skills-sweep ①: objectstack-formula (579 lines) — behavioral-claim verification against @objectstack/formula, calibration flight

[zhuangjianguo]

Member ① of program anchor #13658 (maintainer-ordered published-skills factual sweep). Scope: skills/objectstack-formula/** — 2 .md files, 579 lines.

Method, verdicts, non-vacuity control and PR-body shape: per the anchor. This is the calibration flight: besides its own corrections, its report must state (a) claims inventoried, (b) minutes/tokens per claim class, (c) false/imprecise density — the anchor re-sizes flights ②–⑫ from these numbers.

Implementation ground truth: packages/*/formula* (@objectstack/formula) and its executed test surface; the check:doc-formula-expressions gate already type-checks examples — this flight verifies the BEHAVIORAL half (function semantics, coercion rules, error behavior, operator tables) that compile-validity cannot see.

Governed surface (skills/**): draft PR, human merge; token ratchet is byte-neutral-or-shrinking, ⛔ no ceiling raises.

Generated by Claude Code

[zhuangjianguo]

Claim: PM loop round 7 — dispatch (program #13658 flight ①, calibration)
Session: session_01EXxTW8mvPBhoHxmyPZ63de
Branch: claude/issue-13660-formula-skill-sweep
Worktree: objectstack-issue-13660
Domain: domain:skills
File surface: skills/objectstack-formula/** (.md) ONLY for edits; read-only everywhere else (implementation + tests are evidence, ⛔ not edit targets — an implementation defect found while verifying is a FINDING to file, never a fix in this PR). Stop on breach.
Container & model: M, mode:subagent, model: opus — no path-derived mandate (skills/** is not in the fable globs; clause-② content limb not fired: verifying/correcting published claim text changes no accept/reject behavior); judgment-bearing claim adjudication ⇒ default opus.
Clause-②: yes | no — no
Serial constraints cleared: skills/objectstack-formula/** has no in-flight or queued holder; PR #13577 touches only skills/objectstack-query/** (disjoint). Token ratchet: byte-neutral-or-shrinking, ⛔ no ceiling raise. Batch 2/3 (with the ui fold flight). Closing: Fixes #13660; anchor #13658 referenced with zero closing verbs adjacent.

---

Generated by Claude Code

[zhuangjianguo]

ACCEPT — reviewer: skills lane seat, session session_01EXxTW8mvPBhoHxmyPZ63de (contract-review tier, full diff personally verified). PR #13674, head cb9a3cd44, draft on the manual lane.

Verified: the diff matches all seven claimed corrections exactly (one file, +9/−12, −3 lines, token ratchet byte-flat at 6002/6002 with the gate's own verdict line quoted). The evidence chain is the program spec working as intended: every FALSE claim carries an executed probe or an enumerated schema read (the defaultValue control row — a real CEL slot auto-wraps, defaultValue doesn't — is exactly the discriminating experiment; the phantom template surfaces return zero hits repo-wide; the stdlib pointer disproof is npm pack --dry-run counting 0 src/ entries). Non-vacuity control satisfied by execution with the verbatim overload string. The 26 NOT MEASURABLE claims are recorded with their reasons and locations, not skipped. No implementation defect found — every disagreement was the document wrong about settled code, correctly not "fixed" in code.

Calibration accepted and adopted for the program (recorded on anchor #13658): 197 claims / 6 FALSE (3.0%) / 8× concentration in surfaces-mapping tables (10.5% vs 1.3%) — flights ②–⑫ point execution-priority at tables first. Landing: draft stays; hotlong requested; human merge closes this card.

---

Generated by Claude Code

[claude]

os-dev-report

{
  "issue": 13660,
  "status": "done",
  "branch": "claude/issue-13660-formula-skill-sweep",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/13674",
  "premise_still_valid": true,
  "summary": "Flight 1 (calibration) of the published-skills factual sweep. Inventoried every behavioral claim in skills/objectstack-formula/** (2 files, 579 lines) FIRST, then verified in inventory order against the implementation (@objectstack/formula, @objectstack/lint, @objectstack/spec) by reading plus executed probes — never against another document. 197 behavioral claims inventoried; 165 VERIFIED, 6 FALSE (all corrected in this PR), 26 NOT MEASURABLE. One further falsehood outside the behavioral inventory (a navigation pointer) was also corrected because it was measured, false and byte-shrinking. Token-neutral under the ratchet: 6002/6002, zero headroom, no ceiling raise; lines 554 to 551. Draft PR, skip-changeset applied, no reviewers, auto_merge null, draft re-read as true after the label write. No implementation defect found worth filing: every disagreement was the document being wrong about settled code, not code contradicting itself.",
  "tests": "All at final head cb9a3cd44 (worktree clean vs HEAD). BASELINE @objectstack/formula: 'Test Files 25 passed (25) / Tests 647 passed (647)' before edits; IDENTICAL 25/647 after edits with probes removed — includes skill-catalog-sync.test.ts, the drift-guard pinning the stdlib table against CEL_STDLIB_FUNCTIONS. NON-VACUITY CONTROL (executed): probe of celEngine.evaluate on a null-plus-string source printed 'runtime :: no such overload: dyn' plus the null/string operand types, matching the skill's asserted message byte for byte. Same executed run pinned floor(-1.2)=-2, ceil(-1.2)=-1, addMonths(date 2026-01-31, 1) clamping to Feb 28, today()/daysFromNow(45) on UTC midnight, all 8 advertised cel-js built-ins resolving, all 3 Salesforce-flavor sources failing compile, and PRIOR()/ISBLANK()/isBlnk() failing with the promised no-matching-overload type error. 76 probe assertions ran; 75 passed and the 1 failure WAS the finding (end minus start did not fault). GATES: family derived from the real diff via 'node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack' — 13 families, 12 EXIT=0, 1 NOT MEASURED (check-test-completeness.mjs EXIT=3 = PREREQUISITE NOT MET by its own text, no saved turbo test log). Ratchet's own verdict line: 'check-skills-token-ratchet: skills/objectstack-formula/SKILL.md is 6002 tokens (ceiling 6002; headroom 0)' and '38 authored bundle file(s) within their ceilings'. check:doc-formula-expressions first exited 1 = PREREQUISITE NOT MET (formula and lint unbuilt), not a finding; after building both it printed 'self-test: 58 cases passed' and '22 record-scoped formula example(s) across 426 files / 1451 TS blocks judged clean'. Beyond the derived family, check:skill-examples: '260 prose examples type-check across 3 surface(s)'. check-nul-bytes OK, 7551 files, no raw control bytes. ESLint narrowed with three measured readings: population read from eslint.config.mjs is the ts/tsx/mts/cts/js/jsx/mjs/cjs globs with NO block naming .md; --format json over the diff's only file returns 1 result, 0 messages; type-aware linting is configured 0 times (no projectService, no parserOptions.project), so this diff cannot move any untouched file's verdict.",
  "claims_inventory": {
    "total": 197,
    "by_class": {
      "surfaces_tables (CEL/cron/template slot rows)": 38,
      "stdlib_catalog (function rows, built-ins, sync claims)": 35,
      "mandatory_patterns (incl. the previous-binding table)": 32,
      "legacy_to_CEL_translation_table": 17,
      "cel_syntax_cheatsheet": 14,
      "has_semantics_and_nullable_guard_publish_gate": 11,
      "date_arithmetic_note": 11,
      "core_contract_and_dialects": 11,
      "adr0032_and_verify_your_work": 10,
      "cron_and_template_quick_reference": 9,
      "determinism_contract": 4,
      "brace_trap (map-literal parse error)": 3,
      "null_plus_string_coercion": 2
    },
    "verified": 165,
    "false_fixed": 6,
    "not_measurable": 26,
    "covered_by_gate": "NO class is entirely covered. Two partial coverages, stated per class as required: (1) stdlib_catalog — skill-catalog-sync.test.ts pins that all 30 catalog NAMES appear in the skill, one direction only (catalog subset of skill); it says nothing about the Returns/Notes columns, so all 22 function semantics were still verified by execution. (2) every class containing a fenced ts example — check:skill-examples type-checks 227 blocks on this surface, but compile-validity is never the whole claim: the sweep's 6 falsehoods all sit in table rows and prose that no example block touches, which is precisely why this gate never caught them.",
    "out_of_inventory": "About 118 of 579 lines make no verifiable behavioral claim (front-matter, the strategic-context paragraph, the Skill Boundaries routing table, Open questions, See also, headings and separators) and are excluded by the anchor's rule. One measured falsehood lives there — the stdlib pointed at node_modules paths under @objectstack/formula/src/, but package.json declares files as dist/README/CHANGELOG and 'npm pack --dry-run --json' reports 0 src/ entries, so those paths never ship. Corrected anyway (measured, false, byte-shrinking) and reported separately so the density figure stays honest."
  },
  "calibration": {
    "effort_per_class": "Measured on this flight, one seat, one container. (a) Rows with a directly executable oracle — stdlib functions, template formatters, cron forms, the legacy translation table, ~100 claims: about 1.5 min/claim. They batch: one 76-assertion probe file, one run, one read. (b) Spec-surface rows, 38 claims: about 4 min/claim — each needs locating the schema, confirming the declared type, and often chasing a retirement note in a sibling module. This is where the cost AND the yield are. (c) Prose behavioral paragraphs — has()/null-guard gate, ADR-0032, ~25 claims: about 8 min/claim, because the truth lives in a 700-line lint module with its own surface ledger. (d) Cross-package runtime semantics — the previous-binding table and bulk-write dispatch, ~26 claims: NOT MEASURABLE from a formula-package flight; they need a kernel harness. Estimate 20 min/claim if a flight is scoped to build one. WALL CLOCK for this whole flight: roughly 3.5 hours including 4 builds (spec closure 2m53s, formula 2m40s, lint 12s, client-react 3m09s) and the gate runs. Excluding builds, about 2.5 hours for 197 claims = about 45 sec/claim averaged, but that average is misleading — see the class split.",
    "false_density": "6 FALSE / 197 inventoried = 3.0% overall. NOT uniform, and this is the number that should size flights 2 through 12: 4 of the 6 sit in the surfaces/mapping tables — 4/38 = 10.5% there against 2/159 = 1.3% everywhere else, an ~8x concentration. That reproduces PR #13577's pattern exactly (a mapping column asserting engine behavior). RECOMMENDATION: point the remaining flights at operator/surface/mapping TABLES first and treat prose as lower yield; a table-first pass would have found 4 of 6 for ~20% of the effort. Second pattern worth carrying: 3 of the 6 are PHANTOM ROWS naming schemas or fields that were retired or never existed (etl.schedule, sync.schedule, system/notification's 5 fields, github titleTemplate/bodyTemplate) — cheap to detect with a repo-wide grep for each named identifier, and a candidate for a mechanical gate."
  },
  "files_changed": "skills/objectstack-formula/SKILL.md only (554 to 551 lines; 6002 to 6002 tokens; bundle total 186903 unchanged). references/_index.md is generator-owned and was measured-correct, so untouched. No implementation file touched.",
  "changeset": "none — pure skills/** docs correction, releases nothing from any package. skip-changeset applied via the additive POST labels endpoint; read back and present.",
  "governed_surface_posture": "draft=true, auto_merge=null, not marked ready — human merge. Final re-read after the auto-labeler settled: labels = documentation, size/s, skip-changeset (my skip-changeset survived the size-labeler write, which is why the additive endpoint plus read-back is used), and requested_reviewers = 1 (user hotlong). I did NOT request that review and did not touch it: it appeared between my post-label read and the final read, so it belongs to another actor. Recorded, not corrected. It does not conflict with the governed posture — draft plus human merge is unchanged.",
  "deviations": "1) The dispatch-gates derivation printed a STALE TREE warning: my worktree HEAD is the branch point while origin/main advanced 3 commits during the flight. The single file it flagged as changed is scripts/pm/check-half-states.mjs, a changeset-triggered family — and this diff carries no changeset — so the 13 skills/** families are unaffected. Reported rather than silently accepted. 2) Two gates needed workspace builds before they measured anything (check:doc-formula-expressions, check:skill-examples); their first non-zero exits were PREREQUISITE NOT MET, not reds, and both are green after building. 3) ESLint repo-wide was narrowed with the three-reading declaration above rather than run. 4) OBSERVED PLATFORM BEHAVIOR, not my write: PR creation auto-appended a second session-URL signature footer after the one in my body, so #13674 shows the footer twice. Body lines 1-149 read back byte-identical to what was sent — nothing was sanitized away, including the angle-bracketed CEL overload types inside the fenced blocks. Left alone rather than PATCHed, since editing the body is the documented way to downgrade a session-URL footer. 5) State on the PR that is NOT mine, recorded rather than corrected: a reviewer (hotlong) was requested, and the auto-labeler added documentation and size/s, both after I posted the first read. My skip-changeset label survived that concurrent label write — verified by the final read-back.",
  "mcp_calls": "0 — the whole run used the repo-scoped REST channel, probed first and confirmed working (issue reads, PR body read, PR create, additive label write, comment post). gh is absent in this container. No MCP GitHub tool was called.",
  "open_questions": [],
  "out_of_scope_findings": []
}

Generated by Claude Code

---

Generated by Claude Code