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 +1,229 @@
|
|
|
1
|
+
import { runTeardown } from "../../competition/shared/teardown.js";
|
|
2
|
+
import { buildOperationLifecycleCompleteFields, buildRecordLifecycleCompleteFields, } from "../../domain/shared/lifecycle.js";
|
|
3
|
+
import { writeVerificationArtifact } from "../../domain/verify/competition/artifacts.js";
|
|
4
|
+
import { flushVerificationRecordBuffer, readVerificationRecords, rewriteVerificationRecord, } from "../../domain/verify/persistence/adapter.js";
|
|
5
|
+
import { toErrorMessage } from "../../utils/errors.js";
|
|
6
|
+
import { getVerificationProgrammaticResultPath, getVerificationRubricResultPath, } from "../../workspace/structure.js";
|
|
7
|
+
export const VERIFY_ABORT_DETAIL = "Verification aborted before completion.";
|
|
8
|
+
let activeVerification;
|
|
9
|
+
let terminationInFlight = false;
|
|
10
|
+
export function registerActiveVerification(context) {
|
|
11
|
+
activeVerification = context;
|
|
12
|
+
}
|
|
13
|
+
export function clearActiveVerification(verificationId) {
|
|
14
|
+
if (activeVerification?.verificationId !== verificationId) {
|
|
15
|
+
return;
|
|
16
|
+
}
|
|
17
|
+
if (!terminationInFlight) {
|
|
18
|
+
activeVerification = undefined;
|
|
19
|
+
}
|
|
20
|
+
}
|
|
21
|
+
export async function terminateActiveVerification(status) {
|
|
22
|
+
if (!activeVerification || terminationInFlight) {
|
|
23
|
+
return;
|
|
24
|
+
}
|
|
25
|
+
terminationInFlight = true;
|
|
26
|
+
const context = activeVerification;
|
|
27
|
+
let persistenceError;
|
|
28
|
+
try {
|
|
29
|
+
const existingRecord = await readVerificationRecords({
|
|
30
|
+
root: context.root,
|
|
31
|
+
verificationsFilePath: context.verificationsFilePath,
|
|
32
|
+
limit: 1,
|
|
33
|
+
predicate: (record) => record.sessionId === context.verificationId,
|
|
34
|
+
}).then((records) => records[0]);
|
|
35
|
+
if (!existingRecord) {
|
|
36
|
+
return;
|
|
37
|
+
}
|
|
38
|
+
const completedAt = new Date().toISOString();
|
|
39
|
+
const detail = status === "aborted" ? VERIFY_ABORT_DETAIL : "Verification failed.";
|
|
40
|
+
await materializeTerminalVerificationArtifacts({
|
|
41
|
+
root: context.root,
|
|
42
|
+
verificationId: context.verificationId,
|
|
43
|
+
record: existingRecord,
|
|
44
|
+
status,
|
|
45
|
+
completedAt,
|
|
46
|
+
detail,
|
|
47
|
+
});
|
|
48
|
+
await rewriteVerificationRecord({
|
|
49
|
+
root: context.root,
|
|
50
|
+
verificationsFilePath: context.verificationsFilePath,
|
|
51
|
+
sessionId: context.verificationId,
|
|
52
|
+
mutate: (existing) => {
|
|
53
|
+
const methods = existing.methods.map((method) => finalizeVerificationMethodRef({
|
|
54
|
+
verificationId: context.verificationId,
|
|
55
|
+
method,
|
|
56
|
+
target: existing.target,
|
|
57
|
+
status,
|
|
58
|
+
completedAt,
|
|
59
|
+
detail,
|
|
60
|
+
}));
|
|
61
|
+
const inProgress = existing.status === "queued" || existing.status === "running";
|
|
62
|
+
if (!inProgress) {
|
|
63
|
+
if (methods.every((method, index) => method === existing.methods[index])) {
|
|
64
|
+
return existing;
|
|
65
|
+
}
|
|
66
|
+
return {
|
|
67
|
+
...existing,
|
|
68
|
+
methods,
|
|
69
|
+
};
|
|
70
|
+
}
|
|
71
|
+
return {
|
|
72
|
+
...existing,
|
|
73
|
+
status,
|
|
74
|
+
methods,
|
|
75
|
+
...buildRecordLifecycleCompleteFields({
|
|
76
|
+
existing,
|
|
77
|
+
startedAt: existing.startedAt ?? completedAt,
|
|
78
|
+
completedAt,
|
|
79
|
+
}),
|
|
80
|
+
error: existing.error ?? detail,
|
|
81
|
+
};
|
|
82
|
+
},
|
|
83
|
+
forceFlush: true,
|
|
84
|
+
});
|
|
85
|
+
await flushVerificationRecordBuffer({
|
|
86
|
+
verificationsFilePath: context.verificationsFilePath,
|
|
87
|
+
sessionId: context.verificationId,
|
|
88
|
+
});
|
|
89
|
+
}
|
|
90
|
+
catch (error) {
|
|
91
|
+
persistenceError =
|
|
92
|
+
error instanceof Error ? error : new Error(toErrorMessage(error));
|
|
93
|
+
console.error(`[voratiq] Failed to finalize verification ${context.verificationId}: ${toErrorMessage(error)}`);
|
|
94
|
+
}
|
|
95
|
+
finally {
|
|
96
|
+
try {
|
|
97
|
+
await finalizeRegisteredVerificationTeardown(context);
|
|
98
|
+
}
|
|
99
|
+
finally {
|
|
100
|
+
terminationInFlight = false;
|
|
101
|
+
activeVerification = undefined;
|
|
102
|
+
}
|
|
103
|
+
}
|
|
104
|
+
if (persistenceError) {
|
|
105
|
+
throw persistenceError;
|
|
106
|
+
}
|
|
107
|
+
}
|
|
108
|
+
export async function finalizeActiveVerification(verificationId) {
|
|
109
|
+
if (!activeVerification ||
|
|
110
|
+
activeVerification.verificationId !== verificationId) {
|
|
111
|
+
clearActiveVerification(verificationId);
|
|
112
|
+
return;
|
|
113
|
+
}
|
|
114
|
+
const context = activeVerification;
|
|
115
|
+
try {
|
|
116
|
+
await finalizeRegisteredVerificationTeardown(context);
|
|
117
|
+
}
|
|
118
|
+
finally {
|
|
119
|
+
clearActiveVerification(verificationId);
|
|
120
|
+
}
|
|
121
|
+
}
|
|
122
|
+
function finalizeVerificationMethodRef(options) {
|
|
123
|
+
const { verificationId, method, target, status, completedAt, detail } = options;
|
|
124
|
+
if (method.status !== "queued" && method.status !== "running") {
|
|
125
|
+
return method;
|
|
126
|
+
}
|
|
127
|
+
const artifactPath = method.artifactPath ??
|
|
128
|
+
buildVerificationMethodArtifactPath({
|
|
129
|
+
verificationId,
|
|
130
|
+
method,
|
|
131
|
+
});
|
|
132
|
+
return {
|
|
133
|
+
...method,
|
|
134
|
+
status,
|
|
135
|
+
artifactPath,
|
|
136
|
+
...buildOperationLifecycleCompleteFields({
|
|
137
|
+
existing: method,
|
|
138
|
+
startedAt: method.startedAt ?? completedAt,
|
|
139
|
+
completedAt,
|
|
140
|
+
}),
|
|
141
|
+
error: method.error ?? detail,
|
|
142
|
+
...(method.method === "programmatic" &&
|
|
143
|
+
target.kind !== "run" &&
|
|
144
|
+
method.scope.kind === "run"
|
|
145
|
+
? { scope: { kind: "target" } }
|
|
146
|
+
: {}),
|
|
147
|
+
};
|
|
148
|
+
}
|
|
149
|
+
async function materializeTerminalVerificationArtifacts(options) {
|
|
150
|
+
const { root, verificationId, record, status, completedAt, detail } = options;
|
|
151
|
+
for (const method of record.methods) {
|
|
152
|
+
if (method.status !== "queued" && method.status !== "running") {
|
|
153
|
+
continue;
|
|
154
|
+
}
|
|
155
|
+
const artifactPath = buildVerificationMethodArtifactPath({
|
|
156
|
+
verificationId,
|
|
157
|
+
method,
|
|
158
|
+
});
|
|
159
|
+
if (!artifactPath) {
|
|
160
|
+
continue;
|
|
161
|
+
}
|
|
162
|
+
if (method.method === "rubric") {
|
|
163
|
+
const { template, verifierId } = method;
|
|
164
|
+
if (!template || !verifierId) {
|
|
165
|
+
continue;
|
|
166
|
+
}
|
|
167
|
+
await writeVerificationArtifact({
|
|
168
|
+
root,
|
|
169
|
+
artifactPath,
|
|
170
|
+
artifact: {
|
|
171
|
+
method: "rubric",
|
|
172
|
+
template,
|
|
173
|
+
verifierId,
|
|
174
|
+
generatedAt: completedAt,
|
|
175
|
+
status,
|
|
176
|
+
result: {},
|
|
177
|
+
error: detail,
|
|
178
|
+
},
|
|
179
|
+
});
|
|
180
|
+
continue;
|
|
181
|
+
}
|
|
182
|
+
if (record.target.kind === "run") {
|
|
183
|
+
await writeVerificationArtifact({
|
|
184
|
+
root,
|
|
185
|
+
artifactPath,
|
|
186
|
+
artifact: {
|
|
187
|
+
method: "programmatic",
|
|
188
|
+
generatedAt: completedAt,
|
|
189
|
+
status,
|
|
190
|
+
error: detail,
|
|
191
|
+
target: record.target,
|
|
192
|
+
scope: "run",
|
|
193
|
+
candidates: [],
|
|
194
|
+
},
|
|
195
|
+
});
|
|
196
|
+
continue;
|
|
197
|
+
}
|
|
198
|
+
await writeVerificationArtifact({
|
|
199
|
+
root,
|
|
200
|
+
artifactPath,
|
|
201
|
+
artifact: {
|
|
202
|
+
method: "programmatic",
|
|
203
|
+
generatedAt: completedAt,
|
|
204
|
+
status,
|
|
205
|
+
error: detail,
|
|
206
|
+
target: record.target,
|
|
207
|
+
scope: "target",
|
|
208
|
+
results: [],
|
|
209
|
+
},
|
|
210
|
+
});
|
|
211
|
+
}
|
|
212
|
+
}
|
|
213
|
+
async function finalizeRegisteredVerificationTeardown(context) {
|
|
214
|
+
await runTeardown(context.teardown);
|
|
215
|
+
}
|
|
216
|
+
function buildVerificationMethodArtifactPath(options) {
|
|
217
|
+
const { verificationId, method } = options;
|
|
218
|
+
if (method.method === "programmatic") {
|
|
219
|
+
return getVerificationProgrammaticResultPath(verificationId);
|
|
220
|
+
}
|
|
221
|
+
if (!method.verifierId || !method.template) {
|
|
222
|
+
return undefined;
|
|
223
|
+
}
|
|
224
|
+
return getVerificationRubricResultPath({
|
|
225
|
+
sessionId: verificationId,
|
|
226
|
+
verifierId: method.verifierId,
|
|
227
|
+
template: method.template,
|
|
228
|
+
});
|
|
229
|
+
}
|
|
@@ -0,0 +1,7 @@
|
|
|
1
|
+
import type { VerificationConfig } from "../../configs/verification/types.js";
|
|
2
|
+
export declare function resolveVerifyRubricMaxParallel(options: {
|
|
3
|
+
targetKind: "spec" | "run" | "reduce" | "message";
|
|
4
|
+
verificationConfig: VerificationConfig;
|
|
5
|
+
verifierAgentCount: number;
|
|
6
|
+
requestedMaxParallel?: number;
|
|
7
|
+
}): number;
|
|
@@ -0,0 +1,15 @@
|
|
|
1
|
+
import { resolveEffectiveMaxParallel } from "../shared/max-parallel.js";
|
|
2
|
+
export function resolveVerifyRubricMaxParallel(options) {
|
|
3
|
+
const { targetKind, verificationConfig, verifierAgentCount, requestedMaxParallel, } = options;
|
|
4
|
+
const rubricTemplateCount = targetKind === "spec"
|
|
5
|
+
? verificationConfig.spec.rubric.length
|
|
6
|
+
: targetKind === "run"
|
|
7
|
+
? verificationConfig.run.rubric.length
|
|
8
|
+
: targetKind === "reduce"
|
|
9
|
+
? verificationConfig.reduce.rubric.length
|
|
10
|
+
: verificationConfig.message.rubric.length;
|
|
11
|
+
return resolveEffectiveMaxParallel({
|
|
12
|
+
competitorCount: verifierAgentCount * rubricTemplateCount,
|
|
13
|
+
requestedMaxParallel,
|
|
14
|
+
});
|
|
15
|
+
}
|
|
@@ -0,0 +1,18 @@
|
|
|
1
|
+
import type { ResolvedVerificationTarget, VerificationCompetitiveCandidate } from "../../domain/verify/competition/target.js";
|
|
2
|
+
export type VerifyTargetKind = "spec" | "run" | "reduce" | "message";
|
|
3
|
+
export interface VerifyTargetSelection {
|
|
4
|
+
kind: VerifyTargetKind;
|
|
5
|
+
sessionId: string;
|
|
6
|
+
}
|
|
7
|
+
export interface ResolveVerifyTargetInput {
|
|
8
|
+
root: string;
|
|
9
|
+
specsFilePath: string;
|
|
10
|
+
runsFilePath: string;
|
|
11
|
+
reductionsFilePath: string;
|
|
12
|
+
messagesFilePath: string;
|
|
13
|
+
verificationsFilePath: string;
|
|
14
|
+
target: VerifyTargetSelection;
|
|
15
|
+
}
|
|
16
|
+
export type VerifyCompetitiveCandidate = VerificationCompetitiveCandidate;
|
|
17
|
+
export type ResolvedVerifyTarget = ResolvedVerificationTarget;
|
|
18
|
+
export declare function resolveVerifyTarget(input: ResolveVerifyTargetInput): Promise<ResolvedVerifyTarget>;
|
|
@@ -0,0 +1,420 @@
|
|
|
1
|
+
import { basename } from "node:path";
|
|
2
|
+
import { CliError, RunNotFoundCliError } from "../../cli/errors.js";
|
|
3
|
+
import { TERMINAL_MESSAGE_STATUSES } from "../../domain/message/model/types.js";
|
|
4
|
+
import { readMessageRecords } from "../../domain/message/persistence/adapter.js";
|
|
5
|
+
import { TERMINAL_REDUCTION_STATUSES } from "../../domain/reduce/model/types.js";
|
|
6
|
+
import { readReductionRecords } from "../../domain/reduce/persistence/adapter.js";
|
|
7
|
+
import { RunRecordNotFoundError } from "../../domain/run/model/errors.js";
|
|
8
|
+
import { fetchRunsSafely } from "../../domain/run/persistence/adapter.js";
|
|
9
|
+
import { TERMINAL_SPEC_STATUSES } from "../../domain/spec/model/types.js";
|
|
10
|
+
import { readSpecRecords } from "../../domain/spec/persistence/adapter.js";
|
|
11
|
+
import { readVerificationRecords } from "../../domain/verify/persistence/adapter.js";
|
|
12
|
+
import { TERMINAL_RUN_STATUSES } from "../../status/index.js";
|
|
13
|
+
import { MESSAGE_RESPONSE_FILENAME } from "../../workspace/structure.js";
|
|
14
|
+
export async function resolveVerifyTarget(input) {
|
|
15
|
+
const { target } = input;
|
|
16
|
+
switch (target.kind) {
|
|
17
|
+
case "spec":
|
|
18
|
+
return resolveSpecVerifyTarget(input);
|
|
19
|
+
case "run":
|
|
20
|
+
return resolveRunVerifyTarget(input);
|
|
21
|
+
case "reduce":
|
|
22
|
+
return resolveReductionVerifyTarget(input);
|
|
23
|
+
case "message":
|
|
24
|
+
return resolveMessageVerifyTarget(input);
|
|
25
|
+
}
|
|
26
|
+
}
|
|
27
|
+
async function resolveSpecVerifyTarget(input) {
|
|
28
|
+
const { root, specsFilePath, target } = input;
|
|
29
|
+
const [record] = await readSpecRecords({
|
|
30
|
+
root,
|
|
31
|
+
specsFilePath,
|
|
32
|
+
limit: 1,
|
|
33
|
+
predicate: (entry) => entry.sessionId === target.sessionId,
|
|
34
|
+
});
|
|
35
|
+
if (!record) {
|
|
36
|
+
throw new CliError(`Spec session \`${target.sessionId}\` not found.`, [], [
|
|
37
|
+
"Re-run `voratiq spec` or confirm the session id in `.voratiq/spec/index.json`.",
|
|
38
|
+
]);
|
|
39
|
+
}
|
|
40
|
+
if (!TERMINAL_SPEC_STATUSES.includes(record.status)) {
|
|
41
|
+
throw new CliError(`Spec session \`${target.sessionId}\` is not complete.`, [`Status: \`${record.status}\`.`], ["Wait for the spec to finish before running `voratiq verify`."]);
|
|
42
|
+
}
|
|
43
|
+
return {
|
|
44
|
+
baseRevisionSha: resolveSpecBaseRevisionSha(record),
|
|
45
|
+
competitiveCandidates: record.agents
|
|
46
|
+
.filter((agent) => agent.status === "succeeded" && agent.outputPath)
|
|
47
|
+
.map((agent) => ({
|
|
48
|
+
canonicalId: agent.agentId,
|
|
49
|
+
forbiddenIdentityTokens: [agent.agentId],
|
|
50
|
+
})),
|
|
51
|
+
target: {
|
|
52
|
+
kind: "spec",
|
|
53
|
+
sessionId: record.sessionId,
|
|
54
|
+
},
|
|
55
|
+
specRecord: record,
|
|
56
|
+
};
|
|
57
|
+
}
|
|
58
|
+
async function resolveRunVerifyTarget(input) {
|
|
59
|
+
const { root, runsFilePath, target } = input;
|
|
60
|
+
const { records } = await fetchRunsSafely({
|
|
61
|
+
root,
|
|
62
|
+
runsFilePath,
|
|
63
|
+
runId: target.sessionId,
|
|
64
|
+
filters: { includeDeleted: true },
|
|
65
|
+
}).catch((error) => {
|
|
66
|
+
if (error instanceof RunRecordNotFoundError) {
|
|
67
|
+
throw new RunNotFoundCliError(target.sessionId);
|
|
68
|
+
}
|
|
69
|
+
throw error;
|
|
70
|
+
});
|
|
71
|
+
const record = records[0];
|
|
72
|
+
if (!record) {
|
|
73
|
+
throw new RunNotFoundCliError(target.sessionId);
|
|
74
|
+
}
|
|
75
|
+
if (!isRunStatusCompleteForVerification(record.status)) {
|
|
76
|
+
throw new CliError(`Run \`${target.sessionId}\` is not complete.`, [`Status: \`${record.status}\`.`], ["Wait for the run to finish before running `voratiq verify`."]);
|
|
77
|
+
}
|
|
78
|
+
const candidateIds = [
|
|
79
|
+
...new Set(record.agents.map((agent) => agent.agentId)),
|
|
80
|
+
].sort((left, right) => left.localeCompare(right));
|
|
81
|
+
if (candidateIds.length === 0) {
|
|
82
|
+
throw new CliError(`Run \`${target.sessionId}\` has no candidate agents to verify.`, [], ["Re-run `voratiq run` to generate verifiable candidates."]);
|
|
83
|
+
}
|
|
84
|
+
return {
|
|
85
|
+
baseRevisionSha: record.baseRevisionSha,
|
|
86
|
+
competitiveCandidates: candidateIds.map((candidateId) => ({
|
|
87
|
+
canonicalId: candidateId,
|
|
88
|
+
forbiddenIdentityTokens: collectRunCandidateIdentityTokens({
|
|
89
|
+
runRecord: record,
|
|
90
|
+
candidateId,
|
|
91
|
+
}),
|
|
92
|
+
})),
|
|
93
|
+
target: {
|
|
94
|
+
kind: "run",
|
|
95
|
+
sessionId: record.runId,
|
|
96
|
+
candidateIds,
|
|
97
|
+
},
|
|
98
|
+
runRecord: record,
|
|
99
|
+
};
|
|
100
|
+
}
|
|
101
|
+
function isRunStatusCompleteForVerification(status) {
|
|
102
|
+
return status === "pruned" || TERMINAL_RUN_STATUSES.includes(status);
|
|
103
|
+
}
|
|
104
|
+
async function resolveReductionVerifyTarget(input) {
|
|
105
|
+
const { root, reductionsFilePath, target } = input;
|
|
106
|
+
const [record] = await readReductionRecords({
|
|
107
|
+
root,
|
|
108
|
+
reductionsFilePath,
|
|
109
|
+
limit: 1,
|
|
110
|
+
predicate: (entry) => entry.sessionId === target.sessionId,
|
|
111
|
+
});
|
|
112
|
+
if (!record) {
|
|
113
|
+
throw new CliError(`Reduction session \`${target.sessionId}\` not found.`, [], [
|
|
114
|
+
"Re-run `voratiq reduce` or confirm the session id in `.voratiq/reduce/index.json`.",
|
|
115
|
+
]);
|
|
116
|
+
}
|
|
117
|
+
if (!TERMINAL_REDUCTION_STATUSES.includes(record.status)) {
|
|
118
|
+
throw new CliError(`Reduction session \`${target.sessionId}\` is not complete.`, [`Status: \`${record.status}\`.`], ["Wait for the reduction to finish before running `voratiq verify`."]);
|
|
119
|
+
}
|
|
120
|
+
const referenceRepo = await resolveReductionReferenceRepo({
|
|
121
|
+
...input,
|
|
122
|
+
reductionRecord: record,
|
|
123
|
+
});
|
|
124
|
+
const competitiveCandidates = record.reducers
|
|
125
|
+
.filter((reducer) => reducer.status === "succeeded" && reducer.outputPath)
|
|
126
|
+
.map((reducer) => ({
|
|
127
|
+
canonicalId: reducer.agentId,
|
|
128
|
+
forbiddenIdentityTokens: [reducer.agentId],
|
|
129
|
+
}));
|
|
130
|
+
if (referenceRepo.kind === "git") {
|
|
131
|
+
return {
|
|
132
|
+
baseRevisionSha: referenceRepo.baseRevisionSha,
|
|
133
|
+
competitiveCandidates,
|
|
134
|
+
target: {
|
|
135
|
+
kind: "reduce",
|
|
136
|
+
sessionId: record.sessionId,
|
|
137
|
+
},
|
|
138
|
+
reductionRecord: record,
|
|
139
|
+
};
|
|
140
|
+
}
|
|
141
|
+
return {
|
|
142
|
+
competitiveCandidates,
|
|
143
|
+
target: {
|
|
144
|
+
kind: "reduce",
|
|
145
|
+
sessionId: record.sessionId,
|
|
146
|
+
},
|
|
147
|
+
reductionRecord: record,
|
|
148
|
+
referenceRepoUnavailable: {
|
|
149
|
+
reason: "message-lineage",
|
|
150
|
+
messageSessionId: referenceRepo.messageSessionId,
|
|
151
|
+
},
|
|
152
|
+
};
|
|
153
|
+
}
|
|
154
|
+
async function resolveMessageVerifyTarget(input) {
|
|
155
|
+
const { root, messagesFilePath, target } = input;
|
|
156
|
+
const [record] = await readMessageRecords({
|
|
157
|
+
root,
|
|
158
|
+
messagesFilePath,
|
|
159
|
+
limit: 1,
|
|
160
|
+
predicate: (entry) => entry.sessionId === target.sessionId,
|
|
161
|
+
});
|
|
162
|
+
if (!record) {
|
|
163
|
+
throw new CliError(`Message session \`${target.sessionId}\` not found.`, [], [
|
|
164
|
+
"Re-run `voratiq message` or confirm the session id in `.voratiq/message/index.json`.",
|
|
165
|
+
]);
|
|
166
|
+
}
|
|
167
|
+
if (!TERMINAL_MESSAGE_STATUSES.includes(record.status)) {
|
|
168
|
+
throw new CliError(`Message session \`${target.sessionId}\` is not complete.`, [`Status: \`${record.status}\`.`], [
|
|
169
|
+
"Wait for the message session to finish before running `voratiq verify`.",
|
|
170
|
+
]);
|
|
171
|
+
}
|
|
172
|
+
const recipients = resolveVerifiableMessageRecipients(record);
|
|
173
|
+
if (recipients.length === 0) {
|
|
174
|
+
throw new CliError(`Message session \`${target.sessionId}\` has no verifiable message responses.`, [
|
|
175
|
+
"Verification requires at least one succeeded recipient with a durable `response.md` artifact.",
|
|
176
|
+
], [
|
|
177
|
+
"Re-run `voratiq message` to capture at least one succeeded response before running `voratiq verify`.",
|
|
178
|
+
]);
|
|
179
|
+
}
|
|
180
|
+
return {
|
|
181
|
+
competitiveCandidates: recipients.map((recipient) => ({
|
|
182
|
+
canonicalId: recipient.agentId,
|
|
183
|
+
forbiddenIdentityTokens: [recipient.agentId],
|
|
184
|
+
})),
|
|
185
|
+
target: {
|
|
186
|
+
kind: "message",
|
|
187
|
+
sessionId: record.sessionId,
|
|
188
|
+
},
|
|
189
|
+
messageRecord: record,
|
|
190
|
+
};
|
|
191
|
+
}
|
|
192
|
+
function collectRunCandidateIdentityTokens(options) {
|
|
193
|
+
const { runRecord, candidateId } = options;
|
|
194
|
+
const tokens = new Set();
|
|
195
|
+
tokens.add(candidateId);
|
|
196
|
+
for (const agent of runRecord.agents) {
|
|
197
|
+
if (agent.agentId !== candidateId) {
|
|
198
|
+
continue;
|
|
199
|
+
}
|
|
200
|
+
tokens.add(agent.agentId);
|
|
201
|
+
if (agent.model.trim().length > 0) {
|
|
202
|
+
tokens.add(agent.model);
|
|
203
|
+
}
|
|
204
|
+
}
|
|
205
|
+
return Array.from(tokens);
|
|
206
|
+
}
|
|
207
|
+
async function resolveReductionReferenceRepo(options) {
|
|
208
|
+
const { root, specsFilePath, runsFilePath, reductionsFilePath, messagesFilePath, verificationsFilePath, reductionRecord, seenReductionIds = new Set(), } = options;
|
|
209
|
+
if (seenReductionIds.has(reductionRecord.sessionId)) {
|
|
210
|
+
throw new CliError(`Reduction session \`${reductionRecord.sessionId}\` has a recursive target chain.`, [], [
|
|
211
|
+
"Inspect `.voratiq/reduce/index.json` and repair the reduction target metadata.",
|
|
212
|
+
]);
|
|
213
|
+
}
|
|
214
|
+
seenReductionIds.add(reductionRecord.sessionId);
|
|
215
|
+
switch (reductionRecord.target.type) {
|
|
216
|
+
case "spec": {
|
|
217
|
+
const [record] = await readSpecRecords({
|
|
218
|
+
root,
|
|
219
|
+
specsFilePath,
|
|
220
|
+
limit: 1,
|
|
221
|
+
predicate: (entry) => entry.sessionId === reductionRecord.target.id,
|
|
222
|
+
});
|
|
223
|
+
if (!record) {
|
|
224
|
+
throw new CliError(`Spec session \`${reductionRecord.target.id}\` referenced by reduction \`${reductionRecord.sessionId}\` was not found.`);
|
|
225
|
+
}
|
|
226
|
+
return {
|
|
227
|
+
kind: "git",
|
|
228
|
+
baseRevisionSha: resolveSpecBaseRevisionSha(record, {
|
|
229
|
+
ownerLabel: `reduction \`${reductionRecord.sessionId}\``,
|
|
230
|
+
}),
|
|
231
|
+
};
|
|
232
|
+
}
|
|
233
|
+
case "run": {
|
|
234
|
+
const { records } = await fetchRunsSafely({
|
|
235
|
+
root,
|
|
236
|
+
runsFilePath,
|
|
237
|
+
runId: reductionRecord.target.id,
|
|
238
|
+
filters: { includeDeleted: true },
|
|
239
|
+
}).catch((error) => {
|
|
240
|
+
if (error instanceof RunRecordNotFoundError) {
|
|
241
|
+
throw new RunNotFoundCliError(reductionRecord.target.id);
|
|
242
|
+
}
|
|
243
|
+
throw error;
|
|
244
|
+
});
|
|
245
|
+
const runRecord = records[0];
|
|
246
|
+
if (!runRecord) {
|
|
247
|
+
throw new RunNotFoundCliError(reductionRecord.target.id);
|
|
248
|
+
}
|
|
249
|
+
return {
|
|
250
|
+
kind: "git",
|
|
251
|
+
baseRevisionSha: runRecord.baseRevisionSha,
|
|
252
|
+
};
|
|
253
|
+
}
|
|
254
|
+
case "verify": {
|
|
255
|
+
const [verificationRecord] = await readVerificationRecords({
|
|
256
|
+
root,
|
|
257
|
+
verificationsFilePath,
|
|
258
|
+
limit: 1,
|
|
259
|
+
predicate: (entry) => entry.sessionId === reductionRecord.target.id,
|
|
260
|
+
});
|
|
261
|
+
if (!verificationRecord) {
|
|
262
|
+
throw new CliError(`Verification session \`${reductionRecord.target.id}\` referenced by reduction \`${reductionRecord.sessionId}\` was not found.`);
|
|
263
|
+
}
|
|
264
|
+
switch (verificationRecord.target.kind) {
|
|
265
|
+
case "run": {
|
|
266
|
+
const { records } = await fetchRunsSafely({
|
|
267
|
+
root,
|
|
268
|
+
runsFilePath,
|
|
269
|
+
runId: verificationRecord.target.sessionId,
|
|
270
|
+
filters: { includeDeleted: true },
|
|
271
|
+
}).catch((error) => {
|
|
272
|
+
if (error instanceof RunRecordNotFoundError) {
|
|
273
|
+
throw new RunNotFoundCliError(verificationRecord.target.sessionId);
|
|
274
|
+
}
|
|
275
|
+
throw error;
|
|
276
|
+
});
|
|
277
|
+
const runRecord = records[0];
|
|
278
|
+
if (!runRecord) {
|
|
279
|
+
throw new RunNotFoundCliError(verificationRecord.target.sessionId);
|
|
280
|
+
}
|
|
281
|
+
return {
|
|
282
|
+
kind: "git",
|
|
283
|
+
baseRevisionSha: runRecord.baseRevisionSha,
|
|
284
|
+
};
|
|
285
|
+
}
|
|
286
|
+
case "spec": {
|
|
287
|
+
const [record] = await readSpecRecords({
|
|
288
|
+
root,
|
|
289
|
+
specsFilePath,
|
|
290
|
+
limit: 1,
|
|
291
|
+
predicate: (entry) => entry.sessionId === verificationRecord.target.sessionId,
|
|
292
|
+
});
|
|
293
|
+
if (!record) {
|
|
294
|
+
throw new CliError(`Spec session \`${verificationRecord.target.sessionId}\` referenced by verification \`${verificationRecord.sessionId}\` was not found.`);
|
|
295
|
+
}
|
|
296
|
+
return {
|
|
297
|
+
kind: "git",
|
|
298
|
+
baseRevisionSha: resolveSpecBaseRevisionSha(record, {
|
|
299
|
+
ownerLabel: `verification \`${verificationRecord.sessionId}\``,
|
|
300
|
+
}),
|
|
301
|
+
};
|
|
302
|
+
}
|
|
303
|
+
case "reduce": {
|
|
304
|
+
const [parentReduction] = await readReductionRecords({
|
|
305
|
+
root,
|
|
306
|
+
reductionsFilePath,
|
|
307
|
+
limit: 1,
|
|
308
|
+
predicate: (entry) => entry.sessionId === verificationRecord.target.sessionId,
|
|
309
|
+
});
|
|
310
|
+
if (!parentReduction) {
|
|
311
|
+
throw new CliError(`Reduction session \`${verificationRecord.target.sessionId}\` referenced by verification \`${verificationRecord.sessionId}\` was not found.`);
|
|
312
|
+
}
|
|
313
|
+
return await resolveReductionReferenceRepo({
|
|
314
|
+
root,
|
|
315
|
+
specsFilePath,
|
|
316
|
+
runsFilePath,
|
|
317
|
+
reductionsFilePath,
|
|
318
|
+
messagesFilePath,
|
|
319
|
+
verificationsFilePath,
|
|
320
|
+
reductionRecord: parentReduction,
|
|
321
|
+
seenReductionIds,
|
|
322
|
+
});
|
|
323
|
+
}
|
|
324
|
+
case "message": {
|
|
325
|
+
const [messageRecord] = await readMessageRecords({
|
|
326
|
+
root,
|
|
327
|
+
messagesFilePath,
|
|
328
|
+
limit: 1,
|
|
329
|
+
predicate: (entry) => entry.sessionId === verificationRecord.target.sessionId,
|
|
330
|
+
});
|
|
331
|
+
if (!messageRecord) {
|
|
332
|
+
throw new CliError(`Message session \`${verificationRecord.target.sessionId}\` referenced by verification \`${verificationRecord.sessionId}\` was not found.`);
|
|
333
|
+
}
|
|
334
|
+
return messageRecord.baseRevisionSha
|
|
335
|
+
? {
|
|
336
|
+
kind: "git",
|
|
337
|
+
baseRevisionSha: messageRecord.baseRevisionSha,
|
|
338
|
+
}
|
|
339
|
+
: {
|
|
340
|
+
kind: "none",
|
|
341
|
+
messageSessionId: messageRecord.sessionId,
|
|
342
|
+
};
|
|
343
|
+
}
|
|
344
|
+
}
|
|
345
|
+
throw new CliError(`Verification session \`${verificationRecord.sessionId}\` references an unsupported target kind.`);
|
|
346
|
+
}
|
|
347
|
+
case "reduce": {
|
|
348
|
+
const [parentReduction] = await readReductionRecords({
|
|
349
|
+
root,
|
|
350
|
+
reductionsFilePath,
|
|
351
|
+
limit: 1,
|
|
352
|
+
predicate: (entry) => entry.sessionId === reductionRecord.target.id,
|
|
353
|
+
});
|
|
354
|
+
if (!parentReduction) {
|
|
355
|
+
throw new CliError(`Reduction session \`${reductionRecord.target.id}\` referenced by reduction \`${reductionRecord.sessionId}\` was not found.`);
|
|
356
|
+
}
|
|
357
|
+
return await resolveReductionReferenceRepo({
|
|
358
|
+
root,
|
|
359
|
+
specsFilePath,
|
|
360
|
+
runsFilePath,
|
|
361
|
+
reductionsFilePath,
|
|
362
|
+
messagesFilePath,
|
|
363
|
+
verificationsFilePath,
|
|
364
|
+
reductionRecord: parentReduction,
|
|
365
|
+
seenReductionIds,
|
|
366
|
+
});
|
|
367
|
+
}
|
|
368
|
+
case "message": {
|
|
369
|
+
const [messageRecord] = await readMessageRecords({
|
|
370
|
+
root,
|
|
371
|
+
messagesFilePath,
|
|
372
|
+
limit: 1,
|
|
373
|
+
predicate: (entry) => entry.sessionId === reductionRecord.target.id,
|
|
374
|
+
});
|
|
375
|
+
if (!messageRecord) {
|
|
376
|
+
throw new CliError(`Message session \`${reductionRecord.target.id}\` referenced by reduction \`${reductionRecord.sessionId}\` was not found.`);
|
|
377
|
+
}
|
|
378
|
+
return messageRecord.baseRevisionSha
|
|
379
|
+
? {
|
|
380
|
+
kind: "git",
|
|
381
|
+
baseRevisionSha: messageRecord.baseRevisionSha,
|
|
382
|
+
}
|
|
383
|
+
: {
|
|
384
|
+
kind: "none",
|
|
385
|
+
messageSessionId: messageRecord.sessionId,
|
|
386
|
+
};
|
|
387
|
+
}
|
|
388
|
+
}
|
|
389
|
+
}
|
|
390
|
+
function resolveVerifiableMessageRecipients(record) {
|
|
391
|
+
return record.recipients.flatMap((recipient) => {
|
|
392
|
+
if (recipient.status !== "succeeded" ||
|
|
393
|
+
typeof recipient.outputPath !== "string" ||
|
|
394
|
+
!isCanonicalMessageResponseArtifact(recipient.outputPath)) {
|
|
395
|
+
return [];
|
|
396
|
+
}
|
|
397
|
+
return [
|
|
398
|
+
{
|
|
399
|
+
agentId: recipient.agentId,
|
|
400
|
+
outputPath: recipient.outputPath,
|
|
401
|
+
},
|
|
402
|
+
];
|
|
403
|
+
});
|
|
404
|
+
}
|
|
405
|
+
function isCanonicalMessageResponseArtifact(outputPath) {
|
|
406
|
+
return basename(outputPath) === MESSAGE_RESPONSE_FILENAME;
|
|
407
|
+
}
|
|
408
|
+
function resolveSpecBaseRevisionSha(record, options = {}) {
|
|
409
|
+
if (record.baseRevisionSha) {
|
|
410
|
+
return record.baseRevisionSha;
|
|
411
|
+
}
|
|
412
|
+
const ownerLabel = options.ownerLabel
|
|
413
|
+
? `${options.ownerLabel} targets legacy spec session \`${record.sessionId}\``
|
|
414
|
+
: `Spec session \`${record.sessionId}\``;
|
|
415
|
+
throw new CliError(`${ownerLabel} is missing \`baseRevisionSha\`.`, [
|
|
416
|
+
"This spec record was created before base revisions were persisted for spec sessions.",
|
|
417
|
+
], [
|
|
418
|
+
"Re-run `voratiq spec` to regenerate the spec session before running `voratiq verify`.",
|
|
419
|
+
]);
|
|
420
|
+
}
|