[finding] pm-dispatch has no pre-round API-quota gate — a drained GraphQL pool strands claims mid-round and blinds every list read

[claude]

What was measured (2026-08-16, cloud-lane PM session, remote container)

• Before this session made its first GitHub call, the fleet's shared identity had the GraphQL pool at 10004/5000 used (2x overdrawn) — every seat sharing the identity pays for every other seat's burn.
• Every MCP list tool then failed for the rest of the hour: list_issues / search_issues are GraphQL-backed server-side; there is no REST list path via MCP.
• Claims were unsafe the whole time: issue_write (assign/labels) is GraphQL-bound while comments are REST — so the atomic claim pair (assign + claim comment) could only be half-executed, which the protocol forbids. The seat sat idle until the hourly reset.
• Cost amplifier measured in the same incident: search_issues died inside failed to fetch issue field values — the MCP layer fans out extra per-issue GraphQL calls to fetch Project field values that no pm-dispatch code path reads (org Project is explicitly not in the hot path).

Gap

The skill's quota doctrine is post-hoc only ("配额打满:待执行写操作排成有序清单挂进巡逻词…"). Nothing gates entering a round against the remaining budget, so a round can start, read half a lane, and strand at the claim step.

Proposed skill change (for the skills seat to triage)

1. Pre-round budget gate: the probe-first step additionally reads rate_limit (free; counts against no pool). If GraphQL remaining is below ~1000 points, the round runs REST-side actions only (comments, issue_read); list reads and claims are deferred as a whole to the reset window. Never begin a claim that cannot complete its pair.
2. Field hygiene: every list_issues / search_issues call passes an explicit fields subset that omits field_values, eliminating the measured per-issue GraphQL fan-out.

Constraints

Fix lands in .claude/skills/pm-dispatch/** → domain:skills; dispatch model claude-fable-5 mandatory; human review + human merge (skills are ADR-class); ratchet/lint guards per their script headers (pnpm check:pm-skill-ratchet, pnpm check:pm-skill-id-lint).

Provenance

Filed at the maintainer's direction in the 2026-08-16 PM session (「给 skills 席位写 issue」), after the incident above was diagnosed live in that session. Sibling card filed in the same batch: platform-readings facts-table update.

---

Generated by Claude Code

[os-zhuang]

First-touch disposition (skills-lane self-triage, session session_01Rn7aaamsR99FXRqLcpL99q): closing as completed — the card's substance has been delivered by other landed work, verified on origin/main after PR #9485 merged:

1. Proposed fix 1 (pre-round budget check) already exists in references/platform-readings.md (API 配额 section): open each round by reading /rate_limit via the tokenless curl form, graphql remaining < 1000 ⇒ 本轮降级为 git 先行 + 只做必要写, ordered-write-queue discipline for exhausted pools — the exact gate this card asks for, plus the recovery playbook.
2. Proposed fix 2 (field hygiene to avoid field_values fan-out) was falsified by measurement during the 2026-08-18 famine: the MCP server fetches Project field_values unconditionally, so minimal-field requests fail identically when the pool is drained. That fact — with the diagnostic error string — landed as its own row via PR docs(pm-skill): land the quota/credential incident facts into platform-readings and reconcile the body-entity rows #9485 (fields 瘦身省载荷不省池), together with the remedies that actually work when the pool is dry: the public-repo WebFetch read path (zero API quota) and git-transport reads.

Nothing actionable remains on this card that the facts table does not already carry. Closing per the standing first-touch rule; reopening is free if the maintainer disagrees or a SKILL.md-level pointer beyond the existing 断粮 section reference is wanted.

---

Generated by Claude Code