pm-dispatch SKILL: 5-hour usage-limit outages — queryable signals, cross-wall recovery timer (hourly cron, not send_later chains), wall-death playbook anchor

[os-zhuang]

Filed by the skills seat (#7623, session session_01XoWxhuq3WNNaZUTs6kWdk8) on the maintainer's prompt (skills-seat 专题讨论, 2026-08-11 ~13:4xZ, verbatim, untranslated):

> 项目经理技能还有一个缺口,有时候 Claude code 会到达 5-hour limit,首先项目经理能不能查到相关的数据,其次是到达时间窗口工作就会停,是否应该设置一个1小时的定时以监测时间窗口已经解锁。

The gap is real and already measured at fleet scale: two whole-account token outages on 2026-08-08/09 (~11:0x–12:10Z and ~16:0x–17:1x) each killed four in-flight devs at once (recorded in SKILL.md step 6, 直接验收兜底 section); #6686 is the earlier four-devs-one-wall specimen. What the skill lacks is a single place linking detection readings → cross-wall timer → recovery playbook.

Measured facts (2026-08-11, this container — re-measure at work time)

1. npx ccusage blocks works inside a CCR container (reads local ~/.claude/projects/**.jsonl): it reported the active 5-hour block (start 12:00, 3h21m remaining), per-container token burn (4.1M) and a burn-rate projection. Two readings are genuinely useful: window boundary/remaining time and burn rate (pre-warning).
2. Three blind spots of that reading: single-container only (cloud dev cards burn the same account's allowance in their own containers, invisible here); it estimates cost, not plan allowance — no surface exposes account-level remaining quota to the agent; block start is inferred locally (approximate).
3. The authoritative wall signal is the failure itself: API calls fail at the wall and the host surfaces "limit reached, resets at HH:MM" — i.e. the reset time IS usually available at hit time, just not queryable beforehand.

Deltas to .claude/skills/pm-dispatch/SKILL.md

1. 断粮检测读数 paragraph (likely near step 6's 舰队级死因 text): what a PM can query (ccusage per-container block timing + burn rate, explicitly labelled an estimate with the three blind spots) and cannot (account-level remaining quota); the wall signal and its embedded reset time.
2. Cross-wall recovery timer discipline:
   • On wall-hit — or pre-armed when the burn projection crosses the plan's risk threshold — arm an hourly cron Routine (create_trigger, cron), ⛔ not a send_later chain: a one-shot that fires into the dead window burns its only shot and the chain is severed — exactly the window it exists to cover. The cron re-fires hourly; the first post-unlock firing succeeds, runs recovery, then deletes the cron (self-cleanup is part of the discipline — a survivor cron is an orphan timer).
   • Prefer a one-shot aimed at reset-time + buffer when the reset time was captured from the wall signal (one precise shot beats hourly polling — same instinct as notes 1 读间隔 and notes 3 整点对齐); the hourly cron is the fallback when no reset time is in hand.
   • Fired text follows notes-3 discipline verbatim: opens with 「幂等 —— 动手前先重读状态」, criteria only, no imperatives.
3. Anchor the recovery playbook to this signature: first post-unlock action sequence = probe every in-flight cloud card (they die at the wall, zero information loss iff the draft-PR-early contract held) → direct-acceptance path or transcript-revival per existing step-6 rules → re-arm normal patrols. This is linking existing text, not new rules.

PM 机制假设(须实测,鼓励证伪): whether a send_later that fires during an account-wide outage is lost or retried by the platform is unmeasured — the dev should verify what can be verified cheaply (e.g. doc reading; a live outage test is not required) and word the SKILL text to match the evidence level (if unverifiable, state it as the conservative design rationale, not as a measured fact).

Process constraints (standing rulings — restated, not new)

• Model: model: claude-fable-5 mandatory (card touches .claude/skills/pm-dispatch/**; maintainer 2026-08-10, [batch] PM-skill optimization: maintainer-accepted 9-point plan, consolidated with #5925 item 7 + #6644 L2 + #7055 + #7156 #7341).
• Landing: skills updates are ADR-class (pm-dispatch SKILL: skills updates are ADR-class — human review/merge required (maintainer ruling 2026-08-11) #7548) — draft PR, ⛔ no auto-merge, no queue entry; the maintainer merges by hand.
• Same-file serial: third in the SKILL.md hot-file queue, strictly behind pm-dispatch SKILL structural compression: 3050 → ~1400–1600 lines (PR₂ of batch #7631, maintainer-directed) #7754 (whole-file compression, in flight) → pm-dispatch SKILL: cloud-card comms is event-driven by ruling — roster limitation on record, notes-21 verify-before-toggle amendment, PR subscription as a hard step #7755 (event-driven-comms deltas). ⚠️ This card's section references describe the pre-compression file; locate targets semantically on the post-compression text.

Refs: #7754, #7755 (queue predecessors), #6686 (额度墙 specimen), seat post #7623.

[os-zhuang]

Batch membership: absorbed into the #7885 consolidated dispatch (PR₁ carries Fixes #7765; this card's deltas are authored directly as principles in the rewritten text — detection readings, cross-wall timer discipline, wall-death recovery anchor, and the send_later-during-outage 机制假设 verified at whatever evidence level is cheaply reachable). No longer an independent dispatch candidate; the label stays until the batch PR closes it. Composition and rationale recorded on #7885.

---

Generated by Claude Code