[os-steve]

Part of #18375 — ruling ① D's workflow-text edit only; the .claude/** half is serialised behind five open PRs

Clause-②: no

Scope

One file, one step, message text only: .github/workflows/pr-automation.yml, the
Require a changeset (or the skip-changeset label) step. No if: expression, no status
function, no new or removed step, no job or step renamed (the changeset-check job still
publishes the check-run name Check Changeset, and all 14 step names are byte-identical —
git diff reports zero - name: lines on either side). Ruling ① D is explicitly "no
workflow mechanics change", so the mechanics are untouched on purpose, including the three
suppression layers the card records.

Why this step carries the text

A PR whose only .changeset rows are ones it CHANGED rather than added adds none of its
own, so ADDED counts 0 and this step fails first. Every later step of the job
inherits the implicit success(), so the step that hosts the two-class refusal
(check-empty-changeset.mjs, COLLISION vs DELIBERATE CORRECTION) is skipped and its
remedy text has never printed for the class it was written for. Until the fenced-out
.claude/** half lands, this message is the only place that class reads anything.

Today that message prescribed the skip-changeset label for it, carrying the message's own PREFERRED marker.

The edit

• A leading route 0 discriminator in the job-log body: run
  git diff --name-status over .changeset/*.md against the merge base; every row M or
  D and none A means this PR is correcting somebody else's pending release note.
  Prescription for that class: do not apply skip-changeset; write the confirmation on
  the PR, naming the note and what changed under it; leave this check red. It states why
  the red costs nothing (Check Changeset is not one of the seven required contexts, so
  the approver merges over it), why the label is refused twice over (a false declaration,
  and it exempts the whole job), and that a PR which also releases something of its own
  still takes route 1 for that half.
• The ::error:: annotation (the Checks-tab one-liner) leads with the same discriminator
  before the label sentence, because a trailing caveat is the one shape that does not
  reach an author or an agent skimming the annotation.
• Route 2 keeps its PREFERRED marker for the ordinary releases-nothing case and gains the
  caveat that closes the read-alike trap: a correcting PR also releases nothing of its
  own, so route 2 reads as if it fitted.
• The unsure tie-breaker is scoped to routes 1 and 2, so it cannot overrule route 0.
• One comment sentence above the block ("Both lead with the label.") was updated to stay
  true of the two renderings it describes — declared here rather than done quietly, since
  a comment that describes the message falsely is the same declared-vs-actual defect this
  card is about. Nothing in the repo pins that sentence (git grep hit: 1, this file).

Verification

• node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack at
  c602393c derived 36 commands for this one path; each was run with its exit code
  captured before any pipe. Results are in the report comment on [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; the run was still
  in flight for the tail of the list when this PR was opened and is reported there.
• check-empty-changeset.mjs's consumer battery pins this step from outside: it requires
  exit 1 within 900 characters of the ::error::This PR adds no changeset anchor.
  Measured before and after: the gap was 428, it is now 698 (headroom 202), and the
  battery's regex still matches. The other pins it holds over this job are untouched —
  five failable steps, both live label reads on every one of them, two
  grep -qxF 'skip-changeset' reads, no continue-on-error.
• The advisory claim in the new text is measured, not recalled:
  scripts/check-required-contexts.mjs mentions Check Changeset 0 times, with
  controls lit in the same file (Lint & Repo Gates 24, TypeScript Type Check 18,
  Test Core 15, Governed Surface Queue Guard 6).
• The YAML parses, the job still has 14 steps with unchanged names, and the rendered
  run: script passes bash -n.

Changeset

None, and the skip-changeset label is route 2 for this PR: measured, no package manifest
lists .github or a workflow path in its files[] (positive control: the files[] arrays
of packages/spec and packages/cli read back non-empty). This PR is not the route-0
class either — it adds no changeset and changes none (git diff --name-status over
.changeset/*.md against the merge base returns zero rows).

Acceptance notes

• The other half of ruling ① D (.claude/agents/os-dev.md and the pm-dispatch text
  carrying the one-line rule) is deliberately not here: those are governed paths with
  five open PRs already editing them, and the dispatch fenced them out for a later round.
  [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 therefore stays open, which is why this body says Part of and not a closing
  keyword.
• The L3 step-order finding (zero status functions in changeset-check, so the refusal
  step never runs for this class) stays recorded on [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 and is not fixed here — the
  ruling rules mechanics out. The new text states the consequence instead of routing
  around it.

---

Generated by Claude Code