agents(os-dev): narrow three test obligations — one-time ablation proofs, import-side tests only on public-surface change, no wording pins

[os-steve]

Fixes #16456

Governed surface (.claude/**): draft PR, human merge. One file, .claude/agents/os-dev.md, line-neutral at 403/403: every added line is paid by a fold inside the same block, every line is at or under 120 bytes (LC_ALL=C awk, measured max 120, zero over), and the diff stays in the rules-only register (no dates, quotations, issue numbers, narrative or model names). No other file moves — the measurement below says why.

Three test obligations in the dev contract are narrowed (maintainer-directed, verbatim: 「我想的是测试会不会太多，是否都是必要的，是不是应该砍，每次修改都要完整的测试吗」 → 「同意，这几个都处理，你现在就负责派发」). Line numbers: 改前 = origin/main at 6eba38f5a3, 改后 = ddb4213508. In the quoted lines the two angle-bracket placeholders are spelled YOUR_BRANCH and PATH; the file keeps its own spellings — the GitHub sanitizer eats angle-bracket fragments in bodies.

① Reverse verification and ablation are one-time proofs — os-dev.md L210–214 → L210–213

where	改前	改后	paid by
L210	反向验证(回退修复,看诊断变化)先 commit 修复:恢复只是 git checkout YOUR_BRANCH -- PATH。	反向验证与消融是一次性证明:前后运行引在 PR 正文与报告里;⛔ 不留永久测试文件。	the block's two commit-first lines (old L210 + old L213) fold into new L211
L211	对着未提交的编辑,git checkout origin/main -- PATH 不留任何恢复点。	两者都先 commit 修复再回退或变异:恢复腿指向 HEAD,HEAD 必须先装着你的实现。	old L210's restore spelling YOUR_BRANCH yields to HEAD, which hard line ① (L217) already mandates
L212	恢复机制与字节一致性证明规则见 AGENTS.md;从已 commit 的状态重跑,红/绿数字才可信。	未提交时 git checkout origin/main -- PATH 不留恢复点,git checkout HEAD -- PATH 删的正是实现。	the block's two uncommitted-hazard lines (old L211 + old L214) fold into this one; 「事后检查全绿」 (consequence narrative) deleted
L213	消融同一条,先 commit 再变异:恢复腿指向 HEAD,HEAD 必须先装着你的实现。	恢复机制与字节一致性证明规则见 AGENTS.md;从已 commit 的状态重跑,红/绿数字才可信。	old L212, unchanged, moved down one
L214	对未提交的实现,一次完美的 git checkout HEAD -- PATH 删的正是实现本身,事后检查全绿。	—	folded into new L212

No longer owed: a fixture, spec or self-test case added to the tree to make a reverse-verification or ablation proof permanent — the legs stay mandatory (the how-to at L188 and L214–260 is untouched), and the evidence is the quoted before/after run in the PR body and the report.

② Import-side tests only when a public surface changes — os-dev.md L129–130

where	改前	改后	paid by
L129	受影响包 = CI 会测的包,清单读 TURBO_SCM_BASE="$BASE" pnpm exec turbo ls --affected。	受影响包 = 本包;import 方只在公开面变化时欠测试:spec 契约、发布的 exports、线上形状。	in place — this line was the definition that made L130 operative (affected = CI's --affected graph, importers included), so it is the echo the ruling names
L130	⛔ 不按改了哪些包猜:普通 import 被改模块的包也在清单里,欠它们的测试。	公开面字节不变 ⇒ 只欠本包测试与派生门禁,⛔ 不给每个 import 方补测试。	in place

No longer owed: a test in — or a local test run of — every package that merely imports the changed module while every public surface (a packages/spec contract, a published exports entry, a wire shape) stays byte-identical; own-package tests and the derived gates remain owed. When a public surface does change, importers still owe tests, and the consumer-sweep filter at L185 (pnpm --filter '...@objectstack/PKG') remains the enumerator — the turbo ls --affected spelling leaves with the CI-set definition it served.

③ No pin tests on wording — new os-dev.md L227 (the measurement found no prior carrier)

where	改前	改后	paid by
L227 (new)	—	提示、裁决与错误文案 ⛔ 不 pin,除非消费者解析其原文;断言种类、退出码或具名主体。	the one line freed by ①'s fold (5 → 4)
L228	措辞本身即契约的地方,在 code+status 之上再断言 message 首句,永不取而代之。	unchanged	it is the mechanics of the exception (wording that IS the parsed criterion), read directly under the rule

No longer owed: a pin test on a remedy string, a verdict line, a prompt or an error message's prose; the assertion is on the kind, the exit code or the named subject, and wording is pinned only where a consumer parses it (L228 keeps the code+status-first mechanics for that case).

Measurement — every carrier of the three obligations, by file and line

First pass grep -n -E '欠|import|反向验证|消融|pin|文案|原文' over the five named files, second pass a read of each hit's neighbourhood, third pass a wider grep (受影响|affected|下游|消费者|清扫|判定行|message|wording|fixture|snapshot|措辞|永久) over .claude/agents/os-dev.md, all of .claude/skills/pm-dispatch/ (references and lanes included), AGENTS.md, .claude/skills/dogfood-verification/SKILL.md and the published skills/objectstack-pm-dispatch/.

Moves (all in os-dev.md): L129–130 (②), L210–214 (① — the block gains the one-time-proof rule and folds 5 → 4), new L227 (③).

Read true under the ruling, left in place:

• os-dev.md L185–187 — consumer-sweep direction on a contract tightening: the public-surface case of ②.
• os-dev.md L188 — cross-package type reverse verification: already one-time (「确认转红,再恢复」).
• os-dev.md L225–226, L228 — rejection tests assert code+status; wording only where it is itself contract: ③'s direction.
• os-dev.md L236–237 — fixture sweep by a narrowed rule's consumer radius: a rule narrowing is a public-surface change (②).
• os-dev.md L265–271 — Definition of done test clause and the two packages/cli tier sentences (L131, L266): 「受影响包」 now reads through L129; both read true unchanged.
• os-dev.md L369 — report tests field (「ablation: rebuild + on-disk mutation proof」): the quoted-run evidence of ①.
• references/core-rules.md — no carrier (L61 pin is a version pin; L150 is the report contract).
• references/review-checklist.md L31 (real command + output as test evidence), L53–58 (rejection assertions on code/status), L59–62 (N-package sweep evidence only on a contract tightening) — consistent with ①, ③, ②; no wording-pin demand anywhere in the file.
• references/contract-review.md — no carrier.
• SKILL.md L182–183 (behaviour pins reversed on a contrary fact — behaviour, not wording), L546 (repo-wide pin sweep on a public-semantics flip: the public-surface case), L548 (code+status minimum), L596 — consistent; no line moves.
• references/lanes/services.md L17–18 — 「每条否定性 pin 都要消融验证:被禁行为放回、测试转红、报失败输出、恢复」: already the one-time form of ① (report the failing output, restore).
• references/lanes/spec.md L38–39 — contract-surface cards run consumer-package tests: the public-surface case of ②.
• references/dispatch-runbook.md L223–225 — the pin sweep on a semantics flip; the grep of error messages is the search instrument for existing pins, and L224 asserts substance, not prose.
• AGENTS.md L364–371 — commit-first for reverse verification; consistent with ①.
• published skills/objectstack-pm-dispatch/rules/dev-template.md L71–72 (generic 「run the affected packages'」 with no importer definition) and L95–96 (wording only where it is itself contract) — no echo; skills/** untouched, so the two-readings clause does not apply.

The wording-pin obligation (③) had no carrier in the corpus — it was a habit; rule ③ is added as a fold, next to its nearest neighbour L228.

Line pins and consistency

• check-skill-line-ratchet live: ✓ check-skill-line-ratchet: .claude/agents/os-dev.md is 403 lines (ceiling 403; headroom 0). — no ceiling touched; self-test 155 cases pass.
• Every line ≤ 120 bytes: LC_ALL=C awk over the file → max=120 over120=0 (the file already carried 120-byte lines).
• On-disk proof of the fold (not the editor's exit code): grep -c of six removed anchors → 0 each; six inserted anchors → 1 each.
• git merge-tree --write-tree HEAD origin/main → exit 0, 0 CONFLICT lines against origin/main at 245c6a23cd (tree 4c4791bb1c).

Gates (union run at ddb4213508, the final commit; every exit captured before any pipe)

Derived with node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack (no path passed; change set from the merge base 6eba38f5a: 1 path). Reconciled: ✓ dispatch-gates --ran: 16 derived famil(ies) accounted for — 16 run, 0 NOT-MEASURED. (exit 0).

• pnpm check:pm-skill-ratchet exit 0 — the two lines quoted above; declared cross-file moves: 1, total ceilings down 9 lines (main's state, untouched).
• pnpm check:pm-skill-id-lint exit 0 — ✓ check-skill-id-lint: 26 file(s) clean (pattern /#[0-9]{3,}/g).
• pnpm check:skill-frame-sync exit 0 — 4 axes … binding sentence present in all 2; 4 count mention(s) agree; 75 markdown files scanned for undeclared copies.
• node scripts/pm/check-clause2-carriers.mjs --self-test exit 0 — ✓ check-clause2-carriers self-test: 190 cases pass.
• node scripts/pm/check-governed-merges.mjs --test .claude/agents/os-dev.md → exit 3 — One hit governs the whole PR — 「混合 diff 一条命中即整 PR 分叉」; proportion is not a question. .claude/** ×1.
• pnpm check:pm-governed-merges exit 0 — self-test 274 assertions.
• pnpm check:nul-bytes exit 0 — scanned 8078 text file(s) … no raw ASCII control bytes.
• pnpm check:agent-model-declared exit 0 — 0 justified inherit(s); no definition leaves its tier to the dispatching session.
• pnpm check:agent-test-spelling, check:doc-authoring, check:driver-memory-census, check:refd-timer-probe, check:watch-hint-literal, check-closing-keyword-parity (+ self-test), check-comment-mask-corpus, check-governed-queue-guard --self-test — all exit 0.
• pnpm --filter @objectstack/lint run check:doc-formula-expressions — first run exit 3 PREREQUISITE NOT MET — the workspace package @objectstack/formula is not built (nothing measured); prerequisites built under the verify-lock, re-run: exit 0 — ✓ check:doc-formula-expressions: 22 record-scoped formula example(s) across 430 files / 1371 TS blocks judged clean by @objectstack/formula.
• Whole-repo pnpm lint through bash scripts/pm/os-verify-lock.sh (slot issue-16456, one acquisition, full 540s budget): os-verify-lock: VERDICT command-exit 0 · held the lock 89s (1m29s) · waited 0s (shared-box seconds).

No ablation applies — the diff is instruction text; the proof is the gates above.

维护者速读(草稿)

• 改了什么:dev 契约(.claude/agents/os-dev.md)里三条测试义务收窄:反向验证与消融只是一次性证明、引在 PR 与报告里,不再留永久测试文件;只有公开面(spec 契约、发布的 exports、线上形状)变化时,import 方的包才欠测试;提示、裁决与错误文案不做 pin 测试,除非消费者解析其原文。403 行上限不动,只折不加。
• 为什么改:您的判断「测试会不会太多」有实测支撑——3,388 个测试文件对 2,183 个源文件,每天新增 35–89 个,而这三条义务正是契约自己在要求这个增速。
• 风险与代价(含回滚):风险是 import 方的回归少了一道本地预检,但 CI 仍跑全套、公开面一变 import 方照旧欠测试;回滚只需 revert 这一个 commit。
• 席位意见:
• 你要做的:确认三条新规则的措辞是您的本意(尤其②「公开面」的三项定义),然后人工合并。

---

🤖 Generated with Claude Code

https://claude.ai/code/session_019RfFHiRCSs3JXLK4cwcfox

---

Generated by Claude Code

[claude]

Verdict: ACCEPT (governed — to the approvers)

Implemented-by: os-dev flight Y (fable), branch claude/issue-16456-dev-contract-test-obligations, head ddb4213508
Reviewed-by: pm-dispatch skills seat (contract-tier read in a detached compare worktree at ddb4213508, merge-base 6eba38f5a3)

What was checked

• One file, .claude/agents/os-dev.md, +6/−6, 403/403 lines with every line at or under 120 bytes; rules only, no issue numbers or provenance in the changed lines.
• ② (import-side tests): lines 129–130 now read 「受影响包 = 本包;import 方只在公开面变化时欠测试:spec 契约、发布的 exports、线上形状。」 and 「公开面字节不变 ⇒ 只欠本包测试与派生门禁,⛔ 不给每个 import 方补测试。」 — the ruling's substance; the consumer sweep at line 185 stays the enumerator for the case where a public surface does change.
• ① (ablation as a one-time proof): the commit-first block folds 5 → 4 lines and opens with 「反向验证与消融是一次性证明:前后运行引在 PR 正文与报告里;⛔ 不留永久测试文件。」; the recovery mechanics (commit first, HEAD restore leg, the uncommitted-edit trap) survive in the three lines below it.
• ③ (no wording pins): new line 227 「提示、裁决与错误文案 ⛔ 不 pin,除非消费者解析其原文;断言种类、退出码或具名主体。」, placed beside the existing rule that a message's first sentence is asserted on top of code+status only where the wording is the contract — the two read as one rule. The measurement found no prior carrier of this obligation in the corpus (a habit, not a line), so it is an addition paid by the fold above.
• Carriers outside the five named files (lanes/services.md, lanes/spec.md, dispatch-runbook.md, AGENTS.md, the published dev-template) already read true under the ruling and are listed with line numbers in the PR body; the Definition of done's test clause and both cli tier sentences read true unchanged.
• Gates in the compare worktree: ratchet live green at 403/403 (self-test 155), check-governed-merges --test exit 3 (governed), check:skill-frame-sync and check-clause2-carriers green; merge-tree --write-tree against origin/main clean. CI on the head: 17 success / 12 skipped / 0 failure.

Rulings

• No open question. The dev's three named deviations are accepted as stated: the harness-written co-author trailer on the commit (the AGENTS.md exemption), the 速读 draft in os-dev.md's five-heading form, and keeping the consumer-sweep filter as the enumerator.

Landing

Governed surface: reviewers os-zhuang and hotlong requested, needs-user-decision added, the final 维护者速读 follows this comment. Human merge; the seat never flips ready or queues a governed PR. On merge the seat strips pm:dispatched and clears the assignee on #16456, then dispatches #16388 (queued behind this file).

---

Generated by Claude Code

[claude]

维护者速读(skills 席,2026-09-07T03:1xZ)

事情:按您 02:0xZ 的指示(「同意，这几个都处理，你现在就负责派发」的 C 项),dev 契约 .claude/agents/os-dev.md 收窄三条测试义务,一个文件、+6/−6、403 行钉子不动:① 反向验证与消融只作一次性证明(前后运行引在 PR 正文与报告),⛔ 不再留永久测试文件;② import 方的包只在公开面(spec 契约、发布的 exports、线上形状)变化时欠测试,公开面不变只欠本包测试与派生门禁;③ 提示/裁决/错误文案不做 pin,除非消费者解析其原文。

为什么:测试文件 3,388 对源码 2,183,每天新增 35–89 个,增长是契约义务的函数;抽样 145 个新测试里 import 方补测的只有 6 个,所以 ② 的直接效果小,① 与 ③ 是主要来源。三条都不动「怎么做验证」,只动「留不留下来」。

风险:低。契约收窄意味着某些跨包回归改由合并队列的 CI 捕获(受影响集 + 跨包扫描),而不是由 dev 提前补测;公开面变化时的 import 方义务仍在。文本层面 ratchet / frame-sync / clause-② 载体检查全绿。

要您做什么:一个决定——合并或退回。审批人 os-zhuang / hotlong 已被请求复审。

席位意见:建议合并。三条与您对测试总量的判断一致;若想保守,可只合 ① 与 ③、把 ② 退回(我们会拆成单行 PR)。

---

Generated by Claude Code