[finding] repoRelativeLiterals() 的顶层目录白名单缺 docs,而 docs/… 是可声明的 glob(turbo.json 今天就声明着两条)⇒ 引用整条 docs/… 路径的测试对扁平收集器不可见,门禁读作合规 #18342

Activity

1. os-try-charles

   added

   ci/cd

   tooling

   finding

   on Sep 16, 2026
2. os-bill commented on Sep 18, 2026

   

   os-bill

   on Sep 18, 2026 – with Claude

   Collaborator

   More actions

   Claim: PM loop round 27
   Session: session_01JbZnqu8bt6YqfJsr9vaFb3
   Branch: claude/issue-18342-repo-relative-toplevel-whitelist
   Worktree: objectstack-issue-18342
   Domain: domain:spec
   Seat: domain:spec#2(座位贴 #18549)
   File surface: scripts/check-cross-package-test-inputs.mjs —— ⚠️ 开放并预先申报:.changeset/*.md 与门禁反向要求的派生物。⛔ 只读栅栏见下。
   Container & model: M, mode:subagent, model: default judgement tier
   Clause-②: no (narrowing)
   Thread-read: 5712639960

   ⏱️ 2026-09-18T02:02Z 取数。本席对当前全部 25 个 open PR 逐个算 merge-base..head:scripts/check-cross-package-test-inputs.mjs 与 turbo.json 的持有者各 0 个。⭐ 亮控:同一把仪器对 packages/spec/scripts/build-schemas.ts 读出 [18863] ⇒ 仪器活着,那两个 0 是真零。

   ---

   ⛔ 先说三处卡面已经漂了的地方,⛔ 不要照卡面的数字做

   ⏱️ 2026-09-18T02:02Z 现读 origin/main 18cc3b1dfc:

   卡面写的	现读
   白名单在 :1185	:1369 —— 文件动过了。⇒ ⛔ 行号在你开工那一刻自己取,⛔ 不抄卡、⛔ 不抄本令
   卡面与分诊都写「turbo.json 声明着两条」docs/…(分诊点名 :547 / :633)	三条,而且行号也动了::332 docs/qa/platform-checklist/areas/*.json · :549 docs/audits/2026-07-unknown-key-strictness-ledger.md · :635 docs/**
   验收第 2 条把「还有没有第三个同形缺口」留给承接者去量	⭐ 本席量了,答案是「有,而且不止一个」 —— 见下

   ⭐⭐ 验收第 2 条的答案(本席现读,⛔ 但你仍要自己复量)

   判据取头注自己那句话 —— :1362-1363:

   The list below is every top-level directory a declared glob can name; a new one added to the tree belongs here too.

   ⇒ 把 turbo.json 里每一条 $TURBO_ROOT$/… 的首段拿出来数,和白名单 packages|apps|examples|content|scripts|skills 对齐:

   turbo.json 声明到的首段	条数	白名单认得?
   packages	86	✅
   scripts	55	✅
   examples	18	✅
   content	16	✅
   skills	4	✅
   docs	3	⛔ 不认得 —— 卡面这一条
   .github(:504 .github/workflows/scaffold-e2e.yml)	1	⛔ 不认得 —— ⭐ 卡面没点名的第二个
   .claude(:550 .claude/skills/spec-property-retirement/SKILL.md)	1	⛔ 不认得 —— ⭐ 卡面没点名的第三个
   sdui.manifest.json(:201)	1	⛔ 不认得,⚠️ 但它是仓根一个文件、不是目录 —— 形状不同,见下

   ⚠️⚠️ .claude 那一条要单独当心:.claude/** 是本仓的受管面。一道开始能看见 .claude/… 字面量的门禁,爆炸半径和 docs/ 不是一回事。⇒ 你必须分别量 .github 与 .claude 打开后各自让门禁看见了什么,⛔ 不许把三个目录当一个改动一把塞进去然后只报一个总数。

   ⚠️ sdui.manifest.json 是第四种形状,⛔ 不是第四个目录:头注承诺的是「every top-level directory」,而它是仓根文件。⇒ 头注这句话是否也该覆盖仓根文件,是一个真问题:你判,并在正文里把判断和代价写出来;⛔ 不许默默把它塞进同一条正则,也⛔ 不许默默漏掉。

   ⚠️ 反方向一条,⛔ 免得你当成缺陷:白名单里的 apps 在 turbo.json 里被声明 0 次。⭐ 但 apps/ 是树上真实存在的顶层目录(git ls-tree origin/main 读得到),而头注说的是「a declared glob can name」⇒ ⛔ 这不是缺陷,不要动它。

   本轮要收的东西

   1. 按验收第 2 条的实际读数补白名单 —— ⛔ 不是「把 docs 塞进去就收工」。
   2. ⭐ 验收第 3 条:让头注与白名单机械同步。 ⚠️ 这道门禁已经有自检骨架可用:SELF_TEST_BATTERIES(:401)+ SELF_TEST_BATTERY_FLOOR(:425,现值 12)+ registerCase()。⇒ 新增的电池要登记进 roster 并把地板相应抬高,⛔ 光加 case 不登记等于没有地板。
      ⚠️⚠️ 一个顺序陷阱:若你把一致性检查写成「从 turbo.json 实际出现过的首段反推」,它今天就会因为 .github / .claude 而红。⇒ 它必须和白名单修复同一笔落地,⛔ 不能先落检查再落修复。
   3. ⛔ 不许为了让新覆盖变绿去放宽门禁自己的判据或删自检用例(卡面验收第 4 条,原样转达)。

   ⛔ 只读栅栏与停手线

   • ⛔ turbo.json —— 它是本轮的判据来源,⛔ 不是本轮的改动对象。改它等于改考题。
   • ⭐⭐ 停手线(本轮最重要的一条):卡面「⛔ 没量的部分」自己写着 —— 把这些目录打开后,此前不可见的引用会变可见,其中可能有真的未声明项,而那是「要逐条处置的活,⛔ 不是一行改动」。
     ⇒ 若门禁在补全白名单后对树上现存的引用变红:停下来,把普查交回本席 —— 按目录、按文件、按引用逐条列出,附你判断它是真违规还是假阳。⛔ 不要在本轮里把它们一条条修掉,⛔ 也不要为了让它绿而缩小白名单。⭐ 本席会据你的普查另立卡。
     ⚠️ 反过来:若一条都没变红,那也是读数,而且要带亮控 —— 证明你的新白名单确实能看见东西(见下 LIT),⛔ 不要把一个没被证明能发火的绿当成成功。

   验收

   • ⭐ LIT(必须发火,两个方向都要):① 造一条落在未声明子树里的 docs/…(以及 .github/…、.claude/…)引用 ⇒ 门禁必须红;② 把它声明掉 ⇒ 必须绿。⛔ 只证 ② 不算(卡面验收第 1 条原文)。⚠️ 三个目录各自都要有自己的 LIT,⛔ 不许拿一个证另两个。
   • ⭐ DARK(必须读 0):改前改后对整树跑同一道门禁,已经可见的那些字面量的判定逐字不动。给出改前/改后两组读数,⛔ 不要只说「没变」。
   • ⭐ 收集器行为的直接读数,照卡面 ③ 的做法逐条喂正则并打印结果(⛔ 打印,不是数数):'docs/…' · '.github/…' · '.claude/…' · 发火 'scripts/js-comment-mask.mjs' · 发火 'content/docs/deployment/cli.mdx' · 暗控 'zzqq/nope.md'。
   • --self-test 全绿,且既有 12 条电池一条不少。

   本轮章程增量

   1. PR 正文首行 Fixes #18342;Clause-②: no (narrowing) 行单独占一行、写在行首,⛔ 不加反引号。
   2. ⚠️ 本席申报 no (narrowing) 是有依据的判断、⛔ 不是省事:章程 references/lanes/spec.md:19 —— 「放宽接受集或扩大公开面的卡…即条款②;收窄仍是语义面,不触条款②」。本轮门禁的接受集只会变小(看见得更多 ⇒ 拒得更多),且 scripts/** 在任何包的 files[] 里都没有条目 ⇒ 不碰已发布契约面(SKILL.md:514:「条款②只指已发布契约面」)。⭐ 若你量出任何已发布面被动到,那就是申报错了:停下来交回本席,⛔ 不要自己把它改成 yes。
   3. ⭐ 探针做在你保留的分支上,⛔ 不开一次性分支(容器建得出远端分支却删不掉,git push --delete 与 DELETE /git/refs/… 两道皆 403,⛔ 不重试)。
   4. 提交尾部 ⛔ 不带卡号 trailer;trailer 对不含模型名;⛔ 不设 OS_ALLOW_CARD_TRAILER_PUSH=1,要改用 amend。
   5. PR 正文只写一次,⛔ 不事后 PATCH;要改的在报告里点名,本席代写。
   6. 门禁清单取 node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack 逐条跑,退出码先落盘再 --ran 对账。
   7. ⛔ 上面每一条读数都是本席 ⏱️ 2026-09-18T02:02Z 在 origin/main 18cc3b1dfc 上取的。若你的 base 已经移动,重取,⛔ 不要继承本席的数字当成你自己的读数。

   ---

   Generated by Claude Code
3. os-bill

   added

   pm:dispatched

   and removed

   pm:queue

   on Sep 18, 2026
4. os-bill

   self-assigned this

   on Sep 18, 2026
5. os-bill commented on Sep 18, 2026

   

   os-bill

   on Sep 18, 2026 – with Claude

   Collaborator

   More actions

   os-dev-report

   {
     "issue": 18342,
     "status": "needs_decision",
     "branch": "claude/issue-18342-repo-relative-toplevel-whitelist",
     "pr": "https://github.com/objectstack-ai/objectstack/pull/18869",
     "session": "session_01JbZnqu8bt6YqfJsr9vaFb3",
     "premise_still_valid": true,
     "summary": "Base had moved to f9a9e76fee, so every reading was re-taken; the seat's counts reproduced exactly (86/55/18/16/4 recognised, docs 3, .github 1, .claude 1, sdui.manifest.json 1) and the card's four collector probes reproduced verbatim. repoRelativeLiterals() now builds its regex from an exported RECOGNISED_TOP_LEVEL_DIRECTORIES list that gains docs and .claude; turboRootTopLevelNames() reads turbo.json back and splits it against the git index; a new self-test battery reconciles the two so the header note is checked rather than promised (roster registered, floor 12 to 13, both in the same commit so the ordering trap does not bite). turbo.json was read as criterion only, never modified. .github is measured and DEFERRED, not added: it is the stop-line case and the disposition is the seat's, see open_questions.",
     "tests": "GATES: all 30 families from 'dispatch-gates --commands --repo objectstack-ai/objectstack' run, exit codes captured to disk before any pipe, reconciled with --ran in the tool's own 'command :: exit N' format — it reports '30 derived, 30 run, 0 NOT-MEASURED, 0 UNRUN (a DERIVED zero — none of them is 3)'. Three (check-comment-mask-corpus, check:cli-command-ids, check:driver-memory-census) first returned exit 3 PREREQUISITE NOT MET (no node_modules in a fresh worktree — nothing measured, not a finding); after pnpm install all three re-ran green. SELF-TEST: 229 cases before, 255 after — exactly the +26 the new battery pins; 14 batteries, the 12 pre-existing floors intact. GATE: exit 0, whole-tree output byte-identical before and after ('OK: 29 package(s) ... 13 walked root(s) judged, 3 on ACCEPTED_WALK_RADII; 2023 test file(s) ...'). DARK at literal granularity via findEscapingPackages(): 219 rows before, 220 after, REMOVED 0 — one addition only, the bright control @objectstack/plugin-auth -> docs/qa/platform-checklist/areas/identity-auth.json, already inside that package's declared glob. So the green is proven able to fire, not merely quiet. LIT, both directions, one per directory, each injecting a flat literal into a test that already escapes in a package whose globs are known (no turbo.json edit needed): docs undeclared 'docs/protocol-upgrade-guide.md' exit 1 naming the file / docs declared 'docs/qa/.../automation.json' exit 0; .claude undeclared '.claude/launch.json' exit 1 naming the file / .claude declared '.claude/skills/spec-property-retirement/SKILL.md' exit 0; .github pending exit 0 and .github promoted exit 1 naming all three literals — proving the flat collector is green on .github because the entry is PENDING, not because it is blind. MUTATION DISCIPLINE: every probe ran under a trap restoring on EXIT INT TERM with absolute paths; each mutation proved on disk by git hash-object before/after (a no-op would have been reported as a void reading, not silently re-run) and each restore proved by blob hash equality against HEAD plus an empty 'git diff HEAD', never by an exit code. The per-directory probes ran on pristine HEAD content; the LIT legs ran only AFTER the fix was committed, so restore had a real commit to come back to. LINT, proven narrowing: population 6848 files read from eslint's own config via --format json; narrowed run 1 file, 0 errors 0 warnings; invariance — eslint.config.mjs:328 states there is no parserOptions.project and no typed rules, so this diff cannot move the verdict on any untouched file. Union taken at final HEAD e0f0b9c536. Control-char self-scan of the edited file: clean.",
     "mcp_calls": "0 — no MCP GitHub tool was called, read or write; all GitHub traffic went through the REST proxy with curl/urllib and GITHUB_TOKEN",
     "api_writes": "3 REST writes: POST /repos/objectstack-ai/objectstack/pulls (draft PR 18869), POST /repos/objectstack-ai/objectstack/issues/18869/labels (skip-changeset, additive endpoint, read back — union minus read-back empty, nothing stripped; size/m was added concurrently by the size-labeler and left alone), POST /repos/objectstack-ai/objectstack/issues/18342/comments (this report). Plus 2 git pushes (the empty-branch write-routing probe, then the implementation commit) — git, not REST. No PATCH of the PR body. Budget not exceeded.",
     "open_questions": [
       {
         "question": ".github is the stop-line case: adding it to the whitelist turns the gate RED on two tree literals, and per the dispatch I neither fixed them nor shrank the whitelist. Both are over-collection, NOT reads — packages/cli/test/scaffold-ci-script-parity.test.ts:17 names '.github/workflows/ci.yml' inside a BLOCK COMMENT (its real read is line 74, a reconstructed path under packages/, already declared), and packages/create-objectstack/src/template-consistency.test.ts:376 asserts the SCAFFOLDED project contains that path. Neither opens this repo's workflow; both collide with it because a scaffolded project's CI file has the same repo-relative spelling, and statSync confirms the collision as a real file. Note .github is NOT invisible meanwhile: the reconstruction half already rosters .github/workflows/scaffold-e2e.yml for create-objectstack (declared on create-objectstack#test), so what is deferred is the FLAT spelling alone. How should .github be disposed?",
         "options": [
           "A — keep the landed shape: .github stays in TOP_LEVEL_DIRECTORIES_PENDING carrying its reason, whose exact keys the new battery pins, so the omission is declared and loud rather than silent. Cost: the flat collector stays blind to .github, so a genuine future flat .github read would go unseen until the entry is promoted.",
           "B — promote .github and declare the two globs. Cost: it declares an input that is not one — folding every edit of this repo's CI workflow into @objectstack/cli's and create-objectstack's test hashes to hash a file neither test opens. This is the 'one card at a time' work the dispatch explicitly reserved, so I did not do it.",
           "C — promote .github and remove the false positives at the source instead (mask comments for the flat collector, and/or stop the assertion literal from colliding). Cost: strictly larger than this card, and it changes the collector's documented 'over-collection is harmless' premise, which deserves its own card."
         ],
         "recommendation": "A for this PR, then a follow-up card for C. A is the only option that keeps the gate green while leaving nothing silent, and the census is now on the record with its reason pinned by the battery. C is the real fix, because the two literals expose a genuine limitation — the flat collector reads comments and assertion strings as paths, which is harmless for docs/ but not for a directory whose paths collide with scaffolded output. B is the one I would avoid: it buys green by writing down an input that does not exist."
       },
       {
         "question": "sdui.manifest.json is a FOURTH SHAPE, not a fourth directory — the header note promises 'every top-level DIRECTORY'. I judged it, per the dispatch, rather than folding it in silently or dropping it.",
         "options": [
           "Scope the note to directories and make the repo-root-file class LOUD (what landed): TURBO_ROOT_FILE_INPUTS pins the repo-root files turbo.json declares and the battery asserts the live reading equals it, so a SECOND one reds instead of arriving unnoticed.",
           "Extend the regex to match repo-root files by exact name."
         ],
         "recommendation": "The landed option. A directory anchors the match on its own name followed by a slash, which is what separates docs/x.md from prose; a repo-root file's whole repo-relative path is a BARE FILENAME, indistinguishable from any other quoted string. The cheapest rule that would reach sdui.manifest.json — collect a quoted dotted word — also collects 'package.json', which statSync then confirms as a real repo-root file, so every test that merely names that string would roster it. That is the point where this collector's 'over-collection is harmless' premise breaks: it would force a declaration on a file nothing reads."
       }
     ],
     "out_of_scope_findings": [
       "to file (class (a); dedupe words: repoRoot findUp hand-rolled seed unrecognised, cross-package-test-inputs escape, live-dialect-matrix budget, .github/workflows/ci.yml undeclared) — packages/drivers/driver-sql/src/live-dialect-matrix.budget.test.ts:107 genuinely reads join(repoRoot(), '.github/workflows/ci.yml'), and check:cross-package-test-inputs cannot see that file AT ALL: its repoRoot() at line 38 is a hand-rolled for(;;) walk rather than one of the recognised findUp seed spellings, so the file yields no escape and contributes nothing to the roster. The read is undeclared today and the gate is green over it — the same failure direction as this card (reads as compliant) reached by a DIFFERENT mechanism (seed recognition, not the top-level whitelist), which is why it is not fixed here. Named failing probe: the literal does not appear in the findEscapingPackages() dump either before or after this change, while a declared sibling read in the same package does.",
       "noted, not filed: packages/spec/src/migrations/migrations.test.ts names 'docs/protocol-upgrade-guide.md' in three places and still contributes no roster entry, because that file produces no escape. Not a defect — the flat collector only contributes for files that already escape via a path expression — recorded only so the next reader does not mistake its absence from the dump for a collector bug. Successor who will touch this file: whoever takes the driver-sql finding above, same gate, same dump.",
       "noted, not filed: the PR body carries the platform-appended session-URL footer in addition to the harness attribution I sent; read-back proves the body is my text PLUS that tail, with no mutation, so it is the documented per-surface behaviour and not a body to repair. Successor: none."
     ]
   }
   

   ---

   Generated by Claude Code
6. os-bill commented on Sep 18, 2026

   

   os-bill

   on Sep 18, 2026 – with Claude

   Collaborator

   More actions

   席位裁定:报告 5724348714 的两个 open_questions

   ⏱️ 2026-09-18T02:55Z。两问都归本席(派发令写着「撞上就停下把普查交回本席」,你照做了)。⛔ 两问本席都自己复量过再判,⛔ 不按你的叙述判。

   ① .github 的处置:A —— 保持你交付的样子,并由本席另立 C 的卡

   你给的两条「过度收集」,本席逐条验过,逐字复现

   ⏱️ 2026-09-18T02:55Z 于 origin/main 1a6bc8e8be:

   站点	本席现读	单位 / 什么会让它不是这个值
   packages/cli/test/scaffold-ci-script-parity.test.ts:17	该行在一个 /** … */ 块注释里(注释开于 :3,到 :17 仍未闭合)⇒ 散文,⛔ 不是路径表达式	单位:该文件中位于块注释内的字面量数。若 :3 那个块在 :17 之前闭合,它就会是一条真代码字面量 —— 本席逐行读到 :22 确认未闭合
   同文件 :74	真正的读是 resolve(HERE, '../../..', 'packages/create-objectstack/src/templates/blank/.github/workflows/ci.yml') ⇒ 重建路径,落在 packages/ 下	而它上面 :71-73 的注释自己写着两条 glob 已为 @objectstack/cli 声明并镜像进 turbo.json
   packages/create-objectstack/src/template-consistency.test.ts:375	expect(scaffolded).toContain('.github/workflows/ci.yml'); ⇒ 对被脚手架出来的项目的断言,⛔ 不是对本仓文件的读
   两个文件里对 .github 路径的 fs 调用	0	⭐ 亮控:同一把 grep 对任何 fs 调用在两文件各读出 3 与 17 ⇒ 这个 0 是读数,⛔ 不是死搜索

   ⇒ 两条都不是读。 判定成立。

   ⭐ 而本席另外量到一条,它让 A 比你论证的还要硬

   你在 out_of_scope_findings 里交回的那条,本席复量并确认:packages/drivers/driver-sql/src/live-dialect-matrix.budget.test.ts:107 真的读本仓自己的工作流 ——

   const ci = readFileSync(join(repoRoot(), '.github/workflows/ci.yml'), 'utf8');
   

   而它 :38-45 的 repoRoot() 是一段手写的 for(;;) 上行找 pnpm-workspace.yaml,⛔ 不是被识别的 findUp 种子拼法。

   ⇒ ⭐⭐ 所以:树里确实存在一条真的 flat .github 读,而提升 .github 抓不到它 —— 它根本没产生 escape,压根进不了名册。
   ⇒ ⇒ B 的账本因此更难看:B 花的代价是写下两个不存在的输入,换到的却不是那条真读。⚠️ 本席未重跑 findEscapingPackages() 的 dump 复验「改前改后都不出现」,那一条点名取自你的报告,⛔ 本席标记为未验证。

   裁定与理由

   • A,理由不是「最省事」,是 B 会让门禁往另一个方向撒谎:它会把本仓 CI 工作流的每一次编辑折进 @objectstack/cli 与 create-objectstack 的测试输入哈希,为了一个两者都不打开的文件。⭐ 一道过度声明的门禁,教会下一个读者「声明不作数」——那比一个声明了自己没有的保护更难修。
   • C 是真修,但它比本卡严格更大,而且要动收集器「over-collection is harmless」这条自述前提 ⇒ 另立卡,由本席立。派发令本来就写着「⛔ 不要在本轮里把它们一条条修掉」。
   • ⭐ A 的代价本席认下并记明:flat 拼法对 .github 保持盲。但它⛔ 不是全盲 —— 重建那一半已经为 create-objectstack 名册了 .github/workflows/scaffold-e2e.yml(你量的,本席采信并点名来源)⇒ 被推迟的只有 flat 拼法这一种。
   • ⭐ 而 A 的关键条件是你已经做到的那条:TOP_LEVEL_DIRECTORIES_PENDING 带着理由、且那几个键被新电池钉住 ⇒ 这个省略是被声明的、响的,⛔ 不是静默的。若它只是被漏掉而没有这条钉子,本席会判 C 而不是 A。

   ② sdui.manifest.json:采纳你落地的那一个

   ⇒ 把头注限定到目录,并用 TURBO_ROOT_FILE_INPUTS 把仓根文件这一类钉响(第二个出现时变红,⛔ 不是悄悄到达)。

   本席同意你的判据,并复述本席认为决定性的那一句:目录靠「自己的名字后面跟一个斜杠」锚定匹配,而仓根文件的整条仓内路径就是一个裸文件名,和任何被引号括起来的字符串无法区分。 ⇒ 最便宜的那条能够到 sdui.manifest.json 的规则(收集带点的引号词)同时会收集 'package.json',而 statSync 会确认它是真实存在的仓根文件 ⇒ 任何只是提到那个字符串的测试都会被名册进去。那正是「over-collection is harmless」这条前提破掉的地方。

   ⚠️ 本席同时记下它的代价,免得下一个读者以为这条没有成本:头注从此不覆盖仓根文件,而 turbo.json 今天就声明着一个(sdui.manifest.json,:201)。⇒ 那条覆盖是被 TURBO_ROOT_FILE_INPUTS 的钉子换来的,⛔ 不是被消除的。

   接下来

   • ⛔ 本 PR 不需要任何改动。 你交付的就是裁定的形状。⇒ 本席按常规走落地流程(CI 收敛 → 翻 ready → 入队)。
   • ⭐ 两张卡由本席立,⛔ 不由你:① C(在源头消除 flat 收集器的假阳:给它遮蔽注释、并处理断言字面量的碰撞);② driver-sql 那条(repoRoot() 的手写上行不被种子识别 ⇒ 一条真读对门禁完全不可见)。⚠️ 你那条 noted, not filed(migrations.test.ts 提 docs/protocol-upgrade-guide.md 却不产生名册项)本席也收下,作为 ① 的输入。
   • ⚠️ 你报告里那条关于 PR 正文页脚的观察,本席确认是已记录的平台行为(裸 REST 建 PR 时按送出体尾部追加),⛔ 不是要修的正文。

   _{⏱️ 2026-09-18T02:55Z · domain:spec seat 2 · 座位贴 #18549 · 读数取自 origin/main 1a6bc8e8be}

   ---

   Generated by Claude Code
7. os-bill

   closed this as completedin #18869on Sep 18, 2026
8. os-bill

   removed

   pm:dispatched

   on Sep 18, 2026
9. akarma-synetal

   added a commit that references this issue on Sep 28, 2026

   fix(gates): recognise docs/ and .claude/ path literals, and reconcile…

   ...

   This issue will close once commit 475042c is merged into the 'main' branch.

   475042c