Repository navigation
AGENTS.md:348 counts the drift its own rule forbids as "four times", and the ledger at :353 has not been incremented since #1620
Description
Found while executing #1536 (the docs/feature-inventory.md QUO-010 / QUO-012 rows). Filed rather than swept: that card's declared file surface is docs/feature-inventory.md plus a changeset, and AGENTS.md is not on it.
Measured on origin/main @ df549fef.
What is stale
Documentation-discipline rule 5 — the rule that forbids hand-copying a machine list into prose — states a hand-maintained count of its own precedents, and carries a hand-maintained roster beside it:
AGENTS.md:348 A table transcribed into prose drifts, and this repo has measured it drifting four
AGENTS.md:349 times.
AGENTS.md:353 > Drift measured: #610, #965, #977, #1228. Boundary precedents: #1422 (documentable)
At least two further drifts of exactly that class have been measured since those four, and neither is in the ledger:
docs/feature-inventory.mdPRM-010 hand-copies the demo-staffing headcount ("3 个可登录演示账号"), and PR #1463 makes it five #1464 (closed) —docs/feature-inventory.mdPRM-010 hand-copied the demo-staffing headcount 「3 个可登录演示账号」, which a landing PR made five.docs/feature-inventory.mdQUO-010/QUO-012 creditcrm_productwith three capabilities that were removed #1536 (open, PR docs(inventory): drop the crm_product capabilities the catalog no longer has #1619) —docs/feature-inventory.mdQUO-010 / QUO-012 creditedcrm_productwithbilling_type,unit_of_measureand an inventory capability, none of which is declared. The card's own body names itself "a fifth".
A third is a candidate rather than a certainty, and wants a judgement this card does not make: #1319 (closed) recorded feature-inventory.md's OPP-009 citing a key that no longer exists. That is the same drift shape, but the card bundles it with a code-comment correction, so whether it counts as a ledger entry is a call for whoever maintains the ledger.
Why it is worth a card
The count and the roster are themselves a hand-copied machine fact in the one paragraph that forbids them, so the rule currently under-reports its own evidence. That evidence is the rule's argument: "this repo has measured it drifting four times" is what makes rule 5 more than an assertion, and the true number is at least six.
⛔ This is not a proposal to delete the ledger. The section says outright that the precedent cards are tombstones, kept so the same card is not filed again next quarter — the ledger earns its place. The question is only whether it is maintained on each landing drift, or is a snapshot of the 2026-08-31 ruling that is not meant to grow. If the latter, the sentence at :348 wants wording that does not read as a live tally.
Not in scope
- ⛔ Nothing about the QUO-010 / QUO-012 rows themselves — that is
docs/feature-inventory.mdQUO-010/QUO-012 creditcrm_productwith three capabilities that were removed #1536, already in flight as PR docs(inventory): drop the crm_product capabilities the catalog no longer has #1619. - ⛔ No opinion here on rule 5's boundary test, or on the
#1422against#1329/#1326precedents beside the ledger.
Activity
os-steve commentedon Sep 6, 2026
First-touch grading →
pm:queue+documentation+ typeTask.repo:hotcrmseat R39, 2026-09-06T00:4xZ.⭐ The anti-drift rule's own evidence has drifted. Filed by #1536's dev off its declared surface — correct,
AGENTS.mdwas not on it.Why it queues
Measured, named lines (
:348,:353), and two concrete un-ledgered instances (#1464 closed; #1536 whose own body calls itself "a fifth", landing now as PR #1619). ⇒ lane work. ⛔ Not the decision box: the open question it raises is settled below by rule 5's own criterion, not by a new convention.⭐ Seat ruling on the open question — the COUNT is the defect; the LEDGER is not
The card asks whether the ledger is a live tally to maintain, or a snapshot whose sentence should stop reading as one. Both halves resolve, and differently, under rule 5's own test:
:348's numeral — "drifting four times" — IS the rule-5 defect. It is a machine-derivable fact hand-transcribed into prose, with nothing holding the copy to the original. ⭐ That the violation sits inside the sentence forbidding it is the finding, and the fix is to stop stating a hand-maintained number — say that the drift is measured repeatedly and let the roster below carry the instances.:353's roster is NOT a rule-5 defect and ⛔ must not be deleted. It is a curated tombstone list with no machine source — nobody can derive "which cards were filed for this" from the tree. Rule 5 condemns transcribed machine lists, ⛔ not curated ones. And the section states its purpose plainly: the tombstones exist so the same card is not filed again next quarter. ⇒ a dedup list that stops growing stops working. Keep it, and extend it.⇒ Add #1464 and #1536. The rule's argument gets stronger, not weaker: "at least six" beats "four", and it is the evidence that makes rule 5 more than an assertion.
AGENTS.mdis governed and its chain is now FOUR deep⛔ Never two drafts on one hand-merged governed file. ⛔ Not dispatchable until PR #1603 merges and the chain drains in order.
⭐ Worth stating because it changes the priority of a maintainer action: four cards now sit behind one human merge, and three of them (#1618, #1620, and this lane's share of #1573) are corrections to the agent instruction file itself — the class this repo has repeatedly measured as producing confidently-wrong actions rather than loud failures. ⇒ merging PR #1603 is currently the highest-leverage single action available to the maintainer in this lane. Carried to the round report, ⛔ not nagged on the PR.
Scope note for whoever takes it
⛔ Nothing about the QUO-010 / QUO-012 rows (that is #1536 / PR #1619). ⛔ No opinion on rule 5's boundary test or the
#1422/#1329/#1326precedents beside the ledger — the card correctly excludes both.Generated by Claude Code
os-justin commentedon Sep 9, 2026
Claim: PM loop round 2 — maintainer direct dispatch to the skills seat, authority verbatim: 「看一下 hotcrm 的agents.md 写的太乱了。只留下必要的。包括同时同时审核 hotcrm 的skills 任务，合并派发处理。」 and 「首先要明确 hotcrm 是元数据应用，平台功能应该在 objectstack中开发。 其次 claude.md 是不是直接让他阅读 agents.md 即可，没必要维护两套。所有仓库都有类似的问题」. Fold member of #1823 (chain head; full claim there): rule 5's own hand-maintained drift count and precedent ledger leave the file (premise re-checked: PR #1809 rewrote
:348— the dev reads the current line, not the card's).Session:
session_01MoTv7pn338AZ71owsp19gQBranch:
claude/issue-1823-agents-md-necessary-onlyWorktree:
hotcrm-issue-1823Domain:
repo:hotcrm(single-lane repo; the skills seat acts here on the direct instruction only)File surface:
AGENTS.mdonly for this member (stop on breach; explain in the report)Container & model:
L (a 651-line instruction file rewritten to ≤250 lines + five folded cards + a new CLAUDE.md),mode:subagent(the hotcrm seat's standing 「不要用云卡片」),model: claude-fable-5-1 — --tier REFUSES for a sister repo (Families are a property of THAT repo. A card landing in another repo derives nothing here — assert with --repo to make this checkable. dispatch-gates: REFUSING ), so the tier is the seat's explicit decision: a whole-file rewrite of the agent instruction surface with necessity judgement is protocol-semantics work ⇒ contract-review tier, reviewed in the seat at the same tierClause-②: no
Thread-read: none
Serial constraints cleared:
AGENTS.mdfree per therepo:hotcrmseat's R58 close (5602546839, 13:21Z: 「AGENTS.mdis FREE. Serial queue: #1745 → #1620 → #1618」 — consumed by this fold); no open hotcrm PR touchesAGENTS.md,CLAUDE.md,.github/instructions/**or.github/tasks/**(6 open PRs' file lists read 2026-09-09T13:51Z); newestAGENTS.mdtouch onorigin/main965933bis67bd085(#1820, 2026-09-09) ⇒ same-day churn, the dev works on the merged tree; hotcrm has noscripts/pm/⇒ no H17 index, no verify lock (docs-only, no build).Generated by Claude Code
os-justin commentedon Sep 9, 2026
Landed in the #1823 fold: draft PR #1830 (governed, maintainer merge;
skip-changeset). Fullos-dev-reportis on #1823, comment 5603505779; this card's own acceptance line is in the PR body under Per-member acceptance.Generated by Claude Code
os-justin commentedon Sep 9, 2026
ACCEPT for this fold member — the review record is on the chain head #1823 (5603569687, this seat, 2026-09-09T14:30Z): PR #1830 delivers it — rule 5's drift count and ledger are gone (「measured it drifting repeatedly … This rule binds this file too」); the seat-grading conflict is stated on the PR for the maintainer. Governed terminal (draft,
needs-user-decision, 速读, reviews requested); the card stayspm:dispatcheduntil the merge lands.Generated by Claude Code
os-justin commentedon Sep 9, 2026
Landing record — PR #1830 (
Fixesthis card; fold head #1823 with #1745 · #1620 · #1618 · #1818 · #1681) MERGED 2026-09-09T15:33:23Z through hotcrm's merge queue, squash commit06b9bb327bb0onorigin/main(read at 2026-09-09T15:50Z:AGENTS.md248 lines,CLAUDE.mdis the 6-line pointer). GOVERNED (AGENTS.md,CLAUDE.md,.github/instructions/**): approved, flipped ready and enqueued by os-zhuang (review APPROVED 15:32:55Z;merged_byos-zhuang) — this seat did none of the three. In-seat review ACCEPT 5603569687 on #1823; 速读 5603571614 on the PR. Recorded by the skills seat acting in therepo:hotcrmlane on the maintainer's direct instruction (audit note on the hotcrm seat post #1353, comment 5603003979), sessionsession_01MoTv7pn338AZ71owsp19gQ.Same stroke:
pm:dispatchedremoved and the assignee cleared on this auto-closed card (folded member; its own criterion — theAGENTS.mddrift-count line is gone — was read on the merged tree).Generated by Claude Code