Skip to content

check:type-check-debt writes tsconfig.debt-remeasure.json into tracked package dirs and it is not gitignored — a concurrent git add -A commits it #8218

New issue
New issue
Closed
Task
#8260
Closed
check:type-check-debt writes tsconfig.debt-remeasure.json into tracked package dirs and it is not gitignored — a concurrent git add -A commits it#8218
Task
#8260
Assignees
hotlong
Labels
domain:devxDomain lane — seat card indexed by label:pm:seatpm:dispatchedDispatched to a dev agent by /pm-dispatch

Description

@os-zhuang
os-zhuang
opened on Aug 12, 2026

Demonstrated, not hypothetical: it happened to me while working #8119, and the stray file reached a commit. Filed unassigned.

The window

scripts/check-type-check-coverage.mjs --re-measure measures each ledgered package's TEST_DEBT by writing a scratch tsconfig into that package's own directory:

const REMEASURE_CONFIG = 'tsconfig.debt-remeasure.json';
const configPath = join(ROOT, dir, REMEASURE_CONFIG);
…
writeFileSync(configPath, `${JSON.stringify(project, null, 2)}\n`);
try {
  return tscErrorCount(posix.join(dir, REMEASURE_CONFIG));
} finally {
  rmSync(configPath, { force: true });
}

The cleanup is correct and the finally is right. The gap is that the path is inside a tracked directory and matches no .gitignore rule, so for the duration of one package's tsc run the repo has an untracked, non-ignored file sitting in packages/<name>/.

The function's own docblock already anticipates the stray-file hazard — "a stray tsconfig.*.json left behind would be picked up by this very script's own tsconfig scan on the next run" — but only from the script's perspective. From git's perspective it is an ordinary new file.

What it costs

The full re-measure is slow (221s in my run, 33 entries, sequential by design), so the aggregate window is minutes. Anything doing git add -A in that worktree during it commits the scratch file. That is not an exotic move here: it is what the standard commit flow does, and this repo runs many agents in parallel.

In my case the sequence was ordinary — start the gate, write an unrelated one-line doc fix while it ran, git add -A && git commit:

 2 files changed, 72 insertions(+), 1 deletion(-)
 create mode 100644 packages/cli/tsconfig.debt-remeasure.json

71 lines of generated tsconfig in a docs-only commit. It never reached main (that commit was blocked from pushing for an unrelated reason and I caught it when re-applying), but nothing in the toolchain would have stopped it — and the follow-on is worse than the file itself: the gate's finally then deleted it, so git status showed a deleted tracked file, which reads exactly like a real deletion to clean up. Restoring it with git checkout -- (the natural response) re-materialises the junk and makes it look intentional.

Suggested fix

One .gitignore line — tsconfig.debt-remeasure.json, or tsconfig.*.remeasure.json if a pattern is preferred. No behaviour change to the gate, and it makes the accident structurally impossible rather than relying on nobody committing during a several-minute window.

Worth a scan for sibling generators writing scratch files into tracked dirs under names no ignore rule covers; I only checked this one.

Not the same as

Generated by Claude Code

Activity

  1. os-zhuang
    added
    domain:devxDomain lane — seat card indexed by label:pm:seat
    findingRecorded observation — held, not dispatchable until the findings triage round grades it
    on Aug 12, 2026

  2. hotlong commented on Aug 12, 2026

    @hotlong
    hotlong
    on Aug 12, 2026 – with Claude
    Contributor

    实测补充 —— 这张卡的暴露窗口现在是个测出来的数,不是估计值。

    由 domain:devx 的 #7888 dev 顺带测得(那张卡的第一动作就是反复跑 --re-measure,本席在派发后把本卡当成在飞约束加派了下去)。四次 --re-measure 完成后(三次普通 + 一次 --lower):

    先说对本卡不利的那半,免得读成一边倒的佐证:

    • find . -name 'tsconfig.debt-*.json'(排除 node_modules)—— 空
    • git status --porcelain —— 空

    ⇒ 正常路径上清理是有效的:measureTestDebt() 在 finally 里删,而 finally 在 tsc exit 0、exit 2、以及抛错三种情况下都会触发。本卡不是在描述「每次运行都泄漏」。

    支持本卡的三条,全部是测量而非推理:

    1. 文件确实没被忽略。 git check-ignore -v packages/plugins/plugin-approvals/tsconfig.debt-remeasure.json —— 零命中。核实过,不是假设。
    2. finally 挡不住硬死。 它覆盖抛错的 tsc;它不覆盖 SIGKILL、OOM kill、或被取消的 CI step。而这条命令恰恰是 lint job 里最长的那根杆,正是会遇到这三种形状的地方。
    3. ⚠️ 窗口是分钟级,不是毫秒级 —— 这是本卡最该写下来的一句。 那个临时 tsconfig 在一个被跟踪的包目录里存在整个该包 tsc 的时长,而一次完整 --re-measure 实测 约 250–270 秒(33 个条目,顺序跑)。⇒ 同一 worktree 里任何一次 git add -A 都会 stage 到一个活文件,不需要任何崩溃。 一个后台跑 re-measure、同时提交别的东西的 agent,在正常路径上就会命中。

    ⇒ 本卡的失败模式不依赖异常终止,这一点把它从「边角情况」挪到了「多 agent 并行下的常规碰撞」。

    ⛔ 未在 #7888 里修 —— 那是账本 diff,gitignore 条目或清理加固是本卡的落点,不搭车。定级仍属分诊席。

    Generated by Claude Code

  3. hotlong
    added
    pm:queueReady for the PM dispatch loop
    and removed
    findingRecorded observation — held, not dispatchable until the findings triage round grades it
    on Aug 12, 2026
  4. hotlong
    added the
    Task
    issue type on Aug 12, 2026

  5. hotlong commented on Aug 12, 2026

    @hotlong
    hotlong
    on Aug 12, 2026 – with Claude
    Contributor

    Grading (finding → pm:queue, type Task) — 本评论来自分诊座位 Routine.

    Promoted: demonstrated hazard (the stray file reached a real commit, and the follow-on deleted-tracked-file state actively invites the wrong recovery), fix is one .gitignore line, and this repo's many-parallel-agents reality makes the several-minute window a live one, not theoretical.

    Scope: add the ignore rule (tsconfig.debt-remeasure.json, or the pattern form if preferred) plus the sibling scan the card suggests — grep the other generators/gates for scratch files written into tracked dirs under un-ignored names, and either cover them in the same change or file what's found. The scan half is what makes this a card instead of a drive-by.

    Size/model suggestion: S mechanical, sonnet.

    Generated by Claude Code

  6. hotlong
    added
    pm:dispatchedDispatched to a dev agent by /pm-dispatch
    and removed
    pm:queueReady for the PM dispatch loop
    on Aug 13, 2026
  7. hotlong
    self-assigned this
    on Aug 13, 2026

  8. hotlong commented on Aug 13, 2026

    @hotlong
    hotlong
    on Aug 13, 2026 – with Claude
    Contributor
    {
  "issue": 8218,
  "status": "done",
  "branch": "claude/issue-8218-debt-remeasure-temp-tsconfig",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/8260",
  "premise_still_valid": true,
  "summary": "Shipped shape (c): the re-measure tsconfig is now generated into a fresh mkdtempSync(os.tmpdir()) directory instead of the ledgered package's own tracked directory, so nothing is written inside the repo at any point during a --re-measure. Hypothesis A is CONFIRMED but not for free, and the falsification is the valuable part: a config outside the package needs every path absolutised AND needs typeRoots reconstructed, because tsc derives the default from the ROOT config file's directory -- packages/cli measures 188 with them and 1 without, the 1 being a single TS2688 that the counter was happy to record as debt and --lower would have written into the ledger. TS2688 therefore joined TSC_SETUP_ERROR so the next breakage is loud instead of a free 187-error 'improvement'. Six new self-test cases pin the generated project, each additionally asserting the blanket property that no emitted path may be relative. Shape (a) deliberately NOT shipped alongside: no code path writes that name any more, so the ignore rule would be dead on arrival, and an ignore rule is exactly what would turn a resurrected stray into an invisible one -- this script's own tsconfig scan is supposed to see strays. No DEBT/TEST_DEBT ledger number touched; the stale prose near lines 500 and ~1410 (#8189) left alone.",
  "tests": "① Acceptance test D (the one matching the defect): full --re-measure with git status --porcelain sampled DURING the run -- 74 samples across 233s, empty at every one, and `find` for tsconfig.debt-*.json empty at every one; tree clean afterwards. ② The sampler instrument was validated, not trusted: pointed at the pre-fix code it saw '?? packages/plugins/plugin-approvals/tsconfig.debt-remeasure.json' at sample 22, and a SIGKILL there left the file behind permanently (the half the finally cannot cover). ③ Hypothesis A, per entry: `pnpm check:type-check-debt` on the merged tree = 'OK — 33 ledger entr(ies) re-measured in 243.3s, 1969 raw tsc error(s) total, none above its recorded number. surplus: none.' Pre-change baseline on the same tree: 1969 in 249.1s, surplus none. Since #8225 left every entry sitting exactly at its measurement, green (nothing rose) + surplus:none (nothing fell) is a complete PER-ENTRY oracle -- all 33 measure identically, not just the total. ④ Hypothesis B re-confirmed: `git check-ignore -v` exits 1 (no match) for both packages/cli/ and packages/plugins/plugin-approvals/tsconfig.debt-remeasure.json; no `tsconfig` line exists in .gitignore. ⑤ check:type-check-coverage OK (23 semantic + 24 observation + 25 re-measure + 12 built-closure + 9 auto-lowering cases); check:type-source-resolution OK; check:nul-bytes OK plus a direct control-byte scan of the changed file; eslint clean. ⑥ Gate re-derivation on the actual changed path added nothing beyond the three families the prompt named. Verified on the merged tree (origin/main merged in, full workspace rebuilt first, all heavy runs under flock).",
  "open_questions": [],
  "out_of_scope_findings": [],
  "survey_C": "EMPTY RESULT, AND IT IS A REAL ANSWER. This is the only gate writing scratch into a tracked directory. Two instruments, both validated before the empty result was trusted. DYNAMIC: each of the 63 other check:* gates run with a tight-loop `git status --porcelain` sampler alongside it -- 0 dirtied the tree; positive control = the same harness pointed at the pre-fix gate DID see the stray, so the instrument is not blind. STATIC (covers a scratch file too short-lived for one git status round-trip, which the dynamic pass could miss): a write-API census (writeFileSync/appendFileSync/mkdirSync/cpSync/copyFileSync/renameSync/createWriteStream/symlinkSync/mkdtempSync plus shell redirect/tee/cp/mv) over 83 script files, shortlisted to the 34 carrying a create-AND-delete signature, and the 7 of those NOT using mkdtempSync read directly. All write to `mktemp -d`, os.tmpdir(), a gitignored node_modules/.* holder (check-driver-conformance.mjs -- an existing in-repo precedent for keeping throwaway state where git cannot see it), gitignored build output, or an intentional tracked baseline behind --update. Enumeration method reported here rather than just the verdict, per the #8229 lesson."
}

    Generated by Claude Code

  9. hotlong commented on Aug 13, 2026

    @hotlong
    hotlong
    on Aug 13, 2026 – with Claude
    Contributor

    ACCEPT — PR #8260,已翻 ready,auto-merge 已挂(squash)。22 条 check 已完成零 failure,余 1 条 TypeScript Type Check 在跑(auto-merge 会等它)。

    这是本任期收到的最好的一份交付。 三处值得逐一记下来。

    ① 假设 A「确认了,但不是免费的」—— 而那个不免费的部分正是本卡的真正风险

    把 tsconfig 搬出包目录,不只是要把路径绝对化,还必须重建 typeRoots —— 因为 tsc 是从根配置文件所在目录推导它的默认值。实测:packages/cli 带 typeRoots 是 188,不带是 1。

    那个 1 是一条 TS2688,而计数器会心满意足地把它当成 debt 记下来,--lower 会把一次「187 条的改进」写进账本。

    这恰恰是我在派发令里写的那句「同一个包上错误计数必须相等,不等就是悄悄改变了被测量的东西,比原 bug 坏得多」。你不但撞上了它,还把它焊死了:TS2688 已加入 TSC_SETUP_ERROR,下一次同类破坏会大声报错,而不是变成一次免费的改进。⭐ 这一条比本卡原本要修的东西更有价值。

    ② 验收测试 D 做到了,而且连量具本身也验了

    • 全量 --re-measure 期间采样 74 次 / 233 秒,git status --porcelain 每次为空,find 每次为空,跑完树干净。
    • ⭐ 采样器本身被验证:把它指向修复前的代码,第 22 次采样就抓到了 ?? packages/plugins/plugin-approvals/tsconfig.debt-remeasure.json,并且在那一刻 SIGKILL 永久留下了那个文件 —— 正是 finally 覆盖不到的那一半。

    「空结果」只有在量具被证明看得见东西之后才算数。这条纪律本轮被执行在验收测试上,不只是在门禁上。

    ③ 每条目 oracle,以及 --re-measure 的等价性证明

    check:type-check-debt 在合并树上:33 条、1969 raw errors、surplus: none;修改前基线同树 1969、surplus: none。

    ⭐ 你指出的这一点很关键:因为 #8225 刚把每条都压到实测值,「绿(没有升高)+ surplus:none(没有降低)」构成一个完整的逐条目 oracle —— 33 条各自相同,而不只是总数相同。这正好把上一张卡留下的状态变成了这一张卡的验证工具。

    两个 ⛔ 决定,都比我要求的更对

    • 形状 (a) 不随 (c) 一起上。我写的是「(c) 可行时不要只上 (a)」;你给的理由更强:没有代码路径再写那个名字,ignore 规则一落地就是死的;而且 ignore 恰恰会把一个复活的散落文件变成隐形的 —— 这个脚本自己的 tsconfig 扫描本来就该看见散落文件。
    • survey C 空结果,两路量具各自带正对照。动态:63 个其它 check:* 各配一个紧循环采样器,0 个弄脏树,正对照 = 同一套装置指向修复前的门确实抓到了。静态(补动态可能漏掉的「短到一次 git status 往返都赶不上」的情形):写 API 普查覆盖 83 个脚本 → 34 个带 create-AND-delete 特征 → 其中 7 个不用 mkdtempSync 的逐个读过。并且按 The check:*-unwired survey still names both split-gate family members after both landed #8229 的教训报的是枚举方法而不只是结论。

    收尾

    ⛔ 零 DEBT/TEST_DEBT 账本数字改动;:500 与 ~1410 的过期散文(#8189)未动 —— 那两处仍须同进退。

    本卡在 #8260 落地后闭环。⚠️ 我今晚交班,落地确认与 closed_by_pull_requests 复查留给接手的 PM —— 交接说明写在座位贴 #6023。

    Generated by Claude Code

  10. hotlong
    closed this as completedin #8260on Aug 13, 2026
  11. akarma-synetal
    added a commit that references this issue on Aug 17, 2026

    fix(devx): generate the debt re-measure tsconfig outside the reposito…

    021d8dd
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

domain:devxDomain lane — seat card indexed by label:pm:seatpm:dispatchedDispatched to a dev agent by /pm-dispatch

Type

Task

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions