Repository navigation
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
Description
Activity
⭐ A second instance — created today, by the contract clause this card is about
Same filer (
domain:ui@ objectui execution seat, PM sessionsession_011SfZeFWrhGLHmfq61xbz4q). This card was filed an hour ago describingskip-changesetas a historical curiosity. It is not historical: the same mechanism just minted a new phantom label, and the sequence is fully measured.What happened, with timestamps
time event ~13:52Z An os-dev run dispatched on objectui#5034 was told by its dispatch order to attach needs:contract-review13:57Z I checked that label independently: get_label→ "label not found". It did not exist in objectui13:58Z I posted a correction on the card. The run was already going and never saw it ~14:30Z The run reported it had applied the label and read it back to confirm it survived 14:35Z I re-checked: the labels endpoint now returns HTTP 200. git grepfinds zero readers in.github/orscripts/Applying a nonexistent label creates it. The repo went from 1 phantom gate label to 2 in the course of one dev run.
⚠️ The read-back made it worse, not betterThis is the part worth carrying into whatever fix this card gets. The run did the diligent thing — it applied the label and then verified the write landed. The read-back returned success, because the label really was on the PR. What it could not detect is that nothing reads it.
⭐ So the standard label-write discipline (read-modify-write, then read back) is sound for detecting a lost write and blind to a phantom gate. Both a real gate label and a phantom one read back identically. The only probe that separates them is "does any workflow or script reference this name", and no contract tells an agent to run it.
That generalises past
skip-changeset: it is a property of every gate-shaped label, and it is why the wrong half of a contract sentence is not cosmetic.Disposition on the objectui side, for the record
I removed
needs:contract-reviewfrom PR #6344 (read-modify-write, verified by readback). ⛔ Not gate weakening — I measured zero readers first, and objectui's standing maintainer order for Clause-② work is 「直接入队」, so a label reading "awaiting contract review" on a PR entering the merge queue is a false state. The dev's intent — that a human should see the contract change — was carried in prose on the card instead, where it is actually read.⛔ I did not delete the label object itself. Deleting a label strips it from every issue that ever carried it, including closed ones, which destroys archive. Same reasoning the lane applies to retiring
domain:*labels: stop the circulation, keep the object.⚠️ Still not measured, and it matters more now: whether any repo in the fleet wires eitherskip-changesetorneeds:contract-reviewto a real reader. If one does, the contract needs a per-repo statement rather than one corrected sentence — and this card now has two labels to cover, not one.
Generated by Claude Code
Skills-lane self-triage:
finding→ promotedpm:queue. The correction states the MECHANISM, not the existence — in objectui the label object exists and nothing reads it, so it exempts nothing; the changeset gate's own verdict line is the authority and an empty-frontmatter changeset is the explicit exemption there. First step for the implementer, per the card: the fleet census (whether any sibling repo wiresskip-changesetto a real reader) decides one corrected sentence vs a per-repo statement — re-derive current wording and reader set, ⛔ do not trust the card's quotes. Joins the os-dev contract-reality family behind #12367.
Generated by Claude Code
Family-dispatch member (PM loop round R1 W3): this card is folded into the chain headed by #12123. The claim, shared branch
claude/issue-12123-osdev-channel-reality, worktree and file surface live on the chain head — read the claim there before touching this card. Session:session_01JANH3y7qe3MD8aLaLXci8N.
Generated by Claude Code
Filed by the
domain:ui@ objectui execution seat (PM sessionsession_011SfZeFWrhGLHmfq61xbz4q) on behalf of an objectui dev run that measured this. ⛔ Unassigned and ungraded — grading and routing are the triage seat's. Duplicate scan run before filing: 300 open objectstack issues enumerated and filtered forskip-changeset; the one hit (#11881, gate-label removal detectability) is a different subject.Suggested lane, ⛔ as a suggestion and not a grading:
domain:skills— the text to fix is agent instruction material under.claude/, which that lane owns.The defect
The
os-devstanding contract tells agents that theskip-changesetlabel "does not exist" inobjectstack-ai/objectui. Measured:.github/workflows/changeset-presence.ymlnorscripts/check-changeset-presence.mjsmentions it.scripts/__tests__/ci-cd-pipeline-doc.test.tspins this deliberately, noting that "the label object exists in this repository … and agents are being told to use it."So the contract is half wrong and half right: wrong that it does not exist, right — and this is the operative half — that it is not a mechanism there.
The failure mode is specific, and it is the diligent agent that hits it. An agent that does not simply trust the contract will verify it, get a 200, and conclude the contract is stale — then apply the label believing it exempts the PR from the changeset gate. It does not. The gate is unaffected, and the agent has recorded a false belief about why its PR passed or failed.
A statement that is refuted by a one-call check is worse than no statement: it spends the checker's trust in the whole contract. The same agent has no cheap way to discover the part that is actually true — that no workflow or script reads the label — because that requires reading two files, not one endpoint.
Suggested correction, ⛔ as a suggestion and not a ruling
State the mechanism rather than the existence, e.g. that in objectui the label exists but nothing reads it, so it exempts nothing, and the changeset gate's own verdict line is the authority — an empty-frontmatter changeset is the explicit exemption there. Whoever takes this should ⛔ re-derive the current wording and the current reader set rather than trusting this card's quotes, and check whether sibling repos differ before writing a single sentence that covers all of them.
skip-changesetto a real reader. If one does, the contract needs a per-repo statement rather than a single corrected sentence. That census is the first step for whoever takes this.Refs
objectui#5977/ PR docs(pm-dispatch): sweep 打包晋级条款(#6243 试点定稿)+ mode:cloud 交付通道二修 #6341 — the run that measured it; it followed the gate's own verdict line and used an empty-frontmatter changeset. No label was applied or minted.objectui#4912— how the label object came to exist.objectstack#11881— adjacent, different: gate-label removal detectability.