication exhaustion before readiness, post-deployment preparation failure before candidate creation, and cleanup/recovery aggregation. No beta.40 acceptance, publication, tag, GitHub Release, or soak evidence is reused.
|
|
911
|
+
|
|
912
|
+
## 2026-08-05 version 3.0.0-beta.40 candidate-activation recovery correction
|
|
913
|
+
|
|
914
|
+
The owner-terminal beta.39 candidate activation verified and installed the exact tarball, created the hardened npm and private Wrangler toolchains, and advanced the same-name Worker to beta.39. The candidate daemon then received repeated explicit relay authentication rejection. Activation attempted compatible-service forward recovery, but the ordinary owned-service start helper treated missing readiness as terminal, stopped the provider, and returned only a generic cleanup-incomplete error. The resulting local state contained a committed beta.39 service owner and beta.39 service definition with no loaded daemon, while the Worker was already beta.39. No beta.39 activation record was written. Starting the previous beta.38 service could not recover because the exact-version Worker correctly rejected the older daemon. The exact isolated beta.39 runtime later installed and started its matching service successfully, proving that package identity, service definition, device root, and relay protocol were intact after convergence. This manual recovery restored availability but is not successful candidate activation evidence.
|
|
915
|
+
|
|
916
|
+
Beta.40 corrects both timing and compensation semantics. An explicit candidate authentication rejection still permits only one same-name deployment with the unchanged selected identity. After that single repair write, candidate startup receives ten bounded attempts with exponential delay instead of three sub-second attempts. Other network, TLS, proxy, health, integrity, and policy failures retain their existing non-redeployment classifications. The change addresses eventual convergence without generalizing remote-write retries.
|
|
917
|
+
|
|
918
|
+
Normal service handoff remains strict: `startOwnedServiceRuntime` must observe the exact committed owner, verified service daemon, startup-readiness checkpoint, and matching Worker. Failure compensation is now separate. Once remote preparation has advanced the Worker, recovery invokes only the platform provider start, leaves its retry policy intact, and independently polls the exact candidate daemon and Worker. Recovery is reported only after the daemon is service-owned, relay-ready, and version-matched. If it does not converge, the primary authentication failure and the precise recovery reason are returned together in an `AggregateError`; provider-active state alone is never called success. Fault-injection coverage reproduces the beta.39 failure shape and verifies that the primary authentication cause, missing-daemon state, bounded waits, lock release, and separate strict/recovery start paths remain observable.
|
|
919
|
+
|
|
920
|
+
Beta.39 is blocked. Its manually recovered live runtime is operational evidence only; the failed owner command and missing activation record prohibit acceptance, publication, and soak. Beta.40 later reached an exact ready service only through automatic recovery while the owner command still exited nonzero, so it is also blocked. Beta.41 requires a new exact tarball, successful owner-terminal activation, observed Worker/daemon verification, acceptance record, publication, registry-verified activation, and a new complete seven-day major-version soak.
|
|
921
|
+
|
|
922
|
+
## 2026-08-05 version 3.0.0-beta.39 installed-consumer supply-chain correction
|
|
923
|
+
|
|
924
|
+
A post-publication review found that beta.38's source-workspace audit did not describe the package installed by an ordinary npm consumer. The repository root used `overrides.undici=7.29.0`, so `npm audit`, the workspace SBOM, and all 100 existing checks were green. npm does not apply a dependency package's root overrides when that package is installed by another project or as a global package. A clean beta.38 consumer and the live global installation therefore resolved `wrangler@4.115.0 -> miniflare@4.20260722.1 -> undici@7.28.0`. The consumer audit reported one high and three moderate vulnerable dependency records, including CVE-2026-13697. This is a blocking published-artifact defect and invalidates the beta.38 soak.
|
|
925
|
+
|
|
926
|
+
Beta.39 removes Wrangler from `dependencies` and retains it only as a development dependency for source checks. Runtime deployment uses a separate package-owned control-plane toolchain. The shipped toolchain package manifest and lockfile select Wrangler 4.115.0 while overriding undici to 7.29.0 and sharp to 0.35.3. On first use, the CLI copies those exact bytes into a version-and-digest-keyed owner-only state directory, runs `npm ci` through a located npm 12 CLI, verifies the complete npm tree, requires zero production audit findings, verifies registry signatures, and writes a bounded audit marker. Process-identity locking prevents concurrent partial installations. A template mismatch, malformed marker, invalid dependency edge, wrong version, audit finding, signature failure, or interrupted install fails closed; a later invocation repairs only by reinstalling from the unchanged packaged lock. Online audit/signature evidence is refreshed after 24 hours before further Wrangler use. The private installation does not trust the ambient npm executable: it constructs a digest-keyed hardened npm 12.0.1 from three exact registry tarballs. The npm tarball retains its reviewed CLI while its bundled undici 6.27.0 and brace-expansion 5.0.7 directories are replaced with integrity-pinned 6.28.0 and 5.0.9 packages before the CLI is accepted. CI uses the same bootstrap. The fixed versions remove the advisories found in the stock npm 12.0.1/12.0.2 bundles without performing a mutable global npm installation.
|
|
927
|
+
|
|
928
|
+
The release gate now separately proves the consumer artifact. It packs the current package, installs the tarball into an empty local consumer with lifecycle scripts disabled, verifies the installed package identity, requires a zero-vulnerability production audit, validates that Wrangler/Miniflare are absent and no vulnerable undici is present, and generates and validates a CycloneDX 1.5 SBOM from that installed tree. The global-install smoke test uses the same tarball and additionally builds the private toolchain from the installed package before probing zero-argument startup. CI package-audit and `release:check` run the consumer gate explicitly. Workspace audit/SBOM remain useful development evidence but are no longer treated as proof of the published consumer graph.
|
|
929
|
+
|
|
930
|
+
The review also reproduced Dependabot's Node 24 failure. Only `devEngines.runtime.onFail` changes from `error` to `warn`; Node.js 26 remains mandatory in `engines`, `.node-version`, CI, CLI startup, and doctor. An isolated Node 24.19/npm 12.0.1 reproduction confirmed that metadata and lockfile operations complete with the warning, matching Dependabot's requirement without claiming Node 24 runtime support.
|
|
931
|
+
|
|
932
|
+
Finally, all nested npm operations that must create a real tarball remove inherited dry-run, global, prefix, and workspace configuration. A regression sets `npm_config_dry_run=true` around release-acceptance packing and proves the real tarball hash still matches.
|
|
933
|
+
|
|
934
|
+
## 2026-08-05 version 3.0.0-beta.38 live relay-liveness review
|
|
935
|
+
|
|
936
|
+
The activated beta.37 daemon experienced one real short transport interruption at 2026-08-05 07:38:57 local time (`2026-08-04T23:38:57.893Z`). Authenticated `server_info.daemon.relay_transport` retained close code 1006, category `connection_interrupted`, one reconnect attempt, a 2,631-millisecond outage, and a preceding ready interval of 550,090 milliseconds. The socket returned to ready state with `outage_active=false`. The launchd process remained PID 73042 with `runs=1` and `last exit code = never exited`, excluding daemon crash or service restart as the immediate cause. No stale-connection result was accepted; one `owner_missing_acknowledged` result reflected the documented at-least-once terminal replay case in which the owner had already settled and the duplicate was acknowledged to stop replay.
|
|
937
|
+
|
|
938
|
+
macOS network evidence is temporally specific but not causally complete. The default route was carried by `utun5`. At 07:38:53, four seconds before the WebSocket close, several system connection monitors changed the `utun5` link-quality classification from good to poor. The tunnel was owned by the Karing system extension (`com.nebula.karing.karingServiceSE`), and multiple tunnel-carried connections closed or reopened in the same window. Neither Karing's unified-log stream nor readable recent product logs supplied a node-change or upstream-failure record. The defensible conclusion is therefore OS Wi-Fi/TUN path degradation correlated with the close; the evidence does not prove a Karing implementation defect, a proxy-node switch, Cloudflare failure, DNS/Fake-IP behavior, or another specific upstream trigger.
|
|
939
|
+
|
|
940
|
+
The fresh source review nevertheless found a Worker-side amplification hazard. For every heartbeat, `BridgeRoom` refreshed the hibernation attachment and awaited Durable Object alarm inspection/mutation before sending `pong`. Alarm storage is normally fast, but it is an unnecessary dependency in the liveness acknowledgement path: latency or contention there could make a healthy peer appear silent. The same helper implicitly scheduled an alarm for tool results, after which matched results scheduled it a second time. Beta.38 makes attachment touch synchronous, emits `pong` first, schedules once afterward, and gives terminal results one explicit coalesced schedule. A second fault-directed pass found that the scheduling helper still allowed pre-write failures—such as a durable deadline read or invalid-socket cleanup failure—to escape. That could reject a WebSocket event after `pong`, or abort a newly registered foreground call before its daemon dispatch. Event-time scheduling now catches the complete calculation/transition/write path and reports only a bounded error class; the platform-invoked `alarm()` processor still propagates failures for retry. Architecture tests freeze heartbeat ordering and ownership, while runtime tests inject both storage-write and durable-deadline-read failures. These corrections improve resistance to Worker-side delay and storage faults; they cannot prevent a TCP/WebSocket interruption caused below the application layer.
|
|
941
|
+
|
|
942
|
+
## 2026-08-04 version 3.0.0-beta.37 second-order relay review
|
|
943
|
+
|
|
944
|
+
A fresh review of the unactivated beta.36 candidate did not reuse its green suite as proof. It found six second-order races. An asynchronous authentication proof created for an old socket could reject after reconnect and invoke the fatal path against the replacement. A send failure and two independent specific timeout sources could overwrite an earlier, more causal close category. Moving all invalidation through the idempotent cleanup owner caused runtime-alarm callbacks to enter alarm scheduling recursively. Removing that recursion without recalculating pending state would in turn retain a deadline computed before socket detach changed operation ownership. The hello-time diagnostic snapshot still had `outage_active=true` when stored on a ready socket. Finally, the Worker bounded but did not enumerate the claimed coarse transport error class, allowing arbitrary authenticated-daemon metadata into `server_info`.
|
|
945
|
+
|
|
946
|
+
Beta.37 binds authentication completion to the originating socket, applies a specificity-preserving close-category function, and makes alarm ownership explicit. Runtime-alarm callers invalidate with rescheduling disabled; after all socket transitions, the coordinator recomputes transient and durable deadlines from the resulting state and performs one coalesced alarm write. Socket closure starts before durable detach is awaited, so a storage exception cannot leave the invalidated transport open. A per-socket cleanup registry shares concurrent callback work, retains successful settlement, retries one transient failure in-event, and allows a later callback to retry persistent failure without duplicating disconnected metrics or warnings. Promotion from `probing` to `daemon` converts the reconnect snapshot to recovered state, extends its duration through actual readiness, canonicalizes timestamps, and restricts transport error classes to the stable operational vocabulary. Failed reconnect candidates preserve the prior healthy-ready duration rather than resetting it.
|
|
947
|
+
|
|
948
|
+
New tests reproduce an old-proof rejection after a replacement socket opens, a late send failure plus Worker liveness signal after a readiness timeout, a pending deadline that becomes earlier during candidate invalidation, concurrent cleanup plus persistent-failure retry, failed reconnect attempts after a healthy connection, and the exact ready-attachment diagnostic transition. Architecture tests forbid runtime-alarm invalidation from re-entering alarm scheduling and require the cleanup owner to remain scheduling-free. The existing beta.36 tarball is stale by promotion digest and was never activated, accepted, published, tagged, or deployed; beta.37 requires a new exact candidate and live owner verification.
|
|
949
|
+
|
|
950
|
+
## 2026-08-04 version 3.0.0-beta.36 relay-disconnect review
|
|
951
|
+
|
|
952
|
+
The reported disconnect did not coincide with a daemon restart. Launchd still owned the same process that began at 08:02:02 local time, while the Worker-facing connection identity was established again at 19:36:29. During final verification the same PID and `runs=1` service established another connection at 20:36:53 without a warning-level relay event, proving a second sub-ten-second same-process interruption. The daemon warning log contained earlier 1006 and transport-timeout outages but no record for either brief reconnect because the background service runs at `warn` and suppresses recovered interruptions shorter than ten seconds. The active socket used the operating-system network stack through a TUN route; Karing and Tailscale packet-tunnel components were both present. macOS recorded a Tailscale network-configuration change at 19:32. Around 20:34, shortly before the second reconnect, the system reported Wi-Fi link quality changing from Good to Poor, Tailscale closing and reopening several DERP paths, and repeated Karing network-path checks. This repeated correlation strengthens the system-network/TUN-path churn hypothesis, but no source proved which tunnel, upstream node, DNS/Fake-IP mapping, Worker edge, or TCP event caused either reconnect. The exact external trigger therefore remains unknown and is not represented as a code-proven Karing or Tailscale defect.
|
|
953
|
+
|
|
954
|
+
Source review found four internal defects that made this class of incident harder or more expensive to handle. First, a timeout path could set a specific close category and call `terminate()`, after which a generic WebSocket `error` event overwrote it with `relay_transport_error` before `close`. Second, Cloudflare can deliver both `webSocketError` and `webSocketClose` for one socket; both entered cleanup, repeating disconnected metrics, pending detach attempts, durable storage transactions, and alarm scheduling. Several candidate/readiness invalidation branches also expired or closed a socket outside that cleanup owner. Third, synchronous stale-socket reclamation discarded the asynchronous invalidation Promise with `void`; a Durable Object storage failure could therefore escape as an unhandled rejection and the cleanup was not retained by the event lifecycle. Fourth, the daemon already retained bounded outage state in memory, but a brief recovered interruption was unavailable from the authenticated remote `server_info`; `machine-mcp doctor` created a separate non-relay runtime and still displayed its skipped remote-relay check inside an otherwise green result.
|
|
955
|
+
|
|
956
|
+
Beta.36 keeps the first specific close category, makes attachment expiry return a one-time ownership claim, and routes every Worker invalidation through one cleanup method before closure. Stale-socket cleanup is retained through Durable Object `waitUntil`; its rejection is reduced to the coarse `daemon.socket.cleanup.failed` error class and cannot create a second rejection if observability itself fails. The next authenticated hello carries only schema 1 bounded fields: coarse application network route, outage count/active state/start/duration/attempts, close category/code, coarse transport error class, last disconnect time, and previous ready duration. The Worker validates enums, timestamps, integer ranges, metadata lengths, and schema before storing the summary in its hibernation-safe socket attachment. Only authenticated `server_info.daemon.relay_transport` returns it. Public health and default logs remain unchanged. Doctor now reports `running_service_process_inspected=false` and `remote_relay_inspected=false` and directs operators to authenticated server info.
|
|
957
|
+
|
|
958
|
+
The release dependency audit then found two newly disclosed high-severity transitive issues: `brace-expansion` 5.0.8 through ESLint/minimatch and `undici` 7.28.0 through Wrangler/Miniflare. The ordinary fix safely advanced `brace-expansion` to 5.0.9, while npm's forced remediation proposed an unrelated Wrangler downgrade even though `undici` 7.29.0 is the fixed 7.x release. Beta.36 therefore retains the already tested Wrangler 4.115.0 and Miniflare 4.20260722.1 packages but adds deterministic root overrides for `brace-expansion` 5.0.9 and `undici` 7.29.0. Complete and production dependency audits now report zero vulnerabilities, all 117 registry package signatures verify, and 38 packages carry verified attestations.
|
|
959
|
+
|
|
960
|
+
The existing 120-second same-daemon call recovery remains the continuity boundary. These changes improve causal evidence and remove duplicate Worker work; they do not guarantee that an external MCP host will keep its foreground HTTP/SSE request open during every network transition, and they cannot stabilize a third-party VPN/TUN or edge path. Long or cleanup-sensitive work still belongs in process sessions or managed jobs.
|
|
961
|
+
|
|
962
|
+
## 2026-08-03 version 3.0.0-beta.31 host-delivery margin review
|
|
963
|
+
|
|
964
|
+
The reported “message send timed out” interruption did not coincide with a daemon crash or a current relay outage. Launchd still owned one verified beta.30 daemon process with `runs=1`; Worker and daemon versions matched; and the local security-audit chain recorded a temporally aligned `exec_command` as successfully completed after 83,514 milliseconds. During the incident the Worker showed two durable `exec_command` calls still active, the oldest at roughly 81 seconds. Both later reached terminal state, while the host ended the task. The privacy-preserving audit deliberately omits raw command text, so an exact one-to-one mapping to the UI task cannot be proven; the timestamps and active-call counts nevertheless align. This separates execution completion from message delivery: persistence can preserve a legacy result, but it cannot force a host that has abandoned the response to resume it.
|
|
965
|
+
|
|
966
|
+
The beta.30 claim that 85 seconds was host-safe was therefore too strong. Its maximum execution budget plus the five-second Worker overhead allowed a terminal deadline of 90 seconds, leaving no credible allowance for request admission, Durable Object work, public SSE forwarding, host processing, or final assistant-message submission. The exact host deadline is external and not a Machine Bridge contract, so beta.31 does not claim a mathematically guaranteed bound. It conservatively caps remote synchronous execution at 60 seconds, records the Worker settlement deadline five seconds later, and directs longer work to durable process sessions or managed jobs. Admission and transport latency can consume part of that interval.
|
|
967
|
+
|
|
968
|
+
A second source review found that the first beta.31 candidate did not actually preserve that split. `daemonToolTimeoutMs` returned the execution budget plus overhead, and the Worker reused the same value both for its pending-call deadline and for the daemon `tool_call.timeout_ms`. Tools whose handler had no narrower inner timeout could therefore execute for the entire 65 seconds and consume the intended settlement margin. Beta.31 now computes an explicit `{executionTimeoutMs, settlementTimeoutMs}` budget: the daemon receives 60 seconds at most, while the Worker deadline is recorded at 65 seconds from Worker registration. Tests assert the two values independently for transient and durable stream paths, including the one-second integration timeout case.
|
|
969
|
+
|
|
970
|
+
The same review found a discovery-cache mismatch outside the current repository source: the live beta.30 Worker and canonical catalog enforced an 85-second maximum, while the tool schema supplied to the active ChatGPT conversation still advertised 600 seconds. A 120-second request was consequently generated by the host and rejected before dispatch. Beta.31 publishes a 60-second canonical schema and documents that an already-open host may need rediscovery or reconnection; the Worker remains the authoritative validation boundary. Because Machine Bridge cannot invalidate a host-owned schema cache, it must continue failing oversized requests with `side_effects_started=false`.
|
|
971
|
+
|
|
972
|
+
The same second review rejected the first observability fix as semantically ambiguous. A terminal publication with zero live internal subscribers can be normal: completion may precede subscription, after which the outer Worker obtains the persisted result through a storage-backed HTTP response. Conversely, a successful internal WebSocket send proves only Durable Object-to-outer-Worker queueing, not public SSE consumption or host receipt. Beta.31 therefore exposes separate aggregate counters for terminal publications, live internal-subscriber sends, storage responses, and storage-race sends/failures. `server_info.tool_delivery` explicitly states that host terminal receipt is not observable and that these counters cover legacy resumable Worker-internal storage and subscription transport only. Request identity, tool arguments, command text, and result content remain absent.
|
|
973
|
+
|
|
974
|
+
The resumption configuration retained another stale dependency on the former local 610-second relay envelope. A stream created during legacy prepare but never activated could remain for 730 seconds, consuming one of 64 bounded stream slots, even though hosted settlement now ends by 65 seconds. New unactivated records are capped at 185 seconds: maximum hosted settlement plus the 120-second replay window. Once activated, the existing operation/reconnect expiry algebra remains authoritative and extends the record as required.
|
|
975
|
+
|
|
976
|
+
## 2026-08-02 version 3.0.0-beta.30 interruption and recovery review
|
|
977
|
+
|
|
978
|
+
A repeated-call incident was reproduced while the local launchd daemon remained the same healthy beta.29 process. Recent security-audit records showed that many 20–78 second local calls reached terminal state even when the MCP host reported interruption, and the public MCP endpoint was briefly unreachable while the daemon connection identity remained unchanged. Live Cloudflare sampling after recovery showed no Worker exceptions; internal legacy subscription WebSockets closed frequently as part of their ordinary terminal lifecycle. The evidence therefore separates three cases: local execution failure, Worker/DO delivery interruption after admission, and a connection failure before Worker code runs. Only the first two are observable or repairable inside this repository; a pre-Worker edge/TCP/TLS failure still requires host retry or an independently operated alternate endpoint.
|
|
979
|
+
|
|
980
|
+
The review found five concrete amplification defects. The shared tool catalog advertised 120-second defaults and a 600-second maximum while the Worker rejected values above 85 seconds, allowing the host to generate requests that could never pass remote validation. After the catalog was corrected, the full gate exposed the inverse execution mismatch: `exec_command` and `run_process` still defaulted to 120 seconds locally when relay arguments omitted the field, and `run_local_command` could inherit a 600-second owner manifest even though the Worker waited only 60 or 85 seconds. Legacy internal subscription recovery retried only over roughly 400 milliseconds and every new resume forcibly closed the previous subscriber, allowing concurrent original/resume paths to replace each other. A lost prepare response caused the outer Worker to repeat no request, while a client retry had no explicit argument fingerprint for safe in-flight reattachment. Finally, the stateful burst limiter keyed only by Worker hostname, so every account and client shared one 120-request/minute bucket.
|
|
981
|
+
|
|
982
|
+
Beta.30 makes the foreground catalog truthful at its source: every configurable synchronous tool advertises a maximum of 85 seconds and a 30- or 60-second default on every MCP surface. A shared timeout module now drives both Worker admission and local relay execution; omitted relay values use the advertised default, and registered-command manifests are capped at the remote ceiling before process spawn. Owner-local registered-command invocation remains allowed to use its explicit local manifest budget. Longer remote work uses process sessions or managed jobs. Signed-session legacy preparation now retries a small bounded set of transient DO failures with an exact cloned request; token, signed session, typed JSON-RPC ID, tool name, and a canonical SHA-256 argument fingerprint bind the stream before daemon dispatch and remain through its bounded terminal-retention window, so an identical retry reattaches and changed arguments fail rather than duplicating side effects. Terminal subscriptions retry over a bounded multi-second window, permit four simultaneous receivers, multicast one persisted terminal result, and release only the delivery socket when a public reader disappears. A serialized subscriber-admission gate prevents concurrent limit bypass. Sessionless legacy requests remain independent by design because a shared bearer token and typed ID cannot safely identify one client; the exact-POST reattachment guarantee therefore requires the signed session issued by legacy initialization. Sessionless POST preparation is attempted once so an ambiguous response cannot cause an automatic duplicate side effect. For DPoP-bound access tokens, the outer Worker replaces any public internal header with a fresh random retry ID. The Durable Object transaction atomically consumes the proof and stores `proof replay key -> retry ID`; only that same opaque internal request may reuse the proof for at most four authorization attempts, and the binding is accepted only while the primary nonce record remains present. Different retry IDs, malformed IDs, missing nonce state, and ordinary client replay fail closed.
|
|
983
|
+
|
|
984
|
+
The burst guard is now layered: a 1,200-request/minute route/Worker bucket limits aggregate abuse, while a 120-request/minute route/subject bucket isolates authenticated credentials or an anonymous network identity through a truncated SHA-256 digest. Raw credentials, addresses, request keys, arguments, and results are never used as log fields. Limiter failure remains fail-open so the optional guard cannot become a control-plane outage.
|
|
985
|
+
|
|
986
|
+
A separate 100-second `runtime.event_loop.stall` warning was correlated exactly with macOS sleep from 07:15:31 to the 07:17:40 wake. The unchanged daemon PID and power-management log prove that warning was correct sleep/wake accounting rather than JavaScript blockage. Sleep can still interrupt network transport; same-daemon relay recovery remains bounded by the existing reconnect contract.
|
|
987
|
+
|
|
988
|
+
## 2026-08-01 version 3.0.0-beta.29 security-audit storage review
|
|
989
|
+
|
|
990
|
+
Review of the live 4,096-event audit chain found that moving persistence to a worker thread removed daemon event-loop blocking but did not remove storage amplification. Each sparse batch still read roughly 3 MiB, parsed the complete document, recalculated every SHA-256 link, serialized the complete state, atomically replaced it, and called `fsync`. On the inspected machine, a cold full-chain verification took approximately 1.3 seconds. The previous implementation therefore could accumulate audit work at ordinary interactive call rates even though tool-result delivery itself remained asynchronous.
|
|
991
|
+
|
|
992
|
+
Beta.29 gives each audit worker a verified storage session. The first access performs the complete bounded read and chain verification. Subsequent writes reuse that state only while the on-disk regular-file identity, size, modification time, and metadata-change time remain unchanged. A cross-process writer or external alteration changes that identity and forces a fresh secure read and full hash-chain verification under the owner-state lock. The cache never turns an unverifiable file into an accepted state, and a failed write does not advance the cached state. In the same environment, cached single-record writes fell to roughly 39–264 milliseconds; the remaining cost is full-file serialization, atomic replacement, and `fsync`, not repeated chain verification. A segmented append/checkpoint format remains a possible future optimization if sustained high-rate audit throughput becomes a product requirement.
|
|
993
|
+
|
|
994
|
+
The review also proved that the former retention contract was internally inconsistent. Every event field could be within its legal bound while 4,096 events exceeded the separate 4 MiB file ceiling. The storage layer then rejected the write and left the audit backend unhealthy before reaching its advertised count limit. Retention is now the intersection of the event and byte limits: event encodings are measured exactly, the oldest prefix is removed until both bounds hold, and the anchor advances to the last removed hash so the retained suffix verifies independently. Diagnostics expose `maximum_bytes` alongside the event maximum.
|
|
995
|
+
|
|
996
|
+
Faster writes exposed a pre-existing lock race. After exclusive creation failed with `EEXIST`, a holder could release the lock before the contender read it. The contender treated both `ENOENT` and malformed content as corruption. The lock reader now returns distinct missing, invalid, and valid-owner states; only missing retries, while malformed state continues to fail closed. Cross-worker audit tests exercise this release/acquire window and verify continuous sequence numbers with no lost events.
|
|
997
|
+
|
|
998
|
+
The live review also showed isolate-local event counters that could appear inconsistent with the persistent pending-call snapshot after an isolate restart. That is not durable-call loss: a persisted call may start in one isolate and complete in a later isolate. Beta.29 adds an explicit machine-readable `metric_scope` declaring the current-isolate lifetime and non-closed counter algebra; the persistent pending-call snapshot remains authoritative for active ownership.
|
|
999
|
+
|
|
1000
|
+
This source review does not activate beta.29, replace the running beta.28 daemon, deploy the Worker, modify live audit or legacy authorization files, publish npm, create a tag, or record acceptance.
|
|
1001
|
+
|
|
1002
|
+
## 2026-07-31 version 3.0.0-beta.27 control-plane resilience audit
|
|
1003
|
+
|
|
1004
|
+
A host-pressure incident exposed a causal chain that ordinary timeout tests had not covered: a metadata-heavy directory deletion exceeded its foreground deadline; subsequent minimal calls were delayed; and the daemon later classified the relay as unresponsive. Review of production diagnostics and the exact installed source showed that host filesystem/endpoint-security load was the trigger, but Machine Bridge amplified it. Timeout/cancellation performed synchronous process-table snapshots on the daemon thread, every remote call synchronously reread, rehashed, rewrote, and `fsync`ed the complete bounded audit file before returning a result, the process could remain alive after the timeout response, and heartbeat policy could not distinguish remote silence from a locally stalled event loop. Existing tests proved that `cancel_call` was sent and the Promise rejected; they did not prove post-timeout control-plane availability or completed process drain.
|
|
1005
|
+
|
|
1006
|
+
Beta.27 removes that amplification path. Process-group identity capture and revalidation use bounded asynchronous `ps` execution and retain exact PID/start-time/PGID fail-closed semantics. A capture is initiated before `SIGTERM`; ambiguous identity prevents escalation. Runtime process accounting now keeps calls in a draining state after their protocol owner finishes and exposes pending escalation supervision. The shared admission algebra reserves two of thirty-two Worker pending-call slots and two of sixteen local runtime slots for bounded diagnosis/recovery. Worker admission merges transient Promise calls with durable stream calls by tool under one serialized gate, so ordinary saturation cannot consume the control path before a request reaches the daemon.
|
|
1007
|
+
|
|
1008
|
+
Relay heartbeat policy is now a focused state machine. It measures expected-versus-actual timer delivery, records local event-loop stalls, sends a fresh heartbeat, and applies a bounded recovery interval before treating continued silence as a remote liveness failure. Diagnostics expose only timing/count classes, not network endpoints or message content. Genuine silence still terminates the half-open socket and enters the existing bounded reconnect path.
|
|
1009
|
+
|
|
1010
|
+
Security-audit persistence and startup chain verification are isolated in a dedicated Worker thread. The main runtime begins with an explicit `audit_initializing` projection, projects an allowlisted metadata record, queues it with a fixed capacity, and returns the tool result without waiting for disk I/O. The Worker batches records, serializes cross-process writers, verifies the existing SHA-256 chain, and performs owner-only atomic replacement and `fsync`. Batch acknowledgements avoid per-record duplicate snapshots. Health reports queue depth/capacity, worker readiness, retained events, chain status, and dropped-record count. Failures are warning-rate-limited and include only tool/coarse error class. Command text, arguments, paths, principal identifiers, content, form values, and results remain prohibited; account/client/family identifiers are salted and hashed only inside the audit storage boundary.
|
|
1011
|
+
|
|
1012
|
+
The review also retained strict module budgets by extracting call-capacity, relay-heartbeat, process signal/supervision/snapshot/ownership, and audit dispatch/storage/warning responsibilities rather than raising architecture thresholds. New behavior tests cover local stall recovery, true relay silence, capture-before-signal ordering, asynchronous escalation, draining ownership, end-to-end Worker/local reserved control capacity, mixed transient/durable admission, audit non-blocking dispatch/batching/tamper detection, warning suppression, and privacy projection. Remaining residual risk is explicit: the asynchronous audit queue is not a write-ahead log, so a process or operating-system crash before persistence may lose queued events and the in-process dropped counter cannot observe that loss; CPU, memory, filesystem latency, and network isolation are still OS/container responsibilities; a severely stalled kernel can delay all user processes, but one Machine Bridge call no longer performs known synchronous process enumeration or audit `fsync` on the daemon event loop.
|
|
1013
|
+
|
|
1014
|
+
## 2026-07-29 version 3.0.0-beta.26 GitHub publication ownership audit
|
|
1015
|
+
|
|
1016
|
+
After beta.25 was accepted and merged, several non-interactive paths attempted to start its guarded GitHub prerelease command: foreground MCP execution, detached shell/session variants, launchd handoff, and a managed job. Host lifecycle ownership terminated the long-running descendants before publication, and the managed job preserved a terminal `SIGTERM` result. A later calendar-triggered launchd attempt reached the concurrently prepared beta.26 guard and was rejected before fetch. No beta.25 tag, GitHub Release, or npm version was created. The event showed that candidate acceptance and green exact-commit CI are durable evidence but not proof of fresh operator intent for a later remote publication.
|
|
1017
|
+
|
|
1018
|
+
Beta.26 makes publication a distinct owner-terminal workflow boundary. `--publish`, `--publish-prerelease`, and `--backfill` require TTY-backed stdin/stdout/stderr plus `--owner-terminal-confirm` before fetch, full checks, tag creation, or any remote call. A process-identity lock in an owner-only common Git state directory serializes publication across the main checkout and linked worktrees, and safely reclaims only a verified stale owner. Release failures now throw through the lock scope instead of exiting before cleanup. Ordinary checks, candidates, acceptance, branch/PR work, Bash execution, and npm publication policy remain separate. This does not claim cryptographic human presence: arbitrary same-user code can allocate a pseudo-terminal, so stronger adversarial separation still requires an external protected release environment or OS user-presence mechanism.
|
|
1019
|
+
|
|
1020
|
+
Candidate verification under concurrent simulator and Rust/Wasmtime workloads exposed test-infrastructure false negatives rather than product failures: the publication common-Git-directory probe had a 5-second local metadata deadline, runtime Git success fixtures and the repository-root subprobe used hidden 10-second deadlines beneath 30-60 second Git operations, local self-test success fixtures used 5-30 second process/CLI budgets, the managed-job timeout-tree test read its descendant PID without first observing the fixture checkpoint, and `diagnose_runtime` reused a 5-second process-completion budget that was too short to distinguish scheduler starvation from a broken spawn boundary. Read-only Git metadata probes remain fail-closed with bounded 30-second deadlines; non-timing process and self-test fixtures use named 30-60 second budgets and phase-labelled errors; managed-job and shell process-tree tests wait for descendant-PID publication before timeout/cleanup assertions; maintenance-lock contention uses an explicit parent/child release handshake instead of a wall-clock holder; runtime diagnostics use a separate 30-second direct-process and shell health budget; process-lock and daemon-takeover helper processes explicitly receive `NODE_V8_COVERAGE=""` because Node otherwise reinjects the parent coverage directory into every child, multiplying profiler startup without adding threshold-relevant evidence; the exclusive-create proof uses four simultaneous child processes, preserving the cross-process single-winner invariant without multiplying Node cold-start cost across twelve contenders; daemon readiness and successful takeover/stop paths use a named 30-second observation budget; managed-job runners remain covered, while their trivial marker-writing child steps receive `NODE_V8_COVERAGE=""` and a 120-second success budget. Approval, resource-validation/redaction, bounded-output, discard-output, and cleanup/recovery fixtures use the same named success budget; the managed-job tree fixture uses a 180-second timeout plus a 150-second descendant-readiness window so timeout/tree-kill is evaluated only after the resistant descendant exists; cancellation semantics remain separate. The aggregate-output proof uses four steps with a 600-second observer, and the ordinary terminal observer is 480 seconds, exceeding a three-phase 3×120-second plan plus startup margin without altering production timeout semantics. Managed-job CLI list/inspect/submit/read fixtures use a distinct 120-second observation budget and preserve status/signal/error diagnostics because they validate state projection rather than a latency SLA. These nested CLI probes explicitly receive `NODE_V8_COVERAGE=""`; top-level local-self coverage remains active, and dedicated CLI-entrypoint plus managed-job fixtures retain the relevant gated evidence. The same stress run exposed a product race in the ten-second recovery grace period: a spawned runner could still be initializing before it created `runner.pid`, allowing a concurrent read to relaunch the job as interrupted. Launch now uses an owner-only provisional PID/one-time-token claim published by the parent; the child verifies and atomically upgrades it before executing, and claim collisions terminate the spawned child fail-closed. Under severe scheduler starvation, macOS could leave an already-exited child as a zombie while libuv had not yet delivered `exit` or `close`; the timeout callback now recognizes that exited-but-undrained state and the shared settlement fallback re-reads the real exit code instead of reporting a false timeout. Browser-broker connection, rejection, close, handshake, and state-convergence fixtures use a separate bounded 30-second observation window; the one-second request timeout and normalized browser-operation deadlines remain unchanged. Explicit short refusal, force-escalation, timeout, cancellation, and process-tree semantics remain covered independently.
|
|
1021
|
+
|
|
1022
|
+
## 2026-07-29 version 3.0.0-beta.25 MCP 2026-07-28 migration and repository audit
|
|
1023
|
+
|
|
1024
|
+
This section records the beta.25 migration state and is not the current protocol contract. At beta.25, the audit treated MCP `2026-07-28` as a protocol-state-machine replacement rather than a metadata bump, while temporarily retaining an explicit **stateful** MCP `2025-11-25` compatibility adapter. Current beta.61 deletes that stateful adapter and its session/replay stores entirely; the later remote-HTTP stateless initialization compatibility described in the 2026-08-11 section is a separate bounded interoperability layer and does not reuse those stores. In the beta.25 design, current HTTP/stdio requests carried version and capabilities per request, used `server/discover`, received no MCP session, never entered the older `Last-Event-ID` recovery store, and scoped cancellation to the request or response stream; the older initialization/session path was isolated behind the temporary adapter.
|
|
1025
|
+
|
|
1026
|
+
The review found and corrected several non-obvious protocol defects: header/body version mismatch had to take precedence over unsupported-version classification; a modern `initialize` request was incorrectly entering the legacy adapter; OAuth token plus JSON-RPC ID was incorrectly acting as a global modern request identity; unknown tools and malformed arguments were being wrapped as tool failures instead of protocol `-32602`; `subscriptions/listen` accepted missing or mistyped filters; non-object `structuredContent` values were discarded; and the first modern stream design retained an unresolved terminal Promise across Durable Object events. The final design uses one direct stream request plus a stream-scoped cancellation control that indexes only the active pending call; there is no modern prepare/subscribe descriptor, terminal-result registry, or replay state. Modern same-token/same-ID concurrency, removed methods, malformed non-dispatch, filter validation, arbitrary JSON structured content, and deterministic cancellation-control delivery now have direct regression coverage. Wrangler does not reliably surface raw TCP disconnects to local Worker stream callbacks, so deployed-edge cancellation remains a candidate-activation observation rather than a falsely claimed emulator proof.
|
|
1027
|
+
|
|
1028
|
+
A shared bounded JSON Schema 2020-12 validator now enforces the catalog before Worker dispatch and again at the local runtime boundary. Unsupported dialects or keywords fail during catalog compilation, automatic network `$ref` dereference is impossible, and schema depth/node/pattern/issue plus runtime-step budgets are fixed. The object path no longer calls `Object.keys()` before budgeting: each own property consumes work, so an 8 MiB high-cardinality object cannot bypass the traversal ceiling. Open modern `_meta`, capability-extension, and subscription-filter JSON is independently capped at 4,096 structural nodes, 32 levels, and bounded key lengths; resource subscription lists accept at most 256 bounded strings. One role-aware inspector is shared by modern and legacy Worker dispatch; it rejects hidden tools, missing names, non-object arguments, and schema failures before daemon send or legacy resumption allocation with bounded `-32602` data and `side_effects_started=false`.
|
|
1029
|
+
|
|
1030
|
+
The cancellation review found a DPoP-specific failure that ordinary Bearer tests would miss. The private `modern-cancel` control originally copied the public Authorization and DPoP headers and re-entered access-token verification after the original proof JTI had been consumed. The final design treats the random 256-bit stream ID as an internal-only cancellation capability: caller-supplied control headers are stripped by the outer Worker, cancellation is handled before OAuth inside the Durable Object, and the control request carries no OAuth/DPoP credential. Actual `/mcp` requests now also enforce the Streamable HTTP Origin rule; unrelated or opaque origins fail with 403, while OAuth navigation routes retain their separate behavior. CORS no longer reflects arbitrary `Mcp-Param-*` names and bounds the complete preflight header list.
|
|
1031
|
+
|
|
1032
|
+
The official MCP conformance checkout was run against the real local Wrangler/OAuth/daemon integration through a test-only bearer-injecting loopback proxy. A fresh checkout at `49103de6ed70804e940637bf3e9e29e4a3f54e64` (`0.2.0-alpha.10`) reported fourteen upstream dependency advisories, including seven high-severity findings; those packages remain outside the project dependency graph and candidate. The runner receives only a loopback proxy URL, never the injected bearer, is used as a short-lived audit tool, and is deleted after execution. The driver now rejects missing, symlinked, unlocked, malformed, or dependency-uninstalled checkouts before starting Worker integration, replacing ambiguous `spawn ENOENT` and `tsx not found` failures with explicit diagnostics. `http-header-validation` passes without exclusions. `server-stateless` and `caching` pass with check-scoped expected failures only for conformance-only diagnostic tools and prompt/resource feature families the production server does not advertise. The proxy is loopback-only, accepts only relative `/mcp`, maps to the exact validated HTTPS or loopback upstream path, refuses alternate same-origin paths and absolute-form targets before injecting the bearer, bounds request bodies, settles aborted uploads, and terminates the runner's process tree on timeout. The baseline cannot hide an unrelated regression and becomes stale-failing when an excluded check starts passing. The alpha conformance package is intentionally not a project dependency.
|
|
1033
|
+
|
|
1034
|
+
The dependency pass found Wrangler `4.115.0` as the only newer direct pin. Its bundled workerd remains the already reviewed `1.20260722.1`, while Miniflare advances to `4.20260722.1`; the exact workerd lifecycle-script approval therefore remains unchanged. The relevant Wrangler behavior is bounded `429` handling: reasonable `Retry-After` values are honored, waits above sixty seconds fail fast, and surfaced command failures include `retry_after_ms`. This improves candidate deployment/upload diagnosis without converting a cloud rate limit into an unbounded activation transaction. Complete and production dependency audits remain at zero known vulnerabilities, and registry signature verification reports no missing or invalid signatures.
|
|
1035
|
+
|
|
1036
|
+
The first authoritative full run exposed two verification defects rather than production failures. Under V8 coverage load, the shell process-tree fixture could reach its 200 ms timeout before the helper process had created `child.pid`; the test then misclassified absence of the fixture as a cleanup failure. The fixture now starts the actual bounded operation, waits up to four seconds for a syntactically valid descendant PID, and still relies on the five-second production timeout plus the normal process-tree termination path—there is no blind sleep or leaked process on failure. The next coverage pass correctly identified that the newly exported Worker argument validator lacked a direct call in coverage fixtures. Positive, additional-property, and unknown-tool paths were added, restoring `src/worker/tool-catalog.ts` to 100% function coverage without weakening its 95% gate.
|
|
1037
|
+
|
|
1038
|
+
A later candidate run exposed a second exact-boundary test race. The resistant descendant check waited exactly five seconds after the foreground timeout, equal to the production two-second graceful interval plus the maximum three-second process-ownership verification budget. A slow final `ps` identity probe could therefore complete just after the assertion. The test now imports those production constants and adds a fixed two-second scheduling margin; three consecutive self-test runs pass, and the runtime escalation algorithm and fail-closed ownership checks are unchanged.
|
|
1039
|
+
|
|
1040
|
+
A subsequent candidate gate exposed a managed-job verification race rather than a production cleanup failure. The runner deliberately persists a terminal status with `artifact_cleanup_pending=true` before deleting private runtime resource copies, temporary files, the active plan, PID claim, and cancellation artifacts, then confirms the status with pending cleared. The integration helper returned on the first terminal status and immediately asserted that those artifacts were absent, so scheduler load could expose the valid intermediate checkpoint. Diagnostic repetitions observed pending terminal states directly. The helper now waits for both a terminal status and cleanup confirmation, with deterministic checkpoint assertions and ten consecutive managed-job integration passes. The production persistence ordering and recovery semantics are unchanged.
|
|
1041
|
+
|
|
1042
|
+
A subsequent design review used Miguel Salinas's camelAI Durable Object/Code Mode write-up as a prompt for broader research rather than as an implementation template. Cloudflare's Code Mode and Dynamic Workers material demonstrates that compact code plans and lightweight isolates can reduce tool-context cost, but the feature remains experimental and solves a different trust/deployment problem. Machine Bridge therefore does not replace or demote local Bash, add model-generated JavaScript execution, or add a new sandbox dependency in beta.25. Its owner-facing product value depends on efficient direct shell access.
|
|
1043
|
+
|
|
1044
|
+
The applicable finding was tool-harness scaling. Recent routing studies report material accuracy degradation as catalogs grow and gains from task-specific shortlisting; production work on skill descriptions identifies overlapping descriptions as a distinct collision source; ToolBench-X shows that explicit recovery guidance often matters more than additional inference under tool failures; and ToolPrivacyBench shows that successful completion does not imply need-to-know information flow. The project now returns bounded set-level execution routes, ambiguity and fallbacks; rewrites high-collision descriptions; filters discovery by effective account authority; and exposes failure-aware advice without enforcing a smaller tool surface. Direct Bash remains the general escape hatch. The new bilingual regression corpus fixes expected route behavior for commands, shell, process sessions, jobs, Git/files, browser, applications, diagnostics, and protected resources.
|
|
1045
|
+
|
|
1046
|
+
The same review found that capability resolution used the daemon's full policy when composing application/browser metadata even after the relay had established a narrower account authority. That created an inventory side channel for delegated roles. `LocalRuntime` now passes the request's effective policy to bootstrap/resolution, and application scanning is skipped before touching the OS when the role cannot invoke it. Matching capability fingerprints may omit repeated static instructions, but every task still rescans and reranks dynamic capabilities; the fingerprint is neither authorization nor conversation identity. Because omission turns the fingerprint into a correctness boundary, it now binds target/scope, instruction provenance/precedence, skill sources, and complete command definitions rather than only content hashes and argv. The route engine reads only frozen policy-visible name/title/description metadata rather than cloning complete input schemas on every call, and its envelope identifies its schema plus non-probabilistic score semantics.
|
|
1047
|
+
|
|
1048
|
+
A second supply-chain pass found that an ambiguous unscoped `cyclonedx-npm` command can resolve to a dependency-confusion placeholder that prints a warning yet exits successfully, producing no SBOM. The project does not depend on that package: `sbom:test` now invokes the pinned npm CLI directly, bounds and parses its CycloneDX 1.5 JSON, verifies the current package identity and root dependency graph, rejects local filesystem paths, and is part of the full candidate gate. The CI package-audit job retains its independent temporary-file validation.
|
|
1049
|
+
|
|
1050
|
+
Release-state review found that a pending tarball could remain internally consistent with its manifest after later source edits. Candidate start/activation now recomputes the current promotion-content digest and rejects a mismatch before tarball verification, npm installation, Worker deployment, or service mutation; the stale local candidate was removed. Critical coverage now directly gates the shared protocol/subscription/schema modules, role-aware input inspection, modern controller/proxy/stream, HTTP contract, and candidate guard. The immutable beta.24 identity was not reused; the unreleased work is beta.25. No npm publication, tag, GitHub release, Worker deployment, service replacement, acceptance, commit, or push is part of this audit.
|
|
1051
|
+
|
|
1052
|
+
## 2026-07-28 version 3.0.0-beta.24 candidate-activation convergence audit
|
|
1053
|
+
|
|
1054
|
+
Owner-machine activation of the exact beta.23 candidate exposed a release-state defect that repository-local health checks could not prove. The same-name Worker reported the expected package version and normal health, but the candidate daemon received an HTTP authentication rejection before the Worker registered a WebSocket candidate. A forced redeployment using the unchanged local device identity then completed preflight, challenge authentication, readiness probing, and ordinary MCP delivery without rotating credentials. The supported conclusion is that version health and the locally recorded deployment fingerprint did not prove that the active edge deployment used the intended device-authentication material.
|
|
1055
|
+
|
|
1056
|
+
The failure also exposed unsafe compensation ordering. Activation had already advanced the Worker, then restarted the previous service definition after candidate authentication failed. Because Worker and daemon versions are an exact current-only contract, the restored older daemon could not reconnect and the system was left without a remote control path. A Windows task that exits with code zero but does not remain active is now a failed persistent start, not successful service evidence. Systemd unknown or maintenance states fail before mutation, while activating/reloading states retain restoration intent. Beta.24 separates recovery by commit boundary: before remote preparation, a provider carrying verified restoration intent may be restored; after remote preparation, cleanup installs and starts the compatible candidate service instead of reviving a known-incompatible daemon.
|
|
1057
|
+
|
|
1058
|
+
Candidate activation now treats explicit device-authentication rejection as positive convergence evidence distinct from ambiguous timeout, TLS, proxy, or health failure. It performs at most one same-name, same-identity forced deployment, never requests secret rotation, and bounds candidate startup to three attempts. Persistent failure is still returned to the operator. If remote preparation occurred, the error also states whether a compatible candidate service was started for automatic recovery; installation, start, lock-release, and runtime-stop failures remain aggregated.
|
|
1059
|
+
|
|
1060
|
+
Fault-injection tests reproduce the initial rejection, one repair deployment, successful second authentication, repeated rejection exhaustion, compatible-service forward recovery, service-stop refusal, ambiguous provider state, daemon-owner ambiguity, malformed version/wait/repair inputs, invalid attempt bounds, missing lock-release contracts, and cleanup failure. The activation module is now part of the critical coverage gate at 100% function and 80% branch coverage. No account identifiers, endpoint names, device keys, filesystem paths, command output, or credentials are added to tracked audit evidence or operational logs.
|
|
1061
|
+
|
|
1062
|
+
Remote foreground timeout validation was also tightened at the Worker execution boundary. When `timeout_seconds` is present it must be a numeric safe integer from 1 through 85; strings, fractions, zero, negative values, non-finite values, and larger values are rejected before a daemon `tool_call` with `side_effects_started=false`. This prevents direct JSON-RPC callers from bypassing the generated schema through permissive coercion.
|
|
1063
|
+
|
|
1064
|
+
The process-tree review found that an empty POSIX ownership snapshot still authorized forced escalation while the ChildProcess object had not reported exit. That was inconsistent with the documented fail-closed identity rule and could, under an extreme exit-event/PID-reuse race, signal an unrelated reused process group. Beta.24 now requires a non-empty captured ownership set and exact PID/start-time continuity before `SIGKILL`; failed inspection may leave a resistant descendant for operator cleanup. The child-settlement, process-ownership, and system-route helpers are included in strict checked-JavaScript contracts, and the child-settlement state machine has an independent 100% function/89% branch coverage gate.
|
|
1065
|
+
|
|
1066
|
+
A second full-plan reproduction showed that the first mitigation was incomplete: the detached runner could still remain in `running` beyond twenty seconds with no output. An isolated Node experiment reproduced the causal mechanism. `spawnSync` timeout defaults to `SIGTERM`; when the synchronous child ignores or fails to complete after that signal, Node continues waiting and the runner event loop cannot execute its own settlement timers. The same child returned at the declared boundary when the timeout used `SIGKILL`. Beta.24 therefore hard-bounds synchronous process-identity, process-tree, delegated-sandbox, macOS trust-broker, candidate-activation, prerelease-install, and synchronous verification helpers with `killSignal: "SIGKILL"`. A trust-broker child killed because its synchronous deadline expired is classified as `ETIMEDOUT` before signal-based code-signing diagnostics, so timeout is not misreported as an entitlement failure. Architecture checks reject any listed runtime probe whose timeout count is not matched by a hard-kill signal, and behavior tests execute an uncooperative child that ignores `SIGTERM`.
|
|
1067
|
+
|
|
1068
|
+
Candidate preparation also exposed a managed-job timeout race under repeated full-plan execution. A timed-out step and its descendants had exited, but macOS/libuv occasionally omitted the ChildProcess `close` event after `exit`; the runner therefore retained an unresolved step promise. Separately, process-tree ownership fallback could spend a full three seconds on the process table and another three seconds per captured member. Beta.24 gives each ownership decision one shared three-second monotonic budget and adds a one-second post-`exit` close fallback that preserves normal output-drain priority, destroys only residual handles, and settles through the existing terminal persistence path. Tests cover direct close, exit fallback, close/fallback race suppression, global timeout allocation, descendant death, terminal status, and detached runner exit.
|
|
1069
|
+
|
|
1070
|
+
Beta.23 is blocked and was not accepted, published, or promoted. Beta.24 requires a new exact candidate, owner activation, observed live Worker/service verification, and acceptance before any Git push or prerelease release.
|
|
1071
|
+
|
|
1072
|
+
## 2026-07-28 version 3.0.0-beta.23 workflow-closeout interruption and repository audit
|
|
1073
|
+
|
|
1074
|
+
The current incident was not a local daemon crash. During the reproduced failure, launchd retained the same daemon PID, process start time, and `runs=1`; the machine was awake with active user and Xcode assertions. The remote surface temporarily reported no connected daemon and recovered roughly one minute later. The operating-system default route was carried by Karing's TUN interface, normal Worker health succeeded through that route, and an interface-bound physical-network probe did not. The supported conclusion is therefore a transient relay path failure inside or beyond the system VPN/TUN boundary. Machine Bridge can detect and recover its socket, but cannot select or repair a third-party VPN node.
|
|
1075
|
+
|
|
1076
|
+
Beta.22 still contained an independent amplification defect. The shared local tool schema accepted foreground timeouts up to 600 seconds and commonly defaulted to 120, while the hosted Worker silently capped configurable foreground execution at 85 seconds plus five seconds for terminal delivery. That mismatch could let local mutation work finish before the Worker cancelled the call, producing exactly the misleading state “transactional patch applied, validation interrupted.” Beta.23 makes the remote contract truthful: Worker `tools/list` advertises an 85-second maximum while preserving 30- or 60-second tool defaults, and any larger request fails before daemon dispatch with a structured no-side-effect marker. Local/stdio operation keeps its broader schema because it does not traverse the hosted request boundary.
|
|
1077
|
+
|
|
1078
|
+
The audit added a privacy-bounded system-route diagnostic rather than logging raw networking state. On macOS it invokes only the fixed `/sbin/route -n get default` probe and projects the interface into a coarse route class and interception boolean. Interface names, IP addresses, DNS answers, Worker endpoints, proxy URLs, and credentials are omitted. Unsupported platforms and probe failures degrade to skipped diagnostic checks with coarse error classes.
|
|
1079
|
+
|
|
1080
|
+
The second repository review found one stale closeout sentence in `TESTING.md`; it still required beta.21 gates despite two later candidates. That version-specific assertion is replaced by a current-candidate rule. No new secret-bearing default logs, raw command/result logging, public administration route, unbounded request body, hidden compatibility execution path, or dependency-range relaxation was accepted. The first route-diagnostic draft exceeded the existing `runtime-diagnostics.mjs` line budget; the implementation was extracted into a dedicated boundary instead of weakening the architecture gate.
|
|
1081
|
+
|
|
1082
|
+
This source change does not prove third-party TUN reliability and does not deploy a Worker, activate the candidate, replace the running beta.22 daemon, rotate credentials, publish npm, push Git history, create a tag, or record live acceptance. Those remain explicit owner operations after complete repository verification.
|
|
1083
|
+
|
|
1084
|
+
## 2026-07-28 version 3.0.0-beta.22 ChatGPT interruption and full repository audit
|
|
1085
|
+
|
|
1086
|
+
The visible symptom combined three different mechanisms that must not be collapsed into one diagnosis. First, historical 11–34 minute relay gaps aligned with macOS clamshell sleep and DarkWake intervals; those records are expected machine suspension, not evidence of an active-use Cloudflare or daemon failure. Second, a controlled foreground command requested 120 seconds but the hosted tool path abandoned it at approximately 100 seconds. Third, the live Worker retained durable `exec_command` records after the local runtime had already completed and removed the corresponding operation. The latter two mechanisms explain active-use ChatGPT interruption and ghost pending-call growth.
|
|
1087
|
+
|
|
1088
|
+
The terminal-delivery protocol had a concrete loss window. The local runtime deleted a completed result as soon as `WebSocket.send()` accepted the frame, although that call proves only local queueing. If the connection ended before the Worker committed its generation-guarded terminal storage transaction, the runtime had neither an active operation nor a retained result to replay, while the Worker still owned the durable call. Beta.22 adds `tool_result_ack`: the runtime retains and periodically replays terminal results until acknowledgement; the Worker acknowledges only a transient settlement, a committed durable settlement, or an idempotent already-terminal duplicate. A result from a stale WebSocket generation is never acknowledged. Durable persistence failure now propagates instead of logging and falsely incrementing completion.
|
|
1089
|
+
|
|
1090
|
+
The timeout path also crossed responsibility boundaries. A single call timeout used a 45-second socket-silence heuristic to invalidate the complete daemon connection even though daemon liveness already has an independent 90-second alarm and readiness state machine. A slow command could therefore interrupt unrelated reads and make ChatGPT lose every local surface. Beta.22 removes socket invalidation from both transient and durable call expiry; timeout sends best-effort per-call cancellation only. The liveness alarm remains the sole owner of connection invalidation. Because the hosted foreground tool path was observed to stop near 100 seconds, configurable remote operations are capped at 85 seconds plus five seconds for terminal delivery. Longer work belongs in `start_process`/`read_process` or managed jobs, subject to what the MCP host exposes.
|
|
1091
|
+
|
|
1092
|
+
The broader review found a separate operations defect not named in the report: active daemon logs were safely trimmed before startup but could grow without another bound during a very long service lifetime. Beta.22 reuses the existing secure in-place trim every 15 minutes. The maintenance path preserves regular-file, no-follow, single-link, owner-only mode, schema, UTF-8, and line-boundary checks; failures log only a coarse class. Static review found no new secret-bearing operational logs, raw command/result logging, unbounded Worker body, public control-plane path, or obsolete transient settlement branch. Intentional migration fixtures remain confined to tests and explicit state upgrade code.
|
|
1093
|
+
|
|
1094
|
+
Fault-directed tests cover connected and reconnect result retention, lost acknowledgement heartbeat replay, Worker acknowledgement identity, persistent terminal-write failure, stale-generation non-acknowledgement, host-safe timeout bounds, log-maintenance scheduling/failure containment, and the previous cancellation/reconnect matrix. The repository architecture line budget rejected an oversized first draft of the recovery module; the implementation was reduced below the existing boundary rather than weakening the gate.
|
|
1095
|
+
|
|
1096
|
+
## 2026-07-27 version 3.0.0-beta.21 relay-continuity and repository audit
|
|
1097
|
+
|
|
1098
|
+
The reported symptom was a temporary loss of every local command surface, including `pwd`, followed by automatic daemon reconnection. Process evidence rejects the premise that the local daemon crashed: launchd retained one daemon PID, one start time, and `runs=1` throughout the observed interval. The failing layer was the relay/Worker connection and result-delivery path. Historical service logs also contain abnormal WebSocket closure and heartbeat-outage intervals that later recovered without a daemon restart. The exact network origin of each `1006` remains unknowable from those logs alone; a system VPN/TUN, proxy route, intermediary, or edge connection can all produce the same transport-level symptom.
|
|
1099
|
+
|
|
1100
|
+
A separate failure occurred during this audit: at 2026-07-27 05:24:33 UTC the local service emitted `remote relay protocol error`, exited with code 1, and launchd advanced from one run to two before reconnecting a replacement process. The retained log does not include the raw Worker error code, so that exact triggering frame cannot be reconstructed. Static and red-green analysis nevertheless found a concrete fatal-amplification family: `daemon_transport_error` and `daemon_liveness_timeout`, emitted by the Worker for transient socket failure, fell through the local `handleServerError` default and invoked `failPermanently`; a failed `hello` send was relabelled as authentication failure; and a readiness-probe result lost to an ending socket/session was relabelled as protocol violation. Beta.21 routes all three transport races through ordinary socket teardown and bounded reconnect, classifies error-frame and close-only forms consistently, uses WebSocket 1012 for transient Worker invalidation, and preserves permanent failure only for unknown/incompatible protocol, genuine authentication failure, malformed readiness sequencing, or identity/version mismatch.
|
|
1101
|
+
|
|
1102
|
+
The audit found two Machine Bridge amplification defects. First, the MCP initialize response declared `tools.listChanged=false`, but `tools/list` was derived from the momentary ready-daemon set. A brief relay interruption therefore withdrew every tool except `server_info` without the protocol notification required for a mutable catalog. Beta.21 makes discovery stable by package catalog and account role. This does not broaden authority: every call still requires a live end-to-end-ready daemon, daemon policy/tool permission, and account-role permission. `server_info.tool_delivery` now reports stable advertised and live effective scopes separately, and disconnected calls fail retryably with `unavailable` instead of mutating the catalog.
|
|
1103
|
+
|
|
1104
|
+
Second, streamed calls returned from their initiating Durable Object event but retained ownership in an in-memory pending registry backed by JavaScript timers. That model either prevents hibernation or loses ownership if the object is evicted; an alarm alone cannot recover an in-memory record. Beta.21 persists the call ID, daemon-process identity, random per-WebSocket connection generation, client request key, operation/reconnect deadline, and bounded result-transform metadata in the stream record before dispatch. Durable Object alarms and event-entry sweeps are the cross-event deadline owners. A restarted object can rediscover the call, the same verified daemon process can atomically rebind it, and a delayed close or result from the old socket fails the generation check. Success, rejection, cancellation, send failure, operation timeout, reconnect-grace expiry, and duplicate terminal delivery converge through one guarded terminal transaction. A FIFO admission gate enforces one 32-call ceiling across persistent and JSON-only calls. Active-record expiry advances over every new reconnect and remaining-operation budget, so repeated successful handovers cannot delete a still-owned call. JSON-only calls retain the prior bounded Promise/timer path.
|
|
1105
|
+
|
|
1106
|
+
The broader repository review refused line-budget exceptions and extracted outer Worker routing, pending-call persistence, record validation, stream-index maintenance, durable settlement, result projection, alarm storage, and socket-attachment decoding into focused modules. It removed the obsolete transient `registerEvent` settlement branch, added architecture rules forbidding its return, preserved Node strip-only TypeScript compatibility, and added fault-directed tests for restart recovery, corrupt records, fixed storage-write budgets, alarm expiry, stale generations, exactly-once settlement, stable discovery, and disconnected execution. Logging and privacy rules prohibit arguments, results, request keys, account identifiers, raw call/connection IDs, private paths, and subscriber payloads from operational diagnostics.
|
|
1107
|
+
|
|
1108
|
+
The workflow-level global verification deliberately repeated the complete project-native gate and exposed a separate real cleanup defect after two earlier green full runs. Under load, the POSIX process-tree ownership snapshot could miss a newly spawned descendant or the later full-table `ps` scan could fail; if the direct parent exited after `SIGTERM`, the conservative identity check then skipped `SIGKILL`, leaving an anti-`SIGTERM` descendant reparented to init. The first fix refreshed process-group ownership after graceful termination and performed targeted `ps -p` identity checks for captured PIDs. A later full-plan coverage run still reproduced the boundary on macOS because the initial full-table snapshot itself could time out and produce no ownership to revalidate. Darwin now uses `ps -g <PGID>` for capture and group revalidation, avoiding the unrelated system-wide process table while retaining exact PID/start-time/PGID identity. The self-test wait includes margin beyond the two-second grace plus three-second inspection budget. Repeated complete self-tests and coverage leave no descendant behind, while empty, changed, or reused identity still suppresses escalation.
|
|
1109
|
+
|
|
1110
|
+
A final exact-command preflight exposed a separate namespace regression that unit and package tests had not exercised: the POSIX `defaultStateRoot()` fallback had accidentally been changed to the new machine-service control directory. The standard owner command therefore installed the candidate under `machine-bridge-mcp-control` and failed before foreground ownership inspection because that directory already contained control artifacts. The fallback is restored to `~/.local/state/machine-bridge-mcp`; the control directory remains the sibling `machine-bridge-mcp-control`. Injected POSIX, XDG, and Windows path tests require the two roots to differ, and the exact owner wrapper was executed against the live beta.23 foreground daemon to prove it now installs under the profile state root, rejects before Worker/service mutation, identifies the old runtime, and emits the verified recovery sequence.
|
|
1111
|
+
|
|
1112
|
+
At that historical checkpoint, `npm run check:fast` passed 63 repository tasks; current plan sizes are authoritative only from `scripts/check-plan.mjs` and live check output. Full, dependency, package, Worker dry-run, privacy-history, and workflow-policy verification remain separate evidence and must pass before merge readiness is claimed. This source audit does not deploy a Worker, replace the running daemon, activate a candidate, rotate credentials, publish npm, push Git history, create a tag, or record live acceptance.
|
|
1113
|
+
|
|
1114
|
+
## 2026-07-26 version 3.0.0-beta.18 account-continuity and quota audit
|
|
1115
|
+
|
|
1116
|
+
The repeated hosted-client message `We couldn't connect your account. Please try again.` occurred while the Worker, Durable Object, daemon, Wrangler login, and end-to-end doctor probes were healthy. The remaining credential path had a destructive concurrency assumption: refresh tokens rotated once, and any immediate second use revoked the complete family—including the replacement credentials just returned to a concurrent request. A network retry or two refreshes racing through the same hosted connector could therefore convert a recoverable response race into total account loss. Beta.18 records a bounded source snapshot for 30 seconds and uses a deployment-keyed, domain-separated HMAC to reproduce the original replacement pair for at most two identity-equivalent retry responses. Retries neither create another credential branch nor extend expiration; the Worker returns a retryable 429 after that budget, and retains whole-family revocation for replay after the grace window. Client, resource, scope, account/version/role, deployment token version, and DPoP binding are revalidated on every retry.
|
|
1117
|
+
|
|
1118
|
+
The broader quota review found that beta.17 removed time-proportional stream polling but still forwarded public discovery and random paths to the Durable Object, rewrote liveness alarms on every heartbeat, exposed Preview URLs, and rethrew outer failures as platform exceptions. Beta.18 moves all stateless metadata and 404s to the outer Worker, adds a stateful-route rate-limit binding, coalesces alarm writes, exposes logical storage-write/alarm counters, retries terminal subscription transport failure within a hard bound, converts unexpected edge failures to a privacy-safe structured 502, and keeps the automatically provisioned `workers.dev` endpoint as the complete default deployment path.
|
|
1119
|
+
|
|
1120
|
+
A second fault-directed pass found issues outside the original incident scope. Bounded request readers retained at most the configured bytes but continued draining the rest of an oversized stream, allowing attacker-controlled duration work. Invalid methods on known stateful routes reached the rate limiter and Durable Object before returning 405. Several write-path and recursive-walk checks treated every `lstat`/`opendir` error as “missing,” so permission or I/O failure could be evaluated under a false path state. Repeated platform degradation also emitted one log per request, and newly split security modules were absent from the critical-coverage gate. The final implementation cancels body readers as soon as a declared or observed limit is crossed, rejects invalid methods before stateful dispatch, distinguishes only `ENOENT` from real filesystem failure, exposes partial application/skill discovery warnings, suppresses repeated edge logs with a reported duplicate count, removes an unreachable revoked-family branch and duplicate stream `waitUntil`, and establishes coverage floors for every new boundary module.
|
|
1121
|
+
|
|
1122
|
+
A third pass across local administration, sensitive-resource rollback, and the browser extension found additional boundary defects. The account administration client accepted an unbounded response and silently converted invalid successful JSON into an empty object. Newly generated SSH key files could remain unregistered if state persistence and best-effort cleanup both failed, because cleanup errors were discarded. The browser extension returned arbitrary local/page exception text and debugger fallback details to the remote caller, relied only on the broker for its concurrency ceiling, and the edge error classifier followed cyclic `cause` chains without a depth limit. Skill-symlink warnings also risked embedding absolute paths through native filesystem messages. These paths now use bounded strict JSON, complete two-file key rollback with compound failure, a fixed browser error boundary and fallback reason, an independent 32-request extension ceiling, bounded cycle-aware error traversal, and coarse path-projected discovery warnings.
|
|
1123
|
+
|
|
1124
|
+
Tests cover schema-2 migration, two idempotent concurrent refresh retries that reproduce the original access/refresh pair, retry-budget exhaustion without family revocation, post-grace whole-family revocation, DPoP proof replay and client/resource/scope isolation, static-route/unknown-path/invalid-method DO bypass, rate-limit denial and fail-open binding outage, nested/cyclic/deep quota classification, sanitized and throttled gateway logs, immediate oversized-body cancellation, bounded strict administration responses, complete and incomplete SSH rollback, fail-closed filesystem state inspection, privacy-safe partial discovery reporting, fixed browser-error projection and independent concurrency, fixed normal stream request/write budgets, single keepalive registration, subscription retry bounds, and alarm set/no-op/delete behavior. This source audit does not claim that an in-Worker rate limiter can enforce the account-wide Workers Free daily limit against distributed traffic; it also records the deployment-wide per-location bucket as a localized availability trade-off.
|
|
1125
|
+
|
|
1126
|
+
No live Worker deployment, daemon replacement, npm publication, tag, push, or acceptance is implied by this source audit. The exact beta.18 candidate must be owner-activated and observed before acceptance.
|
|
1127
|
+
|
|
1128
|
+
## 2026-07-25 version 3.0.0-beta.16 pending-call recovery and boundary audit
|
|
1129
|
+
|
|
1130
|
+
The reported incident exposed two separate failure domains. The host/connector layer returned `No shard mapper found` for a temporary high-replication backfill keyspace and then rejected even minimal MCP calls. The exact text did not exist in Machine Bridge source, deployed Worker events, daemon logs, or local command output; Worker HTTP server-error counters remained unchanged. The original shard-routing failure therefore occurred before or outside the deployed Worker/daemon boundary. The evidence is insufficient to identify the owner of that upstream temporary store, so this audit does not misattribute it to Cloudflare, the local daemon, OAuth, or Git. The host path later recovered without credential rotation or state deletion.
|
|
1131
|
+
|
|
1132
|
+
Independent inspection after recovery found a real Machine Bridge amplification defect: two calls remained active and detached beyond the shared two-minute reconnect grace. Pending operation and reconnect expiry used only in-memory `setTimeout` callbacks. Event-driven streamed calls intentionally return from the initiating Durable Object event, so a timer alone was not a valid cross-event lifecycle guarantee. The Worker now records monotonic operation/reconnect deadlines, exposes the earliest remaining duration to one combined Durable Object alarm, and performs an overdue scan at every HTTP and WebSocket event boundary. Timer callbacks remain the low-latency path. Alarm-storage failure emits only a bounded coarse event and does not falsely fail work that was already dispatched; the next event scan is the compensating path.
|
|
1133
|
+
|
|
1134
|
+
A second race existed during verified daemon replacement. The previous implementation rebound only already-detached calls, then closed the incumbent socket. Calls still attached to that socket were detached later by its asynchronous close callback, after the only rebind pass, and could remain orphaned. Same-instance promotion now transfers both attached and detached calls to the candidate before closing the incumbent, sends the complete authoritative `resume_calls` set, preserves the remaining normal timeout, and restores ownership to a still-open incumbent if replacement acknowledgement fails. Different daemon instances remain unable to inherit calls.
|
|
1135
|
+
|
|
1136
|
+
The broader review found a separate prototype-sensitive input defect in managed-job environment maps. POSIX/Windows-compatible variable validation permits names such as `__proto__`, `constructor`, `toString`, and `valueOf`, but validation accumulated them into ordinary JavaScript objects. Assigning `__proto__` invoked the legacy prototype setter rather than creating an own environment entry. Both plain and resource-backed environment maps now use null-prototype dictionaries, with stage/inspection regression coverage for all prototype-shaped keys and the existing duplicate-variable denial.
|
|
1137
|
+
|
|
1138
|
+
Documentation had also drifted: the remote lifecycle still claimed streamed work used Durable Object `waitUntil`, despite beta.15 explicitly removing and architecture tests forbidding that design. Architecture, operations, testing, changelog, and this audit now describe event-settled dispatch, alarm-plus-sweep deadline enforcement, atomic same-instance handover, upstream-host diagnosis, and prototype-safe environment maps. Complete and production dependency audits report zero vulnerabilities; privacy history, critical-module coverage, Worker dry-run, deterministic lifecycle tests, managed-job integration, type checks, lint, and real Wrangler OAuth/MCP integration pass before candidate preparation.
|
|
1139
|
+
|
|
1140
|
+
These source changes do not claim to repair the upstream shard mapper or prove its platform owner. They prevent a transient external failure or socket replacement from leaving permanent pending-call occupancy inside Machine Bridge. No Worker deployment, daemon/service replacement, global installation, credential rotation, push, tag, npm publication, or GitHub Release is performed by the source audit itself. Exact beta.16 candidate activation and owner-machine verification remain mandatory.
|
|
1141
|
+
|
|
1142
|
+
## 2026-07-25 version 3.0.0-beta.15 event-lifecycle audit
|
|
1143
|
+
|
|
1144
|
+
The exact beta.14 candidate was activated by the repository owner. The Worker, launchd daemon, private runtime path, activation record, candidate checksum, `status`, and `doctor` all converged on `3.0.0-beta.14`; the globally installed beta.12 package remained the explicit rollback baseline. A random temporary owner account and OAuth client then exercised the production protocol and were removed after the run. Sequence zero, token/session isolation, disconnect recovery, terminal sequence one, and empty acknowledgement replay all passed.
|
|
1145
|
+
|
|
1146
|
+
The acceptance-blocking beta.14 defect was narrower than beta.13 but still real. The outer Worker owned the public SSE and used short polls, yet the initiating Durable Object fetch created `dispatchJsonRpc(...).then(...)`, attached that unresolved terminal Promise to recovery state, and retained it through the event lifecycle. In production, later requests to the same object—including concurrent `server_info` and the authoritative `notifications/cancelled` request—remained queued until the daemon operation ended. Local Wrangler integration allowed an interleaving that production did not, so beta.14 has no acceptance and is not a valid publication candidate.
|
|
1147
|
+
|
|
1148
|
+
Beta.15 makes streamed daemon execution event-driven. The initiating fetch commits the pending recovery row, registers a pending call whose settlement is a callback rather than a Promise, sends `tool_call`, and immediately returns the internal stream descriptor. A later daemon WebSocket result event resolves or rejects that record and persists the terminal JSON-RPC envelope. Explicit cancellation, timeout, send failure, no-instance socket failure, and reconnect-grace expiry use the same settlement path. Same-instance reconnect only detaches and rebinds the record; it does not manufacture a terminal result. JSON-only clients continue through the ordinary Promise-based call path.
|
|
1149
|
+
|
|
1150
|
+
The recovery store no longer contains live terminal Promises. An active-stream set distinguishes work owned by the current Durable Object instance; a transient terminal map preserves an already produced message only when Durable Object storage fails. A pending stored record with no active owner is still converted to the bounded restart-ambiguity result. Static architecture checks reject reintroduction of a stream-initiation `dispatchJsonRpc` Promise, `resumption.attach`, Durable Object `waitUntil`, or a Promise-valued recovery map. Direct tests prove that registration returns without settlement and that later success, rejection, cancellation, timeout, send failure, transformation failure, persistence failure, and same-instance reconnect each close the stream exactly once.
|
|
1151
|
+
|
|
1152
|
+
This source correction does not accept beta.15, deploy it, replace the login daemon, push Git history, publish npm, create a tag, or create a GitHub Release. A new exact candidate and repeated owner-machine production test are mandatory.
|
|
1153
|
+
|
|
1154
|
+
## 2026-07-25 version 3.0.0-beta.14 production SSE-concurrency audit
|
|
1155
|
+
|
|
1156
|
+
The exact beta.13 tarball was activated through the owner command and converged successfully: the same-name Worker, one launchd daemon, private candidate runtime, service definition, status, and doctor all reported `3.0.0-beta.13`. A temporary random owner account and OAuth client were then used only for bounded live protocol verification and were deleted afterward. Sequence-zero delivery, token/session isolation, disconnect followed by authenticated `GET /mcp`, terminal sequence one, and empty acknowledgement replay all succeeded.
|
|
1157
|
+
|
|
1158
|
+
The same live run exposed a release-blocking production scheduling defect that the local Wrangler integration had not represented. While `BridgeRoom` directly returned an open SSE response, later requests routed to the same Durable Object—including `server_info` and the authoritative session-scoped `notifications/cancelled` notification—did not enter until the stream ended. The daemon call eventually terminated through its own boundary, but Worker observability did not record a successful cancellation. Beta.13 therefore has no acceptance record, is not pushed or published, and is explicitly blocked.
|
|
1159
|
+
|
|
1160
|
+
Beta.14 separates client transport ownership from durable state ownership. The outer stateless Worker creates the public SSE stream. `BridgeRoom` performs OAuth/DPoP authorization, signed MCP-session validation, stream admission, daemon dispatch, explicit cancellation, and terminal persistence, then returns a small internal descriptor. The outer Worker uses one service-binding WebSocket subscription after the authenticated descriptor: pending streams hibernate under `acceptWebSocket()`, terminal state is pushed once, and missing or expired state fails closed. No internal request remains open in the Durable Object for the life of a long tool call. Publicly supplied internal-control headers are stripped before every service-binding forward, so callers cannot select the unauthenticated internal subscribe path.
|
|
1161
|
+
|
|
1162
|
+
The real Wrangler regression keeps the original SSE response open, confirms a concurrent `server_info` sees the pending call, sends `notifications/cancelled`, observes the matching daemon `cancel_call`, and receives a cancelled terminal result on the original stream. Existing disconnect/recovery, wrong-session rejection, sequence-one acknowledgement, CORS, persistence faults, capacity, integrity, and oversized-message tests remain in force. This correction requires a new exact beta.14 candidate, owner activation, and repeated live verification; beta.13 activation evidence cannot be reused.
|
|
1163
|
+
|
|
1164
|
+
## 2026-07-25 version 3.0.0-beta.13 resumable-delivery and self-correction audit
|
|
1165
|
+
|
|
1166
|
+
The exercise began from clean released `main` at `ff3e2de` (`v3.0.0-beta.12`). Fast gates passed before modification, so the review concentrated on transport state, Durable Object persistence, protocol delivery, and live outage evidence rather than ordinary syntax defects.
|
|
1167
|
+
|
|
1168
|
+
Beta.12 correctly stopped treating an HTTP/SSE disconnect as cancellation, but that repaired only execution continuity. Its initial SSE frame had no event cursor, `/mcp` rejected GET, and a terminal response could not be retrieved after the response transport disappeared. Beta.13 implements the Streamable HTTP recovery split: POST commits a pending delivery record before dispatch and emits sequence zero; authenticated GET plus `Last-Event-ID` resumes only the original OAuth-token/MCP-session stream; sequence one is terminal and is not replayed again after acknowledgement. Explicit `notifications/cancelled` remains the sole client cancellation signal.
|
|
1169
|
+
|
|
1170
|
+
The first implementation pushed `BridgeRoom` beyond its architecture line budget. Rather than increasing the limit, authorization, recovery HTTP adaptation, lifecycle persistence, record encoding/integrity, and SSE framing were extracted into focused modules. A later review found that a prefix list would load every stored result value while admitting a new stream, potentially amplifying 64 retained 1.5-MiB results into a large allocation. The design was replaced with a small transactional metadata index and separate value records. Terminal messages carry SHA-256 integrity metadata, result retention is two minutes, capacity is 64, and oversized online results persist only a bounded replay error.
|
|
1171
|
+
|
|
1172
|
+
The review itself encountered a real beta.12 relay outage while a compound remote command attempted both a large source overwrite and tests. The daemon later reconnected with the same launchd PID and run count; the write had not completed, while earlier files remained intact. This exposed a flaw in the development method: a returned test result is not evidence that the preceding remote write persisted when the enclosing transport failed. Engineering rules now require independent anchor/Git verification after remote mutations. The same incident showed that warn-level service logs recorded an outage start but recovery only at info; recovery for an already-warned outage is now warn-level, while brief self-healing interruptions remain debug-only.
|
|
1173
|
+
|
|
1174
|
+
The mandatory candidate preflight then found a newly disclosed high-severity denial-of-service advisory in the development-only chain `eslint 10.7.0 -> minimatch 10.2.5 -> brace-expansion 5.0.7`. The compatible `brace-expansion 5.0.8` release satisfies the existing `^5.0.5` range, so the correction changes only the transitive lock entry and installed development tree. The candidate remains blocked until both complete and production-only registry audits return zero high-severity findings.
|
|
1175
|
+
|
|
1176
|
+
Dependabot PR #56 proposed Wrangler 4.114.0 but intentionally failed the repository gates because it did not update the exact `workerd` lifecycle-script approval and changed package bytes without a new prerelease acceptance. Beta.13 absorbs the update as one reviewed unit: Wrangler 4.114.0, Miniflare 4.20260722.0, workerd 1.20260722.1, and the corresponding platform binaries advance together; `sharp` remains forced to audited 0.35.3. An architecture assertion now derives the resolved workerd version from the lockfile and requires exactly one matching allowlist entry, so a future dependency-only PR cannot silently repeat this partial-update failure. The superseded Dependabot PR is closed only after this integrated change passes the complete local gate.
|
|
1177
|
+
|
|
1178
|
+
Fault-directed tests cover token/session isolation, live and persisted replay, restart ambiguity, persistence failure, capacity/expiry, oversized results, integrity tampering, SSE sequence semantics, CORS, and a real Wrangler disconnect/resume path. Static architecture checks prevent full-value scans, enforce module budgets and dependency direction, and require the storage/isolation anchors. This source audit does not deploy a Worker, replace the running beta.12 daemon, activate a candidate, publish npm, create a tag, or create a GitHub Release. Exact candidate preparation and owner-authorized activation remain separate release stages.
|
|
1179
|
+
|
|
1180
|
+
## 2026-07-23 version 3.0.0-beta.12 ChatGPT task-continuity audit
|
|
1181
|
+
|
|
1182
|
+
The reported symptom was not evidence that the model simply chose to stop. Live beta.11 inspection found a healthy single launchd daemon and matching Worker, but the Worker could retain an orphaned active call after the ChatGPT turn had already ended. The service log independently recorded repeated relay outages with the same daemon process still alive, including an incident where the existing thirty-second recovery deadline cancelled an active call. A routine source search also lost its result through the same path during this audit. These observations separate process health from request continuity: the daemon can keep running while either the ChatGPT-to-Worker HTTP request or the Worker-to-daemon WebSocket disappears.
|
|
1183
|
+
|
|
1184
|
+
The Worker contained a protocol defect at the first boundary. Every remote tool call was returned as one buffered JSON response only after local completion, leaving a long-running request with no response bytes. It also propagated the HTTP request abort signal into the pending-call registry and sent `cancel_call` when that signal fired. An intermediary reclaiming an idle HTTP request was therefore indistinguishable from an explicit user cancellation. This directly explains the visible behavior: the current turn ends without a terminal tool result, while a later “continue” message starts a new model turn that can resume the broader task from conversation state.
|
|
1185
|
+
|
|
1186
|
+
Beta.12 separates transport disposal from MCP cancellation. A client advertising `text/event-stream` receives an immediate SSE priming frame, ten-second keepalive comments, and one terminal JSON-RPC message. The Durable Object registers the completion with `waitUntil`; closing the response stream only stops writes and its heartbeat timer. The pending request remains indexed until a daemon result, timeout, reconnect-grace expiry, or a session-scoped `notifications/cancelled` request. JSON-only clients remain supported, but their HTTP disconnect likewise no longer cancels the underlying MCP request.
|
|
1187
|
+
|
|
1188
|
+
The second failure mechanism was a timing-contract mismatch. Worker detachment and local result retention each used independent thirty-second literals, while observed routed-network outages regularly exceeded that interval. Beta.12 moves heartbeat, reconnect, execution, overhead, and maximum-call values into one packaged shared contract. Same-instance recovery is two minutes. The Worker pauses the remaining operation deadline only while a record is detached, then resumes it after same-instance rebinding; healthy connected calls retain their original timeout. The local relay accepts the same bounded execution envelope. This is bounded continuation, not durable execution across daemon restart or machine failure. Managed jobs remain the durable mechanism for work that must survive those boundaries.
|
|
1189
|
+
|
|
1190
|
+
Regressions exercise content negotiation including `q=0`, immediate stream priming before the daemon replies, keepalive framing, terminal result projection, response-stream cancellation without operation cancellation, real HTTP abort with retained pending/request-key indexes, subsequent explicit cancellation targeting the original daemon call, Worker/local timeout-envelope parity and detached-deadline pause/resume, and the existing same-instance socket-rebind state machine. No beta.12 Worker deployment, daemon/service replacement, global installation, acceptance record, Git push, tag, npm publication, or GitHub Release is implied by this source correction.
|
|
1191
|
+
|
|
1192
|
+
## 2026-07-23 version 3.0.0-beta.11 external-review verification
|
|
1193
|
+
|
|
1194
|
+
The supplied read-only review was rechecked against source, package gates, platform plans, and the repository's explicit product/maintenance invariants rather than accepted as a backlog wholesale. Its central observability concern was correct: Worker field-name redaction was materially weaker than the local logger, so a future event could place a token, email address, credential URL, private-key header, or user-home path in an innocently named string field. The review did not notice a second defect in both structured loggers: sanitized caller fields were spread after core fields and could replace local `timestamp`, `level`, `component`, `message`, or `event` metadata and the corresponding Worker metadata, weakening incident chronology and event identity without executing code.
|
|
1195
|
+
|
|
1196
|
+
Beta.11 moves the portable free-form patterns into one shared module used by both runtimes. Local logging retains environment-derived home aliases and path-key suppression; Worker events gain content-aware value redaction without importing Node-only state. Local and Worker core metadata is assigned after caller fields and therefore remains authoritative. Tests cover sensitive-key redaction, sensitive content under a non-sensitive key, user-home aliases, email/token removal, control fields that attempt to forge event metadata, and unchanged safe route fields.
|
|
1197
|
+
|
|
1198
|
+
The line-cap warning was also directionally correct. `LocalRuntime` was 697 lines against a 700-line guardrail, but an arbitrary split would only hide coupling. Resource lookup, bounded binary/UTF-8 reads, and SSH resource registration form an independent lifecycle used by browser/application injection and the public generation tool. They now live in `runtime-resource-service.mjs`; `LocalRuntime` remains the composition root and falls to 653 lines. A direct boundary test and an independent architecture/package cap protect the extraction.
|
|
1199
|
+
|
|
1200
|
+
Several recommendations were rejected because their premises did not match the current tree or because they would reduce usability without a demonstrated failure mechanism. Linux already has a systemd-user service adapter and detached restart path. macOS and Windows run the complete platform plan, including service, process-lock, sandbox, application, managed-job, filesystem, and full-access behavior tests; Ubuntu alone carries the additional Worker/browser/package integrations. `full-access-test.mjs` is also the implementation behind the supported `machine-mcp full-test` command, not an unused test-only harness. Node 26/npm 12 is an explicitly documented single-baseline policy. Wrangler and the packaged release/deployment helpers support the zero-extra-install first-run and self-contained release state machine; splitting them would add setup and version-skew failure modes for the sole maintainer. Source visibility is not treated as a security boundary. The default canonical `full` profile and absence of per-operation prompts are deliberate usability decisions, so beta.11 makes them harder to misinterpret in diagnostics rather than inserting prompts or narrowing authority.
|
|
1201
|
+
|
|
1202
|
+
The review's suggestion to delete or rotate unreadable managed-job state was not adopted. An unreadable record deliberately blocks destructive state removal because automatically moving or deleting a directory that cannot be safely read can discard active recovery material or cross a replacement race. The bounded 50-job admission limit already prevents unbounded job-count growth; terminal readable records retain seven days, staged secrets retain twenty-four hours, and unreadable records require explicit operator inspection. A future quarantine feature would need token/snapshot-safe ownership, byte accounting, and removal semantics before it could improve rather than weaken this contract.
|
|
1203
|
+
|
|
1204
|
+
The dependency pass found only the next exact Wrangler patch. Updating to 4.113.0 advanced Miniflare/workerd to the 2026-07-21 build; npm initially blocked the new workerd postinstall because the repository intentionally permits scripts by exact package version. Beta.11 updates that allowlist to `workerd 1.20260721.1`, and clean-install/package/Worker dry-run gates must prove the new pin rather than relying on an existing binary. Registry audit, production-only audit, signatures, and attestations remain separate release evidence.
|
|
1205
|
+
|
|
1206
|
+
The second pass found no reason to reduce default authority, require a second reviewer, remove systemd support, broaden compatibility branches, or split the package into separately versioned runtime/deployment products. Remaining residuals are explicit: an authenticated remote owner is intentionally high authority; Cloudflare sampled logs are operational telemetry rather than an audit trail; unknown or transformed secrets can evade regex redaction, so arguments/results remain omitted; browser automation acts inside the real user profile; and OS CPU, memory, syscall, and egress isolation remain external deployment responsibilities. No Worker deployment, daemon/service replacement, global installation, credential rotation, Git push, tag, npm publication, or GitHub Release is implied by this source audit.
|
|
1207
|
+
|
|
1208
|
+
## 2026-07-22 version 3.0.0-beta.7 nested lifecycle PATH audit
|
|
1209
|
+
|
|
1210
|
+
Beta.6 passed its automated gates and exact live activation, but the launchd PATH still contained repository `node_modules/.bin` entries and npm's private `node-gyp-bin`. Source, tarball, and isolated runtime inspection proved that the intended sanitizer was packaged. A direct invocation over the installed plist PATH also produced the expected clean value. The remaining variable was the activation process environment.
|
|
1211
|
+
|
|
1212
|
+
A bounded live probe confirmed the topology: the beta.6 daemon inherited one npm run-script marker from beta.5 activation; starting another npm command prepended a second complete project-bin prefix and marker. The beta.6 algorithm selected the first marker, so it discarded only the outer activation layer and retained the inner inherited layer. This was a deterministic nested-lifecycle defect, not stale package bytes or a launchd rewrite.
|
|
1213
|
+
|
|
1214
|
+
Beta.7 selects the final npm run-script marker and inherits only entries after that boundary. Tests reproduce two nested Unix and Windows npm layers, then a stale candidate runtime and an ordinary user bin. Both npm layers and the stale runtime are rejected; current Node/package paths and the user bin remain. Beta.6 is blocked and has no acceptance record. No beta.7 publication or source release is implied by this correction.
|
|
1215
|
+
|
|
1216
|
+
## 2026-07-22 version 3.0.0-beta.6 service-environment reproducibility audit
|
|
1217
|
+
|
|
1218
|
+
The exact beta.5 candidate completed its full automated gates, isolated installation, same-name Worker deployment, persistent launchd handoff, project/Git verification, idempotent `service start`, and detached `service restart`. The live restart changed the daemon PID, incremented the launchd run count, exited the prior process cleanly, and restored one ready same-version daemon with no pending calls. Schema-4 service logs remained owner-only, single-link, and free of mixed text. These checks validated the beta.5 lifecycle corrections.
|
|
1219
|
+
|
|
1220
|
+
Inspection of the installed launchd definition then found a separate reproducibility defect. Candidate activation is invoked through an npm lifecycle from the source repository and was itself launched by the previous daemon. npm prepended a chain of repository/ancestor `node_modules/.bin` directories and its private `@npmcli/run-script/lib/node-gyp-bin`; the previous daemon contributed its beta.4 candidate bin as the first inherited PATH entry. The service installer accepted every absolute entry. Activation then pruned the beta.4 runtime, leaving a dead path in the persistent launchd definition, while unrelated source-repository and npm-private paths remained ahead of the operator's ordinary PATH. Internal Git/release decisions were already pinned to trusted absolute executables, so no observed authority bypass occurred, but the service environment depended on the mechanism used to invoke installation and no longer represented a reproducible operator environment.
|
|
1221
|
+
|
|
1222
|
+
Beta.6 blocks beta.5 and repairs the common service-path constructor rather than special-casing candidate activation. Node and the current package bin are inserted explicitly. If npm's run-script `node-gyp-bin` marker exists, every preceding lifecycle-injected entry and the marker itself are discarded. Entries in the candidate runtime container are rejected unless they are the current package bin. The remaining absolute inherited user PATH and platform defaults retain their order and are deduplicated. A user who intentionally supplies a `node_modules/.bin` outside an npm lifecycle keeps it; the rule distinguishes provenance using the npm marker instead of banning a path shape globally.
|
|
1223
|
+
|
|
1224
|
+
The regression reproduces the owner-machine topology and proves that Node, the current candidate, and the inherited user bin remain, while the npm project prefix, npm-private marker, and stale prior candidate are removed. Existing launchd/systemd quoting and absolute-only tests remain in force. No beta.6 Worker deployment, persistent service replacement, acceptance record, Git push, tag, npm publication, or GitHub Release is implied by this source correction.
|
|
1225
|
+
|
|
1226
|
+
## 2026-07-22 version 3.0.0-beta.5 relay, lifecycle, and recovery audit
|
|
1227
|
+
|
|
1228
|
+
Beta.4 successfully corrected the delegated-policy and fixed-Git defects found in beta.3, but it did not complete the release contract. During further owner-machine verification, the authenticated relay repeatedly disappeared and later recovered without a daemon PID change. The launchd daemon remained alive, retained the same process start time, and reauthenticated as the same installed beta.4 runtime. The evidence therefore excludes a local daemon crash for the observed incidents. The Worker route was resolved through the machine's system VPN/TUN and synthetic DNS mapping; direct `en0` probes to the public Cloudflare addresses were reset while the VPN path succeeded. The VPN service itself remained nominally connected, so the exact failing internal node cannot be proven from available logs. The supported conclusion is narrower: a system-level routed path can fail while the local process and VPN control plane remain alive, and the previous sixty-second reconnect cap unnecessarily prolonged recovery.
|
|
1229
|
+
|
|
1230
|
+
Beta.5 treats that distinction as an architectural contract. `network_route` now describes only Machine Bridge's application-level selection (`system-network-stack`, `application-http-proxy`, or invalid application proxy), and `network_route_scope` explicitly warns that an OS VPN/TUN may still intercept the first case. Relay status and `diagnose_runtime` retain bounded outage history, close category/code, transport error class, ready duration, and next-retry timing. Reconnect backoff is capped at fifteen seconds. Warn/info events are timestamped NDJSON and omit raw close reasons, proxy endpoints, tool arguments, commands, page data, and results. Heartbeat semantics remain conservative: the daemon sends every twenty-five seconds, requires inbound Worker traffic within seventy-five seconds, and the Worker reclaims a silent socket after ninety seconds.
|
|
1231
|
+
|
|
1232
|
+
The service review found that command semantics and ownership had been conflated. `service start` stopped and replaced an already running daemon, so invoking it through that daemon could destroy its own control path. Machine-global service control also accepted a workspace/state selector that did not necessarily own the installed service. Beta.5 makes start idempotent, isolates explicit restart, verifies the exact service daemon owner, and prevents unrelated state roots from stopping it. macOS/Linux use a detached service-manager handoff; Windows restart fails closed until a behaviorally verified helper boundary exists. Provider status is projected into a small structured record rather than returning launchd/systemd environment and path dumps.
|
|
1233
|
+
|
|
1234
|
+
The state and filesystem review found several variants of the same trust problem. Corrupt workspace state could be backed up and silently replaced on a later startup; an incomplete object could clear the recovery marker; and symlink-only checks did not prevent a workspace or state file from sharing an inode through a hard link. Beta.5 persists a validated recovery marker, requires a complete current-schema workspace/state envelope before unlocking, reads the marker once, and rejects multiple-link inodes at owner-only state, service logs, runner diagnostics, and existing-file path tools. The autostart log transition is strict: both active files are safely opened before mutation, schema 4 is committed only after reset/rotation succeeds, and a failed migration blocks startup rather than destroying incident evidence or mixing formats.
|
|
1235
|
+
|
|
1236
|
+
Long-lived execution received a failure-injection review rather than another happy-path test. Delayed SIGKILL now uses process-start/PGID membership snapshots so it neither misses surviving descendants nor targets a reused identity. Managed-job runners inherit only a minimal control environment unless the accepted plan explicitly requests full environment. Staged plans expire after twenty-four hours. Terminal result, terminal status, private runtime/plan cleanup, and cleanup confirmation are ordered; every write/delete failure is visible, and a valid terminal result can reconstruct status after a crash so finally steps are not repeated. Security-audit writes and legacy-state cleanup share a generic owner-state lock, and uninstall observes those locks plus job transition/recovery locks.
|
|
1237
|
+
|
|
1238
|
+
The browser and protocol review closed several same-machine amplification paths. The extension identity is fixed by the public manifest key and checked in both Origin and hello. Extension and runtime clients use different credentials. Replacing a socket clears its timer and requests, disables stale reconnect, and duplicate request IDs close the protocol. A failed hello, keepalive, result, error, Worker welcome, or Worker pong is treated as transport failure immediately. Mutable broker Sets/Maps are no longer exposed as public compatibility aliases; only bounded load counts are reported.
|
|
1239
|
+
|
|
1240
|
+
Release-control commands now use trusted absolute Git/GitHub CLI executables rather than npm lifecycle PATH. Unknown exceptions no longer export their raw message to a remote caller, and public structured details must be bounded JSON. Unauthorized MCP bodies are streamed to completion without retention before the 401 response, avoiding both buffering and response-after-stream races. Refresh-replay marker pressure revokes the affected family before marker eviction. Delegated macOS execution is enabled only after an allow/deny behavior matrix, not because `/usr/bin/sandbox-exec` can run `/usr/bin/true`.
|
|
1241
|
+
|
|
1242
|
+
The audit does not claim a distributed transaction across Cloudflare Worker deployment and the local service manager. Candidate activation verifies the foreground candidate before handoff and aggregates every local cleanup error, but an already changed Worker plus a failed service handoff may still require explicit operator rollback from the recorded previous runtime and deployment evidence. Likewise, system VPN/TUN node selection remains outside the repository. These residuals are now documented and observable rather than hidden behind labels such as `direct` or "rollback succeeded".
|
|
1243
|
+
|
|
1244
|
+
Beta.4 is blocked. No beta.5 package, Worker deployment, persistent service replacement, acceptance record, Git push, tag, npm publication, or GitHub Release is implied by this source audit. The complete repository gates, exact candidate activation, live same-version verification, and owner-machine acceptance remain mandatory.
|
|
1245
|
+
|
|
1246
|
+
## 2026-07-22 version 3.0.0-beta.4 authority, release, and low-friction hardening audit
|
|
1247
|
+
|
|
1248
|
+
|
|
1249
|
+
The first live candidate activation exposed a compatibility regression before any Worker or daemon handoff: the new shared public-JWK helper serialized members as `kty`, `crv`, `x`, `y`, while version 2 device identifiers used the RFC-style `crv`, `kty`, `x`, `y` order. Because JSON member order participates in the identifier hash, the intact version 2 key pair appeared to have an invalid identifier. Version `3.0.0-beta.2` restored the stable order, locked both local and Worker computations to the version 2 result, rejected the erroneous beta.1 ordering, and marked beta.1 as blocked.
|
|
1250
|
+
|
|
1251
|
+
The second live activation reached the next boundary and failed before Worker deployment or daemon handoff: the runtime-compiled helper was ad-hoc signed and therefore had no provisioning-profile-validated access group for the macOS data-protection Keychain. Persistent Secure Enclave key generation returned `errSecMissingEntitlement` (`-34018`). Version `3.0.0-beta.3` marks beta.2 as blocked, stops automatic Secure Enclave migration, keeps the portable root by default, and requires an explicitly configured provisioned app-like broker plus a real create/delete key probe before enrollment. Both failures validate the prerelease process: neither candidate was published or promoted as stable, and each blocking fix increments the beta version and restarts candidate verification.
|
|
1252
|
+
|
|
1253
|
+
The third live activation completed the exact beta.3 Worker and persistent launchd handoff, survived a controlled service restart, and completed ordinary owner read/write/process calls. The same live review then exposed a delegated-account boundary defect: local `project_overview` returned an already-intersected request policy while the Worker interpreted that field as the daemon ceiling, and fixed read-only Git metadata commands reused the arbitrary-process path, causing reviewer/editor metadata to fail whenever the delegated sandbox was unavailable. Version `3.0.0-beta.4` blocks beta.3, returns explicit effective and daemon fields, and moves implementation-owned Git probes to a validated no-shell minimal-environment internal process path while leaving caller-selected process execution sandbox-gated.
|
|
1254
|
+
|
|
1255
|
+
Repeated relay outages during beta.4 verification were traced to the package installation smoke test rather than the network or Cloudflare. Its isolated zero-argument startup reached foreground preparation, which unconditionally stopped the machine-global service label before the fake Wrangler boundary. The test therefore sent `SIGTERM` to the real beta.3 launchd daemon whenever the full plan reached `install:test`. Foreground startup now checks the exact state/workspace daemon lock and may stop the platform service only when that lock identifies a live, verified `service` process. The smoke test installs a service-manager trap and fails if isolated startup attempts any machine-level service control. A real post-fix installation smoke run preserved the same daemon PID, relay `connected_at`, and readiness state.
|
|
1256
|
+
|
|
1257
|
+
The same live review found an observability contract defect. The launchd definition correctly selected `--log-format json`, but only the structured `event()` path honored it; direct logger methods and the daemon-ready summary still emitted unstructured text without timestamps. Beta.4 makes JSON format authoritative across all logger methods and emits daemon readiness as a single redacted JSON lifecycle event, preserving newline-delimited ingestion and incident-timeline accuracy.
|
|
1258
|
+
|
|
1259
|
+
Immediately before candidate preparation, the registry audit database began reporting GHSA-f88m-g3jw-g9cj against Miniflare’s exact `sharp@0.34.5` dependency. The current and newly released Wrangler lines still selected that vulnerable version, so a Wrangler-only upgrade was insufficient. Beta.4 applies an exact npm override to patched `sharp@0.35.3`, updates the explicit lifecycle-script allowlist, verifies the resolved dependency tree, and requires complete Worker/Miniflare integration plus zero-vulnerability, registry-signature, and provenance-attestation checks before candidate generation.
|
|
1260
|
+
|
|
1261
|
+
The release-process review found a separate governance defect: the repository used the final semantic version while the package was still an unpublished local candidate, and successful live verification could flow directly into a stable tag and `latest` publication. That process had no persistent beta runtime, no registry-verified observation period, no defect-reset rule, and no machine-enforced proof that stable bytes matched the tested prerelease. Version 3 moved the package to the `3.0.0-beta` channel; at that stage the candidate under review was `3.0.0-beta.4`. The release process now has explicit `dev`, `beta`, `rc`, and stable channels. Candidate activation is now one owner-executed command that installs the exact tarball under owner-only state, updates the same-name Worker, proves candidate relay readiness, installs the login service, performs a lock-controlled handoff, verifies the exact background daemon and Worker versions, and exits while the service remains active.
|
|
1262
|
+
|
|
1263
|
+
Formal soak begins only after the exact accepted package is published under the correct npm dist-tag, marked as a GitHub Prerelease, installed from the registry, and reactivated. Major, minor, and patch releases require minimum seven-day, three-day, and one-day intervals. A blocking issue creates a new prerelease and restarts the clock. Stable promotion uses a digest over the npm file inventory with only synchronized version metadata normalized; any runtime, dependency, script, policy, extension, Worker, or packaged-documentation change forces another prerelease. Guarded push, GitHub release creation, npm publication, CI portable acceptance, and stable promotion all enforce these records.
|
|
1264
|
+
|
|
1265
|
+
The version 2 transaction-lease design solved several effect-classification gaps but retained the wrong privilege primitive for delegated accounts: a terminal lease could expand effective behavior after the account role had already been selected. It also required copy-an-ID, approve, and retry interaction, which increased prompt fatigue and encouraged broad temporary grants. Version 3 replaces that model with a request-scoped intersection of daemon capability, immutable account role, trusted OAuth client, account version, refresh-token family, operation invariants, and object ownership. Risk classification remains, but it now supplies hard enforcement and privacy-preserving audit metadata rather than temporary elevation. Legacy lease files remain readable only for migration cleanup and are not consumed by the runtime.
|
|
1266
|
+
|
|
1267
|
+
The review found cross-principal object gaps beyond initial tool dispatch. Interactive processes, retained one-shot output, and managed jobs could outlive the request that created them and were not uniformly bound to the originating account/client. Version 3 binds those objects to account ID, account version, OAuth client ID, and refresh-family ID. Another account, another client for the same account, or a newly issued family cannot read, continue, send input to, cancel, or terminate the object. The same principal is carried from Worker token verification through the relay envelope and local authority context.
|
|
1268
|
+
|
|
1269
|
+
Generic remote `full` file and process authority was also capable of reaching Machine Bridge's own state. Version 3 identifies protected control-plane roots and rejects generic path-based file access even for `owner`. Device-root material and profile state are no longer exposed through ordinary file APIs. This is defense against accidental/tool-level export, not a shell sandbox: an authorized owner interpreter still has the OS user authority and can reach same-user state unless an external OS boundary is used.
|
|
1270
|
+
|
|
1271
|
+
The long-lived account-administration HMAC was replaced by the same root-trust hierarchy used for daemon authentication. The default device root on every platform, including macOS, is the existing owner-only portable P-256 key. At daemon startup it signs a 24-hour ephemeral session certificate without Keychain access or user presence; the private session key remains in memory and signs WebSocket preflight, Worker challenge, reconnects, and account-administration requests. The Worker stores only the root public JWK. Account administration no longer deploys or stores `ACCOUNT_ADMIN_SECRET`; each request binds origin, method, path, body hash, session key ID, timestamp, and nonce to the certified session signature.
|
|
1272
|
+
|
|
1273
|
+
A non-exportable macOS Secure Enclave root is an optional deployment enhancement, not an npm runtime assumption. `MBM_MACOS_TRUST_BROKER` must name an app-like broker with a strict Apple code signature, stable Team ID, provisioning-profile-validated data-protection Keychain access, canonical non-symlink executable path with no group/other write access, and successful temporary create/delete key probe. The root state binds that broker identity and every later use revalidates it. The packaged Swift source and ad-hoc build remain development/protocol fixtures and are intentionally rejected by the production validator.
|
|
1274
|
+
|
|
1275
|
+
When a provisioned broker is configured, root migration and rotation remain two-phase. A candidate root is persisted as pending, its public key is deployed, and Worker health/version convergence is verified before local promotion. Failed probe, upload, interrupted startup, or failed health leaves the old active root intact. `--daemon-only` refuses to activate an undeployed pending root.
|
|
1276
|
+
|
|
1277
|
+
OAuth clients are now persistent trust objects rather than reusable registration shells. First authorization binds one client record to one account, account version, and role. Client revocation independently removes its codes and token records. Refresh rotation remains single-use and family-replay revocation remains fail closed. Optional DPoP ES256 binds supported client token families to a client key while Bearer remains for interoperability. The DPoP pass found that cryptographically valid proofs from arbitrary self-generated keys were consuming the global replay cache before OAuth credential validation, allowing unauthenticated cache exhaustion. Proof verification is now pure; the replay marker is consumed only after access-token or grant/client/account/resource validation succeeds. Unsupported critical JWS semantics are rejected.
|
|
1278
|
+
|
|
1279
|
+
Delegated direct execution received an explicit negative-boundary review. Restricting cwd, HOME, and environment does not prevent an ordinary same-user process from reading state, Keychain, or unrelated home files. Version 3 therefore requires a behaviorally verified OS sandbox. On the current macOS 27 beta host, deny-default `sandbox-exec` policies abort, so delegated process execution fails closed instead of using a misleading blacklist. Owner execution is unchanged. Hard mutually untrusted execution still requires a dedicated account, container, or VM.
|
|
1280
|
+
|
|
1281
|
+
The local security audit is now a bounded SHA-256 hash chain. Entries contain tool identity, risk class, result, timing, byte counts, target digest, and salted principal references. Commands, paths, file contents, form values, and output are excluded. Nonce stores no longer evict unexpired replay markers when full; they reject new entries until capacity is available. DPoP replay capacity is reserved for already authenticated/validated requests rather than arbitrary public proofs.
|
|
1282
|
+
|
|
1283
|
+
The normal interaction target is no prompt for ordinary tool calls, WebSocket reconnect, or portable-root daemon startup. A provisioned Secure Enclave deployment requests one user-presence operation when the root signs a daemon session certificate and may request one for an independent account-administration command or explicit root rotation. Terminal operation approval and `job approve` paths are removed. `stage_job` is a non-executing draft; execution requires owner-authorized `start_job` or explicit local `job submit`.
|
|
1284
|
+
|
|
1285
|
+
Regression evidence includes hard role ceilings under a canonical-full daemon; explicit effective versus daemon-ceiling project snapshots; reviewer/editor fixed Git metadata without arbitrary-process authority; minimal-environment internal process dispatch; cross-account/client/family process and job denial; control-plane path denial; root-certified session preflight/challenge tamper, expiry, wrong-root, and replay tests; signed account administration without a symmetric secret; trusted-client binding and independent revocation; DPoP proof/token binding, post-authorization replay consumption, invalid-grant non-consumption, and critical-header rejection; nonce saturation; audit-chain tamper detection; delegated sandbox behavior; macOS development-broker compilation, production rejection of ad-hoc signing, provisioned-broker path/signature/Team-ID binding, and create/delete probe cleanup; two-phase root deployment state; package/install/stdio/Worker integration; and critical-module coverage gates.
|
|
1286
|
+
|
|
1287
|
+
Residual limits remain explicit. A compromised active owner client can exercise the daemon ceiling without per-operation prompts. Bearer clients remain vulnerable to token possession until expiry or revocation. Portable roots are exportable. A same-user process can inspect daemon memory or interfere with local files and services. Owner execution has no universal CPU, memory, syscall, or network isolation. Public binary-store distribution, external audit anchoring, independent review, and protected publication identities remain optional external hardening rather than release blockers for the self-hosted package.
|
|
1288
|
+
|
|
1289
|
+
Live beta.3 Worker deployment, persistent daemon replacement, ordinary owner-tool verification, and one controlled launchd restart were performed on the owner machine. No account or device-root rotation, Git push, tag, npm publication, GitHub Release, or stable promotion was performed.
|
|
1290
|
+
|
|
3
1291
|
## 2026-07-21 version 2.0.0 remote-authority audit
|
|
4
1292
|
|
|
5
1293
|
The critical question was whether canonical `full` could retain its automation value without allowing one long-lived remote bearer to remain equivalent to unattended local-user control. The pre-2.0 implementation correctly intersected account role with the daemon capability ceiling, but the final authorization decision was still tool-granular: once `exec_command`, unrestricted paths, browser tools, or managed jobs were exposed, the original arguments reached the handler without a local effect-level transaction boundary. The daemon WebSocket also authenticated with one long-lived shared bearer, access tokens lasted thirty days, and account administration sent its long-lived secret directly as a network bearer.
|
|
@@ -16,7 +1304,7 @@ The review found and corrected several additional boundary gaps. Protecting only
|
|
|
16
1304
|
|
|
17
1305
|
Regression evidence includes P-256 preflight/challenge generation and tamper/wrong-key/expiry rejection; real Worker candidate and readiness integration; fifteen-minute access tokens and whole-family revocation after refresh replay; signed administration success, legacy bearer rejection, nonce replay rejection, and malformed-state rejection; account/client lease isolation, automatic workspace behavior, scoped/full windows, compound browser/application authority, external-sensitive path composition, sensitive writes, process/job continuation, symbolic-link and patch-move targets, cross-process mutation serialization, owner-only state, and malformed-record rejection. The complete repository gates remain authoritative before release.
|
|
18
1306
|
|
|
19
|
-
|
|
1307
|
+
At the time of version 2.0, residual limits were explicit. Device and administration private material still lived in owner-only cross-platform state rather than macOS Secure Enclave/Keychain; a determined same-user process can read or alter that state and can interfere with the daemon. `full` still passes the complete parent environment by contract once process authority is locally leased. Machine Bridge still does not provide kernel CPU/memory quotas, syscall filtering, or network egress isolation. DPoP was still future work because hosted MCP clients did not uniformly support it. Logs remain privacy-oriented operational telemetry, not a remote tamper-proof forensic ledger. Signed/notarized root-owned distribution, external audit anchoring, and hard sandboxing remain separate distribution/deployment work rather than being simulated in application code.
|
|
20
1308
|
|
|
21
1309
|
No Worker deployment, daemon/service replacement, global installation, credential rotation in live state, push, tag, npm publication, or GitHub Release is performed by these source changes.
|
|
22
1310
|
|
|
@@ -46,7 +1334,7 @@ Regression evidence covers three layers: deterministic pending-registry detach/r
|
|
|
46
1334
|
|
|
47
1335
|
The release-integrity mechanism correctly bound acceptance to exact package bytes, but the operational role split was wrong. Earlier documentation required the owner to execute candidate and approval commands in a terminal. The explicit owner decision is now that authorization is given in the active conversation; the coding agent starts the exact candidate through Machine Bridge, verifies its version, readiness, deployment identity, and representative functionality, then records acceptance and completes source release operations. This removes terminal authorization as a user-facing dependency without allowing automated tests alone to manufacture acceptance.
|
|
48
1336
|
|
|
49
|
-
Version 1.2.9 codifies that flow. `npm run release:candidate:start -- --allow-worker-deploy` verifies and installs the pending tarball into an ignored isolated prefix and launches it in the foreground without replacing the normal global installation. The explicit flag acknowledges that normal candidate startup may update the configured same-name Worker in place. After explicit owner authorization in the active conversation, the coding agent runs that command through Machine Bridge and must observe the live candidate through Machine Bridge; automated checks or an unobserved process are insufficient. After successful observation, the agent records the new `owner-
|
|
1337
|
+
Version 1.2.9 codifies that flow. `npm run release:candidate:start -- --allow-worker-deploy` verifies and installs the pending tarball into an ignored isolated prefix and launches it in the foreground without replacing the normal global installation. The explicit flag acknowledges that normal candidate startup may update the configured same-name Worker in place. After explicit owner authorization in the active conversation, the coding agent runs that command through Machine Bridge and must observe the live candidate through Machine Bridge; automated checks or an unobserved process are insufficient. After successful observation, the agent records the new `owner-started-agent-verified-local-candidate` marker, commits, pushes through the guarded command, completes the pull request, and runs `npm run release` to create or verify the annotated tag and GitHub Release. The old 1.2.8 owner-recorded marker remains accepted only for historical compatibility.
|
|
50
1338
|
|
|
51
1339
|
npm publication and Worker deployment remain owner-operated live release actions. The coding agent does not perform either as part of `npm run release`. Any packaged-byte change after candidate preparation invalidates the pending or recorded acceptance and requires another owner-authorized, agent-operated and observed candidate run.
|
|
52
1340
|
|
|
@@ -72,7 +1360,7 @@ Version 1.2.7 extracts `process-contract.mjs`, `process-tree.mjs`, and `executio
|
|
|
72
1360
|
|
|
73
1361
|
A separate lifecycle race existed in the browser broker. `stop()` could run while pairing/listen/proxy startup was awaiting I/O; the old start could then finish and repopulate a listener after shutdown. Startup now carries a monotonically increasing generation through every asynchronous boundary. Stop invalidates that generation, and stale completion closes all HTTP/WebSocket transports, pending calls, and routed proxy requests. Browser status also no longer relies on global `fetch` for loopback health: a strict bounded direct-HTTP adapter accepts only canonical `127.0.0.1` and cannot be diverted by environment proxy configuration. Remote Worker health and relay traffic continue to use the documented proxy resolver.
|
|
74
1362
|
|
|
75
|
-
The resource review distinguished application bounds from kernel controls. The implementation has real concurrency, timeout, stdin, output, retention, and job-size limits, but it does not create CPU quotas, resident-memory ceilings, syscall sandboxes, or egress filters. `server_info.runtime.execution_guardrails` now
|
|
1363
|
+
The resource review distinguished application bounds from kernel controls. The implementation has real concurrency, timeout, stdin, output, retention, and job-size limits, but it does not create CPU quotas, resident-memory ceilings, syscall sandboxes, or egress filters. Remote `diagnose_runtime.runtime.execution_guardrails` and local stdio `server_info.runtime.execution_guardrails` now state both categories explicitly so a timeout cannot be misread as a CPU quota and a proxy route cannot be misread as network isolation. Hard isolation remains an external deployment responsibility: dedicated low-privilege account, container, or VM.
|
|
76
1364
|
|
|
77
1365
|
Managed-job launch previously assumed `spawn()` returned a usable child. A system-level launch failure could emit an unhandled child `error` after the job had been reported accepted. Launch now attaches the error observer before returning, refuses acceptance without a positive PID, records a coarse structured failure class, and keeps shell parsing disabled explicitly. The audit also corrected stale `hello_ack`-only readiness wording, incorrect `server_info` field paths, a duplicated plan-retention guard, and documentation that still described daemon takeover as non-escalating.
|
|
78
1366
|
|
|
@@ -217,7 +1505,7 @@ A dormant recovery-safety helper exposed a concrete lifecycle defect: after reco
|
|
|
217
1505
|
|
|
218
1506
|
The process boundary now distinguishes fixed executable plus argv invocation from explicit shell execution. Direct calls reject NULs and oversized executable/argv input and use `spawn(..., { shell: false })`; `exec_command` remains an explicit canonical-full shell capability and is documented as local-user authority. CodeQL exceptions for these intentional boundaries are exact rule/path records with rationales and expiry dates. The CodeQL workflow now parses generated SARIF and fails on every other security-tagged result, preventing a successful analysis upload from masquerading as a clean security gate.
|
|
219
1507
|
|
|
220
|
-
|
|
1508
|
+
For the historical version 1.0 release described in this section, the public contract was current-only to that release: MCP `2025-11-25` was the sole advertised protocol. Existing then-current local state remained usable, so upgrade safety was provided by bounded daemon/Worker version convergence and exact browser-extension version handshakes rather than by retaining still older protocol or lock interpretations. This statement is historical; beta.61 uses MCP `2026-07-28` as its only native state model while remote HTTP exposes only the bounded stateless initialization compatibility described in the current 2026-08-11 audit. Randomized protocol, policy, argv, and real no-shell-evaluation tests supplement deterministic state-machine and cross-platform integration coverage.
|
|
221
1509
|
|
|
222
1510
|
A follow-up control-plane and credential-lifecycle review found four issues outside the original concurrency report. The repository already prohibited hosted GitHub connectors, but that rule was available only after project context loading; shared initialization and built-in instructions now state the fail-closed local `git`/`gh` rule before any GitHub operation. `ENGINEERING.md` and `CONTRIBUTING.md` also contradicted the newer source-release contract by assigning tags and GitHub Releases to the human operator; all normative files now distinguish automated GitHub source completion from explicitly authorized npm and live-machine operations.
|
|
223
1511
|
|
|
@@ -382,3 +1670,13 @@ No review can prove the absence of all defects. Important residual boundaries ar
|
|
|
382
1670
|
## Ongoing review rule
|
|
383
1671
|
|
|
384
1672
|
A future change to locks, state deletion, service lifecycle, detached processes, credentials, browser/app authority, package contents, or public transport must add a behavior-level regression test and update the applicable security/operations documentation. String-presence tests may supplement, but must not replace, executable state-transition tests.
|
|
1673
|
+
|
|
1674
|
+
### 2026-08-06 Durable Objects write-amplification incident
|
|
1675
|
+
|
|
1676
|
+
Live beta.44 observability measured 264 estimated stream rows for 40 calls in one isolate (about 6.6 rows per call before all acknowledgement cleanup). The causal mechanism was a global `mcp-stream-index` row rewritten alongside every per-stream transition. Beta.45 makes per-stream records the sole authority, deletes the legacy index once, meters only committed mutations, and enforces three-row immediate and four-row daemon lifecycle ceilings. The beta.44 acceptance is invalid because the quota repair changes Worker and package bytes; only a newly activated and observed beta.45 candidate may be accepted.
|
|
1677
|
+
|
|
1678
|
+
### 2026-08-07 beta.45 external CI gate incident
|
|
1679
|
+
|
|
1680
|
+
The accepted beta.45 commit passed the exact local 108-task full plan, package/install/SBOM/privacy checks, live Worker and daemon verification, and guarded push. Its first pull-request run nevertheless supplied two independent external counterexamples. On Windows, Git checkout materialized workflow YAML with CRLF because the repository had no workflow-specific `eol=lf` attribute; `workflow-policy:test` correctly rejected the changed bytes. In CodeQL 2.26.2, seven release error sinks were classified as log injection because a sanitized but externally derived diagnostic was interpolated into a log string, and the macOS trust-broker tamper regression was classified as a filesystem race because it reopened a path after the builder had inspected it. These findings are not accepted as exceptions.
|
|
1681
|
+
|
|
1682
|
+
Beta.46 preserves the strict LF-only workflow policy and fixes checkout instead, emits release failures as bounded JSON records after portable redaction, and mutates the test binary through a no-follow validated descriptor. The beta.45 acceptance is invalidated because release scripts are packaged. Local simulation can prove Git attribute resolution, single-line JSON encoding, no-follow descriptor behavior, and project regressions; only the subsequent GitHub Windows and CodeQL runs can close the platform/environment oracle.
|