finding(cli): app-generator's two tests pin a DIFFERENT generator branch depending on cwd — so CI never covers the standalone branch #7807

Activity

1. claude

   added

   tests

   pm:queue

   finding

   priority:p3

   domain:uiobjectui ui stream: fix lands on the published library or apps — objectui execution seatobjectui ui stream: fix lands on the published library or apps — objectui execution seat

   package: cli

   pm:dispatched

   and removed

   pm:queue

   on Sep 5, 2026
2. claude commented on Sep 5, 2026

   

   claudeBot

   on Sep 5, 2026 – with Claude

   ContributorAuthor

   More actions

   Claim: PM loop, round 6 — all-opus lane (Clause-② no)
   Session: session_01KbJQ1y1J12nZxYzFWhP8Q3
   Branch: claude/issue-7807-app-generator-inworkspace-case
   Worktree: objectui-issue-7807
   Domain: domain:ui
   File surface: packages/cli/src/__tests__/app-generator.test.ts · 本卡自己的 .changeset/*.md. ⛔ Stop on breach: 任何非测试文件的改动（含 createTempApp / contextOfCurrentProcess 的实现）、任何 packages/cli/package.json 的 test 脚本改动、任何改动现有那两条测试的行为。Explain in the report.
   Container & model: S, mode:subagent, model: opus
   Clause-②: no —— 新增一条测试用例。不改变契约接受/拒绝行为，不扩大公开面。⚠️ Void condition: 任何 packages/*/src 下非测试文件的改动。
   Serial constraints cleared: **PR #7806（卡 #7799）**已交付，它判定 app-generator.test.ts immune 并未修改它（本卡正是从它的探针读数析出的）；若 #7806 尚未落地也不阻塞（本卡加的是新用例，不碰它判过的东西）。在飞的 #7798 / #7795 / #7744 / #7745 都与 packages/cli/ 不相交。

   ---

   事实（#7799 的席位实测，⛔ 不必重测）

   app-generator.test.ts 在两种 cwd 下都绿（各 41 passed），且经反向探针确认非空过。但探针同时暴露：逐字节相同的那两条测试，在两种调用下钉的是不同分支 ——

   调用	cwd	实际被钉的分支
   pnpm exec vitest run packages/cli/…（仓根，CI 跑的形式）	仓根	IN-WORKSPACE
   pnpm --filter @object-ui/cli test	packages/cli	STANDALONE

   探针读数：仓根 41p；包目录 1 failed | 40 passed，AssertionError: PROBE7799 ambient isMonorepo: expected false to be true。

   ⇒ CI 只跑仓根形式 ⇒ STANDALONE 分支在 CI 里从来没被覆盖过，而且从读代码上看不出来（看起来两条测试覆盖了两个分支）。

   ⛔ 最要紧的一条：不要"修"现有那两条

   contextOfCurrentProcess() 有意在比较的两侧都镜像 ambient cwd，而被测的真实 createTempApp 也用同样方式派生自己的上下文 ⇒ 这个镜像两种 cwd 下都自洽。

   ⚠️ 只把测试那一侧改成从文件派生，会让镜像失同步，把一条本来通过的测试变红。 #7799 的席位据此判它 immune 并停手，本席采纳。⇒ ⛔ 现有两条一个字都不要动。

   交付物

   加第三条用例，用文件派生的根显式钉住 IN-WORKSPACE 分支，保留原有的 ambient 镜像不动 ⇒ 两个分支在任何调用下都被覆盖。

   ⭐ 路径派生用从 import.meta.url 逐段上溯（PR #7796 落地、#7806 沿用的同一个拼法）。⛔ 不要引入第三种拼法；⛔ 不要用 new URL(rel, import.meta.url) —— 它在本仓测试变换里被 Vite 重写成 http://localhost:3000/@fs/…，fileURLToPath 两种 cwd 下都抛。

   ⚠️ 若你判断同时该补一条显式钉住 STANDALONE 的用例，可以做 —— 但分开两条、各自显式，⛔ 不许再让"钉哪个分支"取决于 cwd。

   ⭐ 验收 —— 判据是「两种 cwd 下覆盖相同」

   • 两种 cwd 各跑一次，都绿，且报出文件数/用例数：仓根 pnpm exec vitest run packages/cli/src/__tests__/app-generator.test.ts；包目录 pnpm --filter @object-ui/cli test。⇒ 新用例在两边都跑到、都绿。
   • ⭐ 点亮控制：把新用例的根派生临时打回 ambient cwd 形式，确认它在包目录下变红并指名（因为那时它又会去钉 STANDALONE）；然后在 trap … EXIT INT TERM 下用 git checkout HEAD -- <绝对路径> 还原，以 blob hash 相等 + 空 git diff HEAD 证明 —— ⛔ 证明靠状态，不靠 trap 是否触发；变异前先用 hash 确认落盘（空 hash 读作失败，不读作「没什么可比」）。
   • ⭐ 证明你没有动坏现有那两条：给出改动前后 app-generator.test.ts 在两种 cwd 下的用例数对照（改动前各 41p，改动后应为 41+N，且没有任何一条从通过变成不通过）。
   • pnpm --filter @object-ui/cli type-check（走 turbo），报出任务数与 exit code。

   变更集

   改动落在发版包 src/ 下（虽是测试文件）⇒ 按 AGENTS.md:161 大概率需要。⛔ 不要凭规则推断：跑 node scripts/check-changeset-presence.mjs，判词原样抄进报告；欠就写空 frontmatter。⛔ 永不写 major。

   ⚠️ 五个仪器陷阱

   ① 裸 tsc --noEmit 不是本仓的门（turbo type-check dependsOn ^build，到不了非包目录；构建前的 TS2882 是未满足的前提、不是结果）② 本容器没有「未认证的探测」，匿名样子的 POST 是真写入，探通道只用 GET ③ scripts/__tests__/check-doc-links.test.ts 共享盒上 15s 超时是已知 flake ④ ⛔ new URL(rel, import.meta.url) 被 Vite 重写（见上）⑤ ⛔ 不用两点式 origin/main..HEAD 报文件数（并发落地会虚报，实测 4 vs 真实 2），以记录的 base 或 GitHub 的 changed_files 为准。

   契约复审

   Clause-② 是 no ⇒ 不排在契约复审后面。

   ---

   Generated by Claude Code
3. claude commented on Sep 5, 2026

   

   claudeBot

   on Sep 5, 2026 – with Claude

   ContributorAuthor

   More actions

   os-dev-report

   {
     "issue": 7807,
     "status": "done",
     "branch": "claude/issue-7807-app-generator-inworkspace-case",
     "pr": "https://github.com/objectstack-ai/objectui/pull/7810",
     "premise_still_valid": true,
     "summary": "Added two cases to the 'generation onto disk' block of packages/cli/src/__tests__/app-generator.test.ts that NAME the generator context branch instead of inheriting it from the ambient cwd: one pins IN-WORKSPACE against the file-derived REPO_ROOT (the file's existing import.meta.url walk-up, PR 7796's spelling, not a third one), the other pins STANDALONE against the test's own mkdtemp directory, both sides naming the same path. The existing two cases are byte-identical: the only removed line in the entire diff is the vitest import, which gains vi. process.chdir() is not an available lever here - it throws ERR_WORKER_UNSUPPORTED_OPERATION under the unit project's pool: 'threads' (measured in this container with a worker_threads probe) - so process.cwd() is stubbed for exactly the one synchronous writer call and restored in a finally. I added the second, explicitly-STANDALONE case because one explicit IN-WORKSPACE case alone would not equalise coverage: from the repo root the ambient pair also pins IN-WORKSPACE, so STANDALONE would still be unreached in the only form CI runs. Two files changed (GitHub changed_files = 2, matching the recorded base a617bb8d3): the test file and an empty-frontmatter changeset. No non-test file under packages/*/src was touched, and packages/cli/package.json's test script is untouched - the Clause-2 void condition did not trigger.",
     "tests": "All readings below are at the final commit afa1722c9 (git rev-parse --short HEAD), working tree clean, and every exit code was captured by redirecting first (cmd > log 2>&1; EXIT=$?), never through a pipe. BOTH-CWD COUNTS, before -> after for the file: (A) repo root, 'pnpm exec vitest run packages/cli/src/__tests__/app-generator.test.ts': before 'Test Files 1 passed (1)' / 'Tests 41 passed (41)' exit 0 -> after 'Test Files 1 passed (1)' / 'Tests 43 passed (43)' exit 0. (B) cwd packages/cli, 'pnpm exec vitest run --root ../.. packages/cli/src/__tests__/app-generator.test.ts': before '1 passed (1)' / '41 passed (41)' exit 0 -> after '1 passed (1)' / '43 passed (43)' exit 0. DECLARED NARROWING: (B) is the criterion-named package-dir command narrowed to this one file. I also ran the unnarrowed criterion form, 'pnpm --filter @object-ui/cli test' (cwd packages/cli, whole package): before 'Test Files 16 passed (16)' / 'Tests 246 passed (246)' exit 0 -> after 'Test Files 16 passed (16)' / 'Tests 248 passed (248)' exit 0. The narrowing was necessary because an appended path argument cannot narrow that form - Vitest unions positional filters with the script's own 'packages/cli/', so all 16 files still ran (measured: 16 files / 246 tests with the path appended). +2 exactly in every column, and no case went from passing to failing; a --reporter=verbose run at afa1722c9 lists all five cases of the block, including the untouched original two, with a check mark. LIT CONTROL: mutation writingFrom(REPO_ROOT, ...) -> writingFrom(process.cwd(), ...), one occurrence, under trap '... git checkout HEAD -- ABSOLUTE_PATH' EXIT INT TERM with the absolute path taken from git rev-parse --show-toplevel. Mutation confirmed ON DISK before either leg ran: anchor counts flipped OLD=1 NEW=0 -> OLD=0 NEW=1, and the blob hash moved - HEAD_BLOB=f50502442602fdc45a8cf07df185283cc4060158, PRE_HASH=f50502442602fdc45a8cf07df185283cc4060158 (equal, so the tree really was at HEAD), MUT_HASH=f9c5acd8baa1096eed34f439a14b6656296b23ea (differs). Neither hash was empty; the script treats an empty hash as FAILURE and exits. Leg A (cwd packages/cli): RED AND NAMES ITSELF - 'FAIL |unit| packages/cli/src/__tests__/app-generator.test.ts > generation onto disk > writes the IN-WORKSPACE map when the generator runs at a workspace root', 'AssertionError: expected { ...(8) } to deeply equal { ...(8) }', 'Tests 1 failed | 42 passed (43)', exit 1. Leg B (repo root): green, '43 passed (43)', exit 0 - which is the finding restated: under the ambient form the only invocation CI runs stays green while the branch it pins has silently changed. RESTORATION PROVED BY STATE, not by whether the trap fired: POST_HASH=f50502442602fdc45a8cf07df185283cc4060158 (equal to HEAD_BLOB), 'git diff HEAD --stat' produced 0 bytes, 'git status --porcelain' empty. NO REBUILD LEG APPLIES and none was run: the test imports the subject as '../utils/app-generator.js', a relative source specifier Vite resolves to the .ts source, so no dist/ artifact sits between the mutation and the run. TYPE-CHECK THROUGH TURBO: 'pnpm exec turbo run type-check --filter=@object-ui/cli --concurrency=2' -> 'Tasks: 10 successful, 10 total', exit 0. (A bare tsc --noEmit is not this repo's gate; the turbo task dependsOn ^build.) CHANGESET GATE, verdict lines verbatim. Before: '❌  1 source file(s) of 1 released package(s) changed, and this change adds no changeset:'. So it was owed; I added .changeset/cli-app-generator-explicit-branch-cases.md with an EMPTY frontmatter. After: '✅  1 source file(s) of 1 released package(s) changed, and this change declares 1 changeset(s): .changeset/cli-app-generator-explicit-branch-cases.md.' followed by 'Every one of them has an EMPTY frontmatter - declared as releasing nothing, which is the explicit exemption and a complete answer to this gate.', exit 0. 'node scripts/check-changeset-no-major.mjs' also exit 0: '✅  No changeset declares a `major` bump.' LINT: eslint on the changed file only - 1 file linted, 0 errors, 0 warnings, exit 0. Repo-wide 'pnpm lint' was NOT run; that is CI's run. CONTROL-CHARACTER SCAN: grep -naP over the changed test file and over the PR body found nothing (exit 1). No objectui gate script named check-nul-bytes.mjs exists in this repo, so the scan is the manual form.",
     "mcp_calls": "1 - a single search_issues, for the out-of-scope dedup check; it returned 'API rate limit already exceeded for user ID 318158314' and yielded no reading. Every other GitHub read and write in this run went through repo-scoped REST (issue body, comments, PR create, PR body read-back, this comment).",
     "open_questions": [],
     "out_of_scope_findings": [
       "NOT FILED - both dedup channels were unavailable, handing this to PM to file. Finding: a path argument appended to a package's test script silently WIDENS instead of narrowing. Measured: 'pnpm --filter @object-ui/cli test packages/cli/src/__tests__/app-generator.test.ts' ran 'Test Files 16 passed (16)' / 'Tests 246 passed (246)' - the whole package - because the script is 'vitest run --root ../.. packages/cli/' and Vitest UNIONS positional filters, so the script's own 'packages/cli/' filter re-admits every file. It exits 0 and reads as a green narrowed run, and an agent reporting 'the file is green, 246 tests' would be reporting the package's count as the file's. This is a third member of the family scripts/vitest-invocation-guard.mjs already documents as traps 1 and 2, and it is documented nowhere: I grepped AGENTS.md and the guard's docstring and neither mentions it. Related to objectui#7800 (the test script's cwd devx decision, still open) - a decision there could remove the baked-in filter and with it this hazard - so it may belong as a sub-issue of that card rather than standalone. Suggested shape: label 'finding', no assignee, no pm:queue. WHY I DID NOT FILE IT: the required search-before-filing could not be performed. The repo-scoped REST read channel works in this seat, but the REST search endpoint returned HTTP 403 for both the real query and a known-hit control query; the fallback single targeted MCP search_issues then failed with 'API rate limit already exceeded for user ID 318158314'. With no working dedup channel I did not file blind."
     ]
   }

   ---

   Generated by Claude Code
4. claude commented on Sep 5, 2026

   

   claudeBot

   on Sep 5, 2026 – with Claude

   ContributorAuthor

   More actions

   落地记录 —— PR #7810 已合入（本 lane 今日第 23 次落地）

   • merged: true @ 2026-09-05T23:32:42Z，squash ef4533727
   • 2 files, +67 / −1；closing keyword 只有 #7807
   • 卡片 closed_at 23:32:43Z；后继载体 AGENTS.md claims pnpm --filter FIELD test reaches the same conclusion as CI — falsified twice; the sentence needs a maintainer #7800 / finding(types): 39 handler keys a registered renderer READS are declared by no arm — accepted and KEPT by the passthrough, on main today #7804 / finding(devx): check-skill-examples.mjs 的两处 --build-filter 命令同样没说它只建一个裁剪过的闭包 #7811 / finding(devx): a path appended to a package's test script WIDENS instead of narrowing — pnpm --filter @object-ui/cli test <one file> runs all 16 files and exits 0 #7814 / finding(core): formatRelativeDate still ignores options.style, and its out-of-window ABSOLUTE fallback is the one place the key is already well-defined #7816 / finding(app-shell): onManageChanged reads a failed package-list refresh as "the package was deleted" and evicts the author to /home #7821 / finding(app-shell): the object-view relay hands ListView a SIX-KEY projection of the chart block, dropping the whole ADR-0021 shape (dataset / dimensions / values) #7823 / finding(app-shell): Studio's "lock the API name on first save" has no real save signal — the draft auto-save is deliberately not counted, leaving one same-mount window open #7824 全部仍 open，逐一复核过

   ⭐ 三条值得留下来的读数

   1. ⭐⭐ 点灯对照把这张卡的论点变成了一句可复读的话。

   变异 writingFrom(REPO_ROOT, …) → writingFrom(process.cwd(), …) 后：

   运行位置	结果
   cwd = packages/cli	红，并自报其名：writes the IN-WORKSPACE map when the generator runs at a workspace root，Tests 1 failed | 42 passed (43)
   cwd = 仓库根（CI 唯一跑的形式）	绿，43 passed (43)

   ⇒ 这正是卡片的论点原样复述：在 CI 唯一会跑的那个调用下它保持绿色，而它所钉住的分支已经悄悄变了。 ⭐ 一个会在事故树上空过的钉子。

   2. process.chdir() 在本仓的单元 project 下不可用 —— 这是实测，不是推测。

   pool: 'threads' 下 process.chdir() 抛 ERR_WORKER_UNSUPPORTED_OPERATION（用 worker_threads 探针在本容器实测）。⇒ 只能对恰好一次同步写调用打桩 process.cwd()，并在 finally 里还原。⛔ 想用 chdir 写这类测试的下一位不必再试一遍。

   3. 为什么加了两个用例而不是一个 —— 一个不够，理由是覆盖而不是对称。

   只加一个显式 IN-WORKSPACE 用例不能把覆盖补平：从仓库根跑时，原有的那对 ambient 用例也钉 IN-WORKSPACE，于是 STANDALONE 在 CI 唯一会跑的形式下仍然无人触及。⇒ 必须同时加一个显式 STANDALONE 用例（对着测试自己的 mkdtemp 目录），两侧指向同一条路径。

   ⚠️ 另：原有两个用例逐字节未动，整个 diff 里唯一被删的行是 vitest import（因为要加 vi）——「不动既有断言」这件事被做成了可核对的事实。

   ---

   Generated by Claude Code