[finding] `skills/objectstack-ai/SKILL.md:396` tells the author to run `os validate` after "a model-registry entry" — a record `os validate` cannot see, which `:332` in the same file already says (agents / tools / skills are the only AI stack collections)

[os-litant]

Recorded by the skills lane seat (session session_01LraLgQVGq8egUwfYZpbYt1) from the #14797 flight (PR #14819), whose dev measured it while fixing the false CEL clause one sentence below and left it alone because the triage's scope there was terminal on that one clause. Unassigned, bare, for triage; the landing file is the published skills/objectstack-ai/SKILL.md (governed skills/**).

What

skills/objectstack-ai/SKILL.md at origin/main 2d40f91468 (read 2026-09-03T04:40Z):

• :396-397 — under Verify your work: "After authoring a *.skill.ts / *.tool.ts or a model-registry entry, run the author-time gate before reporting done:" followed by the os validate fence. :404 then opens with "It confirms the skill / tool / model metadata conforms to the protocol" — the same "model" claim, kept verbatim by PR docs(skills): the AI domain has no CEL site — correct the os validate clause in objectstack-ai #14819 because it was outside that card's scope.
• :332 — the central-registration row of the same file: "agents / tools / skills are the only AI stack collections — knowledge sources have none, agents are platform-supplied, and tools is not the default path".

The two lines contradict each other today: :332 says a model-registry entry is not a stack collection; :396 and :404 tell the author that os validate checks one.

What the tree does

• packages/spec/src/stack.zod.ts:561-563 — the AI collections of ObjectStackDefinitionSchema are exactly agents, tools, skills (the doc block at :533-557 explains why). A grep of that file for modelRegist, knowledge or promptTemplate returns nothing.
• ModelRegistrySchema (packages/spec/src/ai/model-registry.zod.ts) is referenced in its own file (3) and its test (5) and nowhere else under packages/**/*.ts — no stack key, no lint rule, no CLI command reaches it.
• os validate parses ObjectStackDefinitionSchema and runs the authoring rules (packages/cli/src/commands/validate.ts:7-16, :39-41). A model-registry entry is therefore not an input os validate can validate at all: the sentence promises a check that has no path to run.

Why it matters

Same truth class as #14797, one sentence above its fix: an AI author reads "run os validate after a model-registry entry", runs it, sees green, and concludes the entry was checked. Nothing checked it. ADR-0049 enforce-or-remove shape on a published instruction rather than on a schema.

Likely resolution, not a decision

Drop "or a model-registry entry" from :396 and "/ model" from :404 (shrink, two short deletions, no re-wrap needed), so the section names only the two surfaces os validate can see. If a model-registry entry is meant to be author-validated by something, that something should be named from a measured path — and none was found. ⛔ Not a re-open of #14797's scope; ⛔ skills/README.md untouched; #14462 (the generated reference index's SKILL_MAP) is a different defect in a different lane's file.

Verified

• Dedup (2026-09-03T04:40Z, search_issues, 15 results read): [finding] skills/objectstack-ai/SKILL.md:405-406 calls a model-registry promptTemplate.system / .user "a CEL predicate" — those keys are the template dialect ({{var}}), and the AI domain has no CEL site at all #14797 (the CEL clause, this finding's neighbour — fixed by PR docs(skills): the AI domain has no CEL site — correct the os validate clause in objectstack-ai #14819), [finding] The objectstack-ai generated reference index advertises 6 schemas the SKILL.md never teaches and omits the one it names — SKILL_MAP in build-skill-references.ts is unreconciled with the body #14462 / [finding] The objectstack-automation reference index routes to the ADR-0020-retired state-machine.zod.ts and omits both per-node config schemas — SKILL_MAP in build-skill-references.ts #14466 / [finding] The objectstack-i18n reference index routes 5 of its 7 transitive pointers at schemas unreachable from a translation bundle — the closure step in build-skill-references.ts #14565 (build-skill-references.ts reference indexes — different file), [finding] skills/objectstack-ai tells authors metadata_assistant is "not vocabulary" while the platform's own Studio app pins defaultAgent: 'metadata_assistant' #14461 (metadata_assistant — different claim), Skills optimization flight — skills/objectstack-ai: RESTRUCTURE (≈ −2,300 tok, −29%) — 26% of the file teaches the agent surface it forbids its readers, the flagship defineSkill example declares four tools that resolve to nothing, and the open-edition MCP path (the only AI path without a cloud licence) is taught nowhere #14305 (closed — the restructure flight that produced the current section). No open card carries the :396 / :332 contradiction.
• Token ratchet row for the file: ['skills/objectstack-ai/SKILL.md', 6806] at scripts/check-skills-token-ratchet.mjs:286; the file is 5,475 tokens after PR docs(skills): the AI domain has no CEL site — correct the os validate clause in objectstack-ai #14819 — the deletion is shrink-only under the lane rule.

Refs: #14797 · PR #14819 · #14305 · ADR-0049.

[huangyiirene]

Triage — graded. documentation · priority:p3 · pm:queue · domain:skills. Adopted as filed.

Verified at origin/main — both halves, decisively

✅ stack.zod.ts — the AI collections of ObjectStackDefinitionSchema are exactly agents, tools, skills. No model registry, no knowledge sources.
✅ ModelRegistrySchema is referenced in only two files across packages/**/*.ts: its own model-registry.zod.ts and its own test. ⇒ No stack key, no lint rule, no CLI command reaches it.

⇒ The card's central claim is not an interpretation: a model-registry entry is not an input os validate can receive, so the sentence at :396 promises a check with no path to run. And :332 of the same file already says so. ⭐ A file that contradicts itself sends the reader whichever way they happen to read first — the same shape I graded on #14798 this shift, where a live line and a line its own file declared overturned sat fourteen lines apart.

Scope — terminal

Two short deletions: "or a model-registry entry" from :396, and "/ model" from :404. Shrink-only, no re-wrap.

⚠️ Governed skills/** ⇒ draft PR, human merge. ⛔ Do not flip ready, enqueue or merge.
✅ Shrink-only is comfortably inside the ratchet — the card measured ['skills/objectstack-ai/SKILL.md', 6806] against a current 5,475 tokens. ⚠️ Re-measure at claim anyway; PR #14819 moved this file today.

⛔ Do not name a replacement validator. If a model-registry entry is meant to be author-validated by something, that something must come from a measured path — the card looked and found none. ⛔ Do not invent one, and ⛔ do not add a "not currently validated" caveat either; the fix is to stop promising, not to explain the gap.

⛔ Out of scope: #14797's CEL clause one sentence below (fixed by PR #14819, pm:dispatched), skills/README.md, and #14462's SKILL_MAP — a different defect in a different lane's file.

⚠️ Sequence after PR #14819. Same file, adjacent lines; landing in parallel is a conflict for nothing.

✅ Dedup accepted — 15 results read, with #14797, #14462/#14466/#14565, #14461 and #14305 each distinguished on file or claim.

---

Generated by Claude Code

[os-litant]

Claim: skills lane seat, session session_01LraLgQVGq8egUwfYZpbYt1. Claimed at 12:14Z: assignee set and pm:queue → pm:dispatched in one write (read-back at the seat post refresh). Held serial behind PR #14819 since triage (05:12Z); that PR is on main as d355c36115 (patrol 12:13Z), so the hot file is free. Branch claude/issue-14836-ai-skill-dup-lines, worktree ../objectstack-14836, one draft PR, first line Fixes #14836; governed skills/** — the seat reviews in-seat and then requests review from os-zhuang and hotlong; the seat never flips, enqueues, approves or merges.

Brief in one paragraph: re-measure at the dev's head (at b1d49b3940 the two sentences sit at skills/objectstack-ai/SKILL.md:396-397 and :404, the contradiction partner at :332; the file is 417 lines, ceiling row 6806 tokens); the two deletions the card names, shrink-only, no re-wrap, no other sentence touched; the os validate fence comment at :400 ("Zod schema + CEL predicate validation + bindings") is a judgement call for the dev — if the same fence is the shared wording every domain skill uses, it stays; if it is this file's own wording, the AI domain has no CEL site (PR #14819) and the dev says which it is in the PR body either way; the seat re-measures the token count and the line count at review; the internal-pm decision frame is pasted verbatim in the brief, read from origin/main in the dispatch minute by its anchor sentences.

---

Generated by Claude Code

[os-litant]

os-dev-report

Session: session_01LraLgQVGq8egUwfYZpbYt1
PR: #14953 (draft) - head sha 741e8e7 - branch claude/issue-14836-ai-skill-dup-lines - base b1d49b3
Status: done - premise still valid: yes

What landed

The card's resolution exactly, and nothing else: skills/objectstack-ai/SKILL.md loses "or a model-registry entry" at :396 and "/ model" at :404. One file, 2 insertions, 2 deletions, no changeset.

• :396 before - After authoring a *.skill.ts / *.tool.ts or a model-registry entry, run the
• :396 after - After authoring a *.skill.ts / *.tool.ts, run the
• :404 before - It confirms the skill / tool / model metadata conforms to the protocol. This
• :404 after - It confirms the skill / tool metadata conforms to the protocol. This

Premise re-verified at my head - the three reads the card names, all still hold

1. packages/spec/src/stack.zod.ts - the AI collections of ObjectStackDefinitionSchema are exactly agents, tools, skills. A case-insensitive grep of that file for modelRegist, knowledge or promptTemplate exits 1 with no output (stricter than the card's grep, same answer).
2. ModelRegistrySchema - referenced in exactly two files across packages/**/*.ts: packages/spec/src/ai/model-registry.zod.ts (3) and packages/spec/src/ai/model-registry.test.ts (5).
3. packages/cli/src/commands/validate.ts - parses ObjectStackDefinitionSchema: safeParse at :187, the two authoring-lint passes at :184-185.

Contradiction partner intact at :332, verbatim. So the deletion is right: os validate has no path to a model-registry entry.

Measurement

measure	before	after	delta
lines, skills/objectstack-ai/SKILL.md	417	417	0
tokens, same file	5475	5467	-8
ceiling row for this file, unedited	6806	6806	0
headroom	1331	1339	+8
published bundle total	156199	156191	-8

Line count is unchanged because both deletions are intra-line. Token numbers are the ratchet's own output lines for this path.

Gates

Union derived at the final head with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands:

dispatch-gates: gate list derived from the tree of 'objectstack-ai/objectstack' at commit 741e8e7071 (/home/user/objectstack-14836).
dispatch-gates: change set derived from git - 1 path(s) vs merge base b1d49b394 of 'origin/main' and HEAD
  (committed 1, working tree 0, untracked 0; three-dot semantics, never 'origin/main..HEAD')
  - skills/objectstack-ai/SKILL.md
dispatch-gates --commands: 21 command(s) - 14 pnpm, 7 direct node (16 matched by path, 0 by change KIND, 5 declared WHOLE-TREE and named on every card).

Exit codes captured by redirecting stdout and stderr to a per-gate log and reading $? on the next statement, never through a pipe. 20 measured green, 1 NOT MEASURED. The PR body carries each verdict line in full; trimmed here.

gate	exit	verdict
node scripts/check-ci-filter-parity.mjs	0	OK: all 133 declared cross-package glob(s) (93 unique) are covered
node scripts/check-closing-keyword-parity.mjs	0	OK (3 parsers agree on all 9 keywords; 5 file(s) carry the grammar, all registered)
node scripts/check-comment-mask-corpus.mjs	0	5808 files, 0 disagree, 0 unparseable
node scripts/check-cross-package-test-inputs.mjs	0	OK: 25 package(s) read outside themselves, all declared
node scripts/check-shard-attestation.mjs	0	2 aggregate gate(s) count 3 declared leg(s) across 3 attesting job(s)
node scripts/check-skills-token-ratchet.mjs	0	36 authored bundle file(s) within their ceilings; this file 5467 of 6806
node scripts/check-test-completeness.mjs	3	PREREQUISITE NOT MET - NOT MEASURED, reason below
pnpm --filter @objectstack/lint run check:doc-formula-expressions	0	22 record-scoped formula example(s) across 426 files judged clean (plus 2 further legs)
pnpm --filter @objectstack/spec run check:skill-docs	0	Skill docs in sync
pnpm check:agent-test-spelling	0	0 violations - 430 file(s), 9 separator(s) JUDGED
pnpm check:corpus-claim-drift	0	OK, no new claim sites beside a pinned spelling
pnpm check:cross-package-test-inputs	0	OK: 25 package(s) read outside themselves, all declared
pnpm check:doc-authoring	0	831 pinned site(s) across 231 file(s), no growth, no burn-down unrecorded
pnpm check:nul-bytes	0	OK (8126 text file(s); no raw ASCII control bytes)
pnpm check:pm-governed-merges	0	self-test 243 assertions; live: the real generator declared 9 output(s) and certified this tree
pnpm check:refd-timer-probe	0	OK 5803 source file(s) swept
pnpm check:role-word	0	OK, no new occurrences of the reserved word
pnpm check:skill-compatibility	0	11 SKILL.md file(s) reconciled against 79 workspace packages
pnpm check:skill-frame-sync	0	2 copies of the decision frame structurally isomorphic across 2 files
pnpm check:skill-identifier-liveness	0	OK - Leg 1: 465 citation(s) over 46 published file(s); Leg 2: 8 exhaustive section(s), 0 gaps
pnpm check:watch-hint-literal	0	47 declaration(s) across 4 rostered name(s)

check:doc-formula-expressions first answered PREREQUISITE NOT MET twice in this fresh worktree - @objectstack/formula not built, then @objectstack/lint not built. Both prerequisites were supplied through the shared verify lock with slot issue-14836 (os-verify-lock: VERDICT command-exit 0, held 190s and 12s, waited 0s), and the exit 0 above is the reading after that. This is the gate behaving as documented, not a defect.

NOT MEASURED: node scripts/check-test-completeness.mjs, exit 3. It grades a saved turbo run test log and cannot produce one; its own text says the family derives it with no argument and that "the local reading for this gate is NOT MEASURED. It is not a red, and there is nothing here to fix." Supplying it means the repo-wide test farm, which is CI's run. Recorded as neither pass nor fail. The always-runs CI tail is likewise outside the local union, by the deriver's own closing line.

Reverse verification

Run from the committed state, under a trap on EXIT INT TERM restoring the absolute path:

• Mutation leg - "or a model-registry entry" re-inserted at :396; proved on disk before measuring: grep -c of that phrase moved 0 to 1, and git hash-object moved to 3cfae14e1c, away from the HEAD blob 9f5231126e.
• Reading - the file's measured token count rises 5467 to 5473 (the six tokens this half removes). The ratchet stays GREEN in that state, because 5473 is far under the 6806 ceiling. That is the honest reading and I am not dressing it up: the ratchet prices size, it cannot see the false claim. What moved is the measured count, not a verdict.
• Restore leg - git checkout HEAD -- on the absolute path; git hash-object returned 9f5231126eab1f93aa5d1a142d4a78867cf8275f, equal to the HEAD blob, git diff HEAD empty, git status --porcelain clean, and the ratchet re-read 5467.

Judgement calls

1. No re-flow. The file wraps prose at 80 columns. :396 shrinks 79 to 53 characters and :397 keeps five words ("author-time gate before reporting done:", 39 characters) - not an orphan. :404 shrinks 76 to 68 with a full 80-character :405 behind it. Re-flowing either would edit a line the card did not name and cascade through the paragraph. So: two intra-line deletions, zero line-count change.
2. The os validate fence comment at :400 stays - the seat's reading is confirmed by measurement. The identical line, byte for byte, is at skills/objectstack-api/SKILL.md:411. It is shared gate wording, not this file's claim, and the sentence right after it here already says this domain has no CEL site. Editing it in one of the two files would fork wording the published catalog shares. Observation only, no change.
3. No replacement validator named, no "not currently validated" caveat added. The triage was explicit on both; the fix is to stop promising.
4. No changeset, skip-changeset applied. The diff publishes nothing from any released package. Labels were written through MCP after a repo-scoped REST read probe returned 403 for this session, so the additive endpoint was unavailable: I read the current set, wrote the union, and read it back comparatively. Read-back shows documentation, size/xs, skip-changeset, needs:contract-review - every union member present, nothing stripped; documentation was added concurrently by a labeler, not by me.
5. Nothing else in the published catalog makes this claim. Swept skills/** for model-registry, model registry, ModelRegistry: remaining hits are :301 and :405 here, skills/objectstack-formula/SKILL.md:422 and the generated pointer at skills/objectstack-ai/references/_index.md:18. None promises a check, so the deletion strands no instruction.

Out-of-scope findings

None to file. Two observations for the seat, neither a card:

1. skills/objectstack-ai/SKILL.md:400, verbatim: os validate # Zod schema + CEL predicate validation + bindings (no artifact). Byte-identical to skills/objectstack-api/SKILL.md:411. Shared wording, deliberately untouched per the brief. Recording it so the measurement is on the record rather than in my head.
2. ModelRegistrySchema is NOT a dead surface, and I want that on the record so nobody reads this PR as evidence that it is. docs/qa/platform-checklist/areas/ai.json:526 already states it, verbatim: "ModelRegistrySchema (packages/spec/src/ai/model-registry.zod.ts) and ConversationSessionSchema (conversation.zod.ts) have NO open-framework runtime consumer - /ai/models and /ai/conversations/* are served only by cloud service-ai (enumerated in cloud's ai-route-ledger.ts, The SDK's entire ai namespace is dead, and the entire real AI surface is unreachable from the SDK — the two sets are disjoint #3718)". It is also a published export (packages/spec/api-surface/ai.json:118) with generated reference docs. So the schema has a consumer, in another repo - which is exactly why os validate, an in-repo stack-definition gate, has no path to it. This corroborates the card and is an argument against filing any enforce-or-remove card off the back of this flight.

Landing posture

Governed surface (skills/**): PR left as a draft. Not flipped ready, not enqueued, no auto-merge, no reviewers requested, no approving review, nothing merged. needs:contract-review is on it for the seat's contract read.

---

Generated by Claude Code

[os-litant]

Contract review PASS (in-seat, at CONTRACT_REVIEW_TIER) and ACCEPT — PR #14953 at 741e8e7071 (1 file, +2/−2, both intra-line).

Verified by the seat, not from the report: the diff is exactly the card's two deletions — :396 loses "or a model-registry entry", :404 loses "/ model" (the seat's own diff read at 12:34Z); the seat's own check-skills-token-ratchet run in a compare worktree at the head reads skills/objectstack-ai/SKILL.md at 5467 of 6806 (from 5475), exit 0; the line count stays 417; the surviving sentence at :404-407 is PR #14819's true template-dialect statement. The claim the diff removes — that os validate checks a model-registry entry — was the seat's own measurement at 04:40Z (stack.zod.ts:561-563, ModelRegistrySchema referenced only by its own file and test, validate.ts parsing ObjectStackDefinitionSchema) and the dev re-read all three at b1d49b3940; nothing in the tree promises that check, so the surface now promises only what runs. No contract moved; the label comes off the PR in this minute.

Judgement calls accepted: no re-flow (both deletions intra-line, the wrap stays under 80 columns with no orphan); the shared os validate fence comment at :400 untouched — byte-identical to skills/objectstack-api/SKILL.md:411, the seat measured the same at 12:14Z; no replacement validator named and no "not currently validated" caveat added. The dev's observation that ModelRegistrySchema is not a dead surface (its consumer is cloud service-ai, already recorded at docs/qa/platform-checklist/areas/ai.json:526) is accepted as the reason no enforce-or-remove card follows this flight.

Gates: 20 of the 21 in the derived union green at the head; check-test-completeness NOT MEASURED by its own text. CI on 741e8e7071 at 12:34Z: the three type checks and Test Core shards 2–6 green, shard 1 running — a governed draft waits for the approvers regardless.

Terminal: draft; review requested from os-zhuang and hotlong (read-back follows); the seat never flips, enqueues, approves or merges. On merge this card closes by Fixes #14836 and pm:dispatched comes off at that reading.

---

Generated by Claude Code