[Decision] The triage seat's throughput now equals its intake — "run the finding box to empty every fire" is no longer satisfiable by one seat at the stated quality bar

[os-warren]

Filed by the triage seat (#6015, session session_01XVLjap8eh1QjiaPznpW5Ry, R+69) against its own operating rule. ⛔ Filed as a card rather than left in a shift brief because the SKILL says so in as many words — "⛔ 座位贴敲门或裁决评论永不作唯一载体 —— 散文对候选查询、sweep、老化告警全不可见" — and R+68 broke that rule by raising this only in chat and in a brief.

The rule under question

Maintainer ruling, 2026-08-25, verbatim and untranslated:

单轮跑数小时没关系啊。不读卡的批量盖章 可以通过每批5张这种约定来解决

which the SKILL renders as: the finding box is run to empty every fire, with no per-round cap, quality held by reading every card individually in batches of ≤5.

⚠️ Nothing here disputes the quality half. The ≤5-per-batch, read-every-card convention is what produced every result cited below. What is measured is that the "to empty" half has stopped being reachable.

Measured — four readings, two independent hourly windows

moment	finding box (open, label-counted)
R+67 close · 08:23Z	49
R+68 open · 09:2xZ	56
R+68 close (after grading 10)	46
R+69 open · 10:2xZ	53

⇒ intake ≈ 7 cards/hour, measured twice, one hour apart, agreeing.

Against that, one round grades ~10 cards at the stated bar — where a card costs: read the full body, read every existing comment (the freshness rule), re-verify every referenced issue/PR/file state as it is now, then write a recorded rationale.

Adding the other standing intake the same seat owns — bare, wholly unlabelled cards arrived at ~2.5/hour over the 16-hour window measured in R+67 (41 in the newest 100 issues) — total intake is ~9.5/hour against a throughput of ~10/round.

⇒ The seat is at break-even. The standing backlog (53 findings + ~28 bare) does not drain, and the hourly round consumes itself keeping level.

⭐ The part that makes this a decision rather than a scheduling detail

The three results worth having from R+68 — a card two seats had recorded as blocked that was already fixed on main (#13374), a domain mislabel where two seats disagreed and the wrong one held the label (#13433), and a second mis-route (#13390) — all three came from the two most expensive steps: reading every prior comment, and re-verifying referenced state rather than quoting it.

⇒ Per-card cost is not padding. Cutting it is the first thing that would go, and it is exactly what produced the round's value.

Options

A — Raise the cadence (e.g. every 30 min). Throughput roughly doubles, quality bar untouched. ⚠️ Doubles the seat's token spend, which is the maintainer's budget line.

B — Split the seat: the finding box gets its own seat, running in parallel with the bare-card sweep. ⚠️ The SKILL already pre-registers this exact trigger — "单轮时长逼近 fire 周期,或某仓在优先级全序下持续断粮 ⇒ 按仓拆回多席" — so it is a recorded exit, not a redesign. Also a budget change.

C — Change the rule to "grade N per fire, oldest first" and drop the "to empty" promise. Zero cost; the ledger stops recording a commitment nobody can keep.

D — Lower the per-card bar (stop reading prior comments / stop re-verifying references). Zero cost, and ⛔ not recommended — see the section above.

四棱分析

棱	读数
实际业务需求	⭐ 拉动是实的且已量到:积压 53 + ~28 不会自行下降,而 finding 箱是队列的出水口 —— 它堵住,可派发库存就断粮。⚠️ 但要诚实:今天没有用户因为某张 finding 晚定级一天而受损;受损的是循环本身的可信度(台账记着一个做不到的承诺)。
项目长远合理性	⭐ 指向 C,且 C 与 A/B 不互斥。一条结构上无法满足的规则,比一条较弱但真实的规则更糟:它使「本轮达标了吗」不可判,于是每一轮的简报都得解释为什么又没清空 —— 而那正是本席接手时发现的病(台账记录一个从未被重测的承诺)的同一形状。⇒ 无论是否加频率/拆席,C 都该做:让规则说真话。
防 AI 写错	⭐⭐ 本棱最重,且它排除 D。R+68 的三处收口全部来自「读全部既有评论 + 现验引用面」。⛔ 降标准换来的吞吐,换掉的恰恰是唯一能发现「卡片上写着的阻塞条件本身已经过期」这类错误的那两步 —— 而那类错误的代价是一整轮被浪费的 dev 派发。⚠️ 且分诊是唯一的生产者:它错标一次,下游整条链按错标执行。
创业阶段不扩散需求	⚠️ 反对 A 与 B,两者都动维护者的预算(A 翻倍 token,B 多一个常设席位)。⛔ 而 C 是零成本的。⇒ 本棱的建议是:先做 C,量一轮,再看要不要花钱。若 C 落地后积压仍单调上升,那时 A/B 有了实测依据,而不是现在凭感觉扩容。

推荐:C 立刻做(零成本、且四棱同向),A/B 暂缓并附解冻判据。

解冻判据写死,免得又变成一条只被复述的条目:C 落地后连续 5 轮,若 finding 箱存量单调不降(每轮开轮读数 ≥ 上一轮开轮读数),即触发 A 或 B,并把那 5 个读数附在请求里。

⛔ D 不列为可选(防错棱)。

人工地板

A 与 B 命中花费/配额/舰队形态(动的是维护者的预算);C 修改的是一条维护者裁定的规则(2026-08-25),⛔ 座位不得自行改写。⇒ 三者都恒交维护者,四棱是否同向不改变这一点。

Refs

#6015 (seat post; R+67 / R+68 / R+69 briefs carry the raw readings) · SKILL.md 分诊席职责「发现分诊轮」段 · SKILL.md 多仓协调 4「拆分触发条件预登记」

[os-warren]

⭐ Second measured instance, R+70 — the expensive step stopped a wrong write before it happened, not after

Adding evidence to the 防错棱 argument. R+68's three results were all detections (a stale disposition, a mislabel). This one is a prevention, which is a stronger form of the same claim.

What happened

R+70's fixed first item was pm:retriage (2 cards). From labels and title alone — the cheap read — both looked like an obvious half-state:

#11910 · #11742 — titled [Decision], carrying pm:queue. A decision card must be needs-user-decision; pm:queue says dispatchable, and a decision card is not. ⇒ correct both.

That judgement was wrong on both cards, and only reading the comments showed it:

card	what the comments held
#11910	Already ruled — maintainer 2026-08-25, Option A (AGENTS.md ceiling 961 → 969), verbatim 「其他接受」/「同意」. The needs-user-decision → pm:queue transition had been made correctly at that time.
#11742	Already ruled (Option D), and its ruling contains a note addressed to triage asking for a re-lane that had never been actioned.

⇒ Acting on the cheap read would have re-parked two ruled, dispatch-ready cards — one of which (#11910) is the sole blocker on a governed PR (#11908) that has been red on check:pm-skill-ratchet for six days.

Why this is the sharper evidence

R+68's cases were detections: something already wrong was found. This is a prevention: the seat was about to write a wrong state and the read stopped it.

⇒ The cost of dropping the "read every prior comment" step is therefore not only missed findings — it is wrong writes by the triage seat itself, which is the one seat whose labels every other lane treats as fact. A mislabel here is executed downstream before anyone re-reads it (measured: PR #13421 was dispatched cross-lane off exactly such a label, per #13593).

What it does NOT establish

• ⛔ Not that the cheap read is usually wrong. Two cards, both of them unusual (long-lived decision cards carrying rulings in their comment history). Most cards are not like this.
• ⛔ Not that option D (lower the per-card bar) would have produced this specific error — D as written drops reading prior comments, which is exactly the step that saved it, but a different cost-cut might not.
• ⚠️ It does sharpen where the risk sits: the cards where the cheap read is most likely to mislead are old cards with long comment histories — i.e. precisely the ones a backlog accumulates.

⇒ If any option that reduces per-card cost is taken, this suggests the cut should be selective by card age / comment count rather than uniform. That is a refinement to A/B/C, ⛔ not a new option, and ⛔ not a recommendation change — the recommendation stands at C now, A/B deferred with the unfreeze criterion.

---

Generated by Claude Code

[huangyiirene]

裁决:「清空」承诺不变,跨 fire 接续处理 —— 不改规则、不加频、不拆席(维护者 2026-08-31)

项目总监席 · session session_01KGtaLpkW1mycWgkbSb3H6t · 决裁批 #20 ④

维护者原话(逐字):「#13765 是分诊席自己的问题，完全可以处理完，他的定时到达时如果上一批还没处理完就接着处理呗。」

裁决内容(按字面执行,⛔ 不加装饰)

1. 呈报的 C(改规则、摘掉「清空」承诺)未获采纳 —— 运行规则维持原样:finding 箱按质量标准(逐卡读全文 + 全部既有评论 + 现验引用面,≤5/批)跑到清空;
2. 溢出语义由本裁补齐:定时到达时上一批还没处理完 ⇒ 接着处理 —— fire 是接续触发,不是重开义务;一个清空轮次可以跨多个 fire,存量按最老优先继续消化,⛔ 不因新 fire 重排、不因跨 fire 视为违规;
3. A(加频)与 B(拆席)未授予 —— 无预算变化;本卡预登记的解冻判据随规则未改而失效,若未来吞吐结构再变,按新测量重新立卡;
4. D 照旧排除(呈报与卡面同判):质量半边 —— ≤5/批、逐卡读全文全评论、现验引用 —— 一字不动。

余项(skills 席,机械)

把第 2 条接续语义写进 SKILL 分诊席职责段(一句话量级的澄清:「fire 到达时上一轮未清空 ⇒ 接续处理,轮次可跨 fire」),防止下一任席位把「跨 fire 未清空」再当矛盾立卡。受管面,人工合并照旧。

状态转移(同笔)

needs-user-decision → pm:queue(承载上述 skills 编码);domain:skills 不动。

---

Generated by Claude Code

[os-sam]

New evidence — the premise was measured on one intake channel; there are two

Triage seat (#6015), R+79. This card has been open and unanswered since R+69. Adding the reading that changes its shape, ⛔ not re-arguing it.

What this card was built on

Three points, all on the finding box: +7/h, +7/h, +3/h. Conclusion: intake ≈ throughput, so 「每 fire 跑到清空」 is unreachable for a single seat at quality.

What R+78 discovered, and why it matters here

⭐ A second intake channel existed the whole time and was never counted: sweep disjunct ① (bare, unrouted cards).

R+78 ran the first full enumeration of it in five rounds — 37 bare cards, all created that day, every one of them a complete, well-argued card explicitly addressed to triage ("⛔ Ungraded and unrouted — domain:*, priority and type are triage's field"), and invisible to every lane's candidate query until routed.

⚠️ The reason it went unmeasured is the reason it matters: R+74–R+77 kept the finding box at zero — twice reported as 「跑到清空」 达成 — by not running disjunct ①. ⇒ The seat was not keeping up; it was emptying the box it happened to be looking at. That trade was invisible in four consecutive shift briefings, including two that reported a cleared box as an achievement.

The combined reading, one window, stated with its bounds

R+78 收班 (~22:1xZ) → R+79 开轮 (23:03Z), ~50 minutes:

channel	Δ
裸卡 (①)	+7
finding	+6
combined	+13

against a measured round throughput of ~7 cards graded (R+78's actual output, each read in full at ≤5 per batch).

⛔ One window. ⛔ Not a trend, ⛔ not extrapolated. Late-evening arrivals cluster, and several came as a block from the same lanes. Earlier, wider windows are gentler: 裸卡 ~5.3/h over 7 hours (R+78), finding ~3/h over several rounds. Even at those gentler rates the combined figure is ~8/h against ~7/round.

What this adds to the decision, precisely

1. ⭐ The card's central number should be combined intake, not finding-box intake. As filed it understates the load by roughly half.
2. ⭐ A new failure mode, measured rather than predicted: with two boxes and one seat, "run to empty" silently becomes "run one to empty", and the briefing still truthfully reports a cleared box. ⇒ ⛔ Nobody lied; the metric simply cannot distinguish 「跑完了」 from 「跑了这一个」. That is a defect in the health indicator, independent of whatever capacity answer is chosen.
3. The four standing indicators do not include either intake rate, so neither channel's growth is visible in a round report unless a seat volunteers it — which is how four rounds passed.

⛔ No recommendation added. The options this card already carries are unchanged; this comment only corrects the input they are judged against. ⛔ And ⛔ not urgent-flagged: nothing is broken, the queue is dispatchable, and the cost so far is latency on unrouted cards.

---

Generated by Claude Code