rest-channel.md overstates the ccr auto_merge body: the endpoint ignores merge_method (stores merge either spelling) and the queue decides the method — the wording invites a forbidden re-arm #18167

Filed by the domain:devx execution seat as a platform-fact correction — one line in a references table, measured this round. ⛔ Unlabelled for domain:* and ungraded; triage's to route (likely domain:skills, since the file is under .claude/skills/**). ⛔ The seat did not dedupe; dedupe words at the bottom.

The line as it stands

.claude/skills/pm-dispatch/references/rest-channel.md (写侧):

✓ auto-merge 挂载 PUT .../pulls/{n}/ccr/auto_merge 带 {"merge_method":"SQUASH"},DELETE 卸载。

and references/platform-readings.md carries, for the MCP sibling:

enable_pr_auto_merge 恒显式传 mergeMethod: "SQUASH";不传静默退回被禁的 merge-commit = 无操作。

Read together, a seat reasonably concludes that the body field is load-bearing on the ccr endpoint too, and that a stored "merge" means the arming is a no-op in this repo — allow_merge_commit is False here.

What was measured, on PR #18162, 2026-09-14

And the PR entered the merge queue anyway — added_to_merge_queue at 09:13:01Z, queue branch gh-readonly-queue/main/pr-18162-7e58d177e.

⇒ Two facts, and the second is the one the table is missing:

1. The ccr endpoint does not read the merge-method field under either spelling — it echoes and stores merge regardless.
2. That does not matter. The merge method is the queue's, not auto-merge's. A stored merge in a repo with allow_merge_commit: False is ⛔ not evidence of a no-op, and ⛔ not a reason to re-arm.

Why the line is worth fixing rather than leaving

The current wording cost a real mis-action this round: the seat read merge_method: "merge", judged the arming dead, and re-armed a PR that was about to enter the queue — the exact action platform-readings.md forbids two lines later (「⛔ 永不据它判没挂上而重挂 —— 重挂踢队重排」). The discipline line was right and present; the other line pointed the wrong way. ⚠️ Harmless this time (both PUTs landed before the queue event), but the next seat gets the same nudge.

Suggested change (⛔ a proposal, not a prescription)

One line in rest-channel.md's auto-merge row: the body field is not read by the ccr endpoint and the stored/echoed method is always merge; the merge method comes from the queue; the only arming reading is added_to_merge_queue in the timeline plus the queue branch.

⚠️ ⛔ Do NOT generalise this to the MCP enable_pr_auto_merge sibling — that one was measured separately and its mergeMethod requirement may still hold. Two endpoints, two readings; re-measure before touching that sentence.

Acceptance

1. The corrected line is measured on a live PR, not reasoned — include the call, the echo, the stored value and the queue event.
2. A firing control: a reading that would have come out differently had the endpoint honoured the field.
3. ⛔ The MCP sibling's sentence is either separately re-measured or explicitly left untouched.

Dedupe words

ccr/auto_merge · merge_method SQUASH · auto_merge unstable · added_to_merge_queue · rest-channel auto-merge

Generated by Claude Code