[dmccoystephenson]

Summary

Three changes to the generated dev-loop template:

1. Phase 9 self-audit becomes a structured rubric. Replaces the free-form bullet list with a 6-item rubric (identity drift, instruction clarity, edge case coverage, phase friction, drift candidates, external-signal quality) — each scored PASS / FAIL / no-signal with a one-line justification. Mirrors the Phase 4 self-review structure shipped in PR Reframe self-review as rubric-based and CI-grounded #15. Grounded in RESEARCH.md §1 and §5 (structured rubrics outperform free-form for LLM critique).

2. Filed gap issues now carry a label from a 5-label taxonomy: template-rule, repo-specific, edge-case, research-gap, process. Makes triage faster and surfaces patterns across child-skill repos.

3. Step 6 creates the labels at repo-creation time so they exist when the first cycle's Phase 9 fires. Uses a single for loop over a colon-delimited list of name:color:description entries.

Closes #24

Test plan

• Phase 9 step 2 replaced with structured rubric (6 items, each with explicit FAIL/PASS/no-signal scoring)
• Phase 9 step 3 updated to require --label argument and lists the 5-label taxonomy inline
• Step 6 has a label-creation loop with the 5 labels, hex colors, and descriptions matching the Phase 9 taxonomy verbatim
• No new {{placeholders}} introduced (universal rule — placeholder count still 22 with TEMPLATE_VERSION and GENERATED_AT from PR Add template-version header to generated skills for drift detection #33)
• Diff 44 lines across 2 files (net +28) — well under scope ceiling
• Localization verified: "Phase 9 — Self-audit" heading at line 397, "### 6 — Create a private GitHub repo" at line 535
• Per the rule shipped in PR Promote 'verify each Closes #N' to template Phase 4 #26, ran gh issue view 24 and confirmed title/body match
• Per the rule shipped in PR Track implementation outcomes back into RESEARCH.md #27, added Implementations entries under RESEARCH.md §1 and §5
• Do-not-auto-merge check (PR Add hard PR scope gates: LOC ceiling, do-not-auto-merge paths #30): modified create-dev-loop.md, RESEARCH.md; no matches

Note on label taxonomy duplication

The taxonomy is listed in two places: Phase 9 (where the agent picks one) and Step 6 (where the agent creates the labels). They must stay in sync. A future PR could DRY this — e.g. a single placeholder {{GAP_LABEL_TAXONOMY}} — but that's overengineering for 5 entries. Listed in both spots verbatim; renaming-siblings-together rule applies if anyone changes the names.

---

drafted by Claude on behalf of Daniel Stephenson

[dmccoystephenson]

Self-review rubric:

• Scope: PASS — 2 files, 44 lines; each change maps to one of #24's 4 AC items or the CLAUDE.md PR-27 Implementations rule.
• Tests-new: PASS — the new self-audit rubric is exercised by this very cycle (I'll run it after merge); the new label taxonomy is exercised by the Step 6 loop syntax (verified shell expansion works with IFS=:).
• Tests-fix: N/A — not a bug fix.
• Sibling structure: PASS — Phase 9 rubric block matches the Phase 4 rubric structure (bold-item-name + colon + question; PASS/FAIL/no-signal scoring); new Step 6 bash block matches the existing Step 6 bash-block style (fenced bash, no surrounding prose interruption).
• Sibling renames: PASS — the label-taxonomy names (template-rule, repo-specific, etc.) appear verbatim in Phase 9 AND Step 6; siblings renamed/added together.
• Docs: PASS — Phase 7 sources (create-dev-loop.md, README.md) reviewed; README is high-level and doesn't enumerate Phase 9 internals (correctly out of scope). RESEARCH.md §1 and §5 get Implementations entries per PR-27 rule.
• Issue resolution: PASS — all 4 AC items met: (1) structured checklist replaces bullet list; (2) substitution table unchanged; (3) gap-issue command takes --label; (4) labels created in Step 6 before first cycle runs.
• Validation: PASS — placeholder count unchanged at 22; Step 6 bash block uses standard \``` plain fence (Step 6 is OUTSIDE the outer template, so no escape needed); Phase 9 fenced block uses leading-backslash escape (Phase 9 is INSIDE the outer template).
• {{EXTERNAL_SIGNAL_LABEL}}: N/A — no CI; manual validation against diff served as anchor.
• Placeholder-consistency: N/A — no template placeholder added.
• Escaping: PASS — confirmed by reading the diff: Step 6 label-creation block uses plain triple-backticks (it's outside the embedded template), Phase 9 issue-create block uses \``bash` with leading backslash (inside template).

Adversarial probe — 'Does the new rubric's External-signal-quality item make sense for repos where the external anchor is just "run the test command locally"?' Considered: yes, the item asks 'did it catch the kinds of issues it's supposed to catch this cycle' — for any anchor (CI, manual validation, local tests), the agent can score it. No exclusion needed.