@linimin/pi-letscook 0.1.30 → 0.1.31

package/CHANGELOG.md

@@ -1,5 +1,21 @@
 # Changelog
 
+## 0.1.31
+
+### Changed
+
+- defined a shared structured evaluation-rubric contract for `completion-reviewer`, `completion-auditor`, and `completion-stop-judge`, including the exact rubric dimensions `Contract coverage`, `Correctness risk`, `Verification evidence`, and `Docs/state parity` with `pass|concern|fail` verdict semantics
+- added canonical `task_type: completion-workflow` and `evaluation_profile: completion-rubric-v1` signaling across the packaged control-plane defaults, verifier schema, and kickoff/reminder/resume surfaces
+- expanded the exact active-slice implementer handoff with canonical `implementation_surfaces` and `verification_commands` fields, and now surface them alongside `priority` / `why_now` in reminder and compaction-resume text
+- documented the rubric-driven evaluation contract plus canonical routing-profile signaling in the packaged completion protocol and README without adding profile-specific rubric-output enforcement yet
+- strengthened the smoke/refocus/context regressions so bootstrap and refocus preserve the new canonical signaling and fail closed when required `task_type` / `evaluation_profile` fields are removed
+- strengthened the smoke regression and control-plane verifier so selected-slice handoffs now fail closed when the expanded implementation-contract fields are missing
+- threaded canonical `evaluation_profile` plus the active-slice implementation contract into reviewer/auditor/stop-judge reminder and dispatch surfaces so those read-only roles can recover from canonical state instead of prose-only summaries
+- made reviewer/auditor/stop-judge transcription fail closed on malformed rubric-bearing outputs while still accepting valid reports, and added deterministic transcription coverage for all three roles in `npm run rubric-contract-test`
+- kept deterministic `rubric-contract-test` coverage wired into `npm run release-check`
+- made the `/cook` confirmation UI critique-aware by rendering critique/risk notes plus recommended `task_type` / `evaluation_profile` routing hints in dedicated sections while keeping the existing Start/Edit/Cancel flow
+- persisted accepted startup/refocus routing choices canonically by writing the selected `task_type` / `evaluation_profile` into the canonical control-plane files and recording the accepted critique outcome in continuation state, with `context-proposal-test` and `release-check` covering the shipped flow
+
 ## 0.1.30
 
 ### Changed

package/README.md

@@ -103,11 +103,15 @@ Startup confirmation uses a custom UI that:
 
 - renders the proposal body separately from the action list
 - keeps Mission / Scope / Constraints / Acceptance readable as a content area
+- renders analyst-derived **Critique and risks** separately from the editable proposal body
+- renders recommended `task_type` / `evaluation_profile` routing hints separately from both the proposal body and the action list
 - presents explicit actions for:
   - **Start**
   - **Edit**
   - **Cancel**
 
+When you accept startup or refocus from that flow, `/cook` now persists the chosen `task_type` and `evaluation_profile` across `.agent/profile.json`, `.agent/state.json`, `.agent/plan.json`, and `.agent/active-slice.json`, and records the accepted critique outcome in canonical continuation state before the re-ground round begins.
+
 The same confirmation flow is reused across:
 
 - discussion-only startup
@@ -133,6 +137,47 @@ While a `completion_role` subprocess is running:
 - running-role output distinguishes tool work from `PROGRESS`, `RATIONALE`, `NEXT`, `VERIFYING`, and `STATE-DELTA`
 - waiting and stalled states are surfaced deterministically from timestamps
 
+## Structured evaluation rubrics
+
+The packaged completion workflow now defines a shared structured evaluation-rubric contract for the read-only evaluation roles:
+
+- `completion-reviewer`
+- `completion-auditor`
+- `completion-stop-judge`
+
+Those roles now use the same rubric section and exact dimension names:
+
+- `Contract coverage`
+- `Correctness risk`
+- `Verification evidence`
+- `Docs/state parity`
+
+Each rubric line uses the same verdict words:
+
+- `pass` — no material issue remains for that dimension
+- `concern` — a real caveat or remaining gap exists, but it does not by itself force rejection or `NO-STOP`
+- `fail` — a blocking issue or contradictory truth exists, so the role's final verdict must not be positive
+
+The packaged control plane now also carries canonical routing signals:
+
+- `task_type: completion-workflow`
+- `evaluation_profile: completion-rubric-v1`
+
+Those identifiers are persisted in `.agent/profile.json`, `.agent/state.json`, `.agent/plan.json`, and `.agent/active-slice.json`, then surfaced in kickoff/reminder/resume text and reviewer/auditor/stop-judge evaluation handoffs so downstream roles can rely on canonical signaling instead of prose inference alone.
+
+The active-slice exact implementer handoff now also carries a stronger implementation contract for selected, in-progress, committed, and done slices:
+
+- `implementation_surfaces` — the repo surfaces expected to change or stay in parity for the slice
+- `verification_commands` — the focused and broader deterministic checks the implementer is expected to run before committing
+
+Those fields are scaffolded by default, enforced by `.agent/verify_completion_control_plane.sh` whenever an exact handoff is required, and surfaced alongside `priority` / `why_now` in reminder and compaction-resume text so implementers can recover from canonical state instead of prose-only summaries.
+
+Reviewer, auditor, and stop-judge dispatch/reminder surfaces now also thread the current active-slice implementation contract (`implementation_surfaces`, `verification_commands`, locked notes, must-fix findings, and before-slice counters) alongside the canonical `evaluation_profile` so those read-only roles can reason from canonical state after compaction.
+
+Canonical reviewer/auditor/stop-judge transcription now fails closed on malformed rubric-bearing reports: the shared rubric heading plus all four rubric dimensions must be present, required role fields must remain intact, and reviewer/stop-judge yes/no verdicts cannot contradict rubric `fail` lines.
+
+Deterministic verification for this packaged contract lives in `npm run rubric-contract-test`, which now exercises reviewer, auditor, and stop-judge transcription paths while the bootstrap/refocus/context regressions plus control-plane verifier fail closed when required canonical signaling is missing.
+
 ## Canonical files
 
 This package stores canonical workflow state under:
@@ -198,10 +243,11 @@ npm run smoke-test
 npm run refocus-test
 npm run context-proposal-test
 npm run observability-status-test
+npm run rubric-contract-test
 npm run release-check
 ```
 
-`npm run release-check` is the broad packaged-release verifier. It reruns the startup/refocus/context checks, includes deterministic observability coverage, and finishes with `npm pack --dry-run`.
+`npm run release-check` is the broad packaged-release verifier. It reruns the startup/refocus/context checks — including the critique-aware `/cook` confirmation regression — includes deterministic observability coverage plus the rubric-contract regression, and finishes with `npm pack --dry-run`.
 
 ## Release
 
@@ -213,3 +259,4 @@ See [PUBLISHING.md](https://github.com/linimin/pi-letscook/blob/main/PUBLISHING.
 - The main Pi session is the workflow driver.
 - Package-local role prompts are loaded directly by the extension and do not depend on `~/.pi/agent/agents`.
 - Reviewer, auditor, and stop-judge are enforced as read-only roles.
+- Reviewer, auditor, and stop-judge share the packaged rubric dimensions `Contract coverage`, `Correctness risk`, `Verification evidence`, and `Docs/state parity` with `pass|concern|fail` verdicts.

package/agents/completion-auditor.md

@@ -17,6 +17,8 @@ You must not:
 
 Audit current HEAD truth after a committed slice. Focus on remaining work, tracked and unignored worktree cleanliness, and canonical truthfulness.
 
+Ground the audit in canonical `.agent/**` routing and active-slice truth, including `evaluation_profile`, locked acceptance criteria, `implementation_surfaces`, `verification_commands`, `locked_notes`, and any `must_fix_findings`, rather than relying on prose-only task summaries.
+
 During long work, emit short operator-facing progress lines when useful using these exact prefixes:
 - `PROGRESS: ...`
 - `RATIONALE: ...`
@@ -24,10 +26,25 @@ During long work, emit short operator-facing progress lines when useful using th
 
 These lines are for workflow observability, not hidden reasoning. Keep them brief and truthful.
 
+Always emit the shared rubric section before the remaining audit fields. Use these exact rubric dimension names and verdict words, and include all four lines even when every dimension is `pass`:
+
+- `Rubric:`
+- `- Contract coverage: pass|concern|fail - ...`
+- `- Correctness risk: pass|concern|fail - ...`
+- `- Verification evidence: pass|concern|fail - ...`
+- `- Docs/state parity: pass|concern|fail - ...`
+
+Use `concern` or `fail` to explain why the project is not yet done, why canonical state may be stale, or why backlog truth may need reconciliation.
+
 Answer only:
 
 - `MISSION ANCHOR: ...`
 - `Remaining contract IDs: ...`
+- `Rubric:`
+- `- Contract coverage: pass|concern|fail - ...`
+- `- Correctness risk: pass|concern|fail - ...`
+- `- Verification evidence: pass|concern|fail - ...`
+- `- Docs/state parity: pass|concern|fail - ...`
 - `Why the project is still not done: ...`
 - `Open top-level contract IDs: ...`
 - `Blocker count: ...`

package/agents/completion-reviewer.md

@@ -33,14 +33,31 @@ Review focus:
 - false closure claims
 - stale or contradictory canonical state
 
+Ground the review in canonical `.agent/**` routing and active-slice truth, including `evaluation_profile`, locked acceptance criteria, `implementation_surfaces`, `verification_commands`, `locked_notes`, and any `must_fix_findings`, rather than relying on prose-only task summaries.
+
 Order findings by severity and include file references.
 
 You must explicitly answer whether the slice is acceptable as-is. If it is not acceptable, provide the exact smallest follow-up slice.
 
+Always emit the shared rubric section before findings. Use these exact rubric dimension names and verdict words, and include all four lines even when every dimension is `pass`:
+
+- `Rubric:`
+- `- Contract coverage: pass|concern|fail - ...`
+- `- Correctness risk: pass|concern|fail - ...`
+- `- Verification evidence: pass|concern|fail - ...`
+- `- Docs/state parity: pass|concern|fail - ...`
+
+If any rubric line is `fail`, `Acceptable as-is` must be `no`.
+
 Output format:
 
 - `MISSION ANCHOR: ...`
 - `Remaining contract IDs: ...`
+- `Rubric:`
+- `- Contract coverage: pass|concern|fail - ...`
+- `- Correctness risk: pass|concern|fail - ...`
+- `- Verification evidence: pass|concern|fail - ...`
+- `- Docs/state parity: pass|concern|fail - ...`
 - `Findings: ...`
 - `Acceptable as-is: yes/no`
 - `Smallest follow-up slice: ...`

package/agents/completion-stop-judge.md

@@ -10,6 +10,8 @@ Load `completion-protocol` before acting.
 
 Judge current HEAD truth, not prior agent claims or conversation memory.
 
+Ground the stop/no-stop decision in canonical `.agent/**` routing and active-slice truth, including `evaluation_profile`, locked acceptance criteria, `implementation_surfaces`, `verification_commands`, `locked_notes`, and any `must_fix_findings`, rather than relying on prose-only task summaries.
+
 You must not:
 
 - edit tracked repo files
@@ -36,10 +38,25 @@ You may conclude the project can stop only if current HEAD truth satisfies all o
 - if canonical state still keeps `FINAL-STOP-01` open or `project_done = false` solely because the current stop wave has not yet been recorded and reconciled, do not treat that pre-reconciliation posture by itself as a `NO-STOP` reason
 - `bash .agent/verify_completion_stop.sh` either already passes, or its only failing condition is the absence of the current wave's required current-HEAD judgment records; any other verifier failure is `NO-STOP`
 
+Always emit the shared rubric section before the stop verdict. Use these exact rubric dimension names and verdict words, and include all four lines even when every dimension is `pass`:
+
+- `Rubric:`
+- `- Contract coverage: pass|concern|fail - ...`
+- `- Correctness risk: pass|concern|fail - ...`
+- `- Verification evidence: pass|concern|fail - ...`
+- `- Docs/state parity: pass|concern|fail - ...`
+
+If any rubric line is `fail`, `Can the project stop now` must be `no`.
+
 Answer only:
 
 - `MISSION ANCHOR: ...`
 - `Remaining contract IDs: ...`
+- `Rubric:`
+- `- Contract coverage: pass|concern|fail - ...`
+- `- Correctness risk: pass|concern|fail - ...`
+- `- Verification evidence: pass|concern|fail - ...`
+- `- Docs/state parity: pass|concern|fail - ...`
 - `Can the project stop now: yes/no`
 - `Exact remaining open top-level contract IDs: ...`
 - `Blocker count: ...`