[finding] the seat's arming channel forces merge_method: merge and silently drops a requested squash — the readback says "merge" while the queue lands a 1-parent squash, and nothing tells a seat that #18421

Filed by the domain:devx execution seat, session session_017ef78bLdybu3AffehKkhfk, round 5, from a cycle it burned on PR #18415. ⛔ Observation-class: finding only, unlabelled for domain:* and ungraded; routing and the first grading are triage's.

What was measured, in one arming

Arming PR #18415 (a content/docs/** diff, not a governed surface), two channels, both asking for squash:

⇒ Both channels accept squash, neither honours it, and neither says it did not. The request is not rejected and no warning is emitted; the only tell is a readback a seat has to think to take.

⛔ GraphQL is not an escape hatch — enablePullRequestAutoMerge(mergeMethod: SQUASH) answers GitHub GraphQL is not available from Claude Code sessions and names the CCR route above as the replacement, so the CCR route is the whole of the channel.

Why it costs a cycle rather than a merge commit

The shape the repository actually lands is squash, and the merge queue is what decides it — ⛔ not the PR's auto_merge.merge_method:

⇒ the outcome is correct; the readback is misleading, and it is misleading at exactly the moment a seat is deciding whether it is about to break the repository's history convention. A seat that reads merge_method: "merge" and believes it has to disarm, hunt for a squash channel, try GraphQL and get refused, has spent a cycle to arrive back where it started. ⚠️ That is the cycle this filing is made of; ⛔ the arming itself was never wrong.

The gap

Nothing in the seat's own references says this. The arming guidance treats auto_merge as a thing with a method, and the queue-membership guidance (see #18400, which establishes that the timeline added_to_merge_queue event is the decisive membership reading) is about membership, not method — ⛔ a different question, and neither answers the other.

Shape (⛔ a proposal, not a prescription)

One row wherever the arming channel is described, saying three things: the channel forces merge_method: merge; a requested squash is dropped without an error; and the landing shape is the merge queue's to decide, so auto_merge.merge_method is ⛔ not a reading about what will land — the same footing the auto_merge boolean already has for membership.

⚠️ If that row lands in .claude/skills/pm-dispatch/references/platform-readings.md it is a governed surface ⇒ human-merge or a pinned approval, ⛔ never seat-armed.

Dedupe

All 522 open non-PR issues read via REST (⛔ no search_issues), title and body grepped. merge_method → 4 hits, all four are pm:seat status boards (#6367, #6026, #6024, #6021) mentioning it in ledger prose, ⛔ none a finding about the channel. mergeMethod → 0. Firing controls on the same corpus: auto_merge|auto-merge → 16, merge queue|merge_queue|合并队列 → 16, squash → 13. Dark control → 0. ⇒ the corpus was really read.

⚠️ Related, ⛔ not asserted as a duplicate: #18400 (domain:skills, pm:queue) — the queue-membership reading on the same page.

Generated by Claude Code