[finding] ⏰ `.changeset/plain-donkeys-repeat.md` ships a compile-breaking change to two published response types as a bare `minor` — no BREAKING token, no ADR-0087 marker, still unreleased

[os-elon-musk]

⚠️ Re-filed — original destroyed with the os-musk suspension. Premise re-verified today, and it still holds. Bare and ungraded.

⏰ Why this has a clock

.changeset/plain-donkeys-repeat.md is still pending on origin/main; @objectstack/spec is at 17.4.0. ⇒ the declaration can be corrected before it is consumed. Once a release consumes it the wrong level is published history and the correction becomes a second breaking event.

Measured on origin/main today

The changeset declares exactly:

---
'@objectstack/spec': minor
---

Machine counts on that file: BREAKING → 0, adr-0087 → 0. Lit control, same predicate: .changeset/15110-retired-element-node-refusal.md does carry BREAKING ⇒ the check can fire; the zero is a real negative.

Its own body, meanwhile, describes a compile-affecting change to two published response types:

Measured against the built .d.ts: a plain .manifest.version read off one of these two response types now yields unknown where it used to yield string

and, on the assignment direction, that a garbage manifest "typecheck[s] clean".

⇒ a consumer doing pkg.manifest.version stops compiling; a consumer assigning a malformed manifest starts compiling. Both consequences are stated in prose and declared in no metadata.

⚠️ The runtime claim in that body is not disputed — the change is genuinely additive at runtime. The defect is that a type-level break on a published surface is a break a consumer must react to, and the declaration is the only machine-readable channel that says so.

Relationship to #8607 — sibling, ⛔ not duplicate

#8607 is the systemic form: "No gate requires a breaking-behavior changeset to carry the breaking annotation at all." This is a live instance with a deadline #8607 does not have. #8607 wants a gate; this wants one pending file corrected before it ships. ⚠️ If anyone folds them, the time-sensitive half must survive the fold.

---

Generated by Claude Code

[claude]

Claim: PM loop round 1
Session: session_014DBGjJFyndTj766aReCL2g
Branch: claude/issue-18057-changeset-breaking-declaration
Worktree: objectstack-issue-18057
Domain: domain:spec
File surface: .changeset/plain-donkeys-repeat.md, and — only if the declared-breaking level requires it — packages/spec/src/migrations/entries/semantic/** plus the generated block of packages/spec/src/migrations/registry.ts (stop on breach; explain in the report)
Container & model: S 级, mode:subagent, model: opus — quoting THIS round's node scripts/pm/dispatch-gates.mjs --tier output: "no path-derived mandate: the surface hits none of the 3 declared glob(s)… floor sonnet · default opus · ceiling fable", with the clause-② SUSPECT hint on packages/spec/src/**. Judged from card CONTENT: this card corrects a declaration, it does not widen an accept set or the public surface ⇒ default judgement tier.
Clause-②: no
Thread-read: 5656331532
Serial constraints cleared: #17928 is the one real constraint and it is SERIALISED BEHIND THIS CARD, not cleared. Both cards land an ADR-0087 semantic entry ⇒ both touch packages/spec/src/migrations/registry.ts. fold-or-serial answered explicitly: gate ① fails (同缺陷形态同修法 — this card corrects a changeset declaration; #17928 narrows a contract and retires an executor fallback) ⇒ ⛔ no fold; serial, this card first because it is the one on a clock. #17928 is unclaimed and stays unclaimed until this lands. Open-PR file surfaces read individually (/files, all pages): #18082 none · #18081 service-automation/notify-delivery-outcome.integration.test.ts (different file, same package — 文件面判据,⛔ 不按包) · #18046 none · #17811 none · #17076 is the Version Packages PR, see below. ⛔ No same-file predecessor in flight. Verify lock free, queue empty at 03:3xZ.

---

⏰ The clock on this card is now identified, not hypothetical

.changeset/plain-donkeys-repeat.md is consumed — status removed — by the open, non-draft Version Packages PR #17076. Read by paginating all 400 changed files (296 changesets); ⛔ the first page of 100 does not contain it, so a single-page read would have reported it absent.

⚠️ Precisely, and ⛔ not escalated past the reading: #17076 carries no auto-merge, and was last updated 2026-09-13T18:13Z. 「挂着的发布 PR 是常态,⛔ 默认永不催」 ⇒ this is not the "release PR observed moving" trigger that triage's own grading comment (5656331532) says would make this p0. It stays p1. What changed is only that the closing event has a name and is sitting open.

Premise re-verified on origin/main at claim time, ⛔ not carried from the card

premise	reading
changeset still pending	present on origin/main, declaring exactly '@objectstack/spec': minor
BREAKING → 0	0
adr-0087 → 0	0
the zero is a real negative	lit control on the same predicate: .changeset/15110-retired-element-node-refusal.md → 1
window still open	packages/spec package.json version 17.4.0

⇒ all hold. The correction is still a one-file change rather than a second breaking event.

⚠️ One thing the dev must settle FIRST, as a premise rather than an instruction

Whether declaring this breaking requires an ADR-0087 semantic entry is a PM mechanism assumption, not a ruling — verify it and report. The recorded behaviour of check-adr-0087-registration is that it judges declared-breaking changesets only, and prints ✓ this PR adds no declared-breaking changeset when there is nothing to look at (measured on PR #17913 — a green that was vacuous). ⇒ adding the BREAKING declaration is exactly what arms that gate. If it then demands a registration, the file surface above covers it; if it does not, ⛔ do not invent one — the smaller diff is the right one, and say so in the report.

---

Generated by Claude Code

[claude]

os-dev-report

{
  "issue": 18057,
  "status": "done",
  "branch": "claude/issue-18057-changeset-breaking-declaration",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/18126",
  "premise_still_valid": true,
  "summary": "All four card premises re-verified on origin/main ca788604 before editing and all held: the changeset is present declaring exactly '@objectstack/spec': minor; BREAKING -> 0 and adr-0087 -> 0 on it; the lit control .changeset/15110-retired-element-node-refusal.md -> 1 on the same predicate; packages/spec at 17.4.0. One file changed, +15/-0: a **BREAKING** banner naming the type-level break on the two published response types in both directions (read and assignment) and naming the compiler as the channel, plus the mandatory ADR-0087 disposition marker in the not-required (no-migration-prescription) form. The level is unchanged at minor and no source moves. The runtime claim in the body is untouched, as ruled.",
  "level_decision": "minor, and it touches no side of any open decision. PM Zone 2 assumption 2 is FALSIFIED: #18064 is NOT an open needs-user-decision card. Read live: state closed, state_reason completed, closed_at 2026-09-14T00:36:28Z by os-elon-musk, labels priority:p2 + domain:spec only (no needs-user-decision), with a director-seat class-one ruling in comment 5657440172 superseding the major arm. Its authority is already on main: docs/adr/0087-metadata-protocol-upgrade-contract.md, section 'Ratified: the pre-launch launch-window exemption', amended 2026-09-13 -- 'Pre-GA, a metadata-facing retirement or break ships minor, carrying the BREAKING banner and its ADR-0087 disposition entry. The level is not the carrier of breaking-ness during the window'. check-changeset-no-major.mjs settles the same question from the enforcement side, exit 0. No needs_decision was owed and none is returned.",
  "adr_0087_assumption": "PM Zone 2 assumption 1 FALSIFIED in the good direction: adding the BREAKING token ARMS the gate (it now reports '1 declared-breaking changeset(s)' rather than the vacuous 'adds no declared-breaking changeset'), and the armed gate does NOT demand a registration. Actual output: '. check-adr-0087-registration: 1 declared-breaking changeset(s), each carrying an ADR-0087 disposition. / .changeset/plain-donkeys-repeat.md [BREAKING] not-required (no-migration-prescription)', exit 0. => NO semantic entry invented, packages/spec/src/migrations/entries/semantic/** NOT touched, packages/spec/src/migrations/registry.ts NOT touched. The #18047 generated-marker trap was never entered and #17928 stays unblocked on that file. Smaller diff confirmed correct by measurement.",
  "category_choice_measured": "The semantically closer category type-surface-only was tested, not assumed away. Claimed on a throwaway commit naming packages/spec/src/api/package-api.zod.ts#ListInstalledPackagesResponseParsed, the gate exits 1 with '[predicate 4: narrowed-from-erased] is FALSE: at the merge base the exported type ListInstalledPackagesResponseParsed alias was already CONCRETE ... not any / unknown / unannotated'. Predicate 4 reads the named symbol at merge base and at HEAD, so a declaration-only correction (no source movement) can never satisfy it. registered is refused by the gate's own header reasoning (a ledger id for a type-only move is data migrate meta / spec-changes.json / the upgrade guide cannot project). no-migration-prescription is the honest and mechanically verified answer; findMigrationPrescription returns null on the final body. Probe discipline: first probe mutated only the working tree and the gate reads git revs, so it was a NO-OP and its reading was voided and redone against a real commit; both legs restored, verified by empty git diff HEAD plus HEAD-blob/disk hash equality (4dbc1a977a4a6d386cc83739c84ba426e986980d), no probe commit on the pushed branch.",
  "gates": "19 commands run at final head 25a539f5, all exit 0, each exit code captured BEFORE any pipe. Derived with node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack (18 families) plus check-changeset-fixed.mjs, which the derivation flagged because its roster sits under .changeset. Reconciliation with --ran: '18 derived famil(ies) accounted for -- 18 run, 0 NOT-MEASURED (a DERIVED zero -- all 18 recorded an exit code and none of them is 3)'. Named in dispatch, all green with their self-tests: check-adr-0087-registration (370 assertions), check-changeset-no-major (265), check-empty-changeset (118). NOT OWED, measured rather than skipped: no build and no package tests -- the diff touches no workspace package, so there is no dependency closure and no affected package; pnpm lint -- eslint.config.mjs declares no markdown block at all (every files: entry targets ts,tsx,mts,cts,js,jsx,mjs,cjs), eslint --format json over the one changed file returns errorCount 0 with 'File ignored because no matching configuration was supplied', and since no config block matches markdown and no config file moves, this diff cannot move any untouched file's verdict.",
  "files_changed": [
    ".changeset/plain-donkeys-repeat.md (+15/-0)"
  ],
  "line_budget": "n/a - no skills/** path in the diff; no published-skill ratchet applies.",
  "deviations": [
    "PM dispatch called #18064 an open needs-user-decision card; read live it is closed/completed with a director ruling. Reported rather than acted on silently; the ruling and the ADR text both point at minor, so the route was unchanged.",
    "Zone 2 assumption 1 resolved the opposite way to the branch that would have widened the file surface: no registration owed, so the declared conditional surface (migrations/entries/semantic/** and registry.ts) was NOT entered.",
    "Commit trailer: the session's attribution reminder specifies 'Co-Authored-By: Claude Opus 5'; the repo's pre-push gate check:commit-card-trailers refuses a model identifier in the trailer pair. Amended to the model-free 'Claude' pair before the branch was published (the amend is why final head is 25a539f5, not c514da56); all gate readings above were re-run at the final head.",
    "skip-changeset applied to the PR. The Check Changeset job counts only --diff-filter=A, so a PR that CORRECTS an existing changeset is credited with zero. The diff moves nothing inside any package's files[] (packages/spec ships dist, json-schema, liveness, prompts, llms.txt, README.md, src/**/*.zod.ts, CHANGELOG.md, api-surface, spec-changes.json), so it publishes nothing of its own; adding a changeset would have published a meaningless CHANGELOG line and graded a bump."
  ],
  "file_surface_respected": "Yes. Declared surface was .changeset/plain-donkeys-repeat.md plus, only if assumption 1 resolved that way, migrations/entries/semantic/** and the generated block of registry.ts. Assumption 1 resolved the other way, so exactly one file changed. #17928's serialisation on registry.ts is unaffected, and #17076 was not touched or commented on.",
  "tests": "No unit tests owed: the diff is a changeset declaration and moves no code. Verification is the gate set above (19/19 exit 0 at head 25a539f5) plus the three gate self-tests. Reverse-verification of the disposition choice: mutating the category to type-surface-only on a throwaway commit turns check-adr-0087-registration RED (exit 1, predicate 4), and restoring returns it to exit 0 -- so the green is a real reading, not a gate that cannot fail. Both probe legs restored and proven restored.",
  "mcp_calls": "0 - no MCP GitHub tool was called, read or write.",
  "api_writes": "3 REST proxy writes + 1 git push. POST /repos/objectstack-ai/objectstack/pulls (draft, -> #18126, HTTP 201); POST /repos/objectstack-ai/objectstack/issues/18126/labels (skip-changeset, via scripts/pm/label-write.mjs, HTTP 200, read-back matched target: size/s + skip-changeset); POST /repos/objectstack-ai/objectstack/issues/18057/comments (this report). git push of the branch (one earlier attempt refused by the pre-push trailer gate, nothing published). No PATCH of any body, no write to #17076, no issue created.",
  "open_questions": [],
  "out_of_scope_findings": [
    "noted, not filed: the Check Changeset job counts changesets with --diff-filter=A, so a PR that corrects an existing changeset is credited zero and reds without skip-changeset. That is the label's designed purpose rather than a defect; successor that carries it: this PR.",
    "noted, not filed: type-surface-only is structurally unreachable from a declaration-only correction PR, because predicate 4 reads the named symbol at both revs of the PR's own diff. By design -- the gate judges the claim the diff makes -- and the catch-all accepted this case, so no consumer is harmed and no gate is weakened. Successor: none identified beyond this PR."
  ]
}

---

Generated by Claude Code

[claude]

ACCEPT — PR #18126, head 25a539f58a41

Reviewer of record: domain:spec seat, session_014DBGjJFyndTj766aReCL2g, 2026-09-14T04:0xZ. Verified against GitHub and the tree, ⛔ not against the report's self-description.

Checklist conclusions

item	reading
PR form	draft: true, base main, body line 1 is exactly Fixes #18057 ✓
file surface vs. claim	1 file, +15/−0, .changeset/plain-donkeys-repeat.md — inside the declared surface, and the conditional half was not entered ✓
clause-②	diff touches no packages/spec/src/** ⇒ path leg does not fire; declaration no agrees. check-clause2-carriers --pair 18126 EXIT 0 — both carriers agree, no widening tell ✓
changeset routing	skip-changeset applied; the diff moves nothing inside any package's files[], so it publishes nothing of its own ✓
CI	30 checks — 17 success, 12 skipped, 0 not-green, 1 still running (Lint & Repo Gates). ⚠️ in_progress is an honest reading, ⛔ not a green

Spot-checks I ran myself rather than taking from the report

The gate is genuinely armed, not vacuously green. This is the one claim worth re-measuring, because a previous round on this board found a check-adr-0087-registration green that was vacuous — it had passed for having nothing to look at. Run at the PR head with --base origin/main:

✓ check-adr-0087-registration: 1 declared-breaking changeset(s), each carrying an ADR-0087 disposition.
    .changeset/plain-donkeys-repeat.md  [BREAKING]  not-required (no-migration-prescription)

⇒ it names the file and the disposition, so it looked at something. ⭐ The dev also proved the green can fail — mutating the category to type-surface-only on a throwaway commit turns it RED on predicate 4, and restoring returns exit 0. A gate that cannot fail is not evidence; this one was shown to fail.

registry.ts untouched, so #17928 is unblocked. Confirmed on the file list. The #18047 generated-marker trap was never entered.

⛔ Two corrections this seat owes, both of them mine

1. My dispatch order was wrong about #18064, and the dev was right to say so. I wrote that #18064 was an open needs-user-decision card and instructed the dev to return needs_decision if the level question touched it. Read live: closed / completed at 2026-09-14T00:36:28Z, labels priority:p2 + domain:spec only — no needs-user-decision. It had closed ~3h before I wrote the claim, and I carried it from a lane listing taken at 00:14Z instead of re-reading the card. That is exactly the rule 「⛔ 写入前对每张卡现读当前状态;任何列表快照读数一律作废」, and I broke it. ⭐ The mechanism assumption was written into the dispatch as falsifiable and the dev falsified it — the protocol worked; my reading did not.

2. My own verification produced a false negative, and I nearly graded the card on it. Checking the dev's cited authority I grepped origin/main for "not the carrier of breaking-ness" and got 0, with a lit control that passed — which looked like a fabricated citation. It is not. The ADR text breaks the line mid-phrase and emphasises the word: The level is\n*not* the carrier of breaking-ness during the window. My control word shared the corpus but not the failure mode — 「同主体 = 同语料、同路径形、同包界、同引法、同失效形态:跨行短语配跨行控制词」. The citation is accurate and the authority is on main: ADR-0087, Ratified: the pre-launch launch-window exemption, Amended 2026-09-13 (#18003) — 「Pre-GA, a metadata-facing retirement or break ships minor, carrying the **BREAKING** banner and its ADR-0087 disposition entry」.

⇒ the level decision is not the dev's judgement call at all; it is what the amended ADR prescribes, and that amendment exists precisely because two director rulings came out opposite on this question (#16929 minor vs #16885 major).

Deviations, all declared and all accepted

• The commit trailer was amended to drop the model identifier — the repo's check:commit-card-trailers refuses one, which conflicts with this session's attribution reminder. Repo gate wins; that is why the final head is 25a539f5. All gate readings were re-run at the final head ✓
• skip-changeset: the Check Changeset job counts --diff-filter=A only, so a PR that corrects an existing changeset is credited zero. Correct use of the label ✓
• Both out_of_scope_findings are noted, not filed with a named successor, and neither is one of the three filing classes ✓

api_writes 3 + one push, mcp_calls 0 — inside the four-write budget.

Landing next: ready → auto-merge. ⛔ Not enqueued while a check is still running; the merge will be verified by content on origin/main, never by the merged field.

---

Generated by Claude Code