docs(pm-dispatch): cloud publish-fallback protocol, probe-control validity, type-check-debt dispatch clause (#8007)

[huangyiirene]

Fixes #8007

⚠️ PR opened by the domain:metadata PM seat on the implementing dev's behalf — and the reason is the subject of this very PR: the objectstack-ai org does not have the Claude GitHub App connected, so cloud dev containers can git push but cannot reach the GitHub API to open a PR, comment, or file an issue. Measured 5 of 5 dispatches this shift.

Authored on claude-fable-5, per this lane's standing rule that any card touching .claude/skills/pm-dispatch/** is filed and dispatched, never PM-edited. That rule was kept even under an explicit instruction to update the skill — the instruction is satisfied through the card, not around it.

What this records

One shift's measured failures, distilled into the skill:

1. Cloud publish-fallback protocol. Dev sessions that finish work and cannot publish it park silently in need_input — one sat idle 1h48m before the PM noticed. The skill now carries: the brief goes in full inside the dispatch prompt (a brief that lives only as a GitHub comment is unreachable — one dev reported literally "issue Product question: an uninstall with no organizationId deletes EVERY organization's rows for that package (measured 5 of 5, including a foreign org's) #7780 comment text unreadable; need GitHub access or paste §1–§6"); every prompt carries a publish-failure clause; and the PM opens the PR from the dev's own changeset, read with git diff origin/main...FETCH_HEAD -- .changeset/ — ⛔ never git ls-tree .changeset/ | head -1, which returns the alphabetically-first pre-existing changeset and produced identical unrelated text for two different branches before it was caught.

2. Probe-control validity. The existing "pair every zero-hit with a positive control" rule needed a sharper clause: a merged branch is auto-deleted, so it is not a valid control. This seat used one, got empty output, and nearly read it as "probe works, nothing there".

3. check:type-check-debt as a dispatch-time gate clause. Three PRs hit this gate in one shift. Packages in the ledger exclude their own tests from tsconfig, and some (@objectstack/metadata-protocol) have no typecheck script at all, so pnpm typecheck stays green while the test layer drifts. It needs a built closure or exit 1 is a missing-module cascade rather than the ledger. ⛔ Never raise an entry to admit a file added in the same PR; ⛔ never --lower other packages' informational lines.

Also carried: the mutually-exclusive-region exemption protocol (declare regions at claim · merge main before the PR · merge again after each sibling lands · queue arbitrates), and the reading that makes it safe — it guards against concurrent editors, not against a finished PR sitting in a merge queue.

Shape

Written as principles distilled in place, without issue-ID citations, matching the direction main took today in #7938 (os-dev) and #7948 (AGENTS.md). Edits are largely in-place rewrites rather than appended sections — 75 insertions against 36 deletions across SKILL.md, references/dispatch-runbook.md, and references/platform-readings.md.

⚠️ #7946 and #7971 both touched this skill earlier today; the dev was instructed to branch from current main and check for overlap rather than re-litigate or duplicate their changes.

⚠️ The dev was explicitly asked for judgement about what to leave out — a skill that accretes every incident becomes unreadable, and #7971 was itself a pruning PR. Its report on which findings it declined, and why, is still owed and will be appended as a comment.

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
objectstack	Ignored		Aug 12, 2026 12:11pm

[huangyiirene]

⛔ HUMAN MERGE REQUIRED — this PR will not be armed, and must not be

The implementing dev surfaced a constraint my dispatch brief did not carry:

Skills updates are ADR-class (maintainer ruling 2026-08-11): human review + human merge — do not queue, do not enable auto-merge.

⇒ Auto-merge is deliberately NOT enabled and this stays a draft. It is waiting on a human, not on CI. Flagging explicitly so it does not read as a stalled PR — and so a later round of mine does not "helpfully" arm it.

CI status

Check Changeset was the only red, and the gate named its own remedy:

If it releases nothing (including any skills/** change — see #5947), apply the skip-changeset label

.claude/skills/** publishes no package, so I applied the label. ⛔ Note the same message forbids the tempting alternative: a newly added empty-frontmatter changeset is rejected (#5471), because it is a real input to changesets/action and an all-empty set stalls the release silently and greenly (#4898). The label is the sanctioned route, and it matches both precedent PRs on this file set today.

Everything else concluded green.

⭐ The dev's judgement — the part worth reviewing

Its full report is on #8007 (5266472956). It declined several findings I handed it, with reasoning I think is right in every case: the "already-answered-question" check as a near-duplicate of existing rules; the unlabeled-filing finding as a platform fact rather than a new rule; the whole cloud-publish complex placed in the runbook rather than SKILL.md and written to name its own retirement, on the ground that it is org infrastructure rather than PM-loop doctrine.

⭐⭐ And it refused to reframe the region exemption as a standing authorization — because I asserted the maintainer "has since authorised it" without supplying a dated verbatim quote, and this file's own convention requires date + 原话. That is a correct evidentiary pushback on its PM, and I am recording it as such rather than overriding it. If the maintainer wants the exemption recorded as standing, a dated verbatim ruling on an issue makes it a one-line follow-up.

⚠️ Two errors in my brief, corrected by the dev

• I named gates that do not exist. check:skill-refs / check:skill-docs are not root scripts. The real ones are the frame-sync / freshness / compatibility trio plus check:pm-skill-ratchet / check:pm-skill-id-lint — which the dev found and ran. I invented plausible names instead of reading package.json.
• SKILL.md is under a shrink-only line ratchet I did not know about. The dev held it at exactly 686 → 686, paying for additions by compressing five passages whose detail already lives verbatim in references/ — the extraction discipline the ratchet script itself prescribes.

Out of scope, filed

check:skill-frame-freshness is red on clean main (fixture drift after the 2026-08-12 principles-only rewrites; it dies in --self-test before scanning). While it is red, every PR touching these files sails past a dead gate. Filed as #8024, verified against an untouched checkout.

---

Generated by Claude Code

[huangyiirene]

Closing — maintainer decision, 2026-08-12

The root cause this PR documents has been fixed at source: the Claude GitHub App is now connected to the objectstack-ai org. The maintainer's call is to close rather than narrow, on the reasoning that the findings may be knock-on effects of that outage and are better re-derived if they recur than carried speculatively in the skill.

I argued for narrowing instead — keeping the probe-control and TEST_DEBT clauses, which are independent of GitHub connectivity — and the maintainer reaffirmed. Recording that plainly so the disagreement is on the record rather than implied, and proceeding.

What is being discarded, stated honestly

Of the four change groups, one was caused by the outage (the cloud publish-fallback protocol) and three were not:

• probe-control validity — a merged branch is auto-deleted, so it is invalid as a positive control. This caught a bad measurement of mine today, and it is a git/GitHub behaviour with no connection to the App.
• the TEST_DEBT dispatch clause — check:type-check-debt, built closure first, and the two anti-patterns. Three PRs hit that gate in one shift, and on one of them the "type errors" were two real API misuses (registerObject called without its required packageId) that vitest structurally cannot see.
• four SKILL.md clauses — the per-hot-file editor cap, the "guards concurrent editors, not queued PRs" reading, re-read the lane from GitHub every round, and the debt-gate summary line.

⚠️ These are not lost. All three survive as ⭐⭐ entries on the seat sticker #6367, which is this lane's operational memory and was refreshed this round. They will keep governing this seat's dispatches; what is lost is sharing them with other lanes through the skill. That is a real but bounded cost, and it is the maintainer's to accept.

Credit where it is due

The implementing dev (on claude-fable-5) did the best judgement work of the shift on this card: it declined several findings I handed it as near-duplicates, placed the cloud complex in the runbook rather than SKILL.md on the grounds that it is org infrastructure, not PM-loop doctrine, wrote that section to name its own retirement condition — which is precisely the event that closes this PR — and held SKILL.md at exactly 686/686 on the shrink-only ratchet by compressing passages whose detail already lives in references/.

It also refused to reframe the region exemption as a standing authorization because I asserted it without a dated verbatim ruling, which the file's own convention requires. That pushback was correct and remains correct.

Its report stands on #8007 (5266472956).

Not affected by this closure

• check:skill-frame-freshness is red on main — self-test fixture drifted after the principles-only rewrites, so the gate never scans #8024 — check:skill-frame-freshness red on clean main; a gate that dies in --self-test before scanning, so every PR touching those files sails past it. Pre-existing, verified on an untouched checkout, unrelated to this PR.

Branch claude/issue-8007-pm-dispatch-cloud-publish is left in place; nothing is force-deleted.

---

Generated by Claude Code