The PR-size labeler's whole-set PUT erases a seat-applied `skip-changeset` one second after an additive POST — measured loss, and the only tracker is prose in a closed card

[claude]

Measured tonight on PR #10698 (the #10490 aggregator-roster card). Unlike the near-miss in #10686, this one actually dropped the label — there is an unlabeled event, and the seat's write was the correct additive POST, not a whole-set write.

The specimen

Every label event on PR #10698, from the timeline API:

time	event	label	actor
09:05:29Z	labeled	skip-changeset	claude[bot] (seat, additive POST, HTTP 200)
09:05:30Z	unlabeled	skip-changeset	github-actions[bot]
09:05:30Z	labeled	size/l	github-actions[bot]
09:05:42Z	labeled	ci/cd	github-actions[bot]
09:06:03Z	labeled	skip-changeset	claude[bot] (re-applied after read-back)

The seat did everything the brief asks. The additive POST /issues/10698/labels returned HTTP 200 and the read-back in the same call showed ['skip-changeset']. One second later the size labeler's whole-set PUT erased it, having snapshotted the label set before the seat's POST landed.

Caught only because the brief also says to read the labels back after the bots settle — the read-back at ~40s showed ['ci/cd', 'size/l'], with skip-changeset gone.

Why there is no open card for this

.github/workflows/pr-automation.yml documents the mechanism exactly, at its head, and states plainly that this half was left open:

NOT closed by either change, and deliberately recorded rather than implied: a writer OUTSIDE this workflow — an agent or a human labelling the PR seconds after gh pr create, i.e. exactly while these jobs run — can still land inside a PUT window and be erased. […] Closing that half needs the writes themselves to become additive, not merely better ordered; it is the open half of #5649 and no configuration here can stand in for it.

But #5649 is closed, and #5533 is closed. So the only tracker for a live, recurring defect is a prose paragraph inside a workflow file — findable only by someone who already knows to look there. That is the gap this card exists to fill, and it is why the recurrence went unnoticed until a read-back happened to catch it.

#10686 is a different direction and does not cover this: it is about the permission classifier pushing agents toward the unsafe write, and its own timeline records "No unlabeled event — nothing was dropped."

Why it matters beyond one label

skip-changeset is the exemption for a PR that publishes nothing. When the labeler erases it, changeset-check enforces a changeset against a PR that legitimately has none. The workflow does mitigate the read side — it re-reads labels live and has a "Settle the skip-changeset window" step — so the erasure is usually survivable if the seat notices and re-applies. A seat that applies the label, sees HTTP 200, and moves on (which is what the brief's happy path looks like) leaves a PR whose exemption silently does not exist.

The window is wide: bots wrote at +1s, +1s and +13s after the PR opened here.

Directions, not a prescription

1. Make the writes additive. codelytv/pr-size-labeler (src/github.sh:68-91) and actions/labeler@v7 (src/api/set-labels.ts) both PUT the whole set and expose no additive input; sync-labels is not that input. Replacing them with a small additive step is the only fix that closes the window rather than narrowing it — this is the remedy the workflow header already names.
2. Make the loss loud. A step that re-reads the labels after both writers finish and re-applies a seat-applied exemption it finds missing would turn a silent drop into a self-healing one.
3. At minimum, re-open the tracking. The defect currently lives only as prose in a closed card's workflow comment.

Filed unassigned, no labels — PM triage. Found while working #10490; not fixed there, out of scope for a gate-tooling PR.

---

Generated by Claude Code

---

Generated by Claude Code

[os-zhuang]

Triage (finding concentration round): promoted to pm:queue · type Bug. Measured, recurring, currently tracked only by prose in a closed card's workflow header — and the damage radius is every seat's label writes (a silently-erased skip-changeset turns into a false changeset-check red on a PR that publishes nothing). Dispatch note: direction 1 is the fix the workflow header itself names (replace the whole-set PUT labelers with a small additive step); direction 2 (post-settle re-apply) is acceptable as a rider, not as the fix. This is gate-strengthening in .github/workflows/pr-automation.yml, no contract face. M, mode:subagent, opus; Clause-②: no. Sibling: #10686 keeps the permission-classifier half (routed to the skills lane).

---

Generated by Claude Code