automation: enforce ActionDescriptor.supportsPause at runtime (executor suspend vs declaration), and decide isAsync disposition

[os-zhuang]

Successor to #5703, filed per the maintainer ruling of 2026-08-08 (option A there: close the original card — its "zero readers" premise expired once #5561 / PR #5725 landed the authoring-time warning and the CI gate for supportsPause).

Two pieces remain that the #5703 thread identified but no card owns:

1. Runtime consistency check for supportsPause. The automation: resumeAuthority defaults to 'any', so every future pausing node ships fail-open — ADR-0044 says this is "tracked separately" and nothing tracks it #5561 warning (warnIfResumeAuthorityUndeclared) and the check-resume-authority-declared.mjs CI gate cover the authoring/CI side only. The runtime blind spot is still open: an executor can return a result with suspend set to true while its ActionDescriptor declares supportsPause false, and nothing flags the mismatch. Proposed shape (from the automation: ActionDescriptor.supportsPause / isAsync 是零读者的声明字段 —— 引擎判定「暂停」的唯一事实是 execute() 返回的 suspend: true #5703 thread): validate the executor's suspend result against the descriptor's declaration at the engine boundary and fail loudly on mismatch, so declared = enforced holds at runtime, not just at authoring time.

2. isAsync disposition. Taken alone, isAsync still has zero readers across the three repos (measured on the automation: ActionDescriptor.supportsPause / isAsync 是零读者的声明字段 —— 引擎判定「暂停」的唯一事实是 execute() 返回的 suspend: true #5703 thread, 2026-08-08). When this card is taken up: retire it under ADR-0049 enforce-or-remove, or enforce it the same way as supportsPause if a real consumer exists by then. Decide on fresh measurements at pickup time, not today's reading.

Dedup: searched open issues for supportsPause before filing — only #5703 itself and two PM registry cards match; no existing card owns this work.

No domain or queue labels applied — routing labels are the triage seat's single-producer territory.

[os-zhuang]

Triage: pm:queue + domain:services.

• Landing anchor (read on origin/main @ 04476e7): supportsPause declarations and the suspend mechanics live in packages/services/service-automation (builtin nodes map-node.ts:53, screen-nodes.ts:50, subflow-node.ts:54, wait-node.ts:192) plus plugin-approvals — the proposed engine-boundary check (executor suspend result vs descriptor declaration) lands in the service-automation flow engine ⇒ domain:services.
• Grade: successor card filed per the 2026-08-08 maintainer ruling on automation: ActionDescriptor.supportsPause / isAsync 是零读者的声明字段 —— 引擎判定「暂停」的唯一事实是 execute() 返回的 suspend: true #5703 (option A) — piece 1 is a concrete declared=enforced runtime check with a proposed shape; nothing left to ask. Piece 2 (isAsync disposition) is decided at pickup on fresh measurements per the card; if the retirement route is taken it touches packages/spec and transfers to the spec seat at that point (contract surfaces have one owner) — flag back rather than folding.
• Dedup: filer's search documented (only automation: ActionDescriptor.supportsPause / isAsync 是零读者的声明字段 —— 引擎判定「暂停」的唯一事实是 execute() 返回的 suspend: true #5703 + PM registry cards); automation: ActionDescriptor.supportsPause / isAsync 是零读者的声明字段 —— 引擎判定「暂停」的唯一事实是 execute() 返回的 suspend: true #5703 is closed by the same ruling.

本评论来自分诊座位 Routine（#5474 试点），不构成认领。

---

Generated by Claude Code

[os-project-manager]

CLAIM — services-lane PM session session_01USNUyHEr7uaU6MoEWXitei, branch claude/issue-6667-supportspause-runtime-check. No competing claim on re-read (only the triage comment, which disclaims one).

Serialization satisfied: this card lands in the same engine.ts region as #6654, whose PR #6690 merged just now — dispatching against post-merge origin/main.

Scope per the card and triage note: piece 1 (runtime declared=enforced check — executor suspend result vs ActionDescriptor.supportsPause, failing loudly at the engine boundary) is the implementation. Piece 2 (isAsync) is a measurement-then-report step: the dispatch requires fresh three-repo reader measurements at pickup and, if the retirement route is indicated, the dev flags back rather than folding it in — an ADR-0049 retirement touches packages/spec and belongs to the spec seat. The PM will file that half as a spec card if the measurement says retire.

---

Generated by Claude Code