Update ruling D (#18375): the maintainer lifts the ban — `skip-changeset` may go on a PR that only corrects a pending changeset; four carriers in the repo still say the opposite

[os-support-ai]

③ — a maintainer-directed task. The maintainer's words, verbatim, with provenance:

• Words: 「立卡更新我的决裁,这两个你直接打标签跟进合并。」
• Who / where: the maintainer, in the chat of domain:spec execution seat 1 (session_013RDBh5DqXd2xnLwvHLgLFr, seat post [PM seat] domain:spec — seat 1 · 🟢 session_013RDBh5DqXd2xnLwvHLgLFr · 未达档(CONTRACT_REVIEW_TIER=fable)⇒ 条款②走隔离达档子代理,⛔ 不席内自审 · 本班 16 落地(最新 #19801 → 9df39341e6)· 0 在飞 · ⏳ 等人 1:#19769 受管面人合 · 🗳️ 决策箱 1:#19461 · 队列无可派(北极星第 3 条 + 围栏 + 他席已认领) #6017), on 2026-09-23 UTC.
• Replying to: the seat's report that PRs docs(changeset): correct two false mechanism claims in a pending connector-ledger note #19746 and docs(changeset): correct four statements in the pending enableOnInstall one-authority note #19747 were red only on Check Changeset (route 0) under ruling D. Before the words above, the maintainer asked whether that ruling was unreasonable and felt like a waste of their time. The seat offered three routes: 甲 keep it; 乙 let an at-tier review PASS stand in for the maintainer's confirmation on wording-only corrections; 丙 restore the label, which the seat advised against. The maintainer directed the label on both PRs and asked for the ruling to be updated.

Filed by that seat. ⛔ Ungraded and unrouted: routing is triage's, and this card spans a governed surface and a non-governed one, so it may be split.

The ruling being updated

Batch #158 item 1 on #18375 (5729188290, maintainer 「同意」 2026-09-18): ② B — a PR that edits an existing changeset is confirmed in writing on the PR, and the gate stays advisory red; ⛔ skip-changeset is never applied to it. ① D — no workflow mechanics change; only the texts say so.

The update, as this seat reads the words (⚠️ the maintainer corrects this if it is wrong)

A PR that only corrects a pending changeset (every .changeset row M/D, none A) may take skip-changeset. No per-PR maintainer confirmation is required. Nothing in the words asks for a workflow mechanics change, so this round changes texts only, the same scope ① D had.

Applied already, before the texts change, on the maintainer's direct instruction: #19746 and #19747 carry skip-changeset and are in the merge queue. Provenance is on each PR (5790571940, 5790573062).

The carriers that still say the opposite (re-read on origin/main, cited by content)

#	file	what it says today	governed?
1	.github/workflows/pr-automation.yml	route 0 of the 「Require a changeset」 message: 「do NOT apply 'skip-changeset' … LEAVE THIS CHECK RED」, and its paragraph 「Ruled on #18375 (ruling D …) the 'skip-changeset' label is never applied to a PR that edits an existing changeset」; the ::error:: line repeats it	no
2	scripts/check-empty-changeset.mjs	the DELIBERATE CORRECTION refusal: 「no label and no diff shape makes that safe … this gate stays red either way」	no
3	.claude/agents/os-dev.md	「PR 改到已有的 .changeset/*.md:⛔ 永不打 skip-changeset,门禁红着是对的,确认走 PR 文字。」	yes (Tier S)
4	.claude/skills/pm-dispatch/SKILL.md	「改到已有 .changeset/*.md 的 PR ⛔ 不打 skip-changeset;确认取维护者原话或出处三件转述。」	yes (Tier S)

⛔ Not a carrier: AGENTS.md's 「A bug fix in a released package takes a patch changeset — never none, and ⛔ never skip-changeset」. That is a different rule, for a PR that ships a fix, and this update does not touch it.

What the executing round must state, so the maintainer knows what the label now waives

The label is read by the changeset job's own guards. When applied, the job is skipped: that is how PR #18765 merged on 2026-09-18 with nothing in the job run. ⇒ The PR that rewrites the texts should list, from the workflow as it stands, which checks in that job stop running on a correction PR once the label is on. For example: whether the edited changeset's frontmatter or bump level is still validated anywhere else. ⛔ It does not change those mechanics unless the maintainer asks. It only states them.

Constraints for whoever takes it

• .claude/** is a governed surface, so rows 3 and 4 land by the maintainer's hand. ⛔ No seat readies, enqueues or arms that PR.
• .claude/skills/pm-dispatch/SKILL.md is at its line ceiling: row 4 is a rewrite of the same line, ⛔ never a net new line.
• Rows 1 and 2 are wording in a workflow message and a script message. ⛔ Zero if:, step or job changes.
• [finding] the skip-changeset label suppresses the DELIBERATE-CORRECTION refusal whose own text says 「no label and no diff shape makes that safe」 — declared contract against enforced behaviour #18375 is closed. The executing round leaves one comment there pointing to this card, so the ruling's home names its successor.

Dedupe

Repo-scoped issue search (mcp__github__search_issues), 「skip-changeset deliberate correction pending changeset ruling D route 0 label confirmation」: 10 shown (total_count 32, ranked by relevance, ⛔ not a count of matches). None is an open card updating this ruling. Nearest: #18375 (closed, the ruling's home), #19729 / #19735 (the cards behind #19746 / #19747), #18160 and #18404 (closed, the same gate's earlier rulings).

---

Generated by Claude Code

[os-support-ai]

Closed not planned — the maintainer chose to keep ruling D

domain:spec execution seat 1 (session_013RDBh5DqXd2xnLwvHLgLFr), 2026-09-23T07:15Z. The maintainer's follow-up in the same chat, verbatim: 「甲、维持现状。 用聊天里一句话确认,」. That selects the seat's option 甲: ruling D (batch #158 item 1 on #18375) stays as written, and a one-sentence confirmation in chat, carried to the PR with provenance, is the confirmation.

⇒ Nothing on this card is owed. The four carriers it listed are correct as they stand. On #19746 and #19747, skip-changeset was removed again, and the maintainer's chat sentence is carried as the confirmation of record (5790662865, 5790664030). Both PRs stay in the merge queue.

---

Generated by Claude Code