Filed and claimed by the repo:objectui seat (session session_01RnQd8iMMUwXQEV1crFmQiQ, 2026-08-15) under the maintainer direct-command channel — the maintainer challenged this rule in the live PM session. Verbatim provenance (untranslated, 2026-08-14/15, this session):
「继续,加快进度,最多并发5个子任务」
「哪里来的波次纪律?为什么要等我执行压缩,你乱说的吧?」
「你说你承认"在本环境里执行得过于机械",以后还会遇到这种问题等我而中断?」
The defect
.claude/skills/pm-dispatch/SKILL.md, section 轮次报告与节奏, sub-section 波次收工点(维护者 2026-08-09 裁定) prescribes: ① 收工点 = 在飞归零 → ② 存档 flush → ③ /compact 由维护者执行 → ④ 压缩后再派下一波,⛔ 不抢派.
Steps ③④ were written for terminal PM sessions, where /compact is human-only and a mid-review auto-compaction can lose in-context judgments. Applied literally in a remote session (auto-summarized, often unattended), they convert a safety checkpoint into an idle stop waiting for a human who is not watching — measured in this seat's session: the seat halted at in-flight-zero on 2026-08-14 until the maintainer asked why it had stopped, twice.
The amendment (direction, for the drafting dev)
- Keep ① and ② unchanged — the in-flight-zero checkpoint and the flush-judgments-to-GitHub obligation are correct everywhere (they are what makes ANY compaction, manual or automatic, lossless).
- Replace ③④: after the flush, the seat proceeds to the next wave immediately. A maintainer-executed
/compact at a flush point remains a terminal-session optimization the maintainer may perform, never a dispatch precondition. ⛔ Waiting for it is removed. - Preserve the existing provenance format: the 2026-08-09 ruling stays cited as superseded-in-part by the 2026-08-14/15 ruling quoted above.
Constraints (binding, from the skill's own guardrails)
- The skill's line-count ratchet (
check:pm-skill-ratchet) only goes down — the rewrite must be net non-increasing. - No issue numbers in operational text (
check:pm-skill-id-lint). - Any card touching
.claude/skills/pm-dispatch/** runs on claude-fable-5 (mandatory tiering clause). - Skills changes are ADR-class: the PR is drafted and pushed by the dev, reviewed by this seat on the card, and left hanging for the maintainer's manual merge — the AI never merges/queues it. That merge is the maintainer's veto window on this very amendment.
Filed and claimed by the
repo:objectuiseat (sessionsession_01RnQd8iMMUwXQEV1crFmQiQ, 2026-08-15) under the maintainer direct-command channel — the maintainer challenged this rule in the live PM session. Verbatim provenance (untranslated, 2026-08-14/15, this session):The defect
.claude/skills/pm-dispatch/SKILL.md, section 轮次报告与节奏, sub-section 波次收工点(维护者 2026-08-09 裁定) prescribes: ① 收工点 = 在飞归零 → ② 存档 flush → ③/compact由维护者执行 → ④ 压缩后再派下一波,⛔ 不抢派.Steps ③④ were written for terminal PM sessions, where
/compactis human-only and a mid-review auto-compaction can lose in-context judgments. Applied literally in a remote session (auto-summarized, often unattended), they convert a safety checkpoint into an idle stop waiting for a human who is not watching — measured in this seat's session: the seat halted at in-flight-zero on 2026-08-14 until the maintainer asked why it had stopped, twice.The amendment (direction, for the drafting dev)
/compactat a flush point remains a terminal-session optimization the maintainer may perform, never a dispatch precondition. ⛔ Waiting for it is removed.Constraints (binding, from the skill's own guardrails)
check:pm-skill-ratchet) only goes down — the rewrite must be net non-increasing.check:pm-skill-id-lint)..claude/skills/pm-dispatch/**runs onclaude-fable-5(mandatory tiering clause).