voratiq 0.1.0-beta.2 → 0.1.0-beta.21
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/README.md +41 -29
- package/dist/agents/launch/chat.d.ts +23 -0
- package/dist/agents/launch/chat.js +44 -0
- package/dist/agents/launch/environment.d.ts +8 -0
- package/dist/{commands/run/agents/workspace-prep.js → agents/launch/environment.js} +5 -27
- package/dist/agents/launch/prompt.d.ts +6 -0
- package/dist/agents/launch/prompt.js +12 -0
- package/dist/agents/launch/provider-state.d.ts +39 -0
- package/dist/agents/launch/provider-state.js +103 -0
- package/dist/agents/runtime/auth.d.ts +27 -0
- package/dist/agents/runtime/auth.js +72 -0
- package/dist/agents/runtime/chat.d.ts +5 -0
- package/dist/agents/runtime/chat.js +7 -0
- package/dist/agents/runtime/errors.d.ts +27 -0
- package/dist/agents/runtime/errors.js +51 -0
- package/dist/{commands/run/agents → agents/runtime}/failures.d.ts +0 -1
- package/dist/agents/runtime/failures.js +136 -0
- package/dist/agents/runtime/harness.d.ts +2 -0
- package/dist/agents/runtime/harness.js +119 -0
- package/dist/{commands/run/agents/sandbox-launcher.d.ts → agents/runtime/launcher.d.ts} +18 -6
- package/dist/{commands/run/agents/sandbox-launcher.js → agents/runtime/launcher.js} +17 -39
- package/dist/{commands/run/agents/workspace-prep.d.ts → agents/runtime/manifest.d.ts} +6 -6
- package/dist/agents/runtime/manifest.js +34 -0
- package/dist/agents/runtime/policy.d.ts +32 -0
- package/dist/agents/runtime/policy.js +240 -0
- package/dist/agents/runtime/registry.d.ts +4 -0
- package/dist/agents/runtime/registry.js +54 -0
- package/dist/{commands/run → agents/runtime}/sandbox.d.ts +8 -2
- package/dist/{commands/run → agents/runtime}/sandbox.js +28 -67
- package/dist/agents/runtime/shim/run-agent-shim.d.ts +1 -0
- package/dist/agents/runtime/shim/run-agent-shim.js +276 -0
- package/dist/agents/runtime/types.d.ts +91 -0
- package/dist/{commands/run/agents → agents/runtime}/watchdog.d.ts +4 -3
- package/dist/{commands/run/agents → agents/runtime}/watchdog.js +155 -26
- package/dist/auth/providers/codex.js +7 -2
- package/dist/auth/providers/gemini.js +14 -6
- package/dist/auth/providers/types.d.ts +1 -0
- package/dist/auth/providers/utils.d.ts +0 -1
- package/dist/auth/providers/utils.js +1 -49
- package/dist/bin.js +369 -71
- package/dist/cli/apply.d.ts +4 -0
- package/dist/cli/apply.js +28 -9
- package/dist/cli/auto.d.ts +32 -0
- package/dist/cli/auto.js +232 -0
- package/dist/cli/contract.d.ts +328 -0
- package/dist/cli/contract.js +480 -0
- package/dist/cli/errors.d.ts +3 -0
- package/dist/cli/errors.js +21 -3
- package/dist/cli/init.d.ts +5 -0
- package/dist/cli/init.js +34 -6
- package/dist/cli/list.d.ts +6 -4
- package/dist/cli/list.js +39 -16
- package/dist/cli/mcp.d.ts +2 -0
- package/dist/cli/mcp.js +16 -0
- package/dist/cli/message.d.ts +28 -0
- package/dist/cli/message.js +147 -0
- package/dist/cli/operator-envelope.d.ts +180 -0
- package/dist/cli/operator-envelope.js +425 -0
- package/dist/cli/output.d.ts +15 -1
- package/dist/cli/output.js +153 -5
- package/dist/cli/prune.d.ts +7 -3
- package/dist/cli/prune.js +57 -12
- package/dist/cli/reduce.d.ts +29 -0
- package/dist/cli/reduce.js +211 -0
- package/dist/cli/root-launcher.d.ts +4 -0
- package/dist/cli/root-launcher.js +15 -0
- package/dist/cli/run.d.ts +27 -1
- package/dist/cli/run.js +108 -16
- package/dist/cli/spec.d.ts +31 -0
- package/dist/cli/spec.js +180 -0
- package/dist/cli/verify.d.ts +35 -0
- package/dist/cli/verify.js +297 -0
- package/dist/commands/apply/command.d.ts +2 -0
- package/dist/commands/apply/command.js +145 -6
- package/dist/commands/apply/errors.d.ts +43 -4
- package/dist/commands/apply/errors.js +100 -22
- package/dist/commands/apply/types.d.ts +2 -1
- package/dist/commands/auto/command.d.ts +145 -0
- package/dist/commands/auto/command.js +433 -0
- package/dist/commands/auto/errors.d.ts +19 -0
- package/dist/commands/auto/errors.js +19 -0
- package/dist/commands/auto/validation.d.ts +14 -0
- package/dist/commands/auto/validation.js +90 -0
- package/dist/commands/fetch.d.ts +2 -2
- package/dist/commands/fetch.js +4 -4
- package/dist/commands/init/agents.d.ts +2 -1
- package/dist/commands/init/agents.js +66 -63
- package/dist/commands/init/command.js +300 -16
- package/dist/commands/init/types.d.ts +18 -7
- package/dist/commands/interactive/lifecycle.d.ts +15 -0
- package/dist/commands/interactive/lifecycle.js +141 -0
- package/dist/commands/list/command.d.ts +10 -3
- package/dist/commands/list/command.js +597 -40
- package/dist/commands/message/command.d.ts +23 -0
- package/dist/commands/message/command.js +215 -0
- package/dist/commands/message/errors.d.ts +9 -0
- package/dist/commands/message/errors.js +20 -0
- package/dist/commands/message/lifecycle.d.ts +14 -0
- package/dist/commands/message/lifecycle.js +128 -0
- package/dist/commands/prune/command.d.ts +2 -1
- package/dist/commands/prune/command.js +61 -10
- package/dist/commands/prune/errors.d.ts +1 -1
- package/dist/commands/prune/errors.js +5 -5
- package/dist/commands/prune/types.d.ts +21 -0
- package/dist/commands/reduce/command.d.ts +26 -0
- package/dist/commands/reduce/command.js +145 -0
- package/dist/commands/reduce/errors.d.ts +17 -0
- package/dist/commands/reduce/errors.js +32 -0
- package/dist/commands/reduce/targets.d.ts +11 -0
- package/dist/commands/reduce/targets.js +271 -0
- package/dist/commands/root-launcher/command.d.ts +31 -0
- package/dist/commands/root-launcher/command.js +233 -0
- package/dist/commands/run/command.d.ts +6 -1
- package/dist/commands/run/command.js +70 -40
- package/dist/commands/run/lifecycle.d.ts +7 -5
- package/dist/commands/run/lifecycle.js +45 -23
- package/dist/commands/run/record-init.d.ts +4 -1
- package/dist/commands/run/record-init.js +5 -2
- package/dist/commands/run/shim/run-agent-shim.d.ts +2 -1
- package/dist/commands/run/shim/run-agent-shim.js +4 -219
- package/dist/commands/run/validation.d.ts +2 -3
- package/dist/commands/run/validation.js +54 -25
- package/dist/commands/shared/max-parallel.d.ts +5 -0
- package/dist/commands/shared/max-parallel.js +15 -0
- package/dist/commands/shared/preview.d.ts +10 -0
- package/dist/commands/shared/preview.js +60 -0
- package/dist/commands/shared/resolve-reduction-competitors.d.ts +15 -0
- package/dist/commands/shared/resolve-reduction-competitors.js +13 -0
- package/dist/commands/shared/resolve-stage-competitors.d.ts +19 -0
- package/dist/commands/shared/resolve-stage-competitors.js +171 -0
- package/dist/commands/shared/session-id.d.ts +1 -0
- package/dist/commands/shared/session-id.js +1 -0
- package/dist/commands/spec/command.d.ts +22 -0
- package/dist/commands/spec/command.js +330 -0
- package/dist/commands/spec/errors.d.ts +11 -0
- package/dist/commands/spec/errors.js +23 -0
- package/dist/commands/verify/agents.d.ts +8 -0
- package/dist/commands/verify/agents.js +29 -0
- package/dist/commands/verify/command.d.ts +23 -0
- package/dist/commands/verify/command.js +168 -0
- package/dist/commands/verify/lifecycle.d.ts +14 -0
- package/dist/commands/verify/lifecycle.js +229 -0
- package/dist/commands/verify/max-parallel.d.ts +7 -0
- package/dist/commands/verify/max-parallel.js +15 -0
- package/dist/commands/verify/targets.d.ts +18 -0
- package/dist/commands/verify/targets.js +420 -0
- package/dist/competition/command-adapter.d.ts +35 -0
- package/dist/competition/command-adapter.js +20 -0
- package/dist/competition/core.d.ts +41 -0
- package/dist/competition/core.js +181 -0
- package/dist/competition/shared/extra-context.d.ts +14 -0
- package/dist/competition/shared/extra-context.js +100 -0
- package/dist/competition/shared/preflight.d.ts +11 -0
- package/dist/competition/shared/preflight.js +39 -0
- package/dist/competition/shared/prompt-helpers.d.ts +16 -0
- package/dist/competition/shared/prompt-helpers.js +27 -0
- package/dist/competition/shared/prune.d.ts +1 -0
- package/dist/competition/shared/prune.js +4 -0
- package/dist/competition/shared/sandbox-policy.d.ts +9 -0
- package/dist/competition/shared/sandbox-policy.js +7 -0
- package/dist/competition/shared/teardown.d.ts +36 -0
- package/dist/competition/shared/teardown.js +101 -0
- package/dist/configs/agents/defaults.d.ts +31 -2
- package/dist/configs/agents/defaults.js +346 -30
- package/dist/configs/agents/errors.js +14 -11
- package/dist/configs/agents/loader.d.ts +11 -1
- package/dist/configs/agents/loader.js +71 -4
- package/dist/configs/agents/types.js +2 -2
- package/dist/configs/environment/detect.js +9 -4
- package/dist/configs/environment/errors.js +4 -4
- package/dist/configs/environment/loader.d.ts +1 -1
- package/dist/configs/environment/loader.js +3 -3
- package/dist/configs/orchestration/bootstrap.d.ts +16 -0
- package/dist/configs/orchestration/bootstrap.js +122 -0
- package/dist/configs/orchestration/errors.d.ts +15 -0
- package/dist/configs/orchestration/errors.js +28 -0
- package/dist/configs/orchestration/loader.d.ts +9 -0
- package/dist/configs/orchestration/loader.js +148 -0
- package/dist/configs/orchestration/types.d.ts +102 -0
- package/dist/configs/orchestration/types.js +65 -0
- package/dist/configs/sandbox/defaults.js +14 -4
- package/dist/configs/sandbox/errors.d.ts +1 -1
- package/dist/configs/sandbox/errors.js +1 -1
- package/dist/configs/sandbox/loader.js +6 -4
- package/dist/configs/sandbox/schemas.js +4 -2
- package/dist/configs/settings/loader.d.ts +7 -0
- package/dist/configs/settings/loader.js +81 -0
- package/dist/configs/settings/types.d.ts +47 -0
- package/dist/configs/settings/types.js +23 -0
- package/dist/configs/verification/errors.d.ts +11 -0
- package/dist/configs/verification/errors.js +21 -0
- package/dist/configs/verification/loader.d.ts +8 -0
- package/dist/configs/verification/loader.js +43 -0
- package/dist/configs/verification/methods.d.ts +35 -0
- package/dist/configs/verification/methods.js +41 -0
- package/dist/configs/verification/programmatic-defaults.d.ts +10 -0
- package/dist/configs/verification/programmatic-defaults.js +42 -0
- package/dist/configs/verification/programmatic-detect.d.ts +10 -0
- package/dist/configs/{evals/detect.js → verification/programmatic-detect.js} +22 -33
- package/dist/configs/verification/types.d.ts +49 -0
- package/dist/configs/verification/types.js +45 -0
- package/dist/contracts/list.d.ts +207 -0
- package/dist/contracts/list.js +154 -0
- package/dist/domain/interactive/model/types.d.ts +104 -0
- package/dist/domain/interactive/model/types.js +83 -0
- package/dist/domain/interactive/persistence/adapter.d.ts +39 -0
- package/dist/domain/interactive/persistence/adapter.js +144 -0
- package/dist/domain/interactive/prompt.d.ts +3 -0
- package/dist/domain/interactive/prompt.js +7 -0
- package/dist/domain/message/competition/adapter.d.ts +36 -0
- package/dist/domain/message/competition/adapter.js +197 -0
- package/dist/domain/message/competition/prompt.d.ts +8 -0
- package/dist/domain/message/competition/prompt.js +29 -0
- package/dist/domain/message/model/mutators.d.ts +17 -0
- package/dist/domain/message/model/mutators.js +107 -0
- package/dist/domain/message/model/types.d.ts +100 -0
- package/dist/domain/message/model/types.js +87 -0
- package/dist/domain/message/persistence/adapter.d.ts +43 -0
- package/dist/domain/message/persistence/adapter.js +124 -0
- package/dist/domain/reduce/competition/adapter.d.ts +42 -0
- package/dist/domain/reduce/competition/adapter.js +826 -0
- package/dist/domain/reduce/competition/output-validation.d.ts +4 -0
- package/dist/domain/reduce/competition/output-validation.js +18 -0
- package/dist/domain/reduce/competition/prompt.d.ts +10 -0
- package/dist/domain/reduce/competition/prompt.js +96 -0
- package/dist/domain/reduce/competition/reduction.d.ts +9 -0
- package/dist/domain/reduce/competition/reduction.js +32 -0
- package/dist/domain/reduce/model/types.d.ts +122 -0
- package/dist/domain/reduce/model/types.js +84 -0
- package/dist/domain/reduce/persistence/adapter.d.ts +43 -0
- package/dist/domain/reduce/persistence/adapter.js +126 -0
- package/dist/domain/run/competition/adapter.d.ts +30 -0
- package/dist/domain/run/competition/adapter.js +39 -0
- package/dist/domain/run/competition/agent-execution.d.ts +20 -0
- package/dist/domain/run/competition/agent-execution.js +45 -0
- package/dist/domain/run/competition/agent-preparation.d.ts +12 -0
- package/dist/domain/run/competition/agent-preparation.js +24 -0
- package/dist/domain/run/competition/agents/artifacts.d.ts +17 -0
- package/dist/domain/run/competition/agents/artifacts.js +173 -0
- package/dist/{commands/run → domain/run/competition}/agents/lifecycle.d.ts +3 -3
- package/dist/{commands/run → domain/run/competition}/agents/lifecycle.js +84 -64
- package/dist/domain/run/competition/agents/post-processing.d.ts +12 -0
- package/dist/domain/run/competition/agents/post-processing.js +4 -0
- package/dist/domain/run/competition/agents/preparation.js +64 -0
- package/dist/{commands/run → domain/run/competition}/agents/run-context.d.ts +9 -16
- package/dist/{commands/run → domain/run/competition}/agents/run-context.js +22 -70
- package/dist/{commands/run → domain/run/competition}/agents/types.d.ts +10 -13
- package/dist/domain/run/competition/agents/workspace.d.ts +21 -0
- package/dist/domain/run/competition/agents/workspace.js +47 -0
- package/dist/{commands/run → domain/run/competition}/errors.d.ts +8 -1
- package/dist/{commands/run → domain/run/competition}/errors.js +39 -9
- package/dist/{commands/run → domain/run/competition}/phases.d.ts +1 -2
- package/dist/domain/run/competition/phases.js +1 -0
- package/dist/domain/run/competition/prompt.d.ts +7 -0
- package/dist/domain/run/competition/prompt.js +27 -0
- package/dist/{commands/run → domain/run/competition}/reports.d.ts +5 -3
- package/dist/{commands/run → domain/run/competition}/reports.js +7 -19
- package/dist/domain/run/competition/termination-state.d.ts +4 -0
- package/dist/domain/run/competition/termination-state.js +12 -0
- package/dist/{records → domain/run/model}/enhanced.d.ts +6 -7
- package/dist/{records → domain/run/model}/enhanced.js +11 -11
- package/dist/{records → domain/run/model}/errors.d.ts +1 -1
- package/dist/{records → domain/run/model}/errors.js +5 -5
- package/dist/{records → domain/run/model}/mutators.d.ts +4 -3
- package/dist/{records → domain/run/model}/mutators.js +58 -36
- package/dist/domain/run/model/types.d.ts +376 -0
- package/dist/domain/run/model/types.js +192 -0
- package/dist/{records/persistence.d.ts → domain/run/persistence/adapter.d.ts} +9 -3
- package/dist/domain/run/persistence/adapter.js +340 -0
- package/dist/domain/run/persistence/error-mapping.d.ts +2 -0
- package/dist/domain/run/persistence/error-mapping.js +17 -0
- package/dist/domain/shared/lifecycle.d.ts +54 -0
- package/dist/domain/shared/lifecycle.js +165 -0
- package/dist/domain/shared/token-usage.d.ts +21 -0
- package/dist/domain/shared/token-usage.js +38 -0
- package/dist/domain/spec/competition/adapter.d.ts +31 -0
- package/dist/domain/spec/competition/adapter.js +196 -0
- package/dist/domain/spec/competition/prompt.d.ts +11 -0
- package/dist/domain/spec/competition/prompt.js +44 -0
- package/dist/domain/spec/model/output.d.ts +13 -0
- package/dist/domain/spec/model/output.js +36 -0
- package/dist/domain/spec/model/types.d.ts +98 -0
- package/dist/domain/spec/model/types.js +84 -0
- package/dist/domain/spec/persistence/adapter.d.ts +51 -0
- package/dist/domain/spec/persistence/adapter.js +140 -0
- package/dist/domain/verify/blinding/aliases.d.ts +7 -0
- package/dist/domain/verify/blinding/aliases.js +23 -0
- package/dist/domain/verify/competition/adapter.d.ts +54 -0
- package/dist/domain/verify/competition/adapter.js +444 -0
- package/dist/domain/verify/competition/artifacts.d.ts +6 -0
- package/dist/domain/verify/competition/artifacts.js +7 -0
- package/dist/domain/verify/competition/blinding.d.ts +24 -0
- package/dist/domain/verify/competition/blinding.js +109 -0
- package/dist/domain/verify/competition/finalize.d.ts +11 -0
- package/dist/domain/verify/competition/finalize.js +65 -0
- package/dist/domain/verify/competition/programmatic.d.ts +15 -0
- package/dist/domain/verify/competition/programmatic.js +352 -0
- package/dist/domain/verify/competition/prompt.d.ts +19 -0
- package/dist/domain/verify/competition/prompt.js +63 -0
- package/dist/domain/verify/competition/rubric.d.ts +23 -0
- package/dist/domain/verify/competition/rubric.js +77 -0
- package/dist/domain/verify/competition/shared-layout.d.ts +121 -0
- package/dist/domain/verify/competition/shared-layout.js +365 -0
- package/dist/domain/verify/competition/target.d.ts +47 -0
- package/dist/domain/verify/competition/target.js +1 -0
- package/dist/domain/verify/model/mutators.d.ts +16 -0
- package/dist/domain/verify/model/mutators.js +126 -0
- package/dist/domain/verify/model/types.d.ts +408 -0
- package/dist/domain/verify/model/types.js +289 -0
- package/dist/domain/verify/persistence/adapter.d.ts +43 -0
- package/dist/domain/verify/persistence/adapter.js +126 -0
- package/dist/domain/verify/programmatic/runner.d.ts +22 -0
- package/dist/domain/verify/programmatic/runner.js +209 -0
- package/dist/domain/verify/rubric-result.d.ts +28 -0
- package/dist/domain/verify/rubric-result.js +121 -0
- package/dist/extra-context/contract.d.ts +17 -0
- package/dist/extra-context/contract.js +60 -0
- package/dist/interactive/index.d.ts +2 -0
- package/dist/interactive/index.js +1 -0
- package/dist/interactive/providers/launch.d.ts +23 -0
- package/dist/interactive/providers/launch.js +203 -0
- package/dist/interactive/providers/mcp.d.ts +13 -0
- package/dist/interactive/providers/mcp.js +547 -0
- package/dist/interactive/providers/shared.d.ts +2 -0
- package/dist/interactive/providers/shared.js +1 -0
- package/dist/interactive/providers.d.ts +3 -0
- package/dist/interactive/providers.js +3 -0
- package/dist/interactive/records.d.ts +2 -0
- package/dist/interactive/records.js +1 -0
- package/dist/interactive/substrate.d.ts +21 -0
- package/dist/interactive/substrate.js +522 -0
- package/dist/interactive/types.d.ts +101 -0
- package/dist/interactive/types.js +1 -0
- package/dist/mcp/server.d.ts +88 -0
- package/dist/mcp/server.js +790 -0
- package/dist/persistence/error-mapping.d.ts +19 -0
- package/dist/persistence/error-mapping.js +44 -0
- package/dist/persistence/errors.d.ts +26 -0
- package/dist/persistence/errors.js +49 -0
- package/dist/persistence/extra-context.d.ts +9 -0
- package/dist/persistence/extra-context.js +60 -0
- package/dist/{records → persistence}/history-lock.js +2 -2
- package/dist/persistence/record-path-schema.d.ts +3 -0
- package/dist/persistence/record-path-schema.js +16 -0
- package/dist/persistence/session-store.d.ts +92 -0
- package/dist/persistence/session-store.js +412 -0
- package/dist/policy/auto.d.ts +13 -0
- package/dist/policy/auto.js +22 -0
- package/dist/policy/index.d.ts +5 -0
- package/dist/policy/index.js +5 -0
- package/dist/policy/resolution.d.ts +6 -0
- package/dist/policy/resolution.js +23 -0
- package/dist/policy/result.d.ts +53 -0
- package/dist/policy/result.js +15 -0
- package/dist/policy/selector.d.ts +11 -0
- package/dist/policy/selector.js +57 -0
- package/dist/policy/verification.d.ts +77 -0
- package/dist/policy/verification.js +365 -0
- package/dist/policy/verifier-selection.d.ts +13 -0
- package/dist/policy/verifier-selection.js +78 -0
- package/dist/preflight/branch.d.ts +9 -0
- package/dist/preflight/branch.js +48 -0
- package/dist/preflight/errors.d.ts +3 -0
- package/dist/preflight/errors.js +10 -3
- package/dist/preflight/index.d.ts +13 -0
- package/dist/preflight/index.js +43 -8
- package/dist/render/interactions/confirmation.js +4 -2
- package/dist/render/transcripts/apply.js +9 -10
- package/dist/render/transcripts/auto.d.ts +27 -0
- package/dist/render/transcripts/auto.js +21 -0
- package/dist/render/transcripts/init.d.ts +4 -15
- package/dist/render/transcripts/init.js +71 -72
- package/dist/render/transcripts/list.d.ts +10 -1
- package/dist/render/transcripts/list.js +121 -15
- package/dist/render/transcripts/message.d.ts +72 -0
- package/dist/render/transcripts/message.js +362 -0
- package/dist/render/transcripts/prune.d.ts +7 -2
- package/dist/render/transcripts/prune.js +64 -17
- package/dist/render/transcripts/reduce.d.ts +74 -0
- package/dist/render/transcripts/reduce.js +395 -0
- package/dist/render/transcripts/root-launcher.d.ts +19 -0
- package/dist/render/transcripts/root-launcher.js +40 -0
- package/dist/render/transcripts/run.d.ts +35 -6
- package/dist/render/transcripts/run.js +241 -165
- package/dist/render/transcripts/shared.d.ts +2 -0
- package/dist/render/transcripts/shared.js +11 -4
- package/dist/render/transcripts/spec.d.ts +74 -0
- package/dist/render/transcripts/spec.js +394 -0
- package/dist/render/transcripts/stage-progress.d.ts +22 -0
- package/dist/render/transcripts/stage-progress.js +6 -0
- package/dist/render/transcripts/update-check.d.ts +2 -0
- package/dist/render/transcripts/update-check.js +22 -0
- package/dist/render/transcripts/verify.d.ts +74 -0
- package/dist/render/transcripts/verify.js +409 -0
- package/dist/render/utils/agents.d.ts +10 -9
- package/dist/render/utils/agents.js +30 -82
- package/dist/render/utils/badges.d.ts +3 -20
- package/dist/render/utils/badges.js +3 -36
- package/dist/render/utils/duration.d.ts +12 -0
- package/dist/render/utils/duration.js +37 -0
- package/dist/render/utils/interactive-frame.d.ts +6 -0
- package/dist/render/utils/interactive-frame.js +38 -0
- package/dist/render/utils/records.js +4 -4
- package/dist/render/utils/runs.d.ts +3 -9
- package/dist/render/utils/runs.js +16 -48
- package/dist/render/utils/stage-output.d.ts +20 -0
- package/dist/render/utils/stage-output.js +44 -0
- package/dist/render/utils/timezone.d.ts +2 -0
- package/dist/render/utils/timezone.js +42 -0
- package/dist/render/utils/transcript-shell.d.ts +66 -0
- package/dist/render/utils/transcript-shell.js +155 -0
- package/dist/render/utils/transcript.d.ts +7 -1
- package/dist/render/utils/transcript.js +12 -2
- package/dist/render/utils/wrap.d.ts +1 -0
- package/dist/render/utils/wrap.js +20 -0
- package/dist/status/colors.d.ts +2 -3
- package/dist/status/colors.js +3 -3
- package/dist/status/index.d.ts +108 -8
- package/dist/status/index.js +164 -5
- package/dist/update-check/checker.d.ts +24 -0
- package/dist/update-check/checker.js +130 -0
- package/dist/update-check/prompt.d.ts +25 -0
- package/dist/update-check/prompt.js +62 -0
- package/dist/update-check/semver.d.ts +17 -0
- package/dist/update-check/semver.js +36 -0
- package/dist/update-check/state-path.d.ts +8 -0
- package/dist/update-check/state-path.js +18 -0
- package/dist/utils/binaries.js +14 -8
- package/dist/utils/errors.d.ts +3 -1
- package/dist/utils/errors.js +3 -1
- package/dist/utils/git.d.ts +10 -0
- package/dist/utils/git.js +15 -3
- package/dist/utils/output.d.ts +5 -1
- package/dist/utils/output.js +4 -2
- package/dist/utils/process.d.ts +2 -1
- package/dist/utils/process.js +7 -3
- package/dist/utils/session-id.d.ts +1 -0
- package/dist/utils/session-id.js +22 -0
- package/dist/utils/slug.d.ts +2 -0
- package/dist/utils/slug.js +15 -0
- package/dist/utils/voratiq-cli-target.d.ts +9 -0
- package/dist/utils/voratiq-cli-target.js +58 -0
- package/dist/workspace/agents.d.ts +13 -16
- package/dist/workspace/agents.js +22 -147
- package/dist/workspace/chat/artifacts.d.ts +9 -0
- package/dist/workspace/chat/artifacts.js +82 -12
- package/dist/workspace/chat/native-usage.d.ts +13 -0
- package/dist/workspace/chat/native-usage.js +60 -0
- package/dist/workspace/chat/sources.d.ts +9 -5
- package/dist/workspace/chat/sources.js +89 -23
- package/dist/workspace/chat/token-usage-result.d.ts +23 -0
- package/dist/workspace/chat/token-usage-result.js +7 -0
- package/dist/workspace/chat/usage-extractor.d.ts +30 -0
- package/dist/workspace/chat/usage-extractor.js +461 -0
- package/dist/workspace/chat/usage-mappings.d.ts +20 -0
- package/dist/workspace/chat/usage-mappings.js +136 -0
- package/dist/workspace/credential-guard.js +1 -1
- package/dist/workspace/dependencies.js +4 -4
- package/dist/workspace/errors.d.ts +5 -0
- package/dist/workspace/errors.js +13 -3
- package/dist/workspace/layout.d.ts +17 -6
- package/dist/workspace/layout.js +51 -32
- package/dist/workspace/promotion.d.ts +32 -0
- package/dist/workspace/promotion.js +34 -0
- package/dist/workspace/prune.d.ts +1 -1
- package/dist/workspace/run.d.ts +1 -3
- package/dist/workspace/run.js +6 -15
- package/dist/workspace/setup.d.ts +8 -0
- package/dist/workspace/setup.js +359 -56
- package/dist/workspace/shim.js +1 -1
- package/dist/workspace/structure.d.ts +91 -26
- package/dist/workspace/structure.js +227 -43
- package/dist/workspace/templates.d.ts +9 -3
- package/dist/workspace/templates.js +26 -15
- package/dist/workspace/verification-defaults.d.ts +12 -0
- package/dist/workspace/verification-defaults.js +1017 -0
- package/package.json +30 -24
- package/dist/cli/review.d.ts +0 -12
- package/dist/cli/review.js +0 -33
- package/dist/commands/errors.d.ts +0 -4
- package/dist/commands/errors.js +0 -7
- package/dist/commands/init/evals.d.ts +0 -4
- package/dist/commands/init/evals.js +0 -219
- package/dist/commands/review/command.d.ts +0 -10
- package/dist/commands/review/command.js +0 -26
- package/dist/commands/run/agent-execution.d.ts +0 -19
- package/dist/commands/run/agent-execution.js +0 -63
- package/dist/commands/run/agents/auth-stage.d.ts +0 -23
- package/dist/commands/run/agents/auth-stage.js +0 -108
- package/dist/commands/run/agents/chat-preserver.d.ts +0 -9
- package/dist/commands/run/agents/chat-preserver.js +0 -35
- package/dist/commands/run/agents/eval-runner.d.ts +0 -19
- package/dist/commands/run/agents/eval-runner.js +0 -27
- package/dist/commands/run/agents/failures.js +0 -32
- package/dist/commands/run/agents/preparation.js +0 -123
- package/dist/commands/run/agents.d.ts +0 -14
- package/dist/commands/run/agents.js +0 -47
- package/dist/commands/run/prompts.d.ts +0 -4
- package/dist/commands/run/prompts.js +0 -16
- package/dist/commands/run/sandbox-registry.d.ts +0 -4
- package/dist/commands/run/sandbox-registry.js +0 -54
- package/dist/configs/evals/defaults.d.ts +0 -8
- package/dist/configs/evals/defaults.js +0 -28
- package/dist/configs/evals/detect.d.ts +0 -10
- package/dist/configs/evals/errors.d.ts +0 -16
- package/dist/configs/evals/errors.js +0 -29
- package/dist/configs/evals/loader.d.ts +0 -9
- package/dist/configs/evals/loader.js +0 -46
- package/dist/configs/evals/types.d.ts +0 -42
- package/dist/configs/evals/types.js +0 -74
- package/dist/evals/runner.d.ts +0 -16
- package/dist/evals/runner.js +0 -132
- package/dist/records/persistence.js +0 -469
- package/dist/records/types.d.ts +0 -255
- package/dist/records/types.js +0 -160
- package/dist/render/transcripts/review.d.ts +0 -2
- package/dist/render/transcripts/review.js +0 -36
- /package/dist/{commands/run → agents/runtime}/shim/agent-manifest.d.ts +0 -0
- /package/dist/{commands/run → agents/runtime}/shim/agent-manifest.js +0 -0
- /package/dist/{commands/run → agents/runtime/shim}/argv.d.ts +0 -0
- /package/dist/{commands/run → agents/runtime/shim}/argv.js +0 -0
- /package/dist/{commands/run/agents → agents/runtime}/types.js +0 -0
- /package/dist/{commands/run → domain/run/competition}/agents/preparation.d.ts +0 -0
- /package/dist/{commands/run/phases.js → domain/run/competition/agents/types.js} +0 -0
- /package/dist/{commands/run → domain/run/model}/id.d.ts +0 -0
- /package/dist/{commands/run → domain/run/model}/id.js +0 -0
- /package/dist/{records → persistence}/history-lock.d.ts +0 -0
|
@@ -0,0 +1,444 @@
|
|
|
1
|
+
import { readFile } from "node:fs/promises";
|
|
2
|
+
import { join, resolve } from "node:path";
|
|
3
|
+
import { detectAgentProcessFailureDetail } from "../../../agents/runtime/failures.js";
|
|
4
|
+
import { runSandboxedAgent } from "../../../agents/runtime/harness.js";
|
|
5
|
+
import { stageExtraContextFiles, } from "../../../competition/shared/extra-context.js";
|
|
6
|
+
import { composeStageSandboxPolicy } from "../../../competition/shared/sandbox-policy.js";
|
|
7
|
+
import { buildUnavailableTokenUsageResult, resolveTokenUsage, } from "../../../domain/shared/token-usage.js";
|
|
8
|
+
import { emitStageProgressEvent } from "../../../render/transcripts/stage-progress.js";
|
|
9
|
+
import { toErrorMessage } from "../../../utils/errors.js";
|
|
10
|
+
import { pathExists } from "../../../utils/fs.js";
|
|
11
|
+
import { prepareScratchAgentWorkspace } from "../../../workspace/agents.js";
|
|
12
|
+
import { extractProviderNativeTokenUsageForSession } from "../../../workspace/chat/native-usage.js";
|
|
13
|
+
import { ensureWorkspaceDependencies } from "../../../workspace/dependencies.js";
|
|
14
|
+
import { buildScopedAgentWorkspacePaths, } from "../../../workspace/layout.js";
|
|
15
|
+
import { getVerificationRubricExecutionDirectoryPath, getVerificationRubricResultPath, VORATIQ_VERIFICATION_DIR, } from "../../../workspace/structure.js";
|
|
16
|
+
import { verificationResultArtifactSchema, } from "../model/types.js";
|
|
17
|
+
import { writeVerificationArtifact } from "./artifacts.js";
|
|
18
|
+
import { assertNoVerificationIdentityLeak, assertRubricResultSelectorsMatchAliasMap, buildForbiddenVerificationIdentityTokens, parseRubricResultPayload, } from "./blinding.js";
|
|
19
|
+
import { buildRubricPrompt } from "./prompt.js";
|
|
20
|
+
import { attachVerifierWorkspaceMounts, buildStagedVerificationInputs, sharedInputsUseReferenceRepo, } from "./shared-layout.js";
|
|
21
|
+
export function createVerifyCompetitionAdapter(input) {
|
|
22
|
+
const { root, verificationId, resolvedTarget, aliasMap, environment, extraContextFiles, sharedInputs, teardown, mutators, renderer, } = input;
|
|
23
|
+
const startedAtByExecutionKey = new Map();
|
|
24
|
+
const tokenUsageResultByExecutionKey = new Map();
|
|
25
|
+
function executionKeyForCandidate(candidate) {
|
|
26
|
+
return `${candidate.template.template}:${candidate.agent.id}`;
|
|
27
|
+
}
|
|
28
|
+
return {
|
|
29
|
+
prepareCandidates: (candidates) => ({
|
|
30
|
+
ready: candidates.map((candidate) => {
|
|
31
|
+
const workspacePaths = buildVerifierRubricWorkspacePaths({
|
|
32
|
+
root,
|
|
33
|
+
verificationId,
|
|
34
|
+
verifierId: candidate.agent.id,
|
|
35
|
+
template: candidate.template.template,
|
|
36
|
+
});
|
|
37
|
+
registerScratchWorkspaceTeardown(teardown, workspacePaths, candidate.agent.id, candidate.template.template);
|
|
38
|
+
return { candidate, workspacePaths };
|
|
39
|
+
}),
|
|
40
|
+
failures: [],
|
|
41
|
+
}),
|
|
42
|
+
queueCandidate: async (candidate) => {
|
|
43
|
+
await mutators.recordMethodSnapshot({
|
|
44
|
+
method: "rubric",
|
|
45
|
+
template: candidate.template.template,
|
|
46
|
+
verifierId: candidate.agent.id,
|
|
47
|
+
scope: resolvedTarget.target.kind === "run"
|
|
48
|
+
? { kind: "run" }
|
|
49
|
+
: { kind: "target" },
|
|
50
|
+
status: "queued",
|
|
51
|
+
});
|
|
52
|
+
emitStageProgressEvent(renderer, {
|
|
53
|
+
type: "stage.candidate",
|
|
54
|
+
stage: "verify",
|
|
55
|
+
candidate: {
|
|
56
|
+
methodKey: `rubric:${candidate.template.template}:${candidate.agent.id}`,
|
|
57
|
+
verifierLabel: candidate.template.template,
|
|
58
|
+
agentLabel: candidate.agent.id,
|
|
59
|
+
status: "queued",
|
|
60
|
+
},
|
|
61
|
+
});
|
|
62
|
+
},
|
|
63
|
+
onCandidateRunning: async (prepared) => {
|
|
64
|
+
const startedAt = new Date().toISOString();
|
|
65
|
+
const executionKey = executionKeyForCandidate(prepared.candidate);
|
|
66
|
+
startedAtByExecutionKey.set(executionKey, startedAt);
|
|
67
|
+
await mutators.recordMethodSnapshot({
|
|
68
|
+
method: "rubric",
|
|
69
|
+
template: prepared.candidate.template.template,
|
|
70
|
+
verifierId: prepared.candidate.agent.id,
|
|
71
|
+
scope: resolvedTarget.target.kind === "run"
|
|
72
|
+
? { kind: "run" }
|
|
73
|
+
: { kind: "target" },
|
|
74
|
+
status: "running",
|
|
75
|
+
startedAt,
|
|
76
|
+
});
|
|
77
|
+
emitStageProgressEvent(renderer, {
|
|
78
|
+
type: "stage.candidate",
|
|
79
|
+
stage: "verify",
|
|
80
|
+
candidate: {
|
|
81
|
+
methodKey: `rubric:${prepared.candidate.template.template}:${prepared.candidate.agent.id}`,
|
|
82
|
+
verifierLabel: prepared.candidate.template.template,
|
|
83
|
+
agentLabel: prepared.candidate.agent.id,
|
|
84
|
+
status: "running",
|
|
85
|
+
startedAt,
|
|
86
|
+
},
|
|
87
|
+
});
|
|
88
|
+
},
|
|
89
|
+
executeCandidate: async (prepared) => {
|
|
90
|
+
const executionKey = executionKeyForCandidate(prepared.candidate);
|
|
91
|
+
const startedAt = startedAtByExecutionKey.get(executionKey) ?? new Date().toISOString();
|
|
92
|
+
const { workspacePaths } = prepared;
|
|
93
|
+
await prepareScratchAgentWorkspace({ paths: workspacePaths });
|
|
94
|
+
await attachVerifierWorkspaceMounts({
|
|
95
|
+
workspacePath: workspacePaths.workspacePath,
|
|
96
|
+
contextPath: workspacePaths.contextPath,
|
|
97
|
+
sharedInputs,
|
|
98
|
+
});
|
|
99
|
+
await ensureWorkspaceDependencies({
|
|
100
|
+
root,
|
|
101
|
+
workspacePath: workspacePaths.workspacePath,
|
|
102
|
+
environment,
|
|
103
|
+
});
|
|
104
|
+
await stageExtraContextFiles({
|
|
105
|
+
contextPath: workspacePaths.contextPath,
|
|
106
|
+
files: extraContextFiles,
|
|
107
|
+
});
|
|
108
|
+
const staged = buildStagedVerificationInputs({
|
|
109
|
+
workspacePaths,
|
|
110
|
+
sharedInputs,
|
|
111
|
+
});
|
|
112
|
+
const prompt = buildRubricPrompt({
|
|
113
|
+
template: prepared.candidate.template,
|
|
114
|
+
target: resolvedTarget.target,
|
|
115
|
+
staged,
|
|
116
|
+
extraContextFiles,
|
|
117
|
+
});
|
|
118
|
+
assertNoVerificationIdentityLeak({
|
|
119
|
+
text: prompt,
|
|
120
|
+
forbidden: buildForbiddenVerificationIdentityTokens({
|
|
121
|
+
resolvedTarget,
|
|
122
|
+
allowed: [
|
|
123
|
+
prepared.candidate.agent.id,
|
|
124
|
+
prepared.candidate.agent.model,
|
|
125
|
+
],
|
|
126
|
+
}),
|
|
127
|
+
});
|
|
128
|
+
const sandboxPolicy = composeStageSandboxPolicy({
|
|
129
|
+
stageWriteProtectedPaths: [
|
|
130
|
+
workspacePaths.contextPath,
|
|
131
|
+
join(workspacePaths.workspacePath, "context"),
|
|
132
|
+
join(workspacePaths.workspacePath, "inputs"),
|
|
133
|
+
sharedInputs.sharedInputsAbsolute,
|
|
134
|
+
...(sharedInputsUseReferenceRepo(sharedInputs)
|
|
135
|
+
? [
|
|
136
|
+
join(workspacePaths.workspacePath, "reference_repo"),
|
|
137
|
+
sharedInputs.referenceRepoAbsolute,
|
|
138
|
+
]
|
|
139
|
+
: []),
|
|
140
|
+
],
|
|
141
|
+
});
|
|
142
|
+
const result = await runSandboxedAgent({
|
|
143
|
+
root,
|
|
144
|
+
sessionId: verificationId,
|
|
145
|
+
sandboxStageId: "verify",
|
|
146
|
+
agent: prepared.candidate.agent,
|
|
147
|
+
prompt,
|
|
148
|
+
environment,
|
|
149
|
+
teardownAuthOnExit: false,
|
|
150
|
+
captureChat: true,
|
|
151
|
+
paths: {
|
|
152
|
+
agentRoot: workspacePaths.agentRoot,
|
|
153
|
+
workspacePath: workspacePaths.workspacePath,
|
|
154
|
+
sandboxHomePath: workspacePaths.sandboxHomePath,
|
|
155
|
+
runtimeManifestPath: workspacePaths.runtimeManifestPath,
|
|
156
|
+
sandboxSettingsPath: workspacePaths.sandboxSettingsPath,
|
|
157
|
+
runtimePath: workspacePaths.runtimePath,
|
|
158
|
+
artifactsPath: workspacePaths.artifactsPath,
|
|
159
|
+
stdoutPath: workspacePaths.stdoutPath,
|
|
160
|
+
stderrPath: workspacePaths.stderrPath,
|
|
161
|
+
},
|
|
162
|
+
extraWriteProtectedPaths: sandboxPolicy.extraWriteProtectedPaths,
|
|
163
|
+
extraReadProtectedPaths: sandboxPolicy.extraReadProtectedPaths,
|
|
164
|
+
});
|
|
165
|
+
const tokenUsageResult = await extractProviderNativeTokenUsageForSession({
|
|
166
|
+
root,
|
|
167
|
+
domain: VORATIQ_VERIFICATION_DIR,
|
|
168
|
+
sessionId: verificationId,
|
|
169
|
+
agentId: prepared.candidate.agent.id,
|
|
170
|
+
provider: prepared.candidate.agent.provider,
|
|
171
|
+
modelId: prepared.candidate.agent.model,
|
|
172
|
+
chatCaptured: result.chat?.captured === true,
|
|
173
|
+
format: result.chat?.format,
|
|
174
|
+
artifactPath: result.chat?.artifactPath,
|
|
175
|
+
});
|
|
176
|
+
tokenUsageResultByExecutionKey.set(executionKey, tokenUsageResult);
|
|
177
|
+
const tokenUsage = resolveTokenUsage(tokenUsageResult);
|
|
178
|
+
const completedAt = new Date().toISOString();
|
|
179
|
+
const artifactPath = getVerificationRubricResultPath({
|
|
180
|
+
sessionId: verificationId,
|
|
181
|
+
verifierId: prepared.candidate.agent.id,
|
|
182
|
+
template: prepared.candidate.template.template,
|
|
183
|
+
});
|
|
184
|
+
if (result.exitCode !== 0 || result.errorMessage) {
|
|
185
|
+
const detectedDetail = result.watchdog?.trigger && result.errorMessage
|
|
186
|
+
? result.errorMessage
|
|
187
|
+
: await detectAgentProcessFailureDetail({
|
|
188
|
+
provider: prepared.candidate.agent.provider,
|
|
189
|
+
stdoutPath: workspacePaths.stdoutPath,
|
|
190
|
+
stderrPath: workspacePaths.stderrPath,
|
|
191
|
+
});
|
|
192
|
+
const detail = detectedDetail ??
|
|
193
|
+
result.errorMessage ??
|
|
194
|
+
`Agent exited with code ${result.exitCode ?? "unknown"}`;
|
|
195
|
+
await writeFailureRubricArtifact({
|
|
196
|
+
root,
|
|
197
|
+
artifactPath,
|
|
198
|
+
verifierId: prepared.candidate.agent.id,
|
|
199
|
+
template: prepared.candidate.template.template,
|
|
200
|
+
generatedAt: completedAt,
|
|
201
|
+
error: detail,
|
|
202
|
+
});
|
|
203
|
+
return {
|
|
204
|
+
template: prepared.candidate.template.template,
|
|
205
|
+
verifierId: prepared.candidate.agent.id,
|
|
206
|
+
status: "failed",
|
|
207
|
+
artifactPath,
|
|
208
|
+
startedAt,
|
|
209
|
+
completedAt,
|
|
210
|
+
error: detail,
|
|
211
|
+
tokenUsage,
|
|
212
|
+
tokenUsageResult,
|
|
213
|
+
};
|
|
214
|
+
}
|
|
215
|
+
const outputPath = resolve(workspacePaths.workspacePath, "result.json");
|
|
216
|
+
if (!(await pathExists(outputPath))) {
|
|
217
|
+
const detail = `Missing result.json. See stderr: ${workspacePaths.stderrPath}`;
|
|
218
|
+
await writeFailureRubricArtifact({
|
|
219
|
+
root,
|
|
220
|
+
artifactPath,
|
|
221
|
+
verifierId: prepared.candidate.agent.id,
|
|
222
|
+
template: prepared.candidate.template.template,
|
|
223
|
+
generatedAt: completedAt,
|
|
224
|
+
error: detail,
|
|
225
|
+
});
|
|
226
|
+
return {
|
|
227
|
+
template: prepared.candidate.template.template,
|
|
228
|
+
verifierId: prepared.candidate.agent.id,
|
|
229
|
+
status: "failed",
|
|
230
|
+
artifactPath,
|
|
231
|
+
startedAt,
|
|
232
|
+
completedAt,
|
|
233
|
+
error: detail,
|
|
234
|
+
tokenUsage,
|
|
235
|
+
tokenUsageResult,
|
|
236
|
+
};
|
|
237
|
+
}
|
|
238
|
+
const raw = await readFile(outputPath, "utf8");
|
|
239
|
+
try {
|
|
240
|
+
assertNoVerificationIdentityLeak({
|
|
241
|
+
text: raw,
|
|
242
|
+
forbidden: buildForbiddenVerificationIdentityTokens({
|
|
243
|
+
resolvedTarget,
|
|
244
|
+
allowed: [
|
|
245
|
+
prepared.candidate.agent.id,
|
|
246
|
+
prepared.candidate.agent.model,
|
|
247
|
+
],
|
|
248
|
+
}),
|
|
249
|
+
});
|
|
250
|
+
const resultPayload = parseRubricResultPayload(raw);
|
|
251
|
+
const artifact = verificationResultArtifactSchema.parse({
|
|
252
|
+
method: "rubric",
|
|
253
|
+
template: prepared.candidate.template.template,
|
|
254
|
+
verifierId: prepared.candidate.agent.id,
|
|
255
|
+
generatedAt: completedAt,
|
|
256
|
+
status: "succeeded",
|
|
257
|
+
result: resultPayload,
|
|
258
|
+
});
|
|
259
|
+
assertRubricResultSelectorsMatchAliasMap({
|
|
260
|
+
artifactPath,
|
|
261
|
+
result: artifact.method === "rubric" ? artifact.result : undefined,
|
|
262
|
+
aliasMap,
|
|
263
|
+
});
|
|
264
|
+
await writeVerificationArtifact({
|
|
265
|
+
root,
|
|
266
|
+
artifactPath,
|
|
267
|
+
artifact,
|
|
268
|
+
});
|
|
269
|
+
return {
|
|
270
|
+
template: prepared.candidate.template.template,
|
|
271
|
+
verifierId: prepared.candidate.agent.id,
|
|
272
|
+
status: "succeeded",
|
|
273
|
+
artifactPath,
|
|
274
|
+
startedAt,
|
|
275
|
+
completedAt,
|
|
276
|
+
tokenUsage,
|
|
277
|
+
tokenUsageResult,
|
|
278
|
+
};
|
|
279
|
+
}
|
|
280
|
+
catch (error) {
|
|
281
|
+
const detail = toErrorMessage(error);
|
|
282
|
+
await writeFailureRubricArtifact({
|
|
283
|
+
root,
|
|
284
|
+
artifactPath,
|
|
285
|
+
verifierId: prepared.candidate.agent.id,
|
|
286
|
+
template: prepared.candidate.template.template,
|
|
287
|
+
generatedAt: completedAt,
|
|
288
|
+
error: detail,
|
|
289
|
+
});
|
|
290
|
+
return {
|
|
291
|
+
template: prepared.candidate.template.template,
|
|
292
|
+
verifierId: prepared.candidate.agent.id,
|
|
293
|
+
status: "failed",
|
|
294
|
+
artifactPath,
|
|
295
|
+
startedAt,
|
|
296
|
+
completedAt,
|
|
297
|
+
error: detail,
|
|
298
|
+
tokenUsage,
|
|
299
|
+
tokenUsageResult,
|
|
300
|
+
};
|
|
301
|
+
}
|
|
302
|
+
},
|
|
303
|
+
onCandidateCompleted: async (prepared, result) => {
|
|
304
|
+
const executionKey = executionKeyForCandidate(prepared.candidate);
|
|
305
|
+
await mutators.recordMethodSnapshot({
|
|
306
|
+
method: "rubric",
|
|
307
|
+
template: prepared.candidate.template.template,
|
|
308
|
+
verifierId: prepared.candidate.agent.id,
|
|
309
|
+
scope: resolvedTarget.target.kind === "run"
|
|
310
|
+
? { kind: "run" }
|
|
311
|
+
: { kind: "target" },
|
|
312
|
+
status: result.status,
|
|
313
|
+
artifactPath: result.artifactPath,
|
|
314
|
+
startedAt: result.startedAt,
|
|
315
|
+
completedAt: result.completedAt,
|
|
316
|
+
tokenUsage: result.tokenUsage,
|
|
317
|
+
...(result.status === "failed" ? { error: result.error } : {}),
|
|
318
|
+
});
|
|
319
|
+
emitStageProgressEvent(renderer, {
|
|
320
|
+
type: "stage.candidate",
|
|
321
|
+
stage: "verify",
|
|
322
|
+
candidate: {
|
|
323
|
+
methodKey: `rubric:${prepared.candidate.template.template}:${prepared.candidate.agent.id}`,
|
|
324
|
+
verifierLabel: prepared.candidate.template.template,
|
|
325
|
+
agentLabel: prepared.candidate.agent.id,
|
|
326
|
+
status: result.status,
|
|
327
|
+
startedAt: result.startedAt,
|
|
328
|
+
completedAt: result.completedAt,
|
|
329
|
+
artifactPath: result.artifactPath,
|
|
330
|
+
tokenUsage: result.tokenUsage,
|
|
331
|
+
tokenUsageResult: result.tokenUsageResult,
|
|
332
|
+
},
|
|
333
|
+
});
|
|
334
|
+
startedAtByExecutionKey.delete(executionKey);
|
|
335
|
+
tokenUsageResultByExecutionKey.delete(executionKey);
|
|
336
|
+
},
|
|
337
|
+
captureExecutionFailure: async ({ prepared, error }) => {
|
|
338
|
+
const executionKey = executionKeyForCandidate(prepared.candidate);
|
|
339
|
+
const startedAt = startedAtByExecutionKey.get(executionKey) ?? new Date().toISOString();
|
|
340
|
+
const completedAt = new Date().toISOString();
|
|
341
|
+
const artifactPath = getVerificationRubricResultPath({
|
|
342
|
+
sessionId: verificationId,
|
|
343
|
+
verifierId: prepared.candidate.agent.id,
|
|
344
|
+
template: prepared.candidate.template.template,
|
|
345
|
+
});
|
|
346
|
+
const detail = toErrorMessage(error);
|
|
347
|
+
const tokenUsageResult = tokenUsageResultByExecutionKey.get(executionKey) ??
|
|
348
|
+
buildUnavailableTokenUsageResult({
|
|
349
|
+
provider: prepared.candidate.agent.provider,
|
|
350
|
+
modelId: prepared.candidate.agent.model,
|
|
351
|
+
message: detail,
|
|
352
|
+
});
|
|
353
|
+
const tokenUsage = resolveTokenUsage(tokenUsageResult);
|
|
354
|
+
await writeFailureRubricArtifact({
|
|
355
|
+
root,
|
|
356
|
+
artifactPath,
|
|
357
|
+
verifierId: prepared.candidate.agent.id,
|
|
358
|
+
template: prepared.candidate.template.template,
|
|
359
|
+
generatedAt: completedAt,
|
|
360
|
+
error: detail,
|
|
361
|
+
});
|
|
362
|
+
await mutators.recordMethodSnapshot({
|
|
363
|
+
method: "rubric",
|
|
364
|
+
template: prepared.candidate.template.template,
|
|
365
|
+
verifierId: prepared.candidate.agent.id,
|
|
366
|
+
scope: resolvedTarget.target.kind === "run"
|
|
367
|
+
? { kind: "run" }
|
|
368
|
+
: { kind: "target" },
|
|
369
|
+
status: "failed",
|
|
370
|
+
artifactPath,
|
|
371
|
+
startedAt,
|
|
372
|
+
completedAt,
|
|
373
|
+
tokenUsage,
|
|
374
|
+
error: detail,
|
|
375
|
+
});
|
|
376
|
+
emitStageProgressEvent(renderer, {
|
|
377
|
+
type: "stage.candidate",
|
|
378
|
+
stage: "verify",
|
|
379
|
+
candidate: {
|
|
380
|
+
methodKey: `rubric:${prepared.candidate.template.template}:${prepared.candidate.agent.id}`,
|
|
381
|
+
verifierLabel: prepared.candidate.template.template,
|
|
382
|
+
agentLabel: prepared.candidate.agent.id,
|
|
383
|
+
status: "failed",
|
|
384
|
+
startedAt,
|
|
385
|
+
completedAt,
|
|
386
|
+
artifactPath,
|
|
387
|
+
tokenUsage,
|
|
388
|
+
tokenUsageResult,
|
|
389
|
+
},
|
|
390
|
+
});
|
|
391
|
+
startedAtByExecutionKey.delete(executionKey);
|
|
392
|
+
tokenUsageResultByExecutionKey.delete(executionKey);
|
|
393
|
+
return {
|
|
394
|
+
template: prepared.candidate.template.template,
|
|
395
|
+
verifierId: prepared.candidate.agent.id,
|
|
396
|
+
status: "failed",
|
|
397
|
+
artifactPath,
|
|
398
|
+
startedAt,
|
|
399
|
+
completedAt,
|
|
400
|
+
error: detail,
|
|
401
|
+
tokenUsage,
|
|
402
|
+
tokenUsageResult,
|
|
403
|
+
};
|
|
404
|
+
},
|
|
405
|
+
sortResults: compareVerificationsByTemplateThenVerifierId,
|
|
406
|
+
};
|
|
407
|
+
}
|
|
408
|
+
async function writeFailureRubricArtifact(options) {
|
|
409
|
+
const { root, artifactPath, verifierId, template, generatedAt, error } = options;
|
|
410
|
+
await writeVerificationArtifact({
|
|
411
|
+
root,
|
|
412
|
+
artifactPath,
|
|
413
|
+
artifact: {
|
|
414
|
+
method: "rubric",
|
|
415
|
+
template,
|
|
416
|
+
verifierId,
|
|
417
|
+
generatedAt,
|
|
418
|
+
status: "failed",
|
|
419
|
+
result: {},
|
|
420
|
+
error,
|
|
421
|
+
},
|
|
422
|
+
});
|
|
423
|
+
}
|
|
424
|
+
function buildVerifierRubricWorkspacePaths(options) {
|
|
425
|
+
const { root, verificationId, verifierId, template } = options;
|
|
426
|
+
return buildScopedAgentWorkspacePaths({
|
|
427
|
+
agentRoot: resolve(root, getVerificationRubricExecutionDirectoryPath({
|
|
428
|
+
sessionId: verificationId,
|
|
429
|
+
verifierId,
|
|
430
|
+
template,
|
|
431
|
+
})),
|
|
432
|
+
});
|
|
433
|
+
}
|
|
434
|
+
function compareVerificationsByTemplateThenVerifierId(left, right) {
|
|
435
|
+
return (left.template.localeCompare(right.template) ||
|
|
436
|
+
left.verifierId.localeCompare(right.verifierId));
|
|
437
|
+
}
|
|
438
|
+
function registerScratchWorkspaceTeardown(teardown, workspacePaths, verifierId, template) {
|
|
439
|
+
const labelPrefix = `${verifierId}/${template}`;
|
|
440
|
+
teardown.addPath(workspacePaths.workspacePath, `${labelPrefix} workspace`);
|
|
441
|
+
teardown.addPath(workspacePaths.contextPath, `${labelPrefix} context`);
|
|
442
|
+
teardown.addPath(workspacePaths.runtimePath, `${labelPrefix} runtime`);
|
|
443
|
+
teardown.addPath(workspacePaths.sandboxPath, `${labelPrefix} sandbox`);
|
|
444
|
+
}
|
|
@@ -0,0 +1,7 @@
|
|
|
1
|
+
import { mkdir, writeFile } from "node:fs/promises";
|
|
2
|
+
import { dirname, resolve } from "node:path";
|
|
3
|
+
export async function writeVerificationArtifact(options) {
|
|
4
|
+
const absolutePath = resolve(options.root, options.artifactPath);
|
|
5
|
+
await mkdir(dirname(absolutePath), { recursive: true });
|
|
6
|
+
await writeFile(absolutePath, `${JSON.stringify(options.artifact, null, 2)}\n`, "utf8");
|
|
7
|
+
}
|
|
@@ -0,0 +1,24 @@
|
|
|
1
|
+
import { parseRubricResultPayload, type RubricResultPayload } from "../rubric-result.js";
|
|
2
|
+
import type { ResolvedVerificationTarget } from "./target.js";
|
|
3
|
+
export declare function buildBlindedAliasMap(resolvedTarget: ResolvedVerificationTarget): Record<string, string> | undefined;
|
|
4
|
+
export declare function aliasForCandidate(canonicalId: string, aliasMap?: Record<string, string>): string;
|
|
5
|
+
export declare function assertRubricResultSelectorsMatchAliasMap(options: {
|
|
6
|
+
artifactPath: string;
|
|
7
|
+
result?: RubricResultPayload;
|
|
8
|
+
aliasMap?: Record<string, string>;
|
|
9
|
+
}): void;
|
|
10
|
+
export declare function buildForbiddenVerificationIdentityTokens(options: {
|
|
11
|
+
resolvedTarget: ResolvedVerificationTarget;
|
|
12
|
+
allowed?: readonly string[];
|
|
13
|
+
}): string[];
|
|
14
|
+
/**
|
|
15
|
+
* Build an overlap-safe allowlist: verifier self-identity tokens are only
|
|
16
|
+
* permitted when they are not also candidate-identity tokens under
|
|
17
|
+
* verification.
|
|
18
|
+
*/
|
|
19
|
+
export declare function overlapSafeAllowlist(allowed: readonly string[], candidateTokens: ReadonlySet<string>): Set<string>;
|
|
20
|
+
export declare function assertNoVerificationIdentityLeak(options: {
|
|
21
|
+
text: string;
|
|
22
|
+
forbidden: readonly string[];
|
|
23
|
+
}): void;
|
|
24
|
+
export { parseRubricResultPayload };
|
|
@@ -0,0 +1,109 @@
|
|
|
1
|
+
import { generateBlindedCandidateAlias } from "../blinding/aliases.js";
|
|
2
|
+
import { parseRubricResultPayload, readRubricResultPreferred, readRubricResultRanking, } from "../rubric-result.js";
|
|
3
|
+
export function buildBlindedAliasMap(resolvedTarget) {
|
|
4
|
+
const candidateIds = resolvedTarget.competitiveCandidates.map((candidate) => candidate.canonicalId);
|
|
5
|
+
if (candidateIds.length === 0) {
|
|
6
|
+
return undefined;
|
|
7
|
+
}
|
|
8
|
+
const aliasMap = {};
|
|
9
|
+
const seen = new Set();
|
|
10
|
+
for (const candidateId of [...new Set(candidateIds)].sort((a, b) => a.localeCompare(b))) {
|
|
11
|
+
const alias = generateBlindedCandidateAlias({ seen });
|
|
12
|
+
seen.add(alias);
|
|
13
|
+
aliasMap[alias] = candidateId;
|
|
14
|
+
}
|
|
15
|
+
return aliasMap;
|
|
16
|
+
}
|
|
17
|
+
export function aliasForCandidate(canonicalId, aliasMap) {
|
|
18
|
+
if (!aliasMap) {
|
|
19
|
+
return canonicalId;
|
|
20
|
+
}
|
|
21
|
+
const match = Object.entries(aliasMap).find(([, value]) => value === canonicalId);
|
|
22
|
+
return match?.[0] ?? canonicalId;
|
|
23
|
+
}
|
|
24
|
+
export function assertRubricResultSelectorsMatchAliasMap(options) {
|
|
25
|
+
const { artifactPath, result, aliasMap } = options;
|
|
26
|
+
if (!aliasMap || !result) {
|
|
27
|
+
return;
|
|
28
|
+
}
|
|
29
|
+
const unknownSelectors = new Set();
|
|
30
|
+
const preferred = readRubricResultPreferred(result);
|
|
31
|
+
if (preferred && !aliasMap[preferred]) {
|
|
32
|
+
unknownSelectors.add(preferred);
|
|
33
|
+
}
|
|
34
|
+
for (const selector of readRubricResultRanking(result) ?? []) {
|
|
35
|
+
if (!aliasMap[selector]) {
|
|
36
|
+
unknownSelectors.add(selector);
|
|
37
|
+
}
|
|
38
|
+
}
|
|
39
|
+
if (unknownSelectors.size === 0) {
|
|
40
|
+
return;
|
|
41
|
+
}
|
|
42
|
+
const selectors = Array.from(unknownSelectors)
|
|
43
|
+
.sort((left, right) => left.localeCompare(right))
|
|
44
|
+
.map((selector) => `\`${selector}\``)
|
|
45
|
+
.join(", ");
|
|
46
|
+
throw new Error(`Invalid verifier result.json for \`${artifactPath}\`: unknown blinded selector(s) ${selectors}.`);
|
|
47
|
+
}
|
|
48
|
+
export function buildForbiddenVerificationIdentityTokens(options) {
|
|
49
|
+
const { resolvedTarget, allowed = [] } = options;
|
|
50
|
+
// Collect all candidate identity tokens first so we can detect overlap
|
|
51
|
+
// with verifier self-identity tokens.
|
|
52
|
+
const candidateTokens = new Set();
|
|
53
|
+
for (const candidate of resolvedTarget.competitiveCandidates) {
|
|
54
|
+
for (const token of candidate.forbiddenIdentityTokens) {
|
|
55
|
+
const normalized = token.toLowerCase().trim();
|
|
56
|
+
if (normalized) {
|
|
57
|
+
candidateTokens.add(normalized);
|
|
58
|
+
}
|
|
59
|
+
}
|
|
60
|
+
}
|
|
61
|
+
// Only allowlist verifier self-identity tokens that do NOT overlap with
|
|
62
|
+
// candidate identity tokens. Overlapping tokens must stay forbidden so
|
|
63
|
+
// that a shared identity between verifier and candidate is still detected
|
|
64
|
+
// as a leakage vector.
|
|
65
|
+
const allowedTokens = overlapSafeAllowlist(allowed, candidateTokens);
|
|
66
|
+
const tokens = new Set();
|
|
67
|
+
for (const candidate of resolvedTarget.competitiveCandidates) {
|
|
68
|
+
for (const token of candidate.forbiddenIdentityTokens) {
|
|
69
|
+
const normalized = token.toLowerCase().trim();
|
|
70
|
+
if (!normalized || allowedTokens.has(normalized)) {
|
|
71
|
+
continue;
|
|
72
|
+
}
|
|
73
|
+
tokens.add(normalized);
|
|
74
|
+
}
|
|
75
|
+
}
|
|
76
|
+
return Array.from(tokens);
|
|
77
|
+
}
|
|
78
|
+
/**
|
|
79
|
+
* Build an overlap-safe allowlist: verifier self-identity tokens are only
|
|
80
|
+
* permitted when they are not also candidate-identity tokens under
|
|
81
|
+
* verification.
|
|
82
|
+
*/
|
|
83
|
+
export function overlapSafeAllowlist(allowed, candidateTokens) {
|
|
84
|
+
return new Set(allowed
|
|
85
|
+
.map((token) => token.toLowerCase().trim())
|
|
86
|
+
.filter((token) => token !== "" && !candidateTokens.has(token)));
|
|
87
|
+
}
|
|
88
|
+
export function assertNoVerificationIdentityLeak(options) {
|
|
89
|
+
const { text, forbidden } = options;
|
|
90
|
+
const haystack = text.toLowerCase();
|
|
91
|
+
const leaks = forbidden.filter((token) => containsBoundedToken(haystack, token));
|
|
92
|
+
if (leaks.length === 0) {
|
|
93
|
+
return;
|
|
94
|
+
}
|
|
95
|
+
const preview = leaks
|
|
96
|
+
.slice(0, 5)
|
|
97
|
+
.map((token) => `\`${token}\``)
|
|
98
|
+
.join(", ");
|
|
99
|
+
throw new Error(`Blinded verification leakage validation failed: forbidden candidate identity token(s) detected: ${preview}${leaks.length > 5 ? ", ..." : ""}.`);
|
|
100
|
+
}
|
|
101
|
+
function containsBoundedToken(text, token) {
|
|
102
|
+
if (!token) {
|
|
103
|
+
return false;
|
|
104
|
+
}
|
|
105
|
+
const escaped = token.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
|
|
106
|
+
const pattern = new RegExp(`(?<![a-z0-9_-])${escaped}(?![a-z0-9_-])`, "iu");
|
|
107
|
+
return pattern.test(text);
|
|
108
|
+
}
|
|
109
|
+
export { parseRubricResultPayload };
|
|
@@ -0,0 +1,11 @@
|
|
|
1
|
+
import type { VerificationMethodResultRef, VerificationStatus } from "../model/types.js";
|
|
2
|
+
import type { ResolvedVerificationTarget } from "./target.js";
|
|
3
|
+
export declare function deriveVerificationStatusFromMethods(methods: readonly VerificationMethodResultRef[]): VerificationStatus;
|
|
4
|
+
export declare function maybePersistSelectedSpecPath(options: {
|
|
5
|
+
root: string;
|
|
6
|
+
verificationsFilePath: string;
|
|
7
|
+
verificationId: string;
|
|
8
|
+
resolvedTarget: ResolvedVerificationTarget;
|
|
9
|
+
aliasMap?: Record<string, string>;
|
|
10
|
+
methods: readonly VerificationMethodResultRef[];
|
|
11
|
+
}): Promise<void>;
|
|
@@ -0,0 +1,65 @@
|
|
|
1
|
+
import { loadVerificationSelectionPolicyOutput } from "../../../policy/index.js";
|
|
2
|
+
import { rewriteVerificationRecord } from "../persistence/adapter.js";
|
|
3
|
+
export function deriveVerificationStatusFromMethods(methods) {
|
|
4
|
+
if (methods.some((method) => method.status === "failed")) {
|
|
5
|
+
return "failed";
|
|
6
|
+
}
|
|
7
|
+
if (methods.some((method) => method.status === "aborted")) {
|
|
8
|
+
return "aborted";
|
|
9
|
+
}
|
|
10
|
+
return "succeeded";
|
|
11
|
+
}
|
|
12
|
+
export async function maybePersistSelectedSpecPath(options) {
|
|
13
|
+
const { root, verificationsFilePath, verificationId, resolvedTarget, aliasMap, methods, } = options;
|
|
14
|
+
if (!("specRecord" in resolvedTarget)) {
|
|
15
|
+
return;
|
|
16
|
+
}
|
|
17
|
+
const record = await loadVerificationSelectionPolicyOutput({
|
|
18
|
+
root,
|
|
19
|
+
record: {
|
|
20
|
+
sessionId: verificationId,
|
|
21
|
+
createdAt: new Date(0).toISOString(),
|
|
22
|
+
status: "succeeded",
|
|
23
|
+
target: resolvedTarget.target,
|
|
24
|
+
methods: [...methods],
|
|
25
|
+
...(resolvedTarget.specRecord.extraContext
|
|
26
|
+
? { extraContext: resolvedTarget.specRecord.extraContext }
|
|
27
|
+
: {}),
|
|
28
|
+
...(resolvedTarget.specRecord.extraContextMetadata
|
|
29
|
+
? {
|
|
30
|
+
extraContextMetadata: resolvedTarget.specRecord.extraContextMetadata,
|
|
31
|
+
}
|
|
32
|
+
: {}),
|
|
33
|
+
...(aliasMap ? { blinded: { enabled: true, aliasMap } } : {}),
|
|
34
|
+
},
|
|
35
|
+
canonicalCandidateIds: resolvedTarget.specRecord.agents.map((agent) => agent.agentId),
|
|
36
|
+
});
|
|
37
|
+
const decision = record.decision;
|
|
38
|
+
if (decision.state !== "resolvable") {
|
|
39
|
+
return;
|
|
40
|
+
}
|
|
41
|
+
const selected = resolvedTarget.specRecord.agents.find((agent) => agent.status === "succeeded" &&
|
|
42
|
+
agent.agentId === decision.selectedCanonicalAgentId &&
|
|
43
|
+
agent.outputPath);
|
|
44
|
+
if (!selected?.outputPath) {
|
|
45
|
+
return;
|
|
46
|
+
}
|
|
47
|
+
await rewriteVerificationRecord({
|
|
48
|
+
root,
|
|
49
|
+
verificationsFilePath,
|
|
50
|
+
sessionId: verificationId,
|
|
51
|
+
mutate: (existing) => {
|
|
52
|
+
if (existing.target.kind !== "spec") {
|
|
53
|
+
return existing;
|
|
54
|
+
}
|
|
55
|
+
return {
|
|
56
|
+
...existing,
|
|
57
|
+
target: {
|
|
58
|
+
...existing.target,
|
|
59
|
+
specPath: selected.outputPath,
|
|
60
|
+
},
|
|
61
|
+
};
|
|
62
|
+
},
|
|
63
|
+
forceFlush: true,
|
|
64
|
+
});
|
|
65
|
+
}
|