[laurigates]

git-merge-hazards described the serial-merge late-commit guard more loosely
than scripts/merge-chain.sh (#3251) enforces it.

The script exempts a commit newer than the review timestamp only if it has
two or more parents and GitHub committed it
(committer.email == noreply@github.com, which is what an update-branch
merge looks like). It never looks at the subject line.

The docs said "refuse any non-merge commit". The hand-run snippet exempted
any commit whose subject starts with Merge. Under that snippet, a conflict
agent's own Merge branch main into … commit would pass, even though it
carries a conflict resolution nobody has reviewed.

Changes

• SKILL.md §6 states the exemption as the script implements it, and why
  your own merge of main is refused. The file stays under the 10,000-char
  ceiling that test-no-force-stack.sh enforces.
• references/serial-merge-chain.md prose matches the script. The guard
  snippet now uses the script's own commits query and the same parent-count
  plus committer test. It stops when the commit list is cut off, and it
  merges pinned to the head it read. For sessions without GraphQL, one line
  gives the REST equivalent.
• references/serial-merge-chain.md has a new "Known limits" section with
  the two feat(git-plugin): add merge-chain.sh merge shepherd to git-merge-hazards #3251 verifier points. Committer email can be forged; checking
  signature { wasSignedByGitHub } would harden that. A merge made in
  GitHub's web conflict editor also passes the exemption, although it
  contains a resolution. merge-chain.sh itself is unchanged.
• references/delta-apply-stack.md and references/no-force-push-stack.md
  no longer claim the guard allows a bot's merge commit, and no longer use
  "non-merge commit" wording.
• Regression check: a new §10 in scripts/tests/test-merge-chain.sh.
  • It greps for the stale wording.
  • It extracts the snippet, runs bash -n on it, and executes it against a
    stub gh. A GitHub-made merge passes; an agent's merge and a
    single-parent Merge … commit are refused.
• A regression-ledger row.

Tests

• test-merge-chain.sh: 116 passed, 0 failed.
• test-no-force-stack.sh: 356 passed, 0 failed.
• test-push-refspec.sh: 18 passed, 0 failed.
• pre-commit passes, including check-docs-index (STATUS=OK).

Fixes #3254

---

Generated by Claude Code

[github-actions]

Plugin Compliance Review

Plugin	plugin.json	Frontmatter	Body	Marketplace	Release Config	Bash Patterns	Descriptions	When-to-Use	Size	Overall
git-plugin	✅	✅	✅	✅	✅	✅	⚠️	⚠️	⚠️	⚠️

Recommendations

• ⚠️ git-plugin/gh-cli-agentic: description is 210 chars — over the 200-char target band (see .claude/rules/skill-quality.md)
• ⚠️ git-plugin/git-push: description is 237 chars — over the 200-char target band (see .claude/rules/skill-quality.md)
• ⚠️ git-plugin/github-issue-autodetect: description is 230 chars — over the 200-char target band (see .claude/rules/skill-quality.md)
• ⚠️ git-plugin/github-pr-title: description is 221 chars — over the 200-char target band (see .claude/rules/skill-quality.md)
• ⚠️ git-plugin/release-please-protection: description is 251 chars — over the 200-char target band (see .claude/rules/skill-quality.md)
• ⚠️ git-plugin/git-upstream-fix-check: SKILL.md has no '## When to Use This Skill' section — recommended where sibling skills compete for the same intent, otherwise the description carries the trigger (see .claude/rules/context-engineering.md)
• ⚠️ git-plugin/git-repo-delete-check: SKILL.md has no '## When to Use This Skill' section — recommended where sibling skills compete for the same intent, otherwise the description carries the trigger (see .claude/rules/context-engineering.md)
• ⚠️ git-plugin/git-issue-scoping: SKILL.md has no '## When to Use This Skill' section — recommended where sibling skills compete for the same intent, otherwise the description carries the trigger (see .claude/rules/context-engineering.md)
• ⚠️ git-plugin/git-pr: SKILL.md is 10549 chars (~2637 tokens, >10000) — consider extracting to REFERENCE.md or scripts/ (ceiling: 26000 chars / ~6500 tokens)
• ⚠️ git-plugin/git-commit-push-pr: SKILL.md is 10627 chars (~2656 tokens, >10000) — consider extracting to REFERENCE.md or scripts/ (ceiling: 26000 chars / ~6500 tokens)
• ⚠️ git-plugin/git-pr-feedback: SKILL.md is 18746 chars (~4686 tokens, >10000) — consider extracting to REFERENCE.md or scripts/ (ceiling: 26000 chars / ~6500 tokens)
• ⚠️ git-plugin/git-commit-workflow: SKILL.md is 11052 chars (~2763 tokens, >10000) — consider extracting to REFERENCE.md or scripts/ (ceiling: 26000 chars / ~6500 tokens)
• ⚠️ git-plugin/git-issue: SKILL.md is 15561 chars (~3890 tokens, >10000) — consider extracting to REFERENCE.md or scripts/ (ceiling: 26000 chars / ~6500 tokens)
• ⚠️ git-plugin/git-upstream-pr-diverged: SKILL.md is 10474 chars (~2618 tokens, >10000) — consider extracting to REFERENCE.md or scripts/ (ceiling: 26000 chars / ~6500 tokens)
• ⚠️ git-plugin/git-conflicts: SKILL.md is 11315 chars (~2828 tokens, >10000) — consider extracting to REFERENCE.md or scripts/ (ceiling: 26000 chars / ~6500 tokens)
• ⚠️ git-plugin/git-issue-scoping: SKILL.md is 10046 chars (~2511 tokens, >10000) — consider extracting to REFERENCE.md or scripts/ (ceiling: 26000 chars / ~6500 tokens)
• ⚠️ git-plugin/git-local-hazards: SKILL.md is 10013 chars (~2503 tokens, >10000) — consider extracting to REFERENCE.md or scripts/ (ceiling: 26000 chars / ~6500 tokens)
• ⚠️ git-plugin/git-triage: SKILL.md is 10007 chars (~2501 tokens, >10000) — consider extracting to REFERENCE.md or scripts/ (ceiling: 26000 chars / ~6500 tokens)

[claude]

Skill Quality Review

• git-plugin/skills/git-merge-hazards/SKILL.md: passes all checks. It has a When to Use decision table and an Agentic Optimizations table. The description is phrased around user intents (merging a PR or a stacked chain, auditing whether a branch landed, merging over red CI). model and effort are unset, so both are inherited, which is appropriate. There is no REFERENCE.md; the supporting files are under references/ and scripts/, which are not part of this checklist.

Approved for skill quality.