omnilane 0.31.0 → 0.32.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/README.md CHANGED
@@ -78,11 +78,11 @@ key with no extra CLI at all):
78
78
  ```mermaid
79
79
  flowchart LR
80
80
  M["main loop<br/><i>any CLI you drive</i>"] --> T{{"routing.yaml<br/>one shared table"}}
81
- T -->|hardest-coding| C1["CodexGPT-5.6 Sol"]
82
- T -->|bulk-mechanical| C2["Codex — GPT-5.6 Terra"]
83
- T -->|taste-final| C3["Claude — Opus 5"]
84
- T -->|long-context| C4["Gemini — 3.1 Pro"]
85
- T -->|live-search| C5["Grok — 4.5"]
81
+ T -->|hardest-coding| C1["ClaudeFable 5.1"]
82
+ T -->|bulk-mechanical| C2["Codex — GPT-5.6 Sol"]
83
+ T -->|taste-final| C3["Claude — Fable 5.1"]
84
+ T -->|long-context| C4["Gemini — 3.7 Flash"]
85
+ T -->|live-search| C5["Grok — 4.6"]
86
86
  T -->|"arbitrate (opt-in)"| C6["vote — 1-4 model panel"]
87
87
  ```
88
88
 
@@ -115,23 +115,25 @@ actually resolves.
115
115
 
116
116
  | Lane | First choice | Backup | When |
117
117
  |---|---|---|---|
118
- | 🔥 hardest-coding | GPT-5.6 Sol (xhigh) | Claude Opus 5 (xhigh) | Hardest implementation, deep root-cause debug, correctness-critical edits |
119
- | 🏗️ bulk-mechanical | GPT-5.6 Terra (max) | Claude Sonnet 5 (high) | Refactors, migrations, tests, review sweeps — mechanical endurance |
120
- | 🧹 triage | GPT-5.6 Luna (medium) | Gemini 3.6 Flash (Low) | High-volume scans, first-pass filtering |
121
- | ⚖️ hard-judgment | Claude Opus 5 (xhigh) | GPT-5.6 Sol (max) | Architecture arbitration, deep reasoning, second opinions |
122
- | ✒️ taste-final | Claude Opus 5 (high) | GPT-5.6 Sol (max) | User-facing prose, prompt/doc polish, style arbitration |
123
- | 💬 consult | Explicit named vendor/model | (no fallback) | Direct natural-language consultation; always keep `--vendor` |
124
- | 🎨 ui-draft | GPT-5.6 Sol (xhigh) | Claude Opus 5 (high) | UI drafts only WITH a design system / reference images |
125
- | 📚 long-context | Gemini 3.1 Pro (High) | GPT-5.6 Sol (high) | 1M-token sweeps, retrieval, and synthesis across long documents; Flash for fast repeated loops |
126
- | ⚡ fast-agentic | GPT-5.6 Luna (max) | Gemini 3.6 Flash (High) | Fast multi-step agentic loops, multimodal checks |
127
- | 📡 live-search | Grok 4.5 | — (off) | Realtime X/web search and social context |
128
- | 🚰 coding-overflow | Grok 4.5 | Kimi K3 → Qwen3 Coder Plus → OpenCode | Codex-quota relief valve for mid-tier coding |
118
+ | 🔥 hardest-coding | Claude Fable 5.1 (xhigh) | GPT-5.6 Sol (xhigh) | Hardest implementation, deep root-cause debug, correctness-critical edits |
119
+ | 🏗️ bulk-mechanical | GPT-5.6 Sol (high) | Gemini 3.7 Flash (High) → Claude Sonnet 5 (high) | Refactors, migrations, tests, review sweeps — mechanical endurance |
120
+ | 🧹 triage | GPT-5.6 Luna (high) | Gemini 3.7 Flash (Low) → Claude Haiku 4.5 | High-volume scans, first-pass filtering |
121
+ | ⚖️ hard-judgment | Claude Fable 5.1 (xhigh) | GPT-5.6 Sol (max) → Grok 4.6 | Architecture arbitration, deep reasoning, second opinions |
122
+ | ✒️ taste-final | Claude Fable 5.1 (high) | GPT-5.6 Sol (max) | User-facing prose, prompt/doc polish, style arbitration |
123
+ | 💬 consult | GPT-5.6 Sol (max) | Claude Fable 5.1 (high) → Grok 4.6 → Gemini 3.7 Flash (High) | Direct named-model consultation; keep `--vendor` to prevent fallback |
124
+ | 🎨 ui-draft | GPT-5.6 Sol (xhigh) | Claude Fable 5.1 (high) | UI drafts only WITH a design system / reference images |
125
+ | 📚 long-context | Gemini 3.7 Flash (Medium) | GPT-5.6 Terra (max) → Claude Opus 5 (medium) | Long-document retrieval and synthesis, ordered on AA-LCR, cost, and throughput |
126
+ | ⚡ fast-agentic | Gemini 3.7 Flash (Medium) | GPT-5.6 Luna (high) | Fast multi-step agentic loops, multimodal checks |
127
+ | 📡 live-search | Grok 4.6 | — (off) | Realtime X/web search and social context |
128
+ | 🚰 coding-overflow | Grok 4.6 | Gemini 3.7 Flash (High) → Kimi K3 → Qwen3 Coder Plus → OpenCode | Codex-quota relief valve for mid-tier coding |
129
129
  | 🗳️ arbitrate | off (opt-in vote panel) | — | Built-in opinion panel for big calls — disabled by default; enable it in `routing.local.yaml`, one call per voter per round |
130
130
 
131
131
  The **backup** is the next candidate in the lane's `routing.yaml` chain — what
132
132
  dispatch falls back to when the first-choice vendor CLI is not installed. Every
133
133
  lane is such a chain; when nothing in it is installed the lane degrades to `off`.
134
134
 
135
+ > **Fable 5.1 is in the defaults — and where Opus 5 still fits.** See the current three-way evidence and Opus override in the [FAQ](#-faq).
136
+
135
137
  ### Natural-language consultation
136
138
 
137
139
  With the `omnilane` skill or `/route`, you can ask normally:
@@ -155,12 +157,12 @@ with who is driving. What changes is which lanes you **self-execute** (you
155
157
  already are that model, so no second call) versus **dispatch**. Your harness's
156
158
  `omnilane` skill applies the right row automatically; this is the human view.
157
159
 
158
- - **Claude Code · Fable 5** — self-execute: hard-judgment, taste-final, the hardest correctness-critical fixes. Dispatch mechanical coding volume → Codex, long-context → Gemini, live-search → Grok.
159
- - **Claude Code · Opus 5** — self-execute: hard-judgment and taste-final. Dispatch bulk coding to Codex lanes, long-context → Gemini, live-search → Grok.
160
- - **Codex · Sol** — self-execute: hardest-coding, hard-judgment, ui-draft. Dispatch taste-final → Claude, long-context → Gemini, live-search → Grok, bulk → Codex Terra.
161
- - **Codex · Terra** — self-execute: bulk-mechanical. Escalate the genuinely hardest pieces to Sol; dispatch taste → Claude, long-context → Gemini, live-search → Grok.
162
- - **Grok Build · Grok 4.5** — self-execute: live-search, coding-overflow (mid-tier coding). Dispatch everything hard to Codex/Claude/Gemini — and verify every API signature and cited fact first.
163
- - **Antigravity · Gemini** — self-execute: long-context and context-heavy agentic work on 3.1 Pro, fast repeated loops on Flash. Dispatch hardest coding/judgment/taste to Codex/Claude; live-search → Grok.
160
+ - **Claude Code · Fable 5.1** — self-execute: hard-judgment, taste-final, hardest-coding. Dispatch bulk → Codex Sol high; long-context and fast loops → Gemini 3.7 Flash; live-search → Grok.
161
+ - **Claude Code · Opus 5** — self-execute: hard-judgment and taste-final when its lower hallucination rate or price is preferred. Dispatch hardest coding Fable 5.1 or Sol, bulk → Sol high, long-context and fast loops → Gemini 3.7 Flash, live-search → Grok.
162
+ - **Codex · Sol** — self-execute: hardest-coding, bulk-mechanical, hard-judgment, ui-draft. Dispatch taste-final → Claude, long-context and fast loops → Gemini 3.7 Flash, live-search → Grok.
163
+ - **Codex · Terra** — self-execute: long-context as the Codex fallback. Bulk-mechanical now defaults to Sol high; escalate hardest pieces to Sol xhigh, taste → Claude, fast loops → Gemini 3.7 Flash, live-search → Grok.
164
+ - **Grok Build · Grok 4.6** — self-execute: live-search and coding-overflow. Dispatch hard coding/judgment/taste to Codex/Claude/Gemini; verify API signatures and cited facts.
165
+ - **Antigravity · Gemini 3.7 Flash** — self-execute: long-context and fast loops at Medium, bulk/overflow at High, triage at Low. Dispatch hardest coding/judgment/taste to Codex/Claude; live-search → Grok.
164
166
 
165
167
  </details>
166
168
 
@@ -489,52 +491,30 @@ with the key you set — those are advise-only and never edit files.
489
491
  </details>
490
492
 
491
493
  <details>
492
- <summary><b>Where is Claude Fable 5? Why is it not in the default table?</b></summary>
494
+ <summary><b>Fable 5.1 is in the defaults and where Opus 5 still fits</b></summary>
493
495
 
494
496
  <br/>
495
497
 
496
- **Because the top Claude tier is usually the main loop itself, not a dispatched
497
- worker.** Lanes exist to send work to a model *other than* the one you are
498
- driving. If Fable 5 is your main loop, routing judgment and taste back to Fable 5
499
- just adds a second call for no gain — which is why the "pick your main model"
500
- list above gives Fable 5 its own row as a **driver**, self-executing
501
- hard-judgment, taste-final, and the hardest correctness-critical fixes.
502
-
503
- **The measurements do not argue for it as a worker either.** On the Artificial
504
- Analysis Intelligence Index (2026-07-24) Opus 5 (max) scores 61 and Fable 5 (max)
505
- scores 60 — Artificial Analysis calls them "effectively tied", and Epoch AI's
506
- Capability Index ranks them the other way (Fable 5 161, Opus 5 159). Call it a
507
- draw on general intelligence. Where they are not tied is agentic professional
508
- output, and Opus 5 leads by a wide margin:
509
-
510
- | Benchmark | Claude Opus 5 (max) | Claude Fable 5 | |
511
- |---|---:|---:|---|
512
- | AA-Briefcase (agentic knowledge work, Elo) | 1720 | 1574 | **+146** |
513
- | GDPval-AA v2 (Elo) | 1861 | 1747 | **+114** |
514
- | Cost per AA-Briefcase task | $17.79 | $22.30 | **-20%** |
515
- | API price, input / output per 1M | $5 / $25 | $10 / $50 | **half** |
516
-
517
- Opus 5's max, xhigh and high tiers sweep the top three AA-Briefcase places, and
518
- its `high` tier still beats Fable 5 at under half the cost per task. So Fable 5
519
- costs twice as much without buying an advantage on any axis a lane is defined
520
- around.
521
-
522
- **What Fable 5 is genuinely better at**: factual breadth. It stays ahead of
523
- Opus 5 on AA-Omniscience, as its size class suggests, and Opus 5 answers more
524
- readily when uncertain — its hallucination rate is 50%, up 14 points from
525
- Opus 4.8. If your task is recall-heavy rather than execution-heavy, name
526
- Fable 5 explicitly:
498
+ Fable 5.1 now leads `hardest-coding`, `hard-judgment`, and `taste-final`.
499
+ At matched xhigh effort it leads Opus 5 on intelligence, agentic work, and
500
+ coding. Sol max remains the far cheaper cross-vendor judgment fallback.
527
501
 
528
- ```bash
529
- dispatch.sh --vendor claude --model claude-fable-5 --effort high consult "…"
530
- ```
502
+ | Benchmark (AA, retrieved 2026-09-02) | Claude Fable 5.1 (xhigh) | Claude Opus 5 (xhigh) | GPT-5.6 Sol (max) |
503
+ |---|---:|---:|---:|
504
+ | Intelligence | 64.8 | 62.5 | 60.9 |
505
+ | Agentic | 59.8 | 58.4 | 57.8 |
506
+ | Coding | 80.7 | 77.0 | 77.4 |
507
+ | Hallucination rate (lower is better) | .71 | **.60** | .92 |
508
+ | AA $/task | $2.65 | $1.80 | **$0.95** |
531
509
 
532
- **This is a cost / main-loop policy choice, not a capability verdict.** Fable 5
533
- is in the configurator's model menu, and one line in `routing.local.yaml`
534
- overrides the default if you disagree:
510
+ Fable 5.1 is not a bulk or triage default: it costs twice Opus 5 per token and
511
+ consumes the most Claude Code subscription quota per turn. Opus 5 remains the
512
+ lower-hallucination, lower-price Claude option, stays in `long-context` at
513
+ medium, and remains selectable everywhere through
514
+ `~/.omnilane/routing.local.yaml`:
535
515
 
536
516
  ```yaml
537
- taste-final: claude claude-fable-5 high
517
+ hard-judgment: claude claude-opus-5 xhigh
538
518
  ```
539
519
 
540
520
  </details>
@@ -598,7 +578,7 @@ Default lane assignments follow Artificial Analysis coding/intelligence data
598
578
  pages) plus published head-to-head reviews; they are opinions, not laws — the
599
579
  configurator and `routing.local.yaml` exist so you can disagree. The full
600
580
  working notes, including per-benchmark caveats, live in
601
- [`docs/model-capabilities-2026-07.md`](docs/model-capabilities-2026-07.md).
581
+ [`docs/model-capabilities-2026-09.md`](docs/model-capabilities-2026-09.md).
602
582
 
603
583
  ## ⚠️ Known limitations
604
584
 
@@ -614,6 +594,12 @@ working notes, including per-benchmark caveats, live in
614
594
 
615
595
  ## 📜 Release history
616
596
 
597
+ ## What's new in v0.32.0
598
+
599
+ - **Full 2026-09 routing re-evaluation.** Fable 5.1 and Gemini 3.7 Flash enter the defaults, backed by the dated Artificial Analysis snapshot.
600
+ - **Catalogs match live CLI surfaces.** Fable 5.1 is added and retired Gemini 3.5 Flash rows are removed.
601
+ - **Opus 5 remains available.** It stays in `long-context` and can override any lane through `routing.local.yaml`.
602
+
617
603
  ## What's new in v0.31.0
618
604
 
619
605
  - **Unlimited goal budgets by default.** `budget_jobs` and `budget_seconds` now persist as JSON `null` and render as `unlimited`; the previous implicit 8-job and 900-second caps are gone. Use `--budget-jobs N` or `--budget-seconds S` to opt in to a hard cap. The repeat-failure fuse remains enabled by default.
@@ -715,7 +701,7 @@ working notes, including per-benchmark caveats, live in
715
701
  latency-bound.
716
702
  - **Lane comments no longer carry numbers.** `routing.yaml` now states why each
717
703
  ordering holds; every score, price and throughput figure lives in
718
- `docs/model-capabilities-2026-07.md` with its retrieval date, so a stale
704
+ `docs/model-capabilities-2026-09.md` with its retrieval date, so a stale
719
705
  figure never requires a routing-table edit.
720
706
  - **A value profile** in `routing.local.yaml.example` trades about one
721
707
  Intelligence Index point for 30-40% off the cost per task.
@@ -774,7 +760,7 @@ working notes, including per-benchmark caveats, live in
774
760
  profiles still pointed at `claude-opus-4-8` and Gemini 3.5 Flash; they now use
775
761
  Claude Opus 5 (with lane-appropriate effort) and Gemini 3.6 Flash.
776
762
  - **Corrected the Intelligence Index figures** in
777
- `docs/model-capabilities-2026-07.md` against the Artificial Analysis source
763
+ `docs/model-capabilities-2026-09.md` against the Artificial Analysis source
778
764
  (index points, not percentages), added the AA-Briefcase / GDPval-AA v2
779
765
  comparison, and recorded the two results that cut against the defaults:
780
766
  Fable 5's lead on factual knowledge and Sol's lead on presentation quality.
@@ -822,7 +808,7 @@ working notes, including per-benchmark caveats, live in
822
808
  - **Five OpenAI-compatible direct-API vendors** — `deepseek`, `zai` (GLM),
823
809
  `mistral`, `groq`, and `cerebras` join `openrouter` as CLI-free lanes (curl +
824
810
  a `<VENDOR>_API_KEY`). A one-line `lib/common.sh` registry entry adds each;
825
- see [`docs/model-capabilities-2026-07.md`](docs/model-capabilities-2026-07.md).
811
+ see [`docs/model-capabilities-2026-09.md`](docs/model-capabilities-2026-09.md).
826
812
  - **Fish shell completion** — `omnilane completion fish | source`.
827
813
 
828
814
  ## What's new in v0.8.3
@@ -846,7 +832,7 @@ working notes, including per-benchmark caveats, live in
846
832
  DeepSeek, Z.ai GLM, Mistral, Groq, and Cerebras. Each needs only `curl` and
847
833
  its `<VENDOR>_API_KEY`; advise/consult only. A one-line `lib/common.sh`
848
834
  registry entry defines each endpoint, key env, and default model. See
849
- [`docs/model-capabilities-2026-07.md`](docs/model-capabilities-2026-07.md).
835
+ [`docs/model-capabilities-2026-09.md`](docs/model-capabilities-2026-09.md).
850
836
  - **`opencode` vendor** — headless dispatch through the OpenCode
851
837
  multi-provider aggregator CLI (`opencode run`). Advise mode pins OpenCode's
852
838
  built-in read-only `plan` agent; work mode uses `--auto`. Joins the default
package/README.zh-CN.md CHANGED
@@ -72,11 +72,11 @@ omnilane 让**任何**一个 agentic CLI 的主循环把子任务分类到通道
72
72
  ```mermaid
73
73
  flowchart LR
74
74
  M["主循环<br/><i>你在用的任一 CLI</i>"] --> T{{"routing.yaml<br/>一张共用路由表"}}
75
- T -->|hardest-coding| C1["CodexGPT-5.6 Sol"]
76
- T -->|bulk-mechanical| C2["Codex — GPT-5.6 Terra"]
77
- T -->|taste-final| C3["Claude — Opus 5"]
78
- T -->|long-context| C4["Gemini — 3.1 Pro"]
79
- T -->|live-search| C5["Grok — 4.5"]
75
+ T -->|hardest-coding| C1["ClaudeFable 5.1"]
76
+ T -->|bulk-mechanical| C2["Codex — GPT-5.6 Sol"]
77
+ T -->|taste-final| C3["Claude — Fable 5.1"]
78
+ T -->|long-context| C4["Gemini — 3.7 Flash"]
79
+ T -->|live-search| C5["Grok — 4.6"]
80
80
  T -->|"arbitrate(可选)"| C6["vote — 1-4 模型评审团"]
81
81
  ```
82
82
 
@@ -103,24 +103,23 @@ flowchart LR
103
103
 
104
104
  | 通道 | 首选模型 | 备选模型 | 用途 |
105
105
  |---|---|---|---|
106
- | 🔥 hardest-coding | GPT-5.6 Sol (xhigh) | Claude Opus 5 (xhigh) | 最难的实现、深度调试、正确性攸关的修改 |
107
- | 🏗️ bulk-mechanical | GPT-5.6 Terra (max) | Claude Sonnet 5 (high) | 重构、迁移、测试、大面积扫描——机械耐力活 |
108
- | 🧹 triage | GPT-5.6 Luna (medium) | Gemini 3.6 Flash (Low) | 高量初筛、第一轮过滤 |
109
- | ⚖️ hard-judgment | Claude Opus 5 (xhigh) | GPT-5.6 Sol (max) | 架构仲裁、深度推理、第二意见 |
110
- | ✒️ taste-final | Claude Opus 5 (high) | GPT-5.6 Sol (max) | 对外文字、prompt 与文档打磨、风格终审 |
111
- | 💬 consult | 明确指定的厂商/模型 | (不降级) | 自然语言直接咨询;必须保留 `--vendor` |
112
- | 🎨 ui-draft | GPT-5.6 Sol (xhigh) | Claude Opus 5 (high) | 有设计规范/参考图时的 UI 出稿;开放式视觉品味交给 taste-final |
113
- | 📚 long-context | Gemini 3.1 Pro (High) | GPT-5.6 Sol (high) | 百万 token 扫读、检索与长文跨段整合;高速重复循环仍优先 Flash |
114
- | ⚡ fast-agentic | GPT-5.6 Luna (max) | Gemini 3.6 Flash (High) | 快速多步骤 agentic 循环、多模态检查 |
115
- | 📡 live-search | Grok 4.5 | —(off) | 实时 X/网络搜索与社群脉络 |
116
- | 🚰 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 |
117
- | 🗳️ 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 |
106
+ | 🔥 hardest-coding | Claude Fable 5.1 (xhigh) | GPT-5.6 Sol (xhigh) | 最难的实现、深度调试、正确性关键的修改 |
107
+ | 🏗️ bulk-mechanical | GPT-5.6 Sol (high) | Gemini 3.7 Flash (High) → Claude Sonnet 5 (high) | 重构、迁移、测试、大范围扫描——机械耐力活 |
108
+ | 🧹 triage | GPT-5.6 Luna (high) | Gemini 3.7 Flash (Low) → Claude Haiku 4.5 | 大量扫描、第一轮筛选 |
109
+ | ⚖️ hard-judgment | Claude Fable 5.1 (xhigh) | GPT-5.6 Sol (max) → Grok 4.6 | 架构裁决、深度推理、第二意见 |
110
+ | ✒️ taste-final | Claude Fable 5.1 (high) | GPT-5.6 Sol (max) | 对外文字、提示词/文档润色、风格裁决 |
111
+ | 💬 consult | GPT-5.6 Sol (max) | Claude Fable 5.1 (high) → Grok 4.6 → Gemini 3.7 Flash (High) | 直接指定模型咨询;保留 `--vendor` 避免降级 |
112
+ | 🎨 ui-draft | GPT-5.6 Sol (xhigh) | Claude Fable 5.1 (high) | 仅在提供设计系统/参考图时生成 UI 草稿 |
113
+ | 📚 long-context | Gemini 3.7 Flash (Medium) | GPT-5.6 Terra (max) Claude Opus 5 (medium) | 长文档提取与综合,按 AA-LCR、成本和吞吐排序 |
114
+ | ⚡ fast-agentic | Gemini 3.7 Flash (Medium) | GPT-5.6 Luna (high) | 高速多步骤工具循环、多模态检查 |
115
+ | 📡 live-search | Grok 4.6 | — (`off`) | 实时 X/网页搜索与社交上下文 |
116
+ | 🚰 coding-overflow | Grok 4.6 | Gemini 3.7 Flash (High) → Kimi K3 → Qwen3 Coder Plus → OpenCode | Codex 配额耗尽时的中量级编码安全阀 |
117
+ | 🗳️ arbitrate | `off`(可选模型评审团) | — | 重大决定的内置意见评审团;默认禁用,在 `routing.local.yaml` 启用,每位评审每轮调用一次 |
118
118
 
119
119
  **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
120
120
  通道都是这样一条链;整条都没装时,通道自动降为 `off`。
121
121
 
122
- > **Claude Fable 5 去哪了?** 默认表刻意不放——理由与完整数据见
123
- > [常见问题](#-常见问题)。
122
+ > **Fable 5.1 已进入默认表——以及 Opus 5 仍适合放在哪里。** 三方数据与 Opus override 见[常见问题](#-常见问题)。
124
123
 
125
124
  ### 自然语言咨询
126
125
 
@@ -141,12 +140,12 @@ flowchart LR
141
140
  你哪些通道**自己做**(你本来就是那个模型,省一次调用)、哪些**派出去**。你 CLI 里
142
141
  的 `omnilane` 技能会自动套对的那一行,这里是给人看的版本。
143
142
 
144
- - **Claude Code · Fable 5** — 自己做:hard-judgment、taste-final、最吃正确性的硬修。派出去:机械编码量 → Codex、长文 → Gemini、实时搜索 → Grok。
145
- - **Claude Code · Opus 5** — 自己做:hard-judgment、taste-final。大量编码走 Codex 通道、长文 → Gemini、实时搜索 → Grok。
146
- - **Codex · Sol** — 自己做:hardest-coding、hard-judgment、ui-draft。派出去:taste-final → Claude、长文 → Gemini、实时搜索 Grok、粗活Codex Terra
147
- - **Codex · Terra** 自己做:bulk-mechanical。真正最硬的往上升给 Sol;taste → Claude、长文 → Gemini、实时搜索 → Grok。
148
- - **Grok Build · Grok 4.5** — 自己做:live-search、coding-overflow(中量级编码)。所有硬活派给 Codex/Claude/Gemini——先验每个 API 签名与引用事实。
149
- - **Antigravity · Gemini** — 自己做:3.1 Pro 的长文与重上下文 agentic 工作、Flash 的高速重复循环。最难的编码/判断/文字派给 Codex/Claude;实时搜索 → Grok。
143
+ - **Claude Code · Fable 5.1**——自己执行:hard-judgment、taste-final、hardest-coding。派发:bulk → Codex Sol high;long-context/高速循环 → Gemini 3.7 Flash;实时搜索 → Grok。
144
+ - **Claude Code · Opus 5**——需要更低幻觉率或价格时,自己执行 hard-judgment、taste-final。最难编码 Fable 5.1 或 Sol;bulk Sol high;long-context/高速循环 → Gemini 3.7 Flash;实时搜索 → Grok。
145
+ - **Codex · Sol**——自己执行:hardest-coding、bulk-mechanical、hard-judgment、ui-draft。派发:taste-final → Claude;long-context/高速循环 → Gemini 3.7 Flash;实时搜索Grok
146
+ - **Codex · Terra**——自己执行 long-context 的 Codex 备用任务;bulk-mechanical 现在默认由 Sol high 处理。最难部分升级到 Sol xhigh,taste → Claude,高速循环 → Gemini 3.7 Flash,实时搜索 → Grok。
147
+ - **Grok Build · Grok 4.6**——自己执行 live-search、coding-overflow。最难的编码/判断/文字交给 CodexClaudeGemini;仍需验证 API 签名和引用事实。
148
+ - **Antigravity · Gemini 3.7 Flash**——自己执行:Medium long-context/高速循环、High bulk/overflow、Low 的 triage。最难编码/判断/文字交给 CodexClaude;实时搜索 → Grok。
150
149
 
151
150
  </details>
152
151
 
@@ -395,45 +394,29 @@ omnilane goal close "$GOAL_ID" --summary "结账集成已稳定"
395
394
  </details>
396
395
 
397
396
  <details>
398
- <summary><b>Claude Fable 5 去哪了?为什么默认表不放它?</b></summary>
397
+ <summary><b>Fable 5.1 已进入默认表——以及 Opus 5 仍适合放在哪里</b></summary>
399
398
 
400
399
  <br/>
401
400
 
402
- **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,
403
- 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
404
- 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
405
- 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、
406
- taste-final、最吃正确性的硬修。
401
+ Fable 5.1 现在领跑 `hardest-coding`、`hard-judgment`、`taste-final`。
402
+ 同为 xhigh 时,它在智能、代理式工作和编码上都领先 Opus 5;Sol max 则保留为
403
+ 便宜得多的跨厂商判断备用项。
407
404
 
408
- **测量数据也不支持把它当工人。** Artificial Analysis 智能指数(2026-07-24)
409
- Opus 5(max)61 分、Fable 5(max)60 分——AA 自己的用词是「实质打平」,而 Epoch AI
410
- 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
411
- 拉开差距的是 agentic 专业产出,而且差很多:
405
+ | 评测(AA,抓取于 2026-09-02) | Claude Fable 5.1 (xhigh) | Claude Opus 5 (xhigh) | GPT-5.6 Sol (max) |
406
+ |---|---:|---:|---:|
407
+ | 智能 | 64.8 | 62.5 | 60.9 |
408
+ | 代理式 | 59.8 | 58.4 | 57.8 |
409
+ | 编码 | 80.7 | 77.0 | 77.4 |
410
+ | 幻觉率(越低越好) | .71 | **.60** | .92 |
411
+ | AA 每任务成本 | $2.65 | $1.80 | **$0.95** |
412
412
 
413
- | 评测 | Claude Opus 5 (max) | Claude Fable 5 | |
414
- |---|---:|---:|---|
415
- | AA-Briefcase(agentic 知识工作,Elo) | 1720 | 1574 | **+146** |
416
- | GDPval-AA v2(Elo) | 1861 | 1747 | **+114** |
417
- | AA-Briefcase 每任务成本 | $17.79 | $22.30 | **-20%** |
418
- | API 单价,输入/输出 每 1M | $5 / $25 | $10 / $50 | **一半** |
419
-
420
- Opus 5 的 max、xhigh、high 三个档位包揽 AA-Briefcase 前三名,连 `high` 档都在
421
- 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
422
- 通道所在意的优势。
423
-
424
- **Fable 5 真正更强的地方**:事实广度。它在 AA-Omniscience 上仍领先 Opus 5
425
- (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
426
- Opus 4.8 高 14 个百分点。任务偏「回想」而非「执行」时,直接点名它:
427
-
428
- ```bash
429
- dispatch.sh --vendor claude --model claude-fable-5 --effort high consult "…"
430
- ```
431
-
432
- **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,
433
- 在 `routing.local.yaml` 写一行就能覆盖默认:
413
+ Fable 5.1 没进入 bulk triage:每 token 价格是 Opus 5 的两倍,而且每轮
414
+ 消耗最多 Claude Code 订阅配额。Opus 5 仍是幻觉率更低、价格更低的 Claude
415
+ 选项,并以 medium 保留在 `long-context`;也能通过
416
+ `~/.omnilane/routing.local.yaml` 放回任意通道:
434
417
 
435
418
  ```yaml
436
- taste-final: claude claude-fable-5 high
419
+ hard-judgment: claude claude-opus-5 xhigh
437
420
  ```
438
421
 
439
422
  </details>
@@ -493,7 +476,7 @@ vendor 一律当成 `work`,而且它只能逐次明确指定,永远不是 lane
493
476
  默认关闭;要用就在 `routing.local.yaml` 写
494
477
  `arbitrate: vote codex,claude,grok -`(从四家里任选 1-4 个评审),
495
478
  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
496
- 但书)见 [`docs/model-capabilities-2026-07.md`](docs/model-capabilities-2026-07.md)。
479
+ 但书)见 [`docs/model-capabilities-2026-09.md`](docs/model-capabilities-2026-09.md)。
497
480
 
498
481
  ## ⚠️ 已知限制
499
482
 
@@ -507,6 +490,12 @@ vendor 一律当成 `work`,而且它只能逐次明确指定,永远不是 lane
507
490
 
508
491
  ## 📜 版本历程
509
492
 
493
+ ## v0.32.0 新功能
494
+
495
+ - **基于 AA 2026-09 快照全面重评路由。** Fable 5.1 和 Gemini 3.7 Flash 进入默认表,数据集中在新的日期化文档。
496
+ - **模型目录同步当前 CLI 接口。** 加入 Fable 5.1,移除 agy 已下架的 Gemini 3.5 Flash 项,并同步投票器。
497
+ - **Opus 5 仍可使用。** 它保留在 `long-context`,也能通过 `routing.local.yaml` 覆盖任意通道。
498
+
510
499
  ## v0.31.0 新功能
511
500
 
512
501
  - **目标预算默认不设上限。** `budget_jobs` 和 `budget_seconds` 现在以 JSON `null` 保存并显示为 `unlimited`;此前隐含的 8 个任务和 900 秒上限已移除。只有使用 `--budget-jobs N` 或 `--budget-seconds S` 才会启用硬性上限;重复失败保险丝不是预算,默认仍然启用。
@@ -582,7 +571,7 @@ vendor 一律当成 `work`,而且它只能逐次明确指定,永远不是 lane
582
571
  的 Agentic Index 上大幅领先 Flash,而且 2026-07-30 降价后每任务成本只剩零头。
583
572
  Flash 只剩吞吐量优势——若你的循环受延迟限制,可在本机覆写把它调回第一。
584
573
  - **lane 注释不再放数字。**`routing.yaml` 只说明每条排序「为什么」成立;所有分数、
585
- 价格与吞吐量连同取数日期,一律放在 `docs/model-capabilities-2026-07.md`。数字过期
574
+ 价格与吞吐量连同取数日期,一律放在 `docs/model-capabilities-2026-09.md`。数字过期
586
575
  不再需要动路由表。
587
576
  - **新增 value profile**(在 `routing.local.yaml.example`):用约一个 Intelligence
588
577
  Index 分数,换每任务成本降三到四成。
@@ -628,7 +617,7 @@ vendor 一律当成 `work`,而且它只能逐次明确指定,永远不是 lane
628
617
  - **修复:`routing.local.yaml.example` 还指着已退场的模型**——起手配置里的
629
618
  `claude-opus-4-8` 全数改为 `claude-opus-5`(并按通道给对应档位),Gemini 3.5
630
619
  Flash 候选改为 3.6 Flash,与 0.10.0 以来的默认值一致。
631
- - **对照原始资料修正智能指数数字**(`docs/model-capabilities-2026-07.md`):
620
+ - **对照原始资料修正智能指数数字**(`docs/model-capabilities-2026-09.md`):
632
621
  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
633
622
  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。
634
623
 
@@ -665,7 +654,7 @@ vendor 一律当成 `work`,而且它只能逐次明确指定,永远不是 lane
665
654
 
666
655
  ## v0.9.0 新功能
667
656
 
668
- - **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)。
657
+ - **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-2026-09.md`](docs/model-capabilities-2026-09.md)。
669
658
  - **fish shell 补全** — `omnilane completion fish | source`。
670
659
 
671
660
  ## v0.8.3 新功能
package/README.zh-TW.md CHANGED
@@ -72,11 +72,11 @@ omnilane 讓**任何**一個 agentic CLI 的主迴圈把子任務分類到通道
72
72
  ```mermaid
73
73
  flowchart LR
74
74
  M["主迴圈<br/><i>你在用的任一 CLI</i>"] --> T{{"routing.yaml<br/>一張共用路由表"}}
75
- T -->|hardest-coding| C1["CodexGPT-5.6 Sol"]
76
- T -->|bulk-mechanical| C2["Codex — GPT-5.6 Terra"]
77
- T -->|taste-final| C3["Claude — Opus 5"]
78
- T -->|long-context| C4["Gemini — 3.1 Pro"]
79
- T -->|live-search| C5["Grok — 4.5"]
75
+ T -->|hardest-coding| C1["ClaudeFable 5.1"]
76
+ T -->|bulk-mechanical| C2["Codex — GPT-5.6 Sol"]
77
+ T -->|taste-final| C3["Claude — Fable 5.1"]
78
+ T -->|long-context| C4["Gemini — 3.7 Flash"]
79
+ T -->|live-search| C5["Grok — 4.6"]
80
80
  T -->|"arbitrate(選配)"| C6["vote — 1-4 模型評審團"]
81
81
  ```
82
82
 
@@ -103,24 +103,23 @@ flowchart LR
103
103
 
104
104
  | 通道 | 首選模型 | 備選模型 | 用途 |
105
105
  |---|---|---|---|
106
- | 🔥 hardest-coding | GPT-5.6 Sol (xhigh) | Claude Opus 5 (xhigh) | 最難的實作、深度除錯、正確性攸關的修改 |
107
- | 🏗️ bulk-mechanical | GPT-5.6 Terra (max) | Claude Sonnet 5 (high) | 重構、搬遷、測試、大面積掃描——機械耐力活 |
108
- | 🧹 triage | GPT-5.6 Luna (medium) | Gemini 3.6 Flash (Low) | 高量初篩、第一輪過濾 |
109
- | ⚖️ hard-judgment | Claude Opus 5 (xhigh) | GPT-5.6 Sol (max) | 架構仲裁、深度推理、第二意見 |
110
- | ✒️ taste-final | Claude Opus 5 (high) | GPT-5.6 Sol (max) | 對外文字、prompt 與文件打磨、風格終審 |
111
- | 💬 consult | 明確點名的廠商/模型 | (不降級) | 自然語言直接諮詢;必須保留 `--vendor` |
112
- | 🎨 ui-draft | GPT-5.6 Sol (xhigh) | Claude Opus 5 (high) | 有設計規範/參考圖時的 UI 出稿;開放式視覺品味交給 taste-final |
113
- | 📚 long-context | Gemini 3.1 Pro (High) | GPT-5.6 Sol (high) | 百萬 token 掃讀、檢索與長文跨段整合;高速重複迴圈仍優先 Flash |
114
- | ⚡ fast-agentic | GPT-5.6 Luna (max) | Gemini 3.6 Flash (High) | 快速多步驟 agentic 迴圈、多模態檢查 |
115
- | 📡 live-search | Grok 4.5 | —(off) | 即時 X/網路搜尋與社群脈絡 |
116
- | 🚰 coding-overflow | Grok 4.5 | Kimi K3 → Qwen3 Coder Plus → OpenCode | Codex 額度吃緊時的中量級編碼溢流道;事實性宣稱須另行查證 |
117
- | 🗳️ arbitrate | off(選配評審團) | — | 內建意見評審團,重大決定用——預設關閉,要用在 `routing.local.yaml` 開;每評審每輪燒一次額度 |
106
+ | 🔥 hardest-coding | Claude Fable 5.1 (xhigh) | GPT-5.6 Sol (xhigh) | 最難的實作、深度除錯、正確性攸關的修改 |
107
+ | 🏗️ bulk-mechanical | GPT-5.6 Sol (high) | Gemini 3.7 Flash (High) → Claude Sonnet 5 (high) | 重構、搬遷、測試、大面積掃描——機械耐力活 |
108
+ | 🧹 triage | GPT-5.6 Luna (high) | Gemini 3.7 Flash (Low) → Claude Haiku 4.5 | 大量掃描、第一輪篩選 |
109
+ | ⚖️ hard-judgment | Claude Fable 5.1 (xhigh) | GPT-5.6 Sol (max) → Grok 4.6 | 架構裁決、深度推理、第二意見 |
110
+ | ✒️ taste-final | Claude Fable 5.1 (high) | GPT-5.6 Sol (max) | 對外文字、提示詞/文件潤飾、風格裁決 |
111
+ | 💬 consult | GPT-5.6 Sol (max) | Claude Fable 5.1 (high) → Grok 4.6 → Gemini 3.7 Flash (High) | 直接點名模型諮詢;保留 `--vendor` 避免降級 |
112
+ | 🎨 ui-draft | GPT-5.6 Sol (xhigh) | Claude Fable 5.1 (high) | 只有在附設計系統/參考圖時才做 UI 草稿 |
113
+ | 📚 long-context | Gemini 3.7 Flash (Medium) | GPT-5.6 Terra (max) Claude Opus 5 (medium) | 長文件擷取與整合,依 AA-LCR、成本與吞吐排序 |
114
+ | ⚡ fast-agentic | Gemini 3.7 Flash (Medium) | GPT-5.6 Luna (high) | 高速多步驟工具迴圈、多模態檢查 |
115
+ | 📡 live-search | Grok 4.6 | — (`off`) | 即時 X/網頁搜尋與社群脈絡 |
116
+ | 🚰 coding-overflow | Grok 4.6 | Gemini 3.7 Flash (High) → Kimi K3 → Qwen3 Coder Plus → OpenCode | Codex 額度用完時的中量級編碼安全閥 |
117
+ | 🗳️ arbitrate | `off`(選配模型評審團) | — | 重大決定的內建意見評審團;預設停用,在 `routing.local.yaml` 啟用,每位評審每輪一次呼叫 |
118
118
 
119
119
  **備選模型**是候選鏈的下一位——首選那家的廠商 CLI 沒裝時,派工就降到它。每條
120
120
  通道都是這樣一條鏈;整條都沒裝時,通道自動降為 `off`。
121
121
 
122
- > **Claude Fable 5 去哪了?** 預設表刻意不放——理由與完整數據見
123
- > [常見問題](#-常見問題)。
122
+ > **Fable 5.1 已進入預設——以及 Opus 5 仍適合放在哪裡。** 三方數據與 Opus override 寫在[常見問題](#-常見問題)。
124
123
 
125
124
  ### 自然語言諮詢
126
125
 
@@ -141,12 +140,12 @@ flowchart LR
141
140
  你哪些通道**自己做**(你本來就是那個模型,省一次呼叫)、哪些**派出去**。你 CLI 裡
142
141
  的 `omnilane` 技能會自動套對的那一列,這裡是給人看的版本。
143
142
 
144
- - **Claude Code · Fable 5** — 自己做:hard-judgment、taste-final、最吃正確性的硬修。派出去:機械編碼量 → Codex、長文 → Gemini、即時搜尋 → Grok。
145
- - **Claude Code · Opus 5** — 自己做:hard-judgment、taste-final。大量編碼走 Codex 通道、長文 → Gemini、即時搜尋 → Grok。
146
- - **Codex · Sol** — 自己做:hardest-coding、hard-judgment、ui-draft。派出去:taste-final → Claude、長文 → Gemini、即時搜尋 Grok、粗活Codex Terra
147
- - **Codex · Terra** 自己做:bulk-mechanical。真正最硬的往上升給 Sol;taste → Claude、長文 → Gemini、即時搜尋 → Grok。
148
- - **Grok Build · Grok 4.5** — 自己做:live-search、coding-overflow(中量級編碼)。所有硬活派給 Codex/Claude/Gemini——先驗每個 API 簽章與引用事實。
149
- - **Antigravity · Gemini** — 自己做:3.1 Pro 的長文與重脈絡 agentic 工作、Flash 的高速重複迴圈。最硬的編碼/判斷/文字派給 Codex/Claude;即時搜尋 → Grok。
143
+ - **Claude Code · Fable 5.1**——自己做:hard-judgment、taste-final、hardest-coding。派出去:bulk → Codex Sol high;long-context/高速迴圈 → Gemini 3.7 Flash;即時搜尋 → Grok。
144
+ - **Claude Code · Opus 5**——需要較低幻覺率或價格時,自己做 hard-judgment、taste-final。最難編碼 Fable 5.1 或 Sol;bulk Sol high;long-context/高速迴圈 → Gemini 3.7 Flash;即時搜尋 → Grok。
145
+ - **Codex · Sol**——自己做:hardest-coding、bulk-mechanical、hard-judgment、ui-draft。派出去:taste-final → Claude;long-context/高速迴圈 → Gemini 3.7 Flash;即時搜尋Grok
146
+ - **Codex · Terra**——自己做 long-context 的 Codex 備援;bulk-mechanical 已改由 Sol high 預設處理。最難處升級 Sol xhigh,taste → Claude,高速迴圈 → Gemini 3.7 Flash,即時搜尋 → Grok。
147
+ - **Grok Build · Grok 4.6**——自己做 live-search、coding-overflow。最難的編碼/判斷/文字交給 CodexClaudeGemini;仍要驗證 API 簽章與引用事實。
148
+ - **Antigravity · Gemini 3.7 Flash**——自己做:Medium long-context/高速迴圈、High bulk/overflow、Low 的 triage。最難編碼/判斷/文字交給 CodexClaude;即時搜尋 → Grok。
150
149
 
151
150
  </details>
152
151
 
@@ -422,45 +421,29 @@ omnilane goal close "$GOAL_ID" --summary "結帳整合已穩定"
422
421
  </details>
423
422
 
424
423
  <details>
425
- <summary><b>Claude Fable 5 去哪了?為什麼預設表不放它?</b></summary>
424
+ <summary><b>Fable 5.1 已進入預設——以及 Opus 5 仍適合放在哪裡</b></summary>
426
425
 
427
426
  <br/>
428
427
 
429
- **因為 Claude 頂級檔通常就是主迴圈本人,不是被派發的工人。** 通道存在的意義,
430
- 是把工作送給「你正在開的那個模型以外」的模型。如果 Fable 5 就是你的主迴圈,把
431
- 判斷和文字再路由回 Fable 5 只是多一次呼叫、毫無增益——所以上面那份「選你的主控
432
- 模型」清單裡,Fable 5 有自己獨立的一列,身分是**主控**:自己做 hard-judgment、
433
- taste-final、最吃正確性的硬修。
428
+ Fable 5.1 現在領頭 `hardest-coding`、`hard-judgment`、`taste-final`。
429
+ 同為 xhigh 時,它在智慧、代理式工作與編碼都領先 Opus 5;Sol max 則保留為
430
+ 便宜許多的跨廠商判斷備援。
434
431
 
435
- **量測數據也不支持把它當工人。** Artificial Analysis 智慧指數(2026-07-24)
436
- Opus 5(max)61 分、Fable 5(max)60 分——AA 自己的用詞是「實質打平」,而 Epoch AI
437
- 的能力指數排序還是反過來的(Fable 5 161、Opus 5 159)。綜合智力就當它平手。真正
438
- 拉開差距的是 agentic 專業產出,而且差很多:
432
+ | 評測(AA,擷取於 2026-09-02) | Claude Fable 5.1 (xhigh) | Claude Opus 5 (xhigh) | GPT-5.6 Sol (max) |
433
+ |---|---:|---:|---:|
434
+ | 智慧 | 64.8 | 62.5 | 60.9 |
435
+ | 代理式 | 59.8 | 58.4 | 57.8 |
436
+ | 編碼 | 80.7 | 77.0 | 77.4 |
437
+ | 幻覺率(越低越好) | .71 | **.60** | .92 |
438
+ | AA 每任務成本 | $2.65 | $1.80 | **$0.95** |
439
439
 
440
- | 評測 | Claude Opus 5 (max) | Claude Fable 5 | |
441
- |---|---:|---:|---|
442
- | AA-Briefcase(agentic 知識工作,Elo) | 1720 | 1574 | **+146** |
443
- | GDPval-AA v2(Elo) | 1861 | 1747 | **+114** |
444
- | AA-Briefcase 每任務成本 | $17.79 | $22.30 | **-20%** |
445
- | API 單價,輸入/輸出 每 1M | $5 / $25 | $10 / $50 | **一半** |
446
-
447
- Opus 5 的 max、xhigh、high 三個檔位包辦 AA-Briefcase 前三名,連 `high` 檔都在
448
- 不到一半的每任務成本下贏過 Fable 5。也就是說,Fable 5 貴一倍,卻換不到任何一條
449
- 通道所在意的優勢。
450
-
451
- **Fable 5 真正比較強的地方**:事實廣度。它在 AA-Omniscience 上仍領先 Opus 5
452
- (符合兩者的規模差),而 Opus 5 在沒把握時更傾向直接作答——幻覺率 50%,比
453
- Opus 4.8 高 14 個百分點。任務偏「回想」而非「執行」時,直接點名它:
454
-
455
- ```bash
456
- dispatch.sh --vendor claude --model claude-fable-5 --effort high consult "…"
457
- ```
458
-
459
- **這是成本與主迴圈策略的選擇,不是能力判決。** 設定選單的模型清單有列 Fable 5,
460
- 在 `routing.local.yaml` 寫一行就能覆寫預設:
440
+ Fable 5.1 沒進 bulk triage:每 token 價格是 Opus 5 的兩倍,而且每回合
441
+ 消耗最多 Claude Code 訂閱額度。Opus 5 仍是較低幻覺率、較低價格的 Claude
442
+ 選項,並以 medium 留在 `long-context`;也能透過
443
+ `~/.omnilane/routing.local.yaml` 放回任何通道:
461
444
 
462
445
  ```yaml
463
- taste-final: claude claude-fable-5 high
446
+ hard-judgment: claude claude-opus-5 xhigh
464
447
  ```
465
448
 
466
449
  </details>
@@ -517,7 +500,7 @@ vendor 一律當成 `work`,而且它只能逐次明確指定,永遠不是 lane
517
500
  預設通道配置依據 Artificial Analysis 2026-07 快照(已對 AA 站上原始紀錄與
518
501
  各廠官方定價頁交叉核對)加上公開對比評測;這些是意見不是定律——
519
502
  設定選單和 `routing.local.yaml` 就是讓你不同意用的。完整工作筆記(含各評測的
520
- 但書)見 [`docs/model-capabilities-2026-07.md`](docs/model-capabilities-2026-07.md)。
503
+ 但書)見 [`docs/model-capabilities-2026-09.md`](docs/model-capabilities-2026-09.md)。
521
504
 
522
505
  ## ⚠️ 已知限制
523
506
 
@@ -531,6 +514,12 @@ vendor 一律當成 `work`,而且它只能逐次明確指定,永遠不是 lane
531
514
 
532
515
  ## 📜 版本歷程
533
516
 
517
+ ## v0.32.0
518
+
519
+ - **依 AA 2026-09 快照全面重評路由。** Fable 5.1 與 Gemini 3.7 Flash 進入預設,數據集中在新的日期化文件。
520
+ - **模型目錄同步實際 CLI 軟體介面。** 加入 Fable 5.1,移除 agy 已下架的 Gemini 3.5 Flash 項目,並同步投票程式。
521
+ - **Opus 5 仍可使用。** 它留在 `long-context`,也能透過 `routing.local.yaml` 覆寫任何通道。
522
+
534
523
  ## v0.31.0 新功能
535
524
 
536
525
  - **目標預算預設無上限。** `budget_jobs` 與 `budget_seconds` 現在會以 JSON `null` 儲存並顯示為 `unlimited`;原先隱含的 8 個工作與 900 秒上限已移除。使用 `--budget-jobs N` 或 `--budget-seconds S` 才會啟用硬性上限;重複失敗保險絲不是預算,預設仍會啟用。
@@ -611,7 +600,7 @@ vendor 一律當成 `work`,而且它只能逐次明確指定,永遠不是 lane
611
600
  的 Agentic Index 上大幅領先 Flash,而且 2026-07-30 砍價後每任務成本只剩零頭。
612
601
  Flash 只剩吞吐量優勢——若你的迴圈受延遲限制,可在本機覆寫把它調回第一。
613
602
  - **lane 註解不再放數字。**`routing.yaml` 只說明每條排序「為什麼」成立;所有分數、
614
- 價格與吞吐量連同取數日期,一律住在 `docs/model-capabilities-2026-07.md`。數字過期
603
+ 價格與吞吐量連同取數日期,一律住在 `docs/model-capabilities-2026-09.md`。數字過期
615
604
  不再需要動路由表。
616
605
  - **新增 value profile**(在 `routing.local.yaml.example`):用約一個 Intelligence
617
606
  Index 分數,換每任務成本降三到四成。
@@ -657,7 +646,7 @@ vendor 一律當成 `work`,而且它只能逐次明確指定,永遠不是 lane
657
646
  - **修正:`routing.local.yaml.example` 還指著已退場的模型**——起手設定檔裡的
658
647
  `claude-opus-4-8` 全數改為 `claude-opus-5`(並依通道給對應檔位),Gemini 3.5
659
648
  Flash 候選改為 3.6 Flash,與 0.10.0 以來的預設值一致。
660
- - **對照原始資料修正智慧指數數字**(`docs/model-capabilities-2026-07.md`):
649
+ - **對照原始資料修正智慧指數數字**(`docs/model-capabilities-2026-09.md`):
661
650
  那是指數點數不是百分比;補上 AA-Briefcase / GDPval-AA v2 對照,並記下兩項與
662
651
  預設值相反的結果:Fable 5 在事實知識領先、GPT-5.6 Sol 在呈現品質領先。
663
652
 
@@ -697,7 +686,7 @@ vendor 一律當成 `work`,而且它只能逐次明確指定,永遠不是 lane
697
686
  - **五個 OpenAI-compatible direct-API vendor** — `deepseek`、`zai`(GLM)、
698
687
  `mistral`、`groq`、`cerebras`,與 `openrouter` 同為免 CLI 通道(curl 加一把
699
688
  `<VENDOR>_API_KEY`);`lib/common.sh` registry 一行即加一個。詳見
700
- [`docs/model-capabilities-2026-07.md`](docs/model-capabilities-2026-07.md)。
689
+ [`docs/model-capabilities-2026-09.md`](docs/model-capabilities-2026-09.md)。
701
690
  - **fish shell 補全** — `omnilane completion fish | source`。
702
691
 
703
692
  ## v0.8.3 新功能
@@ -719,7 +708,7 @@ vendor 一律當成 `work`,而且它只能逐次明確指定,永遠不是 lane
719
708
  同一條免 CLI 直連 API 路徑,對應 OpenAI-compatible 供應商:DeepSeek、Z.ai GLM、
720
709
  Mistral、Groq、Cerebras。各只要 `curl` 加自己那把 `<VENDOR>_API_KEY`;僅限
721
710
  advise/consult。端點、金鑰、預設模型由 `lib/common.sh` 一行 registry 定義。
722
- 詳見 [`docs/model-capabilities-2026-07.md`](docs/model-capabilities-2026-07.md)。
711
+ 詳見 [`docs/model-capabilities-2026-09.md`](docs/model-capabilities-2026-09.md)。
723
712
  - **`opencode` vendor** — 透過 OpenCode 多供應商聚合 CLI 無頭派工
724
713
  (`opencode run`)。advise 模式鎖定內建唯讀 `plan` agent;work 模式
725
714
  用 `--auto`。加入預設 `coding-overflow` 鏈作為最後備援。
package/VERSION CHANGED
@@ -1 +1 @@
1
- 0.31.0
1
+ 0.32.0
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "omnilane",
3
- "version": "0.31.0",
3
+ "version": "0.32.0",
4
4
  "description": "One routing table, every harness — classify subtasks into lanes and dispatch each lane to the best vendor's agentic CLI (Codex, Claude, Gemini, Grok) using your existing subscription logins.",
5
5
  "bin": {
6
6
  "omnilane": "bin/omnilane"
package/plugin.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "$schema": "https://antigravity.google/schemas/v1/plugin.json",
3
3
  "name": "omnilane",
4
- "version": "0.31.0",
4
+ "version": "0.32.0",
5
5
  "description": "One routing table, every harness: classify subtasks into lanes and auto-dispatch each lane to the best vendor CLI (Codex, Claude Code, Grok Build, Antigravity) with background jobs, depth guard, and serialized codex dispatch."
6
6
  }