[Phase 2] Register skill_candidate artifact type (skill rubric + pass/score channel)

[github-actions]

Story

As a Epic #581's strict-improvement skill gate,
I want a skill_candidate artifact type registered with the skill eval rubric and a pass/score output channel,
so that a candidate skill edit is reviewed by the same review brain and gated on a strict improvement.

Acceptance Criteria

1. A skill_candidate registry entry maps to a skill rubric (Epic Initiative: Eval-gated, human-reviewed self-improving skills (SkillOpt-style) #581's eval criteria) and a pass/score output channel.
2. The content_ref adapter accepts a candidate skill edit (diff or file) and presents it to the rubric via scripts/engine.sh.
3. The output channel emits a pass/fail + score the gate can consume (not a GitHub PR review).
4. pr_diff and plan_json behavior are untouched; the registration is purely additive.
5. bats coverage exercises skill_candidate resolution and a sample producing a score; the new test is registered in the lint.yml bats list.

Tasks / Subtasks

• Author the skill rubric under prompts/ from Epic Initiative: Eval-gated, human-reviewed self-improving skills (SkillOpt-style) #581's eval criteria (strict-improvement gate), aligned with that initiative's eval-case format. (AC: test issue from agent #1)
• Add the skill_candidate registry entry {rubric, pass/score channel} and a content_ref adapter for a candidate skill edit. (AC: test issue from agent #1, Go-live improvements for PR review agent #2)
• Implement the pass/score output channel (gate signal, not a GitHub review). (AC: Add @claude delegation, auto-merge, and rebase handling #3)
• Add bats coverage, register it in lint.yml, and confirm the pr_diff and plan_json guards stay green. (AC: Optimize review: small-PR and incremental fast paths #4, feat: add Copilot engine support via REVIEW_ENGINE toggle #5)

Dev Notes

• Epic Initiative: Eval-gated, human-reviewed self-improving skills (SkillOpt-style) #581 ('Eval-gated, human-reviewed self-improving skills') defines the eval criteria; its phase issues — [Phase 1] Define held-out eval-case format + seed prompts/triage.md case set #582 (held-out eval-case format), [Phase 1] Build the deterministic eval scorer + offline tests (triage) #583 (deterministic scorer), [Phase 2] Extend the harness to prompts/deep-review.md via an LLM-judge scorer #585 (LLM-judge for prompts/deep-review.md), [Phase 2] Strict-improvement gate + manual propose->validate->PR runbook with regression-rejection proof #586 (strict-improvement gate) — are where rubric/scorer semantics live. The skill rubric here must align with that eval format, not invent a parallel one.
• Output is a gate signal (pass/score) consumed by Initiative: Eval-gated, human-reviewed self-improving skills (SkillOpt-style) #581's propose->validate->PR runbook — do not post a GitHub review and do not call post-pr-review.sh.
• Additive only: registering skill_candidate must not touch the pr_diff or plan_json registry entries.
• Reuse scripts/engine.sh routing (run_agentic) for invocation.

Project Structure Notes

New rubric prompt under prompts/ (e.g. prompts/skill-review.md); registry entry + adapter + pass/score channel in scripts/lib/. Sequenced after plan_json so the registry grows one additive type at a time.

References

• prompts/deep-review.md
• scripts/engine.sh#run_agentic
• scripts/lib/review-registry.sh
• tests/

Likely target surface

• prompts/skill-review.md
• scripts/lib/review-registry.sh
• tests/

Story prepared by the BMAD Scrum Master (Bob) for epic #610. Status: ready-for-dev.

[don-petry]

Dev-Lead Implementation Plan

Issue: #615 — [Phase 2] Register skill_candidate artifact type (skill rubric + pass/score channel)

Scope

Additively register a third review artifact_type, skill_candidate, in the rubric registry — binding a new skill-eval rubric (aligned to Epic #581's strict-improvement eval criteria) to a pass/score output channel the gate consumes — mirroring the plan_json (#614) pattern, with pr_diff and plan_json left untouched.

Implementation Checklist

• prompts/skill-review.md — new skill rubric: the review brain scores a candidate skill edit (diff or file) against Epic Initiative: Eval-gated, human-reviewed self-improving skills (SkillOpt-style) #581's eval criteria ([Phase 1] Define held-out eval-case format + seed prompts/triage.md case set #582 case format, [Phase 2] Extend the harness to prompts/deep-review.md via an LLM-judge scorer #585 judge scoring), emitting verdict (pass/fail) + numeric score [0,1] + findings. Reuses the evals/judge.md scoring vocabulary; not a parallel format.
• scripts/lib/review-registry.tsv — add one data row skill_candidate<TAB>prompts/skill-review.md<TAB>scripts/post-skill-score.sh; bump schema_version 2 → 3. pr_diff/plan_json rows unchanged.
• scripts/post-skill-score.sh — pass/score output channel (gate signal, not a GitHub review). Validates verdict ∈ {pass,fail} + numeric score ∈ [0,1], writes the score JSON to SKILL_SCORE_OUTPUT for the strict-improvement gate ([Phase 2] Strict-improvement gate + manual propose->validate->PR runbook with regression-rejection proof #586) to consume. Never calls post-pr-review.sh / gh pr review.
• scripts/evals/review-skill.sh — content_ref adapter: accepts a candidate skill edit (diff or file), resolves {rubric, output_channel} from the registry, presents it to the rubric via engine.sh run_agentic (deep tier, no new model selector), delivers via the resolved channel.
• scripts/lib/README.md — document the skill_candidate row in the artifact-contract table.

Tests to Write

• tests/test_skill_review.bats — registry resolution (skill_candidate → rubric + channel), additive guards (pr_diff AND plan_json still resolve), driver wiring (sources registry + engine.sh, calls run_agentic deep tier, never post-pr-review.sh), output-channel scoring over a sample producing a pass/fail + score, and rejection of malformed/out-of-range scores.
• Register tests/test_skill_review.bats in the lint.yml bats list (AC feat: add Copilot engine support via REVIEW_ENGINE toggle #5).

Test Command

bats tests/test_skill_review.bats (full suite via the lint.yml bats list); lint via bash .dev-lead/scripts/dev-lead-lint.sh

[don-petry]

Dev-Lead: Implementation Complete

Registered a third review artifact_type, skill_candidate, additively beside pr_diff and plan_json — mirroring the plan_json (#614) pattern.

Plan Execution

• Authored the skill rubric prompts/skill-review.md — the review brain scores a candidate skill edit (diff or file) against Epic Initiative: Eval-gated, human-reviewed self-improving skills (SkillOpt-style) #581's strict-improvement eval criteria, aligned to evals/judge.md (0–1 score) and prompts/shared.md (risk taxonomy); not a parallel format.
• Added the skill_candidate registry row {rubric: prompts/skill-review.md, output_channel: scripts/post-skill-score.sh} and bumped schema_version 2 → 3.
• Implemented the pass/score output channel scripts/post-skill-score.sh — emits a machine-readable pass/fail + numeric score the strict-improvement gate consumes via SKILL_SCORE_OUTPUT. Never a GitHub review.
• Implemented the content_ref adapter scripts/evals/review-skill.sh — presents the candidate edit to the rubric via engine.sh run_agentic (deep tier, no new model selector), delivers via the registry-resolved channel.
• Added tests/test_skill_review.bats (30 tests) and registered it in the lint.yml bats list.
• Documented the row in scripts/lib/README.md.

Acceptance Criteria

• AC1 — registry maps skill_candidate → skill rubric + pass/score channel. ✓
• AC2 — adapter accepts a candidate skill edit (diff or file), presents it to the rubric via scripts/engine.sh. ✓
• AC3 — channel emits pass/fail + score (gate signal, not a GitHub PR review). ✓
• AC4 — pr_diff and plan_json rows byte-for-byte unchanged; purely additive (guarded by tests). ✓
• AC5 — bats exercises skill_candidate resolution and a sample producing a score; registered in lint.yml. ✓

Test Results

tests/test_skill_review.bats: 1..30  — all ok (30/30)

Full lint.yml bats list (27 files): 477/477 ok
dev-lead-lint.sh: shellcheck (60 files) + agent profiles — all checks passed
lint.yml: valid YAML

Files Changed

• prompts/skill-review.md (new): strict-improvement skill rubric.
• scripts/post-skill-score.sh (new): pass/score output channel; validates verdict ∈ {pass,fail} and score ∈ [0,1].
• scripts/evals/review-skill.sh (new): content_ref adapter reusing engine.sh deep-tier routing.
• tests/test_skill_review.bats (new): 30 tests — registry resolution, additive guards for pr_diff/plan_json, driver wiring, output-channel scoring + rejection of malformed/out-of-range scores.
• scripts/lib/review-registry.tsv: +1 data row, schema_version 2 → 3.
• scripts/lib/README.md: documented the skill_candidate row and helper examples.
• .github/workflows/lint.yml: registered the new bats file.

Notes

• The output channel validates well-formedness (pass/fail + numeric in-range score) but does not cross-check verdict against the score — the strict-improvement gate ([Phase 2] Strict-improvement gate + manual propose->validate->PR runbook with regression-rejection proof #586) owns threshold/baseline comparison, consistent with how post-plan-findings.sh defers verdict semantics. No evals/ paths are touched, so the holdout guard is not engaged.