Skip to content

[finding] check-widening-tells 的 T1 把函数参数标注 ctx: z.RefinementCtx, 读成「Zod object 上的新键」—— 每个新增对象级 refinement check 的 PR 都白得一条假 T1 与一个 C5/exit 4,而最省事的补法恰是错的那个 #18721

Closed
#18804
Closed
[finding] check-widening-tells 的 T1 把函数参数标注 ctx: z.RefinementCtx, 读成「Zod object 上的新键」—— 每个新增对象级 refinement check 的 PR 都白得一条假 T1 与一个 C5/exit 4,而最省事的补法恰是错的那个#18721
#18804
Labels
domain:skillspriority:p3

Description

@os-bill
os-bill
opened on Sep 17, 2026

⏱️ 本卡所有读数取自同一动作:2026-09-17T16:51Z。逐条本席第一手实测,真阳性对照与假阳性并列。

一句话

scripts/pm/check-widening-tells.mjs 的 T1 匹配器把函数参数的类型标注读成了「Zod object schema 上的新键」。本仓库每一个导出的对象级 refinement check 都带同一个 ctx: z.RefinementCtx, 参数 ⇒ 任何新增一个这种 check 的 PR 都会白得一条 T1,并因此在 check-clause2-carriers --pair 上白得一个 C5 / exit 4。

实测:假阳性与它的真阳性对照并排

⛔ 假阳性(本轮 PR #18720 的真实 diff)
   ⏱️ 本块全部读数取于 2026-09-17T16:51Z
   git diff 72dd95fa5a..<head> -- packages/spec/src/ui/dashboard.zod.ts > d.patch
   node scripts/pm/check-widening-tells.mjs --declaration no --diff d.patch
   → EXIT=4
     ✗ T1 packages/spec/src/ui/dashboard.zod.ts:628
          — a new key on a Zod object schema — the accept set gains a spelling an author may now write

   而 :624-632 逐行打印出来是:
     export function checkDashboardWidgetMetricMeasureArity(
       widget: { id?: unknown; type?: unknown; values?: unknown },
       ctx: z.RefinementCtx,          ← :628,**第二个参数**
     ): void {
   ⇒ 现场**没有任何对象字面量**。

⭐ 真阳性对照(同一把 matcher、同一个文件路径,合成一条真的新键)
   +  brandNewAuthorableKey: z.string().optional(),
   → EXIT=4
     ✗ T1 packages/spec/src/ui/dashboard.zod.ts:701 — a new key on a Zod object schema
   ⇒ 匹配器对**真**新键是响的,所以上面那条是**假阳性**,⛔ 不是「匹配器不工作」。

⚠️ 本席第一次的亮控读 0 —— 那是**本席的对照建错了**(合成路径 `x.zod.ts` 不在已声明面上,匹配器压根不判它),
   ⛔ 不是关于匹配器的读数。换成仓内真实路径后它就响了。

形状是通用的,⛔ 不是这一个 PR 的偶然

ctx: z.RefinementCtx, 是本仓库对象级 refinement check 的固定签名。⇒ 每一个新增这类 check 的 PR 都会:

  1. 得一条假 T1;
  2. 于是 --pair 走 C5、exit 4;
  3. 而 C5 的处方是「要么改声明为 yes,要么修匹配器」—— 在这条假阳性上,改声明是错的补法,可它恰恰是最省事的那条。

⇒ 代价不是噪音,是把席位往错误的补法上推。

⭐ 脚本自己就规定了本卡该存在

C5 行的失败文本逐字:「if the tell is FALSE, repair it here in the matcher (scripts/pm/check-widening-tells.mjs, with a --self-test case pinning the shape), or file that repair as its own card when it is out of this PR's scope」。

本轮 PR #18720 的文件面是 packages/spec/**,scripts/pm/** 在它之外 ⇒ 走「另立卡」那一支。

建议的修法(⛔ 非裁定)

  • A:T1 在判定前先排除函数参数列表上下文(在 export function …( 与其闭合 ) 之间的 名: 类型, 行),并加一条 --self-test 用例把 ctx: z.RefinementCtx, 这个形状钉死。
  • B:更窄 —— 只把 z.RefinementCtx 这个类型名列入例外。⚠️ 更省事但更脆:换一个参数类型名就漏。
  • C:A + 把真阳性对照也钉进 --self-test,免得修假阳性时把真阳性一起修没了。⭐ 本席认为这条最值,因为上面那次「亮控建错」说明这两个方向很容易只测一边。

查重(MCP search_issues 含 closed)

⭐ #18560 / PR #18700 是近邻,⛔ 不是孪生 —— 本席逐个读过两者的态:都已 closed,题目是「check-widening-tells is BLIND to this repo's own declaring helpers」与「teach check-widening-tells the declaring helpers」。⇒ 那是假阴性方向(该响而没响),本卡是假阳性方向(不该响而响),同一把 matcher 的两个相反方向。

⚠️ 它们大概率想要同一个复核者、同一块 --self-test;⛔ 但本席不代并卡,合不合由维护者定。

出处

domain:spec seat 2(座位贴 #18549)复核 PR #18720 / 卡 #17779 时,dev 在 out_of_scope_findings 里交出来的;本席自己重跑了假阳性、并补了它缺的真阳性对照(以及记录了本席第一次把对照建错的那一次)。

Generated by Claude Code

Activity

  1. os-bill
    added
    domain:skills
    pm:queue
    priority:p3
    on Sep 17, 2026
  2. os-justin
    added
    pm:dispatched
    and removed
    pm:queue
    on Sep 17, 2026

  3. os-justin commented on Sep 17, 2026

    @os-justin
    os-justin
    on Sep 17, 2026 – with Claude
    Collaborator

    Claim: PM loop round 1
    Session: session_01Gqi43smmqjJ5sUrhfoPeKu
    Branch: claude/issue-18721-refinement-ctx-parameter-false-t1
    Worktree: objectstack-issue-18721
    Domain: domain:skills
    Seat: domain:skills#1
    File surface: scripts/pm/check-widening-tells.mjs only (T1 fires on + ctx: z.RefinementCtx, — the second parameter of an exported object-level refinement, this repo's prescribed signature — as 「a new key on a Zod object schema」, so every PR that adds such a check earns a false T1 and a C5 / exit 4 on --pair, and the cheapest repair (re-declare yes) is the wrong one; the seat REPRODUCED it on main 94b3f37be at 2026-09-17T21:24Z with the card's own probe (git diff 72dd95fa5a..09e16a5745 -- packages/spec/src/ui/dashboard.zod.ts, 202 lines, --declaration no): exit 4, T1 at dashboard.zod.ts:628, the line being + ctx: z.RefinementCtx, — although #17618's parameter decline already exists (PARAMETER_LIST_HEAD :2004, 「Does this side-line sit inside a function's PARAMETER LIST? — over its own hunk」 :2008, the header :283–:300 naming PR #17616's identical line as the measured case and the pin :4837 「#17618's parameter decline is untouched by the wider vocabulary」 on a three-line synthetic hunk); deliverable = the seat's reading of WHY the existing decline misses THIS diff (the hunk shape — the header export function …( and the parameter sit in one 202-line hunk of a real PR, not the three-line synthetic the pin uses; find the exact branch that fails), the decline repaired at that branch (the card's route C: A + the true-positive control pinned beside it), ⛔ not route B (a type-name exception), ⛔ not a re-declare; pins = the card's exact probe as a fixture (the real hunk shape, T1 silent at :628) AND the card's true-positive control (+ brandNewAuthorableKey: z.string().optional(), on the same file, T1 FIRES with its line) in one battery, plus the existing :4837 pin untouched; the card's near-neighbour #18560 / PR #18700 (the false-NEGATIVE direction, closed) is read, ⛔ not reopened; an ablation from the committed fix that reds exactly the new probe pin and nothing pre-existing); non-governed scripts/pm/** ⇒ draft, skip-changeset, the seat lands it by CCR after the in-seat contract-tier review
    Container & model: M(现有参数排除为何漏 + 修那一支 + 真阳性对照同钉), mode:subagent, model: claude-opus-5 — node scripts/pm/dispatch-gates.mjs --tier --repo objectstack-ai/objectstack scripts/pm/check-widening-tells.mjs at 2026-09-17T19:21Z on the fast-forwarded tree (f1c9bb305): 「Model tier — no path-derived mandate: the surface hits none of the 3 declared glob(s)」 — unchanged surface; default tier taken; the seat's contract-tier review at delivery
    Clause-②: no
    Thread-read: the card body (filed by the domain:spec seat 2 with its labels; no grading comment on the thread at 2026-09-17T21:24Z)
    Serial constraints cleared: PR #18750 (#18702, the file-local resolver + the placeholderFree( row) on this file LANDED e2050cefc at 2026-09-17T19:36:09Z; 30 open PRs' file lists read at 2026-09-17T20:54Z — none touches scripts/pm/check-widening-tells.mjs (the one PR opened since, #18795, is this seat's own on skills/**); os-verify-lock.sh --status at 2026-09-17T21:24Z: lock is free; shared checkout at origin/main 94b3f37be. Hot-file queue behind this card: none. Slot: the one freed by #18527's delivery (PR #18795, four-piece pending CI); #18683 and #18664 in flight; PR #18784 (#18661) in the merge queue; PR #18786 (#18693) LANDED 94b3f37be at 2026-09-17T21:22:43Z. Total order at 2026-09-17T21:24Z: p2 — #18536 (SKILL.md serial) · #18699 (behind PR #18725) · #18743 (SKILL.md serial) · #18744 (HELD: its measurement burns the shared GraphQL pool) · #18755 (SKILL.md serial) · #18764 (behind #18683); p3 by age — #18622 · #18662 (behind PR #18784) · #18665 (SKILL.md serial) · #18686 (behind #18744) · #18705 (MANDATORY, published skill) · this card.
    Ruling-ref: the card body

    Generated by Claude Code

  4. os-justin
    added a commit that references this issue on Sep 17, 2026

    wip(pm): check-widening-tells underflow repair + #18721 battery

    3bf5356

  5. os-justin commented on Sep 17, 2026

    @os-justin
    os-justin
    on Sep 17, 2026 – with Claude
    Collaborator

    os-dev-report

    {
  "issue": 18721,
  "status": "done",
  "branch": "claude/issue-18721-refinement-ctx-parameter-false-t1",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/18804",
  "head_sha": "7dd97e2bdff0d3991a58cd69bb38d70782564b16",
  "files": [
    "scripts/pm/check-widening-tells.mjs"
  ],
  "session": "session_01Gqi43smmqjJ5sUrhfoPeKu",
  "premise_still_valid": true,
  "summary": "The card's premise reproduces on the CURRENT tip (origin/main 7572329069, not only the card's 94b3f37be): the rebuilt probe exits 4 with T1 at dashboard.zod.ts:628 on `+  ctx: z.RefinementCtx,`, and the true-positive control on the same real file fires at :701 on the same run - a false positive, not a dead instrument. WHY, measured: `enclosingDelimiter` abandoned its walk (answering `null`, which every caller reads as 'keep the tell firing') at the FIRST closer arriving with an empty stack, and a real hunk's first line is CONTEXT - here `  });`, the tail of the previous declaration - so the reading was over 108 lines before the hunk showed the `export function ...(` head. Repaired at that branch (route A / C, NOT route B's type-name exception): an underflow drops the closer and the walk continues, because the openers a hunk shows are strictly inside the ones it did not, so the shown stack is a suffix of the real one; an empty shown stack still answers `null`, so the reading stays positive-evidence-only. Pinned in one new 14-case battery in both directions, including the removed-side half where the repair closes a false NEGATIVE too. SCHEMA_PROPERTY_FORMS untouched at 10 rows; #18560 / PR #18700 and #18702 / PR #18750 rows, fixtures and batteries untouched. Assignee was empty at pickup and was left empty, as the os-dev definition prescribes.",
  "measurements": {
    "before_false_positive": "git diff 72dd95fa5a..09e16a5745 -- packages/spec/src/ui/dashboard.zod.ts (202 lines, ONE hunk) | node scripts/pm/check-widening-tells.mjs --declaration no --diff d.patch -> exit 4, 'T1 packages/spec/src/ui/dashboard.zod.ts:628 - a new key on a Zod object schema', patch line 116 = '+  ctx: z.RefinementCtx,'. Taken on origin/main 7572329069.",
    "before_true_positive_control": "'+  brandNewAuthorableKey: z.string().optional(),' inserted into DashboardWidgetSchema in the worktree as a REAL git diff on the REAL path, then `git checkout HEAD -- packages/spec/src/ui/dashboard.zod.ts` -> exit 4, 'T1 packages/spec/src/ui/dashboard.zod.ts:701'. The matcher works; the :628 row is false. (The card's own first control was a mis-build on a synthetic path; this one is on the probe's path.)",
    "why_with_proving_line": "Instrumented over the real patch: newFile[0] kind=context text='  });' is the hunk's FIRST line and is THE PROVING LINE - its two closers underflow a stack that has seen no opener, so enclosingDelimiter returned null at the branch `if (stack.length === 0) return null;` (:1985 on main). inParameterList -> false, enclosingDelimiter -> null, while the SAME walk restarted at the function head (newFile index 108) answers {opener:'(',head:'export function checkDashboardWidgetMetricMeasureArity'}. Not the 202-line hunk length, not the object-literal type on the first parameter (its braces close on their own line), not the distance to the head - one branch.",
    "repair": "scripts/pm/check-widening-tells.mjs enclosingDelimiter(): `if (stack.length === 0) return null;` -> `if (stack.length === 0) continue;` plus its reasoning comment; docblock rewritten; new header section 'The eleventh accidental variable #18721 removed - a hunk's LEADING CONTEXT'; the #17618 header paragraph gains the one-sentence correction that its decline did not reach a real diff until this round.",
    "pins": "New battery \"#18721 - a hunk's LEADING CONTEXT is not a reason to abandon the parameter reading\", 14 cases, registered in SELF_TEST_BATTERIES at 14. Both directions in one battery: PR #18720's own hunk reduced only as far as the failing branch requires (leading context + function head + object-literal-typed first parameter + the ctx line) silent at :628 with a CLEAN verdict; the true-positive control on the same file FIRING at :701; a new key behind the same underflowing context still firing; a real key after the parameter list closes still firing with the shape member as the reported row; the branch itself (opener after underflow / underflow with no opener / no leak past the close / no hunk crossing); and the OLD side, where the repair removes a phantom #16943 budget so a genuine key FIRES. The existing #18560-battery pin (#17618's parameter decline untouched by the wider vocabulary) and #17618's own underflow pin are byte-unchanged and green.",
    "self_test": "node scripts/pm/check-widening-tells.mjs --self-test -> exit 0 :: 473 cases pass (459 before, +14). No pre-existing case changed its verdict.",
    "after_readings": "probe -> exit 0 (clean, no widening tell); true-positive control -> exit 4 at :701; node scripts/pm/check-clause2-carriers.mjs --pair 18804 -> exit 0 (the C5 the card describes is gone).",
    "ablation": "From the COMMITTED fix (HEAD 7dd97e2bdf), under `trap restore EXIT INT TERM` with absolute paths: mutation `continue` -> `return null` applied by a node literal replace that refuses any anchor count other than 1; on-disk proof = anchor counts removed-text 1->0, injected-text 0->1, blob 025f8e51ec -> 9459620c63. Readings under the ablation: --self-test exit 1 with 6 of 473 cases failed - ALL SIX in the new #18721 battery, ZERO pre-existing; probe exit 4 (T1 back at :628); control exit 4. Restore verified by hash (9459620c63 -> 025f8e51ec == HEAD blob) and by `git diff HEAD` = 0 lines. No build/dist exists for a scripts/pm/*.mjs file, so ablation-dist-preflight is not applicable and the hash+anchor-count proof is the on-disk evidence. An earlier attempt was a NO-OP (perl with a double-escaped pattern: anchor counts 1->1, blob unchanged); the script's own guard refused it and exited non-zero rather than reporting a reading - recorded rather than silently re-run.",
    "gates": "Derived from the worktree with `node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack` (no hand-fed paths; 1 path vs merge base 757232906). 29 families, ALL exit 0. Reconciled with --ran: 29 derived, 29 run, 0 NOT-MEASURED, 0 UNRUN (a DERIVED zero - every row carries its exit code). node scripts/check-ci-filter-parity.mjs :: 0 | check-closing-keyword-parity.mjs :: 0 | check-closing-keyword-parity.mjs --self-test :: 0 | check-comment-mask-corpus.mjs :: 0 | check-declaration-mirrors.mjs :: 0 | check-declaration-mirrors.mjs --self-test :: 0 | check-scripts-symbol-anchors.mjs :: 0 | check-scripts-symbol-anchors.mjs --self-test :: 0 | check-self-test-wired.mjs :: 0 | check-self-test-wired.mjs --self-test :: 0 | check-self-test-workflow-commands.mjs :: 0 | check-self-test-workflow-commands.mjs --self-test :: 0 | check-whole-set-label-write.mjs :: 0 | check-whole-set-label-write.mjs --self-test :: 0 | scripts/pm/bare-root-worklist.mjs --self-test :: 0 | pnpm check:agent-test-spelling :: 0 | check:bash32-floor :: 0 | check:cli-command-ids :: 0 | check:cross-package-test-inputs :: 0 | check:driver-memory-census :: 0 | check:entry-guard :: 0 | check:nul-bytes :: 0 | check:parse-guard :: 0 | check:pm-dispatch-gates :: 0 (DETACHED, 742.1s, 1809 cases) | check:pm-widening-tells :: 0 | check:pnpm-filter-targets :: 0 | check:ratchet-remedy-authority :: 0 | check:refd-timer-probe :: 0 | check:watch-hint-literal :: 0. PLUS, outside the derived set: pnpm check:pm-clause2-carriers :: 0 (838 cases, the gate that consumes this file's verdict) and repo-wide `pnpm lint` (eslint . --no-inline-config) :: 0, run on the final commit 7dd97e2bdf - a full run, not a narrowing. ONE red in the first sweep, fixed: check-scripts-symbol-anchors.mjs exit 1 - the new header cited a path:LINE form and a line number is not an anchor form; rewritten to the symbol anchor packages/spec/src/ui/dashboard.zod.ts#checkDashboardWidgetMetricMeasureArity, green on re-run. NOT MEASURED, by the deriver's own account: 2 families taking a value from the workflow, 1 path-scheduled CI job (Test Core shards), 11 wide-population families, 53 artifact-roster families and 14 changeset-pending families are each outside the 29 and are CI's.",
    "pr_state": "Draft, base main, `Fixes #18721` as body line 1, `Clause-②: no` in the body AT CREATION, label `skip-changeset` (additive POST, read back: labels = [skip-changeset], no drift). Body read back in full after the write: byte-identical to what was sent except the trailing newline GitHub strips; exactly ONE footer, session-URL form; no angle-bracket fragment survives anywhere in it. `needs:contract-review` is NOT on the PR - neither added nor removed by this run."
  },
  "tests": "node scripts/pm/check-widening-tells.mjs --self-test -> exit 0 :: 473 cases pass (459 -> 473). Probe after the repair -> exit 0 clean; true-positive control -> exit 4 at dashboard.zod.ts:701; check-clause2-carriers --pair 18804 -> exit 0. 29 derived gate families all exit 0, reconciled with --ran (0 UNRUN, 0 NOT-MEASURED, derived); plus check:pm-clause2-carriers :: 0 and repo-wide pnpm lint :: 0 at 7dd97e2bdf. Ablation from the committed fix: mutation proved on disk by blob hash (025f8e51ec -> 9459620c63) and anchor counts (1->0 / 0->1), --self-test exit 1 with exactly 6 of 473 failing and all six in the new battery, restore verified by hash back to 025f8e51ec and `git diff HEAD` = 0 lines, under a trap.",
  "mcp_calls": "0 - no MCP GitHub tool was called, read or write.",
  "api_writes": "3 REST proxy writes: POST /repos/objectstack-ai/objectstack/pulls (draft PR 18804); POST /repos/objectstack-ai/objectstack/issues/18804/labels (skip-changeset, additive, Content-Type: application/json); POST /repos/objectstack-ai/objectstack/issues/18721/comments (this report). Reads: GET the card, its comments, the PR and its labels. Plus git pushes on the feature branch only (one force-with-lease on my own unshared branch to squash a wip commit; nobody else has pushed it and the PR carries no reviewer).",
  "open_questions": [],
  "out_of_scope_findings": [
    "noted, not filed: the #17618 battery's pin \"a closer with an empty stack is UNDERFLOW - the hunk began inside something it never showed\" still asserts `=== null` and is still green and still literally true, but its sentence now describes the ANSWER rather than a walk that stops. Left byte-unchanged on purpose. Successor: the next repair round on scripts/pm/check-widening-tells.mjs (this file has taken #18560, #18640, #18702 and now #18721), i.e. the domain:skills seat.",
    "noted, not filed: PARAMETER_LIST_HEAD reads the callable head only from text on the SAME line as the `(`, so a declaration split across lines (`export function` on one line, `  checkThing(` on the next) still leaves the tell firing. Pre-existing, deliberate per that constant's own docblock ('a prefix this list does not know leaves the tell where it was' - the loud direction), and unchanged by this round. Successor: the same next repair round.",
    "PROCESS NOTE for the seat: the PR body carries no `## Acceptance notes` section - the two rows above were written after the body was posted, and the os-dev definition forbids me PATCHing a PR body after creation. If the seat wants them on the PR, they belong in a review comment, verbatim from this report."
  ]
}

    Generated by Claude Code

  6. os-justin commented on Sep 17, 2026

    @os-justin
    os-justin
    on Sep 17, 2026 – with Claude
    Collaborator

    ACCEPT — in-seat contract-tier review of PR #18804 (head 7dd97e2bd, report 5721757240) by the skills seat, 2026-09-17T22:13Z. Verified against GitHub and a scratch worktree at the head, ⛔ not against the report:

    Generated by Claude Code

  7. os-justin
    closed this as completedin #18804on Sep 17, 2026

  8. os-justin commented on Sep 17, 2026

    @os-justin
    os-justin
    on Sep 17, 2026 – with Claude
    Collaborator

    LANDED — PR #18804 (#18721, check-widening-tells' parameter decline survives a hunk's leading context) merged through the queue at 2026-09-17T22:39:56Z (merged_at), squash a7bafc29a (single parent e7eb4e918, an ancestor of origin/main; enclosingDelimiter's underflow branch reads continue on main — an underflowing closer is dropped and the walk goes on, an empty shown stack still answers null; the #18721 battery of 14 registered, its --self-test at this head answered 473 cases); armed by this seat through the CCR route after the in-seat ACCEPT 5721940740 — ready 2026-09-17T22:13:59Z, added_to_merge_queue 2026-09-17T22:14:01Z. The card auto-closed completed at 2026-09-17T22:39:57Z on the PR's Fixes; pm:dispatched is stripped in the same act. Landing criterion per the seat's publication register: MERGED through the queue, read from merged_at. The seat's own three probe readings stand: the card's probe exit 4 at :628 on main before, exit 0 at the head after, and the true-positive control at :701 firing at the head — so every PR adding an object-level refinement with ctx: z.RefinementCtx stops earning a false T1 and a C5 on --pair. Serial behind it on this file: none queued.

    Generated by Claude Code

  9. os-justin
    removed
    pm:dispatched
    on Sep 17, 2026
  10. akarma-synetal
    added a commit that references this issue on Sep 28, 2026

    fix(pm): check-widening-tells' parameter decline survives a hunk's le…

    a7bafc2
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    domain:skillspriority:p3

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions