ci: a PR body that says "do not merge" holds nothing (#5154 merged by app/cursor)

[kyle-sexton]

Decision packet from the 2026-09-29 audit of the Cursor agent run. Nothing is implemented; the question is the owner's.

What happened

PR #5154's body said "Superseded by #5059 ... Do not merge this draft." app/cursor merged it on 2026-09-28 at 10:42 UTC and it shipped playbooks 0.13.17 (units of #4027). The PR carried no do-not-merge label, and the label is the only thing ci-status reads.

Two audit groups asked for a guard. Playbooks asked ci to make a PR body containing "Do not merge" fail ci-status or auto-apply the label. Claude-config asked source-control to make such a body block the unattended merge lane. The lane side is independent of this question: app/cursor merged #5154, not the babysit lane.

Why this is the owner's decision

1. docs/conventions/loop-lane/README.md:169-175 says the do-not-merge label is the only cross-lane hold and that prose on a PR is advisory. Each option below either adds a second signal or changes who applies the first.

2. The pr-contract step that reads the label lives in the ci-status composite of melodic-software/ci-workflows (pinned at 4610c31e, v0.27.1, in .github/workflows/ci.yml), not in this repo. Changing what it reads is a change there plus a repin here.

3. A workflow in this repo that applies the label with GITHUB_TOKEN does not re-run ci-status: events triggered by that token create no new workflow runs, except workflow_dispatch and repository_dispatch (GitHub docs, "Triggering a workflow"; cited in the .github/workflows/dependabot-plugin-release.yml comment on PR ci: re-pin gitleaks composite, extend the Windows lane, wire unwired suites #5316, and the same question as ci: every Dependabot PR touching plugins/** fails the changelog-parity bump gate, because Dependabot cannot bump a manifest or write a changelog entry #4144). A token that does re-trigger is a credential only the owner can create.

4. The phrase is common in legitimate PR bodies. gh pr list --state all --search '"do not merge" in:body' on 2026-09-29, filtered to a literal case-insensitive match:

   • Anywhere in the body: 63 PRs (22 merged, 39 closed, 2 open).
   • At the start of a line: 22 PRs (8 merged, 14 closed). Seven of the eight merged (feat(naming): research-backed criteria, structured briefs, rejection protocol (0.2.0) #341, feat(re-anchor): add use-your-skills and reuse-or-replace, extend two correctors #769, feat(claude-memory): add stateless skill #775, feat(session-flow): add running-retro skill #777, chore(claude-lanes): re-pin the review lanes to ci-workflows v0.10.2 #1990, chore(claude-lanes): re-pin the review lanes to ci-workflows v0.11.0 #2079, ci: parallelise lint, the PR test path and the Windows lane #4197) were holds that were lifted by merging: the merged body still carries the phrase. chore: apply changelog rows for CLI surfaces and managed connectors (#4027) #5154 is the eighth.

   A body gate would have stopped those seven merges until someone edited the body to release the hold.

Options

• (A) Keep the convention. The label stays the only hold and a lane or person that decides not to merge applies it. No change.
• (B) ci-status fails while a body line starts with "Do not merge" (case-insensitive) and passes once the line is edited out. Change in ci-workflows pr-contract, a repin here, and line 169 of the loop-lane README updated to name the body line. Releasing a hold then means editing the body.
• (C) A workflow here applies the label when the body says it and removes it when the phrase goes. Needs an App token or PAT so the label event re-runs ci-status (operator credential, as in ci: every Dependabot PR touching plugins/** fails the changelog-parity bump gate, because Dependabot cannot bump a manifest or write a changelog entry #4144 option B). With GITHUB_TOKEN the label lands and ci-status stays green.
• (D) Instruct the agents that open PRs (root AGENTS.md, and the Cursor agent's own instructions): a PR that is superseded or must not merge gets the label, not prose. No infrastructure; it depends on the agent following it.

Recommendation (a recommendation, not a decision)

D now; revisit B if a second PR merges after saying it must not.

Basis: the counts in item 4 (a body gate changes how seven past holds would have been released) and docs/conventions/loop-lane/README.md:169-175 (the label is already the hold). D is reversible and touches no shared infrastructure; B and C need a change in another repo or a credential.

What would change it: a second such merge, or the owner wanting prose enforced server-side (then B, with the pattern narrowed to a marker the owner picks).

Question for the owner

Which option stands: A, B, C or D? If B, which pattern: any line starting with "Do not merge", or a marker such as an HTML comment that the agents are told to write?

Unblocks

The playbooks and claude-config merge-guard requests recorded on PR #5316, and whether the babysit skill body needs a body-text rule of its own.

Related

Refs #4027 (the change #5154 carried), #4144 (token that re-triggers CI), #4670 (the contract-only ci-status stale-read defect: a change to what pr-contract reads should land with that upstream fix, in the same repo and repin).

[kyle-sexton]

Owner decision (2026-09-29): Option D. Tell agents in AGENTS.md to apply the do-not-merge label rather than rely on prose in the PR body; no change to ci-status and no new workflow. Owner note: accept recommendation, D now; revisit B (ci-status fails on a body line starting "Do not merge") if a second PR merges after saying it must not.
Next: an agent adds the label rule to root AGENTS.md and to the Cursor agent instructions, then closes this issue when merged.