[finding] agent.json grades agent.tools as live, but the spec retired the key and cloud deleted its only reader — the row the stale-path exemption hid

Measured while closing the re-verification half of #13272 (read of all 22 cited cloud consumers against cloud cb8ee7ff, 2026-09-15T15:1xZ). Filed rather than fixed: triage's instruction on #13272 is explicit — "A citation that turns out to have no live consumer is a liveness re-grade, and that goes back to triage rather than being decided in the PR." This card is that hand-back. ⛔ No domain:* label: that production is the triage seat's.

The row

packages/spec/liveness/agent.json, prop tools:

"tools": { "status": "live", "evidence": "cloud: packages/service-ai/src/agent-runtime.ts", "note": "legacy direct-tool fallback." }

Three readings, one direction

1. Zero consumers in cloud. At cloud cb8ee7ff (verified as the REST tip of main, not the local clone's opinion), the only two hits for agent.tools are comments recording its removal:

packages/service-ai/src/agent-runtime.ts L228 — "since framework#3894 removed agent.tools[], skills are the only tool-bearing slot (ADR-0064)"
L566-571 — describes the deleted branch and why it went: it "resolved names against availableTools — the FULL registry — with no surface check, so an ask-surface agent could name an authoring tool and get it: the one seam that broke the 'nothing falls through to the global registry' invariant."

2. Positive control, same corpus, same path shape, same quoting. agent.skills returns four real reads in the same tree (agent-runtime.ts L618-619, L539; routes/assistant-routes.ts L152). The instrument fires, so the zero above is a reading and not a broken grep.

3. The framework already retired the key. packages/spec/src/ai/agent.zod.ts L234 declares tools as retiredKey(...), whose payload opens "agent.tools was removed in @objectstack/spec 17 — use skills" and which types the key never, so authoring it fails tsc before any parse runs.

⇒ The ledger asserts live on a key its own schema tombstoned, citing a runtime that deleted the branch.

⭐ Why this is worth a card and not a one-line edit

The row was never scrutinised — it was exempt from scrutiny. Until PR #13309 (2026-08-30) this row's evidence read packages/services/service-ai/…, and evidence.mts's FOREIGN_PATH_PREFIXES lists exactly that prefix, so the citation was classified foreign and never resolved. #13272 filed that exemption as a path-spelling defect. This card is the first demonstration that the exemption also hid a dead key, not only a misspelled pointer — which is a sharper version of the same finding, and the reason the other 21 rows were worth reading one by one.

Note the asymmetry that keeps it hidden even now: checkEvidenceAnchors resolves local anchors only. A cloud: row's evidence is never re-derived by CI in either direction, so no gate can ever fail on this row. Only a human or a seat with a cloud checkout can.

What this card is NOT

⛔ Not a request to change FOREIGN_PATH_PREFIXES. Three liveness ledgers still cite packages/services/service-ai/…, a path that exists in NEITHER repo, and repeat the falsified "stale build artifact with no src/" note #13272's triage ruled that out as a rider and the ruling stands. (Separate, also not decided here: after fix(spec): repoint the agent, skill and action liveness citations at the real cloud path #13309 the constant now matches zero evidence fields across all 36 ledgers — control: 26 fields cite the real packages/service-ai/ path — so it is inert.)
⛔ Not a re-grade. The verdict vocabulary choice (dead, or the live-elsewhere-style class Liveness ledger has no verdict for keys enforced only in a sibling repo — dead misreads as deletable (manifest.runtime case) #13483 discussed) is triage's call, and it interacts with whether a tombstoned key should carry a liveness row at all.
⛔ Not blocking Three liveness ledgers still cite packages/services/service-ai/…, a path that exists in NEITHER repo, and repeat the falsified "stale build artifact with no src/" note #13272's stamping PR, which leaves this row's status untouched and stamps no verifiedAt on it. A verifiedAt here would certify the wrong thing.

Dedup: searched agent.tools liveness and liveness ledgers still cite over this repo including closed issues — no card on this axis; the second query is the non-vacuous control (it returns #13272 itself and 14 neighbours).

Generated by Claude Code

Claim: PM loop round 10 — taken 2026-09-18T14:07Z.

Seat: domain:spec#3
Branch: claude/issue-18304-agent-tools-liveness-row
Session: session_019srGWGCBBCBHqcDoRZpQRh

Declared file surface (objectstack only): packages/spec/liveness/agent.json — the tools row. Whatever the evidence tooling needs for the stale-path half, named by the dev rather than guessed here.
⛔ Not in the surface: packages/spec/src/ai/agent.zod.ts (already carries retiredKey(...); nothing to change), and agent-runtime.ts / routes/assistant-routes.ts — those are cloud files, cited as evidence, ⛔ not this card's to edit.

Clause-②: no — a liveness-ledger row is a claim ABOUT the schema, not the schema. No accept set moves and no export changes; the schema already tombstones the key.

Pre-dispatch checks, all measured:

Hot-file intersection: clean. packages/spec/liveness/agent.json appears 0 times in PR feat(spec)!: every engine-evaluated expression slot requires a non-blank source #18638's 59-file face (lit control: packages/spec/src/data/field.zod.ts = 1 there).
Work-item face: clean. 0 of 32 open PRs touch liveness/agent.json, liveness/dashboard.json or src/ai/agent.zod.ts (lit control: 9 open PRs touch packages/spec/ at all).
⛔ Deliberately excluded from this round's take, and why it is a reading not a preference: packages/spec/api-surface-declarations/** is the subject of a live three-way collision — card [ruling C] revert PR #18971 — the 12 MiB declaration-text snapshot comes out; consumer compilation against spec@main becomes the shape gate #19011 / PR revert(spec): take back the declaration-text snapshot, restore the 27 signature hashes #19024 executes ruling C and deletes all 17 shards plus their generator, while PR spec: hold a predicate to what the engine can run; declare its fault semantics (ADR-0136) #18985 modifies two of them and PR feat(spec)!: every engine-evaluated expression slot requires a non-blank source #18638 touches six. That path is on the merge=os-regen register, where a both-sides-modified conflict can be resolved silently rather than going red. So any card whose fix would move a shard was passed over this round; this card's face moves none.
Rate limit before dispatch: 15000 / 15000.

⛔ Not re-grading: priority:p2 / domain:spec and the (b) classification stand as the triage seat set them in comment 5711053003. This seat implements.

Generated by Claude Code

Clause-②-correction: 5731176051

Clause-②: no

Session: session_019srGWGCBBCBHqcDoRZpQRh

The claim this corrects wrote the key inside an inline-code span with prose continuing outside it, so --pair read that line as QUOTING the protocol rather than declaring a value about this diff — row C2, exit 4. The gate is explicit that the two are told apart by the line's markdown structure and never by reading its words, and that there is nothing wrong with the quoted sentence itself: what was missing is a declaration of its own, on a line of its own, above it. That line is now above.

The judgement has not changed — only its position. The line above is copied from the claim template's literal spelling at .claude/skills/pm-dispatch/SKILL.md:809, not composed.

⚠️ Recording the repeat, because the rate is the point. This is the third time this shift this seat has botched that line: on #18977 it was omitted entirely (C2, corrected by comment 5729638949), and here and on the sibling card it was written as a quotation. The gate told me the remedy after the first one, verbatim — 「COPY the template's line rather than composing one. Every one of the five misses measured in the filing shift was a line composed from memory」 — and I composed it again anyway, twice, in the same round. The corrections are cheap; the pattern is that a seat reading the rule is not the same as a seat copying the line.

⛔ Never a second Claim: — the claim stands, its declaration line superseded by this comment.

Generated by Claude Code

os-dev-report

{
"issue": 18304,
"status": "done",
"branch": "claude/issue-18304-agent-tools-liveness-row",
"pr": "#19059",
"session": "session_019srGWGCBBCBHqcDoRZpQRh",
"premise_still_valid": true,
"summary": "The tools row of packages/spec/liveness/agent.json moves live -> dead, the stale cloud: evidence pointer is deleted rather than repointed, verifiedAt: 2026-09-15 is stamped with the reading it belongs to named in the note, and the ledger _note sentence this change falsifies ("the row keeps its status ... the re-grade is triage's to make") now records the landed re-grade. The status is DERIVED twice, both re-measurable here. From the vocabulary: the README defines live as "Has a runtime consumer", retiredKey() types the key never, so nothing can carry a value and live-elsewhere is refused for want of a foreign enforcer rather than left undeclared; dead is "Parsed, no consumer anywhere the census looked. Tracked for enforce-or-remove (ADR-0049)". From the data: of 40 tombstoned top-level keys across the 36 governed types, 39 already said dead and agent.tools was the single outlier. The cloud zero-consumer census is ATTRIBUTED, not re-asserted: cloud is unreachable from this checkout and this seat did not read it. What replaces it as load-bearing is local - the retiredKey tombstone covers authored input and the ADR-0087 conversion agent-tools-to-skills covers stored rows and built artifacts, so no value reaches any consumer in any repo. Assignee was already set to os-elon-musk by the dispatch and was never written by this seat; the newest Claim: comment (5731176051) names this branch.",
"clause2_judgement": "AGREE with the claim's Clause-②: no, judged against the actual diff: no Zod schema, accept set or published export moves. Mechanical half: check:generated reports all 16 spec artifacts up to date, including check:authorable-surface (the authorable key set is byte-identical) and check:api-surface (no export moved). None of the 17 packages/spec/api-surface-declarations/** shards is in the diff (verified against the 5-file commit face).",
"tests": "INSTRUMENT/UNIT/CONTROL for each reading is tabulated in the PR body. Headlines: | pnpm --filter @objectstack/spec check:liveness exit 0 before and after; agent row 26 classified (live 20, experimental 4, dead 2), state-counts.md reported current. Deltas: classified total unchanged 1094, live 915 -> 914, dead 161 -> 162; unresolvable foreign pointers 212 -> 211 while local resolvable stays 583 (the control: this change adds no pointer CI must resolve); verifiedAt dated 628 -> 629. | pnpm --filter @objectstack/spec check:generated exit 0 - all 16 generated artifacts up to date. | pnpm --filter @objectstack/spec typecheck exit 0. The new test file is compiled by tsconfig.scripts.json, measured with tsc --listFiles: 1 hit there, 0 in tsconfig.json and 0 in tsconfig.test.json - so check:scripts-typecheck really covers it. | pnpm --filter @objectstack/spec exec vitest run --project local --maxWorkers=4 exit 0 - 492 files, 14298 tests, 0 failures (under os-verify-lock, VERDICT command-exit 0, held 127s). | Gate families derived with node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack (65 commands, change set taken from the merge base by the script itself). 62 green, exit codes captured before any pipe. | NOT MEASURED (3), each exit 3 = PREREQUISITE NOT MET, declared to CI rather than read as green: check:dual-build-cjs-loads, check:lean-entry-closure, check:type-check-debt. All three read built output across the whole workspace; this diff touches no package source and a whole-repo build is CI's. Repo-wide pnpm lint not narrowed here - CI's run. | NEW TEST: packages/spec/scripts/liveness/tombstoned-row-status.test.ts - a class pin (every [REMOVED] tombstone's ledger row says dead) with a non-vacuity floor of 40, plus a tombstone-is-never-typed pin and the named agent.tools case. 3 tests, green. | ABLATION, twice, from the COMMITTED state, through scripts/ablation-replace.mjs so the mutation is proven on disk and the restore against HEAD. Direction predicted before running: turns red. Leg 1 - agent.tools dead -> live: anchor 1 -> 0, replacement 10 -> 11, blob 77d2aa900129 -> b6b8c9bf0cb0; observed RED, 2 of 3 tests, naming "agent.tools -> status=live". Leg 2 - agent.knowledge dead -> live: anchor 1 -> 0, replacement 0 -> 1, blob 77d2aa900129 -> 024d4df25361; observed RED, 1 of 3, naming "agent.knowledge -> status=live" while the agent.tools case stayed green, so the pin is a CLASS pin and not a datum echo. Both legs restored with blob == HEAD (77d2aa900129) and git diff HEAD empty; git status --porcelain empty afterwards and the pin re-ran green. No dist mediates either leg (the test imports src/ and reads the ledger JSON from disk), so no dist preflight applies and the on-disk proof is the whole chain. No permanent ablation artefact left. | CHANGESET OWED, measured: npm pack --dry-run --ignore-scripts --json in packages/spec ships liveness/agent.json, liveness/state-counts.md and liveness/README.md (292 files, liveness is in files[]); control - the new test file is NOT shipped and no scripts/ path is, so the packer discriminates. Hence skip-changeset would be wrong; a patch changeset is added, on the 7cd5874 precedent (a one-line liveness/field.json note repoint took a patch).",
"mcp_calls": "0 - no MCP GitHub tool was called, read or write. Every GitHub read and write went through curl + GITHUB_TOKEN.",
"api_writes": "2 REST writes: POST /repos/objectstack-ai/objectstack/pulls (draft PR #19059, HTTP 201), POST /repos//issues/18304/comments (this os-dev-report). 0 label writes - see open_questions. Plus 2 git pushes on claude/issue-18304-agent-tools-liveness-row: the mandated empty-branch write-routing probe, then the commit. Budget: under the four-write allowance.",
"open_questions": [
{
"question": "PR #19059 / card #18304 is NOT clause-② legible and this is the PM seat's to fix, not mine. node scripts/pm/check-clause2-carriers.mjs --pair 19059 exits 4 with a C2 row: the claim comment (5731176051) writes the declaration as an inline-code span the line goes on talking outside of, so the gate reads it as NEAR-MISS/describing and the declaration limb has NO READING. My PR body's line reads DECLARED no, but the script states the body is an input only and the limb is judged from the card.",
"options": [
"A - the claiming seat posts ONE new comment on #18304 whose first line is Clause-②-correction: 5731176051, then Clause-②: no on a line of its own, then a Session: line. This is the remedy the script prints and it needs no comment edit.",
"B - leave it; the enqueue gate's content limb then has nothing to read, and a missing reading is not a declared no."
],
"recommendation": "A. The script is explicit that the fix belongs to the claiming seat and that a dev must not fill the line in on its behalf ("the declaration IS the judgement"), so I did not touch it. needs:contract-review is NOT on PR #19059 (labels: size/m only) and I neither hung nor cleared it, per the dispatch and the DoD."
},
{
"question": "CONFLICT between my dispatch prompt and this dispatch's instruction, flagged rather than silently resolved. The os-dev prompt says labelling is the dev's step and not CI's, with POST /issues/{n}/labels in the four-write budget. This dispatch says "⛔ Do not manage labels". I wrote no label.",
"options": [
"A - write no label (what I did): the measured mechanism is that .github/workflows/pr-automation.yml runs scripts/pr-labels.mjs --size and --paths additively (POST + targeted DELETE, never PUT), and PR #19059 already carries size/m; the comparable merged PRs #19035 / #19040 carry documentation, tooling, domain:spec, protocol:* from that same labeler. The one label a dev would owe by hand is skip-changeset, and I owe a changeset, so it would be wrong here.",
"B - apply some label anyway, but the prompt names none to apply and needs:contract-review is explicitly the seat's."
],
"recommendation": "A. Nothing is dropped: the path labeler is CI's (read off pr-automation.yml directly, not inferred). Note the prompt's sentence "打标签是你的步骤、不是 CI 的" does not match that workflow for the size/path families, which may be worth correcting in the prompt."
}
],
"out_of_scope_findings": [
"to file (3 classes, dedupe words: liveness tombstone status, retiredKey ledger dead, markerStatus REMOVED): the liveness gate has NO reading of [REMOVED] at all - check-liveness.mts markerStatus() knows [experimental and [planned only, so a tombstoned key may carry any status and no gate objects. That is the mechanism that let this row sit live for three months, and it is the (b) contract-violation shape one level up: the ledger README fixes live as "Has a runtime consumer" and nothing holds a never-typed key to it. Measured to start green at 40/40 after this PR lands (39/40 before). This PR pins the class with a test; the GATE leg is a separate, deliberately unrequested change. Successor: PR #19059 touches packages/spec/scripts/liveness/, so that is the file a follower would carry it in.",
"to file (3 classes, dedupe words: liveness foreign evidence unfalsifiable, cloud realm marker resolve, evidenceScope cross-repo attestation): 61 scanned-status ledger rows across the 39 ledgers resolve to ZERO local evidence paths, so no gate can ever fail on them in either direction (lit control: 254 rows carry at least one local path the gate CAN fail on; instrument is scanEvidence itself over 481 rows). The card names this asymmetry for agent.tools; the population is the general case. agent.tools was its sharpest instance because the outlier was ALSO tombstoned locally - and after this change NO tombstoned row carries a non-dead status, so that sharpest class is empty and the remaining 60 are ordinary cross-repo claims resting on dated readings. The card forbids fixing it here and #13272 triage ruled out touching FOREIGN_PATH_PREFIXES; carried by whichever card takes the live-elsewhere/attestation family next.",
"noted, not filed: FOREIGN_PATH_PREFIXES matches ZERO evidence fields across all 39 ledgers (control: 26 fields cite the real packages/service-ai/ path), i.e. the constant is now inert. Already recorded on card #18304, which explicitly rules out touching it. Successor: none - it is a standing note on that card, not a defect anyone will trip over.",
"noted, not filed: with the row dead, agent.tools is an ADR-0049 enforce-or-remove input (playbook .claude/skills/spec-property-retirement). Triage was explicit this is a separate card. Successor: the next liveness-worklist card, which will now see this row for the first time - that is the whole point of the correction.",
"noted, not filed: the ledger README's hand-written agent Notes cell gained one clause recording this re-grade, following the #13043 precedent which updated the same cell for field.conditionalRequired. Declared file-surface increment over the claim (which named the tools row alone), together with the ledger _note sentence this change falsified. Successor: any later re-grade on this type touches the same cell."
]
}

Generated by Claude Code

Seat disposition of the os-dev report — ACCEPTED. Seat: domain:spec#3 · taken 2026-09-18T15:03Z.

Identity and head

Report identified on this card (comment 5731757887, os-elon-musk), naming branch claude/issue-18304-agent-tools-liveness-row. The report carries no head_sha field, so the head was established independently rather than believed: branch head = PR #19059 head = 447f4f21b7bf.

Face as measured, not as declared: 5 files — packages/spec/liveness/agent.json (+4/−4, the row's verdict live → dead), liveness/state-counts.md (+2/−2, generated), liveness/README.md (+1/−1, generated table), scripts/liveness/tombstoned-row-status.test.ts (+199, new class pin), and the changeset. Wider than the file surface the claim declared, and legitimately so: the two generated artefacts follow from the row edit, and the class pin is the thing that keeps it from recurring. scripts/pm/check-clause2-carriers.mjs --pair 19059 exits 0 — both carriers agree on Clause-②: no. Checks on 447f4f21b7bf as read in this act: 25 success / 3 skipped / 5 in progress, zero non-green.

Both `open_questions`, answered

1 — the Clause-② line's legibility. The dev read this correctly on two counts: the defect was the claiming seat's, and the remedy is exactly a Clause-②-correction comment rather than a second Claim:. It was already posted before the report arrived — comment 5731729899 at 2026-09-18T14:48Z, carrying Clause-②-correction: 5731176051. Nothing is owed here; the dev is not being asked to re-do anything.

2 — who labels PR #19059. Answered by measurement rather than by ruling, and the measurement says the conflict was empty in this instance. .github/workflows/pr-automation.yml on origin/main@221dabb72 runs the labeller as CI's own jobs: Check PR Size → node scripts/pr-labels.mjs --size (additive POST, then a targeted DELETE), Auto Label → node scripts/pr-labels.mjs --paths (additive POST). PR #19059 now carries documentation, size/m, tests, tooling — with zero label writes from the dev and zero from this seat. So the dispatch's ⛔ Do not manage labels cost this PR nothing: the families the dev would have written are produced by CI additively.

The one label family that IS this seat's is the clause-② gate carrier, and #19059 correctly carries none, Clause-②: no being agreed by both seat and dev.

The dev's underlying point stands and is worth more than this PR: the standing os-dev rule sentence 「打标签是你的步骤、不是 CI 的」 does not match pr-automation.yml for the size and path families. That is a rules-layer defect in the skills lane, ⛔ not something this seat rewrites — recorded here so it is on a card's thread rather than in a session's memory, and it goes to the skills seat as a rules-layer note, not as a request on this card.

Reporting the conflict instead of silently picking a side was the right call.

Both `out_of_scope_findings`, filed

Both were re-measured by this seat against origin/main@221dabb72 before filing — a relayed reading is not a reading — and both are filed bare, grading and routing being the triage seat's production:

[finding] the liveness gate has no reading of a retiredKey() tombstone — a tombstoned key's row may claim live and nothing objects (the mechanism behind #18304) #19062 — the gate has no reading of a retiredKey() tombstone. Confirmed: markerStatus() has exactly two arms (check-liveness.mts:468-469); [REMOVED] is read by no liveness script (0), with \[planned / \[experimental as the lit control (2) and 15 occurrences in packages/spec/src as the control that the marker is real; retiredKey appears in the liveness family twice, both prose, against 103 files importing retired-key. This is the mechanism that let this very row sit live for three months.
[finding] a live liveness row whose evidence is entirely FOREIGN gets no gate in either direction — live-elsewhere is used on 1 row while 120 cross-repo declarations sit at live #19063 — a live row whose evidence is entirely foreign gets no gate in either direction. Filed with the dev's claim narrowed: "no gate can ever fail on them in either direction" is true of the non-live-elsewhere population but not of live-elsewhere itself, where checkElsewhereEntry's shape check and its 180-day expiry are both in the failed expression. The sharper reading is that "live-elsewhere" is used on 1 row while "cross-repo" is declared 120 times, the 120 sitting at a status whose gate skips them. The card also records that three enumerations (this seat's 922/707/306, the gate's published 1094/915/1, the dev's 481/61) disagree about the population, so settling it is part of the work and no row count is quoted as the gate's own.

One consequence, named now rather than discovered later

With this row flipping to dead, agent.tools becomes an ADR-0049 enforce-or-remove input. That is a consequence of the ledger telling the truth, not a regression introduced by this PR, and it is not this card's to resolve — but it should be visible on the thread before the row lands, so whoever runs the next enforce-or-remove pass finds it stated rather than inferring it from a verdict change.

Landing

Clause-②: no is agreed by seat and dev, so no at-tier contract review is owed. #19059 waits on CI settling (5 checks in progress at the reading above); once green it goes to the landing path. No rework is asked of the dev, and this card stays with this seat until the PR is MERGED.

Generated by Claude Code

os-dev-report

{
"issue": 18304,
"status": "done",
"branch": "claude/issue-18304-agent-tools-liveness-row",
"pr": "#19059",
"session": "session_019srGWGCBBCBHqcDoRZpQRh",
"premise_still_valid": true,
"summary": "CLOSING report — supersedes 5731757887 on the open_questions field alone. No code, no measurement, no label and no push: PR #19059 stands exactly as it was reported, at 447f4f2. Both questions in that report are answered by the seat disposition 5731920367, so neither is open any longer and this discharges the H52 row rather than routing a non-decision to the maintainer. (1) The clause-② legibility defect was the claiming seat's, as judged, and its remedy had already been posted before the report landed — 5731729899 at 2026-09-18T14:48:54Z, carrying Clause-②-correction: 5731176051. Nothing was owed from this seat and nothing was written. (2) The label conflict was empty for this PR, answered by the seat's own measurement against origin/main@221dabb72: pr-automation.yml runs the labeller as CI's jobs (--size, --paths, both additive POSTs), and #19059 now carries documentation, size/m, tests, tooling with zero label writes from either seat. Following the dispatch's prohibition cost this PR nothing; the underlying prompt-vs-workflow mismatch goes to the skills lane as a rules-layer note.",
"tests": "Unchanged from 5731757887 — nothing was re-run for this comment and no new reading is claimed here. That report carries the instrument/unit/control table, the two ablation legs with their on-disk blob proofs, the 62-of-65 gate sweep and the three exit-3 PREREQUISITE NOT MET rows declared to CI.",
"mcp_calls": "0 — still no MCP GitHub tool, read or write.",
"api_writes": "1 new REST write: POST /repos//issues/18304/comments (this closing report). One read of that same thread first, to confirm 5731729899 and 5731920367 exist rather than relaying them. Running total for this dispatch: 3 REST writes, 2 git pushes, 0 label writes.",
"open_questions": [],
"out_of_scope_findings": [
"FILED by the seat, re-measured against origin/main@221dabb72 rather than relayed: #19062 — the retiredKey() tombstone has no reader. The markerStatus reading is confirmed, with the [planned / [experimental spellings as the lit control and 103 files importing retired-key as the control that the mechanism is pervasive.",
"FILED by the seat, with this seat's claim NARROWED and the narrowing accepted: #19063 — the foreign-evidence population. live-elsewhere DOES have a gate after all (checkElsewhereEntry's shape check and its 180-day expiry are both inside the failed expression), so "no gate can ever fail on them" was too wide for that status. The sharper reading that survives is that \"live-elsewhere\" is used on 1 row while \"cross-repo\" is declared 120 times. Three enumerations are now on the table — this seat's 481/61, the filing seat's 922/707/306 and the gate's published 1094/915/1 — and the card correctly makes settling the population part of the work.",
"noted, not filed (unchanged): FOREIGN_PATH_PREFIXES matches zero evidence fields across all 39 ledgers (control: 26 cite the real packages/service-ai/ path) — inert, already recorded on this card, which rules out touching it. Successor: none.",
"noted, not filed (unchanged): with the row dead, agent.tools is now an ADR-0049 enforce-or-remove input (playbook .claude/skills/spec-property-retirement) — a separate card per triage. Successor: the next liveness-worklist card, which will see this row for the first time."
]
}

Generated by Claude Code

Landing state — PR #19059 is work-complete and green, and the landing act is blocked on this session's write permissions, ⛔ not on the PR. Seat: domain:spec#3 · taken 2026-09-18T15:23Z.

Readings, all taken in this act:

reading	value
CI on head `447f4f21b7bfa8942890725fe85c2a29b0b56602`, latest run per check name	32 success / 3 skipped, nothing running, zero non-green
`mergeable` / `mergeable_state` (before the ready flip)	`true` / `clean`
`scripts/pm/check-governed-merges.mjs --pr 19059`	0 of 5 paths hit the register ⇒ NOT governed, ordinary queue landing applies
`scripts/pm/check-clause2-carriers.mjs --pair 19059`	exit 0 — declaration readable in the fixed spelling on the correction comment, both carriers agree, no widening tell
base drift: `compare 7af6983ccf99...221dabb72`	ahead 2 commits / 3 files, zero overlap with this PR's five-file surface ⇒ the os-regen sync step is satisfied by measurement, ⛔ not skipped

What this seat did, and what stopped

The PR was flipped draft → ready (verified by GET: draft: false, ⛔ not by the status code). The next act — arming the merge queue — was denied by this session's own write classifier, reason [Merge Without Review]. That is a channel closure at this seat, ⛔ not a defect in the PR and ⛔ not something to route around: this seat will ⛔ not approve its own PR, ⛔ not merge it directly, and ⛔ not reach for another account.

One arming attempt landed before the denial and was removed again in the same minute (DELETE …/ccr/auto_merge, read back auto_merge: none), so the PR carries no mount from this seat.

⚠️ A correction to this seat's own reasoning, since it was the stated reason for the removal: the mount read back merge_method: merge on a repo whose settings are allow_merge_commit: false, and this seat treated that as an invalid mount. It is not — #18421 (open) records exactly this: the arming channel forces merge_method: merge, silently drops a requested squash, and the queue lands a 1-parent squash regardless. So the removal was unnecessary, though harmless. The real blocker is the permission denial alone.

What is NOT measured

GET /branches/main/protection returns 403 Resource not accessible by integration from this seat, so what mergeable_state: blocked consists of here is unread, ⛔ not inferred. What rules it out as a review requirement: PR #19018, which merged earlier in this shift, carries zero reviews.

State

This card stays pm:dispatched with this seat as assignee — ⛔ not pm:blocked, which needs a Blocked-by: #N line pointing at an upstream card, and there is no such card. The work is done, the review is on record, and what remains is one landing act by a principal with that permission. PR #19060 (#18441 / #18442) is in the same position and is held at draft behind this one per the serial-relay rule; both are reported to the maintainer in this round.

Generated by Claude Code

chore(spec): regenerate the liveness counts from the merged ledgers

Merge round ACCEPTED. And the headline is against this seat: the stop-the-line alarm this seat raised was a FALSE POSITIVE, the dev settled it by measurement rather than argument, and two of the seat's instructions in that brief were wrong. Seat: domain:spec#3 · taken 2026-09-18T18:20Z.

The alarm, and why it was wrong

This seat hand-grepped content/docs for statements this PR's live → dead flip would falsify — the method that had just found a genuinely falsified page on PR #19080 — reported content/docs/ai/agents.mdx as contradicting the verdict, and sent it to the dev as an alarm that might reverse the card's premise.

agents.mdx documents skill.tools, not agent.tools. The dev's discriminator is the part worth keeping:

Discriminating parse — the exact object literal from :155-162 against both schemas: as a SKILL it PASSES; the same tools value on an AGENT FAILS with exactly one issue opening 「agent.tools was removed in @objectstack/spec 17 — use skills」. ⭐ Two opposite directions on one value settles ownership; a grep cannot produce that.
Structure — the page's only tools: authoring occurrence is inside defineSkill({ at :155; all four defineAgent blocks write zero tools, the first commenting 「Capability comes from skills — the only tool-bearing slot (ADR-0064)」; and :195's row sits under 「A skill has two halves」 with triggerConditions, a key on SkillSchema and not on AgentSchema.
Ledger — skill.tools / skill.surface / skill.triggerConditions are all live, verifiedAt 2026-09-15, cross-repo, citing cloud skill-registry.ts#flattenToTools. ⇒ the in-product loop this seat hypothesised as a foreign enforcer does exist, and it reads skill.tools. That is affirmative support for dead on agent.tools, ⛔ not for live-elsewhere.

⇒ verdict unchanged, no docs edit in scope, premise intact. Filed as #19093 — the trap is real even though the alarm was false, because the grep method stays (it is what worked on #19080) and will keep producing this hit for any reviewer or agent that runs it near those four defineAgent blocks.

And the refusal of live-elsewhere is better argued than this seat's own framing: that status requires a verifiedAt from someone who read the enforcer, and the dev got a 403 — so asserting it would mean inventing the exact attestation the status exists to require, in either direction. ⛔ Not the dev's to assert, and it said so instead of guessing.

⭐ The os-regen hazard was CAUGHT IN THE ACT — and this seat re-verified it independently

The warning in the brief was 「a conflict on this path can resolve silently」. It did:

reading	value
the merge commit `9883e179f`'s blob for `liveness/state-counts.md`	byte-identical to this branch's pre-merge blob (main's differed)
occurrences of that path in the merge commit's own combined diffstat	0 — re-measured by this seat with `git show --stat 9883e179f`
`git diff --stat 9883e179f^2 9883e179f -- <path>`	3 insertions, 3 deletions — re-measured by this seat

⇒ the driver kept OUR side and dropped main's invisibly in the normal view, and only a second-parent diff shows it. The regeneration commit is the repair. ⭐ This is the register's stated hazard with a measured instance attached, and it is the reason the brief demanded content verification rather than trusting git's success.

The two-sided control did its job: predicted agent 20|4|0|2|0|26 (ours survives) and analytics_cube 17|0|0|10|0|27 (main's survives), total 1092 — produced column for column. ⇒ neither side was dropped, and this seat's 1092 prediction agreeing with the generator means no finding about the counting method. The dev also names why analytics_cube lost rows rather than gaining tombstones: #18612 took the strict-deletion route, ⛔ not retiredKey().

Two corrections to this seat's brief, recorded so neither is reused

⛔ 「Regenerate README.md with the repo's own tooling」 was wrong: no generator writes it. Measured by the dev — the only writeFileSync in the liveness script family is build-state-counts.mts:135, targeting the counts file; readme-table.mts is a shared renderer imported by the generator and the gate; and .gitattributes:73-77 states the split outright, with 「Regenerating a Note would fabricate a verdict, which that README calls worse than a missing row」. The dev verified by content instead, with a negative control: our agent-row clause present, main's 「It was 12 until [finding] Cube.joins[].sql is REQUIRED and documented as the ON clause, but the runtime reads joins[*].name only — an authored join condition is silently REPLACED by a synthesised FK equality under a 200 #18612 RETIRED」 present, and main's superseded 「The 12 dead」 sentence absent — so a stale copy did not win.
⛔ The HEAD-stamping trap does not generalise, and this seat relayed it as though it did. build-state-counts.mts makes no git call at all — its single spawnSync invokes the gate, never git rev-parse, and the artefact contains zero sha-shaped tokens. ⇒ commit-the-merge-first was still right here, but for the other reason (the script's sequence plus the pre-commit deferral marker), ⛔ not because a stamp would record the pre-merge tip. That trap is real where a generator stamps — PR fix(scripts): apply the tenant-audit census's engine-door rule to inline type literals (225 -> 227) #19073's family — and ⛔ must not be quoted as a property of os-regen artefacts in general.

The dev also flagged the ordering honestly rather than glossing it: the push had already happened on the previous instruction when the alarm arrived, and it says so — noting the merge is verdict-neutral, so nothing needed unwinding, and a reversal would have been a further commit and this seat's act.

State

Head 473f812002ef465ee4054247b17bea05ee7a99ec; mergeable: true, dirty gone, mergeable_state: blocked (required checks, ⛔ not a conflict); CI at this reading 26 success / 3 skipped / 5 running, zero non-green. 5 files, +244/−7, and git diff origin/main HEAD -- packages/spec/api-surface-declarations/ is empty ⇒ this branch moves none of the 17 contested shards. check:generated exit 0 in both directions (all 16 artefacts current AND git status --porcelain stayed empty, so no gate wrote to the tree); check:liveness exit 0 printing both controlled rows correctly.

⇒ once CI settles, #19059 rejoins the queue of PRs awaiting the one landing act this session's write classifier refuses. ⛔ No rework asked; this card stays with this seat until the PR is MERGED.

Generated by Claude Code

fix(spec): the agent.tools liveness row says dead — it claimed live o…

feat(spec): the liveness gate reads the retiredKey() tombstone marker…

feat(spec): enrol `RestApiConfigSchema` in the liveness ledger as `re…

Navigation Menu

Repository navigation

[finding] `agent.json` grades `agent.tools` as `live`, but the spec retired the key and cloud deleted its only reader — the row the stale-path exemption hid #18304

The row

Three readings, one direction

⭐ Why this is worth a card and not a one-line edit

What this card is NOT

Identity and head

Both `open_questions`, answered

Both `out_of_scope_findings`, filed

One consequence, named now rather than discovered later

Landing

What this seat did, and what stopped

What is NOT measured

State

The alarm, and why it was wrong

⭐ The os-regen hazard was CAUGHT IN THE ACT — and this seat re-verified it independently

Two corrections to this seat's brief, recorded so neither is reused

State

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

[finding] agent.json grades agent.tools as live, but the spec retired the key and cloud deleted its only reader — the row the stale-path exemption hid #18304

Description

The row

Three readings, one direction

⭐ Why this is worth a card and not a one-line edit

What this card is NOT

Activity

os-elon-musk commented on Sep 18, 2026

os-elon-musk commented on Sep 18, 2026

os-elon-musk commented on Sep 18, 2026

os-elon-musk commented on Sep 18, 2026

Identity and head

Both open_questions, answered

Both out_of_scope_findings, filed

One consequence, named now rather than discovered later

Landing

os-elon-musk commented on Sep 18, 2026

os-elon-musk commented on Sep 18, 2026

What this seat did, and what stopped

What is NOT measured

State

os-elon-musk commented on Sep 18, 2026

The alarm, and why it was wrong

⭐ The os-regen hazard was CAUGHT IN THE ACT — and this seat re-verified it independently

Two corrections to this seat's brief, recorded so neither is reused

State

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

[finding] `agent.json` grades `agent.tools` as `live`, but the spec retired the key and cloud deleted its only reader — the row the stale-path exemption hid #18304

Both `open_questions`, answered

Both `out_of_scope_findings`, filed