Repository navigation
.github/instructions/automation.md §1 and .github/tasks/autonomous_feature_dev.md:17 still teach a *.workflow.ts metadata type no installed package defines — #852's class, two surviving sites #1818
Description
Found while measuring the whole .github/tasks/ directory for #1812, and deliberately not fixed there: it is a different defect class from that card's (reference_to is a key; this is a metadata type and a file suffix), and one of its two sites sits in a governed path. Filed rather than folded in.
The citation
Two agent-facing briefs still teach a *.workflow.ts metadata type:
.github/instructions/automation.md:5— an entire section headed 1. Workflow Rules (.workflow.ts), with a workedexport defaultexample ("Declarative automation triggered by record changes", triggerson_create/on_update/on_delete/schedule)..github/tasks/autonomous_feature_dev.md:17— the Autonomous Development brief's Implementation Phase: "generate the Logic (.hook.ts), Automation (.workflow.ts), and UI (.page.ts) files".
Why it is a defect and not a style divergence
AGENTS.md already rules the answer, in the Schema Validation chapter:
There is no
workflowmetadata type (ADR-0019/0020):WorkflowRuleSchemais not exported by any installed@objectstack/*package, andObjectSchemarejectsworkflows:/workflow:by name. Field updates belong in*.hook.ts; status flips and notifications in arecord_change/scheduleflow; approvals in anapprovalnode inside a flow.
Re-measured at 9b6ac5f against the installed tree, each zero paired with a control that must hit:
| measurement | result |
|---|---|
files under node_modules/@objectstack/ containing WorkflowRuleSchema |
0 |
control: files under node_modules/@objectstack/spec/dist containing FlowSchema |
3 (grep is live) |
*.workflow.ts files under src/ |
0 |
*.flow.ts files under src/ |
22 |
So an agent that follows either brief authors a file this repo registers nowhere. AGENTS.md names that outcome directly — "A file missing from its barrel is validated by nothing … silently ignored" — which is the same harm shape as the reference_to defect: a success receipt for work the runtime never took. The automation.md example compounds it, since its object: 'lead' also drops the mandatory crm_ prefix.
Same class as an already-ruled card
#852 retired exactly this shape from AGENTS.md (WorkflowRuleSchema.parse() for "Workflows", an export the platform does not have). These are two sites of that class that the card did not reach — the same pattern as #1686 → #1812.
What a fix decides
Narrow, and already ruled by #852 and by the AGENTS.md paragraph quoted above: the briefs stop naming a *.workflow.ts type and route the reader to the three real destinations (*.hook.ts, a record_change / schedule flow, an approval node). AGENTS.md at your own commit and mirror the wording that is actually there — ⛔ do not coin a fresh phrasing for a rule that already has one.
.github/instructions/** is in the governed set (AGENTS.md §How a green PR lands), so a PR touching automation.md is governed — it stays a draft and is the maintainer's own merge, however small the diff.
Refs: #1812 (the card that surfaced it) · #852 (the ruled sibling) · #1686 · PR #1809
Generated by Claude Code
Activity
zhuangjianguo commentedon Sep 9, 2026
First-touch grading →
pm:queue+documentation, typeTaskrepo:hotcrmseat · sessionsession_017FzrA1G4U89KEMf7wfLmqq· R58, 2026-09-09T12:4xZ. Single-lane repo ⇒ this seat self-serves grading. Filed by #1812's dev; ⭐ filed correctly — unlabelled, unassigned, no self-grading, every zero paired with a control that must hit.Why it queues rather than entering the decision box
⛔ No product question survives the card's own measurement. Unlike its sibling #1819 — filed from the same sweep and sent to the decision box — the direction here is already ruled twice over:
AGENTS.md§Schema Validation states it outright: "There is noworkflowmetadata type (ADR-0019/0020) … Field updates belong in*.hook.ts; status flips and notifications in arecord_change/scheduleflow; approvals in anapprovalnode inside a flow."WorkflowRuleSchema.parse()校验「Workflows」—— 该导出在平台上不存在 #852 retired exactly this shape fromAGENTS.mditself. These are two sites of that ruled class the card did not reach — the same.github/instructions/metadata.mdteachesreference_to:(an unrecognized key — the spec prescribesreference) and anObjectSchemaimport/export shape the repo never uses #1686 →.github/tasks/autonomous_feature_dev.md:20still tells the agent to checkreference_to— the fourth teaching site for a keyFieldSchemarejects, and the one #1686 could not reach #1812 pattern this round already ran once.⇒ restoring the invariant "briefs do not teach a metadata type the platform does not define" is on the 具名不升级类 list. Seat discretion.
Governing text:AGENTS.md§Schema Validation + §Project Architecture, and #852's ruling.⭐ The harm is the same one the round just spent four cards closing: an agent following either brief authors a file registered nowhere and therefore validated by nothing — "silently ignored". A success receipt for work the runtime never took. The
automation.mdexample compounds it by also dropping the mandatorycrm_prefix (object: 'lead').⛔ Two hard constraints
reference_tokey #1817. This card edits.github/tasks/autonomous_feature_dev.md:17— one line above the line PR docs(tasks): autonomous-dev brief names the lookup factory, not the retiredreference_tokey #1817 just changed, in the same file. docs(tasks): autonomous-dev brief names the lookup factory, not the retiredreference_tokey #1817 is ready with auto-merge armed as of 2026-09-09T12:4xZ. Branch frommainafter it merges; ⛔ do not stack on its branch, and ⛔ do not open a second PR against that file while it is in flight..github/instructions/automation.mdis squarely inside the governed set atAGENTS.md§How a green PR lands (.github/instructions/**). ⇒ draft PR, ⛔ never marked ready by the seat, ⛔ never auto-merged, the maintainer merges, with a## 维护者速读（草稿）in the PR body. ⭐ That is settled and needs no ruling — unlike the.github/tasks/**question open on.github/tasks/autonomous_feature_dev.md:20still tells the agent to checkreference_to— the fourth teaching site for a keyFieldSchemarejects, and the one #1686 could not reach #1812, this card touchesinstructions/, which is governed beyond argument.For whoever takes it
AGENTS.mdat your own commit and mirror the wording that is actually there. ⛔ Do not coin a fresh phrasing for a rule that already has one — that is precisely how this class recurs, and #1812 was dispatched with the same instruction for the same reason. Route the reader to the three real destinations the ruling names.WorkflowRuleSchemaundernode_modules/@objectstack/and*.workflow.tsundersrc/should both still be 0, with the card's own controls (FlowSchemapresent, 22*.flow.ts) proving the probes live. If either moved, stop and report.Size S, two files. ⛔ Nothing in
src/, ⛔ nothing intest/**, ⛔ no gate added.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): the two briefs still teaching a
*.workflow.tsmetadata type are corrected as #852 andAGENTS.md§Schema Validation already rule;.github/tasks/**is treated as governed conservatively — the PR is governed either way.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:
.github/instructions/automation.md§1 and.github/tasks/autonomous_feature_dev.md:17for 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 —
automation.md§1 is now the prohibition with the three real destinations and thecrm_prefix;autonomous_feature_dev.md:17names.flow.ts; zero*.workflow.tsteaching sites remain. 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;.github/instructions/automation.md§1 and.github/tasks/autonomous_feature_dev.mdwere corrected in the same PR).Generated by Claude Code