skills: triage idle/mutex probe loads two tools by name — stop paying the 55-schema ToolSearch ticket on self-retire rounds

[os-warren]

出处:分诊席 2026-08-20 自测(维护者转交问「分诊的 token 消耗合理吗」):自退轮总耗 ~117k,其中 ~81k(70%)是 ToolSearch 一次性注入 55 个 GitHub MCP 工具 schema 的固定门票,而互斥/自退判断实际只用 list_issues + issue_read 两个工具(真探针 ~14k,符合「空转座位成本上界 = 一次查询」的设计意图)。按 hourly fire、多数为空转估算:每天 ~200 万 token 纯耗在空转轮的 schema 注入上。优化方向由分诊席自己提出,skills 座定级落实。

Scope —— .claude/skills/pm-dispatch/SKILL.md 分诊座位职责(治理面,人合)

加一句工具加载纪律(措辞由 dev 依文件语域定):fire 开局的互斥检查/自退判断只按名加载所需工具(ToolSearch 用 select:mcp__github__list_issues,mcp__github__issue_read 形式),判定「本轮有活」之后才加载其余;⛔ 开局不做泛关键词 ToolSearch(那会注入全家桶 schema)。适用面写清:这是分诊 fresh session 的开局纪律;执行座位(常驻会话)与 dev(需全工具面)不受此约束。

预期效果写进条文作可验判据:空转轮 ~12 万 → ~4 万以内(分诊下轮自测读数回报)。

Serial

SKILL.md 空闲(PR #10089 已合)。治理面终态:draft + review request os-zhuang。模型档:clause ① fable-mandatory。

Graded pm:queue, S(skills 座,own-lane,session ae0d8bd1-5667-5f58-8ff8-fa22ec154325)。

[os-warren]

Claim: session ae0d8bd1-5667-5f58-8ff8-fa22ec154325 · branch claude/issue-10104-triage-tool-ticket

折叠检查:SKILL.md 空闲(PR #10089 已合并,#10081 已收账),队列无同文件第二卡 → 单卡。治理面终态:draft + review request os-zhuang。模型档:clause ① fable-mandatory。

---

Generated by Claude Code

[os-warren]

{
 "issue": 10104,
 "status": "done",
 "branch": "claude/issue-10104-triage-tool-ticket",
 "pr": "https://github.com/objectstack-ai/objectstack/pull/10105",
 "premise_still_valid": true,
 "summary": "Added one tool-loading discipline clause to the 分诊座位职责 section of .claude/skills/pm-dispatch/SKILL.md: the triage fresh session's opening mutex-check/self-retire probe loads only list_issues + issue_read by exact name (select: form of ToolSearch); the full toolset loads only after the round is judged live; broad keyword ToolSearch at session start forbidden. Applicability boundary stated (triage fresh sessions only; standing execution seats and devs exempt) and the measured acceptance criterion written into the clause (idle round ~120k -> under ~40k, read back by the triage seat's next self-measurement). Line budget re-measured on fresh main: 681/682 ceiling = 1 line headroom; clause is one physical line (net +1, matching the file's long-line convention), spending the headroom exactly - no dedup payment needed, file now 682/682.",
 "line_counts": { "before": 681, "after": 682, "ceiling": 682, "payment_audit": "none needed - net +1 paid from the 1-line headroom" },
 "added_sentence": "**工具加载纪律(fire 开局)**:互斥检查/自退判断只按名加载所需工具 —— `ToolSearch` 用 `select:mcp__github__list_issues,mcp__github__issue_read` 形式,判定「本轮有活」之后才加载其余;⛔ 开局不做泛关键词 ToolSearch(一次注入全家桶 schema —— 分诊席 2026-08-20 自测:空转轮 ~12 万 token,~8 万是这张门票);可验判据:空转轮 ~4 万以内,分诊下轮自测读数回报。只约束分诊 fresh session 的开局;执行座位(常驻会话)与 dev(需全工具面)不受此约束。",
 "tests": "All at final head 1d9108789, exit codes captured before any pipe. dispatch-gates.mjs (no args): 'change set derived from git — 1 path(s) vs merge base 7552e0337', 7 families matched, exit 0. Verdict lines: ratchet '✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/SKILL.md is 682 lines (ceiling 682; headroom 0).' (re-run on final head after reverse verification); id-lint '✓ check-skill-id-lint: 15 file(s) clean (pattern /#[0-9]{3,}/g).'; governed-prose '✓ check-governed-prose: 2 instruction surface(s) name all 5 registered governed surfaces … and claim no others.'; governed-merges '✓ check-governed-merges --self-test: 81 assertions'; frame-sync '✓ check-skill-frame-sync: 4 copies of the decision frame are structurally isomorphic across 3 files'; doc-authoring '✓ doc authoring guard: 379 files clean — no bare metadata literals.'; doc-formula-expressions '✓ check:doc-formula-expressions: 22 record-scoped formula example(s) across 404 files / 1414 TS blocks judged clean by @objectstack/formula.' (after building lint dependency closure). All exit 0.",
 "reverse_verification": "Directions predicted before running. (1) Ratchet ablation: predicted over-ceiling red naming the file; observed '✗ check-skill-line-ratchet: .claude/skills/pm-dispatch/SKILL.md is 683 lines; the ratchet ceiling is 682.' exit 1; restored byte-identical (git diff --exit-code = 0 vs committed fix). (2) id-lint positive control: predicted injected #-number turns red naming the file; observed '✗ check-skill-id-lint: .claude/skills/pm-dispatch/SKILL.md: 1 issue-ID citation(s)' at line 249, exit 1; restored byte-identical (git status --porcelain empty), re-run green exit 0.",
 "labels_read_back": ["documentation", "size/xs", "skip-changeset"],
 "pr_state_read_back": { "draft": true, "state": "open", "requested_reviewers": ["os-zhuang"], "head_sha": "1d9108789bd345733b10be39c35a0e11c59958d9" },
 "open_questions": [],
 "out_of_scope_findings": []
}

---

Generated by Claude Code

---

Generated by Claude Code

[os-warren]

ACCEPT — PR #10105 复核通过(治理面:保持 draft、review request os-zhuang 已挂、等维护者合并)。

复核记录(session ae0d8bd1-5667-5f58-8ff8-fa22ec154325,契约复审档 fable):

• 亲取 diff:一行入分诊职责节首段后,四要件齐 —— select: 精确形式(两工具名逐字)、判活后再加载其余、⛔ 开局泛关键词、适用边界(仅分诊 fresh session;常驻座位与 dev 豁免)✓;实测出处与可验判据(空转 ~4 万以内,分诊下轮自测回报)写进条文 —— 条文自带验收,不靠记忆 ✓。
• 行数:681→682 恰花 headroom,无付账需求,净 +1 合规 ✓。反向验证双腿(棘轮 683 红/恢复、id-lint 注入阳性对照证明门禁在读本文件)全中 ✓。7 族判定行逐字 ✓。
• 落地后的闭环:分诊下轮自测读数回报 = 本卡效果实测;若未达 ~4 万,读数进卡再迭代。

---

Generated by Claude Code