@bjoernrochel/sevro 0.1.0-rc.8 → 1.1.0
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/CONTRIBUTING.md +52 -6
- package/README.md +12 -7
- package/dist/examples/basic/graded.json +13 -0
- package/dist/examples/basic/host.js +9 -0
- package/dist/examples/basic/host.js.map +1 -0
- package/dist/examples/basic/prompt-only.json +7 -0
- package/dist/examples/recipes/candidate.js +23 -0
- package/dist/examples/recipes/candidate.js.map +1 -0
- package/dist/examples/recipes/cases/coding-change.json +21 -0
- package/dist/examples/recipes/cases/host-tools.json +21 -0
- package/dist/examples/recipes/cases/mise.json +21 -0
- package/dist/examples/recipes/cases/node.json +21 -0
- package/dist/examples/recipes/cases/python-uv.json +21 -0
- package/dist/examples/recipes/cases/semantic.json +17 -0
- package/dist/examples/recipes/coding-change/check.js +16 -0
- package/dist/examples/recipes/coding-change/check.js.map +1 -0
- package/dist/examples/recipes/executable.js +20 -0
- package/dist/examples/recipes/executable.js.map +1 -0
- package/dist/examples/recipes/extension-artifacts.js +56 -0
- package/dist/examples/recipes/extension-artifacts.js.map +1 -0
- package/dist/examples/recipes/extension-methods.js +167 -0
- package/dist/examples/recipes/extension-methods.js.map +1 -0
- package/dist/examples/recipes/extension.js +28 -0
- package/dist/examples/recipes/extension.js.map +1 -0
- package/dist/examples/recipes/modules.js +9 -0
- package/dist/examples/recipes/modules.js.map +1 -0
- package/dist/examples/recipes/repository-skill/.agents/skills/startup-guide/SKILL.md +7 -0
- package/dist/examples/recipes/repository-skill/.claude/skills/startup-guide/SKILL.md +7 -0
- package/dist/examples/recipes/repository-skill/README.md +3 -0
- package/dist/examples/recipes/run-inputs.js +144 -0
- package/dist/examples/recipes/run-inputs.js.map +1 -0
- package/dist/examples/recipes/run-options.js +103 -0
- package/dist/examples/recipes/run-options.js.map +1 -0
- package/dist/examples/recipes/run.js +62 -0
- package/dist/examples/recipes/run.js.map +1 -0
- package/dist/examples/recipes/runtime/empty.json +1 -0
- package/dist/examples/recipes/runtime/goals.json +1 -0
- package/dist/examples/recipes/runtime/mise.json +6 -0
- package/dist/examples/recipes/runtime/path.json +1 -0
- package/dist/examples/recipes/runtime/plugins.json +4 -0
- package/dist/examples/recipes/runtime/python-uv.json +10 -0
- package/dist/examples/recipes/semantic-judge.js +27 -0
- package/dist/examples/recipes/semantic-judge.js.map +1 -0
- package/dist/examples/templates/coding/adapter.js +14 -0
- package/dist/examples/templates/coding/adapter.js.map +1 -0
- package/dist/examples/templates/coding/case.json +15 -0
- package/dist/examples/templates/repository-skill/.agents/skills/startup-guide/SKILL.md +6 -0
- package/dist/examples/templates/repository-skill/extension-data.js +61 -0
- package/dist/examples/templates/repository-skill/extension-data.js.map +1 -0
- package/dist/examples/templates/repository-skill/extension.js +29 -0
- package/dist/examples/templates/repository-skill/extension.js.map +1 -0
- package/dist/examples/templates/response/adapter.js +13 -0
- package/dist/examples/templates/response/adapter.js.map +1 -0
- package/dist/examples/templates/response/case.json +13 -0
- package/dist/src/advisory.js +86 -0
- package/dist/src/advisory.js.map +1 -0
- package/dist/src/candidate-artifacts.js +139 -0
- package/dist/src/candidate-artifacts.js.map +1 -0
- package/dist/src/candidate-codex-state.js +110 -0
- package/dist/src/candidate-codex-state.js.map +1 -0
- package/dist/src/case-validation.js +92 -0
- package/dist/src/case-validation.js.map +1 -0
- package/dist/src/cli/cancellation.js +21 -0
- package/dist/src/cli/cancellation.js.map +1 -0
- package/dist/src/cli/command.js +60 -0
- package/dist/src/cli/command.js.map +1 -0
- package/dist/src/cli/completion.js +14 -0
- package/dist/src/cli/completion.js.map +1 -0
- package/dist/src/cli/demo.js +45 -0
- package/dist/src/cli/demo.js.map +1 -0
- package/dist/src/cli/evaluation.js +40 -0
- package/dist/src/cli/evaluation.js.map +1 -0
- package/dist/src/cli/extension.js +54 -0
- package/dist/src/cli/extension.js.map +1 -0
- package/dist/src/cli/help.js +190 -0
- package/dist/src/cli/help.js.map +1 -0
- package/dist/src/cli/hosts.js +33 -0
- package/dist/src/cli/hosts.js.map +1 -0
- package/dist/src/cli/init-settings.js +29 -0
- package/dist/src/cli/init-settings.js.map +1 -0
- package/dist/src/cli/init.js +82 -0
- package/dist/src/cli/init.js.map +1 -0
- package/dist/src/cli/invocation.js +431 -0
- package/dist/src/cli/invocation.js.map +1 -0
- package/dist/src/cli/loading.js +110 -0
- package/dist/src/cli/loading.js.map +1 -0
- package/dist/src/cli/preparation.js +88 -0
- package/dist/src/cli/preparation.js.map +1 -0
- package/dist/src/cli/presentation.js +61 -0
- package/dist/src/cli/presentation.js.map +1 -0
- package/dist/src/cli/progress.js +5 -0
- package/dist/src/cli/progress.js.map +1 -0
- package/dist/src/cli/run-manifest.js +108 -0
- package/dist/src/cli/run-manifest.js.map +1 -0
- package/dist/src/cli/run-options.js +367 -0
- package/dist/src/cli/run-options.js.map +1 -0
- package/dist/src/cli/run-overrides.js +51 -0
- package/dist/src/cli/run-overrides.js.map +1 -0
- package/dist/src/cli/settings.js +127 -0
- package/dist/src/cli/settings.js.map +1 -0
- package/dist/src/cli/summary.js +28 -0
- package/dist/src/cli/summary.js.map +1 -0
- package/dist/src/cli.js +20 -0
- package/dist/src/cli.js.map +1 -0
- package/dist/src/contract-types.js +2 -0
- package/dist/src/contract-types.js.map +1 -0
- package/dist/src/engine.js +4 -0
- package/dist/src/engine.js.map +1 -0
- package/dist/src/evaluation/advisory.js +85 -0
- package/dist/src/evaluation/advisory.js.map +1 -0
- package/dist/src/evaluation/check-plan.js +292 -0
- package/dist/src/evaluation/check-plan.js.map +1 -0
- package/dist/src/evaluation/collection.js +61 -0
- package/dist/src/evaluation/collection.js.map +1 -0
- package/dist/src/evaluation/context.js +2 -0
- package/dist/src/evaluation/context.js.map +1 -0
- package/dist/src/evaluation/grading-host.js +28 -0
- package/dist/src/evaluation/grading-host.js.map +1 -0
- package/dist/src/evaluation/grading.js +154 -0
- package/dist/src/evaluation/grading.js.map +1 -0
- package/dist/src/evaluation/invocation-plan.js +312 -0
- package/dist/src/evaluation/invocation-plan.js.map +1 -0
- package/dist/src/evaluation/plan.js +376 -0
- package/dist/src/evaluation/plan.js.map +1 -0
- package/dist/src/evaluation/progress.js +14 -0
- package/dist/src/evaluation/progress.js.map +1 -0
- package/dist/src/evaluation/resources.js +159 -0
- package/dist/src/evaluation/resources.js.map +1 -0
- package/dist/src/evaluation/run.js +148 -0
- package/dist/src/evaluation/run.js.map +1 -0
- package/dist/src/evaluation/scheduler.js +27 -0
- package/dist/src/evaluation/scheduler.js.map +1 -0
- package/dist/src/evaluation/semantic.js +183 -0
- package/dist/src/evaluation/semantic.js.map +1 -0
- package/dist/src/evaluation/trial.js +593 -0
- package/dist/src/evaluation/trial.js.map +1 -0
- package/dist/src/evaluation/types.js +20 -0
- package/dist/src/evaluation/types.js.map +1 -0
- package/dist/src/evidence/host-results.js +67 -0
- package/dist/src/evidence/host-results.js.map +1 -0
- package/dist/src/evidence/retained-artifacts.js +17 -0
- package/dist/src/evidence/retained-artifacts.js.map +1 -0
- package/dist/src/evidence/trial-writer.js +131 -0
- package/dist/src/evidence/trial-writer.js.map +1 -0
- package/dist/src/evidence/usage.js +14 -0
- package/dist/src/evidence/usage.js.map +1 -0
- package/dist/src/executable.js +20 -0
- package/dist/src/executable.js.map +1 -0
- package/dist/src/extension-client.js +212 -0
- package/dist/src/extension-client.js.map +1 -0
- package/dist/src/extension-session.js +236 -0
- package/dist/src/extension-session.js.map +1 -0
- package/dist/src/fixtures/advisory.js +206 -0
- package/dist/src/fixtures/advisory.js.map +1 -0
- package/dist/src/fixtures/bin.js +41 -0
- package/dist/src/fixtures/bin.js.map +1 -0
- package/dist/src/fixtures/cleanup.js +43 -0
- package/dist/src/fixtures/cleanup.js.map +1 -0
- package/dist/src/fixtures/create.js +94 -0
- package/dist/src/fixtures/create.js.map +1 -0
- package/dist/src/fixtures/generated.js +181 -0
- package/dist/src/fixtures/generated.js.map +1 -0
- package/dist/src/fixtures/git-hooks.js +16 -0
- package/dist/src/fixtures/git-hooks.js.map +1 -0
- package/dist/src/fixtures/preparation.js +201 -0
- package/dist/src/fixtures/preparation.js.map +1 -0
- package/dist/src/fixtures/repository.js +255 -0
- package/dist/src/fixtures/repository.js.map +1 -0
- package/dist/src/fixtures/setup.js +175 -0
- package/dist/src/fixtures/setup.js.map +1 -0
- package/dist/src/fixtures/tools.js +57 -0
- package/dist/src/fixtures/tools.js.map +1 -0
- package/dist/src/fixtures/workspace-state.js +155 -0
- package/dist/src/fixtures/workspace-state.js.map +1 -0
- package/dist/src/generated/cli-result.cjs +2 -0
- package/dist/src/generated/extension.cjs +2 -0
- package/dist/src/generated/report.cjs +2 -0
- package/dist/src/generated/run-evidence.cjs +2 -0
- package/dist/src/generated/runtime.cjs +2 -0
- package/dist/src/git.js +72 -0
- package/dist/src/git.js.map +1 -0
- package/dist/src/graders/artifact.js +68 -0
- package/dist/src/graders/artifact.js.map +1 -0
- package/dist/src/graders/git-head.js +112 -0
- package/dist/src/graders/git-head.js.map +1 -0
- package/dist/src/graders/output.js +345 -0
- package/dist/src/graders/output.js.map +1 -0
- package/dist/src/graders/semantic.js +132 -0
- package/dist/src/graders/semantic.js.map +1 -0
- package/dist/src/graders/shell.js +321 -0
- package/dist/src/graders/shell.js.map +1 -0
- package/dist/src/host-diagnostic-redaction.js +52 -0
- package/dist/src/host-diagnostic-redaction.js.map +1 -0
- package/dist/src/host-diagnostics.js +248 -0
- package/dist/src/host-diagnostics.js.map +1 -0
- package/dist/src/hosts/claude-continuation.js +89 -0
- package/dist/src/hosts/claude-continuation.js.map +1 -0
- package/dist/src/hosts/claude-credential.js +121 -0
- package/dist/src/hosts/claude-credential.js.map +1 -0
- package/dist/src/hosts/claude-events.js +136 -0
- package/dist/src/hosts/claude-events.js.map +1 -0
- package/dist/src/hosts/claude-keychain.js +100 -0
- package/dist/src/hosts/claude-keychain.js.map +1 -0
- package/dist/src/hosts/claude-login-recovery.js +67 -0
- package/dist/src/hosts/claude-login-recovery.js.map +1 -0
- package/dist/src/hosts/claude-login-refresh.js +146 -0
- package/dist/src/hosts/claude-login-refresh.js.map +1 -0
- package/dist/src/hosts/claude-login-store.js +124 -0
- package/dist/src/hosts/claude-login-store.js.map +1 -0
- package/dist/src/hosts/claude-login.js +74 -0
- package/dist/src/hosts/claude-login.js.map +1 -0
- package/dist/src/hosts/claude-native-goal.js +138 -0
- package/dist/src/hosts/claude-native-goal.js.map +1 -0
- package/dist/src/hosts/claude-process.js +10 -0
- package/dist/src/hosts/claude-process.js.map +1 -0
- package/dist/src/hosts/claude-refresh-lock.js +76 -0
- package/dist/src/hosts/claude-refresh-lock.js.map +1 -0
- package/dist/src/hosts/claude-repository-invocation.js +229 -0
- package/dist/src/hosts/claude-repository-invocation.js.map +1 -0
- package/dist/src/hosts/claude-settings.js +156 -0
- package/dist/src/hosts/claude-settings.js.map +1 -0
- package/dist/src/hosts/claude-skill-activations.js +131 -0
- package/dist/src/hosts/claude-skill-activations.js.map +1 -0
- package/dist/src/hosts/claude-state.js +237 -0
- package/dist/src/hosts/claude-state.js.map +1 -0
- package/dist/src/hosts/claude-tool-calls.js +150 -0
- package/dist/src/hosts/claude-tool-calls.js.map +1 -0
- package/dist/src/hosts/claude.js +373 -0
- package/dist/src/hosts/claude.js.map +1 -0
- package/dist/src/hosts/codex-app-server.js +764 -0
- package/dist/src/hosts/codex-app-server.js.map +1 -0
- package/dist/src/hosts/codex-configuration.js +91 -0
- package/dist/src/hosts/codex-configuration.js.map +1 -0
- package/dist/src/hosts/codex-events.js +104 -0
- package/dist/src/hosts/codex-events.js.map +1 -0
- package/dist/src/hosts/codex-exec-controls.js +434 -0
- package/dist/src/hosts/codex-exec-controls.js.map +1 -0
- package/dist/src/hosts/codex-invocation.js +54 -0
- package/dist/src/hosts/codex-invocation.js.map +1 -0
- package/dist/src/hosts/codex-marketplace.js +185 -0
- package/dist/src/hosts/codex-marketplace.js.map +1 -0
- package/dist/src/hosts/codex-native-calls.js +516 -0
- package/dist/src/hosts/codex-native-calls.js.map +1 -0
- package/dist/src/hosts/codex-process.js +10 -0
- package/dist/src/hosts/codex-process.js.map +1 -0
- package/dist/src/hosts/codex-profile.js +206 -0
- package/dist/src/hosts/codex-profile.js.map +1 -0
- package/dist/src/hosts/codex-session-files.js +89 -0
- package/dist/src/hosts/codex-session-files.js.map +1 -0
- package/dist/src/hosts/codex-state.js +310 -0
- package/dist/src/hosts/codex-state.js.map +1 -0
- package/dist/src/hosts/codex.js +527 -0
- package/dist/src/hosts/codex.js.map +1 -0
- package/dist/src/hosts/isolation-common.js +43 -0
- package/dist/src/hosts/isolation-common.js.map +1 -0
- package/dist/src/hosts/isolation-roots.js +64 -0
- package/dist/src/hosts/isolation-roots.js.map +1 -0
- package/dist/src/hosts/isolation.js +12 -0
- package/dist/src/hosts/isolation.js.map +1 -0
- package/dist/src/hosts/linux-sandbox.js +127 -0
- package/dist/src/hosts/linux-sandbox.js.map +1 -0
- package/dist/src/hosts/mac-sandbox.js +137 -0
- package/dist/src/hosts/mac-sandbox.js.map +1 -0
- package/dist/src/hosts/native-controls.js +151 -0
- package/dist/src/hosts/native-controls.js.map +1 -0
- package/dist/src/hosts/native-request.js +33 -0
- package/dist/src/hosts/native-request.js.map +1 -0
- package/dist/src/hosts/process-lifecycle.js +167 -0
- package/dist/src/hosts/process-lifecycle.js.map +1 -0
- package/dist/src/hosts/protected-worktrees.js +81 -0
- package/dist/src/hosts/protected-worktrees.js.map +1 -0
- package/dist/src/hosts/skill-hook-catalog.js +100 -0
- package/dist/src/hosts/skill-hook-catalog.js.map +1 -0
- package/dist/src/hosts/skill-hook-handler.js +228 -0
- package/dist/src/hosts/skill-hook-handler.js.map +1 -0
- package/dist/src/hosts/skill-hooks.js +423 -0
- package/dist/src/hosts/skill-hooks.js.map +1 -0
- package/dist/src/hosts/workspace-fingerprint.js +72 -0
- package/dist/src/hosts/workspace-fingerprint.js.map +1 -0
- package/dist/src/identity.js +167 -0
- package/dist/src/identity.js.map +1 -0
- package/dist/src/instrumentation.js +83 -0
- package/dist/src/instrumentation.js.map +1 -0
- package/dist/src/native-transcript-state.js +85 -0
- package/dist/src/native-transcript-state.js.map +1 -0
- package/dist/src/native-transcript-view.js +65 -0
- package/dist/src/native-transcript-view.js.map +1 -0
- package/dist/src/native-transcripts.js +208 -0
- package/dist/src/native-transcripts.js.map +1 -0
- package/dist/src/paths.js +13 -0
- package/dist/src/paths.js.map +1 -0
- package/dist/src/process-lifecycle.js +101 -0
- package/dist/src/process-lifecycle.js.map +1 -0
- package/dist/src/provenance.js +263 -0
- package/dist/src/provenance.js.map +1 -0
- package/dist/src/prune.js +228 -0
- package/dist/src/prune.js.map +1 -0
- package/dist/src/report.js +276 -0
- package/dist/src/report.js.map +1 -0
- package/dist/src/resolved-case.js +29 -0
- package/dist/src/resolved-case.js.map +1 -0
- package/dist/src/results-lock.js +40 -0
- package/dist/src/results-lock.js.map +1 -0
- package/dist/src/results.js +127 -0
- package/dist/src/results.js.map +1 -0
- package/dist/src/run-owner.js +283 -0
- package/dist/src/run-owner.js.map +1 -0
- package/dist/src/runner-paths.js +11 -0
- package/dist/src/runner-paths.js.map +1 -0
- package/dist/src/runtime/config.js +323 -0
- package/dist/src/runtime/config.js.map +1 -0
- package/dist/src/runtime/discovery-apple.js +65 -0
- package/dist/src/runtime/discovery-apple.js.map +1 -0
- package/dist/src/runtime/discovery-providers.js +37 -0
- package/dist/src/runtime/discovery-providers.js.map +1 -0
- package/dist/src/runtime/discovery.js +40 -0
- package/dist/src/runtime/discovery.js.map +1 -0
- package/dist/src/runtime/hook-processes.js +81 -0
- package/dist/src/runtime/hook-processes.js.map +1 -0
- package/dist/src/runtime/hooks.js +304 -0
- package/dist/src/runtime/hooks.js.map +1 -0
- package/dist/src/runtime/paths.js +244 -0
- package/dist/src/runtime/paths.js.map +1 -0
- package/dist/src/runtime/query.js +76 -0
- package/dist/src/runtime/query.js.map +1 -0
- package/dist/src/runtime/seeds.js +98 -0
- package/dist/src/runtime/seeds.js.map +1 -0
- package/dist/src/runtime/shared-cache.js +81 -0
- package/dist/src/runtime/shared-cache.js.map +1 -0
- package/dist/src/runtime/state.js +106 -0
- package/dist/src/runtime/state.js.map +1 -0
- package/dist/src/runtime-process.js +148 -0
- package/dist/src/runtime-process.js.map +1 -0
- package/dist/src/schema-types.js +2 -0
- package/dist/src/schema-types.js.map +1 -0
- package/dist/src/schema.js +16 -0
- package/dist/src/schema.js.map +1 -0
- package/dist/src/shell.js +5 -0
- package/dist/src/shell.js.map +1 -0
- package/dist/src/sqlite-lock.js +18 -0
- package/dist/src/sqlite-lock.js.map +1 -0
- package/dist/src/storage.js +21 -0
- package/dist/src/storage.js.map +1 -0
- package/dist/src/transcript-retention.js +109 -0
- package/dist/src/transcript-retention.js.map +1 -0
- package/dist/src/value-guards.js +26 -0
- package/dist/src/value-guards.js.map +1 -0
- package/docs/README.md +19 -15
- package/docs/architecture.md +141 -17
- package/docs/assets/README.md +5 -4
- package/docs/builtin-graders-v1.md +9 -2
- package/docs/codex-host.md +155 -57
- package/docs/contracts.md +2 -1
- package/docs/creating-cases.md +10 -4
- package/docs/development-cli.md +197 -24
- package/docs/documentation-quality.md +16 -5
- package/docs/evaluation-workflow.md +110 -0
- package/docs/extension-protocol-v1.md +26 -3
- package/docs/getting-started.md +29 -7
- package/docs/guide-evaluation.md +34 -2
- package/docs/identity-v1.md +29 -3
- package/docs/installing.md +35 -17
- package/docs/licensing.md +2 -2
- package/docs/native-controls.md +76 -5
- package/docs/native-hosts.md +197 -8
- package/docs/reading-results.md +102 -0
- package/docs/recipes.md +490 -0
- package/docs/releases.md +48 -13
- package/docs/report-v1.md +29 -1
- package/docs/results-v1.md +122 -4
- package/docs/run-settings-v1.md +147 -0
- package/docs/runtime-v1.md +255 -22
- package/docs/specs/repository-guide.md +1 -1
- package/docs/troubleshooting.md +12 -12
- package/docs/typescript-quality.md +106 -26
- package/examples/recipes/candidate.ts +26 -0
- package/examples/recipes/cases/coding-change.json +21 -0
- package/examples/recipes/cases/host-tools.json +21 -0
- package/examples/recipes/cases/mise.json +21 -0
- package/examples/recipes/cases/node.json +21 -0
- package/examples/recipes/cases/python-uv.json +21 -0
- package/examples/recipes/cases/semantic.json +17 -0
- package/examples/recipes/coding-change/check.ts +23 -0
- package/examples/recipes/executable.ts +19 -0
- package/examples/recipes/extension-artifacts.ts +65 -0
- package/examples/recipes/extension-methods.ts +174 -0
- package/examples/recipes/extension.ts +38 -0
- package/examples/recipes/modules.ts +9 -0
- package/examples/recipes/repository-skill/.agents/skills/startup-guide/SKILL.md +7 -0
- package/examples/recipes/repository-skill/.claude/skills/startup-guide/SKILL.md +7 -0
- package/examples/recipes/repository-skill/README.md +3 -0
- package/examples/recipes/run-inputs.ts +183 -0
- package/examples/recipes/run-options.ts +132 -0
- package/examples/recipes/run.ts +77 -0
- package/examples/recipes/runtime/empty.json +1 -0
- package/examples/recipes/runtime/goals.json +1 -0
- package/examples/recipes/runtime/mise.json +6 -0
- package/examples/recipes/runtime/path.json +1 -0
- package/examples/recipes/runtime/plugins.json +4 -0
- package/examples/recipes/runtime/python-uv.json +10 -0
- package/examples/recipes/semantic-judge.ts +29 -0
- package/examples/templates/coding/adapter.ts +17 -0
- package/examples/templates/coding/case.json +15 -0
- package/examples/templates/repository-skill/.agents/skills/startup-guide/SKILL.md +6 -0
- package/examples/templates/repository-skill/extension-data.ts +64 -0
- package/examples/templates/repository-skill/extension.ts +32 -0
- package/examples/templates/response/adapter.ts +12 -0
- package/examples/templates/response/case.json +13 -0
- package/package.json +23 -9
- package/schemas/report-v1.schema.json +9 -0
- package/schemas/run-evidence-v1.schema.json +2 -1
- package/schemas/runtime-v1.schema.json +43 -7
- package/src/advisory.ts +2 -2
- package/src/candidate-artifacts.ts +42 -11
- package/src/candidate-codex-state.ts +185 -0
- package/src/case-validation.ts +73 -50
- package/src/cli/cancellation.ts +20 -0
- package/src/cli/command.ts +87 -0
- package/src/cli/completion.ts +15 -0
- package/src/cli/demo.ts +61 -0
- package/src/cli/evaluation.ts +55 -0
- package/src/cli/extension.ts +79 -0
- package/src/cli/help.ts +199 -0
- package/src/cli/hosts.ts +55 -0
- package/src/cli/init-settings.ts +31 -0
- package/src/cli/init.ts +112 -0
- package/src/{cli-invocation.ts → cli/invocation.ts} +215 -218
- package/src/cli/loading.ts +151 -0
- package/src/cli/preparation.ts +154 -0
- package/src/cli/presentation.ts +93 -0
- package/src/cli/progress.ts +6 -0
- package/src/cli/run-manifest.ts +128 -0
- package/src/cli/run-options.ts +403 -0
- package/src/cli/run-overrides.ts +69 -0
- package/src/cli/settings.ts +154 -0
- package/src/cli/summary.ts +43 -0
- package/src/cli.ts +14 -592
- package/src/contract-types.ts +94 -0
- package/src/engine.ts +9 -5
- package/src/evaluation/advisory.ts +150 -0
- package/src/{evaluation-check-plan.ts → evaluation/check-plan.ts} +103 -67
- package/src/evaluation/collection.ts +88 -0
- package/src/evaluation/context.ts +10 -0
- package/src/{evaluation-grading-host.ts → evaluation/grading-host.ts} +5 -5
- package/src/evaluation/grading.ts +235 -0
- package/src/{evaluation-invocation-plan.ts → evaluation/invocation-plan.ts} +147 -58
- package/src/{evaluation-plan.ts → evaluation/plan.ts} +208 -141
- package/src/evaluation/progress.ts +19 -0
- package/src/evaluation/resources.ts +241 -0
- package/src/evaluation/run.ts +206 -0
- package/src/{evaluation-semantic.ts → evaluation/semantic.ts} +77 -87
- package/src/evaluation/trial.ts +800 -0
- package/src/{evaluation-types.ts → evaluation/types.ts} +79 -29
- package/src/evidence/host-results.ts +85 -0
- package/src/evidence/retained-artifacts.ts +18 -0
- package/src/evidence/trial-writer.ts +203 -0
- package/src/evidence/usage.ts +17 -0
- package/src/executable.ts +19 -0
- package/src/extension-client.ts +40 -108
- package/src/extension-session.ts +24 -48
- package/src/{advisory-fixture.ts → fixtures/advisory.ts} +12 -72
- package/src/{fixture-bin.ts → fixtures/bin.ts} +1 -1
- package/src/{fixture-cleanup.ts → fixtures/cleanup.ts} +13 -7
- package/src/fixtures/create.ts +183 -0
- package/src/{generated-fixture.ts → fixtures/generated.ts} +10 -53
- package/src/{git-hooks.ts → fixtures/git-hooks.ts} +1 -1
- package/src/{preparation.ts → fixtures/preparation.ts} +6 -16
- package/src/{repository-fixture.ts → fixtures/repository.ts} +30 -64
- package/src/{fixture-setup.ts → fixtures/setup.ts} +36 -53
- package/src/{fixture-tools.ts → fixtures/tools.ts} +1 -1
- package/src/{workspace-state.ts → fixtures/workspace-state.ts} +5 -5
- package/src/generated/cli-result.d.cts +2 -1
- package/src/generated/extension.d.cts +2 -1
- package/src/generated/report.cjs +1 -1
- package/src/generated/report.d.cts +2 -1
- package/src/generated/run-evidence.cjs +1 -1
- package/src/generated/run-evidence.d.cts +2 -1
- package/src/generated/runtime.cjs +1 -1
- package/src/generated/runtime.d.cts +2 -1
- package/src/git.ts +94 -0
- package/src/graders/git-head.ts +31 -85
- package/src/graders/shell.ts +54 -71
- package/src/host-diagnostic-redaction.ts +52 -0
- package/src/host-diagnostics.ts +331 -0
- package/src/hosts/claude-continuation.ts +57 -40
- package/src/hosts/claude-credential.ts +160 -63
- package/src/hosts/claude-events.ts +34 -18
- package/src/hosts/claude-keychain.ts +79 -21
- package/src/hosts/claude-login-recovery.ts +85 -0
- package/src/hosts/claude-login-refresh.ts +179 -0
- package/src/hosts/claude-login-store.ts +162 -0
- package/src/hosts/claude-login.ts +107 -0
- package/src/hosts/claude-native-goal.ts +15 -12
- package/src/hosts/claude-process.ts +10 -50
- package/src/hosts/claude-refresh-lock.ts +82 -0
- package/src/hosts/claude-repository-invocation.ts +9 -34
- package/src/hosts/claude-settings.ts +46 -9
- package/src/hosts/claude-skill-activations.ts +169 -0
- package/src/hosts/claude-state.ts +28 -15
- package/src/hosts/claude-tool-calls.ts +13 -10
- package/src/hosts/claude.ts +121 -81
- package/src/hosts/codex-app-server.ts +78 -52
- package/src/hosts/codex-configuration.ts +5 -10
- package/src/hosts/codex-exec-controls.ts +516 -0
- package/src/hosts/codex-marketplace.ts +12 -10
- package/src/hosts/codex-native-calls.ts +15 -122
- package/src/hosts/codex-process.ts +10 -183
- package/src/hosts/codex-profile.ts +36 -21
- package/src/hosts/codex-session-files.ts +1 -9
- package/src/hosts/codex-state.ts +134 -18
- package/src/hosts/codex.ts +210 -183
- package/src/hosts/isolation-common.ts +8 -11
- package/src/hosts/isolation-roots.ts +5 -4
- package/src/hosts/linux-sandbox.ts +15 -2
- package/src/hosts/mac-sandbox.ts +87 -18
- package/src/hosts/native-controls.ts +72 -22
- package/src/hosts/native-request.ts +55 -0
- package/src/hosts/process-lifecycle.ts +215 -49
- package/src/hosts/protected-worktrees.ts +3 -2
- package/src/hosts/skill-hook-catalog.ts +131 -0
- package/src/hosts/skill-hook-handler.ts +374 -0
- package/src/hosts/skill-hooks.ts +585 -0
- package/src/instrumentation.ts +7 -8
- package/src/native-transcript-state.ts +22 -3
- package/src/native-transcript-view.ts +52 -13
- package/src/native-transcripts.ts +26 -11
- package/src/paths.ts +22 -0
- package/src/process-lifecycle.ts +150 -0
- package/src/provenance.ts +34 -13
- package/src/prune.ts +300 -0
- package/src/report.ts +63 -36
- package/src/resolved-case.ts +1 -1
- package/src/results-lock.ts +42 -0
- package/src/results.ts +14 -18
- package/src/run-owner.ts +15 -9
- package/src/runner-paths.ts +21 -0
- package/src/{runtime-config.ts → runtime/config.ts} +160 -46
- package/src/{runtime-discovery-apple.ts → runtime/discovery-apple.ts} +3 -3
- package/src/{runtime-discovery-providers.ts → runtime/discovery-providers.ts} +3 -9
- package/src/{runtime-discovery.ts → runtime/discovery.ts} +4 -4
- package/src/{runtime-hook-processes.ts → runtime/hook-processes.ts} +3 -2
- package/src/{runtime-hooks.ts → runtime/hooks.ts} +40 -28
- package/src/{runtime-paths.ts → runtime/paths.ts} +5 -19
- package/src/{runtime-query.ts → runtime/query.ts} +5 -5
- package/src/{runtime-seeds.ts → runtime/seeds.ts} +3 -3
- package/src/runtime/shared-cache.ts +130 -0
- package/src/{runtime-state.ts → runtime/state.ts} +28 -11
- package/src/runtime-process.ts +194 -0
- package/src/schema-types.ts +73 -161
- package/src/shell.ts +4 -0
- package/src/sqlite-lock.ts +25 -0
- package/src/transcript-retention.ts +146 -0
- package/src/value-guards.ts +22 -0
- package/docs/assets/sevro-logo.png +0 -0
- package/docs/assets/sevro-symbol.svg +0 -1
- package/docs/assets/sevro-wordmark.svg +0 -1
- package/docs/cache-ownership-validation.md +0 -49
- package/docs/claude-continuation-validation.md +0 -47
- package/docs/concurrency-validation.md +0 -86
- package/docs/configuration-root-validation.md +0 -70
- package/docs/engine-slice.md +0 -114
- package/docs/evidence/runtime-startup-30/README.md +0 -185
- package/docs/evidence/runtime-startup-30/after.json +0 -319
- package/docs/evidence/runtime-startup-30/baseline-timeout.jsonl +0 -2
- package/docs/evidence/runtime-startup-30/before.json +0 -403
- package/docs/evidence/runtime-startup-30/consumer-smoke.json +0 -63
- package/docs/evidence/runtime-startup-30/full-policy-diagnostic.jsonl +0 -3
- package/docs/evidence/runtime-startup-30/minimal-policy-diagnostic.jsonl +0 -3
- package/docs/evidence/runtime-startup-30/plugin-template-assessment.json +0 -20
- package/docs/evidence/runtime-startup-30/seedless-timeout.jsonl +0 -2
- package/docs/evidence.md +0 -17
- package/docs/fixture-cleanup-validation.md +0 -73
- package/docs/guide-verification.json +0 -239
- package/src/evaluation-context.ts +0 -259
- package/src/evaluation-fixture.ts +0 -302
- package/src/evaluation-run.ts +0 -200
- package/src/evaluation-trial.ts +0 -1016
- package/src/hosts/claude-nested-skills.ts +0 -272
- package/src/hosts/codex-command-output.ts +0 -510
- package/src/hosts/codex-executor-command.ts +0 -291
- package/src/hosts/codex-skill-reads.ts +0 -783
- /package/src/{trial-scheduler.ts → evaluation/scheduler.ts} +0 -0
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"claude-state.js","sourceRoot":"","sources":["claude-state.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,kBAAkB,EAAE,6BAA0B;AACvD,OAAO,EAAE,YAAY,EAAE,oBAAiB;AACxC,OAAO,EAAE,UAAU,EAAE,oBAAiB;AACtC,OAAO,EACL,4BAA4B,EAC5B,0BAA0B,GAC3B,kCAA+B;AAChC,OAAO,EACL,EAAE,EACF,KAAK,EACL,QAAQ,EACR,IAAI,EACJ,OAAO,EACP,SAAS,GACV,yBAAyB;AAC1B,OAAO,EAAE,SAAS,EAAE,OAAO,EAAE,IAAI,EAAE,kBAAkB;AAErD,OAAO,EAAE,iBAAiB,EAAE,sCAAmC;AAC/D,OAAO,EACL,0BAA0B,EAC1B,kBAAkB,GACnB,uCAAoC;AACrC,OAAO,EACL,8BAA8B,EAC9B,wBAAwB,GACzB,iCAA8B;AAC/B,OAAO,EAAE,iBAAiB,EAAE,mBAAmB,EAAE,4BAAyB;AAC1E,OAAO,EACL,iBAAiB,IAAI,MAAM,EAC3B,2BAA2B,GAC5B,4BAAyB;AAC1B,OAAO,EAAE,mBAAmB,EAAE,4BAAyB;AACvD,OAAO,EACL,oBAAoB,EACpB,gBAAgB,GACjB,mCAAgC;AACjC,OAAO,EAAE,yBAAyB,EAAE,+BAA4B;AAChE,OAAO,EAAE,kBAAkB,EAAE,6BAA0B;AACvD,OAAO,EAAE,wBAAwB,EAAE,6BAA0B;AAiB7D,SAAS,iBAAiB,CAAC,OAAgB;IACzC,OAAO,kBAAkB,CAAC,OAAO,CAAC,aAAa,CAAC,CAAC;AACnD,CAAC;AAED,KAAK,UAAU,wBAAwB,CACrC,SAAiB,EACjB,OAA0B,EAC1B,eAAuB;IAEvB,MAAM,KAAK,CAAC,IAAI,CAAC,SAAS,EAAE,SAAS,CAAC,EAAE,EAAE,IAAI,EAAE,KAAK,EAAE,CAAC,CAAC;IACzD,MAAM,WAAW,GAAG,MAAM,QAAQ,CAAC,IAAI,CAAC,SAAS,EAAE,SAAS,CAAC,CAAC,CAAC;IAC/D,MAAM,IAAI,GAAG,IAAI,CAAC,SAAS,EAAE,MAAM,CAAC,EAClC,IAAI,GAAG,IAAI,CAAC,SAAS,EAAE,KAAK,CAAC,CAAC;IAChC,MAAM,OAAO,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,KAAK,CAAC,IAAI,EAAE,EAAE,IAAI,EAAE,KAAK,EAAE,CAAC,CAAC,CAAC,CAAC;IAC5E,MAAM,cAAc,GAAG,MAAM,yBAAyB,CACpD,IAAI,CAAC,WAAW,EAAE,QAAQ,CAAC,EAC3B,OAAO,CAAC,cAAc,EACtB,EAAE,MAAM,EAAE,OAAO,CAAC,MAAM,EAAE,UAAU,EAAE,eAAe,EAAE,CACxD,CAAC;IACF,OAAO;QACL,WAAW;QACX,IAAI;QACJ,IAAI;QACJ,cAAc;QACd,UAAU,EAAE,cAAc,CAAC,cAAc;KAC1C,CAAC;AACJ,CAAC;AAED,KAAK,UAAU,oBAAoB,CACjC,OAAgB,EAChB,OAA0B,EAC1B,UAAkB;IAElB,MAAM,UAAU,GAAG,MAAM,wBAAwB,CAAC;QAChD,cAAc,EAAE,OAAO,CAAC,OAAO,CAAC,OAAO,CAAC,UAAU,CAAC,CAAC,CAAC;QACrD,SAAS,EAAE,OAAO,CAAC,SAAS;QAC5B,WAAW,EAAE,OAAO,CAAC,WAAW;QAChC,WAAW,EAAE,OAAO,CAAC,WAAW;QAChC,eAAe,EAAE;YACf,GAAG,OAAO,CAAC,wBAAwB;YACnC,GAAG,iBAAiB,CAAC,OAAO,CAAC;SAC9B;KACF,CAAC,CAAC;IACH,MAAM,KAAK,GAAG,OAAO,CAAC,aAAa,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,YAAY,CAAC,UAAU,CAAC,CAAC;IAC5E,MAAM,SAAS,GAAG,MAAM,iBAAiB,EAAE,CAAC;IAC5C,IAAI,KAAK,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,KAAK,SAAS,IAAI,MAAM,CAAC,IAAI,EAAE,UAAU,CAAC,CAAC;QACtE,MAAM,IAAI,KAAK,CAAC,gDAAgD,CAAC,CAAC;IACpE,MAAM,MAAM,GAAG,MAAM,QAAQ,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC;IAC9C,MAAM,eAAe,GAAG,2BAA2B,CACjD,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,EACjB,KAAK,EACL,OAAO,CAAC,OAAO,CAAC,aAAa,CAAC,CAC/B,CAAC;IACF,IAAI,eAAe,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,MAAM,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC;QACtD,MAAM,IAAI,KAAK,CAAC,mDAAmD,CAAC,CAAC;IACvE,OAAO,KAAK,CAAC;AACf,CAAC;AAED,KAAK,UAAU,wBAAwB,CACrC,OAA0B,EAC1B,KAAe;IAEf,MAAM,SAAS,GAAG,OAAO,CAAC,eAAe;QACvC,CAAC,CAAC,MAAM,QAAQ,CAAC,OAAO,CAAC,eAAe,CAAC;QACzC,CAAC,CAAC,IAAI,CAAC;IACT,IACE,SAAS;QACT,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,MAAM,CAAC,IAAI,EAAE,SAAS,CAAC,CAAC;YAC5C,CAAC,CAAC,MAAM,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC,WAAW,EAAE,CAAC;QAEzC,MAAM,IAAI,KAAK,CAAC,wDAAwD,CAAC,CAAC;IAC5E,OAAO,SAAS,CAAC;AACnB,CAAC;AAED,KAAK,UAAU,oBAAoB,CACjC,OAA0B,EAC1B,KAAe,EACf,SAAiB;IAEjB,IAAI,CAAC,OAAO,CAAC,UAAU;QAAE,OAAO,IAAI,CAAC;IACrC,MAAM,KAAK,GAAG,MAAM,QAAQ,CAAC,OAAO,CAAC,UAAU,CAAC,CAAC;IACjD,IACE,KAAK,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,MAAM,CAAC,IAAI,EAAE,KAAK,CAAC,CAAC;QACzC,CAAC,CAAC,MAAM,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,WAAW,EAAE;QAElC,MAAM,IAAI,KAAK,CAAC,6CAA6C,CAAC,CAAC;IACjE,MAAM,WAAW,GAAG,IAAI,CAAC,SAAS,EAAE,MAAM,EAAE,eAAe,CAAC,CAAC;IAC7D,MAAM,KAAK,CAAC,WAAW,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE,CAAC,CAAC;IAC3D,MAAM,EAAE,CAAC,KAAK,EAAE,IAAI,CAAC,WAAW,EAAE,UAAU,CAAC,EAAE;QAC7C,SAAS,EAAE,IAAI;QACf,KAAK,EAAE,KAAK;QACZ,YAAY,EAAE,IAAI;KACnB,CAAC,CAAC;IACH,OAAO,WAAW,CAAC;AACrB,CAAC;AAED,SAAS,iBAAiB,CACxB,KAA2D,EAC3D,OAAgB,EAChB,SAAwB,EACxB,OAAsB;IAEtB,OAAO;QACL,GAAG,KAAK,CAAC,cAAc,CAAC,WAAW;QACnC,GAAG,OAAO,CAAC,aAAa,EAAE,WAAW;QACrC,gCAAgC,EAAE,GAAG;QACrC,IAAI,EAAE,UAAU,CAAC,OAAO,EAAE,SAAS,CAAC;QACpC,IAAI,EAAE,OAAO,CAAC,GAAG,CAAC,IAAI,IAAI,GAAG;QAC7B,IAAI,EAAE,OAAO,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,EAAE,WAAW,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,IAAI;QACvD,MAAM,EAAE,KAAK,CAAC,IAAI;QAClB,iBAAiB,EAAE,OAAO,CAAC,KAAK,CAAC,UAAU,CAAC;QAC5C,QAAQ,EAAE,GAAG;QACb,GAAG,wBAAwB,CAAC,OAAO,CAAC;KACrC,CAAC;AACJ,CAAC;AAED,SAAS,UAAU,CAAC,OAAgB,EAAE,SAAwB;IAC5D,MAAM,IAAI,GACR,OAAO,CAAC,aAAa,EAAE,WAAW,CAAC,IAAI;QACvC,OAAO,CAAC,GAAG,CAAC,IAAI;QAChB,eAAe,CAAC;IAClB,OAAO,CAAC,OAAO,CAAC,aAAa,EAAE,SAAS,EAAE,IAAI,CAAC;SAC5C,MAAM,CAAC,OAAO,CAAC;SACf,IAAI,CAAC,SAAS,CAAC,CAAC;AACrB,CAAC;AAED,SAAS,wBAAwB,CAC/B,OAAsB;IAEtB,OAAO,OAAO;QACZ,CAAC,CAAC;YACE,YAAY,EAAE,IAAI,CAAC,OAAO,EAAE,UAAU,CAAC;YACvC,sBAAsB,EAAE,IAAI,CAAC,OAAO,EAAE,qBAAqB,CAAC;YAC5D,UAAU,EAAE,GAAG;YACf,uBAAuB,EAAE,GAAG;SAC7B;QACH,CAAC,CAAC,EAAE,CAAC;AACT,CAAC;AAED,8EAA8E;AAC9E,MAAM,CAAC,KAAK,UAAU,kBAAkB,CACtC,SAAiB,EACjB,OAA0B,EAC1B,OAAgB,EAChB,UAAoB,EACpB,UAAkB;IAElB,MAAM,6BAA6B,CAAC,OAAO,EAAE,SAAS,CAAC,CAAC;IACxD,MAAM,KAAK,GAAG,MAAM,wBAAwB,CAAC,SAAS,EAAE,OAAO,EAAE,UAAU,CAAC,CAAC;IAC7E,MAAM,WAAW,GAAG,MAAM,sBAAsB,CAC9C,SAAS,EACT,KAAK,CAAC,UAAU,EAChB,OAAO,CACR,CAAC;IACF,MAAM,KAAK,GAAG,MAAM,oBAAoB,CAAC,OAAO,EAAE,OAAO,EAAE,KAAK,CAAC,UAAU,CAAC,CAAC;IAC7E,MAAM,SAAS,GAAG,MAAM,wBAAwB,CAAC,OAAO,EAAE,KAAK,CAAC,CAAC;IACjE,MAAM,OAAO,GAAG,MAAM,oBAAoB,CAAC,OAAO,EAAE,KAAK,EAAE,OAAO,CAAC,SAAS,CAAC,CAAC;IAC9E,MAAM,YAAY,GAAG,IAAI,CAAC,KAAK,CAAC,WAAW,EAAE,eAAe,CAAC,CAAC;IAC9D,MAAM,KAAK,GAAG,MAAM,gBAAgB,CAClC,OAAO,EACP,SAAS,EACT,KAAK,CAAC,WAAW,EACjB,UAAU,EACV,KAAK,CACN,CAAC;IACF,MAAM,gBAAgB,GAAG,KAAK,EAAE,WAAW,IAAI,UAAU,CAAC;IAC1D,MAAM,GAAG,GAAG,iBAAiB,CAAC,KAAK,EAAE,OAAO,EAAE,SAAS,EAAE,OAAO,CAAC,CAAC;IAClE,MAAM,gBAAgB,GAAG,MAAM,6BAA6B,CAC1D,OAAO,EACP,GAAG,EACH,OAAO,EACP,SAAS,CACV,CAAC;IACF,MAAM,UAAU,GAAG,MAAM,gBAAgB,CACvC,OAAO,CAAC,aAAa,EACrB,OAAO,CAAC,WAAW,EACnB,KAAK,CACN,CAAC;IACF,MAAM,SAAS,CACb,YAAY,EACZ,IAAI,CAAC,SAAS,CACZ,kBAAkB,CAChB,KAAK,CAAC,WAAW,EACjB,KAAK,CAAC,UAAU,EAChB,CAAC,GAAG,UAAU,EAAE,GAAG,gBAAgB,CAAC,EACpC;QACE,GAAG,KAAK;QACR,GAAG,oBAAoB,CAAC,OAAO,CAAC,aAAa,EAAE,OAAO,CAAC,WAAW,CAAC;KACpE,EACD,WAAW,CAAC,MAAM,EAClB;QACE,GAAG,WAAW,CAAC,SAAS;QACxB,GAAG,wBAAwB,CAAC,OAAO,CAAC,uBAAuB,CAAC;QAC5D,GAAG,0BAA0B,CAAC,OAAO,CAAC,sBAAsB,CAAC;KAC9D,EACD,gBAAgB;QACd,CAAC,CAAC,EAAE,IAAI,EAAE,gBAAgB,EAAE,SAAS,EAAE,OAAO,CAAC,SAAS,EAAE,UAAU,EAAE;QACtE,CAAC,CAAC,SAAS,EACb,OAAO,CAAC,SAAS,CAClB,CACF,EACD,EAAE,IAAI,EAAE,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE,CAC5B,CAAC;IACF,MAAM,CAAC,MAAM,CAAC,GAAG,EAAE,WAAW,CAAC,WAAW,CAAC,CAAC;IAC5C,MAAM,CAAC,MAAM,CACX,GAAG,EACH,8BAA8B,CAAC,OAAO,CAAC,uBAAuB,CAAC,EAC/D,4BAA4B,CAAC,OAAO,CAAC,sBAAsB,CAAC,CAC7D,CAAC;IACF,IAAI,OAAO;QAAE,MAAM,KAAK,CAAC,IAAI,CAAC,OAAO,EAAE,WAAW,CAAC,EAAE,EAAE,IAAI,EAAE,KAAK,EAAE,CAAC,CAAC;IACtE,OAAO;QACL,UAAU,EAAE,KAAK,CAAC,UAAU;QAC5B,YAAY;QACZ,GAAG;QACH,KAAK;QACL,UAAU,EAAE,gBAAgB;KAC7B,CAAC;AACJ,CAAC;AAED,0FAA0F;AAC1F,KAAK,UAAU,6BAA6B,CAC1C,OAAgB,EAChB,SAAiB;IAEjB,IACE,CAAC,OAAO,CAAC,aAAa;QACtB,CAAC,0BAA0B,CAAC,kBAAkB,EAAE,EAAE,OAAO,CAAC,SAAS,CAAC,CAAC,MAAM;QAE3E,OAAO;IACT,MAAM,aAAa,GAAG,MAAM,OAAO,CAAC,GAAG,CACrC,CAAC,aAAa,EAAE,qBAAqB,CAAC,CAAC,GAAG,CAAC,wBAAwB,CAAC,CACrE,CAAC;IACF,MAAM,SAAS,GAAG,OAAO,CAAC,SAAS,CAAC,CAAC;IACrC,IACE,aAAa;SACV,IAAI,EAAE;SACN,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,MAAM,CAAC,IAAI,EAAE,SAAS,CAAC,IAAI,MAAM,CAAC,SAAS,EAAE,IAAI,CAAC,CAAC;QAErE,MAAM,IAAI,KAAK,CACb,4DAA4D,CAC7D,CAAC;AACN,CAAC;AAED,KAAK,UAAU,wBAAwB,CAAC,IAAY;IAClD,IAAI,CAAC;QACH,OAAO,CAAC,IAAI,EAAE,MAAM,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC;IACtC,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QACf,IAAK,KAA+B,CAAC,IAAI,KAAK,QAAQ;YACpD,MAAM,IAAI,KAAK,CAAC,oDAAoD,EAAE;gBACpE,KAAK;aACN,CAAC,CAAC;QACL,OAAO,CAAC,IAAI,CAAC,CAAC;IAChB,CAAC;AACH,CAAC;AAED,KAAK,UAAU,sBAAsB,CACnC,SAAiB,EACjB,UAAkB,EAClB,OAAgB;IAEhB,MAAM,MAAM,GAAG,OAAO,CAAC,aAAa,CAAC;IACrC,IAAI,CAAC,MAAM,EAAE,iBAAiB;QAC5B,OAAO,EAAE,MAAM,EAAE,WAAW,EAAE,EAAE,EAAE,SAAS,EAAE,EAAE,EAAE,CAAC;IACpD,MAAM,IAAI,GAAG,IAAI,CAAC,SAAS,EAAE,aAAa,CAAC,CAAC;IAC5C,MAAM,QAAQ,GAAG,IAAI,CAAC,IAAI,EAAE,UAAU,CAAC,CAAC;IACxC,MAAM,KAAK,CAAC,QAAQ,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE,CAAC,CAAC;IACxD,MAAM,OAAO,CAAC,QAAQ,EAAE,IAAI,CAAC,OAAO,CAAC,UAAU,CAAC,EAAE,UAAU,CAAC,CAAC,CAAC;IAC/D,MAAM,OAAO,GAAG,IAAI,CAAC,IAAI,EAAE,WAAW,CAAC,CAAC;IACxC,MAAM,MAAM,GAAG,4BAA4B,UAAU,CAAC,IAAI,CAAC,IAAI,CAAC;IAChE,MAAM,SAAS,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE,IAAI,EAAE,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE,CAAC,CAAC;IAC9D,MAAM,SAAS,CAAC,IAAI,CAAC,IAAI,EAAE,SAAS,CAAC,EAAE,MAAM,EAAE,EAAE,IAAI,EAAE,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE,CAAC,CAAC;IAC5E,OAAO;QACL,MAAM;QACN,SAAS,EAAE,CAAC,QAAQ,EAAE,IAAI,CAAC;QAC3B,WAAW,EAAE;YACX,4BAA4B,EAAE,QAAQ;YACtC,QAAQ,EAAE,OAAO;YACjB,OAAO,EAAE,IAAI;SACd;KACF,CAAC;AACJ,CAAC;AAED,KAAK,UAAU,gBAAgB,CAC7B,OAAgB,EAChB,SAAiB,EACjB,WAAmB,EACnB,UAAoB,EACpB,KAAe;IAEf,MAAM,MAAM,GAAG,OAAO,CAAC,aAAa,CAAC;IACrC,IAAI,CAAC,MAAM,IAAI,CAAC,mBAAmB,CAAC,MAAM,CAAC;QAAE,OAAO,SAAS,CAAC;IAC9D,OAAO,iBAAiB,CAAC;QACvB,SAAS,EAAE,OAAO,CAAC,SAAS;QAC5B,SAAS;QACT,WAAW;QACX,WAAW,EAAE,UAAU;QACvB,cAAc,EAAE,CAAC,WAAW,EAAE,GAAG,KAAK,EAAE,GAAG,kBAAkB,CAAC,MAAM,CAAC,CAAC;QACtE,MAAM;KACP,CAAC,CAAC;AACL,CAAC;AAED,KAAK,UAAU,6BAA6B,CAC1C,OAAgB,EAChB,GAA2B,EAC3B,OAAsB,EACtB,SAAiB;IAEjB,MAAM,QAAQ,GAAG,MAAM,mBAAmB,CACxC,OAAO,CAAC,SAAS,EACjB,mBAAmB,CAAC,OAAO,CAAC,EAC5B,OAAO,CAAC,WAAW,IAAI,WAAW,EAClC,sBAAsB,CAAC,OAAO,EAAE,SAAS,CAAC,CAC3C,CAAC;IACF,IAAI,OAAO,CAAC,aAAa,EAAE,CAAC;QAC1B,MAAM,CAAC,MAAM,CAAC,GAAG,EAAE,QAAQ,CAAC,WAAW,CAAC,CAAC;QACzC,GAAG,CAAC,IAAI,GAAG,CAAC,OAAO,CAAC,aAAa,EAAE,QAAQ,CAAC,WAAW,CAAC,IAAI,IAAI,GAAG,CAAC,IAAI,CAAC;aACtE,MAAM,CAAC,OAAO,CAAC;aACf,IAAI,CAAC,SAAS,CAAC,CAAC;QACnB,OAAO,QAAQ,CAAC,IAAI,CAAC;IACvB,CAAC;IACD,mCAAmC,CAAC,GAAG,EAAE,QAAQ,EAAE,OAAO,CAAC,CAAC;IAC5D,OAAO,SAAS,CAAC;AACnB,CAAC;AAED,SAAS,sBAAsB,CAAC,OAAgB,EAAE,SAAiB;IACjE,OAAO,OAAO,CAAC,aAAa;QAC1B,CAAC,CAAC,CAAC,OAAO,CAAC,oBAAoB,IAAI,IAAI,CAAC,SAAS,EAAE,SAAS,CAAC,CAAC;QAC9D,CAAC,CAAC,SAAS,CAAC;AAChB,CAAC;AAED,SAAS,mCAAmC,CAC1C,GAA2B,EAC3B,QAAyD,EACzD,OAAsB;IAEtB,IAAI,CAAC,OAAO;QAAE,GAAG,CAAC,IAAI,GAAG,QAAQ,CAAC,IAAI,CAAC;IACvC,GAAG,CAAC,MAAM,GAAG,QAAQ,CAAC,IAAI,CAAC;AAC7B,CAAC;AAED,SAAS,mBAAmB,CAAC,OAAgB;IAC3C,OAAO,CACL,OAAO,CAAC,aAAa,IAAI;QACvB,MAAM,EAAE,kBAA2B;QACnC,WAAW,EAAE,EAAE;QACf,aAAa,EAAE,EAAE;KAClB,CACF,CAAC;AACJ,CAAC","sourcesContent":["import { runtimeSeedSources } from \"../runtime/config\";\nimport { minimalRoots } from \"../paths\";\nimport { shellQuote } from \"../shell\";\nimport {\n candidateArtifactEnvironment,\n candidateArtifactReadRoots,\n} from \"../candidate-artifacts\";\nimport {\n cp,\n mkdir,\n realpath,\n stat,\n symlink,\n writeFile,\n} from \"node:fs/promises\";\nimport { delimiter, dirname, join } from \"node:path\";\nimport type { HostAdapter } from \"../engine\";\nimport { nativeStateParent } from \"../native-transcript-state\";\nimport {\n workspaceBoundaryException,\n workspaceStatePath,\n} from \"../fixtures/workspace-state\";\nimport {\n candidateTranscriptEnvironment,\n candidateTranscriptRoots,\n} from \"../native-transcripts\";\nimport { prepareHookMounts, runtimeHooksEnabled } from \"../runtime/hooks\";\nimport {\n insideRuntimeRoot as inside,\n runtimeExecutableProtection,\n} from \"../runtime/paths\";\nimport { prepareRuntimeState } from \"../runtime/state\";\nimport {\n sharedCacheDenyRoots,\n sharedCacheRoots,\n} from \"../runtime/shared-cache\";\nimport { stageClaudeAuthentication } from \"./claude-credential\";\nimport { claudeHostSettings } from \"./claude-settings\";\nimport { evaluationProtectedRoots } from \"./isolation-roots\";\ntype Request = Parameters<HostAdapter[\"run\"]>[0];\nexport interface ClaudeHostOptions {\n binary: string;\n model: string;\n effort: string;\n projectRoot: string;\n resultsRoot: string;\n additionalProtectedRoots: string[];\n credentialFile?: string;\n uvCacheDir?: string;\n toolchainBinDir?: string;\n projectSettings?: boolean;\n timeoutMs?: number;\n retainHostDiagnostics?: boolean;\n}\n\nfunction claudeSeedSources(request: Request): string[] {\n return runtimeSeedSources(request.runtimePolicy);\n}\n\nasync function createClaudePrivateState(\n stateRoot: string,\n options: ClaudeHostOptions,\n loginLifetimeMs: number,\n) {\n await mkdir(join(stateRoot, \"private\"), { mode: 0o700 });\n const privateRoot = await realpath(join(stateRoot, \"private\"));\n const home = join(stateRoot, \"home\"),\n temp = join(stateRoot, \"tmp\");\n await Promise.all([home, temp].map((path) => mkdir(path, { mode: 0o700 })));\n const authentication = await stageClaudeAuthentication(\n join(privateRoot, \"config\"),\n options.credentialFile,\n { binary: options.binary, lifetimeMs: loginLifetimeMs },\n );\n return {\n privateRoot,\n home,\n temp,\n authentication,\n credential: authentication.credentialFile,\n };\n}\n\nasync function protectedClaudeRoots(\n request: Request,\n options: ClaudeHostOptions,\n credential: string,\n) {\n const candidates = await evaluationProtectedRoots({\n ownedStateRoot: dirname(dirname(dirname(credential))),\n workspace: request.workspace,\n projectRoot: options.projectRoot,\n resultsRoot: options.resultsRoot,\n additionalRoots: [\n ...options.additionalProtectedRoots,\n ...claudeSeedSources(request),\n ],\n });\n const roots = request.runtimePolicy ? candidates : minimalRoots(candidates);\n const namespace = await nativeStateParent();\n if (roots.some((root) => root !== namespace && inside(root, credential)))\n throw new Error(\"Claude private state overlaps a protected root\");\n const binary = await realpath(options.binary);\n const executableRoots = runtimeExecutableProtection(\n [dirname(binary)],\n roots,\n Boolean(request.runtimePolicy),\n );\n if (executableRoots.some((root) => inside(root, binary)))\n throw new Error(\"Claude executable resides inside a protected root\");\n return roots;\n}\n\nasync function claudeToolchainDirectory(\n options: ClaudeHostOptions,\n roots: string[],\n): Promise<string | null> {\n const directory = options.toolchainBinDir\n ? await realpath(options.toolchainBinDir)\n : null;\n if (\n directory &&\n (roots.some((root) => inside(root, directory)) ||\n !(await stat(directory)).isDirectory())\n )\n throw new Error(\"Claude toolchain directory is unavailable or protected\");\n return directory;\n}\n\nasync function prepareClaudeRuntime(\n options: ClaudeHostOptions,\n roots: string[],\n workspace: string,\n): Promise<string | null> {\n if (!options.uvCacheDir) return null;\n const cache = await realpath(options.uvCacheDir);\n if (\n roots.some((root) => inside(root, cache)) ||\n !(await stat(cache)).isDirectory()\n )\n throw new Error(\"Claude UV cache is unavailable or protected\");\n const runtimeRoot = join(workspace, \".git\", \"sevro-runtime\");\n await mkdir(runtimeRoot, { recursive: true, mode: 0o700 });\n await cp(cache, join(runtimeRoot, \"uv-cache\"), {\n recursive: true,\n force: false,\n errorOnExist: true,\n });\n return runtimeRoot;\n}\n\nfunction claudeEnvironment(\n state: Awaited<ReturnType<typeof createClaudePrivateState>>,\n request: Request,\n toolchain: string | null,\n runtime: string | null,\n): Record<string, string> {\n return {\n ...state.authentication.environment,\n ...request.runtimePolicy?.environment,\n CLAUDE_CODE_SUBPROCESS_ENV_SCRUB: \"1\",\n PATH: claudePath(request, toolchain),\n LANG: process.env.LANG ?? \"C\",\n HOME: runtime ? join(runtime, \"host-home\") : state.home,\n TMPDIR: state.temp,\n CLAUDE_CONFIG_DIR: dirname(state.credential),\n NO_COLOR: \"1\",\n ...claudeRuntimeEnvironment(runtime),\n };\n}\n\nfunction claudePath(request: Request, toolchain: string | null): string {\n const path =\n request.runtimePolicy?.environment.PATH ??\n process.env.PATH ??\n \"/usr/bin:/bin\";\n return [request.fixtureBinDir, toolchain, path]\n .filter(Boolean)\n .join(delimiter);\n}\n\nfunction claudeRuntimeEnvironment(\n runtime: string | null,\n): Record<string, string> {\n return runtime\n ? {\n UV_CACHE_DIR: join(runtime, \"uv-cache\"),\n UV_PROJECT_ENVIRONMENT: join(runtime, \"project-environment\"),\n UV_OFFLINE: \"1\",\n PYTHONDONTWRITEBYTECODE: \"1\",\n }\n : {};\n}\n\n/** `lifetimeMs` is the access-token lifetime the complete execution needs. */\nexport async function prepareClaudeState(\n stateRoot: string,\n options: ClaudeHostOptions,\n request: Request,\n pluginDirs: string[],\n lifetimeMs: number,\n) {\n await requireClaudeRuntimeNamespace(request, stateRoot);\n const state = await createClaudePrivateState(stateRoot, options, lifetimeMs);\n const transcripts = await claudeTranscriptAccess(\n stateRoot,\n state.credential,\n request,\n );\n const roots = await protectedClaudeRoots(request, options, state.credential);\n const toolchain = await claudeToolchainDirectory(options, roots);\n const runtime = await prepareClaudeRuntime(options, roots, request.workspace);\n const settingsPath = join(state.privateRoot, \"settings.json\");\n const hooks = await claudeHookMounts(\n request,\n stateRoot,\n state.privateRoot,\n pluginDirs,\n roots,\n );\n const effectivePlugins = hooks?.directories ?? pluginDirs;\n const env = claudeEnvironment(state, request, toolchain, runtime);\n const runtimeWriteRoot = await applyClaudeRuntimeEnvironment(\n request,\n env,\n runtime,\n stateRoot,\n );\n const cacheRoots = await sharedCacheRoots(\n request.runtimePolicy,\n request.runtimeRole,\n roots,\n );\n await writeFile(\n settingsPath,\n JSON.stringify(\n claudeHostSettings(\n state.privateRoot,\n state.credential,\n [...pluginDirs, ...effectivePlugins],\n [\n ...roots,\n ...sharedCacheDenyRoots(request.runtimePolicy, request.runtimeRole),\n ],\n transcripts.policy,\n [\n ...transcripts.readRoots,\n ...candidateTranscriptRoots(request.candidateTranscriptRoot),\n ...candidateArtifactReadRoots(request.candidateArtifactRoots),\n ],\n runtimeWriteRoot\n ? { root: runtimeWriteRoot, workspace: request.workspace, cacheRoots }\n : undefined,\n request.workspace,\n ),\n ),\n { flag: \"wx\", mode: 0o600 },\n );\n Object.assign(env, transcripts.environment);\n Object.assign(\n env,\n candidateTranscriptEnvironment(request.candidateTranscriptRoot),\n candidateArtifactEnvironment(request.candidateArtifactRoots),\n );\n if (runtime) await mkdir(join(runtime, \"host-home\"), { mode: 0o700 });\n return {\n credential: state.credential,\n settingsPath,\n env,\n hooks,\n pluginDirs: effectivePlugins,\n };\n}\n\n/** Claude's implicit /tmp/claude writes cannot contain the read-denied peer namespace. */\nasync function requireClaudeRuntimeNamespace(\n request: Request,\n stateRoot: string,\n) {\n if (\n !request.runtimePolicy &&\n !workspaceBoundaryException(workspaceStatePath(), request.workspace).length\n )\n return;\n const implicitRoots = await Promise.all(\n [\"/tmp/claude\", \"/private/tmp/claude\"].map(implicitClaudeWritePaths),\n );\n const namespace = dirname(stateRoot);\n if (\n implicitRoots\n .flat()\n .some((root) => inside(root, namespace) || inside(namespace, root))\n )\n throw new Error(\n \"Claude namespace overlaps an implicit native writable root\",\n );\n}\n\nasync function implicitClaudeWritePaths(root: string): Promise<string[]> {\n try {\n return [root, await realpath(root)];\n } catch (cause) {\n if ((cause as NodeJS.ErrnoException).code !== \"ENOENT\")\n throw new Error(\"Claude implicit native writable root is unreadable\", {\n cause,\n });\n return [root];\n }\n}\n\nasync function claudeTranscriptAccess(\n stateRoot: string,\n credential: string,\n request: Request,\n) {\n const policy = request.runtimePolicy;\n if (!policy?.nativeTranscripts)\n return { policy, environment: {}, readRoots: [] };\n const view = join(stateRoot, \"native-home\");\n const projects = join(view, \"projects\");\n await mkdir(projects, { recursive: true, mode: 0o700 });\n await symlink(projects, join(dirname(credential), \"projects\"));\n const startup = join(view, \"shell-env\");\n const script = `export CLAUDE_CONFIG_DIR=${shellQuote(view)}\\n`;\n await writeFile(startup, script, { flag: \"wx\", mode: 0o600 });\n await writeFile(join(view, \".zshenv\"), script, { flag: \"wx\", mode: 0o600 });\n return {\n policy,\n readRoots: [projects, view],\n environment: {\n SEVRO_NATIVE_TRANSCRIPT_ROOT: projects,\n BASH_ENV: startup,\n ZDOTDIR: view,\n },\n };\n}\n\nasync function claudeHookMounts(\n request: Request,\n stateRoot: string,\n privateRoot: string,\n pluginDirs: string[],\n roots: string[],\n) {\n const policy = request.runtimePolicy;\n if (!policy || !runtimeHooksEnabled(policy)) return undefined;\n return prepareHookMounts({\n workspace: request.workspace,\n stateRoot,\n privateRoot,\n pluginRoots: pluginDirs,\n protectedRoots: [privateRoot, ...roots, ...runtimeSeedSources(policy)],\n policy,\n });\n}\n\nasync function applyClaudeRuntimeEnvironment(\n request: Request,\n env: Record<string, string>,\n runtime: string | null,\n stateRoot: string,\n): Promise<string | undefined> {\n const isolated = await prepareRuntimeState(\n request.workspace,\n claudeCommandPolicy(request),\n request.runtimeRole ?? \"candidate\",\n claudeRuntimeDirectory(request, stateRoot),\n );\n if (request.runtimePolicy) {\n Object.assign(env, isolated.environment);\n env.PATH = [request.fixtureBinDir, isolated.environment.PATH ?? env.PATH]\n .filter(Boolean)\n .join(delimiter);\n return isolated.root;\n }\n applyLegacyClaudeCommandEnvironment(env, isolated, runtime);\n return undefined;\n}\n\nfunction claudeRuntimeDirectory(request: Request, stateRoot: string) {\n return request.runtimePolicy\n ? (request.candidateRuntimeRoot ?? join(stateRoot, \"runtime\"))\n : undefined;\n}\n\nfunction applyLegacyClaudeCommandEnvironment(\n env: Record<string, string>,\n isolated: Awaited<ReturnType<typeof prepareRuntimeState>>,\n runtime: string | null,\n): void {\n if (!runtime) env.HOME = isolated.home;\n env.TMPDIR = isolated.temp;\n}\n\nfunction claudeCommandPolicy(request: Request) {\n return (\n request.runtimePolicy ?? {\n format: \"sevro.runtime.v1\" as const,\n environment: {},\n readOnlyRoots: [],\n }\n );\n}\n"]}
|
|
@@ -0,0 +1,150 @@
|
|
|
1
|
+
import { sha256 } from "../identity.js";
|
|
2
|
+
import { isRecord } from "../value-guards.js";
|
|
3
|
+
import { summarizeClaudeEvents, parseClaudeEvents, } from "./claude-events.js";
|
|
4
|
+
const MAX_RETAINED_CALLS = 128;
|
|
5
|
+
function label(value, max = 256) {
|
|
6
|
+
return typeof value === "string" && boundedLabel(value, max) ? value : null;
|
|
7
|
+
}
|
|
8
|
+
/** Retain ordered host call metadata, never tool prompts or results. */
|
|
9
|
+
export function claudeToolCallsObservation(stream, exitCode) {
|
|
10
|
+
const events = parseClaudeEvents(stream);
|
|
11
|
+
const summary = summarizeClaudeEvents(events, exitCode);
|
|
12
|
+
const state = {
|
|
13
|
+
calls: [],
|
|
14
|
+
malformed: false,
|
|
15
|
+
truncated: false,
|
|
16
|
+
ordinal: 0,
|
|
17
|
+
};
|
|
18
|
+
for (const line of events.lines) {
|
|
19
|
+
collectEvent(line, state);
|
|
20
|
+
if (state.truncated)
|
|
21
|
+
break;
|
|
22
|
+
}
|
|
23
|
+
return {
|
|
24
|
+
id: "sevro.claude.tool-calls",
|
|
25
|
+
completeness: callCompleteness(summary.complete, state),
|
|
26
|
+
data: {
|
|
27
|
+
method: "stream_tool_calls",
|
|
28
|
+
calls: state.calls,
|
|
29
|
+
truncated: state.truncated,
|
|
30
|
+
},
|
|
31
|
+
};
|
|
32
|
+
}
|
|
33
|
+
function callCompleteness(complete, state) {
|
|
34
|
+
return complete && !state.malformed && !state.truncated
|
|
35
|
+
? "complete"
|
|
36
|
+
: "partial";
|
|
37
|
+
}
|
|
38
|
+
function collectEvent(line, state) {
|
|
39
|
+
if (!line.parsed) {
|
|
40
|
+
state.malformed = true;
|
|
41
|
+
return;
|
|
42
|
+
}
|
|
43
|
+
const event = line.value;
|
|
44
|
+
if (!isRecord(event) || event.type !== "assistant")
|
|
45
|
+
return;
|
|
46
|
+
collectAssistant(event, state);
|
|
47
|
+
}
|
|
48
|
+
function collectAssistant(event, state) {
|
|
49
|
+
const blocks = isRecord(event.message)
|
|
50
|
+
? event.message.content
|
|
51
|
+
: null;
|
|
52
|
+
if (!Array.isArray(blocks)) {
|
|
53
|
+
state.malformed = true;
|
|
54
|
+
return;
|
|
55
|
+
}
|
|
56
|
+
const actor = actorContext(event, state);
|
|
57
|
+
for (const block of blocks) {
|
|
58
|
+
collectBlock(block, actor, state);
|
|
59
|
+
if (state.truncated)
|
|
60
|
+
break;
|
|
61
|
+
}
|
|
62
|
+
}
|
|
63
|
+
function actorContext(event, state) {
|
|
64
|
+
const parentToolUseId = label(event.parent_tool_use_id, 128);
|
|
65
|
+
const actor = parentToolUseId ? "nested" : "parent";
|
|
66
|
+
if (event.parent_tool_use_id !== undefined &&
|
|
67
|
+
event.parent_tool_use_id !== null &&
|
|
68
|
+
actor === "parent")
|
|
69
|
+
state.malformed = true;
|
|
70
|
+
return { actor, parentToolUseId };
|
|
71
|
+
}
|
|
72
|
+
function collectBlock(value, actor, state) {
|
|
73
|
+
if (!isRecord(value) || value.type !== "tool_use")
|
|
74
|
+
return;
|
|
75
|
+
state.ordinal++;
|
|
76
|
+
if (!["Skill", "Agent", "Task"].includes(String(value.name)))
|
|
77
|
+
return;
|
|
78
|
+
if (state.calls.length >= MAX_RETAINED_CALLS) {
|
|
79
|
+
state.truncated = true;
|
|
80
|
+
return;
|
|
81
|
+
}
|
|
82
|
+
collectSupportedCall(value, actor, state);
|
|
83
|
+
}
|
|
84
|
+
function collectSupportedCall(value, actor, state) {
|
|
85
|
+
const input = isRecord(value.input) ? value.input : {};
|
|
86
|
+
if (value.name === "Skill") {
|
|
87
|
+
collectSkill(input, actor, state);
|
|
88
|
+
return;
|
|
89
|
+
}
|
|
90
|
+
const call = agentCall(value, input, actor, state.ordinal);
|
|
91
|
+
if (!validAgentCall(call, input))
|
|
92
|
+
state.malformed = true;
|
|
93
|
+
state.calls.push(call);
|
|
94
|
+
}
|
|
95
|
+
function collectSkill(input, actor, state) {
|
|
96
|
+
const invocation = label(input.skill);
|
|
97
|
+
if (!invocation) {
|
|
98
|
+
state.malformed = true;
|
|
99
|
+
return;
|
|
100
|
+
}
|
|
101
|
+
state.calls.push({
|
|
102
|
+
ordinal: state.ordinal,
|
|
103
|
+
...actor,
|
|
104
|
+
name: "Skill",
|
|
105
|
+
skill: invocation.slice(invocation.lastIndexOf(":") + 1),
|
|
106
|
+
invocation,
|
|
107
|
+
});
|
|
108
|
+
}
|
|
109
|
+
function boundedPrompt(input) {
|
|
110
|
+
if (typeof input.prompt !== "string")
|
|
111
|
+
return null;
|
|
112
|
+
return Buffer.byteLength(input.prompt, "utf8") <= 1024 * 1024
|
|
113
|
+
? input.prompt
|
|
114
|
+
: null;
|
|
115
|
+
}
|
|
116
|
+
function promptDigest(prompt) {
|
|
117
|
+
return prompt ? sha256(prompt) : null;
|
|
118
|
+
}
|
|
119
|
+
function firstLineDigest(prompt) {
|
|
120
|
+
return prompt ? sha256(prompt.split(/\r?\n/, 1).join("")) : null;
|
|
121
|
+
}
|
|
122
|
+
function agentCall(value, input, actor, ordinal) {
|
|
123
|
+
const prompt = boundedPrompt(input);
|
|
124
|
+
return {
|
|
125
|
+
ordinal,
|
|
126
|
+
...actor,
|
|
127
|
+
name: "Agent",
|
|
128
|
+
toolUseId: label(value.id, 128),
|
|
129
|
+
subagentType: label(input.subagent_type ?? input.subagentType),
|
|
130
|
+
runInBackground: typeof input.run_in_background === "boolean"
|
|
131
|
+
? input.run_in_background
|
|
132
|
+
: null,
|
|
133
|
+
model: input.model === undefined ? null : label(input.model, 128),
|
|
134
|
+
promptSha256: promptDigest(prompt),
|
|
135
|
+
promptFirstLineSha256: firstLineDigest(prompt),
|
|
136
|
+
};
|
|
137
|
+
}
|
|
138
|
+
function validAgentCall(call, input) {
|
|
139
|
+
const required = [call.toolUseId, call.subagentType, call.promptSha256];
|
|
140
|
+
return (required.every(Boolean) &&
|
|
141
|
+
call.runInBackground !== null &&
|
|
142
|
+
validAgentModel(input, call.model));
|
|
143
|
+
}
|
|
144
|
+
function validAgentModel(input, model) {
|
|
145
|
+
return input.model === undefined || model !== null;
|
|
146
|
+
}
|
|
147
|
+
function boundedLabel(value, max) {
|
|
148
|
+
return (value.length >= 1 && value.length <= max && /^[A-Za-z0-9_.:-]+$/.test(value));
|
|
149
|
+
}
|
|
150
|
+
//# sourceMappingURL=claude-tool-calls.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"claude-tool-calls.js","sourceRoot":"","sources":["claude-tool-calls.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,MAAM,EAAE,uBAAoB;AACrC,OAAO,EAAE,QAAQ,EAAE,2BAAwB;AAC3C,OAAO,EACL,qBAAqB,EACrB,iBAAiB,GAGlB,2BAAwB;AAEzB,MAAM,kBAAkB,GAAG,GAAG,CAAC;AAG/B,SAAS,KAAK,CAAC,KAAc,EAAE,GAAG,GAAG,GAAG;IACtC,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,YAAY,CAAC,KAAK,EAAE,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,IAAI,CAAC;AAC9E,CAAC;AA0CD,wEAAwE;AACxE,MAAM,UAAU,0BAA0B,CACxC,MAA6B,EAC7B,QAAgB;IAEhB,MAAM,MAAM,GAAG,iBAAiB,CAAC,MAAM,CAAC,CAAC;IACzC,MAAM,OAAO,GAAG,qBAAqB,CAAC,MAAM,EAAE,QAAQ,CAAC,CAAC;IACxD,MAAM,KAAK,GAAc;QACvB,KAAK,EAAE,EAAE;QACT,SAAS,EAAE,KAAK;QAChB,SAAS,EAAE,KAAK;QAChB,OAAO,EAAE,CAAC;KACX,CAAC;IACF,KAAK,MAAM,IAAI,IAAI,MAAM,CAAC,KAAK,EAAE,CAAC;QAChC,YAAY,CAAC,IAAI,EAAE,KAAK,CAAC,CAAC;QAC1B,IAAI,KAAK,CAAC,SAAS;YAAE,MAAM;IAC7B,CAAC;IACD,OAAO;QACL,EAAE,EAAE,yBAAyB;QAC7B,YAAY,EAAE,gBAAgB,CAAC,OAAO,CAAC,QAAQ,EAAE,KAAK,CAAC;QACvD,IAAI,EAAE;YACJ,MAAM,EAAE,mBAAmB;YAC3B,KAAK,EAAE,KAAK,CAAC,KAAK;YAClB,SAAS,EAAE,KAAK,CAAC,SAAS;SAC3B;KACF,CAAC;AACJ,CAAC;AAED,SAAS,gBAAgB,CACvB,QAAiB,EACjB,KAAgB;IAEhB,OAAO,QAAQ,IAAI,CAAC,KAAK,CAAC,SAAS,IAAI,CAAC,KAAK,CAAC,SAAS;QACrD,CAAC,CAAC,UAAU;QACZ,CAAC,CAAC,SAAS,CAAC;AAChB,CAAC;AAED,SAAS,YAAY,CAAC,IAAqB,EAAE,KAAgB;IAC3D,IAAI,CAAC,IAAI,CAAC,MAAM,EAAE,CAAC;QACjB,KAAK,CAAC,SAAS,GAAG,IAAI,CAAC;QACvB,OAAO;IACT,CAAC;IACD,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC;IACzB,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,CAAC,IAAI,KAAK,WAAW;QAAE,OAAO;IAC3D,gBAAgB,CAAC,KAAK,EAAE,KAAK,CAAC,CAAC;AACjC,CAAC;AAED,SAAS,gBAAgB,CAAC,KAAY,EAAE,KAAgB;IACtD,MAAM,MAAM,GAAY,QAAQ,CAAC,KAAK,CAAC,OAAO,CAAC;QAC7C,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,OAAO;QACvB,CAAC,CAAC,IAAI,CAAC;IACT,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,EAAE,CAAC;QAC3B,KAAK,CAAC,SAAS,GAAG,IAAI,CAAC;QACvB,OAAO;IACT,CAAC;IACD,MAAM,KAAK,GAAG,YAAY,CAAC,KAAK,EAAE,KAAK,CAAC,CAAC;IACzC,KAAK,MAAM,KAAK,IAAI,MAAM,EAAE,CAAC;QAC3B,YAAY,CAAC,KAAK,EAAE,KAAK,EAAE,KAAK,CAAC,CAAC;QAClC,IAAI,KAAK,CAAC,SAAS;YAAE,MAAM;IAC7B,CAAC;AACH,CAAC;AAED,SAAS,YAAY,CAAC,KAAY,EAAE,KAAgB;IAClD,MAAM,eAAe,GAAG,KAAK,CAAC,KAAK,CAAC,kBAAkB,EAAE,GAAG,CAAC,CAAC;IAC7D,MAAM,KAAK,GAAG,eAAe,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,QAAQ,CAAC;IACpD,IACE,KAAK,CAAC,kBAAkB,KAAK,SAAS;QACtC,KAAK,CAAC,kBAAkB,KAAK,IAAI;QACjC,KAAK,KAAK,QAAQ;QAElB,KAAK,CAAC,SAAS,GAAG,IAAI,CAAC;IACzB,OAAO,EAAE,KAAK,EAAE,eAAe,EAAE,CAAC;AACpC,CAAC;AAED,SAAS,YAAY,CAAC,KAAc,EAAE,KAAY,EAAE,KAAgB;IAClE,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU;QAAE,OAAO;IAC1D,KAAK,CAAC,OAAO,EAAE,CAAC;IAChB,IAAI,CAAC,CAAC,OAAO,EAAE,OAAO,EAAE,MAAM,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;QAAE,OAAO;IACrE,IAAI,KAAK,CAAC,KAAK,CAAC,MAAM,IAAI,kBAAkB,EAAE,CAAC;QAC7C,KAAK,CAAC,SAAS,GAAG,IAAI,CAAC;QACvB,OAAO;IACT,CAAC;IACD,oBAAoB,CAAC,KAAK,EAAE,KAAK,EAAE,KAAK,CAAC,CAAC;AAC5C,CAAC;AAED,SAAS,oBAAoB,CAC3B,KAAY,EACZ,KAAY,EACZ,KAAgB;IAEhB,MAAM,KAAK,GAAG,QAAQ,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC;IACvD,IAAI,KAAK,CAAC,IAAI,KAAK,OAAO,EAAE,CAAC;QAC3B,YAAY,CAAC,KAAK,EAAE,KAAK,EAAE,KAAK,CAAC,CAAC;QAClC,OAAO;IACT,CAAC;IACD,MAAM,IAAI,GAAG,SAAS,CAAC,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,CAAC,OAAO,CAAC,CAAC;IAC3D,IAAI,CAAC,cAAc,CAAC,IAAI,EAAE,KAAK,CAAC;QAAE,KAAK,CAAC,SAAS,GAAG,IAAI,CAAC;IACzD,KAAK,CAAC,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AACzB,CAAC;AAED,SAAS,YAAY,CAAC,KAAY,EAAE,KAAY,EAAE,KAAgB;IAChE,MAAM,UAAU,GAAG,KAAK,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;IACtC,IAAI,CAAC,UAAU,EAAE,CAAC;QAChB,KAAK,CAAC,SAAS,GAAG,IAAI,CAAC;QACvB,OAAO;IACT,CAAC;IACD,KAAK,CAAC,KAAK,CAAC,IAAI,CAAC;QACf,OAAO,EAAE,KAAK,CAAC,OAAO;QACtB,GAAG,KAAK;QACR,IAAI,EAAE,OAAO;QACb,KAAK,EAAE,UAAU,CAAC,KAAK,CAAC,UAAU,CAAC,WAAW,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;QACxD,UAAU;KACX,CAAC,CAAC;AACL,CAAC;AAED,SAAS,aAAa,CAAC,KAAY;IACjC,IAAI,OAAO,KAAK,CAAC,MAAM,KAAK,QAAQ;QAAE,OAAO,IAAI,CAAC;IAClD,OAAO,MAAM,CAAC,UAAU,CAAC,KAAK,CAAC,MAAM,EAAE,MAAM,CAAC,IAAI,IAAI,GAAG,IAAI;QAC3D,CAAC,CAAC,KAAK,CAAC,MAAM;QACd,CAAC,CAAC,IAAI,CAAC;AACX,CAAC;AAED,SAAS,YAAY,CAAC,MAAqB;IACzC,OAAO,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC;AACxC,CAAC;AAED,SAAS,eAAe,CAAC,MAAqB;IAC5C,OAAO,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC,KAAK,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC;AACnE,CAAC;AAED,SAAS,SAAS,CAChB,KAAY,EACZ,KAAY,EACZ,KAAY,EACZ,OAAe;IAEf,MAAM,MAAM,GAAG,aAAa,CAAC,KAAK,CAAC,CAAC;IACpC,OAAO;QACL,OAAO;QACP,GAAG,KAAK;QACR,IAAI,EAAE,OAAO;QACb,SAAS,EAAE,KAAK,CAAC,KAAK,CAAC,EAAE,EAAE,GAAG,CAAC;QAC/B,YAAY,EAAE,KAAK,CAAC,KAAK,CAAC,aAAa,IAAI,KAAK,CAAC,YAAY,CAAC;QAC9D,eAAe,EACb,OAAO,KAAK,CAAC,iBAAiB,KAAK,SAAS;YAC1C,CAAC,CAAC,KAAK,CAAC,iBAAiB;YACzB,CAAC,CAAC,IAAI;QACV,KAAK,EAAE,KAAK,CAAC,KAAK,KAAK,SAAS,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,KAAK,CAAC,KAAK,CAAC,KAAK,EAAE,GAAG,CAAC;QACjE,YAAY,EAAE,YAAY,CAAC,MAAM,CAAC;QAClC,qBAAqB,EAAE,eAAe,CAAC,MAAM,CAAC;KAC/C,CAAC;AACJ,CAAC;AAED,SAAS,cAAc,CAAC,IAAe,EAAE,KAAY;IACnD,MAAM,QAAQ,GAAG,CAAC,IAAI,CAAC,SAAS,EAAE,IAAI,CAAC,YAAY,EAAE,IAAI,CAAC,YAAY,CAAC,CAAC;IACxE,OAAO,CACL,QAAQ,CAAC,KAAK,CAAC,OAAO,CAAC;QACvB,IAAI,CAAC,eAAe,KAAK,IAAI;QAC7B,eAAe,CAAC,KAAK,EAAE,IAAI,CAAC,KAAK,CAAC,CACnC,CAAC;AACJ,CAAC;AAED,SAAS,eAAe,CAAC,KAAY,EAAE,KAAoB;IACzD,OAAO,KAAK,CAAC,KAAK,KAAK,SAAS,IAAI,KAAK,KAAK,IAAI,CAAC;AACrD,CAAC;AAED,SAAS,YAAY,CAAC,KAAa,EAAE,GAAW;IAC9C,OAAO,CACL,KAAK,CAAC,MAAM,IAAI,CAAC,IAAI,KAAK,CAAC,MAAM,IAAI,GAAG,IAAI,oBAAoB,CAAC,IAAI,CAAC,KAAK,CAAC,CAC7E,CAAC;AACJ,CAAC","sourcesContent":["import { sha256 } from \"../identity\";\nimport { isRecord } from \"../value-guards\";\nimport {\n summarizeClaudeEvents,\n parseClaudeEvents,\n type ClaudeEvents,\n type ClaudeEventLine,\n} from \"./claude-events\";\n\nconst MAX_RETAINED_CALLS = 128;\ntype Entry = Record<string, unknown>;\n\nfunction label(value: unknown, max = 256): string | null {\n return typeof value === \"string\" && boundedLabel(value, max) ? value : null;\n}\n\ninterface SkillCall {\n ordinal: number;\n actor: \"parent\" | \"nested\";\n parentToolUseId: string | null;\n name: \"Skill\";\n skill: string;\n invocation: string;\n}\n\ninterface AgentCall {\n ordinal: number;\n actor: \"parent\" | \"nested\";\n parentToolUseId: string | null;\n name: \"Agent\";\n toolUseId: string | null;\n subagentType: string | null;\n runInBackground: boolean | null;\n model: string | null;\n promptSha256: string | null;\n promptFirstLineSha256: string | null;\n}\n\nexport interface ClaudeToolCallsObservation {\n id: \"sevro.claude.tool-calls\";\n completeness: \"complete\" | \"partial\";\n data: {\n method: \"stream_tool_calls\";\n calls: Array<SkillCall | AgentCall>;\n truncated: boolean;\n };\n}\n\ninterface CallState {\n calls: Array<SkillCall | AgentCall>;\n malformed: boolean;\n truncated: boolean;\n ordinal: number;\n}\ntype Actor = { actor: \"parent\" | \"nested\"; parentToolUseId: string | null };\n\n/** Retain ordered host call metadata, never tool prompts or results. */\nexport function claudeToolCallsObservation(\n stream: string | ClaudeEvents,\n exitCode: number,\n): ClaudeToolCallsObservation {\n const events = parseClaudeEvents(stream);\n const summary = summarizeClaudeEvents(events, exitCode);\n const state: CallState = {\n calls: [],\n malformed: false,\n truncated: false,\n ordinal: 0,\n };\n for (const line of events.lines) {\n collectEvent(line, state);\n if (state.truncated) break;\n }\n return {\n id: \"sevro.claude.tool-calls\",\n completeness: callCompleteness(summary.complete, state),\n data: {\n method: \"stream_tool_calls\",\n calls: state.calls,\n truncated: state.truncated,\n },\n };\n}\n\nfunction callCompleteness(\n complete: boolean,\n state: CallState,\n): \"complete\" | \"partial\" {\n return complete && !state.malformed && !state.truncated\n ? \"complete\"\n : \"partial\";\n}\n\nfunction collectEvent(line: ClaudeEventLine, state: CallState): void {\n if (!line.parsed) {\n state.malformed = true;\n return;\n }\n const event = line.value;\n if (!isRecord(event) || event.type !== \"assistant\") return;\n collectAssistant(event, state);\n}\n\nfunction collectAssistant(event: Entry, state: CallState): void {\n const blocks: unknown = isRecord(event.message)\n ? event.message.content\n : null;\n if (!Array.isArray(blocks)) {\n state.malformed = true;\n return;\n }\n const actor = actorContext(event, state);\n for (const block of blocks) {\n collectBlock(block, actor, state);\n if (state.truncated) break;\n }\n}\n\nfunction actorContext(event: Entry, state: CallState): Actor {\n const parentToolUseId = label(event.parent_tool_use_id, 128);\n const actor = parentToolUseId ? \"nested\" : \"parent\";\n if (\n event.parent_tool_use_id !== undefined &&\n event.parent_tool_use_id !== null &&\n actor === \"parent\"\n )\n state.malformed = true;\n return { actor, parentToolUseId };\n}\n\nfunction collectBlock(value: unknown, actor: Actor, state: CallState): void {\n if (!isRecord(value) || value.type !== \"tool_use\") return;\n state.ordinal++;\n if (![\"Skill\", \"Agent\", \"Task\"].includes(String(value.name))) return;\n if (state.calls.length >= MAX_RETAINED_CALLS) {\n state.truncated = true;\n return;\n }\n collectSupportedCall(value, actor, state);\n}\n\nfunction collectSupportedCall(\n value: Entry,\n actor: Actor,\n state: CallState,\n): void {\n const input = isRecord(value.input) ? value.input : {};\n if (value.name === \"Skill\") {\n collectSkill(input, actor, state);\n return;\n }\n const call = agentCall(value, input, actor, state.ordinal);\n if (!validAgentCall(call, input)) state.malformed = true;\n state.calls.push(call);\n}\n\nfunction collectSkill(input: Entry, actor: Actor, state: CallState): void {\n const invocation = label(input.skill);\n if (!invocation) {\n state.malformed = true;\n return;\n }\n state.calls.push({\n ordinal: state.ordinal,\n ...actor,\n name: \"Skill\",\n skill: invocation.slice(invocation.lastIndexOf(\":\") + 1),\n invocation,\n });\n}\n\nfunction boundedPrompt(input: Entry): string | null {\n if (typeof input.prompt !== \"string\") return null;\n return Buffer.byteLength(input.prompt, \"utf8\") <= 1024 * 1024\n ? input.prompt\n : null;\n}\n\nfunction promptDigest(prompt: string | null): string | null {\n return prompt ? sha256(prompt) : null;\n}\n\nfunction firstLineDigest(prompt: string | null): string | null {\n return prompt ? sha256(prompt.split(/\\r?\\n/, 1).join(\"\")) : null;\n}\n\nfunction agentCall(\n value: Entry,\n input: Entry,\n actor: Actor,\n ordinal: number,\n): AgentCall {\n const prompt = boundedPrompt(input);\n return {\n ordinal,\n ...actor,\n name: \"Agent\",\n toolUseId: label(value.id, 128),\n subagentType: label(input.subagent_type ?? input.subagentType),\n runInBackground:\n typeof input.run_in_background === \"boolean\"\n ? input.run_in_background\n : null,\n model: input.model === undefined ? null : label(input.model, 128),\n promptSha256: promptDigest(prompt),\n promptFirstLineSha256: firstLineDigest(prompt),\n };\n}\n\nfunction validAgentCall(call: AgentCall, input: Entry): boolean {\n const required = [call.toolUseId, call.subagentType, call.promptSha256];\n return (\n required.every(Boolean) &&\n call.runInBackground !== null &&\n validAgentModel(input, call.model)\n );\n}\n\nfunction validAgentModel(input: Entry, model: string | null): boolean {\n return input.model === undefined || model !== null;\n}\n\nfunction boundedLabel(value: string, max: number): boolean {\n return (\n value.length >= 1 && value.length <= max && /^[A-Za-z0-9_.:-]+$/.test(value)\n );\n}\n"]}
|
|
@@ -0,0 +1,373 @@
|
|
|
1
|
+
import { existsSync } from "node:fs";
|
|
2
|
+
import { readFile, realpath, rm, stat } from "node:fs/promises";
|
|
3
|
+
import { dirname, isAbsolute, join } from "node:path";
|
|
4
|
+
import { allocateNativeState } from "../native-transcript-state.js";
|
|
5
|
+
import { failedNativeTranscripts, retainEnabledNativeTranscripts, } from "../native-transcripts.js";
|
|
6
|
+
import { fixtureParts } from "../fixtures/preparation.js";
|
|
7
|
+
import { releaseRuntimeHooks, runtimeHookObservation, runtimeHooksEnabled, } from "../runtime/hooks.js";
|
|
8
|
+
import { insideRuntimeRoot as inside } from "../runtime/paths.js";
|
|
9
|
+
import { runtimeObservations } from "../runtime/state.js";
|
|
10
|
+
import { isRecord } from "../value-guards.js";
|
|
11
|
+
import { runClaudeTurns } from "./claude-continuation.js";
|
|
12
|
+
import { validateNativeHostRequest, validHostTimeout } from "./native-request.js";
|
|
13
|
+
import { summarizeClaudeEvents } from "./claude-events.js";
|
|
14
|
+
import { observeClaudeNativeGoal } from "./claude-native-goal.js";
|
|
15
|
+
import { runClaudeProcess } from "./claude-process.js";
|
|
16
|
+
import { claudeRepositoryInvocationObservation, verifyClaudeRepositoryInvocation, } from "./claude-repository-invocation.js";
|
|
17
|
+
import { prepareClaudeState } from "./claude-state.js";
|
|
18
|
+
import { cleanupNativeState, createHostDiagnostics, diagnoseHostRequest, diagnosticConfiguration, } from "../host-diagnostics.js";
|
|
19
|
+
import { CLAUDE_AUTH_VARIABLES, CLAUDE_LOGIN_MARGIN_MS, } from "./claude-credential.js";
|
|
20
|
+
import { claudeToolCallsObservation } from "./claude-tool-calls.js";
|
|
21
|
+
import { claudeSkillActivationsObservation } from "./claude-skill-activations.js";
|
|
22
|
+
import { claudeNativeControls } from "./native-controls.js";
|
|
23
|
+
const DEFAULT_TIMEOUT_MS = 10 * 60_000;
|
|
24
|
+
async function pluginDirectories(workspace, declaration) {
|
|
25
|
+
const root = await realpath(workspace);
|
|
26
|
+
requireUniquePluginDeclaration(declaration);
|
|
27
|
+
const paths = new Set(declaration.artifactPaths);
|
|
28
|
+
const directories = [];
|
|
29
|
+
for (const relativeRoot of declaration.artifactRoots)
|
|
30
|
+
directories.push(await declaredPluginDirectory(root, relativeRoot, paths));
|
|
31
|
+
requireDisjointPluginDirectories(directories);
|
|
32
|
+
for (const artifactPath of paths)
|
|
33
|
+
await verifyPluginArtifact(root, artifactPath, declaration, directories);
|
|
34
|
+
return directories;
|
|
35
|
+
}
|
|
36
|
+
function requireUniquePluginDeclaration(declaration) {
|
|
37
|
+
if (new Set(declaration.artifactRoots).size !==
|
|
38
|
+
declaration.artifactRoots.length ||
|
|
39
|
+
new Set(declaration.artifactPaths).size !== declaration.artifactPaths.length)
|
|
40
|
+
throw new Error("Claude plugin declaration contains duplicates");
|
|
41
|
+
}
|
|
42
|
+
async function declaredPluginDirectory(root, relativeRoot, paths) {
|
|
43
|
+
fixtureParts(relativeRoot);
|
|
44
|
+
if (!paths.has(`${relativeRoot}/.claude-plugin/plugin.json`))
|
|
45
|
+
throw new Error("Claude plugin manifest is undeclared");
|
|
46
|
+
const actual = await realpath(join(root, relativeRoot));
|
|
47
|
+
if (!inside(root, actual))
|
|
48
|
+
throw new Error("Claude plugin directory escapes workspace");
|
|
49
|
+
if (!(await stat(join(actual, ".claude-plugin", "plugin.json"))).isFile())
|
|
50
|
+
throw new Error("Claude plugin manifest is missing");
|
|
51
|
+
return actual;
|
|
52
|
+
}
|
|
53
|
+
function requireDisjointPluginDirectories(directories) {
|
|
54
|
+
if (new Set(directories).size !== directories.length)
|
|
55
|
+
throw new Error("Claude plugin directories overlap");
|
|
56
|
+
if (directories.some((item) => directories.some((other) => item !== other && inside(other, item))))
|
|
57
|
+
throw new Error("Claude plugin directories overlap");
|
|
58
|
+
}
|
|
59
|
+
async function verifyPluginArtifact(root, artifactPath, declaration, directories) {
|
|
60
|
+
fixtureParts(artifactPath);
|
|
61
|
+
const index = declaration.artifactRoots.findIndex((item) => artifactPath.startsWith(`${item}/`));
|
|
62
|
+
const directory = directories.at(index);
|
|
63
|
+
if (index < 0 || directory === undefined)
|
|
64
|
+
throw new Error("Claude plugin artifact is outside a declared directory");
|
|
65
|
+
if (!inside(directory, await realpath(join(root, artifactPath))))
|
|
66
|
+
throw new Error("Claude plugin artifact escapes workspace");
|
|
67
|
+
}
|
|
68
|
+
function pluginSelection(request) {
|
|
69
|
+
const selected = request.explicitSkillInvocation;
|
|
70
|
+
if (!selected || selected.scope === "repository")
|
|
71
|
+
return null;
|
|
72
|
+
return selected;
|
|
73
|
+
}
|
|
74
|
+
function requirePluginInvocation(selected, prompt) {
|
|
75
|
+
const { token, pluginName, skillName } = selected;
|
|
76
|
+
if (token !== `/${pluginName}:${skillName}` ||
|
|
77
|
+
prompt.split(token).length - 1 !== 1 ||
|
|
78
|
+
!/^[a-z][a-z0-9-]*$/.test(pluginName) ||
|
|
79
|
+
!/^[A-Za-z0-9._-]+$/.test(skillName))
|
|
80
|
+
throw new Error("invalid Claude explicit skill invocation");
|
|
81
|
+
}
|
|
82
|
+
function skillPathDeclared(request, index, skillName) {
|
|
83
|
+
const relativeRoot = request.claudePluginDirs?.artifactRoots[index];
|
|
84
|
+
return !!request.claudePluginDirs?.artifactPaths.includes(`${relativeRoot}/skills/${skillName}/SKILL.md`);
|
|
85
|
+
}
|
|
86
|
+
async function verifyInvocation(request, directories) {
|
|
87
|
+
const selected = pluginSelection(request);
|
|
88
|
+
if (!selected)
|
|
89
|
+
return;
|
|
90
|
+
requirePluginInvocation(selected, request.prompt);
|
|
91
|
+
for (const [index, root] of directories.entries()) {
|
|
92
|
+
if (!skillPathDeclared(request, index, selected.skillName))
|
|
93
|
+
continue;
|
|
94
|
+
if (await mountedPluginSkill(root, selected))
|
|
95
|
+
return;
|
|
96
|
+
}
|
|
97
|
+
throw new Error("invoked Claude skill is absent from the package");
|
|
98
|
+
}
|
|
99
|
+
async function mountedPluginSkill(root, selected) {
|
|
100
|
+
let manifest;
|
|
101
|
+
try {
|
|
102
|
+
manifest = JSON.parse(await readFile(join(root, ".claude-plugin", "plugin.json"), "utf8"));
|
|
103
|
+
}
|
|
104
|
+
catch (error) {
|
|
105
|
+
throw new Error("invalid Claude plugin manifest", { cause: error });
|
|
106
|
+
}
|
|
107
|
+
if (!isRecord(manifest) || manifest.name !== selected.pluginName)
|
|
108
|
+
return false;
|
|
109
|
+
const skill = await stat(join(root, "skills", selected.skillName, "SKILL.md")).catch(() => null);
|
|
110
|
+
return !!skill?.isFile();
|
|
111
|
+
}
|
|
112
|
+
function validClaudeOptions(options, timeoutMs) {
|
|
113
|
+
return (availableClaudeSandbox() &&
|
|
114
|
+
requiredClaudePaths(options) &&
|
|
115
|
+
optionalClaudePaths(options) &&
|
|
116
|
+
!!options.model &&
|
|
117
|
+
validClaudeEffortTimeout(options.effort, timeoutMs));
|
|
118
|
+
}
|
|
119
|
+
function availableClaudeSandbox() {
|
|
120
|
+
if (process.platform === "darwin")
|
|
121
|
+
return existsSync("/usr/bin/sandbox-exec");
|
|
122
|
+
return (process.platform === "linux" &&
|
|
123
|
+
existsSync("/usr/bin/bwrap") &&
|
|
124
|
+
existsSync("/usr/bin/socat"));
|
|
125
|
+
}
|
|
126
|
+
function requiredClaudePaths(options) {
|
|
127
|
+
return [
|
|
128
|
+
options.binary,
|
|
129
|
+
options.projectRoot,
|
|
130
|
+
options.resultsRoot,
|
|
131
|
+
...options.additionalProtectedRoots,
|
|
132
|
+
].every(isAbsolute);
|
|
133
|
+
}
|
|
134
|
+
function optionalClaudePaths(options) {
|
|
135
|
+
return [
|
|
136
|
+
options.credentialFile,
|
|
137
|
+
options.uvCacheDir,
|
|
138
|
+
options.toolchainBinDir,
|
|
139
|
+
].every((path) => path === undefined || isAbsolute(path));
|
|
140
|
+
}
|
|
141
|
+
function validClaudeEffortTimeout(effort, timeoutMs) {
|
|
142
|
+
return !!effort && validHostTimeout(timeoutMs);
|
|
143
|
+
}
|
|
144
|
+
/** Run Claude with explicit plugins and its native tool sandbox. */
|
|
145
|
+
export function createClaudeHost(options) {
|
|
146
|
+
const timeoutMs = options.timeoutMs ?? DEFAULT_TIMEOUT_MS;
|
|
147
|
+
if (!validClaudeOptions(options, timeoutMs))
|
|
148
|
+
throw new Error("invalid Claude host configuration");
|
|
149
|
+
return {
|
|
150
|
+
id: "sevro.host.claude",
|
|
151
|
+
model: options.model,
|
|
152
|
+
effort: options.effort,
|
|
153
|
+
...(options.retainHostDiagnostics
|
|
154
|
+
? {
|
|
155
|
+
configuration: diagnosticConfiguration(options.retainHostDiagnostics),
|
|
156
|
+
}
|
|
157
|
+
: {}),
|
|
158
|
+
hostCapabilities: [
|
|
159
|
+
"sevro.host.native-goal",
|
|
160
|
+
"sevro.host.runtime",
|
|
161
|
+
"sevro.host.native-transcripts",
|
|
162
|
+
"sevro.host.hooks",
|
|
163
|
+
"sevro.host.skill-activations",
|
|
164
|
+
"sevro.claude.plugin-dirs",
|
|
165
|
+
"sevro.claude.explicit-invocation",
|
|
166
|
+
...(options.projectSettings
|
|
167
|
+
? ["sevro.claude.repository-invocation"]
|
|
168
|
+
: []),
|
|
169
|
+
"sevro.claude.tool-calls",
|
|
170
|
+
"sevro.host.native-controls",
|
|
171
|
+
"sevro.host.continuation",
|
|
172
|
+
"sevro.claude.continuation",
|
|
173
|
+
],
|
|
174
|
+
run(request) {
|
|
175
|
+
return runClaudeRequest(options, timeoutMs, request);
|
|
176
|
+
},
|
|
177
|
+
};
|
|
178
|
+
}
|
|
179
|
+
async function repositoryMount(request, options) {
|
|
180
|
+
if (request.explicitSkillInvocation?.scope !== "repository")
|
|
181
|
+
return null;
|
|
182
|
+
if (!options.projectSettings)
|
|
183
|
+
throw new Error("Claude repository invocation requires project setting sources");
|
|
184
|
+
return verifyClaudeRepositoryInvocation(request);
|
|
185
|
+
}
|
|
186
|
+
async function runClaudeRequest(options, timeoutMs, request) {
|
|
187
|
+
const diagnostics = createHostDiagnostics("claude", options.retainHostDiagnostics, request);
|
|
188
|
+
return diagnoseHostRequest(diagnostics, async () => {
|
|
189
|
+
diagnostics.stage = "request-validation";
|
|
190
|
+
await diagnostics.readCredentials(options.credentialFile);
|
|
191
|
+
return executeClaudeRequest(options, timeoutMs, request, diagnostics);
|
|
192
|
+
});
|
|
193
|
+
}
|
|
194
|
+
/** Every turn shares one access token, so it must outlive all of them. */
|
|
195
|
+
function claudeLoginLifetime(timeoutMs, request) {
|
|
196
|
+
const turns = request.followUpPrompt === undefined ? 1 : 2;
|
|
197
|
+
return timeoutMs * turns + CLAUDE_LOGIN_MARGIN_MS;
|
|
198
|
+
}
|
|
199
|
+
async function executeClaudeRequest(options, timeoutMs, request, diagnostics) {
|
|
200
|
+
const { pluginDirs, repositoryInvocation } = await validateClaudeRequest(options, request);
|
|
201
|
+
diagnostics.stage = "state-allocation";
|
|
202
|
+
const stateRoot = await allocateNativeState("claude-", request.runtimePolicy ? request.workspace : undefined);
|
|
203
|
+
let hooks;
|
|
204
|
+
try {
|
|
205
|
+
diagnostics.stage = "state-preparation";
|
|
206
|
+
const state = await prepareClaudeState(stateRoot, options, request, pluginDirs, claudeLoginLifetime(timeoutMs, request));
|
|
207
|
+
diagnostics.registerCredentials(state.env, CLAUDE_AUTH_VARIABLES);
|
|
208
|
+
hooks = state.hooks;
|
|
209
|
+
const run = claudeProcessRunner(options, timeoutMs, request, state, diagnostics);
|
|
210
|
+
const execution = await runClaudeTurns({
|
|
211
|
+
prompt: request.prompt,
|
|
212
|
+
followUpPrompt: request.followUpPrompt,
|
|
213
|
+
workspace: request.workspace,
|
|
214
|
+
run,
|
|
215
|
+
});
|
|
216
|
+
const result = await claudeResult(execution, claudeObservationRoot(stateRoot, state.credential, request), request, repositoryInvocation);
|
|
217
|
+
await diagnostics.retainFailure(result, "Claude host execution did not complete");
|
|
218
|
+
await addHookObservation(result, state.hooks);
|
|
219
|
+
await retainEnabledNativeTranscripts(result, request.runtimePolicy, join(stateRoot, "native-home", "projects"), "claude", request.runtimeRole);
|
|
220
|
+
return result;
|
|
221
|
+
}
|
|
222
|
+
catch (cause) {
|
|
223
|
+
return await failedNativeTranscripts(cause, request.runtimePolicy, join(stateRoot, "native-home", "projects"), "claude", request.runtimeRole, diagnostics);
|
|
224
|
+
}
|
|
225
|
+
finally {
|
|
226
|
+
await cleanupNativeState(diagnostics, async () => {
|
|
227
|
+
await releaseRuntimeHooks(hooks);
|
|
228
|
+
await rm(stateRoot, { recursive: true, force: true });
|
|
229
|
+
});
|
|
230
|
+
}
|
|
231
|
+
}
|
|
232
|
+
function claudeProcessRunner(options, timeoutMs, request, state, diagnostics) {
|
|
233
|
+
return (prompt, session) => {
|
|
234
|
+
diagnostics.stage = "claude-execution";
|
|
235
|
+
return runClaudeProcess({
|
|
236
|
+
diagnostics,
|
|
237
|
+
argv: claudeTurnArgs(options, prompt, session, state.settingsPath, state.pluginDirs),
|
|
238
|
+
cwd: request.workspace,
|
|
239
|
+
env: state.env,
|
|
240
|
+
timeoutMs,
|
|
241
|
+
signal: request.signal,
|
|
242
|
+
});
|
|
243
|
+
};
|
|
244
|
+
}
|
|
245
|
+
async function validateClaudeRequest(options, request) {
|
|
246
|
+
await validateNativeHostRequest(request, "Claude");
|
|
247
|
+
const pluginDirs = request.claudePluginDirs
|
|
248
|
+
? await pluginDirectories(request.workspace, request.claudePluginDirs)
|
|
249
|
+
: [];
|
|
250
|
+
await verifyInvocation(request, pluginDirs);
|
|
251
|
+
requireIsolatedHookSources(request, options);
|
|
252
|
+
const repositoryInvocation = await repositoryMount(request, options);
|
|
253
|
+
return { pluginDirs, repositoryInvocation };
|
|
254
|
+
}
|
|
255
|
+
function requireIsolatedHookSources(request, options) {
|
|
256
|
+
if (runtimeHooksEnabled(request.runtimePolicy) && options.projectSettings)
|
|
257
|
+
throw new Error("native goals require isolated hook sources");
|
|
258
|
+
}
|
|
259
|
+
async function addHookObservation(result, hooks) {
|
|
260
|
+
if (!hooks)
|
|
261
|
+
return;
|
|
262
|
+
result.observations?.push(await runtimeHookObservation(hooks));
|
|
263
|
+
}
|
|
264
|
+
function claudeTurnArgs(options, prompt, session, settingsPath, pluginDirs) {
|
|
265
|
+
return [
|
|
266
|
+
options.binary,
|
|
267
|
+
"-p",
|
|
268
|
+
prompt,
|
|
269
|
+
"--output-format",
|
|
270
|
+
"stream-json",
|
|
271
|
+
"--verbose",
|
|
272
|
+
"--model",
|
|
273
|
+
options.model,
|
|
274
|
+
"--effort",
|
|
275
|
+
options.effort,
|
|
276
|
+
"--permission-mode",
|
|
277
|
+
"dontAsk",
|
|
278
|
+
"--tools",
|
|
279
|
+
"Bash,Read,Edit,Skill,Agent",
|
|
280
|
+
"--setting-sources",
|
|
281
|
+
options.projectSettings ? "project" : "",
|
|
282
|
+
"--settings",
|
|
283
|
+
settingsPath,
|
|
284
|
+
"--strict-mcp-config",
|
|
285
|
+
"--mcp-config",
|
|
286
|
+
'{"mcpServers":{}}',
|
|
287
|
+
"--no-chrome",
|
|
288
|
+
...(session ? [session.option, session.id] : []),
|
|
289
|
+
...pluginDirs.flatMap((path) => ["--plugin-dir", path]),
|
|
290
|
+
];
|
|
291
|
+
}
|
|
292
|
+
function claudeObservationRoot(stateRoot, credential, request) {
|
|
293
|
+
return request.runtimePolicy?.nativeTranscripts === true
|
|
294
|
+
? join(stateRoot, "native-home")
|
|
295
|
+
: dirname(credential);
|
|
296
|
+
}
|
|
297
|
+
async function repositoryObservation(execution, configRoot, request, invocation, tools) {
|
|
298
|
+
if (!invocation)
|
|
299
|
+
return null;
|
|
300
|
+
return claudeRepositoryInvocationObservation({
|
|
301
|
+
invocation,
|
|
302
|
+
stream: execution.initialEvents ?? execution.events,
|
|
303
|
+
configRoot: configRoot,
|
|
304
|
+
workspace: request.workspace,
|
|
305
|
+
tools: execution.initialEvents
|
|
306
|
+
? claudeToolCallsObservation(execution.initialEvents, 0)
|
|
307
|
+
: tools,
|
|
308
|
+
});
|
|
309
|
+
}
|
|
310
|
+
async function claudeObservations(execution, configRoot, request, invocation) {
|
|
311
|
+
const tools = claudeToolCallsObservation(execution.events, execution.code);
|
|
312
|
+
const repository = await repositoryObservation(execution, configRoot, request, invocation, tools);
|
|
313
|
+
return [
|
|
314
|
+
tools,
|
|
315
|
+
...runtimeObservations(request.runtimePolicy, request.runtimeRole),
|
|
316
|
+
await observeClaudeNativeGoal(configRoot, execution.followUpEvents ?? execution.events),
|
|
317
|
+
claudeNativeControls(execution.events, execution.code),
|
|
318
|
+
...(repository ? [repository] : []),
|
|
319
|
+
...(execution.continuation ? [execution.continuation] : []),
|
|
320
|
+
claudeSkillActivationsObservation(execution.events, execution.code),
|
|
321
|
+
];
|
|
322
|
+
}
|
|
323
|
+
function claudeArtifacts(execution) {
|
|
324
|
+
return [
|
|
325
|
+
{ id: "sevro.claude.events", bytes: Buffer.from(execution.out, "utf8") },
|
|
326
|
+
...(execution.initialOut !== undefined
|
|
327
|
+
? [
|
|
328
|
+
{
|
|
329
|
+
id: "sevro.claude.initial-events",
|
|
330
|
+
bytes: Buffer.from(execution.initialOut, "utf8"),
|
|
331
|
+
},
|
|
332
|
+
]
|
|
333
|
+
: []),
|
|
334
|
+
...(execution.followUpOut !== undefined
|
|
335
|
+
? [
|
|
336
|
+
{
|
|
337
|
+
id: "sevro.claude.follow-up-events",
|
|
338
|
+
bytes: Buffer.from(execution.followUpOut, "utf8"),
|
|
339
|
+
},
|
|
340
|
+
]
|
|
341
|
+
: []),
|
|
342
|
+
...(execution.err
|
|
343
|
+
? [
|
|
344
|
+
{
|
|
345
|
+
id: "sevro.claude.stderr",
|
|
346
|
+
bytes: Buffer.from(execution.err, "utf8"),
|
|
347
|
+
},
|
|
348
|
+
]
|
|
349
|
+
: []),
|
|
350
|
+
];
|
|
351
|
+
}
|
|
352
|
+
function measuredClaudeUsage(summary, execution) {
|
|
353
|
+
return {
|
|
354
|
+
inputTokens: execution.sessionResultsBound === false ? null : summary.inputTokens,
|
|
355
|
+
outputTokens: execution.sessionResultsBound === false ? null : summary.outputTokens,
|
|
356
|
+
costUsd: execution.sessionResultsBound === false ? null : summary.costUsd,
|
|
357
|
+
usageComplete: execution.sessionResultsBound !== false && summary.usageComplete,
|
|
358
|
+
};
|
|
359
|
+
}
|
|
360
|
+
async function claudeResult(execution, configRoot, request, invocation) {
|
|
361
|
+
const summary = summarizeClaudeEvents(execution.events, execution.code);
|
|
362
|
+
const observations = await claudeObservations(execution, configRoot, request, invocation);
|
|
363
|
+
return {
|
|
364
|
+
finalMessage: summary.finalMessage,
|
|
365
|
+
complete: summary.complete,
|
|
366
|
+
executionFailed: !summary.complete,
|
|
367
|
+
observations,
|
|
368
|
+
artifacts: claudeArtifacts(execution),
|
|
369
|
+
actualCondition: "passive",
|
|
370
|
+
...measuredClaudeUsage(summary, execution),
|
|
371
|
+
};
|
|
372
|
+
}
|
|
373
|
+
//# sourceMappingURL=claude.js.map
|