You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Closeout audit blocks on harness handoffs and declared Node/BAML outputs #1139
In Consiliency/jev-assembler#1, the STATE gate and CI pass, but phase-loop-closeout-audit --repo . exits 1 on ignored outputs generated by the phase verification workflow. The audit reports .dev-skills/ (harness handoffs), .baml/, .cache/, baml_sdk/, dist/, and generated/baml/ as unknown_ignored. The prior STATE runner ended with dirty_worktree_conflict while its verification had an unrelated BAML gate failure; that BAML failure has since been fixed and the public gate now passes.
Repro: run npm ci, npm run bootstrap:baml, npm test, npm run verify:phase -- STATE, then phase-loop-closeout-audit --repo . in the STATE worktree. The audit exits 1 despite a clean tracked tree. .dev-skills/handoffs is written by the installed execute-phase workflow itself.
Expected
The closeout audit should recognize harness-owned handoffs and support a provenance-checked way to classify declared, deterministic tool outputs without broadly trusting every ignored .cache or dist path. Unknown ignored data must remain fail-closed.
Acceptance
.dev-skills/handoffs/** is classified as runner-owned when generated by the harness.
Project build outputs can be classified from explicit producer/plan evidence without making arbitrary ignored files safe.
Regression covers a passing Node/BAML phase verification that leaves ignored generated SDK, dist, and cache outputs, followed by a successful closeout audit.
Local recovery for Consiliency/jev-assembler#1: moved the nine ignored output groups to a private workspace backup without deleting them, reran phase-loop-closeout-audit --repo . (exit 0), committed the refreshed tracked evidence, and ran phase-loop reconcile --verification-status passed with the passing runner-owned verification.json from the relaunch. STATE is now recorded complete and Consiliency/jev-assembler#1 is ready for review. The classifier gap remains reproducible on a normal Node/BAML verification run, which recreates those outputs.
Consiliency/jev-assembler#1 merged with STATE gate and CI passing. For local closeout, we preserved generated BAML/Node outputs and harness handoffs outside the STATE worktree before the audit and worktree prune; this was a workaround, not a classifier fix. The requested provenance-aware classification and regression in this issue remain open.
Reproduced again while executing TX for Consiliency/jev-assembler#4. The phase-loop required closeout audit exited 1 and classified ignored generated outputs (.baml/, .cache/, .dev-skills/, dist/, and generated/baml/) as unknown, reporting dirty_worktree_conflict despite phase-owned source changes and captured test results. This is the same classifier behavior seen in STATE and CORE. I preserved the generated outputs outside the worktree. TX has separate substantive acceptance blockers, so this comment is limited to the repeat audit defect; it is not a request to mark TX complete.
Picked up on 2026-09-30, and an implementation lane has started. The design is provenance-based and general:
Harness handoffs are identified by the harness's own writer record, not by directory name.
Project build outputs are accepted only when a declaration in the plan or config covers them AND the runner's evidence ties them to the declared producer run.
Anything else stays unknown_ignored and fails closed.
It includes a Node/BAML-shaped regression test. A draft PR will follow through the governed board. Once it lands, Consiliency/jev-assembler phases shouldn't need the manual move-aside workaround.
While closing Consiliency/jev-assembler#10 in a newly created, initially clean owned worktree, the installed phase-loop-closeout-audit returned exit 1 for actual outputs of frozen npm installation, TypeScript build, the pinned BAML compatibility check and resolver-generated handoffs. Runtime source is release 0.7.21; the same classifier on merged Consiliency/agent-harness main recognizes Python caches and node_modules but has no measured producer mechanism for these outputs.
Actual buckets: runner_owned .phase-loop/; tool_cache node_modules/; unknown_ignored .baml/.gitignore, .baml/cache/, .baml/profiles-v1.lock, .baml/profiles-v1/, .cache/, .dev-skills/, dist/, generated/baml/. No contents, profile values or credentials were inspected. The target plan and actual sealed commands own/generated these outputs. All current source hashes and runner validation pass.
The module also ignores --help and audits the current directory instead, returning a failure without usage text.
Desired bounded fix
Add supported producer attribution for measured Node/BAML build outputs and resolver-owned handoffs, with deny-by-default path containment. Do not broadly trust arbitrary .cache contents or files/symlinks borrowing a cache name. Keep unknown user/private outputs blocking and add negative cases alongside the real tooling cases. Give --help an actual read-only usage response.
Current workaround
Preserve only this run's generated outputs in its declared .phase-loop/runs/... archive, recording original paths and producers; rerun the unchanged classifier. No classifier verdict, evidence seal, credential value or source code is changed. This is a separate runtime usability gap; it does not add source scope to the fixture repair.
Activity
ViperJuice commentedon Sep 28, 2026
Local recovery for Consiliency/jev-assembler#1: moved the nine ignored output groups to a private workspace backup without deleting them, reran
phase-loop-closeout-audit --repo .(exit 0), committed the refreshed tracked evidence, and ranphase-loop reconcile --verification-status passedwith the passing runner-ownedverification.jsonfrom the relaunch. STATE is now recorded complete and Consiliency/jev-assembler#1 is ready for review. The classifier gap remains reproducible on a normal Node/BAML verification run, which recreates those outputs.ViperJuice commentedon Sep 28, 2026
Consiliency/jev-assembler#1 merged with STATE gate and CI passing. For local closeout, we preserved generated BAML/Node outputs and harness handoffs outside the STATE worktree before the audit and worktree prune; this was a workaround, not a classifier fix. The requested provenance-aware classification and regression in this issue remain open.
ViperJuice commentedon Sep 28, 2026
Reproduced again while executing TX for Consiliency/jev-assembler#4. The phase-loop required closeout audit exited 1 and classified ignored generated outputs (
.baml/,.cache/,.dev-skills/,dist/, andgenerated/baml/) as unknown, reportingdirty_worktree_conflictdespite phase-owned source changes and captured test results. This is the same classifier behavior seen in STATE and CORE. I preserved the generated outputs outside the worktree. TX has separate substantive acceptance blockers, so this comment is limited to the repeat audit defect; it is not a request to mark TX complete.ViperJuice commentedon Sep 30, 2026
Picked up on 2026-09-30, and an implementation lane has started. The design is provenance-based and general:
unknown_ignoredand fails closed.It includes a Node/BAML-shaped regression test. A draft PR will follow through the governed board. Once it lands, Consiliency/jev-assembler phases shouldn't need the manual move-aside workaround.
ViperJuice commentedon Oct 2, 2026
Observed gap
While closing Consiliency/jev-assembler#10 in a newly created, initially clean owned worktree, the installed
phase-loop-closeout-auditreturned exit 1 for actual outputs of frozen npm installation, TypeScript build, the pinned BAML compatibility check and resolver-generated handoffs. Runtime source is release 0.7.21; the same classifier on merged Consiliency/agent-harness main recognizes Python caches and node_modules but has no measured producer mechanism for these outputs.Actual buckets: runner_owned
.phase-loop/; tool_cachenode_modules/; unknown_ignored.baml/.gitignore,.baml/cache/,.baml/profiles-v1.lock,.baml/profiles-v1/,.cache/,.dev-skills/,dist/,generated/baml/. No contents, profile values or credentials were inspected. The target plan and actual sealed commands own/generated these outputs. All current source hashes and runner validation pass.The module also ignores
--helpand audits the current directory instead, returning a failure without usage text.Desired bounded fix
Add supported producer attribution for measured Node/BAML build outputs and resolver-owned handoffs, with deny-by-default path containment. Do not broadly trust arbitrary
.cachecontents or files/symlinks borrowing a cache name. Keep unknown user/private outputs blocking and add negative cases alongside the real tooling cases. Give--helpan actual read-only usage response.Current workaround
Preserve only this run's generated outputs in its declared
.phase-loop/runs/...archive, recording original paths and producers; rerun the unchanged classifier. No classifier verdict, evidence seal, credential value or source code is changed. This is a separate runtime usability gap; it does not add source scope to the fixture repair.ViperJuice commentedon Oct 3, 2026
Fixed by agent-harness#1189, merged as 9135190. Follow-ups are tracked separately.
Merge remote-tracking branch 'origin/main' into claude/1135-baml-v1
fix(closeout): action lines name valid closeout fields; an exception …