skills-sweep ⑩: objectstack-ai (691 lines, 3 files) — behavioral-claim verification, content-class execution-first

[huangyiirene]

Member card of program #13658 (published-skills factual sweep). Flight ⑩ target: skills/objectstack-ai/** — 3 files, 691 lines at origin/main this hour: SKILL.md (602) · references/_index.md (45, generator-owned — sync gate only, never hand-edit) · evals/README.md (44).

Scope: verify every behavioral claim against the implementation per the program method as BOUND by #13658 comments 5474435934 / 5474464789 / 5475032239 + calibrations 5476328676 / 5476875577 / 5477729758 / 5479170359 / 5479891720 (content-class first; both error directions; enumeration-vs-schema-member-set standing probe; "silently / ignored / dropped" seed; spec-.describe()/CLI-help twin rule; eval rubrics FULL weight). Likely implementation homes: packages/spec/src/ai/** (agent/AI schemas if present — locate, don't assume), knowledge-* packages, embedder-openai, packages/mcp, AI-related plugin/service packages. ⚠️ Boundary note: some AI capability is out-of-repo (cloud side) — a claim about an emitter the open framework does not ship is the flight-④ ai:routes class: settle what the OPEN repo enforces, record cloud-side behavior as NOT MEASURABLE rather than false.

Constraints: published governed surface — draft PR, human merge; ceilings unchanged; net line budget ≤ 0; NOT MEASURABLE recorded. Clause ② CONTENT limb (批 #12): needs:contract-review to PR AND this card same-stroke.

Part of #13658; closes with the flight's PR (Fixes).

Generated by Claude Code

[huangyiirene]

Claim: PM loop round 1
Session: session_01EnE7G31tqbxN1rqpQmzurT
Branch: claude/issue-13841-skills-sweep-ai
Worktree: objectstack-issue-13841
Domain: domain:skills
File surface: skills/objectstack-ai/** (stop on breach; explain in the report)
Container & model: M, mode:subagent, model: opus — no path-derived mandate on skills/** per this morning's dispatch-gates.mjs --tier posture.
Clause-②: yes — 批 #12 CONTENT-limb criterion; both carriers same-stroke at PR creation; this seat reviews at CONTRACT_REVIEW_TIER at ACCEPT.
Serial constraints cleared: zero open PRs touch skills/objectstack-ai/** (listings read this hour); sibling flight ⑨ (#13840, skills/objectstack-upgrade/**) runs this same wave — disjoint, no read-coupling.

---

Generated by Claude Code

[huangyiirene]

os-dev-report

{
  "issue": 13841,
  "status": "done",
  "branch": "claude/issue-13841-skills-sweep-ai",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/13867",
  "premise_still_valid": true,
  "summary": "Flight 10 of the published-skills factual sweep swept skills/objectstack-ai (3 files, 691 lines: SKILL.md 602 + references/_index.md 45 generator-owned + evals/README.md 44 — PM's face assumption re-derived from the fetched tree and CONFIRMED exactly). 11 distinct false facts across 21 landing sites were corrected, ~4.3% distinct / 8.2% by site against ~255 inventoried behavioral claims — inside the 1.5-10% working range. Sharpest: every Action was taught as AUTO-exposed to AI when ai.exposed defaults to FALSE and ai.description (40+ chars) is required (the wrong-default class); the retired agent.tools and skill.triggerPhrases keys were taught as live (5 + 1 landing sites, both now parse errors, probe-confirmed); the built-in-skills table marked three cloud skills OSS and a callout claimed the open framework ships the ask agent, contradicting the SAME document 90 lines earlier and settled by git ls-files grep -ic service-ai returning 0; confirmText was taught as a destructive signal after ruling #7828 removed that leg; and agent guardrails/memory/structuredOutput were prescribed as controls while the schema marks all three not-enforced and the liveness ledger records no runtime reader. Correcting F5b also surfaced a SHIPPED open-edition capability the document hid: a skill's instructions are projected onto the MCP prompts primitive by packages/mcp/src/skill-prompts.ts. Net -15 tokens / -10 lines; both ratcheted files stay under their existing ceilings and NO ceiling was touched (the posture all eight preceding flights landed with — none of their commits touched the ratchet script). ~59 claims recorded NOT MEASURABLE (the flight-4 ai:routes boundary: the whole in-product AI runtime is out-of-repo), each listed in the PR body rather than skipped. Draft PR, human merge only; needs:contract-review attached to BOTH carriers in the same stroke as PR creation, read back on both and confirmed present; this seat does not self-clear it. Claim comment: the PM's dispatch-atom comment 5479904482 on #13841 names THIS session id (session_01EnE7G31tqbxN1rqpQmzurT) and this exact branch, so it IS this seat's claim comment (the flight-5 shared-session form); assignee arrived empty (PM-side half state) and I did not write it. File surface held to skills/objectstack-ai/** with zero breach — no ratchet constant, no spec source, no changeset.",
  "tests": "All heavy runs went through bash scripts/pm/os-verify-lock.sh with OS_VERIFY_LOCK_SLOT=issue-13841; every exit code captured by redirect BEFORE any pipe. Union re-run and every figure below is from HEAD 6e6d02bd1 (git rev-parse --short HEAD at the final commit; nothing was measured on an earlier tree). BUILDS FIRST: pnpm --filter '@objectstack/spec...' build exit=0, then pnpm exec turbo run build --concurrency=2 --filter='./packages/*' --filter='./packages/*/*' exit=0 (70/70 tasks successful) — so no probe read a stale .d.ts. GATE FAMILIES from node scripts/pm/dispatch-gates.mjs with NO path argument (script takes its own change set from the merge base; exit=0; header names the tree it answered for: objectstack-ai/objectstack at 6e6d02bd1), harvested with --commands so neither of the two spellings is dropped. 14 families run: 13 exit=0, and node scripts/check-test-completeness.mjs exit=3 which is NOT MEASURED by the gate's own printed verdict ('PREREQUISITE NOT MET ... the local reading for this gate is NOT MEASURED. It is not a red') because it grades a saved turbo test log only CI produces — recorded as not-measured, never as green or red. Green judgment lines quoted from the gates themselves: check-skills-token-ratchet '38 authored bundle file(s) within their ceilings; 11 generator-owned file(s) measured, not ratcheted'; check-nul-bytes 'OK (scanned 7601 text file(s) -- no raw ASCII control bytes)' exit=0, plus a manual control-byte scan of both edited files (grep -naP over the 0x00-0x1f/0x7f class) returning no matches. BEYOND THE DERIVED LIST because the diff edits prose INSIDE an os:check fence: pnpm --filter @objectstack/spec run check:skill-examples exit=0, verdict line 'OK 260 prose examples type-check across 3 surface(s) -- every marked block parsed, so tsc ran the SEMANTIC pass on all of them'. POPULATION CONTROL for that gate (binding per #13658 comment 5476328676): population is NON-EMPTY for this surface — 'skills + docs (@objectstack/spec): 227 block(s)' of 260, and this package contributes 4 os:check fences; not an empty-population reading. Its first invocation exited 1 on 'packages/client-react/dist holds no .d.ts declarations — the package is not built', a REFUSED verdict rather than a finding, cleared by the workspace build above and re-run green. EXECUTED PROBES (the non-vacuity control the program requires): probe 1, 27 assertions against the built dist through @objectstack/spec/ai and /ui — three positive controls parsed (minimal valid skill/agent/tool), then every key this document taught as authorable was REJECTED with its located prescription: skill.triggerPhrases ('was removed in @objectstack/spec 17.0.0 (audit close-out) -- phrases were never matched against the user's message'), agent.tools ('was removed in @objectstack/spec 17 -- use skills'), skill.permissions, all five retired tool keys (category/permissions/active/builtIn/requiresConfirmation), agent.knowledge, top-level agent.temperature; and the asserted defaults and member sets all held (skill.surface 'ask', skill.active true, planning.maxIterations 10, agent.surface 'ask', structuredOutput strict=false/maxRetries=3/retryOnValidationFailure=true and no retry object, knowledge refresh.onRecordChange=true and aiExposed=true, trigger operators exactly eq/neq/in/not_in/contains). Probe 2, 12 assertions, 12 pass / 0 fail: inline agent model.provider accepts exactly openai/azure_openai/anthropic/local while ModelProviderSchema also accepts google/cohere/huggingface/custom; a plain action carries NO ai block (ai=undefined, nothing auto-exposes it); action.ai.exposed default is FALSE; ai.exposed:true without ai.description is refused at ActionSchema ('ai.description is required (40+ chars) when ai.exposed is true'); actionLooksDestructive with confirmText alone = FALSE while mode:'delete' = true and variant:'danger' = true and ai.requiresConfirmation overrides = false; MCP binding approval accepts 'always' with default 'never'. INSTRUMENT FAULT DISCLOSED: probe 1's first pass read the two provider enums by reaching into a Zod internal (.def.entries) and got undefined through the lazySchema proxy, and its third apparent failure tested ai.exposed's description requirement at the wrong schema level (it is an ActionSchema refine, not an ActionAiSchema one) — all three were instrument faults, corrected in probe 2 by parsing each candidate instead of introspecting; none was a finding. Probe 3, the open-edition capability F5b adds, proven by a real test run: pnpm --filter @objectstack/mcp exec vitest run --maxWorkers=2 src/skill-prompts.test.ts -> 'Test Files 1 passed (1) / Tests 12 passed (12)', including 'drops the tool-binding half -- it is cloud-runtime-only', 'does not project an inactive skill' and 'does not project a skill without instructions'. TOKEN RATCHET before/after, both files: SKILL.md 6806/6806 headroom 0 -> 6791/6806 headroom 15; evals/README.md 315/315 -> 315/315; references/_index.md 796 tokens generator-owned, measured not ratcheted and NOT hand-edited (check:skill-frame-sync exit=0 proves it is in sync with gen:skill-refs). ESLINT NARROWING, DECLARED with all three legs of evidence: (1) the population is read from eslint's OWN config resolution, not from my guess — npx eslint --no-inline-config --format json on both changed paths returns, per file, 'File ignored because no matching configuration was supplied'; (2) the count is read from the --format json output: 2 results, 0 errors, 0 lint warnings (2 ignore notices), exit=0; (3) invariance — this repo's flat config supplies no matcher for .md at all, so a Markdown-only diff cannot move any judgment on any file eslint does lint, and no type-aware linting is reachable for a file with no matching config. The repo-wide pnpm lint run stays CI's, as the local-scope rule allows. NOT RUN and stated as such: nothing else; no gate result in this report comes from a bare $? read behind a pipe.",
  "mcp_calls": "7 — 1 issue_read get_comments (#13658 method-comment tail), 1 search_issues (dedup, non-empty so no separate control needed), 1 issue_write create (#13865), 1 create_pull_request (#13867), 2 issue_write label unions (PR + card), 1 add_issue_comment (this report). Every card/PR/comment READ went through the zero-quota public payload channel and git; the repo-scoped REST probe returned 403 'GitHub access is not enabled for this session' at the top of the run, so the REST channel was unavailable to this seat and the MCP+payload degradation ladder was used throughout — declared here as a channel switch.",
  "open_questions": [],
  "out_of_scope_findings": [
    "filed as #13865: ActionAiSchema.requiresConfirmation's JSDoc in packages/spec/src/ui/action.zod.ts still names confirmText as a destructive signal, contradicting actionLooksDestructive in packages/runtime/src/action-execution.ts and maintainer ruling #7828 Option A — the spec-side twin of correction F8, filed per the binding twin rule (#13658 comment 5479170359) so the same false fact is not left published one surface over. Single-site drift: the two other in-repo carriers (packages/mcp/src/mcp-server-runtime.ts and the runtime pin test) already state the corrected rule. Filed unassigned, no pm:queue label, in the repo where the fix lands.",
    "NOT filed, recorded for the roster instead: search_knowledge sits in PLATFORM_TOOLS_BY_PACKAGE and is referenced by this skill's Knowledge section, yet no row of the built-in-skills table claims an owning skill for it. The skill-to-tool assignment lives in the cloud repo, so naming an owner from here would be inventing one — recorded in the PR body's NOT MEASURABLE section as a note for a later flight or the optional harness flight, not as a card.",
    "NOT filed, checked and cleared: content/docs/releases/v17.mdx says authoring agent.tools 'is not a parse error -- the key is tombstoned and stripped'. A direct AgentSchema.parse DOES reject it (probe-confirmed), but the agent-tools-to-skills conversion in packages/spec/src/conversions/registry.ts carries retiredFromLoadPath:true, so the release note is true of the LOAD path it describes. No card — and releases pages are never edited from a code PR in any case."
  ]
}

Generated by Claude Code

---

Generated by Claude Code