finding(process): a closing keyword in a merged PR body does not close the card in this repo — 29 of 29 measured closes were performed by hand, 0 by a commit, and the same is true of the bare `#N` form

[claude]

Filed by the domain:spec @ objectui execution seat, 2026-09-14T08:48Z, from the post-merge bookkeeping of PR objectui#9470. ⛔ Bare finding — no domain:*, no priority, no type, no assignee.

The claim

A closing keyword in a merged PR body does not close the referenced card in this repository. Every card close in the measured sample was performed by hand, minutes after the merge, by a person or a bot.

Measured

Population: 300 closed pull requests fetched over 3 pages of /pulls?state=closed&sort=updated, of which the merged ones were scanned for (?:Fixes|Closes|Resolves|Fixed|Closed|Resolved)\s+<ref>. Reading taken 2026-09-14T08:46–08:48Z.

The instrument is the issue timeline's closed event, not the issue's state. ⭐ That distinction is the whole measurement: a close performed by a landing commit records a commit_id on that event; a close performed by a person or a bot records an actor and leaves commit_id null. state: closed alone cannot tell the two apart, which is why counting closed cards proves nothing here.

spelling	pairs measured	closed BY COMMIT	closed BY ACTOR
Fixes objectui#N	3 (all of them)	0	3
Fixes #N (bare)	25 (sampled from 300)	0	25
Fixes objectstack-ai/objectui#N	1 (all of them)	0	1

29 of 29 closed by an actor. Zero by a commit. The closers are spread across five accounts — os-tesla, baozhoutao, os-zhuang, os-warren, claude[bot] — so this is not one seat's habit; it is the repository's actual behaviour, already being absorbed by everyone who lands work here.

The specific case that started this, with its timings:

PR #9470 merged 2026-09-14T08:43:33Z, body carries  "Fixes objectui#8653."
issue #8653 closed 2026-09-14T08:45:25Z  by claude[bot]  commit_id=None     <- by hand, ~2 min later
PR #9009 merged 2026-09-10T18:52:15Z -> #8623 closed 18:54:52Z by actor, commit_id=None
PR #9007 merged 2026-09-10T18:51:55Z -> #8554 closed 18:55:41Z by actor, commit_id=None

⭐ Why this costs more than one manual close

It explains a recurring residue rather than being a new one. This seat had recorded, six times over, that pm:dispatched is "never auto-stripped by a merged Fixes". The label survives because nothing closed the card — the close and the pm:* strip are the same missing step, not two.

⇒ operationally, for every seat: a merged closing keyword is not bookkeeping. The close, the state_reason: completed, and the pm:* strip are all owed by hand, and a seat that trusts the keyword leaves a closed-in-fact card sitting open + pm:dispatched on the board, where the next round reads it as live work.

⚠️ A hypothesis this card had and DISCARDED — recorded so the next reader does not re-run it

The first reading was n=1: Fixes objectui#8653 did not close #8653, and objectui#N is not one of the two reference forms GitHub documents (#N and owner/repo#N). The hypothesis wrote itself — this repo's full-name PM convention silently disables closing keywords.

⛔ It is false, and two separate errors produced it:

1. The subject sample was contaminated by the measurer. Widening to 390 merged PRs found 3 objectui#N pairs, all state: closed — which looked like a refutation, except that one of the three was decision(plugin-list): title is read undeclared off ListViewSchema — and rowActionDefs is read at a site with NO docblock and NO pin, unlike the grid's ruled twin (6 of 7, objectui#8327) #8653, which this seat had hand-closed two minutes earlier. A census that includes the measurer's own intervention cannot distinguish the mechanism from the measurer.
2. state was the wrong instrument. Only the timeline's commit_id separates an auto-close from a manual one. Switching instruments turned "all three closed, hypothesis refuted" into "all three closed by hand, and so is every control" — a different and much larger finding.

⭐ The control is what settled it: the bare #N form, which GitHub certainly does resolve, behaves identically — 25 of 25 by actor. So the reference spelling is not the variable, and a card filed on the first hypothesis would have sent someone to rewrite every PR template for nothing.

⚠️ NOT measured

• The cause. No repository setting, workflow, or merge-queue behaviour was inspected. The merge commit for fix(plugin-list): retire the legacy title alias and pin the rowActionDefs exemption #9470 demonstrably does carry Fixes objectui#8653. in its body on main, so the keyword reaches the default branch and the close still does not happen. Why is unmeasured, and ⛔ this card asserts no cause.
• Whether a non-queue merge in this repo behaves differently — the sample may contain no such merge, and none was isolated.
• 25 of 300 bare-form pairs were sampled, not all 300; the other two spellings were measured exhaustively (3 and 1).
• An earlier pass reported 16 bare-form pairs as state: None. Those are unreadable API responses, ⛔ not counter-examples, and they are excluded rather than counted either way.
• Whether the sibling repositories behave the same. This reading is objectui-only.

Dedup

Two phrasings of this subject returned 0. ⚠️ The control is same-repo and same-instrument but not same-subject: queries about other CI/tooling subjects in this repo return results (a coverage-gate query returned 12 including objectui#9271; a renderer-cast query returned 4), so the instrument is answering in this repo. On that basis the zero is treated as a reading, ⛔ with the weaker control stated rather than implied.

---

Generated by Claude Code