[dmccoystephenson]

Summary

Bundles the 14 consolidated template-rule issues into create-dev-loop.md in one wave, the documented promotion pattern (CONVENTIONS §"Promoting a carry-over"; the May-25 wave bundled 8 the same way). Each fix is applied once to the template; the per-instance retrofit wave follows as one small PR per lagging skill.

Issue	Phase	Change
#36	header	version/generated-at comments moved below the H1 so commands-format installs show the description, not the comment
#39	8	review-ready hand-off = valid terminal state; never --admin; explicit human auth satisfies a do-not-auto-merge hold
#40	4 / 1	self-review via plain PR comment (reviews API is blocked); backfill on carried-over PRs; inline comments stay in-hunk
#41	1	runtime working-tree resolution (env var → path → clean clone → fresh clone); new REPO_ENV_VAR placeholder
#42	5 / 1	autonomous multi-cycle batch mode + stop condition; externally-directed skip reason recorded once
#43	1	classify harness-blocked ops (CLAUDE.md edits, external submodule reg) at triage
#44	4	UNVERIFIED handling + scope-gated merge when the anchor can't run
#45	1 / 8	adopt a concurrent session's PR; harmless worktree --delete-branch failure
#46	4	green CI ≠ verification when CI scope excludes the changed files
#47	4	empirical Tests-fix (stash-and-run FAIL→PASS), not judged
#48	3	scope formatting to changed files
#49	3	scope ceiling counts non-test LOC; allow dependency-coupled batches
#50	6	don't hardcode the Co-Authored-By model; defer to the harness
#51	3	stage by name, never git add -A

Research grounding (per CLAUDE.md)

• Phase 4: make Tests-fix empirical (stash-and-run), not judged #47 (empirical Tests-fix) and Phase 4 self-review: use a plain PR comment, backfill on carried-over PRs, fold out-of-hunk notes into the body #40 (single bounded self-review) operationalize RESEARCH.md §1/§5 — intrinsic self-critique without an external signal is unreliable; grounding the rubric on an executed check (stash-and-run) rather than judgment is the corrective.
• Phase 3: scope ceiling should count non-test LOC and allow dependency-coupled batches #49 (scope ceiling) refines the RESEARCH.md §2 merge-rate-vs-PR-size finding (counts non-test LOC; the hard stop is unchanged).
• External anchor unavailable in sandbox: define UNVERIFIED handling and scope-gated merge #44/Phase 4/8: green CI is not verification when CI scope excludes the changed files #46 sharpen the "external anchor" definition from §1.
• The remaining items (Template-version HTML comment at file top breaks skill-list description for commands-format installs #36, Phase 8: codify review-ready hand-off as a valid terminal state; never --admin; explicit human auth satisfies a do-not-auto-merge hold #39, Phase 1: resolve the working tree robustly instead of hardcoding an absolute path #41, Loop control: define autonomous multi-cycle batch mode and externally-directed-cycle skip-reason behavior #42, Triage: recognize harness-blocked operations (CLAUDE.md edits, external submodule registration) up front #43, Concurrent-session / worktree handling: adopting another session's PR; harmless --delete-branch local failure #45, Phase 3: scope formatting to changed files; don't let a tree-wide formatter pollute a focused PR #48, Phase 6: don't hardcode the Co-Authored-By model name; defer to the harness git rule #50, Git hygiene: stage by name, never git add -A (leaks .claude/ harness state) #51) are operational/harness-hygiene rules surfaced by per-loop self-audits with no specific RESEARCH.md finding behind them — recorded honestly rather than back-justified. No RESEARCH.md entries added.

Validation

No automated suite (per CLAUDE.md). Structural validation on create-dev-loop.md: outer code fences balanced (13 pairs), escaped inner fences balanced (14 pairs), embedded template closes cleanly, new REPO_ENV_VAR placeholder has a Step-4 table row, no leftover formal-review API call or hardcoded model name. README + CLAUDE.md updated 1:1 for the #36 header move.

Closes #36, #39, #40, #41, #42, #43, #44, #45, #46, #47, #48, #49, #50, #51

🤖 Generated with Claude Code

---

drafted by Claude on behalf of Daniel Stephenson

[dmccoystephenson]

Self-review rubric (anchor = structural validation; this repo has no CI suite, per CLAUDE.md):

• Anchor: PASS — fences balanced (13 outer pairs / 14 escaped inner pairs), embedded template closes at the No issues… line, no unsubstituted breakage introduced.
• Scope: PASS — every hunk maps to a Closes #N; the only non-create-dev-loop.md edits are the README + CLAUDE.md lines describing the Template-version HTML comment at file top breaks skill-list description for commands-format installs #36 header position. No unrelated rewrites of untouched phases.
• Placeholder contract: PASS — the one new token {{REPO_ENV_VAR}} has a Step-4 table row; ${{REPO_ENV_VAR}} renders to ${SLUG_DIR} after substitution.
• Phase numbering: PASS — Phases 1–10 intact; no renumbering.
• Issue resolution: PASS — all 14 issues' named surface areas are actually changed (verified each against its issue body): Template-version HTML comment at file top breaks skill-list description for commands-format installs #36 header order, Phase 8: codify review-ready hand-off as a valid terminal state; never --admin; explicit human auth satisfies a do-not-auto-merge hold #39 Phase-8 hand-off, Phase 4 self-review: use a plain PR comment, backfill on carried-over PRs, fold out-of-hunk notes into the body #40 plain-comment self-review (old reviews-API block removed), Phase 1: resolve the working tree robustly instead of hardcoding an absolute path #41 resolver + table row, Loop control: define autonomous multi-cycle batch mode and externally-directed-cycle skip-reason behavior #42 Phase-5 batch mode, Triage: recognize harness-blocked operations (CLAUDE.md edits, external submodule registration) up front #43 triage classification, External anchor unavailable in sandbox: define UNVERIFIED handling and scope-gated merge #44 UNVERIFIED edge case, Concurrent-session / worktree handling: adopting another session's PR; harmless --delete-branch local failure #45 concurrent-session adoption, Phase 4/8: green CI is not verification when CI scope excludes the changed files #46 CI-scope clause, Phase 4: make Tests-fix empirical (stash-and-run), not judged #47 stash-and-run, Phase 3: scope formatting to changed files; don't let a tree-wide formatter pollute a focused PR #48 scoped formatting, Phase 3: scope ceiling should count non-test LOC and allow dependency-coupled batches #49 non-test LOC ceiling, Phase 6: don't hardcode the Co-Authored-By model name; defer to the harness git rule #50 deferred trailer (hardcoded name removed), Git hygiene: stage by name, never git add -A (leaks .claude/ harness state) #51 stage-by-name.
• Docs: PASS — README Step-3 wording and CLAUDE.md version-comment note both updated for Template-version HTML comment at file top breaks skill-list description for commands-format installs #36; substitution table updated for Phase 1: resolve the working tree robustly instead of hardcoding an absolute path #41.
• Dogfooding: PASS — this PR was staged by name (Git hygiene: stage by name, never git add -A (leaks .claude/ harness state) #51) and this self-review is a plain comment (Phase 4 self-review: use a plain PR comment, backfill on carried-over PRs, fold out-of-hunk notes into the body #40).

Out-of-diff note (folded into body, not inline): the per-instance retrofit wave to the new baseline is the follow-up — one small PR per lagging skill, tracked separately.

One honest soft spot: #48's "scope formatting to changed files" is phrased against black/autoflake examples carried from the originating issue; per-instance retrofits should re-tune those to each repo's real formatter (or mark n/a where there is none).

---

drafted by Claude on behalf of Daniel Stephenson