dsh-fast-mode 0.0.0-stage → 1.5.2

package/CHANGELOG.md

@@ -0,0 +1,627 @@
+# Changelog
+
+## 1.5.2
+
+- **新增实验旋钮 `workspaceMap`（默认关闭）与 `mapEntries`（默认 60）**：会话第一次工具调用之前，harness 把工作目录走一遍（深度 2、≤`mapEntries` 条、跳过 `node_modules`/`.git`/`dist`/`build`/`coverage`/`__pycache__`/`.venv`，正文 ≤1.4 kB），把「相对路径 + 字节数」和看得见的测试入口作为一条通知交给模型。它不结束回合、不改完成判定、不动 `reasoningEffort`、不加请求字段；它只是把模型自己会用 `ls`/`find` 换来的同一份清单先给一次。实现 `src/map.js`，接入 `index.js` 的 `agent/pre-step`，配置见 `src/config.js`、`src/effort.js`。
+- **为什么默认关闭（两轮实测，23 个有效单元）**：W10 与 W11 两块把 `def15` 与 `def15 + workspaceMap` 同块对照。合并 23 个单元：**墙钟 0.927×、输出 0.896×、步数 1.099×**，逐格墙钟中位 **0.791×**、13/23 格更快、隐藏测试零红。方向是对的（输出确实少 10 %），但**步数反而多 10 %**，且反向尾部很重——`stategraph` 与 `codeimpl` 的个别格 1.4×–1.77×，两臂 p90 从 369 s 变成 559 s、最慢 510 s 变 714 s。按「长任务不能变慢」的门槛，它不够格进默认，留作旋钮并公开数字。
+- **修掉一个会让新机制静默失效的缺陷**：运行目录在会话创建之后才放 fixture，第一步读工作目录可能读到空目录；旧写法在这种情况下也把「已下发」标掉，于是该会话再也不会拿到清单。现在只有真正构建出清单才记录已下发（`index.js` 的 `mapSent.add` 前判空），MAPDBG4 单格验证：`built=0`、`built=0`、`built=291`，会话里恰好一条 `Auto map:` 通知。
+- **投递审计工具化**：`bench/audit-delivery.py --tag W10 --needle "Auto map: one inventory"` 一条命令按「任务 × 对 × 臂」列出文案命中次数。W10 结果是 `map15` 10/10 命中、`def15` 0/10 —— 对照臂干净、处理臂每格恰好一次，所以那一块的差值不是「提醒没发」的噪声。协议见 `bench/README.md` 第 15–17 条。
+- 数字与逐格表见 `BENCHMARK.md` 的「筛选块 W10」与「复现块 W11」；原始结果 `bench/results/suite-W10.jsonl`、`bench/results/suite-W11.jsonl`，`python3 bench/report.py --files bench/results/suite-W11.jsonl --baseline def15` 可原样复算。
+
+## 1.5.1
+
+**干净重基线：10 家族 30 个单元，总墙钟 0.72×、逐格中位 0.64×，有效单元 29/29 全绿。默认配置不变。三块独立复现到 0.72× / 0.73× / 0.73×。**
+
+这一版没有改默认行为，改的是**测量**：把两个会污染数字的缺陷修掉，然后用当前默认（`enabled: true` 的出厂配置，即 1.5.0 的 `def15`）重新跑一遍。1.5.0 承诺的「不按步数收尾」没有变。
+
+- **干净重跑（60 次运行 / 30 个单元 / 10 个家族）**：`off` 3434.7 s / 591 244 输出 token / 390 步；`def15` 2472.5 s（**0.72×**）/ 439 708 token（**0.74×**）/ 276 步（**0.71×**）。逐格墙钟中位 **0.64×**、均值 0.84×、90 % 区间 [0.59, 0.79]、30 格里 24 格更快；输出 token 逐格中位 0.69×。分家族从 `migrate` 0.45×、`shellops` 0.57×、`codebase` 0.55× 到 `perfrefactor` 1.18×、`csvpipe` 1.12×——短任务本来就没有可省的重复验证，收益不稳定。
+- **质量门禁：有效单元 29/29 全绿**（30 个单元里 1 个 `debugfix` p1 的 `off` 因会话启动崩溃作废，不算红）：隐藏测试在 agent 看不到的临时副本里执行，两臂通过率都是 1.000，没有「基线绿而该臂红」。有限任务上的结论，不是质保证明。
+- **判据收紧为「有效单元」**：从 1.5.1 起，一个单元只在 `exit == 0 && 步数 ≥ 1 && 输出 > 0 && 判分完成 && 未泄漏` 时计数；崩溃行整格作废并点名。旧口径会把崩溃当成「0 步的红」或「输出 0 的极快运行」，两种都会污染比值和门禁。W3/W5/W7 的准确说法因此是 29/29、19/19、17/17，而不是 60/60。
+- **三块重复现 + 两个机制被证伪，一个机制撤回为「未测」**：W5 同块对照 `def15` 0.73×；W6 筛查里 `openBatch` 0.77×，W7 独立复测回到 0.91×、对默认头对头 1.136×（W7 会话里确认提醒真的下发过，淘汰成立）；W7 再复现 `def15` 0.73×（三块合并 65 个有效单元 0.74×）。同一块里的 `batchVerify` 臂记到 0.84×，但事后逐会话检索发现它 20 个会话**零下发**，那个数字测的不是机制，所以 `batchVerify` 从「已证伪」撤回到**未测**，默认仍关、修复后重测。三个未晋升的机制都留在 `bench/arms/` 作消融臂。
+- **W7 的 7 行崩溃已点名**：`stategraph` p2 三臂、`codebase` p2 的 `def15`+`open15`、`shellpipeline` p2 的 `off`+`def15`，全部 `exit=1` / 1 步 / 0 输出 / ~80 s，工具未执行、产物为空，与机制无关。
+- **对照块 W8b（60 次运行 / 20 个单元）复现逐格中位**：`def15` 对 `off` 逐格中位 **0.717×**、几何均值 0.819×、19 格里 13 格更快，输出逐格中位 0.823×、步数逐格中位 1.000×；这一块总量只有 0.934×（反向尾部更重），所以总量与逐格中位要一起看。`ledger` p2 的 `off` 行泄漏（读了隐藏测试）整格作废，**59 个有效单元零红**。
+- **新增可选 `preferPatch`（默认关闭）**：一次 ≥ 2 kB 的整文件写入之后，下一步发一次「改已存在的文件时只发改动的行」的提醒。文件结果逐字节相同，所以这条只谈形式，不要求重做、也不提前收尾；每轮最多一次。
+- **修掉一个提醒投递缺陷**：同一步里调度结果被多个钩子读取，只有其中一个会真正投递提醒，其余只读钩子也调用了同一个一次性提醒函数，把提醒标成「已用」却从未显示给模型。W8b 的 `pp16` 臂因此 20 个会话里**零下发**，那一臂的数字作废、不计入任何结论（`pp16` 对 `def15` 的中位差是噪声）。修复为只读调用传 `{emit: false}`，提醒只在真正投递时消耗；单格验证同格修复前 0 次下发、修复后 1 次下发，另有回归测试。
+- **修掉的测量缺陷**：① 运行目录移出仓库（`$TMPDIR/dsh-runs/`），产物里出现 `hidden/`、`ref/`、`bench/suite/`、`MANIFEST.json` 字样的行标 `leak` 丢弃；② 指标按 `session id` + 开跑前差集取行，不再按字节偏移；③ 判分在 `copytree` 出来的临时目录里做（`shellops` 会把产物目录 `chmod` 成只读，旧写法直接 `EACCES` 中断批次）；④ `bench/MANIFEST.json` 用整套 fixture 的 sha256 校验，改过 fixture 就拒绝开跑。
+- **查清 W1 那格「红」是题面歧义**：`ledger` 的 `test_round_half_up` 用 `offset: "0.005"` 要求 Decimal 精确参与四舍五入，题面只写「四舍五入到分」+`parse_amount` 抛 `ValueError`，两种读法都成立。参考实现 22 条断言全绿，说明不是快速模式降质。题面已补明规则里的系数/偏移按 `Decimal` 原样解析、只有最终结果取两位小数；fixture 摘要 `2699cd64…` → `2d7546f8…`。
+- **没变的约束**：不改 `reasoningEffort`、不加请求体字段（`service_tier`/`max_tokens`）、不按步数结束回合、不用计数器决定“做完了没有”。快的地方仍然是调度器省下来的重复验证与空转，不是压缩模型思考。
+- 数字与单元级表格见 `BENCHMARK.md` 的 `1.5.1`；原始结果 `bench/results/suite-W3.jsonl`，`python3 bench/report.py --files bench/results/suite-W3.jsonl --baseline off` 可原样复算。
+
+## 1.5.0
+
+**默认不再按步数收尾：什么时候结束由任务状态和完成证据决定，不由计数器决定。**
+
+1.4.4 把「第 10 步提醒收尾」设成默认，速度确实上去了，但它对**还在推进的长任务**同样开火：第 10 步一到就要求停止扩大范围，等于用一个和完成度无关的数字替模型决定“做完了没有”。1.5.0 取消这条默认。
+
+- **默认无收尾线、无步数上限**：`finishLineSteps` 与 `stepBudget` 默认都是 `0`（关闭）。只有用户自己把某个值设成正数时，才各发一次对应提醒；`0` 就是关，不再回落到旧默认值。
+- **无进展只是诊断，不是判决**：连续若干步没有产出新结果时，只发一次「换方法」提醒（后台跑+轮询、把脚本写到文件再执行、缩小到一条路径），**不结束回合**。任务做完就正常报告，不会被提醒逼停。
+- **检查复用绑定产物版本**：同一份产物上逐字重复一条已通过的检查，才提示「这个检查在当前文件上已经通过，结果可复用，把这一步花在剩下的工作上」。任何写操作、编辑、或改写类 shell 命令都会让该版本失效，旧结果立即作废；逐字重复的检查命令本身不会让自己失效（否则 `npm test` 永远无法触发这条提示）。
+- **新增 `outputBudget`（默认关闭）**：本轮累计输出 token（上游 `usage.outputTokens`，不是估算）达到该值，调度器发一次「收掉当前闭环、跑一次证明它的检查、汇报」，每轮最多一次。它盯的是真实解码量而不是步数——三步写出三万 token 的任务很贵，二十步只写两千 token 的任务不贵；但 token 同样是成本而非完成度证据，所以默认 `0`，只作对照臂。对照臂 `outputBudget: 8000` 在 `codeimpl`/`migrate`/`codebase` 各 3 对上跑过（9 个完整单元、27 次运行全绿）：总量 0.68× 墙钟 / 0.70× 输出，其中 `codeimpl` 输出再降 23 %（0.60× → 0.46×），而单步写入短的 `codebase` 反而 1.16× 输出——所以它是可配项，不是默认。
+- **新增 `shapeGate`（默认关闭）**：本轮累计工具调用 ≥ 4 且写入类（不在只读名单里的工具）占比 ≥ 60 % 时，**整轮不再注入任何提醒**，退回原生。它给写密集型任务一条「别管我」的退路——这类任务没有可并行的调用，每条提醒都只是一次用不上的往返。判据只用已有工具名，不发请求、不加请求体字段；默认关，因为它同时会压掉本可救长尾的收尾提醒，尚未单独跑对照臂。
+> **撤回（2026-10-26）**：1.5.0 这一节的实验数字（171 次运行、54 个单元，`fl10` 0.66× / `open15` 0.69×、以及 `def15` 的 0.86× 重基线）来自两个有缺陷的批次——运行目录在仓库内（会话可读到 `hidden/` 与 `ref/`）、指标按字节偏移切片（把同机器其他会话的步数算进来）。**这些数字不再作为速度声明**，只保留当时的决策记录；干净重基线见 1.5.1。
+>
+- **本轮主实验：9 家族 × 6 对三臂交错 ABBA（171 次运行、54 个完整单元）**。`off` 7841.3 s / 2 035 299 输出 token / 3960 步；`fl10` **0.66× 墙钟 / 0.68× 输出 / 0.67× 步数**（逐对 44/54 更快）；`open15` **0.69× / 0.71× / 0.71×**（40/54）。分家族上两臂互补：`open15` 在 `csvpipe` 0.46×、`perfrefactor` 0.62×、`stategraph` 0.63×、`shellops` 输出 0.37× 上更狠，`fl10` 在 `shellpipeline` 0.52×、`ledger` 0.62×、`codeimpl` 0.68×、`debugfix` 0.69× 上更好；两者都在某处留了输出回升（`open15` 在 `debugfix` 1.24× / `migrate` 1.07×，`fl10` 在 `stategraph` 1.16×）。`ledger` 的 22 条断言对模型偏紧：前 6 对里 p1（`fl10` 红）、p2 与 p6（`off` 红）整对作废，补跑 p7–p9 才凑满每臂 6 个完整单元。`fl10` 在 H2、`open15` 在 H3 各出现一次「基线绿而该臂红」，按规则都不做默认。
+- **两个更快的配置都没通过门禁，1.5.0 默认不变**。`fl10`：`ledger` p1 基线与默认都绿、它红 21/22（重跑 5 次全绿，1/6 失败率）；`open15`：补测 5 个家族、30 个单元后相对默认 `def15` 是 **1.03× 墙钟 / 0.99× 输出 / 0.95× 步数**，`shellops` p7 基线绿而它红，`csvpipe` p5 用掉 325 步 / 499 s / 5.2 倍输出 token。按「基线绿而该臂红即不合格」，两者都退回实验旋钮，`finishLineSteps: 0` 与 `openBatch: false` 保持默认——宁慢不降质。分家族与单元级数据见 `BENCHMARK.md` 的 `H2`/`H3`。
+- **进展按结果内容判定**：用结果内容的摘要（digest）而不是输出长度判断“这一步有没有新东西”。`completed 10 / 99` → `11 / 99` 这种等长轮询算进展，不会被误判成卡住；判定单位是整步，不是单次调用。
+- **证据优先级**：最近三次调用都报错时，扇出宽度直接回到 1，压过“请求措辞像扇出”的分类。
+- **看门狗取消后等真实收尾**：取消一个卡住的调用后，先等它真正停下（窗口随预算缩放，250 ms–5 s）再返回失败，并把 `quiesced` 记进日志。以前是“先告诉模型已取消，清理还在跑”，下一个步骤可能和上一个进程并发写同一个目录。
+- **`enabled: false` 是真的关**：关闭态下看门狗预算为 0，不再按档位表在两分钟后取消调用。
+- **测试与 CI**：223 个测试在 Node 20.20.2、22.23.4、24 上全绿。顺带修掉 Node 20/22 上 `node:test` 报 “Promise resolution is still pending but the event loop has already resolved” 并取消 17 个用例的问题（看门狗定时器不再 `unref`，它本来就会被 `clearTimeout`）。
+- **措辞更正**：不再声称“请求体逐字节不变”——真实契约是**不加请求字段**（不改 `reasoningEffort`、不注入 `service_tier` / `max_tokens`）；开着时模型输入里多出的内容是本插件自己的调度提示与提醒，关掉后不存在。
+- 移除从未被读取的常量 `AUTO_LIMITS.parallelMin`。
+- 1.4.4 的 1.86×（7 家族 14 对、`finishLineSteps: 10`）保留为**已退役默认的消融证据**，不是 1.5.0 的速度声明。~~1.5.0 的重基线：9 家族 54 个完整单元、隐藏测试全绿，总量 0.86× 墙钟 / 0.95× 输出 / 0.85× 步数~~ **（2026-10-26 撤回：运行目录曾放在仓库内、指标按字节偏移切片，两个缺陷都会污染数字；干净重基线见 1.5.1。）**
+- **计量日志有上限**：`~/.dsh/fast-mode-metrics.jsonl` 每步一行、只追加，装几个月会无限长（本机 7.6 MB）。现在超过 8 MB 就把当前文件改名成 `.1`、新开一份继续写，最多留两代（约 16 MB）；`bin/report.mjs` 会把两代一起读，轮转不丢历史。
+
+## 1.4.4
+
+**收尾线定版：7 家族 14 对中位 1.86×、输出 −52%、正确性 14/14 零损失。**
+
+- 新增 `finishLineSteps` 并**设为默认 10**：第 10 步由调度器注入一次「别再扩大范围、不重跑已通过的检查、只修还在报的错、然后报告」，不阻断工具、不截断步骤。E1 七家族 14 对 ABBA：中位 **1.86×**、**13/14 更快**（符号检验 **p=0.002**）、输出 **−52 %**、步数 **−10.0**、隐藏测试 **14/14 全绿**；旧默认 `late` 同轮 1.31×（11/14，p=0.057）。
+- 同轮直接配对（`fl10` ÷ `late`，14 对）：中位 **1.27×**、11/14 更快（p=0.057）、输出 **−22 %**、步数中位不变——收尾线省的是**重复验证的输出**，不是步骤。
+- 新增 `stepBudget`（默认 30，可配 4–80）：硬预算只发一次「停止扩大范围、跑一次验证、直接报告」，仍不阻断工具，因为硬截断会把任务切在编辑中途。
+- 被否掉的候选写进 BENCHMARK：`stepBudget: 15`（C1 2.04× / D1 1.26×，短任务桶 0.43–0.77×）、`finishLineSteps: 8`（C1 1.63× / D1 1.50×，6–10 步桶 0.44×）、两条叠加（D1 1.14×）。
+- 归因沿用 B1：`late-guid0`（关掉提示段）0.96× vs `late` 1.26×——提示段是载荷；并行宽度 `parallel: 1` 1.10× 不是。
+- 语料：5 轮 ABBA、272 次运行、7 个家族、隐藏测试确定性评分，全部 `exit 0`；`migrate` 的 12 次“红”经查是隐藏测试过严（只接受 `address.city`），已归一化为 `missing_field:address`，不是 agent 的错。
+- 文档：`docs/REFERENCE.md` / `.en.md` 补上两个旋钮的默认值与实测，`outputDiscipline` 的默认值从旧文案的 `true` 更正为实际的 `late`。
+- 205→206 个测试（默认值、第 10 步触发、每轮一次、步边界）。
+
+## 1.4.3
+
+**速度随步数出现：长任务 1.49×（迟到介入 2.68×），正确性 16/16 与 12/12 全对。**
+
+- 步数分桶（80 对 ABBA 语料）：**1–5 步 1.00×**（47 对）、**6–10 步 1.23×**（33 对，符号检验 p=0.035）、11 步以上无数据。`corr(输出 token 变化, 加速) = −0.77`——赢的那些对都是模型写得少，不是并发。
+- 输出纪律单独 A/B（12 对，深 fixture）：输出 token **−15.4%**（未达预设 20% 门槛）、墙钟 1.13× 但只有 7/12 对更快（p=0.774）、正确性 **96/96 不变**。结论：单看它自己，与 ±30% 的任务内噪声分不开，不作为独立卖点。
+- 新增长任务 fixture：从零实现 `ledger.py`（6 个函数、会计括号负数/货币符号/Decimal/ROUND_HALF_UP/并列取最小等边界），**隐藏 22 条断言**由 runner 事后执行。8 对 ABBA：off 215.2 s / 16.5 步 / 32 718 token → on 144.5 s / 10.5 步 / 24 274 token = **1.49×**、步数 **8/8 对下降**（p=0.008）、输出 **−25.8%**、**16/16 次运行 22/22 全对**。
+- 对照：上一版“12 个 bug”的修复型 fixture 4–6 步就被做完，实测 **0.94×、输出 +11%**——它不是长任务，正好印证分桶：步数不够就没有可省的。
+- 新增 `outputDiscipline: 'late'` 并**设为默认**：系统提示不再放文风规则（短任务零成本），第 4 步由调度器注入一次压缩提醒（`src/auto.js` 的 `compress-output`，每轮一次）。6 对 A/B：184.2 s → 68.6 s = **2.68×**、步数 10.5→7.5、输出 **−52%**、**12/12 次运行 22/22 全对**。`true` 仍可全程生效，`false` 关闭。
+- 短任务对照（4–6 步 fixture，6 对）：late 13.5 s vs 关 17.6 s（中位 1.31×、配对中位 1.26×、4/6 对更快、步数不变）——在 ±30% 噪声内，可读作“短任务没有付出代价”，与“短任务不介入”的产品设想一致。
+- 撤回旧说法：配置说明里的“每步输出 640→190 token、墙钟快约三分之一”从未被单独测过，已换成上面的实测数字。
+- `src/auto.js`：新增 `notePolicy()` 与 `compress-output` 指令，`AUTO_LIMITS.disciplineAfterSteps = 4`，按轮重置 `disciplineSent`。
+- 198→202 个测试（三档取值、每轮只提醒一次、新默认值）。
+
+## 1.4.2
+
+**量化并发，并撤回一个自己造出来的结论。**
+
+- `shell_parallel` 改为**默认关闭**（`shellParallel: false`）。6 个 5 秒作业的 fixture 上三对 ABBA：原生 9.1 s vs 开工具 26.3 s（**0.35×**）。工具本身是并发的（5.1 s 跑完），但基线模型本来就用 `&` + `wait` 自并行，多出的工具诱发一次“再复核”的额外步骤。
+- 提示段只在工具真的注册时才提 `shell_parallel`（`src/guidance.js` 拆出 `SHELL_JOBS`），避免指着不存在的工具。
+- 新增 `guidance` 开关并**修好透传**：`resolveConfig` 之前没把它（和新加的 `shellParallel`）返回出去，导致一次 54 次运行的三臂实验实际上只有两臂配置相同。该实验降级为**噪声标定**：同配置在 n=6 下能给出 0.75× 与 1.01× 两种“效应” → 小任务每任务中位数噪声约 ±30%。此前“notext 1.01× / text 0.88×”的说法**作废**。
+- 新增实测：4 个子代理各 `sleep 20`，整段跨度 **21 s**（串行 80 s，约 3.8×）——并发原语可用，瓶颈在模型是否拆活。
+- 新增实测：把 4 个模块的真实分析交给子代理扇出，**0.25×**（中位 36.3 s vs 9.2 s，步数 5.5→28.5，输出 token 1.5 K→9.6 K）。并发是真的（4 个孩子的 llm 时间确实重叠），但每个子代理的固定成本在小任务上吃掉全部收益——扇出只在深子任务上有意义。
+- 新增实测（pilot，2 对）：深子任务扇出仍 **0.60×**（中位 48.2 s vs 28.8 s，步数 12.5→43，输出 token 5.1 K→16.9 K），且两次里有一次 4 个模块全算错（串行两次全对）。会话取证：4 个子会话在 3 毫秒内同时创建，派发确实并行；亏在每个孩子要 15–25 步才做完一个模块，而串行 11–14 步做完全部四个。扇出到此结案。
+- 196→198 个测试（提示段门控、工具默认值）。
+- 新增人群统计，否掉“重复调用短路”：7168 次工具调用里严格重复 0.1%、宽松 4.4%，且都是 `ls`/`cat`/`python3 gen.py` 这类便宜或有状态的命令——省不到每步 0.5 s 的 TTFT，还会把有状态命令误当纯读。不做。
+
+## 1.4.1
+
+**诚实修正：扩大样本后 1.4.0 的 1.86× 被证伪，并修掉由此暴露的回归。**
+
+- `src/shape.js`：新增 `PER_ITEM` 形状（`one object keyed by file` 这类逐项输出表）判**原生**。1.4.0 的“一次抓全”在这个形状上实测 **0.85×**（15 对）：八次读并成一步，但该步要一次解码整张表，输出 token 1069 → 1197。
+- 实测扩样（76 对，ABBA，全部 PASS）：**中位 1.06×，CI 0.996–1.155×，45/76 对变快，均值 1.15×**。t1 0.85×→1.02×（修掉回归）、t3 1.77×→1.03×（15 对）、t9 写扇出 0.98×、t10 读写混合 1.11×。
+- 结论写在 `BENCHMARK.md` 顶部：墙钟由输出 token 主导，只有“步数下降且单步输出不涨”的形状（shell 扇出）能稳定赚到 ~8%；t6/t8 的 1.4×/1.3× 仍只有 5 对，只能当方向。
+- 181→185 个测试（新增逐项输出、聚合报告两个用例）。
+
+## 1.4.0
+
+**三步合并成一步：形状第一步就判、已知目标一次抓全、聚合用一条命令。**
+
+- `src/guidance.js`：新增 GATHER 规则（已知目标必须进第一个块；>10 个目标或求聚合时改用一条 `rg -c` / `grep -r` / 只打印的短脚本；独立 shell 作业可以同块，不许写成 `&&` 长链）。
+- `src/auto.js`：分类出的宽度**整轮保持**，不再只在第一步生效；同轮后续步骤不会因为上一次调用是写或单条 shell 就退回原生。
+- `src/shape.js`：shell + 扇出（每个目录一个作业）判为宽；纯 shell、串行、聚合之外仍原生。
+- 实测 43 对（8 任务，ABBA，120/120 PASS）：**中位加速比 1.09×（CI 1.03–1.33×），30/43 对变快**，平均步数 5.5 → 4.7；t1 1.86×、t3 1.77×、t6 1.43×、t8 1.29×，t2/t4/t7 为设计的无操作对照（约 1.0×）。
+- 代价：输入 token 上升（抓全 + 聚合），输出 token 略降；单任务 5–8 对，方差仍大于单任务效应。
+
+## 1.3.0
+
+**第一步就知道该不该宽：形状从请求本身的措辞读出来，不再付探测费。**
+
+- 新增 `src/shape.js`：一次正则遍历判断请求形状。串行词（chain / one by one / 依次 / 逐个 / 依赖链）→ 原生；shell 词（bash / python3 / 执行 / 运行）→ 原生（宽块在 shell 上本来就会串行化）；扇出词（independently / every file / across / per file / 每个 / 所有）或 ≥4 个被指名的目标 → 宽度 8；其余一律原生。
+- 分类结果由 `noteEntry` 记入该轮，只被本轮第一步消费；同一轮后续步骤不能中途改判。
+- 因此**新会话不再有 4 宽探测**：判不出形状的请求直接原生，和没装插件同宽；判得出扇出的请求第一步就宽。
+- 试跑（8 任务 × 4 对 ABBA，32 对，64 次运行，质量 63/64，唯一失败在对照组）：中位加速比 1.08×（CI 0.88–1.21×），18/32 对变快；t1 从 0.78× 回到 1.16×、t4 从 0.71× 回到 1.20×，即“不再对非扇出任务下批处理指令”确实拿回了探测的成本。样本仍小，作为 pilot 而非结论。
+- 新增 11 个形状测试（含中文扇出/串行、链式命名多文件的陷阱），全套 183 通过。
+
+## 1.2.1
+
+**把"独立读块"变成程序化判据，并且真的用 8 个任务量过一遍。**
+
+- 调度器新增 `lastStepDistinct`：这一步所有调用都在 `SAFE_CALLS` 里，且参数签名两两不同，才算"独立读块"。两个调用读同一路径会被判成链而不是批，不进入宽模式。
+- 新的一轮请求如果没有并行形状的历史，宽度直接是 1（原生、整段提示不渲染），不再每轮先付一次 4 宽的探测；只有会话历史里出现过独立读块或连续纯读才提速。
+- 独立读块一旦出现（≥2 个调用）就立刻 → 8，不需要等第二步；单调用步骤即使只读也保持原生。
+- 实测（8 任务 × 5 对 ABBA，80 次真实 headless 运行，质量 80/80，同模型同 `effort: medium`）：40 对合并中位加速比 0.96×（CI 0.80–1.14×），19/40 对变快；可并行形状 t2 1.14×、t5 1.14×、t6 1.28×，摘要/审计形状 t1 0.78×、t4 0.71×。**结论：没有普遍变快，收益集中在独立读。**
+- 局限已写进 README：判据只能用"已经发生的调用"，单轮任务只看到新会话那一次 4 宽探测，形状证据来不及积累。下一步是把请求入口的形状（独立文件/目录数）也变成信号，并对宽度做在线选择。
+
+## 1.2.0
+
+**只在值得的地方提速。** 调度器不再固定给一个"快"的宽度，而是每一步先判断形状：只有证据表明这一步是"多个互不依赖的读"才要宽，其余保持原生（宽度 1 = 整段指导不渲染）。
+
+- 新会话先用 4 做一次探测；连续两步纯读 → 8；上一步同块 ≥3 个读 → 8；跨轮记住上一轮末尾的宽读，下一轮开局直接 8。
+- 出现写调用、单调用步骤、或连续 3 步无进展 / 连续 3 次报错 → 回到原生，不再要求批处理，也不再付这段提示的输入钱。
+- 实测（4 任务 × 6 对 ABBA，同模型同 effort，质量 24/24）：可并行的 t2 墙钟 0.69×、输出 0.72×；t1 0.89×；t4 1.04×；无可并行形状的 t3 1.24×。24 对中位 0.96×，方差大于效应 —— 不发布单一加速比。
+- 不新增配置项；旧 `auto` 含义保持不变，行为按上面的证据规则细化。
+
+## 1.1.0
+
+**只剩一个开关。** 闪电本身就是开/关：点一下开，再点一下关。196 px 的面板整个删除——没有弹层、没有下拉、没有第二层界面，也没有一件只有面板里才能做的事。
+
+- `client.js`：删除 `FastPanel`、菜单开关逻辑、点击外部/Esc 关闭、`aria-haspopup`/`aria-expanded`、`data-open`、全部 `.fm-menu*` 样式与 `menu.*` 词条（941 → 716 行）。按钮保留 `aria-pressed` 与 tooltip；未就绪时先重读再切换，拒绝写入仍在 tooltip 说明原因。
+- **说明搬进插件介绍**：`meta.description` 与配置项描述现在完整讲清“同强度、并行 2–8、压输出、卡顿上限、每会话独立、重启记得”，不再占输入区空间。
+- **每会话记忆**：一次点击只写本会话在 `sessions` 列里的条目连同 `enabled`/`gear`；`settings` 服务把 volatile 配置原子写回 profile patch，所以重启、重开会话、切标签页都回到该会话自己的选择（例如 medium + fast）。
+- 测试重写为 20 个直点开关的用例（单击即写、二次关闭、写出的三个字段里没有 effort、拒绝回滚、只读不写、字典无 `menu.*`、CSS 无 `.fm-menu`、按会话/重启恢复）。全套 168 通过。
+
+## 1.0.1
+
+**修复"每轮 30 步"实际变成"整段会话 30 步"。**
+
+- `index.js` 现在在 `session/event` 的 `turn/start` 上调用 `noteTurn()`：每轮清零 `turnSteps`、`stepCalls`、`lastStepCalls`，并重新武装停止线；跨轮保留调度记忆（试过什么、卡了多久）。
+- 修掉一个连带问题：`stop-and-report` 的去重原先按"上一次指令种类"跨轮生效，导致第二条长轮次永远不再触发。去重现在只依赖本轮的 `stopped` 标志，`turn/start` 会把它复位。
+- 新增回归测试：一轮 30 步触发一次停止提示 → `turn/start` 后下一步不再立即停止 → 再走 30 步仍会触发。
+- `docs/REFERENCE*.md` 中原先记录的"已知实现缺口"随之删除。
+
+## 1.0.0
+
+**Fast 模式从此不碰思考强度。** 0.9.x 之前它按档位改写 `reasoningEffort`，最高档甚至发 `off`——这是错的。用户选了多深的思考，就应该是多深；开启快模式改变的是往返次数，不是智力。1.0.0 把整条降档机制**删掉**（不是关掉、不是加开关），`agent/request` 钩子现在是纯透传：`return resolved`。
+
+### 删掉了什么
+
+- 整套 effort 策略：`floor`、`decideEffort`、每步降一档的阶梯、`ludicrous` 发 `reasoningEffort: 'off'`、以及所有按档位写死的 effort 表。配置里再写 `floor`/`ladder`/`aliases` 不会生效，源码里也不再有 `reasoningEffort:` 赋值（有一条测试扫描源码来守住这点）。
+- 面板里的五档列表、速度/倍数估算、每档的等待上限表。它们展示的是已经不存在的选择。
+
+### 现在靠什么变快
+
+同样的思考强度、同样的模型、同样的供应商，只改三件事：
+
+1. **更宽的工具并行**：一步里允许发多个互不依赖的调用，宽度 2–8 由调度器按最近几步的表现自动决定（默认 4，同一步内连续两步顺利后放宽，报错或卡住时收窄）。
+2. **更少过程文字**：输出纪律只收叙述性文字，答案本身一个字不删。
+3. **更少的卡顿**：前台命令有可取消的上限（`auto` 下 120 s 起、按最近 p90 实际延迟在 2–5 分钟自适应；`safe` 不限），卡住会取消并把"换个做法"写回给模型；另有 6 步无进展告警、12 步停止、每轮 30 步后不再放宽的循环保护。
+
+代价是 token 可能更多（并行调用、一次读多处），这是明确的交易：**用 token 换墙钟时间，不用智力换**。
+
+### 面板与档位
+
+- 闪电点开只有一个开关、一行调度状态（引用的上限从配置里读，不再写死）、两行说明。196 px 不变，仍然跟着会话走。
+- `gear` 对外只剩 `auto` 有意义；旧配置里的 `fast`/`turbo`/`max`/`ludicrous` 会被当作 `auto` 并打一条弃用日志，而不是静默降思考强度。`safe` 保留为实验对照臂（固定并行 2、不设上限、不注入提示）。
+- 升级后不会再有"每次进来思考强度都变了"：胶囊显示的是你自己选的档，重启也一样。
+
+### 保留
+
+- 0.9.1 的 v4 会话格式修复（`kind: 'plugin:fast-mode'`）与回归测试。
+- 每会话开关、指标日志、报告与实验 CLI、压缩类插件共存、与 `dsh-batch-tool-calls`/`dsh-ops` 的互斥声明。
+
+## 0.9.1
+
+**Upgrade if you installed 0.9.0.** Its post-compaction reminder could fail a
+whole turn. Nothing else changed.
+
+- **The fix.** The two notices were built with `source: { kind: 'plugin', plugin: 'fast-mode' }`,
+  the pre-v4 shape. DSH's v4 session encoder refuses that pair —
+  `SessionFormatError: format v4 message requires a producer-owned source kind`
+  (`@deepseek-ai/dsh-session-format-v3-to-v4`, `assertV4MessageSources`) — and it
+  refuses it while persisting the event, outside the `agent/pre-step` handler
+  that built the message, so the failure landed on the turn rather than on the
+  notice. A conversation that compacted and then resumed died at the first step
+  after the checkpoint. The source is now `{ kind: 'plugin:fast-mode' }`, which is
+  also the kind v3→v4 migration derives for rows written by older versions, so
+  old transcripts read back under one shape.
+- **Why it cannot come back quietly.** `createNoticeMessage` now validates its
+  source against the mirrored v4 rule before returning, and throws *there* — the
+  caller's `try/catch` then drops the notice and the step proceeds, which is the
+  failure mode this plugin is allowed to have. `isProducerOwnedSource` mirrors
+  the encoder's rule, and `test/notices.test.js` pins it: the retired
+  `{ kind: 'plugin', plugin }` shape must be refused, `plugin:fast-mode` accepted,
+  no leftover `plugin` field. Verified against the released codec extracted from
+  `app.asar`: both notices encode, both retired shapes throw the exact error above.
+- **Existing sessions are unaffected.** The bad row was only ever written by the
+  v4 encoder, which refused it, so nothing was persisted. v0-format sessions under
+  `--private-tmp--` that still carry `kind: "plugin"` rows go through the
+  documented v3→v4 migration, which rewrites them to `plugin:<name>`.
+
+## 0.9.0
+
+Tool wait is the third lever, and the panel stops claiming measurements it never
+made.
+
+- **A stall guard bounds one tool call.** `tools/execute` now races a guarded
+  shell-style call (`bash`, `script`, `monitor`, `pwsh`…) against a per-gear
+  budget: 300 / 180 / 120 / 90 s for Cruise / Sport / Track / Redline, and no
+  watch at all on Eco. On expiry the plugin aborts *its own* controller — never
+  the caller's signal, which the registry re-fuses anyway — gives the tool 250 ms
+  to unwind, restores the caller's signal, and returns a `ToolExecutionFailure`
+  telling the model to move the work to `run_in_background`, pass a `script_path`
+  instead of a heredoc, narrow the probe, and never re-issue the same command
+  unchanged. Waiting tools (`ask_user_question`, `wait_agent`, `job_output`,
+  `job_list`, `task_board_run`, `workflow`, `subagent`, `subagent_fork`) and the
+  file tools are exempt. The ceiling is a **policy budget, not a measurement**;
+  `stallGuard: false` or `toolTimeoutMs: -1` disables it, `toolTimeoutMs: 90000`
+  applies one number to every gear.
+- **The prompt half ships with it.** Four rules joined the execution-speed
+  section: bound any command that can hang, prefer `run_in_background`, write
+  long/quoted/multi-line commands to a file instead of a heredoc, and change the
+  command before retrying one that was cancelled or came back malformed.
+- **`dsh-batch-tool-calls` and `dsh-ops` must not be enabled together with it.**
+  The batch plugin is a static, gear-blind subset of this guidance; the ops plugin
+  adds two shells plus Windows-only binaries and conflicting timeout advice. The
+  host reads the profile's dependency files once on load (and probes for live
+  `ops_*` tools); if either is present it logs one warning and appends one notice
+  to the next step, then goes quiet. It never hard-fails the user. Both reference
+  manuals document the conflict.
+- **The panel says "estimate", and shows the one number that is not.** The footer
+  now reads `倍率预估 · 单次调用 ≤ 3 分` instead of `倍率 = 相对 Eco 的实测中位数`:
+  our own runs do not support a per-gear speed or token ratio — on single-file
+  tasks every enabled gear lands within noise of `safe` — so the multipliers are
+  labelled as estimates while the stall ceiling is quoted straight from the
+  host-side table for the selected gear.
+- **The five gears have names now: Eco / Cruise / Sport / Track / Redline.** The
+  ids are unchanged (`safe` / `fast` / `turbo` / `max` / `ludicrous`), so existing
+  patches, settings and experiment files keep working — only the labels moved.
+- **The gear menu labels its own columns.** The 0.8.0 popover printed
+  `1.6x · 1.5x` with no header, so nothing said which number was the speed and
+  which was the token cost. A `Gear / speed / token` header row now shares one
+  CSS grid with the five gear rows: the numbers sit under their labels, right
+  aligned in tabular figures, in both locales. Each row's `aria-label` names the
+  columns too, so the pair is not opaque to a screen reader either.
+- **The fine print wraps instead of running off the edge.** The old single line
+  (`倍率预估：速度 · token * 关闭思考，质量会下降`) was `white-space: nowrap`
+  inside a 196 px popover and got clipped mid-sentence. It is two short lines in
+  a footer above a hairline now — the estimate note (which also carries the
+  selected gear's stall ceiling) and the warning — and both wrap.
+- **README: 484 lines to 98.** The front page keeps the principle, the measured
+  table, thirty-second setup, the five gears, the coexistence rules and what is
+  next, one short paragraph each. Configuration, logging, measurement instruments,
+  experiment methodology, safety boundary, limits and the test guide moved to
+  `docs/REFERENCE.md`. The result table states the honest part in the open:
+  turbo is 537 ms slower and 5 % fatter per step — it wins on step count.
+
+198 tests, green. `test/stallguard.test.js` (12) covers the budget table, the
+signal dance, the failure text, the frozen notice messages and the peer scan;
+`test/plugin.test.js` adds the wiring cases (registration, a hung `bash`
+cancelled at `toolTimeoutMs: 20`, passthrough, the post-compaction reminder);
+one `test/client.test.js` case pins the footer to the host-side budget table so
+the two halves cannot drift.
+
+## 0.8.0
+
+Everything up to 0.7.0 optimised effort, and Experiment E showed effort is not the
+lever on a self-hosted route: raising it does not shorten a session. What does
+shorten a session is fewer round trips and less visible text. 0.8.0 pulls both,
+and it measures the result instead of asserting it.
+
+- **The execution-speed guidance is one system-prompt section, not a rewritten
+  prompt.** `fast-mode:execution-speed` at order 520 states the batching rules,
+  the gear's parallel target, the speculative-prefetch line and the
+  output-discipline layers. It is injected only for gears marked `directive`, so
+  `safe` injects nothing at all — it stays the pre-0.8.0 behaviour and works as
+  the control arm.
+- **Output discipline is the largest single lever.** `outputDiscipline` (default
+  `true`, not volatile, so changing it needs a reload) tells the model not to
+  narrate around tool calls, not to restate file contents, and not to write
+  end-of-step progress reports. Measured: per-step output **640 → 190 tokens**.
+- **Gears gain a `parallel` target and a `discipline` level**: safe 2 / fast 4 /
+  turbo 6 / max 8 / ludicrous 8 per step; safe injects no discipline, fast gets
+  the light version (three prohibitions), turbo and above the full version plus
+  speculative prefetch. An explicit `parallel` overrides the target — `gear:
+  turbo` + `parallel: 8` fits two more independent calls into the same round
+  trip and never adds a step.
+- **A warning when the profile caps the target.** If `agentLoop.maxParallelToolCalls`
+  sits below what the gear asks for, one warn names the cap and the line to
+  change in `cordis.patch.yml`; it fires once per distinct `parallel` value, not
+  once per request.
+- **Measurement: `bin/experiment.mjs`, and Experiment E.** 16 same-task sessions
+  assigned by hash rather than clicked, `safe` vs `turbo`: median Σ step time
+  25.9 s → 15.6 s (−40 %), median steps 7.5 → 4.5, per-session output tokens
+  3,798 → 2,180 (−43 %), Mann–Whitney p = 0.008, and all ten answer facts agreed.
+  Raw data and the caveats (uneven arms, small n) are in BENCHMARK.md.
+- **The trade, stated plainly**: this plugin does not save tokens; it spends
+  tokens to buy time. Every lever here adds tokens to the same route's request,
+  and none of them works by making the model dumber.
+
+173 tests (8 new in `test/guidance.test.js`); the suite is green.
+
+## 0.7.0
+
+Fast mode was a global switch wearing a per-conversation costume: the bolt sits in
+one composer, so clicking it read as "this conversation", and it changed every
+conversation in the app. Two sessions side by side, one clicked, both accelerated —
+and nothing in the UI said so.
+
+- **The switch and the gear are per conversation now.** A click writes that
+  conversation's own `{enabled, gear}` entry; the panel says `仅本会话生效` /
+  `This conversation only` under the switch, so the scope is stated where the
+  decision is made instead of being inferred.
+- **A conversation with no entry inherits the shared pair**, which doubles as the
+  record of the last choice made anywhere. A brand-new conversation therefore
+  starts from the last selection rather than from the schema default — the
+  behaviour asked for, and the reason the shared fields still exist.
+- **Precedence is explicit**: experiment assignment > that conversation's entry >
+  shared default. An assigned trial still outranks the UI, because a click that
+  quietly overrode the arm would empty the control group of the very trial the
+  user asked for.
+- **One click, one patch.** The shared `enabled` / `gear` and the `sessions` entry
+  travel in a single `settings/update` carrying the revision that was read. Two
+  writes would mean a window where the panel shows one gear and the request runs
+  on another, plus a second revision race; a refused write rolls back all three
+  fields and states the reason in the panel.
+- **The column is bounded and forgiving.** `sessions` holds at most 200 entries,
+  oldest pruned first by `ts` (a pruned conversation just falls back to the
+  shared default). A hand-edited or truncated value degrades to "no overrides"
+  and never takes the chip down.
+- **The two copies of the format are tested against each other.** The host reads
+  the column (`src/state.js`), the browser half writes it (`client.js`), and a
+  drift between them is a switch that looks on and changes nothing — so the
+  shipped `client.js` is loaded in a `vm` and fed the same fixtures as the host
+  module. Writing the test found one real drift (the browser copy accepted an
+  empty session id the host drops), which is now fixed on both sides.
+- The conversation id reaches the client half through the slot's own scoped
+  `inject(sessionId)`; the chip also accepts a standard `sessionId` prop, so a
+  release that stops injecting one spelling still behaves.
+- One caveat, stated where a user will hit it: `sessions` renders as a text input
+  on the plugin settings page, because this build of schemastery has no
+  `.hidden()`. Its description says it is maintained automatically; editing it by
+  hand is how a conversation ends up running a gear nobody chose.
+
+165 tests.
+
+## 0.6.2
+
+The measurement was wrong in the one way that matters most: it counted time but
+not tokens, and it lost which step each row belonged to.
+
+- **Read the event payload where it is.** `session/event` carries its payload in
+  `event.data` (`{turn, step, usage, …}`); the recorder was reading `event.turn`
+  and `event.usage` off the envelope, so every row from 0.6.0 and 0.6.1 had
+  `turn`/`step` missing and `input`/`output` at zero while timings looked fine.
+  Rows now carry step numbers and real token counts, and both the harness
+  spelling (`inputTokens`, `reasoningTokens`, `cacheReadTokens`) and the gateway
+  spelling (`uncachedInputTokens`, `thoughtTokens`, `cachedReadTokens`) are
+  folded.
+- Two regression tests feed the wrapped shape, so this cannot come back quietly.
+- Rows already written to `~/.dsh/fast-mode-metrics.jsonl` keep their zeros;
+  delete the file or filter by `ts` when comparing against new data.
+
+147 tests.
+
+## 0.6.1
+
+0.6.0 could compare gears, but not against the thing they are supposed to beat.
+An assignment always switched the mode on, so a randomized trial had no baseline:
+the only "off" sessions were ones somebody happened to run before starting the
+experiment, which is not an arm — it is a convenience sample.
+
+- **`"off"` is a legal arm.** The assignment file and
+  `dsh-fast-mode-experiment --gears` now accept `off`, meaning the control group:
+  the session runs with the plugin switched off **because the experiment said so**,
+  not because a switch was left alone. `off` is always accepted even when the
+  options do not list it, and the CLI now defaults to
+  `off,safe,fast,turbo,max`, so a fresh `start` is a controlled trial rather than
+  a ranking. `start` says which kind of run it wrote: with `off` it promises a
+  baseline, without it it warns that the report can only rank arms against each
+  other.
+- **The control arm is deliberately invisible to the switch.** A session assigned
+  to `off` returns the request untouched even when the global switch is on — the
+  one place where the UI can disagree with what a session actually does, which
+  the experiment output and both READMEs now state plainly.
+- Also lands in this release: `bin/experiment.mjs` is documented and exposed as
+  the `dsh-fast-mode-experiment` bin with an npm script, and `--lan` follows
+  `--since` into the dashboard's own `1d`/`7d`/`30d` window (default `7d`) rather
+  than always asking for one day and quietly printing nothing.
+
+145 tests.
+
+## 0.6.0
+
+0.5.2 made the switch work. 0.6.0 makes the claim checkable: fast mode measures
+itself, assigns gears to sessions for comparison, and reports the result from a
+terminal. It also adds one gear above `max`, because on the official route that
+is the only honest way past the 2× ceiling.
+
+- **Measurement is built in and on by default.** A `session/event` tap folds
+  every step into `~/.dsh/fast-mode-metrics.jsonl`, one JSON line per completed
+  step, for sessions with the switch **off** as well as on — a comparison needs
+  both sides, and nobody remembers to enable measurement before an A/B. A row
+  carries `v, ts, sid, turn, step, mode` (`on|off|unknown`), `gear` (or `off`),
+  `effort`, `route`, `ttftMs` (step start → first assistant message), `llmMs`
+  (step start → last assistant message), `stepMs` (the whole step), `calls`,
+  `errors`, `input`, `output`, `reasoning`, `cacheRead`, `cacheWrite` and
+  `assistantMessages`. `metrics: false` disables it; a write failure is
+  swallowed and never fails a request.
+- **Experiment assignment without a UI.** An optional
+  `~/.dsh/fast-mode-bench.json` (`benchmarkFile`) assigns gears per session:
+  `{"enabled": true, "gears": ["safe","fast","turbo","max"], "sessions": {"<session id>": "turbo"}}`.
+  An explicit pin wins; otherwise the gear is `FNV-1a(session id) mod gears`.
+  The file is re-read on a 1 s TTL, so an edit applies to the next request, and
+  deleting it returns control to the configured switch. A session an experiment
+  assigns runs on that assignment even with the switch off — letting a forgotten
+  click decide the arm would silently put the session in the control group.
+  Unknown gear names are dropped with one warning.
+- **The experiment has a command, not a JSON file you edit by hand.**
+  `bin/experiment.mjs` (package bin `dsh-fast-mode-experiment`) writes the
+  assignment file, reports how many sessions and steps each arm has collected so
+  far (`status`, including whether a disabled-state sample exists at all), and
+  refuses to overwrite a running experiment without `--force`. An experiment is
+  what makes the numbers trustworthy: the arm is chosen by the session-id hash,
+  not by whoever wants the better result.
+- **A report command instead of a spreadsheet.** `bin/report.mjs`
+  (package bin `dsh-fast-mode-report`) reads the JSONL and prints a per-stance
+  table plus a speedup table against the `off` baseline. The unit is the step;
+  session wall clock is not used, because it mostly measures the human. Flags:
+  `--file`, `--since 2d|12h|90m`, `--json`, `--lan`, `--help`. `--lan` joins
+  local rows to the self-hosted usage dashboard by session UUID using
+  `USAGE_TOKEN` (optional `USAGE_BASE`): read-only, nothing leaves the LAN, and
+  the token is never printed. The dashboard's own window (`1d`/`7d`/`30d`)
+  follows `--since` instead of always asking for one day. An unreachable,
+  slow or unauthorised dashboard degrades to one note; it never costs you the
+  local report.
+- **New gear `ludicrous` (极速), above `max`.** It is the only gear whose floor
+  is `off`, i.e. it may set `reasoningEffort: 'off'`, with a fan-out target of 8
+  and `deepAfter: 1`. That is not "a bit faster than max": the official route
+  declares `off/low/high/max`, so with thinking on the only legal cut is
+  `high → low` (≈2×) and 3–4× requires turning thinking off, which this gear
+  does. The panel therefore marks it `极速*` / `Ludicrous*` and the footer says
+  `* 关闭思考，质量会下降`. The plugin still never emits an id the route does
+  not declare — on a route without `off`, the gear degrades to what the ladder
+  offers.
+- **The chip is one bolt, and the panel is behind it.** The caret and its inline
+  dropdown are gone. Clicking the bolt opens a panel holding the on/off switch,
+  the five gear rows (each with its estimated multipliers, the current one
+  marked) and the fine print, plus a status line that states the reason when a
+  write is refused. Clicking does not toggle; picking a gear while fast mode is
+  off still arms `enabled` in the same write.
+- **The gear is a name.** `gear` is a schema union of
+  `safe/fast/turbo/max/ludicrous`; a number or a typo is a load-time error, as
+  `src/config.js` already documented. The policy layer still normalises
+  defensively, so a bad value can never become "unlimited", but it no longer
+  silently becomes `safe` either.
+- **Measurement paths are composition knobs.** `metricsFile` and
+  `benchmarkFile` are deliberately not `.volatile()`: they are read at load and
+  do not belong on a live form. Moving the JSONL is a `cordis.patch.yml` edit.
+
+Tests 142: the measurement tap proves off sessions are measured too (one row per
+step, first-token vs last-message latency in a multi-message step, token and
+error sums, per-session isolation, a missing directory created on first write,
+an empty path disabling the sink); the assignment proves a stable hash that
+covers every gear, an explicit pin, a single-session pin, a live 1 s re-read,
+inertness without a file or with `enabled: false`, and one complaint for a
+broken file; the report proves both tables, a one-sided refusal in words,
+`--since` including the wider dashboard window, `--json`, `--help` reading no
+file, `--lan` without a token and with one (the token is never printed), and an
+unreachable dashboard degrading to a note; the client proves the contract after
+the caret removal — one icon, clicking opens the panel and writes nothing by
+itself, the switch writes once and a second click turns it back off, five rows
+with estimates and an asterisk on `ludicrous`, and picking a gear while already
+on leaves the toggle alone; and the schema proves an unknown gear name is a
+load-time error, measurement is on by default, and exactly
+`enabled`/`gear`/`parallel` are volatile.
+
+## 0.5.2
+
+0.5.1 shipped the switch but left it inert, and the reason was one wrong field
+access:
+
+- **The remote protocol wraps every answer in `{ ok, value }`.** The chip read
+  `remote.settings.describe().namespaces` straight off the envelope, so it always
+  got `undefined`, classified this plugin as "not configured", and rendered a
+  `disabled` button. Chrome suppresses `title` on a disabled button, so hovering
+  said nothing, and the click handler's guard swallowed every click — a control
+  that was fully explained inside the code and completely dead on screen.
+  `unwrap()` now handles the envelope (and a bare payload, so a service stub
+  still works), `ok: false` surfaces `error.message`, and every remote call has a
+  deadline so a pending bridge becomes a visible error instead of an endless
+  "loading".
+- **The chip is never disabled.** Read-only, unknown and missing-namespace states
+  all stay clickable and open the menu with the reason in words.
+- **The gear lives on the chip.** The caret opens a dropdown listing the four
+  gears, each row showing its estimated speed and token multiplier
+  (`速度 ~1.6× · token ~1.5×`) and marking the current one. Picking a gear while
+  fast mode is off turns it on in the same write. There is deliberately no gear
+  form on the plugin page: this harness renders no generic schema form, so one
+  working entry point beats a second half-working one.
+- **Host-side traceability.** With `debug` on, `apply` writes one line per load
+  before the `enabled` check (`apply v0.5.2 entered enabled=… gear=…`), then one
+  more saying it registered nothing; after two seconds it also reports whether
+  `settings.describe()` actually serves this namespace. A future "the chip does
+  nothing" report is now one log line away from an answer.
+
+Tests 89: the client suite runs the real bundle against the real wire envelope,
+including the bare-payload tolerance, the refused write, the read-only chip, the
+missing namespace, and the gear multipliers.
+
+## 0.5.1
+
+0.5.0 shipped a settings switch that could not be switched. Two defects, both
+about how the `settings` service treats a schema field:
+
+- **The three form fields were not `.volatile()`.** `settings.describe()` projects
+  a Config through `volatileForm`, which keeps only nodes carrying
+  `meta.volatile`; with none, the whole entry disappeared from the form — so
+  there was no settings page, the composer chip read it as "absent" and stayed
+  greyed out, and `settings.update()` would have refused the write with
+  `Config field "enabled" is not volatile`. `enabled`, `gear` and `parallel` are
+  now declared volatile, which is also what lets an edit land on a running fiber
+  without remounting the plugin.
+- **A volatile field is a reference, not a value.** The Loader hands a running
+  plugin a frozen `{get()}` reference so it can swap the value in place;
+  `config.enabled === true` would therefore have been false with the switch on.
+  `resolveConfig` now dereferences at that single boundary (and still accepts a
+  plain config object), and the module documents the protocol.
+
+Tests 80: the schema now has to prove every form field is volatile, and that a
+reference from the Loader reads as an armed switch.
+
+## 0.5.0
+
+Fast mode becomes something you switch, not something an install decides for you.
+
+- **Opt-in.** `enabled` now defaults to `false`, and `apply()` registers nothing —
+  no hooks, no prompt section, no discovery — until something sets it to `true`.
+  Only a literal `true` arms it; `1` or `"yes"` do not.
+- **A lightning switch in the composer.** New client half (`client.js`, served
+  through `dsh.client`) registers on `conversation.input.activity`, beside the
+  model selector: hollow bolt = off, filled black = on, the tooltip carries the
+  gear and the reason when it is not clickable. It reads and writes the plugin's
+  own `enabled` through `remote.settings.describe()` / `update()`, so the chip,
+  the generated Settings page, and `cordis.patch.yml` are three views of one
+  field. Clicks are optimistic and then verified by polling `describe()`; a failed
+  write reverts the icon and surfaces the error in the tooltip. A registration
+  taken by another plugin, an unreadable document, or a read-only layer degrades
+  to a disabled chip, never to an error thrown into the composer.
+- **A config schema.** `Config` (schemastery) declares only the three fields a
+  form should show — `enabled`, `gear`, `parallel`. Undeclared patch-level keys
+  (`floor`, `holdTools`, `debug`, …) pass through untouched, so existing
+  compositions keep working.
+- **One dependency.** `@deepseek-ai/schemastery` (`~3.18.5`), which every shipped
+  DSH plugin that exposes a settings page already declares; the profile already
+  has it hoisted, so this adds no new runtime.
+- **Tests.** 78 (up from 64): the default-off contract, non-boolean configs not
+  arming the policy, real schemastery validation of the schema (`gear: warp`
+  fails loudly, unknown keys survive), the client-half manifest contract
+  (`dsh.client.platform`, `exports["./client"]`, packed `client.js`, bundle id,
+  and the slot being the one that never shadows the model selector), and eight
+  tests that run the shipped `client.js` verbatim in a `vm` sandbox with a stub
+  `__ModuleLoader__` and a fake `settings` remote — proving the bundle parses,
+  declares its loader contract, paints hollow when off and filled when on, sends
+  exactly one `settings/update` with the revision it read, rolls back and explains
+  itself when that write is refused, and never renders a clickable chip over a
+  read-only layer or an absent entry.
+
+## 0.4.0
+
+Speed becomes a stance, not a single threshold, and the plugin now installs itself ON.
+
+- **Gears.** `gear: 1..4` (`safe` / `fast` / `turbo` / `max`) selects a coherent
+  stance: how early a clean streak earns the second effort notch (`deepAfter`
+  6/4/3/2), how many observations count (`window` 8/6/5/4), and how many
+  independent calls the model is asked to batch per step (none/4/6/8).
+  `fast` is the shipped default; `safe` is byte-for-byte 0.3.0 behaviour.
+- **An explicit fan-out directive.** Gears 2–4 register a short system-prompt
+  section (`fast-mode:execution-speed`, order 520) that tells the model to batch
+  independent calls. The number in the text follows the gear, or `parallel`.
+- **A loud cap warning.** If `agentLoop.maxParallelToolCalls` is below the gear's
+  target, the plugin logs the exact `cordis.patch.yml` line that raises it, once.
+- **Display metadata.** `icon.png` at the package root (PNG, 11,607 bytes) plus
+  `locale/zh.json` and `locale/en.json` with `meta.title` / `meta.description`, so
+  the Plugin Manager and Settings show a real icon and localized copy.
+- **Docs.** Chinese README is now the default, English moved to `README.en.md`,
+  and both document the gears, the parallel lever, official-API support, and how
+  to compose with context-compression plugins.
+- **Tests.** 64 (up from 55), including the manifest checks (icon size and type,
+  locale metadata, `files[]` completeness) and the guidance registration path.
+
+## 0.3.0
+
+- Discover the route's real effort ladder through `llm.resolveModelInfo()` instead
+  of assuming a vocabulary; never emit a level the route did not declare.
+- Explicit `ladder`, documented-alias folding, per-route startup logging, and a
+  `route=` field in the debug line.
+- `holdTools` keeps full thinking after context-management tools, which look
+  "clean" from the outside while being anything but.
+
+## 0.2.0
+
+- Observe `tools/pre-execute` / `tools/post-execute` for struggle signals (error,
+  slow call, oversized result, repeated identical call) and lower the effort only
+  while the streak stays clean.
+
+## 0.1.0
+
+- First working cut: per-request reasoning-effort adjustment for DSH, zero
+  dependencies, no session writes.