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 +1,15 @@
|
|
|
1
|
+
import type { EnvironmentConfig } from "../../../configs/environment/types.js";
|
|
2
|
+
import type { VerificationConfig } from "../../../configs/verification/types.js";
|
|
3
|
+
import type { VerifyProgressRenderer } from "../../../render/transcripts/verify.js";
|
|
4
|
+
import type { VerificationRecordMutators } from "../model/mutators.js";
|
|
5
|
+
import type { VerificationRecord } from "../model/types.js";
|
|
6
|
+
import type { ResolvedVerificationTarget } from "./target.js";
|
|
7
|
+
export declare function executeAndPersistProgrammaticMethod(options: {
|
|
8
|
+
root: string;
|
|
9
|
+
verificationId: string;
|
|
10
|
+
resolvedTarget: ResolvedVerificationTarget;
|
|
11
|
+
verificationConfig: VerificationConfig;
|
|
12
|
+
environment: EnvironmentConfig;
|
|
13
|
+
mutators: VerificationRecordMutators;
|
|
14
|
+
renderer?: VerifyProgressRenderer;
|
|
15
|
+
}): Promise<VerificationRecord["methods"][number] | undefined>;
|
|
@@ -0,0 +1,352 @@
|
|
|
1
|
+
import { mkdir, mkdtemp, readFile, rm } from "node:fs/promises";
|
|
2
|
+
import { dirname, join, resolve } from "node:path";
|
|
3
|
+
import { runPreparedWithLimit } from "../../../competition/core.js";
|
|
4
|
+
import { emitStageProgressEvent } from "../../../render/transcripts/stage-progress.js";
|
|
5
|
+
import { toErrorMessage } from "../../../utils/errors.js";
|
|
6
|
+
import { pathExists } from "../../../utils/fs.js";
|
|
7
|
+
import { createDetachedWorktree, getGitStderr, removeWorktree, runGitCommand, } from "../../../utils/git.js";
|
|
8
|
+
import { buildAgentWorkspacePaths } from "../../../workspace/layout.js";
|
|
9
|
+
import { getVerificationProgrammaticResultPath, VORATIQ_VERIFICATION_DIR, } from "../../../workspace/structure.js";
|
|
10
|
+
import { executeProgrammaticChecks } from "../programmatic/runner.js";
|
|
11
|
+
import { writeVerificationArtifact } from "./artifacts.js";
|
|
12
|
+
const PROGRAMMATIC_CANDIDATE_MAX_PARALLEL = 2;
|
|
13
|
+
export async function executeAndPersistProgrammaticMethod(options) {
|
|
14
|
+
const { root, verificationId, resolvedTarget, verificationConfig, environment, mutators, renderer, } = options;
|
|
15
|
+
const methodPlan = resolveProgrammaticMethodPlan({
|
|
16
|
+
resolvedTarget,
|
|
17
|
+
verificationConfig,
|
|
18
|
+
});
|
|
19
|
+
if (methodPlan.kind === "none") {
|
|
20
|
+
return undefined;
|
|
21
|
+
}
|
|
22
|
+
const startedAt = new Date().toISOString();
|
|
23
|
+
const artifactPath = getVerificationProgrammaticResultPath(verificationId);
|
|
24
|
+
const methodScope = "runRecord" in resolvedTarget
|
|
25
|
+
? { kind: "run" }
|
|
26
|
+
: { kind: "target" };
|
|
27
|
+
await mutators.recordMethodSnapshot({
|
|
28
|
+
method: "programmatic",
|
|
29
|
+
slug: "programmatic",
|
|
30
|
+
scope: methodScope,
|
|
31
|
+
status: "running",
|
|
32
|
+
startedAt,
|
|
33
|
+
});
|
|
34
|
+
emitStageProgressEvent(renderer, {
|
|
35
|
+
type: "stage.candidate",
|
|
36
|
+
stage: "verify",
|
|
37
|
+
candidate: {
|
|
38
|
+
methodKey: "programmatic",
|
|
39
|
+
verifierLabel: "programmatic",
|
|
40
|
+
status: "running",
|
|
41
|
+
startedAt,
|
|
42
|
+
artifactPath,
|
|
43
|
+
},
|
|
44
|
+
});
|
|
45
|
+
try {
|
|
46
|
+
if (methodPlan.kind === "run" && "runRecord" in resolvedTarget) {
|
|
47
|
+
const candidates = await runPreparedWithLimit({
|
|
48
|
+
prepared: resolvedTarget.target.candidateIds,
|
|
49
|
+
maxParallel: PROGRAMMATIC_CANDIDATE_MAX_PARALLEL,
|
|
50
|
+
executePrepared: async (candidateId) => await executeProgrammaticChecksForRunCandidate({
|
|
51
|
+
root,
|
|
52
|
+
verificationId,
|
|
53
|
+
artifactPath,
|
|
54
|
+
candidateId,
|
|
55
|
+
checks: methodPlan.checks,
|
|
56
|
+
runId: resolvedTarget.target.sessionId,
|
|
57
|
+
runRecord: resolvedTarget.runRecord,
|
|
58
|
+
baseRevisionSha: resolvedTarget.baseRevisionSha,
|
|
59
|
+
environment,
|
|
60
|
+
}),
|
|
61
|
+
});
|
|
62
|
+
await writeVerificationArtifact({
|
|
63
|
+
root,
|
|
64
|
+
artifactPath,
|
|
65
|
+
artifact: {
|
|
66
|
+
method: "programmatic",
|
|
67
|
+
generatedAt: new Date().toISOString(),
|
|
68
|
+
target: resolvedTarget.target,
|
|
69
|
+
scope: "run",
|
|
70
|
+
candidates,
|
|
71
|
+
},
|
|
72
|
+
});
|
|
73
|
+
const completedRef = {
|
|
74
|
+
method: "programmatic",
|
|
75
|
+
slug: "programmatic",
|
|
76
|
+
scope: methodScope,
|
|
77
|
+
status: "succeeded",
|
|
78
|
+
artifactPath,
|
|
79
|
+
startedAt,
|
|
80
|
+
completedAt: new Date().toISOString(),
|
|
81
|
+
};
|
|
82
|
+
await mutators.recordMethodSnapshot(completedRef);
|
|
83
|
+
emitStageProgressEvent(renderer, {
|
|
84
|
+
type: "stage.candidate",
|
|
85
|
+
stage: "verify",
|
|
86
|
+
candidate: {
|
|
87
|
+
methodKey: "programmatic",
|
|
88
|
+
verifierLabel: "programmatic",
|
|
89
|
+
status: "succeeded",
|
|
90
|
+
startedAt,
|
|
91
|
+
completedAt: completedRef.completedAt,
|
|
92
|
+
artifactPath,
|
|
93
|
+
},
|
|
94
|
+
});
|
|
95
|
+
return completedRef;
|
|
96
|
+
}
|
|
97
|
+
throw new Error(`Programmatic verification is only supported for run targets; \`${resolvedTarget.target.kind}\` targets must use rubric verification.`);
|
|
98
|
+
}
|
|
99
|
+
catch (error) {
|
|
100
|
+
const completedAt = new Date().toISOString();
|
|
101
|
+
const errorMessage = toErrorMessage(error);
|
|
102
|
+
await writeFailureProgrammaticArtifact({
|
|
103
|
+
root,
|
|
104
|
+
artifactPath,
|
|
105
|
+
resolvedTarget,
|
|
106
|
+
generatedAt: completedAt,
|
|
107
|
+
error: errorMessage,
|
|
108
|
+
});
|
|
109
|
+
const failedRef = {
|
|
110
|
+
method: "programmatic",
|
|
111
|
+
slug: "programmatic",
|
|
112
|
+
scope: methodScope,
|
|
113
|
+
status: "failed",
|
|
114
|
+
artifactPath,
|
|
115
|
+
startedAt,
|
|
116
|
+
completedAt,
|
|
117
|
+
error: errorMessage,
|
|
118
|
+
};
|
|
119
|
+
await mutators.recordMethodSnapshot(failedRef);
|
|
120
|
+
emitStageProgressEvent(renderer, {
|
|
121
|
+
type: "stage.candidate",
|
|
122
|
+
stage: "verify",
|
|
123
|
+
candidate: {
|
|
124
|
+
methodKey: "programmatic",
|
|
125
|
+
verifierLabel: "programmatic",
|
|
126
|
+
status: "failed",
|
|
127
|
+
startedAt,
|
|
128
|
+
completedAt,
|
|
129
|
+
artifactPath,
|
|
130
|
+
},
|
|
131
|
+
});
|
|
132
|
+
return failedRef;
|
|
133
|
+
}
|
|
134
|
+
}
|
|
135
|
+
async function executeProgrammaticChecksForRunCandidate(options) {
|
|
136
|
+
const { root, verificationId, artifactPath, candidateId, checks, runId, runRecord, baseRevisionSha, environment, } = options;
|
|
137
|
+
const paths = buildAgentWorkspacePaths({
|
|
138
|
+
root,
|
|
139
|
+
runId,
|
|
140
|
+
agentId: candidateId,
|
|
141
|
+
});
|
|
142
|
+
const runAgentRecord = runRecord.agents.find((agent) => agent.agentId === candidateId);
|
|
143
|
+
if (!runAgentRecord) {
|
|
144
|
+
throw new Error(`Run \`${runId}\` is missing candidate metadata for \`${candidateId}\`.`);
|
|
145
|
+
}
|
|
146
|
+
const workspace = await resolveProgrammaticCandidateWorkspace({
|
|
147
|
+
root,
|
|
148
|
+
verificationId,
|
|
149
|
+
candidateId,
|
|
150
|
+
paths,
|
|
151
|
+
runId,
|
|
152
|
+
baseRevisionSha,
|
|
153
|
+
diffCaptured: runAgentRecord.artifacts?.diffCaptured ?? false,
|
|
154
|
+
});
|
|
155
|
+
try {
|
|
156
|
+
const logsDirectory = resolve(root, dirname(artifactPath), candidateId);
|
|
157
|
+
const outcome = await executeProgrammaticChecks({
|
|
158
|
+
checks,
|
|
159
|
+
cwd: workspace.cwd,
|
|
160
|
+
root,
|
|
161
|
+
logsDirectory,
|
|
162
|
+
environment,
|
|
163
|
+
envDirectoryGuard: {
|
|
164
|
+
trustedAbsoluteRoots: workspace.trustedAbsoluteRoots,
|
|
165
|
+
includeHomeForPythonStack: true,
|
|
166
|
+
failOnDirectoryPreparationError: true,
|
|
167
|
+
},
|
|
168
|
+
});
|
|
169
|
+
return {
|
|
170
|
+
candidateId,
|
|
171
|
+
results: outcome.results,
|
|
172
|
+
};
|
|
173
|
+
}
|
|
174
|
+
finally {
|
|
175
|
+
await workspace.cleanup();
|
|
176
|
+
}
|
|
177
|
+
}
|
|
178
|
+
async function resolveProgrammaticCandidateWorkspace(options) {
|
|
179
|
+
const { root, verificationId, candidateId, paths, runId, baseRevisionSha, diffCaptured, } = options;
|
|
180
|
+
if (await pathExists(paths.workspacePath)) {
|
|
181
|
+
return {
|
|
182
|
+
cwd: paths.workspacePath,
|
|
183
|
+
trustedAbsoluteRoots: [
|
|
184
|
+
paths.workspacePath,
|
|
185
|
+
paths.agentRoot,
|
|
186
|
+
paths.sandboxHomePath,
|
|
187
|
+
],
|
|
188
|
+
cleanup: async () => { },
|
|
189
|
+
};
|
|
190
|
+
}
|
|
191
|
+
const reconstructionRoot = resolve(root, ".voratiq", VORATIQ_VERIFICATION_DIR, "sessions", verificationId, "programmatic", "reconstructed");
|
|
192
|
+
await mkdir(reconstructionRoot, { recursive: true });
|
|
193
|
+
const candidateRoot = await mkdtemp(join(reconstructionRoot, `${candidateId}-`));
|
|
194
|
+
const reconstructedWorkspacePath = resolve(candidateRoot, "workspace");
|
|
195
|
+
let worktreeCreated = false;
|
|
196
|
+
try {
|
|
197
|
+
try {
|
|
198
|
+
await createDetachedWorktree({
|
|
199
|
+
root,
|
|
200
|
+
worktreePath: reconstructedWorkspacePath,
|
|
201
|
+
baseRevision: baseRevisionSha,
|
|
202
|
+
});
|
|
203
|
+
}
|
|
204
|
+
catch (error) {
|
|
205
|
+
throw new Error(`Failed to reconstruct temporary verification workspace for run \`${runId}\` candidate \`${candidateId}\` at base \`${baseRevisionSha}\`: ${toErrorMessage(error)}`);
|
|
206
|
+
}
|
|
207
|
+
worktreeCreated = true;
|
|
208
|
+
if (diffCaptured) {
|
|
209
|
+
const diffAbsolutePath = resolve(paths.artifactsPath, "diff.patch");
|
|
210
|
+
if (!(await pathExists(diffAbsolutePath))) {
|
|
211
|
+
throw new Error(`Run \`${runId}\` candidate \`${candidateId}\` is missing required durable diff artifact \`${diffAbsolutePath}\`.`);
|
|
212
|
+
}
|
|
213
|
+
const diffContent = await readFile(diffAbsolutePath, "utf8");
|
|
214
|
+
if (diffContent.trim().length > 0) {
|
|
215
|
+
await applyDurableCandidatePatch({
|
|
216
|
+
runId,
|
|
217
|
+
candidateId,
|
|
218
|
+
workspacePath: reconstructedWorkspacePath,
|
|
219
|
+
diffAbsolutePath,
|
|
220
|
+
});
|
|
221
|
+
}
|
|
222
|
+
}
|
|
223
|
+
}
|
|
224
|
+
catch (error) {
|
|
225
|
+
if (worktreeCreated) {
|
|
226
|
+
await removeWorktree({
|
|
227
|
+
root,
|
|
228
|
+
worktreePath: reconstructedWorkspacePath,
|
|
229
|
+
}).catch(() => { });
|
|
230
|
+
}
|
|
231
|
+
await rm(candidateRoot, { recursive: true, force: true }).catch(() => { });
|
|
232
|
+
throw error;
|
|
233
|
+
}
|
|
234
|
+
return {
|
|
235
|
+
cwd: reconstructedWorkspacePath,
|
|
236
|
+
trustedAbsoluteRoots: [
|
|
237
|
+
reconstructedWorkspacePath,
|
|
238
|
+
paths.agentRoot,
|
|
239
|
+
paths.sandboxHomePath,
|
|
240
|
+
],
|
|
241
|
+
cleanup: async () => {
|
|
242
|
+
await removeWorktree({
|
|
243
|
+
root,
|
|
244
|
+
worktreePath: reconstructedWorkspacePath,
|
|
245
|
+
}).catch(() => { });
|
|
246
|
+
await rm(candidateRoot, { recursive: true, force: true }).catch(() => { });
|
|
247
|
+
},
|
|
248
|
+
};
|
|
249
|
+
}
|
|
250
|
+
async function applyDurableCandidatePatch(options) {
|
|
251
|
+
const { runId, candidateId, workspacePath, diffAbsolutePath } = options;
|
|
252
|
+
try {
|
|
253
|
+
await runGitCommand(["apply", "--whitespace=nowarn", "--", diffAbsolutePath], {
|
|
254
|
+
cwd: workspacePath,
|
|
255
|
+
});
|
|
256
|
+
}
|
|
257
|
+
catch (error) {
|
|
258
|
+
const detail = getGitStderr(error) ?? toErrorMessage(error);
|
|
259
|
+
throw new Error(`Failed to apply retained diff for run \`${runId}\` candidate \`${candidateId}\`: ${detail}`);
|
|
260
|
+
}
|
|
261
|
+
}
|
|
262
|
+
function resolveProgrammaticMethodPlan(options) {
|
|
263
|
+
const { resolvedTarget, verificationConfig } = options;
|
|
264
|
+
if (resolvedTarget.target.kind === "run") {
|
|
265
|
+
const checks = verificationConfig.run.programmatic.reduce((plan, entry) => {
|
|
266
|
+
const command = entry.command?.trim();
|
|
267
|
+
if (command) {
|
|
268
|
+
plan.push({ slug: entry.slug, command });
|
|
269
|
+
}
|
|
270
|
+
return plan;
|
|
271
|
+
}, []);
|
|
272
|
+
if (checks.length === 0) {
|
|
273
|
+
return { kind: "none" };
|
|
274
|
+
}
|
|
275
|
+
return {
|
|
276
|
+
kind: "run",
|
|
277
|
+
checks,
|
|
278
|
+
};
|
|
279
|
+
}
|
|
280
|
+
if (resolvedTarget.target.kind === "message") {
|
|
281
|
+
return { kind: "none" };
|
|
282
|
+
}
|
|
283
|
+
const configuredChecks = extractUnsupportedProgrammaticChecks(resolvedTarget.target.kind === "spec"
|
|
284
|
+
? verificationConfig.spec
|
|
285
|
+
: verificationConfig.reduce);
|
|
286
|
+
if (configuredChecks.length === 0) {
|
|
287
|
+
return { kind: "none" };
|
|
288
|
+
}
|
|
289
|
+
const checks = configuredChecks.reduce((plan, entry) => {
|
|
290
|
+
const command = entry.command?.trim();
|
|
291
|
+
if (command) {
|
|
292
|
+
plan.push({ slug: entry.slug, command });
|
|
293
|
+
}
|
|
294
|
+
return plan;
|
|
295
|
+
}, []);
|
|
296
|
+
if (checks.length === 0) {
|
|
297
|
+
return { kind: "none" };
|
|
298
|
+
}
|
|
299
|
+
return {
|
|
300
|
+
kind: "unsupported",
|
|
301
|
+
targetKind: resolvedTarget.target.kind,
|
|
302
|
+
checks,
|
|
303
|
+
};
|
|
304
|
+
}
|
|
305
|
+
function extractUnsupportedProgrammaticChecks(stageConfig) {
|
|
306
|
+
const rawEntries = Reflect.get(stageConfig, "programmatic");
|
|
307
|
+
if (!Array.isArray(rawEntries)) {
|
|
308
|
+
return [];
|
|
309
|
+
}
|
|
310
|
+
return rawEntries.flatMap((entry) => {
|
|
311
|
+
if (!entry || typeof entry !== "object") {
|
|
312
|
+
return [];
|
|
313
|
+
}
|
|
314
|
+
const slugValue = Reflect.get(entry, "slug");
|
|
315
|
+
const commandValue = Reflect.get(entry, "command");
|
|
316
|
+
const slug = typeof slugValue === "string" ? slugValue : undefined;
|
|
317
|
+
const command = typeof commandValue === "string" ? commandValue : undefined;
|
|
318
|
+
return slug ? [{ slug, ...(command ? { command } : {}) }] : [];
|
|
319
|
+
});
|
|
320
|
+
}
|
|
321
|
+
async function writeFailureProgrammaticArtifact(options) {
|
|
322
|
+
const { root, artifactPath, resolvedTarget, generatedAt, error } = options;
|
|
323
|
+
if ("runRecord" in resolvedTarget) {
|
|
324
|
+
await writeVerificationArtifact({
|
|
325
|
+
root,
|
|
326
|
+
artifactPath,
|
|
327
|
+
artifact: {
|
|
328
|
+
method: "programmatic",
|
|
329
|
+
generatedAt,
|
|
330
|
+
status: "failed",
|
|
331
|
+
error,
|
|
332
|
+
target: resolvedTarget.target,
|
|
333
|
+
scope: "run",
|
|
334
|
+
candidates: [],
|
|
335
|
+
},
|
|
336
|
+
});
|
|
337
|
+
return;
|
|
338
|
+
}
|
|
339
|
+
await writeVerificationArtifact({
|
|
340
|
+
root,
|
|
341
|
+
artifactPath,
|
|
342
|
+
artifact: {
|
|
343
|
+
method: "programmatic",
|
|
344
|
+
generatedAt,
|
|
345
|
+
status: "failed",
|
|
346
|
+
error,
|
|
347
|
+
target: resolvedTarget.target,
|
|
348
|
+
scope: "target",
|
|
349
|
+
results: [],
|
|
350
|
+
},
|
|
351
|
+
});
|
|
352
|
+
}
|
|
@@ -0,0 +1,19 @@
|
|
|
1
|
+
import type { ResolvedExtraContextFile } from "../../../competition/shared/extra-context.js";
|
|
2
|
+
import type { VerificationTarget } from "../model/types.js";
|
|
3
|
+
import type { StagedVerificationInputs } from "./shared-layout.js";
|
|
4
|
+
export interface RubricTemplateContents {
|
|
5
|
+
template: string;
|
|
6
|
+
prompt: string;
|
|
7
|
+
rubric: string;
|
|
8
|
+
schema: string;
|
|
9
|
+
}
|
|
10
|
+
export declare function loadRubricTemplate(options: {
|
|
11
|
+
root: string;
|
|
12
|
+
template: string;
|
|
13
|
+
}): Promise<RubricTemplateContents>;
|
|
14
|
+
export declare function buildRubricPrompt(options: {
|
|
15
|
+
template: RubricTemplateContents;
|
|
16
|
+
target: VerificationTarget;
|
|
17
|
+
staged: StagedVerificationInputs;
|
|
18
|
+
extraContextFiles: readonly ResolvedExtraContextFile[];
|
|
19
|
+
}): string;
|
|
@@ -0,0 +1,63 @@
|
|
|
1
|
+
import { readFile } from "node:fs/promises";
|
|
2
|
+
import { resolve } from "node:path";
|
|
3
|
+
import { appendConstraints, appendOutputRequirements, buildWorkspaceArtifactRequirements, } from "../../../competition/shared/prompt-helpers.js";
|
|
4
|
+
import { toExtraContextContextSubpath } from "../../../extra-context/contract.js";
|
|
5
|
+
import { VORATIQ_VERIFICATION_DIR } from "../../../workspace/structure.js";
|
|
6
|
+
export async function loadRubricTemplate(options) {
|
|
7
|
+
const { root, template } = options;
|
|
8
|
+
const base = resolve(root, ".voratiq", VORATIQ_VERIFICATION_DIR, "templates", template);
|
|
9
|
+
const [prompt, rubric, schema] = await Promise.all([
|
|
10
|
+
readFile(resolve(base, "prompt.md"), "utf8"),
|
|
11
|
+
readFile(resolve(base, "rubric.md"), "utf8"),
|
|
12
|
+
readFile(resolve(base, "schema.yaml"), "utf8"),
|
|
13
|
+
]);
|
|
14
|
+
return { template, prompt, rubric, schema };
|
|
15
|
+
}
|
|
16
|
+
export function buildRubricPrompt(options) {
|
|
17
|
+
const { template, target, staged, extraContextFiles } = options;
|
|
18
|
+
const lines = [];
|
|
19
|
+
lines.push(`# Verifier Template: ${template.template}`, "", template.prompt.trimEnd(), "", template.rubric.trimEnd(), "", "## Inputs");
|
|
20
|
+
if (staged.kind === "spec") {
|
|
21
|
+
lines.push(`- Base repository snapshot (read-only): \`${staged.referenceRepoPath}/\``, `- Original description: \`${staged.descriptionPath}\``, "- Drafts:", ...staged.candidates.map((candidate) => {
|
|
22
|
+
const parts = [`spec: \`${candidate.specPath}\``];
|
|
23
|
+
if (candidate.specDataPath) {
|
|
24
|
+
parts.push(`metadata: \`${candidate.specDataPath}\``);
|
|
25
|
+
}
|
|
26
|
+
return ` - ${candidate.alias} (${parts.join(", ")})`;
|
|
27
|
+
}));
|
|
28
|
+
}
|
|
29
|
+
else if (staged.kind === "run") {
|
|
30
|
+
lines.push(`- Base repository snapshot (read-only): \`${staged.referenceRepoPath}/\``, `- Selected spec: \`${staged.specPath}\``, "- Candidates:", ...staged.candidates.map((candidate) => {
|
|
31
|
+
const parts = [];
|
|
32
|
+
if (candidate.diffPath)
|
|
33
|
+
parts.push(`diff: \`${candidate.diffPath}\``);
|
|
34
|
+
if (candidate.summaryPath) {
|
|
35
|
+
parts.push(`summary: \`${candidate.summaryPath}\``);
|
|
36
|
+
}
|
|
37
|
+
const suffix = parts.length > 0 ? ` (${parts.join(", ")})` : "";
|
|
38
|
+
return ` - ${candidate.alias}${suffix}`;
|
|
39
|
+
}));
|
|
40
|
+
}
|
|
41
|
+
else if (staged.kind === "message") {
|
|
42
|
+
lines.push(`- Original message prompt: \`${staged.promptPath}\``, "- Candidate responses:", ...staged.candidates.map((candidate) => ` - ${candidate.alias}: \`${candidate.responsePath}\``));
|
|
43
|
+
}
|
|
44
|
+
else if (staged.kind === "reduce-no-reference") {
|
|
45
|
+
lines.push(`- No repository snapshot is available for this reduction target because its lineage reaches message session \`${staged.referenceRepoUnavailable.messageSessionId}\` without a durable base revision.`, "- Reduction candidates:", ...staged.candidates.map((candidate) => ` - ${candidate.alias}: \`${candidate.reductionPath}\``));
|
|
46
|
+
}
|
|
47
|
+
else {
|
|
48
|
+
lines.push(`- Base repository snapshot (read-only): \`${staged.referenceRepoPath}/\``, "- Reduction candidates:", ...staged.candidates.map((candidate) => ` - ${candidate.alias}: \`${candidate.reductionPath}\``));
|
|
49
|
+
}
|
|
50
|
+
lines.push("", "## Output schema", "Return JSON matching this YAML schema exactly:", "", "```yaml", template.schema.trimEnd(), "```");
|
|
51
|
+
appendConstraints(lines);
|
|
52
|
+
appendOutputRequirements(lines, buildWorkspaceArtifactRequirements([
|
|
53
|
+
{
|
|
54
|
+
instruction: "Save the structured rubric result as JSON",
|
|
55
|
+
path: "result.json",
|
|
56
|
+
},
|
|
57
|
+
]));
|
|
58
|
+
lines.push("", "Target metadata:", `- kind: ${target.kind}`, `- sessionId: ${target.sessionId}`);
|
|
59
|
+
if (extraContextFiles.length > 0) {
|
|
60
|
+
lines.push("", "Extra context files:", ...extraContextFiles.map((file) => `- \`context/${toExtraContextContextSubpath(file.stagedRelativePath)}\` (source: \`${file.displayPath}\`)`), "- Treat these files as supplemental context for this invocation.");
|
|
61
|
+
}
|
|
62
|
+
return `${lines.join("\n")}\n`;
|
|
63
|
+
}
|
|
@@ -0,0 +1,23 @@
|
|
|
1
|
+
import type { ResolvedExtraContextFile } from "../../../competition/shared/extra-context.js";
|
|
2
|
+
import type { TeardownController } from "../../../competition/shared/teardown.js";
|
|
3
|
+
import type { AgentDefinition } from "../../../configs/agents/types.js";
|
|
4
|
+
import type { EnvironmentConfig } from "../../../configs/environment/types.js";
|
|
5
|
+
import type { VerificationConfig } from "../../../configs/verification/types.js";
|
|
6
|
+
import type { VerifyProgressRenderer } from "../../../render/transcripts/verify.js";
|
|
7
|
+
import type { VerificationRecordMutators } from "../model/mutators.js";
|
|
8
|
+
import type { VerificationRecord } from "../model/types.js";
|
|
9
|
+
import type { ResolvedVerificationTarget } from "./target.js";
|
|
10
|
+
export declare function executeAndPersistRubricMethods(options: {
|
|
11
|
+
root: string;
|
|
12
|
+
verificationId: string;
|
|
13
|
+
resolvedTarget: ResolvedVerificationTarget;
|
|
14
|
+
verificationConfig: VerificationConfig;
|
|
15
|
+
verifierAgents: readonly AgentDefinition[];
|
|
16
|
+
aliasMap?: Record<string, string>;
|
|
17
|
+
environment: EnvironmentConfig;
|
|
18
|
+
extraContextFiles: readonly ResolvedExtraContextFile[];
|
|
19
|
+
maxParallel: number;
|
|
20
|
+
teardown: TeardownController;
|
|
21
|
+
mutators: VerificationRecordMutators;
|
|
22
|
+
renderer?: VerifyProgressRenderer;
|
|
23
|
+
}): Promise<VerificationRecord["methods"]>;
|
|
@@ -0,0 +1,77 @@
|
|
|
1
|
+
import { executeCompetitionWithAdapter } from "../../../competition/command-adapter.js";
|
|
2
|
+
import { createVerifyCompetitionAdapter, } from "./adapter.js";
|
|
3
|
+
import { loadRubricTemplate } from "./prompt.js";
|
|
4
|
+
import { prepareSharedVerificationInputs } from "./shared-layout.js";
|
|
5
|
+
export async function executeAndPersistRubricMethods(options) {
|
|
6
|
+
const { root, verificationId, resolvedTarget, verificationConfig, verifierAgents, aliasMap, environment, extraContextFiles, maxParallel, teardown, mutators, renderer, } = options;
|
|
7
|
+
const rubricTemplates = resolvedTarget.target.kind === "spec"
|
|
8
|
+
? verificationConfig.spec.rubric
|
|
9
|
+
: resolvedTarget.target.kind === "run"
|
|
10
|
+
? verificationConfig.run.rubric
|
|
11
|
+
: resolvedTarget.target.kind === "reduce"
|
|
12
|
+
? verificationConfig.reduce.rubric
|
|
13
|
+
: verificationConfig.message.rubric;
|
|
14
|
+
if (resolvedTarget.target.kind === "message" &&
|
|
15
|
+
rubricTemplates.length === 0) {
|
|
16
|
+
throw new Error("Message verification requires at least one configured rubric template under `message.rubric` in `.voratiq/verification.yaml`.");
|
|
17
|
+
}
|
|
18
|
+
if (rubricTemplates.length === 0 || verifierAgents.length === 0) {
|
|
19
|
+
return [];
|
|
20
|
+
}
|
|
21
|
+
const sharedInputs = await prepareSharedVerificationInputs({
|
|
22
|
+
root,
|
|
23
|
+
verificationId,
|
|
24
|
+
resolvedTarget,
|
|
25
|
+
environment,
|
|
26
|
+
aliasMap,
|
|
27
|
+
});
|
|
28
|
+
for (const worktreePath of sharedInputs.worktreesToRemove) {
|
|
29
|
+
teardown.addWorktree({
|
|
30
|
+
root,
|
|
31
|
+
worktreePath,
|
|
32
|
+
label: "detached reference worktree",
|
|
33
|
+
});
|
|
34
|
+
}
|
|
35
|
+
teardown.addPath(sharedInputs.sharedRootAbsolute, "shared verification inputs");
|
|
36
|
+
const loadedTemplates = await Promise.all(rubricTemplates.map(async (rubric) => ({
|
|
37
|
+
rubric,
|
|
38
|
+
template: await loadRubricTemplate({
|
|
39
|
+
root,
|
|
40
|
+
template: rubric.template,
|
|
41
|
+
}),
|
|
42
|
+
})));
|
|
43
|
+
const candidates = loadedTemplates.flatMap(({ template }) => verifierAgents.map((agent) => ({
|
|
44
|
+
agent,
|
|
45
|
+
template,
|
|
46
|
+
})));
|
|
47
|
+
const executions = await executeCompetitionWithAdapter({
|
|
48
|
+
candidates,
|
|
49
|
+
maxParallel,
|
|
50
|
+
adapter: createVerifyCompetitionAdapter({
|
|
51
|
+
root,
|
|
52
|
+
verificationId,
|
|
53
|
+
resolvedTarget,
|
|
54
|
+
aliasMap,
|
|
55
|
+
environment,
|
|
56
|
+
extraContextFiles,
|
|
57
|
+
sharedInputs,
|
|
58
|
+
teardown,
|
|
59
|
+
mutators,
|
|
60
|
+
renderer,
|
|
61
|
+
}),
|
|
62
|
+
});
|
|
63
|
+
return executions.map((execution) => ({
|
|
64
|
+
method: "rubric",
|
|
65
|
+
template: execution.template,
|
|
66
|
+
verifierId: execution.verifierId,
|
|
67
|
+
scope: resolvedTarget.target.kind === "run"
|
|
68
|
+
? { kind: "run" }
|
|
69
|
+
: { kind: "target" },
|
|
70
|
+
status: execution.status,
|
|
71
|
+
artifactPath: execution.artifactPath,
|
|
72
|
+
startedAt: execution.startedAt,
|
|
73
|
+
completedAt: execution.completedAt,
|
|
74
|
+
tokenUsage: execution.tokenUsage,
|
|
75
|
+
...(execution.status === "failed" ? { error: execution.error } : {}),
|
|
76
|
+
}));
|
|
77
|
+
}
|
|
@@ -0,0 +1,121 @@
|
|
|
1
|
+
import type { EnvironmentConfig } from "../../../configs/environment/types.js";
|
|
2
|
+
import { type AgentWorkspacePaths } from "../../../workspace/layout.js";
|
|
3
|
+
import type { ResolvedVerificationTarget } from "./target.js";
|
|
4
|
+
export type SharedVerificationInputs = {
|
|
5
|
+
kind: "spec";
|
|
6
|
+
sharedRootAbsolute: string;
|
|
7
|
+
sharedInputsAbsolute: string;
|
|
8
|
+
referenceRepoAbsolute: string;
|
|
9
|
+
worktreesToRemove: readonly string[];
|
|
10
|
+
candidates: readonly {
|
|
11
|
+
alias: string;
|
|
12
|
+
hasSpecData: boolean;
|
|
13
|
+
}[];
|
|
14
|
+
} | {
|
|
15
|
+
kind: "run";
|
|
16
|
+
sharedRootAbsolute: string;
|
|
17
|
+
sharedInputsAbsolute: string;
|
|
18
|
+
referenceRepoAbsolute: string;
|
|
19
|
+
worktreesToRemove: readonly string[];
|
|
20
|
+
candidates: readonly {
|
|
21
|
+
alias: string;
|
|
22
|
+
hasDiff: boolean;
|
|
23
|
+
hasSummary: boolean;
|
|
24
|
+
}[];
|
|
25
|
+
} | {
|
|
26
|
+
kind: "reduce";
|
|
27
|
+
sharedRootAbsolute: string;
|
|
28
|
+
sharedInputsAbsolute: string;
|
|
29
|
+
referenceRepoAbsolute: string;
|
|
30
|
+
worktreesToRemove: readonly string[];
|
|
31
|
+
candidates: readonly {
|
|
32
|
+
alias: string;
|
|
33
|
+
}[];
|
|
34
|
+
} | {
|
|
35
|
+
kind: "reduce-no-reference";
|
|
36
|
+
sharedRootAbsolute: string;
|
|
37
|
+
sharedInputsAbsolute: string;
|
|
38
|
+
worktreesToRemove: readonly string[];
|
|
39
|
+
referenceRepoUnavailable: {
|
|
40
|
+
reason: "message-lineage";
|
|
41
|
+
messageSessionId: string;
|
|
42
|
+
};
|
|
43
|
+
candidates: readonly {
|
|
44
|
+
alias: string;
|
|
45
|
+
}[];
|
|
46
|
+
} | {
|
|
47
|
+
kind: "message";
|
|
48
|
+
sharedRootAbsolute: string;
|
|
49
|
+
sharedInputsAbsolute: string;
|
|
50
|
+
worktreesToRemove: readonly string[];
|
|
51
|
+
promptAbsolute: string;
|
|
52
|
+
candidates: readonly {
|
|
53
|
+
alias: string;
|
|
54
|
+
}[];
|
|
55
|
+
};
|
|
56
|
+
export type StagedVerificationInputs = {
|
|
57
|
+
kind: "spec";
|
|
58
|
+
referenceRepoPath: string;
|
|
59
|
+
descriptionPath: string;
|
|
60
|
+
candidates: readonly {
|
|
61
|
+
alias: string;
|
|
62
|
+
specPath: string;
|
|
63
|
+
specDataPath?: string;
|
|
64
|
+
}[];
|
|
65
|
+
} | {
|
|
66
|
+
kind: "run";
|
|
67
|
+
referenceRepoPath: string;
|
|
68
|
+
specPath: string;
|
|
69
|
+
candidates: readonly {
|
|
70
|
+
alias: string;
|
|
71
|
+
diffPath?: string;
|
|
72
|
+
summaryPath?: string;
|
|
73
|
+
}[];
|
|
74
|
+
} | {
|
|
75
|
+
kind: "reduce";
|
|
76
|
+
referenceRepoPath: string;
|
|
77
|
+
candidates: readonly {
|
|
78
|
+
alias: string;
|
|
79
|
+
reductionPath: string;
|
|
80
|
+
}[];
|
|
81
|
+
} | {
|
|
82
|
+
kind: "reduce-no-reference";
|
|
83
|
+
referenceRepoUnavailable: {
|
|
84
|
+
reason: "message-lineage";
|
|
85
|
+
messageSessionId: string;
|
|
86
|
+
};
|
|
87
|
+
candidates: readonly {
|
|
88
|
+
alias: string;
|
|
89
|
+
reductionPath: string;
|
|
90
|
+
}[];
|
|
91
|
+
} | {
|
|
92
|
+
kind: "message";
|
|
93
|
+
promptPath: string;
|
|
94
|
+
candidates: readonly {
|
|
95
|
+
alias: string;
|
|
96
|
+
responsePath: string;
|
|
97
|
+
}[];
|
|
98
|
+
};
|
|
99
|
+
export declare function prepareSharedVerificationInputs(options: {
|
|
100
|
+
root: string;
|
|
101
|
+
verificationId: string;
|
|
102
|
+
resolvedTarget: ResolvedVerificationTarget;
|
|
103
|
+
environment: EnvironmentConfig;
|
|
104
|
+
aliasMap?: Record<string, string>;
|
|
105
|
+
}): Promise<SharedVerificationInputs>;
|
|
106
|
+
export declare function cleanupSharedVerificationInputs(options: {
|
|
107
|
+
root: string;
|
|
108
|
+
sharedInputs: SharedVerificationInputs;
|
|
109
|
+
}): Promise<void>;
|
|
110
|
+
export declare function buildStagedVerificationInputs(options: {
|
|
111
|
+
workspacePaths: AgentWorkspacePaths;
|
|
112
|
+
sharedInputs: SharedVerificationInputs;
|
|
113
|
+
}): StagedVerificationInputs;
|
|
114
|
+
export declare function sharedInputsUseReferenceRepo(sharedInputs: SharedVerificationInputs): sharedInputs is Extract<SharedVerificationInputs, {
|
|
115
|
+
kind: "spec" | "run" | "reduce";
|
|
116
|
+
}>;
|
|
117
|
+
export declare function attachVerifierWorkspaceMounts(options: {
|
|
118
|
+
workspacePath: string;
|
|
119
|
+
contextPath: string;
|
|
120
|
+
sharedInputs: SharedVerificationInputs;
|
|
121
|
+
}): Promise<void>;
|