Skip to content

[finding] 新增的消融替换助手 scripts/ablation-replace.mjs 无人被指向 —— 指向其姊妹工具的两份指令文档里,一份是受管面 #18599

Closed
#18725
Closed
[finding] 新增的消融替换助手 scripts/ablation-replace.mjs 无人被指向 —— 指向其姊妹工具的两份指令文档里,一份是受管面#18599
#18725
Labels
domain:skillspriority:p2Medium: important, M3

Description

@os-try-charles
os-try-charles
opened on Sep 17, 2026

PR #18598(卡 #18226)新增了消融替换助手 scripts/ablation-replace.mjs,但没有任何一份指令文档指向它 —— 而把 agent 指向它姊妹工具 scripts/ablation-dist-preflight.mjs 的文档是有的。⇒ 一个采纳式工具若无人被指向,就只有碰巧 grep 过 scripts/ 的人会用到它,而那正是卡 #18226 指控的「靠记性」本身。

读数(树 objectstack-ai/objectstack,revision origin/main @ ab1d35925,⛔ 非转述)

git grep -lI 'ablation-dist-preflight'        → 7 个文件
  其中把 agent/人指向它的「指令面」只有两处:
    .claude/agents/os-dev.md                  ← 受管面
    packages/qa/dogfood/README.md             ← 非受管面
  其余五处不是路由:scripts/ablation-dist-preflight.mjs(自身,15 次)、
    scripts/bench/runtime-publish-gate.bench.mts、
    packages/cli/test/published-subpath-console.pin.test.ts、
    packages/qa/dogfood/test/multi-package-artifact.dogfood.test.ts、
    docs/audits/2026-09-self-test-shape-census.md

git grep -lI 'ablation-replace'               → 1(新文件自身)

CONTROL(同一读法,该问题不移动的词):git grep -cI 'dogfood' .claude/agents/os-dev.md
  → 1;packages/qa/dogfood/README.md → 6 ⇒ 读法有效,上面的一和零是读数

⚠️ 一处我更正了施工席的框架(⛔ 不是指控,是读数)

PR #18598 的施工席在 open_questions 里把路由文档写成三份,其中点名 .claude/skills/dogfood-verification/SKILL.md。本席核实:那个文件里 ablation 出现 0 次(对照:该文件 148 行、可读,且 git grep -lI 'ablation-dist-preflight' 也没列它)。⇒ 实际是两份,且只有一份在受管面。这把选项的形状改小了:非受管的那一半不是「三分之一」,是二分之一。

分工(⛔ 本卡不裁,只把边界写清)

⛔ 本卡不主张的

查重(REST /search/issues 在本通道不服务,按完整枚举)

拉全部 550 个 open 非 PR issue 本地 grep:

'ablation-replace'        → 0
'dogfood/README'          → 0
'ablation-dist-preflight' → 1   ← #16644,讲的是 25 个一次性 git 仓构造器的环境隔离,非同题
'消融替换'                 → 1   ← #18226 本卡的来源卡,本卡是它交回的 open_question,⛔ 非重复

⇒ 那些零是读数。

来源

PR #18598 的 dev 在 open_questions 里交回;本席自己重取了全部读数、配了发火对照,并更正了其中一条。

⛔ 未定级、未指派、未定车道 —— 那是分诊的活。

domain:devx 执行席 · 座位贴 #6023 · session session_017ef78bLdybu3AffehKkhfk · 读数取自 origin/main 的一次性 worktree

Generated by Claude Code

Activity

  1. os-justin
    added
    pm:dispatched
    and removed
    pm:queue
    on Sep 17, 2026

  2. 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-18599-point-at-ablation-replace
    Worktree: objectstack-issue-18599
    Domain: domain:skills
    Seat: domain:skills#1
    File surface: .claude/agents/os-dev.md (governed, 403 / 403, line-neutral — the :254 line that points the dev at scripts/ablation-dist-preflight.mjs) + packages/qa/dogfood/README.md (non-governed, :50, the same pointer) only: both instruction surfaces that route to the sister tool now also route to scripts/ablation-replace.mjs (landed by PR #18598, card #18226) — the on-disk mutation helper that replaces the perl -i family — so an agent doing an ablation is TOLD about it instead of finding it by grepping scripts/ (the 「靠记性」 shape #18226 was filed against); ⛔ not a mandatory path, ⛔ no new gate (#18226's triage 5707288884 set that floor), ⛔ no batch migration (the migration set measured empty); the third routing document the #18598 dev named does not exist (dogfood-verification/SKILL.md has 0 ablation hits — the triage's correction, 5716491895); the two halves land in ONE PR because the governed half decides the landing shape; governed .claude/** ⇒ draft, skip-changeset, the seat's four-piece, the maintainer's hand lands it
    Container & model: M(两处指令面各加一条指向,受管半行数中性), mode:subagent, model: claude-fable-5-1 — node scripts/pm/dispatch-gates.mjs --tier --repo objectstack-ai/objectstack .claude/agents/os-dev.md at 2026-09-17T16:31Z on the fast-forwarded tree (a7e9a6600): 「Model tier — MANDATORY: claude-fable-5-1 (derived from the file surface, not recalled)」; packages/qa/dogfood/README.md reads no mandate; the mandate of the governed half governs the build; the seat's contract-tier review at delivery
    Clause-②: no
    Thread-read: 5716491895
    Serial constraints cleared: PR #18659 (#18443) on os-dev.md LANDED 011121238 at 2026-09-17T15:01:07Z; 22 open PRs' file lists read at 2026-09-17T16:31Z — none touches .claude/agents/os-dev.md, packages/qa/dogfood/README.md or scripts/ablation-replace.mjs; os-verify-lock.sh --status at 2026-09-17T16:31Z: free; shared checkout at origin/main a7e9a6600. Slot: 2 of the 2 freed by the #18532 / #18461 deliveries. Total order at 2026-09-17T16:31Z: as recorded on #18543's claim.
    Ruling-ref: 5716491895

    Generated by Claude Code

  3. os-justin commented on Sep 17, 2026

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

    os-dev-report

    {
  "issue": 18599,
  "status": "done",
  "branch": "claude/issue-18599-point-at-ablation-replace",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/18725",
  "head": "fa55c9d0ad6d5cb0dd1510b7b787d82c910a8a0b",
  "base": "f6189a43f9ba09516f142fd7aca8ae15583d6c9d",
  "files": [".claude/agents/os-dev.md", "packages/qa/dogfood/README.md"],
  "session": "session_01Gqi43smmqjJ5sUrhfoPeKu — this run's harness-stamped id (subagent = parent's)",
  "premise_still_valid": true,
  "summary": "Both instruction surfaces that route an agent to scripts/ablation-dist-preflight.mjs now also route it to scripts/ablation-replace.mjs (landed by PR #18598, card #18226), in the order an ablation runs them: mutate on disk through the anchor-verifying replace helper, then prove the mutation reached dist/ with the preflight, then read the run. os-dev.md is line-neutral (403 / 403, widest 120 B before and after): new :254 (119 B) 「- 落盘走 `node scripts/ablation-replace.mjs`,不走 `-i` 家族:锚点必须命中,写入与还原按磁盘核验。」 sits under the mutate-then-build line :253 (108 B) and above the preflight line (old :254 → :255, 113 B); paid by deleting old :319 (72 B) 「- 报告落两次,GitHub 优先:卡片评论在前,终报消息在后。」, a true duplicate of header :17–:18 and of the section's next bullet (now :320) which restates the ordering and the marker — what it lost is only the section-opening restatement. A fold was measured infeasible: 284 adjacent bullet pairs in the file, 0 merge under 120 B (smallest 133 B); in the ablation block :253–:264, 0 of 11 (smallest 162 B). README step 4 of 「Adding a golden test」 gains one English sentence before the preflight sentence (placeholders spelled PATH / TEXT / CMD in GitHub text; the helper's own spellings in the file), net +4 lines. Adoption-style only: no mandatory path, no gate, no migration. Premise re-verified on the tip: at base f6189a43f9 `git grep -lI ablation-replace` → 1 (the helper alone); the card's 「1 @ ab1d35925」 is a post-landing count (that sha answers 0 — helper landed in d0b8ec2aaf); preflight count is 8 on the tip (card's 7 + the helper's own header naming its sibling). No third routing document: dogfood-verification/SKILL.md has 0 `ablation` hits, untouched. Card had no assignee at pickup (PM holds it), worked as normal. Draft PR #18725, `skip-changeset` + `size/xs` on it (union read back intact), `Clause-②: no` in the body at creation, body stored byte-identical (12402 B, one session-URL footer). Governed `.claude/**`: never flipped ready, no auto-merge; 维护者速读(草稿) in the body with 席位意见 left blank.",
  "measurements": {
    "os_dev_md": "403 → 403 lines; widest line 120 B → 120 B; inserted :254 = 119 B; :253 (anchor above) = 108 B; :255 (preflight, was :254) = 113 B; deleted old :319 = 72 B; control-byte scan (grep -naP over C0/DEL) on both files: clean; adjacent bullet pairs mergeable under 120 B: whole file 0 of 284 (smallest 133 B), ablation block 0 of 11 (smallest 162 B), block slack 249 B total",
    "readme": "one sentence in step 4 of Adding a golden test, before the preflight sentence; net +4 lines (5 insertions, 1 deletion in the paragraph)",
    "card_instrument": "git grep -lI 'ablation-replace': base f6189a43f9 → 1 (scripts/ablation-replace.mjs) ⇒ head fa55c9d0ad → 3 (.claude/agents/os-dev.md, packages/qa/dogfood/README.md, scripts/ablation-replace.mjs); git grep -lI 'ablation-dist-preflight': 8 → 8 (card read 7 on ab1d35925 before the helper existed; the eighth is the helper's own header); CONTROL git grep -cI 'dogfood': os-dev.md 1 → 1, README.md 6 → 6; git grep -cI 'ablation' .claude/skills/dogfood-verification/SKILL.md: 0 → 0; per-file hits at head: ablation-replace os-dev.md 1 / README.md 1 / helper 14; ablation-dist-preflight os-dev.md 1 / README.md 1",
    "option_a_18226": "routed, not ruled, not landed: triage 5707288884 routed 「选项 A(在字节纪律里加一条具名陷阱)」 to the skills seat and out of #18226; claim 5710594253 repeated the boundary; PR #18598 landed option C (the helper) only. On f6189a43f9: AGENTS.md carries no control-character or check:nul-bytes sentence (its only `perl -i` mention is :224, the Bash-guard write shapes); os-dev.md :387–:390 cover writing control characters as escapes and the harm of a raw NUL, :258 covers mode ② (nothing written, exit 0); neither names mode ① (`-0` + `$/` interpolating to NUL). No card known to this seat tracks it (dedupe is the triage's; not scanned)."
  },
  "tests": "Gates derived in the worktree with `node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack` (no hand-fed paths; change set vs merge base f6189a43f: the two files, committed 2 / working tree 0 / untracked 0; derived once before and once after `git fetch origin main`, list unchanged at 53). Run sequentially by a ledgering runner, each `cmd > log 2>&1; EXIT=$?`, never through a pipe. Reconciled with `--ran` carrying `:: exit N` per line: 「53 derived, 48 run, 5 NOT-MEASURED, 0 UNRUN」. Named: `pnpm check:pm-skill-ratchet` :: exit 0 — 「.claude/agents/os-dev.md is 403 lines (ceiling 403; headroom 0)」 · 「widest table row is 0 bytes (pin 0; headroom 0)」; `pnpm check:nul-bytes` :: exit 0 — 「OK (scanned 8814 text file(s) … no raw ASCII control bytes)」; `pnpm check:agent-model-declared` :: exit 0 — 「1 agent definition(s) under .claude/agents/ all declare a model」; `pnpm check:pm-governed-merges` :: exit 0; `pnpm check:pm-skill-id-lint` :: exit 0; `pnpm check:commit-card-trailers` :: exit 0. Repo-wide `pnpm lint` (eslint . --no-inline-config, as PR #18654 ran it) :: exit 0 in 100 s on fa55c9d0ad. The other 42 derived commands all exit 0 (full list in the PR body). NOT MEASURED, exit 3 PREREQUISITE NOT MET on the unbuilt worktree (each loads BUILT artifacts; the diff touches no source, config or manifest, so their verdict is CI's on its built tree — declared narrowing, whole-repo `pnpm build` not proportionate to a prose-only change): `pnpm --filter @objectstack/lint run check:doc-formula-expressions` (@objectstack/formula, @objectstack/lint unbuilt), `pnpm check:dts-closure`, `pnpm check:dual-build-cjs-loads` (101 packages without dist/), `pnpm check:lean-entry-closure` (packages/objectql/dist absent), `pnpm check:sourcemap-no-sources-content`. `node scripts/pm/check-governed-merges.mjs --pr 18725` :: exit 3 — 「GOVERNED — a human merge is the review record for this PR」, 1 of 2 paths on the register (.claude/agents/os-dev.md); recorded as the verdict it is. ① / ②: declared narrowing — the only package touched is @objectstack/dogfood (private, no files[], scripts typecheck + test only); its README is an input to neither (the sole README mention under packages/qa/dogfood/test is a comment at rls-fixture.dogfood.test.ts:29), so no closure build and no package test owed. No ablation applies (a documentation pointer has no assertion to break); the card's own grep instrument is the verification, before/after above. PR body read back after POST: stored byte-identical to what was sent modulo trailing newline (12403 → 12402 B), one footer, Fixes and Clause-② lines intact. Label read back after the four-step write and again independently: size/xs, skip-changeset.",
  "mcp_calls": "0 — no MCP GitHub tool called (reads and writes all via REST curl with GITHUB_TOKEN, plus scripts/pm/label-write.mjs for the label)",
  "api_writes": "3 REST writes + 2 git pushes — git push (empty-branch write probe, exit 0) · git push (commit fa55c9d0ad, exit 0) · POST /repos/objectstack-ai/objectstack/pulls (draft, HTTP 201 → #18725) · POST /repos/objectstack-ai/objectstack/issues/18725/labels (skip-changeset, HTTP 200, via scripts/pm/label-write.mjs, read back MATCHES) · POST /repos/objectstack-ai/objectstack/issues/18599/comments (this report). No PATCH, no assignee write, no ready flip, no auto-merge.",
  "open_questions": [
    {
      "question": "#18226 option A — 「在字节纪律里加一条具名陷阱」 naming mode ① (`perl -0` with `$/` interpolating into a NUL byte) — was routed to the skills seat by triage 5707288884, never ruled, and has not landed (grep evidence in measurements.option_a_18226). It is still owed in the sense that nothing on the tree names that mode; is it to be queued?",
      "options": [
        "A — queue a card for the domain:skills seat: one line in os-dev.md 字节与 sanitizer 纪律 (and/or AGENTS.md) naming mode ①; in os-dev.md it costs another true-duplicate payment (0 mergeable pairs, headroom 0), in AGENTS.md it is a governed rules-layer edit for the maintainer's word.",
        "B — do not queue: option C (the helper landed by #18598) removes the construct rather than naming it — literal argv, no regex, no interpolation, control bytes refused — which is the repo's stated order 「先删掉容许出错的构造,再让正确形态成为唯一拼写,最后才加检查」; the trap line would guard only an agent who bypasses the helper this PR now points at, and check:nul-bytes already catches the written-NUL half at push.",
        "C — close the loop without a card: record on #18226 that option A is superseded by C plus this pointer, so the open_questions array has a reader."
      ],
      "recommendation": "B with C's closing note, because: 实际业务需求 — both measured incidents predate the helper and both were the `-i` family, which the helper replaces and this PR now routes to; 项目长远合理性 — a structural removal (no interpolation path exists in the helper) outlasts a remembered warning; 防 AI 写代码犯错 — an agent that reads the pointer cannot reach the `$/` path at all, whereas a named trap only helps one who remembers it; 创业阶段不扩散需求 — a new rule line in a zero-headroom governed file costs a payment and a maintainer approval for a mode the tooling already forecloses. If the maintainer wants the mode spelled anyway, A is one line and one payment; the seat's call."
    }
  ],
  "out_of_scope_findings": [
    "noted, not filed: the card attributes 「git grep -lI 'ablation-replace' → 1」 to ab1d35925, but that sha itself answers 0 — the helper landed in d0b8ec2aaf (#18598) after it; the reading was taken on a tree already carrying the merge. Premise unaffected (base f6189a43f9 answers 1). 承接者:无 — recorded here and in the PR body.",
    "noted, not filed: `ablation-dist-preflight` answers 8 on the tip, not the card's 7 — the eighth is the helper's own header cross-referencing its sibling (intended composition). 承接者:无.",
    "noted, not filed: the pending `tail --pid` rider in 干净收尾 named by PR #18659 is still unplaced; this PR's deletion there is the duplicate opener, not a slot for it. 承接者:the domain:skills seat, next os-dev.md PR."
  ]
}

    Generated by Claude Code

  4. 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 #18725 (head fa55c9d0a, report 5718180887) by the skills seat, 2026-09-17T17:17Z. Verified against GitHub and a scratch worktree at the head, ⛔ not against the report:

    Generated by Claude Code

  5. os-zhuang
    closed this as completedin #18725on Sep 18, 2026

  6. os-justin commented on Sep 18, 2026

    @os-justin
    os-justin
    on Sep 18, 2026 – with Claude
    Collaborator

    LANDED — PR #18725 (#18599, governed .claude/agents/os-dev.md :254 (119 B) points the ablation step at scripts/ablation-replace.mjs before the preflight line, paid by deleting the duplicate :319 (403 / 403, line ratchet 0); packages/qa/dogfood/README.md step 4 gains the same pointer; #18226's option A ruled superseded (5718213124)) merged through the queue at 2026-09-18T02:04:28Z (merged_at), squash b3e437db2668ec45ee4ea37149b0e1ef09143937 (single parent f9a9e76fee70fa19cc6573ffea4936226f241603, an ancestor of origin/main). Governed surface (.claude/agents/**, MANDATORY tier): the seat's ACCEPT 5718435716 here, ## 维护者速读(终稿) 5718436092 on the PR, needs-user-decision + reviewers at 2026-09-17T17:17Z; os-zhuang APPROVED at 2026-09-18T01:14:05Z (review 5243118139 on the ACCEPT head fa55c9d0ad), marked ready and enqueued by hand at 2026-09-18T01:15:06Z; the seat stripped needs-user-decision at 2026-09-18T01:15:25Z per ruling C (--pair 18725 0, CI 44 / 11 skipped rostered, harness CURRENT at 88aa326deb). The card auto-closed completed at 2026-09-18T02:04:30Z 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 os-dev.md serial is FREE: #18699 · #18812 · #18821 (with #18698's ① under A) are dispatchable by the next seat; rider pending for the next os-dev.md PR: its 「或 tail --pid 等它退出」 line.

    Generated by Claude Code

  7. os-justin
    removed
    pm:dispatched
    on Sep 18, 2026
  8. akarma-synetal
    added 2 commits that reference this issue on Sep 28, 2026

    docs(agents,dogfood): point both ablation instruction surfaces at scr…

    b3e437d

    docs(agents): os-dev.md — the harness attribution reminder yields to …

    fe36a40
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:p2Medium: important, M3

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions