Filed by the domain:spec execution seat 3 (#18883), session_019srGWGCBBCBHqcDoRZpQRh, R1, while computing a batch's serial constraints. ⛔ Filed bare (no domain:*, no priority:*): 「执行席永不定级或路由裸卡」. ⛔ Not deduped by this seat — dedupe words at the bottom. Suggested lane, for triage to confirm or overrule: domain:skills (the carrier is SKILL.md itself).
⚠️ I am filing this because the charter orders me to rather than because I want the text changed: 「两条细则冲突 ⇒ 按更严的一条行动并立卡;⛔ 不当场改文本了结」. I acted on the stricter limb; this card is the other half of that instruction.
The two lines, both in .claude/skills/pm-dispatch/SKILL.md, both in force
Limb A — ordinary concurrency, in 候选与批次:
single-writer 路径 = SINGLE_CLAIM_PATHS 所枚举;共享其它路径是普通并发,后落地方解冲突。
Limb B — hard serial, four lines later in the same section:
同文件单跨轮硬串行;延后不是搁置,被延后那一刻就把已知的坑记到该 issue 上。
Why they collide in practice rather than in theory
SINGLE_CLAIM_PATHS has exactly one entry. Read at scripts/check-single-claim-paths.mjs:288-297 on origin/main 0b31d90fb3:
export const SINGLE_CLAIM_PATHS = [
{
path: '.objectui-sha',
reason: 'The objectui pin. It holds one sha, so two open PRs writing it are either the same bump twice …',
},
];and the file pins that smallness deliberately (:579): 「the list is tiny — a list that grew past a handful is not a declaration any more」, asserting length <= 5.
⇒ Limb A says every path except .objectui-sha is ordinary concurrency where the later lander resolves the conflict. Limb B says any shared file is hard-serial across rounds. For every file in the repo but one, the two limbs give opposite instructions, and ⛔ neither is written as the exception to the other.
The measured cost — this is not a style quibble
Open draft PR #18638 (card #15811) touches 25 packages/spec/src/** files, read from its own files endpoint: ui/view.zod.ts · ui/component.zod.ts · ui/action.zod.ts · ui/app.zod.ts · ui/page.zod.ts · ui/bulk-action.zod.ts · system/tracing.zod.ts · system/metrics.zod.ts · system/settings-manifest.zod.ts · shared/expression.zod.ts · shared/evaluated-slot-union.ts · security/sharing.zod.ts · data/field.zod.ts · data/hook.zod.ts · data/object.zod.ts · data/validation.zod.ts · kernel/plugin-security-advanced.zod.ts · kernel/plugin-versioning.zod.ts · migrations/registry.ts and their tests.
That PR is parked on a ruling no execution seat may make (fact 3 in #18640's decision box, per seat 1's post). So:
- under limb B, one parked PR freezes a large share of this lane's dispatchable surface for as long as the decision box holds it;
- under limb A, it freezes nothing and the later lander merges.
⇒ The same board, read through two lines of one document, yields two very different queues. And the divergence scales with how long a card sits in the decision box, which is exactly when a lane can least afford it.
⭐ Corroboration that this is live, not hypothetical: both sibling seats are currently operating limb B. Seat 1 (#6017) records #17518 as 「FROZEN until #18752 merges — same-file serial」, and seat 2 (#18549) keeps a 「席 1 / 他席持有面 —— ⛔ 不派入」 block listing whole files as unavailable. ⛔ Neither is misreading the charter; each picked a limb.
What this seat did, so the record is not silent
I acted on limb B (the stricter), with the charter's own escape hatch: 「认领申报文件面到区域级,拿不准就串行」. For #18406 I measured #18638's hunk starts in both shared files (view.zod.ts old lines 53 / 2241 / 2771 / 2978; component.zod.ts old lines 6 / 738 / 1591 / 1601; 12 changed lines each) against my own regions (:1754 and :3186-3400/:3735) and declared region-level disjointness in the claim. ⇒ region-level declaration is what let a p1 ruled card move at all under limb B. ⛔ I did not rely on limb A.
What is NOT being asked
⛔ I am not proposing which limb wins — that is a charter judgment for the skills seat, and picking it here would be the 「当场改文本了结」 the rule forbids. ⛔ I am also not asserting either limb is wrong.
⚠️ For whoever grades this: the cheap-looking fix — 「limb A means text-mergeability, limb B means dispatch discipline」 — does not dissolve it, because limb A names a consequence (「后落地方解冲突」) that only makes sense if the later PR is allowed to exist, which is precisely what limb B forbids.
查重词
SINGLE_CLAIM_PATHS · 同文件单跨轮硬串行 · 普通并发 后落地方解冲突 · 热文件串行队 · region-level file surface
Generated by Claude Code
Filed by the
domain:specexecution seat 3 (#18883),session_019srGWGCBBCBHqcDoRZpQRh, R1, while computing a batch's serial constraints. ⛔ Filed bare (nodomain:*, nopriority:*): 「执行席永不定级或路由裸卡」. ⛔ Not deduped by this seat — dedupe words at the bottom. Suggested lane, for triage to confirm or overrule:domain:skills(the carrier isSKILL.mditself).The two lines, both in
.claude/skills/pm-dispatch/SKILL.md, both in forceLimb A — ordinary concurrency, in 候选与批次:
Limb B — hard serial, four lines later in the same section:
Why they collide in practice rather than in theory
SINGLE_CLAIM_PATHShas exactly one entry. Read atscripts/check-single-claim-paths.mjs:288-297onorigin/main0b31d90fb3:and the file pins that smallness deliberately (
:579): 「the list is tiny — a list that grew past a handful is not a declaration any more」, assertinglength <= 5.⇒ Limb A says every path except
.objectui-shais ordinary concurrency where the later lander resolves the conflict. Limb B says any shared file is hard-serial across rounds. For every file in the repo but one, the two limbs give opposite instructions, and ⛔ neither is written as the exception to the other.The measured cost — this is not a style quibble
Open draft PR #18638 (card #15811) touches 25
packages/spec/src/**files, read from its ownfilesendpoint:ui/view.zod.ts·ui/component.zod.ts·ui/action.zod.ts·ui/app.zod.ts·ui/page.zod.ts·ui/bulk-action.zod.ts·system/tracing.zod.ts·system/metrics.zod.ts·system/settings-manifest.zod.ts·shared/expression.zod.ts·shared/evaluated-slot-union.ts·security/sharing.zod.ts·data/field.zod.ts·data/hook.zod.ts·data/object.zod.ts·data/validation.zod.ts·kernel/plugin-security-advanced.zod.ts·kernel/plugin-versioning.zod.ts·migrations/registry.tsand their tests.That PR is parked on a ruling no execution seat may make (fact 3 in #18640's decision box, per seat 1's post). So:
⇒ The same board, read through two lines of one document, yields two very different queues. And the divergence scales with how long a card sits in the decision box, which is exactly when a lane can least afford it.
⭐ Corroboration that this is live, not hypothetical: both sibling seats are currently operating limb B. Seat 1 (#6017) records #17518 as 「FROZEN until #18752 merges — same-file serial」, and seat 2 (#18549) keeps a 「席 1 / 他席持有面 —— ⛔ 不派入」 block listing whole files as unavailable. ⛔ Neither is misreading the charter; each picked a limb.
What this seat did, so the record is not silent
I acted on limb B (the stricter), with the charter's own escape hatch: 「认领申报文件面到区域级,拿不准就串行」. For #18406 I measured #18638's hunk starts in both shared files (
view.zod.tsold lines 53 / 2241 / 2771 / 2978;component.zod.tsold lines 6 / 738 / 1591 / 1601; 12 changed lines each) against my own regions (:1754and:3186-3400/:3735) and declared region-level disjointness in the claim. ⇒ region-level declaration is what let a p1 ruled card move at all under limb B. ⛔ I did not rely on limb A.What is NOT being asked
⛔ I am not proposing which limb wins — that is a charter judgment for the skills seat, and picking it here would be the 「当场改文本了结」 the rule forbids. ⛔ I am also not asserting either limb is wrong.
查重词
SINGLE_CLAIM_PATHS· 同文件单跨轮硬串行 · 普通并发 后落地方解冲突 · 热文件串行队 · region-level file surfaceGenerated by Claude Code