skills-sweep ④: objectstack-platform (2,633 lines, 8 files) — behavioral-claim verification, content-class execution-first

[huangyiirene]

Member card of program #13658 (published-skills factual sweep). Flight ④ target: skills/objectstack-platform/** — 8 files, 2,633 lines, enumerated at origin/main this hour:

• SKILL.md (1,286) · rules/plugin-lifecycle.md (340) · references/plugin-hooks.md (328) · rules/service-registry.md (319) · rules/bootstrap-patterns.md (142) · rules/plugin-hooks-events.md (108) · references/_index.md (61) · evals/README.md (49)

Scope: verify every behavioral claim against the implementation per the program method as BOUND by #13658 comments 5474435934 / 5474464789 / 5475032239 plus the flight-③ calibration comment 5476328676 (working false-density range 1.5–6%; eval rubrics inventoried at full weight; check:skill-examples empty-population rule; cross-file contradiction scan first; both-error-direction parse probes; executed probes for behavior-bearing claims).

Constraints: published governed surface — draft PR(s), human merge, .md content goes to maintainer review; token-ratchet ceilings unchanged; net line budget at most 0; NOT MEASURABLE recorded in the report, never annotated into the files. At 2,633 lines this flight may ship up to two sequential PRs at whole-file boundaries (Part of until the last, which closes). Corrections that make falsifiable operator/contract semantic claims fall under clause ②'s CONTENT limb (批 #12 on #13588): the PR creator attaches needs:contract-review in the same stroke the PR exists.

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

Generated by Claude Code

[huangyiirene]

Claim: PM loop round 1
Session: session_01EnE7G31tqbxN1rqpQmzurT
Branch: claude/issue-13747-skills-sweep-platform
Worktree: objectstack-issue-13747
Domain: domain:skills
File surface: skills/objectstack-platform/** (stop on breach; explain in the report)
Container & model: M (claim verification with executed probes — judgment tier), mode:subagent, model: opus — per this morning's dispatch-gates.mjs --tier posture: no path-derived mandate on skills/**, tier is the PM's judgment call.
Clause-②: yes — under 批 #12's CONTENT-limb criterion (falsifiable operator/contract semantic claims in published skills/**), sweep corrections are exactly that class; needs:contract-review attaches to the PR in the same stroke it exists, and this seat runs the review chain at CONTRACT_REVIEW_TIER at ACCEPT.
Serial constraints cleared: none — zero open PRs touch skills/objectstack-platform/** (flight ③'s PR #13740 is skills/objectstack-query/**; #13588's two branches are references/** + scripts/; #13605 is scripts/pm/** — all disjoint, no read-coupling).

---

Generated by Claude Code

[huangyiirene]

Dev claim — flight ④ (skills/objectstack-platform/**)

Session: session_01EnE7G31tqbxN1rqpQmzurT
Branch: claude/issue-13747-skills-sweep-platform (pushed empty at claim time as the write-route probe — green)
Worktree: objectstack-issue-13747, BASE 787d757405db4f3ebbc6ca811948af4438a3fe7a

Read myself and verified untruncated: #13747 body + all 6 timeline items; #13658 body + all 7 comments (24 timeline items, 17 events). Binding method comments 5474435934 / 5474464789 / 5475032239 / 5476328676 in hand.

Channel note: repo-scoped REST is 403 for this seat (session gate closed) — reads run on the zero-quota single-card payload channel, writes on MCP. The #13658 payload page truncated the timeline at 15 of 24 items, so the two late binding comments (5475032239, 5476328676) came from one targeted MCP get_comments call; MCP call count is reported in the final JSON.

Governed posture acknowledged: draft PR(s) only, needs:contract-review attached in the same stroke the PR is created, never self-cleared; token-ratchet ceilings untouched; net line budget across the package at most 0.

---

Generated by Claude Code

[huangyiirene]

os-dev-report

{
  "issue": 13747,
  "status": "done",
  "branch": "claude/issue-13747-skills-sweep-platform",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/13760",
  "premise_still_valid": true,
  "summary": "Flight 4 swept skills/objectstack-platform/** (8 files, 2,633 lines — face re-derived at origin/main and matching the dispatch enumeration exactly). 318 behavioral claims inventoried across 19 classes; 6 distinct false facts at 10 landing sites (1.9% distinct / 3.1% sites — the low half of the 1.5-6% working range, and roughly half of flight 3's density). Corrected in draft PR #13760: (1) rules/service-registry.md taught that a missing 'job' service is auto-filled by createMemoryJob() — it is deliberately NOT (maintainer ruling 2026-08-22), getService('job') throws, and the falsehood had FOUR landing sites in that one file (table row, section headline, criticality table, factory list); (2) SKILL.md's defineStack top-level key enumeration omitted the live v17 'server' key while the next paragraph tells readers an undeclared top-level key fails the load; (3) SKILL.md called the object-level 'namespace' field 'deprecated and ignored by the runtime' when it is retired per ADR-0129 D3 and REFUSED at parse — contradicting SKILL.md's own objects[] refusal posture thirty lines earlier; (4) the logger level enum omitted 'silent', which the same file then uses eight times; (5) Plugin.type was enumerated as 5 of 8 values in SKILL.md and 7 of 8 in rules/plugin-lifecycle.md, a cross-file contradiction settled against PluginSchema; (6) the 'Complete Working Example' package.json pinned ^16.0.0-rc.1 and claimed to mirror the blank template, which pins ^17.0.0 and ships a different dependency set. Five of the six are the deader-than-real direction that flight 2 flagged as the majority class — each hides shipped capability or understates a hard failure. Net line delta across the package 0 (2,633 to 2,633); net tokens -24 by the ratchet's own verdict lines; NO ceiling raised — every file was already at zero headroom, so every edit was made byte-shrinking in place. PM mechanism assumptions: file/line face VERIFIED exactly; zero open PRs touched this package (flight 3's #13740 merged mid-flight and is disjoint — merged up to it before the final derivation); repo-scoped REST is 403 for this seat (session gate), so reads ran on the zero-quota payload channel and writes on MCP. One PM assumption FALSIFIED: check:skill-examples population for this package is NOT empty — SKILL.md carries two os:check fences and both type-check, so the gate is applicable and was run green rather than recorded as empty-population. evals/README.md is clean: all 9 rubric claims VERIFIED (flight 3 found 4 false sites in its equivalent).",
  "tests": "All gate results read from each gate's own printed verdict line, exit codes captured before any pipe. Gate families derived from the REAL diff, never recalled: `node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands`. The FIRST derivation printed a STALE TREE warning naming 6 changed files that ARE the families, so I fetched and merged origin/main (d7ca6dc7, clean) and re-derived on the merged tree — no warning, 13 families, --repo assertion held. UNION RE-RUN AT THE FINAL COMMIT f6c81ba5: check-ci-filter-parity 0 | check-cross-package-test-inputs 0 | check-shard-attestation 0 | check-skills-token-ratchet 0 | check-test-completeness 3 | check:doc-formula-expressions 0 | check:agent-test-spelling 0 | check:cross-package-test-inputs 0 | check:doc-authoring 0 | check:pm-governed-merges 0 | check:role-word 0 | check:skill-compatibility 0 | check:skill-frame-sync 0. Plus two families the path derivation does not name but the diff visibly implicates: check:skill-examples 0 ('260 prose examples type-check across 3 surface(s)') and check:skill-refs 0 ('9 generated files in sync with packages/spec'). check-test-completeness exit 3 is NOT MEASURED, not a red — it grades a saved turbo test log, was handed none, and prints 'PREREQUISITE NOT MET ... Nothing was measured' before parsing a line; its own text instructs recording it as NOT MEASURED when run from the derived family. Two other gates first returned NOT MEASURED for the same reason and were made real by building their prerequisites under the shared verify lock: check:doc-formula-expressions needed @objectstack/formula and @objectstack/lint built; check:skill-examples needed @objectstack/client-react built (it refuses rather than false-greens on a missing dist, by design). Ratchet verdict lines quoted: 'skills/objectstack-platform/SKILL.md is 12687 tokens (ceiling 12705; headroom 18)' / plugin-lifecycle.md 2404 of 2408 / service-registry.md 2329 of 2331 / '38 authored bundle file(s) within their ceilings'. check:doc-authoring caught a defect I introduced and it is the reason for the second commit: my first draft of the job row cited an internal issue id, which skills/** forbids because the catalog ships whole into customer context windows — 'Internal issue-id reference(s) in the PUBLISHED skill catalog: skills/objectstack-platform/rules/service-registry.md:171'. Fixed, re-run green: '48 published skill files clean — no internal issue-id references'. EXECUTED PROBES (three throwaway probe files under packages/qa/dogfood, run against freshly built dist for @objectstack/core and @objectstack/spec, all deleted before the gate derivation — confirmed by a clean `git status --porcelain`): P1 core-fallback injection on ObjectKernel with only a data provider — metadata/cache/queue/i18n RESOLVED, job/auth/realtime THREW '[Kernel] Service ... not found', and CORE_FALLBACK_FACTORIES printed from the built artifact as exactly 'metadata,cache,queue,i18n'. P2 LiteKernel — cache/metadata/i18n all THREW (no fallbacks), registerServiceFactory/checkPluginHealth/getPluginMetrics all undefined, use() returns the kernel and is not a Promise. P3 registerService duplicate THREW '[Kernel] Service (name) already registered' (the exact string the skill quotes), getService on a null-valued key THREW, getServices() returned a Map. P4 hook order fired 'kernel:ready > kernel:bootstrapped > kernel:listening > kernel:shutdown' and a 'data:beforeInsert' handler registered and NEVER fired; getState ran idle/running/stopped. P5 circular plugin dependencies THREW on BOTH kernels. P6/P7 ServiceLifecycle = SINGLETON/TRANSIENT/SCOPED; all six LogLevel members accepted. Spec probe: ObjectStackDefinitionSchema's real shape parsed to 43 keys — all 39 documented keys REAL (zero phantom rows, unlike flight 1), 4 undocumented (server, onEnable, runtimeModule, viewItems); every negative key claim VERIFIED REFUSED (policies, roles, workflows, approvals, driver, featureFlags, features) with the exact 'Unrecognized key(s) on this stack definition' wording the skill quotes; objects[]/field refusal VERIFIED; SeedMode = insert/update/upsert/replace/ignore with mode 'upsert' and externalId 'name' defaulted in the parsed output; composeStacks error/override/merge and manifest first/last/index VERIFIED; ADR-0048 namespace enforcement VERIFIED on all four cases (prefixed accepted, bare refused, double-underscore refused, sys_ exempt). Connectors probe: the warn-then-drop line the skill quotes reproduced BYTE FOR BYTE — 'defineStack: connectors.stripe.bogusKey: (key) is not a declared connector key, so its value is dropped at load.' NO ABLATION was run and none is claimed: I added no gate, so there is nothing whose ability to fail needs proving. The equivalent control is differential and lives inside a single probe run — the four services documented as auto-injected RESOLVED while the one falsely documented as auto-injected THREW alongside the two documented as having no fallback, in the same kernel instance, off a dist built this session at the branch point. That is a positive read of the live map, not an inference from absence.",
  "mcp_calls": "9 — issue_read/get_comments (the 13658 payload page truncated its timeline at 15 of 24 items, so the two late BINDING comments came from one targeted call), add_issue_comment (claim), create_pull_request, issue_read/get_labels (failed loudly on a PR number — the documented routing signal, rerouted to the payload channel), get_label (verifying needs:contract-review exists rather than silently creating a typo), issue_write (label union), search_issues (dedup), issue_write (filed 13762), add_issue_comment (this report). Everything else — both card bodies, all comments, PR label read-back, PR body read-back — ran on the zero-quota payload channel because repo-scoped REST answers 403 for this seat.",
  "open_questions": [],
  "out_of_scope_findings": [
    "filed as #13762: Plugin.type's TSDoc in packages/core/src/types.ts under-enumerates the value set (7 of 8 — omits `objectql`, which ObjectQLPlugin actually declares); the field is typed `string`, so the comment is the only enumeration an author sees, and it is the likely source of the two divergent copies this PR corrects in the published catalog. Labelled `finding`, no assignee. Mechanically it is #13678 territory: pinning the comment against CORE_PLUGIN_TYPES closes the class, not just this instance."
  ],
  "not_measurable": [
    "SKILL.md's top-level key enumeration also omits `onEnable` and `runtimeModule`, both real and both accepted by defineStack. NOT corrected, deliberately: the sentence frames the list as metadata kinds, `onEnable` is a function member, and `runtimeModule` is documented in the schema as `Set by objectstack build; do not author by hand`. Whether the framing covers them is an editorial judgment, not a measurement — left for maintainer review. (`viewItems`, the fourth undocumented key, is a z.never() tombstone and is correctly omitted.)",
    "`ai:routes` is in the spec's kernel-bus registry and the runtime dispatcher listens for it, but no open-framework code triggers it (emitter is out-of-repo cloud AI). Both event tables omit it. The table's scope sentence says `the events the open framework actually fires`, which arguably makes the omission correct — editorial, not measurable here.",
    "`external-datasource:ready` and `datasource-admin:ready` are real registered {service}:ready names absent from the doc's three-item `e.g.` list. An `e.g.` list is not falsifiable by omission; recorded, not corrected.",
    "~12 claims about production boot behaviour — the ordering of the 11 auto-registered plugins in the boot-sequence diagram, `os start`'s exact non-zero exit on a busy port, and the dev port-hop end to end — need a real CLI/kernel boot harness. This is #13658's `cross-package runtime semantics` class (the addendum's ~20 min/claim row); flight 1 recorded 26 and flight 2 ~92, so the program-wide count is now well past the ~50 threshold the addendum named as the point a dedicated harness flight pays for itself.",
    "`node scripts/check-test-completeness.mjs` — NOT MEASURED locally (grades a saved turbo test log; none exists outside CI). Recorded per the gate's own instruction, not as a red."
  ],
  "claim_counts": {
    "total_claims_inventoried": 318,
    "distinct_false_facts": 6,
    "false_landing_sites": 10,
    "false_density_distinct_pct": 1.9,
    "false_density_sites_pct": 3.1,
    "verified": 302,
    "not_measurable": 16,
    "by_class": {
      "core-service criticality and fallback injection": "10 claims, 1 false fact at 4 sites",
      "defineStack top-level surface and refusal postures": "50 claims, 1 false fact at 1 site",
      "object/field/manifest/namespace authoring semantics": "10 claims, 1 false fact at 1 site",
      "plugin interface shape and lifecycle": "14 claims, 1 false fact at 2 sites",
      "logger configuration": "3 claims, 1 false fact at 1 site",
      "example project package.json and version pins": "8 claims, 1 false fact at 1 site",
      "kernel/DI runtime API surface": "18 claims, 0 false",
      "kernel bus events, payloads and ordering": "26 claims, 0 false",
      "ObjectKernel config options and defaults": "10 claims, 0 false",
      "drivers and HTTP adapters": "20 claims, 0 false",
      "templates and project layout": "10 claims, 0 false",
      "CLI command surface incl. negative claims": "40 claims, 0 false",
      "well-known plugin names and service keys": "17 claims, 0 false",
      "seed data": "7 claims, 0 false",
      "composeStacks": "5 claims, 0 false",
      "feature flags": "8 claims, 0 false",
      "references/_index.md schema pointers": "38 claims, 0 false",
      "evals/README.md rubrics (full weight)": "9 claims, 0 false",
      "health, observability and ops pitfalls": "15 claims, 0 false"
    }
  },
  "line_budget": {
    "package_lines_before": 2633,
    "package_lines_after": 2633,
    "net_line_delta": 0,
    "net_token_delta": -24,
    "per_file_tokens_vs_ceiling": {
      "SKILL.md": "12687 / 12705 (-18)",
      "rules/plugin-lifecycle.md": "2404 / 2408 (-4)",
      "rules/service-registry.md": "2329 / 2331 (-2)"
    },
    "ceilings_raised": 0
  },
  "files_changed": [
    "skills/objectstack-platform/SKILL.md",
    "skills/objectstack-platform/rules/plugin-lifecycle.md",
    "skills/objectstack-platform/rules/service-registry.md"
  ],
  "deviations": [
    "Shipped ONE PR, not the two the card permits. The whole correction set is 3 files and fits one whole-file-boundary PR; a second would have been packaging, not sequencing.",
    "Repo-scoped REST is 403 for this seat (session gate: 'GitHub access is not enabled for this session'), confirmed on both a read probe at start and the additive label POST at PR time. Channel switch declared: card/program reads and both PR read-backs ran on the zero-quota payload channel; writes ran on MCP.",
    "PR labels were written by the documented FALLBACK path, not the preferred additive REST POST: MCP read current -> union -> whole-group write -> compare read-back. Read leg used the payload channel's PR-side spelling (data-name=) after issue_read/get_labels failed loudly on a PR number. Read-back returned exactly union(read set, target) — documentation, size/s, needs:contract-review, skip-changeset — and a DELAYED second confirmation after worktree teardown still shows all four, with the PR still DRAFT.",
    "Merged origin/main into the branch mid-flight because the first gate derivation warned STALE TREE and named 6 changed files that ARE the families it derives from. Clean merge; all gates and the ratchet re-run after it, and the union re-run at the final commit f6c81ba5.",
    "Ran two gate families the path derivation does not name (check:skill-examples, check:skill-refs) because the diff visibly implicates them — the derived list is a clue, not a spec.",
    "Built four packages under the shared verify lock (core, spec, formula+lint, client-react+client) rather than reporting their gates as NOT MEASURED. Two of those gates refuse rather than false-green on a missing dist, by design; leaving them unbuilt would have made the report's green half fictional.",
    "The `auth` and `data` rows in service-registry.md's fallback table were reworded in the same stroke, and these are NOT falsehood corrections — both were true. `auth` gained the getService-throws consequence so it reads identically to the corrected `job` row (they behave identically), and `data` shed a redundant clause. Declared here so the falsehood count is not inflated by them."
  ]
}

---

Generated by Claude Code

[huangyiirene]

ACCEPT — PR #13760; governed md ⇒ parked draft, both approvers requested; contract-review PASS on the PR

Reviewed against GitHub, not the report: draft ✓ · Fixes #13747 + Part of #13658 ✓ · 3 files, all inside skills/objectstack-platform/ ✓ · labels {documentation, size/s, skip-changeset} with needs:contract-review attached at creation and cleared at PASS (comment 5476869480 on the PR carries the trace and the five independently re-measured anchors). Budget exemplary: net lines 0, net tokens −24, zero ceilings raised — every edit paid byte-shrinking in place on zero-headroom files.

Spot readings beyond the anchors: the differential control inside P1 (four documented-auto-injected services RESOLVE while the falsely-documented one THROWS beside the two documented-as-absent, same kernel instance, dist built this session) is a positive read of the live map — the right instrument, no ablation owed. The check:doc-authoring catch (internal issue-id briefly in published text → gate red → fixed → green) is recorded as the farm working. One PM assumption falsified honestly: check:skill-examples population here is NOT empty (2 os:check fences, both type-check, gate run green rather than recorded empty). The auth/data row rewordings are declared as non-falsehood edits — accepted, they make the corrected job row's consequence uniform.

Deviations accepted: one PR instead of the permitted two (3 files fit one whole-file boundary); mid-flight merge-up for a non-stale derivation; carrier attached PR-side only (the double-carrier card copy was skipped — moot at clear, but flights ⑤+ should attach both per 载体纪律).

Out-of-scope finding verified: #13762 exists (finding, unassigned — packages/core territory, central triage routes it). NOT-MEASURABLE ×16 recorded with reasons; the cross-package runtime class is now well past the program's ~50-claim harness threshold — flagged on #13658.

Card → pm:awaiting-maintainer; closes via Fixes when the maintainer merges #13760. Flight ⑤ dispatch is merge-cadence-gated (6 drafts already waiting).

---

Generated by Claude Code