spec-property-retirement: the consumer walk rebuilds the whole dependency closure per lap (~6 min on a shared box) — default to an affected-scoped walk against the merge base (#16104 section 3.2)

[claude]

Split out of #16104 (section 3, item 2) by the skills seat at flight P's ACCEPT, as that card's claim prescribed.

The reading

The spec-property-retirement playbook's consumer walk rebuilds the dependency closure per walk: on the spec seat's 2026-09-05 shift each lap cost about six minutes on the shared box (measured on one retirement card). The walk's coverage does not depend on rebuilding everything — an affected-scoped walk (turbo ls --affected against the merge base, then the walk over that set) reaches the same consumers.

What

The playbook's walk rule names the affected-scoped form as the walk's default and keeps the full closure as the fallback when the affected set cannot be computed (a shallow clone, a missing merge base). Line-neutral at the pinned count; the row-bytes pin (326 after PR #16357) untouched; rules-only register.

Acceptance

• check-skill-line-ratchet (the playbook's line and row pins), check-skill-id-lint, check-governed-merges --test exit 3; whole-repo lint.
• The report shows one retirement walk timed both ways on the same card (the measurement is the acceptance).
• Governed (.claude/**): draft PR, seat review, os-zhuang + hotlong, human merge.

Refs #16104.

---

Generated by Claude Code

[claude]

Claim: PM loop round 8 (continued) — flight Q, one governed skills card: the spec-property-retirement playbook's consumer walk rebuilds the whole dependency closure per lap (about six minutes on the shared box, measured by the spec seat on one retirement card); the playbook names an affected-scoped walk against the merge base (turbo ls --affected) as the walk's default, with the full closure as the fallback when the affected set cannot be computed. The acceptance is a measurement: one walk timed both ways on the same card.
Session: session_019RfFHiRCSs3JXLK4cwcfox
Branch: claude/issue-16389-retirement-walk-affected-scope
Worktree: objectstack-issue-16389
Domain: domain:skills
File surface: .claude/skills/spec-property-retirement/SKILL.md (line-pinned; line-neutral by folds inside the walk rule's block; the row-bytes pin — 326 after PR #16357 — untouched). ⛔ No other file; ⛔ not the ratchet map.
Container & model: S, mode:subagent, opus — dispatch-gates.mjs --tier: no path-derived mandate on this playbook; a procedure rule, no contract move.
Clause-②: no
Serial: no open PR touches this playbook (PR #16357 moves only its row-bytes pin in the ratchet map). GOVERNED (.claude/**) ⇒ draft PR, in-seat review, os-zhuang + hotlong, human merge, skip-changeset; at ACCEPT the seat adds needs-user-decision + the 维护者速读. The PR carries Fixes #16389.

---

Generated by Claude Code

[claude]

os-dev-report

PREMISE FALSIFIED — no PR opened, per ruling item 1. No rule in the playbook prescribes rebuilding a dependency closure per walk lap, and the proposed remedy is measurably a no-op for a packages/spec retirement: the affected set and the consumer closure are the same set, and cold-time to the second. Evidence, then the JSON.

1 · The grep — every command the playbook names

.claude/skills/spec-property-retirement/SKILL.md at origin/main 0374bcba9f, 337 lines, unedited by this run. grep -noE '[^`]*(pnpm|turbo|npx|tsc|os migrate)[^`]*' returns, in full:

line	command
255	pnpm --filter @objectstack/spec gen:schema
264 / 266	pnpm i18n:extract · pnpm check:i18n
296	pnpm check:adr-0087-registration
314	pnpm --filter @objectstack/spec build — the file's ONLY build
315	pnpm --filter @objectstack/spec check:generated
316	pnpm check:i18n && pnpm --filter @objectstack/spec test
75 / 86 / 283 / 334	tsc — named as a mechanism, never as a command line
279	turbo.json — a config file in the radius rule; the playbook never invokes turbo

Zero turbo run. Zero unfiltered build. The one build is filtered to one package — and @objectstack/spec has 0 workspace dependencies (measured), so pnpm --filter '@objectstack/spec^...' build resolves to 0 no packages. There is no upstream closure to rebuild at all.

2 · What the playbook actually says about walking consumers

• 75-77 (§1): retiredKey() makes every authoring site fail to compile — 「让墓碑替你找调用者,再去人肉搜」. A mechanism statement. No command, no scope.
• 283 (§4 Examples): 「墓碑路线上 tsc 替你找齐」. Same shape.
• 301-305 (§4 dogfood): the one place the playbook does scope a consumer run — 「推送前跑它的定向套件,不管你的 import 图怎么说」. A targeted suite: the opposite of a closure rebuild.
• 278-281 — the rule the finding names, docs(skills): make the retirement pin's walk radius a declared radius #15566 (landed 88bc10e840, 「make the retirement pin's walk radius a declared radius」): 「半径按包申报一次:写进 scripts/cross-package-test-inputs.mjs … 并配齐 turbo.json 的 PKG#test inputs」. It prescribes a declaration, not a run. Its cost is turbo cache invalidation of @objectstack/spec#test on later PRs — which is exactly what the spec: the rest of the incident-response, training and change-management families — every remaining key and all fifteen defs — has zero readers; whole-def enforce-or-remove is the open question left after #14477 #15513 dev's own open question said (「the cache-invalidation cost lands on every PR」), not a per-lap rebuild for the retiring dev.
• 327-330 (§6) is decisive against adding the proposed rule here: 「四个陷阱不写在这里,因为它们对每个 dev 都成立、锚点在别处 … 锚点是 AGENTS.md 的「Touched packages/spec?」一节与 .claude/agents/os-dev.md 的「本地验证范围」;⛔ 这里不留第二份拷贝。」 And .claude/agents/os-dev.md:129 already carries the affected rule: 「受影响包 = CI 会测的包,清单读 TURBO_SCM_BASE="$BASE" pnpm exec turbo ls --affected」. The candidate line would be a second copy of a rule living at the anchor this file points to — forbidden by the file's own text.

3 · The measurement (dry walk, synthetic retirement, this box)

Instrument: one byte-level edit under packages/spec/src (a comment appended to data/document.zod.ts), which reproduces a retirement's turbo footprint exactly. Mutation proven on disk each lap (marker_count=1, hash-object differs from the HEAD blob); restored by an EXIT INT TERM trap using absolute paths and git checkout HEAD --; restore proven by git diff HEAD empty and hash-object equal to the HEAD blob. ⛔ Nothing real retired.

Sets — identical, member by member

reading	value
all packages (turbo ls)	79
affected (TURBO_SCM_BASE=BASE turbo ls --affected, dry edit present)	75
consumer closure (pnpm --filter '...@objectstack/spec')	75
diff of the two sorted member lists	IDENTICAL SETS
control — affected set on a clean tree	0 no packages (lit)
the 4 packages NOT affected	@objectstack/console, @objectstack/refd-timer-testkit, @objectstack/sdui-parser, create-objectstack — i.e. exactly the four that do not consume spec
--dry=json task graphs of both command forms (docs excluded)	78 tasks each, diff IDENTICAL TASK SETS

Timings — all under scripts/pm/os-verify-lock.sh (slot issue-16389), waited 0s each, VERDICT lines quoted; shared-box seconds, as the lock's own banner requires me to say

lap	command	tasks	cold wall
the playbook's own §5 line	pnpm --filter @objectstack/spec build	1 package	187s (3m07s)
affected-scoped walk	TURBO_SCM_BASE=BASE turbo run build --affected --filter='!@objectstack/docs'	72	491s (8m11s)
full consumer closure, forced cold	turbo run build --filter='...@objectstack/spec' --filter='!@objectstack/docs' --force	72	491s (8m11s)
full closure, unforced, run straight after the affected lap	same, no --force	72	0.187s — 72 cached, 72 total >>> FULL TURBO

Three readings, one conclusion. The two forms are 491s vs 491s — equal to the second. The unforced closure lap replayed 72 of 72 tasks from cache, which is hash-level proof that the affected-scoped walk left the closure form nothing to build — no coverage lost, and no time saved, because there was none to save.

4 · Where the six minutes actually comes from

1. The dev's own command, from the dev-agent file — not the playbook. The only 8-minute command here is a downstream consumer sweep, and that instruction lives at .claude/agents/os-dev.md:185: 「消费者清扫的 filter 方向是前缀:pnpm --filter '...@objectstack/PKG' = 下游消费者」. The playbook never mentions it.
2. "dependency closure" is the wrong direction for this card. @objectstack/spec is the graph root — 0 workspace deps. The expensive direction is downstream, and downstream of spec is 75 of 79 packages. Any walk of spec's consumers is a near-whole-repo build by construction, which is why affected-scoping cannot cut it.
3. The only genuinely wider form is an unfiltered pnpm build (78 packages = 79 minus docs). Affected-scoping it recovers exactly the four non-consumer packages above. That is the "a turbo run build without a filter" limb of ruling item 1 — a habit, at a ceiling of 4 packages of savings.

The finding's own words — 「an affected-scoped walk … would cut it without losing coverage」 — are half right and half wrong in a way only a measurement separates: no coverage is lost (true, and now proven at the task-hash level), and nothing is cut (false).

5 · Noted, not filed (threshold #16351)

The playbook's §5 mislabels check:skill-examples. Lines 320-323 name five audits 「五个纯源码审计 … 它们没有生成器,那里的失败是真发现、不是过期产物」, and line 297 mandates 「整组跑,永不单点」. But packages/spec/scripts/check-skill-examples.ts reads the built dist/*.d.ts and resolves packages/client-react (its own header: 「Because it reads the built dist/*.d.ts, this runs AFTER the workspace build」, CLIENT_REACT_DIR at :291). On a workspace where only spec is built it exits 3 PREREQUISITE NOT MET — which the playbook's sentence would have an executor read as a real finding, and whose natural repair is a wider build. The #15513 dev hit precisely this (「3 NOT MEASURED … check:skill-examples: packages/client-react/dist unbuilt」). Doc-text accuracy, not a runtime defect ⇒ under #16351 clause 2 this rides as a note, not a card; near neighbours #15285 / #15328 cover the exit-code half for other gates. The seat files it at ACCEPT if it grades it class (a).

{
  "issue": 16389,
  "status": "done",
  "branch": "claude/issue-16389-retirement-walk-affected-scope",
  "pr": null,
  "premise_still_valid": false,
  "summary": "The card's premise is falsified and no PR was opened, per ruling item 1. The spec-property-retirement playbook prescribes no dependency-closure rebuild: its complete command inventory is six pnpm lines, of which exactly one is a build (line 314, `pnpm --filter @objectstack/spec build`, one package), with zero `turbo run` and zero unfiltered builds; its consumer-walk language (lines 75-77, 283) names `tsc` as a mechanism with no command and no scope, and the single place it does scope a consumer run scopes it to a targeted suite (line 305). Rule #15566, which the finding names as the cost's source, prescribes a declaration in `cross-package-test-inputs.mjs` and `turbo.json`, not a run — its cost is turbo cache invalidation of `@objectstack/spec#test` on later PRs, which is what the #15513 dev's own open question recorded. The cost comes from the downstream consumer sweep prescribed in `.claude/agents/os-dev.md:185`, not from this file, and it is irreducible: `@objectstack/spec` has 0 workspace dependencies (the upstream closure is empty) while 75 of 79 packages consume it, so the affected set IS the consumer closure. Adding the candidate rule would also be a second copy of `os-dev.md:129`, which the playbook's own section 6 forbids by name.",
  "tests": "No file was edited; `git status` clean, 0 commits beyond BASE 0374bcba9f, and `node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack` answers with its own line 'this branch changes nothing against origin/main (merge base 0374bcba9) — nothing to derive' (exit 2, captured before any pipe), so no gate family is owed and the playbook's 337/337 line pin and row pin are untouched by construction. MEASUREMENT (the card's acceptance), all heavy laps through `bash scripts/pm/os-verify-lock.sh` slot `issue-16389`, each VERDICT line read, each 'waited 0s': (1) playbook's own line `pnpm --filter @objectstack/spec build` — VERDICT command-exit 0, held 187s (3m07s), 1 package; (2) affected-scoped walk `TURBO_SCM_BASE=BASE pnpm exec turbo run build --affected --filter='!@objectstack/docs'` — VERDICT command-exit 0, held 491s (8m11s), turbo 'Tasks: 72 successful, 72 total / Cached: 1 cached'; (3) full consumer closure forced cold `turbo run build --filter='...@objectstack/spec' --filter='!@objectstack/docs' --force` — VERDICT command-exit 0, held 491s (8m11s), 'Cached: 0 cached, 72 total'; (4) the same closure command unforced immediately after lap 2 — 1s, 'Tasks: 72 successful / Cached: 72 cached, 72 total / Time: 187ms >>> FULL TURBO', i.e. the affected walk left the closure form nothing to build (hash-level proof of no coverage difference). SETS: `turbo ls` 79 packages; `TURBO_SCM_BASE=BASE turbo ls --affected` 75; `pnpm --filter '...@objectstack/spec' list --depth -1 --json` 75; `diff` of the two sorted member lists prints nothing — IDENTICAL SETS; the 4 non-affected are console, refd-timer-testkit, sdui-parser, create-objectstack, exactly the packages that do not consume spec; `--dry=json` graphs of both command forms are 78 tasks each and diff clean (IDENTICAL TASK SETS). Lit control: on a clean tree the same affected command prints '0 no packages', so the 75 is the dry edit's doing. DRY WALK, not a real retirement: the instrument is one comment appended to packages/spec/src/data/document.zod.ts; each lap proved the mutation on disk before measuring (marker grep count 1 AND `git hash-object` 0b7ed38ff… differing from the HEAD blob a4906b326…), restored via an `EXIT INT TERM` trap running `git checkout HEAD -- REL` with an absolute REPO_ROOT, and the restore was proven by observation, not exit code: `git diff HEAD` empty and `git hash-object` back to a4906b326…, verified after every lap. Shared-box caveat stated as the lock's banner requires: the lock excludes other locked runs only, so these are shared-box seconds; the load-bearing comparison is the 491s/491s RATIO and the 72/72 cache replay, both of which survive contention.",
  "mcp_calls": "5 — five MCP GitHub reads (issue 16389 get + comments, 16104 get, 15513 get + comments) taken before the REST probe; every read and the report write after it went through the repo-scoped REST channel, which probed HTTP 200 (`gh` is absent in this container). Dedup for the noted item: REST listing of all 667 open issues, 7 pages, grepped locally.",
  "open_questions": [
    {
      "question": "The measurement kills the remedy but not the cost: a spec retirement's consumer walk is ~8 minutes on this box and cannot be scoped away, because `packages/spec` is the graph root (0 workspace deps, 75 of 79 packages consume it). What, if anything, should change?",
      "options": [
        "A — close #16104 section 3.2 as measured-and-refuted, change no file: the playbook prescribes no closure rebuild, os-dev.md:129 already carries the affected rule, and affected-scoping a spec walk saves 0 of 72 tasks.",
        "B — A, plus one line in `.claude/agents/os-dev.md` (NOT this playbook) recording that for a graph-root package the affected set equals the consumer closure, so no one re-derives this measurement.",
        "C — attack the real cost instead: the walk rebuilds 72 packages to run a `tsc`-shaped absence sweep, and `turbo run typecheck --affected` (dependsOn ^build, so still a build) is the same shape. A cheaper instrument for 'did any consumer keep authoring the key' is the tree-scoped absence pin the playbook already mandates at lines 274-281, which needs no build at all."
      ],
      "recommendation": "A, with B as a cheap rider if the seat wants the reading preserved. On 实际业务需求: the remedy serves no measured need — the 8 minutes is irreducible for this card class, so the rule would buy nothing while costing 337-line-pinned budget. On 长远合理性: adding it to this file is a second copy of os-dev.md:129, which the playbook's own section 6 forbids by name; the anchor discipline is the sustainable shape. On 防 AI 写错: a rule that reads as a saving but delivers none is exactly the kind of text that teaches an executor to trust a scope narrowing it has not measured — the opposite of the tightening this axis wants; the honest reading (affected == closure, so narrow nothing) is the safer instruction. On 创业阶段不扩散: no user is pulling on this, the inflow is fleet-manufactured, and #16351 has just ruled inflow down — the cheapest correct move is to close the item with the measurement attached. C is a genuinely different (and possibly better) card, but it is a new question about the walk's instrument, not this card, and it needs its own ruling. ⚠ Note on B: it touches `.claude/agents/os-dev.md`, outside this card's declared one-file surface, so it needs the seat's dispatch rather than this branch."
    }
  ],
  "out_of_scope_findings": [
    "noted, not filed: the playbook's section 5 (lines 320-323) calls `check:skill-examples` one of five 纯源码审计 with 没有生成器 whose failures are 真发现, and line 297 mandates running the group whole — but `packages/spec/scripts/check-skill-examples.ts` reads the built dist/*.d.ts and resolves packages/client-react (its header says it runs AFTER the workspace build; CLIENT_REACT_DIR at :291), so on a spec-only build it exits 3 PREREQUISITE NOT MET, which that sentence would have an executor read as a real finding. Measured by the #15513 dev too (3 NOT MEASURED, client-react/dist unbuilt). Doc-text accuracy rather than a runtime defect ⇒ #16351 clause 2 note, not a card; #15285 and #15328 already cover the exit-code half for sibling gates.",
    "noted, not filed: #16104 section 3's first mechanizable item (os-dev step ① is vacuous off packages/spec/src) is corroborated exactly here — `pnpm --filter '@objectstack/spec^...' build` resolves to '0 no packages' because spec has 0 workspace dependencies. Already filed as part of #16104; no second card."
  ]
}

---

Generated by Claude Code

[claude]

Premise falsified by measurement — closed not planned (skills seat, 2026-09-06T17:3xZ); ruling A on the dev's question

Release: session_019RfFHiRCSs3JXLK4cwcfox · 因 premise falsified · 去向 closed not_planned, no PR (a legal terminal per the dev contract).

Flight Q's reading, verified by the seat against the report's own numbers: the spec-property-retirement playbook prescribes NO dependency-closure rebuild — its command inventory holds one single-package build (pnpm --filter @objectstack/spec build) and zero turbo run; the six-minute laps the spec seat measured come from the downstream consumer sweep the dev contract prescribes (os-dev.md, the consumer sweep line), and that cost is irreducible for this card class because packages/spec is the graph root — 0 workspace dependencies, 75 of 79 packages consume it — so the affected set IS the consumer closure. Measured both ways on this box through the verify lock with a dry edit (a comment appended to a spec file, restored by hash): affected-scoped build 491 s / 72 tasks, forced cold closure build 491 s / 72 tasks, member sets identical (turbo ls --affected 75 = pnpm list '...@objectstack/spec' 75, diff empty), --dry=json task graphs identical (78 = 78), and the closure form replays 72/72 from cache after the affected run. Rule #15566 prescribes a declaration (cross-package-test-inputs.mjs + turbo.json), not a run.

• A — close, change no file: the candidate rule would save 0 of 72 tasks and would be a second copy of the dev contract's affected rule, which the playbook's own section 6 forbids by name.
• B (record that a graph-root package's affected set equals its consumer closure) and C (a cheaper instrument for the consumer sweep — the tree-scoped absence pin the playbook already mandates, which needs no build) are recorded as riders on os-dev: the "build spec" verification step passes without measuring anything off packages/spec/src — condition it on the file surface or replace it by the dispatch-gates derivation (#16104 section 3.1) #16388, the open os-dev.md card on the neighbouring step; neither is a card of its own under the admission threshold.

The two noted, not filed items stand as notes: the playbook's section-5 sentence that reads a check:skill-examples prerequisite exit as a real finding (a doc-accuracy note, the exit-code half already covered by #15285 / #15328), and the corroboration of #16388.

---

Generated by Claude Code