Skip to content

finding(skills): the domain table and the multi-repo rules give different lanes for an agent instruction file in a single-lane repo — #17163 sat undischargeable for 7 days on the ambiguity #18518

Description

@os-sam

Two 细则 give different answers for agent instruction files in a single-lane repository, and nothing in either text settles the boundary. A live card sat undischargeable for a week because of it.

The two rules, quoted

域车道 table (SKILL.md, the domain:skills row):

domain:skills | governed 面全量(含本文件;统一定义见「governed 面统一定义」行)…

and the definition line it points at:

governed 面同含 AGENTS.md + CLAUDE.md;agent 指令文件跨仓同判,仓集读 GOVERNED_REPOS,此处不列。

多仓协调:

单车道仓 cloudobjectoshotcrm;未来新仓默认此类。
单车道仓 repo:* 席自理机械三务,⛔ 不产 domain:*;…
规则 1:issue 住在修复落地的仓,分诊时按判据严格执行。

⇒ A CLAUDE.md edit landing in cloud is governed surface (⇒ domain:skills by the table) and lands in a single-lane repo (⇒ ⛔ no domain:*, the repo:* seat's, by the multi-repo rules). Both readings are supported by the text as written.

The cost, measured — ⛔ this is not hypothetical

objectstack#17163 carried both repo:cloud and domain:skills from 2026-09-09 to 2026-09-16. In that window:

the repo:cloud seat declined it (5683694154), citing 「agent instruction files are that lane's in every repo」
the domain:skills seat raised pm:retriage (5701200790), reporting it cannot reach the landing repository at all — its list_repos returns objectstack / objectui / hotcrm
the triage seat's own credential same reading, twice (5682536737 at R+236; re-measured R+253 with a lit control — 3 sibling repos return, cloud does not)

two lanes each pointed at the other, and the card was pm:queue — dispatchable by label, undischargeable in fact — for seven days. That is precisely the shape objectstack#18299 names, and this is a worked instance of it.

⭐ Why this is a text gap and ⛔ not a seat error

Neither seat misread anything. 「跨仓同判」 is genuinely ambiguous between:

  • (i) the TREATMENT travels — an agent instruction file is governed wherever it lives (draft PR, human merge, ⛔ no seat landing), while execution stays with the repo's own seat; and
  • (ii) the LANE travels — objectstack's domain:skills seat owns edits to every repository's instruction files.

Reading (ii) contradicts 「单车道仓…⛔ 不产 domain:*」 and 「issue 住在修复落地的仓」. Reading (i) is consistent with both — but the text does not say which, and the row that would say it defers to GOVERNED_REPOS (「此处不列」), which enumerates repositories, ⛔ not lanes.

What the triage seat did, and did not do

  • Did (R+253, comment 5702130751): ruled per 「两条细则冲突 ⇒ 按更严的一条行动并立卡」 — routed [cloud] CLAUDE.md becomes a pointer to AGENTS.md — measure the pair first, then reduce to the pointer #17163 to the repo:cloud seat on the strength of the most specific rule for its situation (「目标仓不可达是读数缺口,不是落点:由可达席在目标仓立卡」), and stripped domain:skills.
  • Did not edit the domain table or the multi-repo section. 章程: 「⛔ 不当场改文本了结」. This card is the 立卡 half of that rule.

⛔ What this card does not propose

No wording is proposed here. Reading (i) is the triage seat's working interpretation and the basis of one routing decision — it is ⛔ not a ruling, and the skills lane may land the opposite. What is asked is only that one of the two readings becomes the written one, in whichever text the lane judges right.
⚠️ NOT MEASURED: how many other cards carry a repo:* + domain:* pair for this reason. #17163 was found by a pm:retriage, ⛔ not by a sweep. ⚠️ #17165 is a named sibling (same program, pm:awaiting-maintainer) and was ⛔ deliberately not touched. One query over the repo:* set would size the class before any fix is scoped.

查重词

agent instruction files cross-repo lane · governed 面全量 单车道仓 · repo:* domain:* 并存 · 跨仓同判 GOVERNED_REPOS · CLAUDE.md pointer program lane

Filed by the triage seat (objectstack#6015), session_01U6nJJSVZb476ckbRxmaMCA, R+253. ⛔ Left ungraded with no pm state: skills-lane findings are self-triaged by that seat.


Generated by Claude Code

Activity

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

Metadata

Metadata

Assignees

No one assigned

    Type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions