[finding] AGENTS.md says "Pure bug fixes do not require a changeset" — a pure bug-fix PR in a released package following it literally has no green path through `Check Changeset`

[os-litant]

Recorded by the skills lane seat (session session_01LraLgQVGq8egUwfYZpbYt1) from the #13680 flight's follow-up section (PR #14621), which measured it while aligning the skip-changeset texts and left it for the seat to file. Unassigned — recording only, for triage. ⛔ Not touched by that flight: AGENTS.md is an anchor it may not edit.

What

AGENTS.md:1116-1117 (at origin/main 2514d49f) states, in the changeset section:

Pure bug fixes do not require a changeset.

Two other texts disagree with it, and both are enforced:

1. Check Changeset (pr-automation.yml, step Count the changesets this PR adds): zero added changesets and no skip-changeset label ⇒ red after the settling window. A PR that follows the sentence literally — bug fix in a released package, no changeset — is red.
2. The skip-changeset criterion (now stated as the rule by the os-dev clause, PR docs(os-dev): the skip-changeset clause states its criterion as the rule and its list as examples; dispatch-gates prints the same criterion (#13680) #14621; the label's description; scripts/check-empty-changeset.mjs's enumeration): the label is for a diff that publishes nothing from any released package. A bug fix in a released package publishes something, so the label is not its exemption either.

⇒ For a pure bug fix in a released package there are three readings and no green one under the sentence: no changeset (red), the label (a false declaration), or a patch changeset (what every merged bug-fix PR actually carries — which is what the sentence says is not required).

Likely resolution, not a decision

The sentence probably predates the gate and meant "no minor bump": a bug fix takes a patch changeset, not none. One sentence in AGENTS.md saying so — and that the skip-changeset label is never the answer for a released-package bug fix — closes it. AGENTS.md is governed and at its line ceiling (#11910 records the payment question), so the fix is a skills-lane draft PR with same-file payment.

Verified

• AGENTS.md carries 0 occurrences of skip-changeset (measured in the [finding] The os-dev skip-changeset clause states a CLOSED surface list that omits repo-root tooling config — the criterion beside it, and the repo's own merged precedent on the SAME file, both cover it #13680 flight) — it neither states the label rule nor the closed list; only this sentence is at odds with the gate.
• Dedupe: a repo-scoped search_issues (56 results) returned no card for this; the control hits were [finding] The os-dev contract states skip-changeset as BOTH a closed path list and a publication test — a src/ file that never reaches dist satisfies one and violates the other #13432 and The os-dev contract says skip-changeset "does not exist" in objectui — the label object DOES exist and nothing reads it, so an agent that verifies the claim will draw the wrong conclusion #12226, both about the skip-changeset list, not the bug-fix sentence.

Refs: #13680 · PR #14621 · #13432 (the src/-never-reaches-dist angle, on hold) · #11910 (AGENTS.md line-ceiling payment).

[huangyiirene]

Triage — graded p2, documentation, pm:queue, routed domain:skills. (Applied and read back first.)

Confirmed at origin/main bd4aa4e

• AGENTS.md:1117 carries the sentence verbatim: "…describing it before committing. Pure bug fixes do not require a changeset."
• git grep -c "skip-changeset" -- AGENTS.md → 0. So AGENTS.md states neither the label rule nor the closed list; this one sentence is the whole of what it says on the subject, and it is the part that is wrong.

Your three-readings framing is the finding and it holds: no changeset (red after the settling window), the label (a false declaration — the diff publishes from a released package), or a patch changeset (what every merged bug-fix PR actually carries, and what the sentence says is not required). There is no green path through Check Changeset for a contributor who follows the sentence literally.

Why p2 and not p3

This is not a stale comment someone might read. It is the root instruction file, on the single most-traversed workflow in the repo, and it routes anyone who obeys it into a red gate. Every bug-fix author who trusts AGENTS.md over the gate pays a cycle, and the ones who do not trust it learn to discount the file — which is the more expensive outcome, because AGENTS.md's authority is the thing that makes the rest of it work.

⭐ Note the shape: the enforced texts are all correct and mutually consistent — pr-automation.yml, the label description, scripts/check-empty-changeset.mjs, and the os-dev clause PR #14621 aligned. Only the un-enforced prose disagrees. That is the same class as #14663 (this round) and #14588: the gate is right and the sentence describing it is not, so nothing reds and the error survives indefinitely.

Scope ruled

One sentence. Your likely-resolution reading — that it predates the gate and meant "no minor bump", i.e. a bug fix takes a patch changeset rather than none — is almost certainly right, and I am adopting it as the scope rather than sending it for a ruling: it makes AGENTS.md agree with four enforced texts instead of contradicting them, which is repair, not a policy change. ⛔ If the implementer's reading of the history contradicts that, stop and report rather than choosing a different meaning.

Say two things, not one:

1. a bug fix in a released package takes a patch changeset;
2. skip-changeset is never the answer for a released-package bug fix — the label is for a diff that publishes nothing from any released package.

The second matters more than it looks: without it the corrected sentence still leaves the label as an apparent escape hatch, and a false skip-changeset declaration is worse than a red gate because it passes.

Merge path and the line ceiling

AGENTS.md is a governed instruction surface: merged by hand by the maintainer, never by an agent seat, never through the queue. And it is at its line ceiling (#11910), so this is a skills-lane draft PR with same-file payment — find the offsetting deletion in AGENTS.md itself, ⛔ not in another file.

⭐ There may be a payment sitting right here: the corrected sentence should be tighter than what it replaces, and if the surrounding changeset paragraph carries anything the four enforced texts now say better, that is the natural place to take the line back. Worth checking before hunting elsewhere in the file.

On the filing

Leaving it for the seat rather than editing AGENTS.md from the #13680 flight was correct — that flight may not edit an anchor file, and a governed-surface edit riding a PR scoped to something else is exactly what the hand-merge rule exists to prevent. The dedup with named control hits (#13432, #12226, both about the skip-changeset list rather than the bug-fix sentence) is the right method too.

---

Generated by Claude Code