[dmccoystephenson]

Summary

Adds two structural gates to the generated dev-loop template, grounded in RESEARCH.md §2 (autonomous-agent PRs fail at higher rates when large or touching sensitive paths):

1. Phase 3 scope ceiling. Before pushing, check the cumulative diff. Soft ceiling: ~400 net LOC, ~10 files. Hard stop: ~800 LOC, ~20 files. If exceeded, drop an issue or split the PR.

2. Phase 8 do-not-auto-merge check. Before invoking gh pr merge, compare modified paths against a list. Universal entries cover .github/workflows/*, security/, and >50-line deletions. New {{DO_NOT_AUTO_MERGE}} placeholder lets each generated skill add repo-specific paths (e.g. plugin.yml, pom.xml).

The bug-fix-requires-test gate mentioned in #17's proposal section is deferred to #18 (which has its own AC for a Phase 8 regression-test gate); avoiding overlap.

Closes #17

Test plan

• Phase 3 has scope-ceiling check with git diff --stat command and explicit "what to do if exceeded" language (drop an issue or split)
• Phase 8 has do-not-auto-merge path check with git diff --name-only command, before gh pr merge
• New {{DO_NOT_AUTO_MERGE}} placeholder appears in template body AND in Step 4 substitution table (verified with grep -oP '\{\{[A-Z_]+\}\}')
• New {{#if DO_NOT_AUTO_MERGE}} conditional wraps repo-specific section so generated skills with no repo-specific entries don't render an empty block
• This PR's own diff respects the new ceiling: 27 net LOC across 3 files (well under 400/10 soft, 800/20 hard)
• README.md "Generated skill structure" gets a bullet for the new repo-specific encoding ({{DO_NOT_AUTO_MERGE}}), keeping it in sync with the template
• Per the rule shipped in PR Promote 'verify each Closes #N' to template Phase 4 #26, ran gh issue view 17 and confirmed title/body match
• Per the rule shipped in PR Track implementation outcomes back into RESEARCH.md #27, added an Implementations entry under RESEARCH.md §2

Fixture-validation note

#17's test plan asks for "a fixture issue that intentionally would produce a too-large diff" to confirm the gate stops it. Out of scope for this PR (would require a fixture repo + tooling that doesn't exist). The gate is currently authority-based, not enforced — a future improvement would be a pre-push hook that fails the push if git diff --stat exceeds the threshold. Filing as a follow-up after merge.

---

drafted by Claude on behalf of Daniel Stephenson

[dmccoystephenson]

Self-review rubric:

• Scope: PASS — 3 files modified (create-dev-loop.md, README.md, RESEARCH.md); every change traces to #17 AC, the CLAUDE.md PR-27 rule (Implementations entry), or doc-sync (README's Generated skill structure list).
• Tests-new: PASS — the new {{DO_NOT_AUTO_MERGE}} placeholder has a substitution-table row with format guidance; the scope ceiling is exercised by this PR's own git diff --stat origin/main (27 LOC, 3 files — under the soft ceiling).
• Tests-fix: N/A — not a bug fix.
• Sibling structure: PASS — new Phase 3 'Scope ceiling' paragraph matches the 'Bold label. explanation + fenced command' pattern of its siblings; new Phase 8 do-not-auto-merge block similarly uses the bold-label + fenced-command pattern.
• Sibling renames: N/A — no identifiers renamed.
• Docs: PASS — README updated to add 'Do-not-auto-merge paths' to the Generated-skill-structure list (it's a new repo-specific encoding); RESEARCH.md §2 gets an Implementations entry for PR #30 per the CLAUDE.md rule.
• Issue resolution: PASS-with-deferral — #17 ACs 1-3 met (ceiling check, do-not-auto-merge check, new placeholder in table); AC 4 (fixture-issue test) explicitly deferred in PR body, since no fixture tooling exists.
• Validation: PASS — placeholder grep confirms {{DO_NOT_AUTO_MERGE}} exists in template body, has matching {{#if DO_NOT_AUTO_MERGE}} conditional wrap, and has substitution-table row; new fenced blocks (lines 194-196, 334-336) are properly escaped with leading-backslash triple-backticks.
• {{EXTERNAL_SIGNAL_LABEL}}: N/A — this repo has no CI; manual validation against the diff text is the anchor (which I just did).
• Placeholder-consistency: PASS — {{DO_NOT_AUTO_MERGE}} appears in template body + conditional + substitution-table row; no orphans.
• Escaping: PASS — both new fenced code blocks (one in Phase 3 line 194, one in Phase 8 line 334) use \``` (leading backslash on each triple-backtick), confirmed by reading the surrounding lines.

Adversarial probe — 'Is the 400 LOC ceiling arbitrary?' Considered: yes it is — RESEARCH.md §2 cites 'agent PRs that are larger are unmerged more often' but doesn't give a specific threshold. The 400/800 numbers come from §6's reference to the SmartBear/Microsoft 200-400 LOC industry sweet spot, which is medium-confidence. Documented in the PR body that this is a defensible heuristic, not a measured threshold. Followup-issue tracking the empirical calibration could be filed but the current numbers are reasonable starting points.