omnilane 0.10.2 → 0.10.3

package/README.md

@@ -21,49 +21,30 @@ or any hosted model via OpenRouter — on the subscriptions you already pay for,
 
 ---
 
-## 2026-07-25 update
+## 🤔 What is omnilane?
 
-- Added `claude-opus-5` to default routing: first choice for `hard-judgment` and `taste-final`, plus a fallback for the hardest coding work.
-- Expanded `omnilane configure` across all 13 providers: 106 selectable model entries, complete live catalogs for Codex, Claude Code, Grok Build, and Antigravity, plus verified OpenRouter/OpenCode shortcuts. Custom model IDs remain available through `c`.
+**The problem.** You already drive an AI coding assistant — Claude Code, Codex,
+Cursor, Gemini CLI. Each one talks to a single model family. So every task you
+give it runs on that one model, whether or not it is the right tool: a
+throwaway file rename burns your most expensive model, and a genuinely hard
+architecture question runs on whatever you happened to open.
 
----
-## 👋 New here?
-
-You already drive an AI coding assistant — **Claude Code, Codex, Cursor, Gemini
-CLI**, and the like. Each one talks to a single AI model, and picking the best
-model for each task is left to you.
+**What omnilane does.** It gives your assistant a routing table. Work gets
+sorted into **lanes** — hardest coding, bulk mechanical, triage, hard judgment,
+final polish — and each lane names the model that is best (and cheapest) for
+it. Your assistant keeps the lanes it is already good at and hands the rest to
+another vendor's CLI in the background, using the logins you already have.
 
-**omnilane picks for you.** For every piece of work it routes the task to the
-model that is best (and cheapest) for it — tough coding to a top coder, quick
-checks to a fast cheap model, long documents to a big-context model — all on the
-subscriptions and API keys you already pay for. Use the built-in defaults, or
-tweak one small file. Nothing new to babysit (it runs behind your existing tool),
-and `./install.sh --uninstall` removes it cleanly.
+**What it is not.** Not a proxy, not a new subscription, not another service to
+keep alive. It is a table plus a dispatch script that runs behind the tool you
+already use. `./install.sh --uninstall` removes every trace.
 
-**[⬇ Jump to the 60-second start](#-60-second-start)**
-
-## What's new in v0.10.0
+**You do not need every vendor.** Each lane is a fallback chain. Install one
+CLI or seven — dispatch picks the first candidate you actually have, and a lane
+with nothing available simply turns off. The default table works on a single
+subscription.
 
-- **Gemini 3.6 Flash defaults** — the gemini candidates in `fast-agentic`,
-  `triage`, and `bulk-mechanical` (and the `Gemini Flash` alias) now run
-  Gemini 3.6 Flash (released 2026-07-21): fewer output tokens, a lower output
-  price, and the fastest output speed measured by Artificial Analysis.
-- **Evidence re-audit** — routing comments, model capability notes, and the
-  Gemini price table refreshed against official sources (2026-07-21/22).
-
-## What's new in v0.9.1
-
-- **Fix:** `configure set` no longer deletes hand-written comments from
-  `routing.local.yaml` — it rewrites only its own stamp header and the lane
-  being replaced.
-
-## What's new in v0.9.0
-
-- **Five OpenAI-compatible direct-API vendors** — `deepseek`, `zai` (GLM),
-  `mistral`, `groq`, and `cerebras` join `openrouter` as CLI-free lanes (curl +
-  a `<VENDOR>_API_KEY`). A one-line `lib/common.sh` registry entry adds each;
-  see [`docs/model-capabilities-2026-07.md`](docs/model-capabilities-2026-07.md).
-- **Fish shell completion** — `omnilane completion fish | source`.
+**[⬇ Jump to the 60-second start](#-60-second-start)** · **[❓ Read the FAQ](#-faq)**
 
 ## ⚡ 60-second start
 
@@ -127,7 +108,10 @@ flowchart LR
 
 </div>
 
-## 🛤️ Lanes (defaults — run `scripts/dispatch.sh --list` for your effective table)
+## 🛤️ Lanes
+
+Defaults below — run `scripts/dispatch.sh --list` for the table your machine
+actually resolves.
 
 | Lane | First choice | Backup | When |
 |---|---|---|---|
@@ -148,14 +132,6 @@ The **backup** is the next candidate in the lane's `routing.yaml` chain — what
 dispatch falls back to when the first-choice vendor CLI is not installed. Every
 lane is such a chain; when nothing in it is installed the lane degrades to `off`.
 
-> **Where is Claude Fable 5?** Deliberately not in the defaults: the top
-> Claude tier is usually the *main loop itself*, not a dispatched worker, and
-> it prices above Opus. This is a cost / guardrail / main-loop policy choice,
-> not a capability verdict; Anthropic positions Fable 5 above Opus 5. It is
-> offered in the configurator's model menu —
-> route to it if you disagree (e.g. `taste-final: claude claude-fable-5 high`
-> in `routing.local.yaml`).
-
 ### Natural-language consultation
 
 With the `omnilane` skill or `/route`, you can ask normally:
@@ -238,13 +214,15 @@ optional per-CLI **routing reminder**: a marked, reversible block appended to
 each CLI's instruction file (`~/.claude/CLAUDE.md`, `~/.codex/AGENTS.md`,
 `~/.grok/Agents.md`, `~/.gemini/GEMINI.md` — paths may vary across CLI
 versions) so the main loop remembers to consult the table; non-interactive
-installs can pass `OMNILANE_HOOKS=all|none|claude,codex`. Manual wiring:
+installs can pass `OMNILANE_HOOKS=all|none|claude,codex`.
 
 Use `./install.sh --check` for a read-only drift report. Add `--dry-run` to an
 install or `--uninstall` to preview every checkout-owned file action.
 Rollback the installer-owned links and marked reminders with
 `./install.sh --uninstall`.
 
+Manual wiring:
+
 - **Claude Code**: install as a plugin (ships the skill + `/route`,
   `/route-jobs` commands, and a `SessionStart` hook that auto-injects the
   routing reminder at session open — no CLAUDE.md edit needed), or drop
@@ -394,7 +372,7 @@ code passes through.
   non-Git work running through the existing per-call watchdog path, which emits
   its own warning if no watchdog tool exists.
   Expiry cleans the supervised process group and returns 124. For a deep audit
-  of a fubon-autotrade-sized repository, start around 2–4 hours (7200–14400s)
+  of a large repository, start around 2–4 hours (7200–14400s)
   with a 30-minute per-call watchdog; these are recommendations, not defaults.
 - **Background lifecycle** — `--background` workers run in their own process
   group and survive the caller's exit; killed workers record an exit code, and
@@ -402,12 +380,144 @@ code passes through.
 - **Payload caps** — oversized task text is truncated head+tail before it can
   blow a worker's context.
 
+## ❓ FAQ
+
+<details>
+<summary><b>Do I need all of these subscriptions?</b></summary>
+
+<br/>
+
+No. Every lane is a fallback chain, and dispatch picks the first candidate
+whose CLI is actually installed. With one subscription the whole table collapses
+onto that vendor; lanes with nothing available turn off rather than failing.
+`omnilane doctor` shows exactly what your machine can reach today, and
+`routing.local.yaml.example` ships starter profiles for common situations
+(Claude-only, Codex-heavy, no-Codex).
+
+</details>
+
+<details>
+<summary><b>Does omnilane send my code somewhere new?</b></summary>
+
+<br/>
+
+No new destination. Dispatch shells out to vendor CLIs you already installed
+and logged into, so your code reaches exactly the vendors you already use.
+Runners strip API-key environment variables before invoking a subscription CLI,
+so a stray key cannot silently switch you onto pay-per-token billing. The one
+exception is the direct-API vendor family (`openrouter`, `deepseek`, `zai`,
+`mistral`, `groq`, `cerebras`), which by definition calls that provider's API
+with the key you set — those are advise-only and never edit files.
+
+</details>
+
+<details>
+<summary><b>Where is Claude Fable 5? Why is it not in the default table?</b></summary>
+
+<br/>
+
+**Because the top Claude tier is usually the main loop itself, not a dispatched
+worker.** Lanes exist to send work to a model *other than* the one you are
+driving. If Fable 5 is your main loop, routing judgment and taste back to Fable 5
+just adds a second call for no gain — which is why the "pick your main model"
+list above gives Fable 5 its own row as a **driver**, self-executing
+hard-judgment, taste-final, and the hardest correctness-critical fixes.
+
+**The measurements do not argue for it as a worker either.** On the Artificial
+Analysis Intelligence Index (2026-07-24) Opus 5 (max) scores 61 and Fable 5 (max)
+scores 60 — Artificial Analysis calls them "effectively tied", and Epoch AI's
+Capability Index ranks them the other way (Fable 5 161, Opus 5 159). Call it a
+draw on general intelligence. Where they are not tied is agentic professional
+output, and Opus 5 leads by a wide margin:
+
+| Benchmark | Claude Opus 5 (max) | Claude Fable 5 | |
+|---|---:|---:|---|
+| AA-Briefcase (agentic knowledge work, Elo) | 1720 | 1574 | **+146** |
+| GDPval-AA v2 (Elo) | 1861 | 1747 | **+114** |
+| Cost per AA-Briefcase task | $17.79 | $22.30 | **-20%** |
+| API price, input / output per 1M | $5 / $25 | $10 / $50 | **half** |
+
+Opus 5's max, xhigh and high tiers sweep the top three AA-Briefcase places, and
+its `high` tier still beats Fable 5 at under half the cost per task. So Fable 5
+costs twice as much without buying an advantage on any axis a lane is defined
+around.
+
+**What Fable 5 is genuinely better at**: factual breadth. It stays ahead of
+Opus 5 on AA-Omniscience, as its size class suggests, and Opus 5 answers more
+readily when uncertain — its hallucination rate is 50%, up 14 points from
+Opus 4.8. If your task is recall-heavy rather than execution-heavy, name
+Fable 5 explicitly:
+
+```bash
+dispatch.sh --vendor claude --model claude-fable-5 --effort high consult "…"
+```
+
+**This is a cost / main-loop policy choice, not a capability verdict.** Fable 5
+is in the configurator's model menu, and one line in `routing.local.yaml`
+overrides the default if you disagree:
+
+```yaml
+taste-final: claude claude-fable-5 high
+```
+
+</details>
+
+<details>
+<summary><b>Why do the Claude lanes use <code>xhigh</code> instead of <code>max</code>?</b></summary>
+
+<br/>
+
+Because more effort is not monotonically better. Anthropic documents `xhigh` as
+the starting point for coding and agentic work, `high` as the floor for other
+intelligence-sensitive work, and `max` as the setting for cases where
+correctness outweighs cost. Independent testing agrees: on Vals.ai's Vibe Code
+Bench, Opus 5 scores 89.8% at `high` but only 88.3% at `xhigh` and 88.4% at
+`max` — the top tiers produce more elaborate solutions that fail more often.
+Raise any lane locally if your workload disagrees:
+
+```bash
+omnilane configure set hard-judgment "claude claude-opus-5 max"
+```
+
+</details>
+
+<details>
+<summary><b>What happens when a lane's first-choice CLI is missing?</b></summary>
+
+<br/>
+
+Dispatch walks the chain and uses the first vendor you have. Inspect the
+decision without spending a call:
+
+```bash
+scripts/dispatch.sh --explain hardest-coding   # candidate-by-candidate trace
+scripts/dispatch.sh --list                     # whole effective table
+scripts/dispatch.sh --dry-run hardest-coding "…"   # fully resolved plan, no provider call
+```
+
+</details>
+
+<details>
+<summary><b>Can a dispatched worker edit my files?</b></summary>
+
+<br/>
+
+Only if you ask for it. Dispatch defaults to `advise`, a read-only mode enforced
+per vendor (read-only sandbox, plan mode, or read-only tool set depending on the
+CLI). Editing requires both `--mode work` and an explicit `--workdir`. Workers
+also cannot dispatch again — the depth guard refuses nested fan-out with exit 86,
+so one command can never spiral into a chain of agents spending your quota.
+
+</details>
+
 ## 📊 Defaults and provenance
 
 Default lane assignments follow Artificial Analysis coding/intelligence data
 (2026-07 snapshot, cross-checked against AA site records and vendor pricing
 pages) plus published head-to-head reviews; they are opinions, not laws — the
-configurator and `routing.local.yaml` exist so you can disagree.
+configurator and `routing.local.yaml` exist so you can disagree. The full
+working notes, including per-benchmark caveats, live in
+[`docs/model-capabilities-2026-07.md`](docs/model-capabilities-2026-07.md).
 
 ## ⚠️ Known limitations
 
@@ -423,8 +533,71 @@ configurator and `routing.local.yaml` exist so you can disagree.
 
 ## 📜 Release history
 
+## What's new in v0.10.3
+
+- **Restructured READMEs in all five languages** — the reader now meets a plain
+  "what is this and why would I want it" section first, version history is
+  consolidated at the bottom instead of interrupting the introduction, and a new
+  FAQ answers the questions that kept coming up: do I need every subscription,
+  where does my code go, why is Fable 5 not in the table, why `xhigh` and not
+  `max`, what happens when a CLI is missing, can a worker edit files.
+- **Fixed: plugin manifests advertised a stale version** — `plugin.json` and
+  `.claude-plugin/plugin.json` still reported `0.10.0` after the 0.10.1 and
+  0.10.2 releases, so plugin installs showed the wrong version.
+- **Fixed: `routing.local.yaml.example` shipped retired models** — the starter
+  profiles still pointed at `claude-opus-4-8` and Gemini 3.5 Flash; they now use
+  Claude Opus 5 (with lane-appropriate effort) and Gemini 3.6 Flash.
+- **Corrected the Intelligence Index figures** in
+  `docs/model-capabilities-2026-07.md` against the Artificial Analysis source
+  (index points, not percentages), added the AA-Briefcase / GDPval-AA v2
+  comparison, and recorded the two results that cut against the defaults:
+  Fable 5's lead on factual knowledge and Sol's lead on presentation quality.
+
+## What's new in v0.10.2
+
+- **Claude effort on `hardest-coding` and `hard-judgment` moved from `max` to
+  `xhigh`**, matching Anthropic's documented guidance for Claude Opus 5: start
+  at `xhigh` for coding and agentic work, keep `high` as the floor for other
+  intelligence-sensitive work, and reserve `max` for cases where correctness
+  outweighs cost. Raise it back per lane with
+  `omnilane configure set <lane> "<spec>"`.
+- **Fixed two dead CHANGELOG compare links** that pointed at a `v0.10.0` tag
+  which was never published.
+
+## What's new in v0.10.1
+
+- **`claude-opus-5` joins the default table** as first choice for
+  `hard-judgment` and `taste-final`, plus a fallback for the hardest coding work.
+- **`omnilane configure` covers all 13 providers** with 106 selectable model
+  entries — current native catalogs for Codex, Claude Code, Grok Build and
+  Antigravity, plus verified OpenRouter/OpenCode shortcuts. Custom model IDs
+  remain available through `c`.
+
 <details>
-<summary>Older releases (v0.8.3 and earlier)</summary>
+<summary>Older releases (v0.10.0 and earlier)</summary>
+
+## What's new in v0.10.0
+
+- **Gemini 3.6 Flash defaults** — the gemini candidates in `fast-agentic`,
+  `triage`, and `bulk-mechanical` (and the `Gemini Flash` alias) now run
+  Gemini 3.6 Flash: fewer output tokens, a lower output price, and the fastest
+  output speed measured by Artificial Analysis.
+- **Evidence re-audit** — routing comments, model capability notes, and the
+  Gemini price table refreshed against official sources.
+
+## What's new in v0.9.1
+
+- **Fix:** `configure set` no longer deletes hand-written comments from
+  `routing.local.yaml` — it rewrites only its own stamp header and the lane
+  being replaced.
+
+## What's new in v0.9.0
+
+- **Five OpenAI-compatible direct-API vendors** — `deepseek`, `zai` (GLM),
+  `mistral`, `groq`, and `cerebras` join `openrouter` as CLI-free lanes (curl +
+  a `<VENDOR>_API_KEY`). A one-line `lib/common.sh` registry entry adds each;
+  see [`docs/model-capabilities-2026-07.md`](docs/model-capabilities-2026-07.md).
+- **Fish shell completion** — `omnilane completion fish | source`.
 
 ## What's new in v0.8.3
 

package/README.zh-CN.md

@@ -21,41 +21,25 @@
 
 ---
 
-## 2026-07-25 更新
+## 🤔 omnilane 是什么?
 
-- 默认路由加入 `claude-opus-5`：成为 `hard-judgment`、`taste-final` 第一顺位，也纳入最高难度编程任务的备选。
-- `omnilane configure` 已扩展全部 13 个提供商：共 106 个可选模型，完整收录 Codex、Claude Code、Grok Build、Antigravity 实时列表，并加入已验证的 OpenRouter／OpenCode 快捷项；仍可用 `c` 输入自定义模型 ID。
+**问题在哪。** 你已经在用某个 AI 写代码助手——**Claude Code、Codex、Cursor、
+Gemini CLI** 之类。每一个都只接一个模型家族,所以你交代的每件事都跑在那同一个
+模型上,不管它合不合适:随手改个文件名烧掉最贵的模型,真正难的架构问题却刚好落在
+你当下开着的那个。
 
----
-## 👋 第一次用?
-
-你已经在用某个 AI 写程式助手——**Claude Code、Codex、Cursor、Gemini
-CLI** 之类。每个一次只接一个 AI 模型,而「每件工作该用哪个模型最好」得你自己决定。
-
-**omnilane 帮你决定。** 每一件工作,它会自动派给对那件事最强(也最省)的模型——硬派程式
-交给顶尖程式模型、随手的小检查交给又快又便宜的、长文件交给大脉络模型——全部用你本来就在
-付费的订阅与 API 金钥。用内建预设值就好,或改一个小设定档就能微调。不用另外顾一套东西
-(它躲在你现有工具背后跑),`./install.sh --uninstall` 可干净移除。
-
-**[⬇ 直接跳到 60 秒上手](#-60-秒上手)**
-
-## v0.10.0 新功能
+**omnilane 做什么。** 它给你的助手一张路由表。工作被分进**通道**——最难的实现、
+机械粗活、初筛、硬判断、文字终审——每条通道指名对那件事最强(也最省)的模型。
+助手保留自己本来就擅长的通道,其余用你既有的登录,在后台交给别家厂商的 CLI。
 
-- **Gemini 3.6 Flash 默认路由**——`fast-agentic`、`triage`、`bulk-mechanical`
-  的 gemini 候选(与 `Gemini Flash` 别名)改用 2026-07-21 发布的 Gemini 3.6
-  Flash:输出 token 更少、输出单价更低、Artificial Analysis 实测输出速度第一。
-- **证据重审计**——路由注释、模型能力笔记与 Gemini 价格表对官方来源刷新
-  (2026-07-21/22)。
-
-## v0.9.1 新功能
+**它不是什么。** 不是 proxy、不是另一笔订阅、不是又一个要维护的服务。它就是一张表
+加一支派工脚本,躲在你现有工具背后跑。`./install.sh --uninstall` 可完全清除。
 
-- **修复**:`configure set` 不再删除 `routing.local.yaml` 中手写的注释——
-  只改写自身的戳记行与被替换的 lane。
+**你不需要每一家订阅。** 每条通道都是候选链,派工时自动采用本机实际装了的第一个
+候选。装一家或七家都行,整条链都没有的通道就自动关闭,而不是报错。只有一份订阅
+时,整张默认表会收敛到那一家。
 
-## v0.9.0 新功能
-
-- **新增 5 个 OpenAI-compatible direct-API vendor** — `deepseek`、`zai`(GLM)、`mistral`、`groq`、`cerebras`,与 `openrouter` 同为免 CLI 通道(curl 加一把 `<VENDOR>_API_KEY`);`lib/common.sh` registry 一行即加一个。详见 [`docs/model-capabilities-2026-07.md`](docs/model-capabilities-2026-07.md)。
-- **fish shell 补全** — `omnilane completion fish | source`。
+**[⬇ 直接跳到 60 秒上手](#-60-秒上手)** · **[❓ 看常见问题](#-常见问题)**
 
 ## ⚡ 60 秒上手
 
@@ -132,13 +116,11 @@ flowchart LR
 | 🚰 coding-overflow | Grok 4.5 | Kimi K3 → Qwen3 Coder Plus → OpenCode | Codex 额度吃紧时的中量级编码溢流道;事实性声明须另行查证 |
 | 🗳️ arbitrate | off(可选评审团) | — | 内置意见评审团,重大决定用——默认关闭,要用在 `routing.local.yaml` 打开;每评审每轮烧一次额度 |
 
-**备选模型**是候选链的下一位——首选那家的厂商 CLI 没装时,派发就降到它。
+**备选模型**是候选链的下一位——首选那家的厂商 CLI 没装时,派发就降到它。每条
+通道都是这样一条链;整条都没装时,通道自动降为 `off`。
 
-> **Claude Fable 5 去哪了?** 默认表刻意不放:Claude 顶级档通常就是*主循环本人*,
-> 不是被派发的工人,且定价高于 Opus。这是成本/护栏/主循环策略,不代表能力
-> 不如 Opus;Anthropic 将 Fable 5 定位在 Opus 5 之上。设置菜单的模型清单里有它——
-> 不同意就自己路由过去(例如在 `routing.local.yaml` 写
-> `taste-final: claude claude-fable-5 high`)。
+> **Claude Fable 5 去哪了?** 默认表刻意不放——理由与完整数据见
+> [常见问题](#-常见问题)。
 
 ### 自然语言咨询
 
@@ -343,6 +325,121 @@ configure.sh set|get|unset|list|diff LANE [SPEC]    # 非交互编辑/查看 rou
   而不是永远显示 `running`。
 - **任务载荷上限** — 过大的任务文本自动头尾截断,防止撑爆工作端上下文。
 
+## ❓ 常见问题
+
+<details>
+<summary><b>这些订阅我全都要有吗?</b></summary>
+
+<br/>
+
+不用。每条通道都是候选链,派工时采用本机实际装了的第一个候选。只有一份订阅时,
+整张表会收敛到那一家;整条链都没有的通道自动关闭,不会报错。运行
+`omnilane doctor` 可以看到这台机器现在实际接得到什么,`routing.local.yaml.example`
+也附了常见情形的起手配置(只有 Claude、以 Codex 为主、没有 Codex)。
+
+</details>
+
+<details>
+<summary><b>omnilane 会不会把我的代码送到新的地方?</b></summary>
+
+<br/>
+
+不会多出新的去处。派工是调用你早就装好也登录过的厂商 CLI,所以代码只会到达
+你本来就在用的那几家。执行器在调用订阅制 CLI 前会剥掉 API 密钥环境变量,避免
+一把残留的密钥把你悄悄切到按 token 计费。唯一的例外是 direct-API vendor 家族
+(`openrouter`、`deepseek`、`zai`、`mistral`、`groq`、`cerebras`),它们本来就是
+拿你配置的密钥直调该提供商的 API——这几家只做 advise,不会改文件。
+
+</details>
+
+<details>
+<summary><b>Claude Fable 5 去哪了?为什么默认表不放它?</b></summary>
+
+<br/>
+
+**因为 Claude 顶级档通常就是主循环本人,不是被派发的工人。** 通道存在的意义,
+是把工作送给「你正在开的那个模型以外」的模型。如果 Fable 5 就是你的主循环,把
+判断和文字再路由回 Fable 5 只是多一次调用、毫无增益——所以上面那份「选你的主控
+模型」清单里,Fable 5 有自己独立的一行,身份是**主控**:自己做 hard-judgment、
+taste-final、最吃正确性的硬修。
+
+**测量数据也不支持把它当工人。** Artificial Analysis 智能指数(2026-07-24)
+Opus 5(max)61 分、Fable 5(max)60 分——AA 自己的用词是「实质打平」,而 Epoch AI
+的能力指数排序还是反过来的(Fable 5 161、Opus 5 159)。综合智力就当它平手。真正
+拉开差距的是 agentic 专业产出,而且差很多:
+
+| 评测 | Claude Opus 5 (max) | Claude Fable 5 | |
+|---|---:|---:|---|
+| AA-Briefcase(agentic 知识工作,Elo) | 1720 | 1574 | **+146** |
+| GDPval-AA v2(Elo) | 1861 | 1747 | **+114** |
+| AA-Briefcase 每任务成本 | $17.79 | $22.30 | **-20%** |
+| API 单价,输入/输出 每 1M | $5 / $25 | $10 / $50 | **一半** |
+
+Opus 5 的 max、xhigh、high 三个档位包揽 AA-Briefcase 前三名,连 `high` 档都在
+不到一半的每任务成本下赢过 Fable 5。也就是说,Fable 5 贵一倍,却换不到任何一条
+通道所在意的优势。
+
+**Fable 5 真正更强的地方**:事实广度。它在 AA-Omniscience 上仍领先 Opus 5
+(符合两者的规模差),而 Opus 5 在没把握时更倾向直接作答——幻觉率 50%,比
+Opus 4.8 高 14 个百分点。任务偏「回想」而非「执行」时,直接点名它:
+
+```bash
+dispatch.sh --vendor claude --model claude-fable-5 --effort high consult "…"
+```
+
+**这是成本与主循环策略的选择,不是能力判决。** 设置菜单的模型清单里有 Fable 5,
+在 `routing.local.yaml` 写一行就能覆盖默认:
+
+```yaml
+taste-final: claude claude-fable-5 high
+```
+
+</details>
+
+<details>
+<summary><b>Claude 那几条通道为什么用 <code>xhigh</code> 而不是 <code>max</code>?</b></summary>
+
+<br/>
+
+因为推理档位不是越高越好。Anthropic 官方把 `xhigh` 定为编码与 agentic 工作的
+起手档位,`high` 是其他吃智力任务的下限,`max` 保留给「正确性重于成本」的场合。
+第三方实测也一致:Vals.ai 的 Vibe Code Bench 上,Opus 5 在 `high` 拿 89.8%,
+`xhigh` 只有 88.3%、`max` 88.4%——最高档倾向产出更繁复的解,反而更常出错。
+你的工作类型如果不同意,单条通道自己拉高:
+
+```bash
+omnilane configure set hard-judgment "claude claude-opus-5 max"
+```
+
+</details>
+
+<details>
+<summary><b>通道首选的 CLI 没装会怎样?</b></summary>
+
+<br/>
+
+派工会沿着候选链往下走,用你手上有的第一家。不花任何额度就能先看决策:
+
+```bash
+scripts/dispatch.sh --explain hardest-coding   # 逐候选解释
+scripts/dispatch.sh --list                     # 整张生效表
+scripts/dispatch.sh --dry-run hardest-coding "…"   # 完整解析后的计划,不调用模型
+```
+
+</details>
+
+<details>
+<summary><b>被派工的模型会不会乱改我的文件?</b></summary>
+
+<br/>
+
+除非你明说要它改。派工默认是 `advise` 只读模式,而且是逐厂商实现的(只读沙箱、
+plan 模式,或只给只读工具集)。要改文件必须同时给 `--mode work` 和明确的
+`--workdir`。工作端也不能再往外派——深度守卫会用退出码 86 拒绝嵌套派工,一道
+命令不可能失控变成一整串 AI 烧你的额度。
+
+</details>
+
 ## 📊 默认值与数据来源
 
 默认通道配置依据 Artificial Analysis 2026-07 快照(已对 AA 站上原始记录与
@@ -350,7 +447,8 @@ configure.sh set|get|unset|list|diff LANE [SPEC]    # 非交互编辑/查看 rou
 设置菜单和 `routing.local.yaml` 就是让你不同意用的。评审团(arbitrate)
 默认关闭;要用就在 `routing.local.yaml` 写
 `arbitrate: vote codex,claude,grok -`(从四家里任选 1-4 个评审),
-或改用 `exec` 厂商指向你自己的多模型审查闸脚本。
+或改用 `exec` 厂商指向你自己的多模型审查闸脚本。完整工作笔记(含各评测的
+但书)见 [`docs/model-capabilities-2026-07.md`](docs/model-capabilities-2026-07.md)。
 
 ## ⚠️ 已知限制
 
@@ -364,8 +462,58 @@ configure.sh set|get|unset|list|diff LANE [SPEC]    # 非交互编辑/查看 rou
 
 ## 📜 版本历程
 
+## v0.10.3 新功能
+
+- **五种语言的 README 全面重整**——文档开头改成先讲清楚「这是什么、我为什么会
+  想要它」,版本历程全部收拢到最下方,不再打断开头的介绍;新增常见问题,回答
+  一直被问到的几件事:是不是每家订阅都要有、代码会被送去哪、为什么默认表没有
+  Fable 5、为什么用 `xhigh` 而不是 `max`、首选 CLI 没装会怎样、被派工的模型会不会
+  改文件。
+- **修复:插件清单的版本号没跟上**——`plugin.json` 与
+  `.claude-plugin/plugin.json` 在 0.10.1、0.10.2 发布后仍写着 `0.10.0`,导致插件
+  安装显示错误版本。
+- **修复:`routing.local.yaml.example` 还指着已退场的模型**——起手配置里的
+  `claude-opus-4-8` 全数改为 `claude-opus-5`(并按通道给对应档位),Gemini 3.5
+  Flash 候选改为 3.6 Flash,与 0.10.0 以来的默认值一致。
+- **对照原始资料修正智能指数数字**(`docs/model-capabilities-2026-07.md`):
+  那是指数点数不是百分比;补上 AA-Briefcase / GDPval-AA v2 对照,并记下两项与
+  默认值相反的结果:Fable 5 在事实知识领先、GPT-5.6 Sol 在呈现质量领先。
+
+## v0.10.2 新功能
+
+- **`hardest-coding` 与 `hard-judgment` 的 Claude 档位由 `max` 降为 `xhigh`**,
+  对齐 Anthropic 对 Claude Opus 5 的官方建议:编码与 agentic 工作从 `xhigh` 起跳,
+  `high` 是其他吃智力任务的下限,`max` 保留给正确性重于成本的场合。要拉回去用
+  `omnilane configure set <通道> "<配置>"`。
+- **修掉两条死的 CHANGELOG 比较链接**——它们指向从未发布的 `v0.10.0` tag。
+
+## v0.10.1 新功能
+
+- **默认路由加入 `claude-opus-5`**:成为 `hard-judgment`、`taste-final` 第一顺位,
+  也纳入最高难度编程任务的备选。
+- **`omnilane configure` 已扩展全部 13 个提供商**:共 106 个可选模型,完整收录
+  Codex、Claude Code、Grok Build、Antigravity 实时列表,并加入已验证的
+  OpenRouter／OpenCode 快捷项;仍可用 `c` 输入自定义模型 ID。
+
 <details>
-<summary>旧版本(v0.8.3 以前)</summary>
+<summary>旧版本(v0.10.0 以前)</summary>
+
+## v0.10.0 新功能
+
+- **Gemini 3.6 Flash 默认路由**——`fast-agentic`、`triage`、`bulk-mechanical`
+  的 gemini 候选(与 `Gemini Flash` 别名)改用 Gemini 3.6 Flash:输出 token
+  更少、输出单价更低、Artificial Analysis 实测输出速度第一。
+- **证据重审计**——路由注释、模型能力笔记与 Gemini 价格表对官方来源刷新。
+
+## v0.9.1 新功能
+
+- **修复**:`configure set` 不再删除 `routing.local.yaml` 中手写的注释——
+  只改写自身的戳记行与被替换的 lane。
+
+## v0.9.0 新功能
+
+- **新增 5 个 OpenAI-compatible direct-API vendor** — `deepseek`、`zai`(GLM)、`mistral`、`groq`、`cerebras`,与 `openrouter` 同为免 CLI 通道(curl 加一把 `<VENDOR>_API_KEY`);`lib/common.sh` registry 一行即加一个。详见 [`docs/model-capabilities-2026-07.md`](docs/model-capabilities-2026-07.md)。
+- **fish shell 补全** — `omnilane completion fish | source`。
 
 ## v0.8.3 新功能