omnilane 0.21.0 → 0.32.0

package/README.zh-TW.md

@@ -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
 ```mermaid
 flowchart LR
     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"}}
-    T -->|hardest-coding| C1["Codex — GPT-5.6 Sol"]
-    T -->|bulk-mechanical| C2["Codex — GPT-5.6 Terra"]
-  T -->|taste-final| C3["Claude — Opus 5"]
-    T -->|long-context| C4["Gemini — 3.1 Pro"]
-    T -->|live-search| C5["Grok — 4.5"]
+    T -->|hardest-coding| C1["Claude — Fable 5.1"]
+    T -->|bulk-mechanical| C2["Codex — GPT-5.6 Sol"]
+    T -->|taste-final| C3["Claude — Fable 5.1"]
+    T -->|long-context| C4["Gemini — 3.7 Flash"]
+    T -->|live-search| C5["Grok — 4.6"]
     T -->|"arbitrate(選配)"| C6["vote — 1-4 模型評審團"]
 ```
 
@@ -103,24 +103,23 @@ flowchart LR
 
 | 通道 | 首選模型 | 備選模型 | 用途 |
 |---|---|---|---|
-| 🔥 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 |
-| 🏗️ 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 |
-| 🧹 triage | GPT-5.6 Luna (medium) | Gemini 3.6 Flash (Low) | 高量初篩、第一輪過濾 |
-| ⚖️ hard-judgment | Claude Opus 5 (xhigh) | GPT-5.6 Sol (max) | 架構仲裁、深度推理、第二意見 |
-| ✒️ 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 |
-| 💬 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` |
-| 🎨 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 |
-| 📚 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 |
-| ⚡ 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 |
-| 📡 live-search | Grok 4.5 | —(off) | 即時 X/網路搜尋與社群脈絡 |
-| 🚰 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 |
-| 🗳️ 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 |
+| 🔥 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 |
+| 🏗️ 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 |
+| 🧹 triage | GPT-5.6 Luna (high) | Gemini 3.7 Flash (Low) → Claude Haiku 4.5 | 大量掃描、第一輪篩選 |
+| ⚖️ 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 |
+| ✒️ 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 |
+| 💬 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 |
+| 🎨 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 |
+| 📚 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 |
+| ⚡ 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 |
+| 📡 live-search | Grok 4.6 | — (`off`) | 即時 X／網頁搜尋與社群脈絡 |
+| 🚰 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 |
+| 🗳️ 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 |
 
 **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
 通道都是這樣一條鏈;整條都沒裝時,通道自動降為 `off`。
 
-> **Claude Fable 5 去哪了?** 預設表刻意不放——理由與完整數據見
-> [常見問題](#-常見問題)。
+> **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)。
 
 ### 自然語言諮詢
 
@@ -141,12 +140,12 @@ flowchart LR
 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
 的 `omnilane` 技能會自動套對的那一列,這裡是給人看的版本。
 
-- **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。
-- **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。
-- **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。
-- **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。
-- **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。
-- **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。
+- **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。
+- **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。
+- **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。
+- **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。
+- **Grok Build · Grok 4.6**——自己做 live-search、coding-overflow。最難的編碼／判斷／文字交給 Codex、Claude、Gemini；仍要驗證 API 簽章與引用事實。
+- **Antigravity · Gemini 3.7 Flash**——自己做：Medium 的 long-context／高速迴圈、High 的 bulk／overflow、Low 的 triage。最難編碼／判斷／文字交給 Codex、Claude；即時搜尋 → Grok。
 
 </details>
 
@@ -378,6 +377,22 @@ scripts/jobs.sh retry "$ID" --background
 
 閒置時不會發出 API 呼叫，也不會增加 API 費用。預設若 900 秒內沒有新信箱訊息或新結果事件，工作程序會自動收尾；整體工作逾時仍是外層上限。處理完成可提早執行 `close`。對已結束或不是即時信箱的工作使用 `jobs.sh send`，會明確報錯並失敗。送出後不需追蹤的工作、沒有即時支援的供應商，或必須從乾淨狀態重跑的情況都不適用；請使用新的派工，或在工作完成後使用 `retry`。
 
+## 🎯 目標編排
+
+`omnilane goal` 是工頭式工作台帳。迴圈由呼叫端負責，也就是開啟目標的代理工作階段或終端機前的人：派出一份工作，從完成信箱或 `omnilane jobs wait` 收回結果，判斷下一份工作，再重複執行。工作數與秒數預算預設都不設上限；只有傳入 `--budget-jobs N` 或 `--budget-seconds S` 時，才會啟用對應的硬上限。omnilane 只負責記帳；每次 goal dispatch 前會檢查呼叫端設定的上限與預設啟用的重複失敗熔斷器，工頭關閉目標時才彙整報告。
+
+```bash
+GOAL_ID="$(omnilane goal open "修好不穩定的結帳整合" \
+  --budget-jobs 4 --budget-seconds 900 --workdir /path/to/repo)"
+JOB_ID="$(omnilane goal dispatch "$GOAL_ID" --mode work hardest-coding \
+  "重現結帳失敗，完成最小修正並驗證")"
+omnilane jobs wait "$JOB_ID" --timeout 900
+omnilane goal note "$GOAL_ID" "結帳整合測試已通過"
+omnilane goal close "$GOAL_ID" --summary "結帳整合已穩定"
+```
+
+目標狀態存放在 `$OMNILANE_HOME/goals/<goal-id>/`。用 `goal status` 可查看預算用量、熔斷次數，以及每份工作陸續寫入的中繼資料與結束狀態。`goal close` 會寫入 `report.md` 並印出路徑。單一而且作法明確的工作直接派工即可；有傳入預算旗標時，該上限是硬限制，不代表保證完成。
+
 ## ❓ 常見問題
 
 <details>
@@ -406,45 +421,29 @@ scripts/jobs.sh retry "$ID" --background
 </details>
 
 <details>
-<summary><b>Claude Fable 5 去哪了?為什麼預設表不放它?</b></summary>
+<summary><b>Fable 5.1 已進入預設——以及 Opus 5 仍適合放在哪裡</b></summary>
 
 <br/>
 
-**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,
-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
-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
-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、
-taste-final、最吃正確性的硬修。
-
-**量測數據也不支持把它當工人。** Artificial Analysis 智慧指數(2026-07-24)
-Opus 5(max)61 分、Fable 5(max)60 分——AA 自己的用詞是「實質打平」,而 Epoch AI
-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
-拉開差距的是 agentic 專業產出,而且差很多:
-
-| 評測 | Claude Opus 5 (max) | Claude Fable 5 | |
-|---|---:|---:|---|
-| AA-Briefcase(agentic 知識工作,Elo) | 1720 | 1574 | **+146** |
-| GDPval-AA v2(Elo) | 1861 | 1747 | **+114** |
-| AA-Briefcase 每任務成本 | $17.79 | $22.30 | **-20%** |
-| API 單價,輸入/輸出 每 1M | $5 / $25 | $10 / $50 | **一半** |
-
-Opus 5 的 max、xhigh、high 三個檔位包辦 AA-Briefcase 前三名,連 `high` 檔都在
-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
-通道所在意的優勢。
+Fable 5.1 現在領頭 `hardest-coding`、`hard-judgment`、`taste-final`。
+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
+便宜許多的跨廠商判斷備援。
 
-**Fable 5 真正比較強的地方**:事實廣度。它在 AA-Omniscience 上仍領先 Opus 5
-(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
-Opus 4.8 高 14 個百分點。任務偏「回想」而非「執行」時,直接點名它:
+| 評測（AA，擷取於 2026-09-02） | Claude Fable 5.1 (xhigh) | Claude Opus 5 (xhigh) | GPT-5.6 Sol (max) |
+|---|---:|---:|---:|
+| 智慧 | 64.8 | 62.5 | 60.9 |
+| 代理式 | 59.8 | 58.4 | 57.8 |
+| 編碼 | 80.7 | 77.0 | 77.4 |
+| 幻覺率（越低越好） | .71 | **.60** | .92 |
+| AA 每任務成本 | $2.65 | $1.80 | **$0.95** |
 
-```bash
-dispatch.sh --vendor claude --model claude-fable-5 --effort high consult "…"
-```
-
-**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,
-在 `routing.local.yaml` 寫一行就能覆寫預設:
+Fable 5.1 沒進 bulk 或 triage：每 token 價格是 Opus 5 的兩倍，而且每回合
+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
+選項，並以 medium 留在 `long-context`；也能透過
+`~/.omnilane/routing.local.yaml` 放回任何通道：
 
 ```yaml
-taste-final: claude claude-fable-5 high
+hard-judgment: claude claude-opus-5 xhigh
 ```
 
 </details>
@@ -501,7 +500,7 @@ vendor 一律當成 `work`,而且它只能逐次明確指定,永遠不是 lane
 預設通道配置依據 Artificial Analysis 2026-07 快照(已對 AA 站上原始紀錄與
 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——
 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
-但書)見 [`docs/model-capabilities-2026-07.md`](docs/model-capabilities-2026-07.md)。
+但書)見 [`docs/model-capabilities-2026-09.md`](docs/model-capabilities-2026-09.md)。
 
 ## ⚠️ 已知限制
 
@@ -515,6 +514,23 @@ vendor 一律當成 `work`,而且它只能逐次明確指定,永遠不是 lane
 
 ## 📜 版本歷程
 
+## v0.32.0
+
+- **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。
+- **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。
+- **Opus 5 仍可使用。** 它留在 `long-context`，也能透過 `routing.local.yaml` 覆寫任何通道。
+
+## v0.31.0 新功能
+
+- **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。
+- **管線中的目標狀態不再誤判失敗。** `omnilane goal status` 的消費端提早關閉管線時，現在會以狀態碼 0 結束，不再引發 `BrokenPipeError`，因此 `| head` 與 `| grep -q` 可在 `pipefail` 下正常運作。
+
+## v0.30.0 新功能
+
+- **目標台帳。** `omnilane goal open` 建立預設不限制工作數與秒數的目標台帳；`goal dispatch` 會在每份工作執行前檢查呼叫端設定的工作數或總經過時間上限，以及預設啟用的重複失敗熔斷器。`goal note` 保留呼叫端敘事，`goal status` 顯示預算與各工作紀錄，`goal close` 會寫入 `goals/<id>/report.md`。
+- **迴圈由呼叫端掌握。** 開啟目標的工作階段或使用者負責選擇、派工、檢視與收尾；omnilane 不會執行內建的規劃模型。
+- **doctor 檢查。** `omnilane doctor` 現在會檢查目標編排功能。
+
 ## v0.21.0 新功能
 
 - **明確選擇工作階段模式。** 可用 `dispatch --live` 要求常駐工作階段，或以 `--single-shot` 強制單次派工；對不支援即時工作階段的供應商，`--live` 會立即失敗並列出可用供應商。
@@ -584,7 +600,7 @@ vendor 一律當成 `work`,而且它只能逐次明確指定,永遠不是 lane
   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。
   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。
 - **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、
-  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
+  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
   不再需要動路由表。
 - **新增 value profile**(在 `routing.local.yaml.example`):用約一個 Intelligence
   Index 分數,換每任務成本降三到四成。
@@ -630,7 +646,7 @@ vendor 一律當成 `work`,而且它只能逐次明確指定,永遠不是 lane
 - **修正:`routing.local.yaml.example` 還指著已退場的模型**——起手設定檔裡的
   `claude-opus-4-8` 全數改為 `claude-opus-5`(並依通道給對應檔位),Gemini 3.5
   Flash 候選改為 3.6 Flash,與 0.10.0 以來的預設值一致。
-- **對照原始資料修正智慧指數數字**(`docs/model-capabilities-2026-07.md`):
+- **對照原始資料修正智慧指數數字**(`docs/model-capabilities-2026-09.md`):
   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
   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。
 
@@ -670,7 +686,7 @@ vendor 一律當成 `work`,而且它只能逐次明確指定,永遠不是 lane
 - **五個 OpenAI-compatible direct-API vendor** — `deepseek`、`zai`(GLM)、
   `mistral`、`groq`、`cerebras`,與 `openrouter` 同為免 CLI 通道(curl 加一把
   `<VENDOR>_API_KEY`);`lib/common.sh` registry 一行即加一個。詳見
-  [`docs/model-capabilities-2026-07.md`](docs/model-capabilities-2026-07.md)。
+  [`docs/model-capabilities-2026-09.md`](docs/model-capabilities-2026-09.md)。
 - **fish shell 補全** — `omnilane completion fish | source`。
 
 ## v0.8.3 新功能
@@ -692,7 +708,7 @@ vendor 一律當成 `work`,而且它只能逐次明確指定,永遠不是 lane
   同一條免 CLI 直連 API 路徑,對應 OpenAI-compatible 供應商:DeepSeek、Z.ai GLM、
   Mistral、Groq、Cerebras。各只要 `curl` 加自己那把 `<VENDOR>_API_KEY`;僅限
   advise/consult。端點、金鑰、預設模型由 `lib/common.sh` 一行 registry 定義。
-  詳見 [`docs/model-capabilities-2026-07.md`](docs/model-capabilities-2026-07.md)。
+  詳見 [`docs/model-capabilities-2026-09.md`](docs/model-capabilities-2026-09.md)。
 - **`opencode` vendor** — 透過 OpenCode 多供應商聚合 CLI 無頭派工
   (`opencode run`)。advise 模式鎖定內建唯讀 `plan` agent;work 模式
   用 `--auto`。加入預設 `coding-overflow` 鏈作為最後備援。

package/VERSION

@@ -1 +1 @@
-0.21.0
+0.32.0

package/bin/omnilane

@@ -76,9 +76,14 @@ omnilane — one routing table, every harness
   omnilane list                         effective routing table
   omnilane route [--vendor V] [flags] LANE "TASK"
                                         dispatch or consult a model
-  omnilane goal "TEXT" [--budget-jobs N] [--budget-seconds S] [--workdir DIR]
-                                        run bounded sequential goal loop
-  omnilane goal status GOAL_ID           inspect goal budget and last action
+  omnilane goal open "TEXT" [--budget-jobs N] [--budget-seconds S] [--workdir DIR]
+                                        open a goal; budgets default to unlimited
+  omnilane goal dispatch GOAL_ID [dispatch.sh args...]
+                                         budget-gate and start one background job
+  omnilane goal note GOAL_ID "TEXT"       append a timestamped foreman note
+  omnilane goal status GOAL_ID           inspect budgets, fuse trips, and recorded jobs
+  omnilane goal close GOAL_ID [--summary "TEXT"]
+                                         write the final goal report
   omnilane jobs  [--json] COMMAND [ARGS]  list, status, result, stats, recommend, wait, or audit
   omnilane mcp                          start the MCP stdio server
   omnilane completion bash|zsh|fish      print a shell completion script

package/hooks/routing-instruction.md

@@ -1,18 +1,27 @@
 <!-- omnilane-routing:start -->
 ## omnilane — model routing (persistent reminder)
 
-Before delegating any subtask or choosing a model for a piece of work,
-consult the omnilane routing table: run `omnilane list` (or
-`scripts/dispatch.sh --list` inside the omnilane repo) and classify the
-subtask into a lane. If the lane's first available model is the one you are
-running as, self-execute; otherwise dispatch it headlessly:
+Implementation work — code edits, new files, tests, builds, deploys — is
+dispatched by default, even when the lane's first available model is the one
+you are running as. Consult the routing table with `omnilane list` (or
+`scripts/dispatch.sh --list` inside the omnilane repo), classify the subtask
+into a lane, then dispatch it headlessly:
 
     omnilane route [--vendor V] [--mode work] [--workdir DIR] <lane> "<task>"
 
+Advise mode is the default; pass `--mode work` only with an explicit
+`--workdir`. The commander self-executes only reserved items: planning and
+decomposition, writing task briefs, reviewing reports, acceptance checks,
+replies to the operator, git commit/push, read-only verification, and fixes
+of one line or less. "This lane is mine, so I'll do it myself" is not a
+valid reason to skip dispatch.
+
 If the user explicitly names Claude, Codex, Grok, Gemini, or a canonical model
 alias, use the omnilane skill's consult rules and keep `--vendor` in the
 dispatch; an explicit target must not silently fall back.
 
-Lane definitions, modes, and safety rules live in the `omnilane` skill.
-Workers must never dispatch again (nested dispatch is refused, exit 86).
+Lane definitions, modes, per-model rows, and safety rules live in the
+`omnilane` skill — load it and apply the row for the model you are running
+as; legacy model-routing skill variants are retired. Workers must never
+dispatch again (nested dispatch is refused, exit 86).
 <!-- omnilane-routing:end -->

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnilane",
-  "version": "0.21.0",
+  "version": "0.32.0",
   "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.",
   "bin": {
     "omnilane": "bin/omnilane"

package/plugin.json

@@ -1,6 +1,6 @@
 {
   "$schema": "https://antigravity.google/schemas/v1/plugin.json",
   "name": "omnilane",
-  "version": "0.21.0",
+  "version": "0.32.0",
   "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."
 }

package/routing.local.yaml.example

@@ -8,7 +8,7 @@
 # consult is a multi-vendor direct-target chain. configure.sh intentionally
 # skips it because that menu writes one candidate per lane. If overriding it,
 # retain every vendor you want to address by name:
-# consult: codex gpt-5.6-sol max | claude claude-opus-5 high | grok grok-4.5 - | gemini "Gemini 3.1 Pro (High)" -
+# consult: codex gpt-5.6-sol max | claude claude-fable-5-1 high | grok grok-4.6 - | gemini "Gemini 3.7 Flash (High)" -
 
 # ── Starter profiles ─────────────────────────────────────────────
 # Uncomment ONE block that matches what you actually subscribe to.
@@ -26,10 +26,10 @@
 # live-search:     off - -
 # coding-overflow: off - -
 
-# Profile: Claude Code (Fable 5) as the main loop — let Fable keep judgment/taste,
+# Profile: Claude Code (Fable 5.1) as the main loop — let Fable keep judgment/taste,
 # push only coding volume out to Codex.
-# hard-judgment:   claude claude-fable-5 high
-# taste-final:     claude claude-fable-5 high
+# hard-judgment:   claude claude-fable-5-1 high
+# taste-final:     claude claude-fable-5-1 high
 
 # Profile: Codex-heavy (Sol main) — keep the hard lanes on Codex, Claude for taste.
 # taste-final:     claude claude-opus-5 high

package/routing.yaml

@@ -8,39 +8,37 @@
 #   same table degrades gracefully when you only subscribe to one or two vendors.
 # Override any line in ~/.omnilane/routing.local.yaml (same format; local wins).
 # Every benchmark score, price and throughput figure behind these orderings lives in
-# docs/model-capabilities-2026-07.md, with the date it was retrieved. The comments below
+# docs/model-capabilities-2026-09.md, with the date it was retrieved. The comments below
 # deliberately carry no numbers: they state WHY a lane is ordered the way it is, which
 # stays true for months, while the numbers move every few weeks. Change an ordering and
 # you update the doc; a figure going stale should never need a routing-table edit.
-# (Audited 2026-07-12; re-audited 2026-07-25, 2026-08-02 and 2026-08-03.)
-# defaults follow Artificial Analysis data, 2026-07
+# (Audited 2026-07-12; re-audited 2026-07-25, 2026-08-02, 2026-08-03, and 2026-09-02.)
+# defaults follow Artificial Analysis data, 2026-09
 # snapshot. Verified against AA site records + vendor pricing pages: Intelligence &
 # Coding indexes and 7:2:1 blended prices all match (AA field price1mBlended7To2To1);
 # coding cost-per-task is chart-only (not independently reconstructed). Prices are
 # standard short-context API tier — on subscription CLIs treat $ as relative ranking.
 # Your own job outcomes (~/.omnilane/jobs/) outrank these priors; edit lanes to match.
 
-hardest-coding:  codex gpt-5.6-sol xhigh | claude claude-opus-5 xhigh  # ordered on coding capability specifically, not general intelligence. Sol dropped from max to xhigh on 2026-08-03: on AA's per-effort Coding Index, Sol at xhigh outscores both Sol at max and every Claude tier, at a third less cost — max buys overthinking here, not accuracy. xhigh is also Anthropic's documented starting point for coding/agentic work. Keep Sol first for the established Codex harness lane.
-bulk-mechanical: codex gpt-5.6-terra max | claude claude-sonnet-5 high | gemini "Gemini 3.6 Flash (High)" -  # ordered on endurance per dollar: Terra leads Sonnet 5 on both intelligence and cost per task
-triage:          codex gpt-5.6-luna medium | gemini "Gemini 3.6 Flash (Low)" - | claude claude-haiku-4-5 -  # high-volume scans, ordered on cost per task: Luna is the cheapest model at its intelligence tier by a wide margin
-hard-judgment:   claude claude-opus-5 xhigh | codex gpt-5.6-sol max   # ordered on agentic knowledge work, where Opus 5 leads Sol on AA's benchmarks. xhigh per Anthropic guidance (high is the documented floor for intelligence-sensitive work; max is for correctness-over-cost only) — raise to max locally via `omnilane configure set` if your workload needs it.
-taste-final:     claude claude-opus-5 high | codex gpt-5.6-sol max    # user-facing prose, prompt/doc polish, Chinese phrasing, style arbitration
-consult:         codex gpt-5.6-sol max | claude claude-opus-5 high | grok grok-4.5 - | gemini "Gemini 3.1 Pro (High)" -  # direct named-model consultation; use --vendor to prevent fallback
-ui-draft:        codex gpt-5.6-sol xhigh | claude claude-opus-5 high  # only with a design system / reference images; open-ended visual taste -> taste-final
-long-context:    gemini "Gemini 3.1 Pro (High)" - | codex gpt-5.6-sol high | claude claude-opus-5 high  # all have 1M context; ordered on AA-LCR, which scores exactly this lane's work — extracting and synthesising across long documents — and where Gemini leads both fallbacks. Corrected 2026-08-03: this comment used to send multi-hop synthesis to the Claude candidate on second-hand prior-generation figures, and current first-party per-effort data reverses that, so the two fallbacks swapped. Caveat in docs: AA-LCR runs at 10k-100k tokens, so nothing here settles behaviour at a full 1M
-fast-agentic:    codex gpt-5.6-luna max | gemini "Gemini 3.6 Flash (High)" -  # fast multi-step tool loops. Reordered 2026-08-03: Luna leads Flash on agentic benchmarks AND costs a fraction as much per task, so Flash's remaining edge is raw throughput alone. Keep Flash first only if your loops are latency-bound. Both take image input, so the lane's multimodal checks are unaffected
-live-search:     grok grok-4.5 - | off          # native X/web search lane; no real substitute
-coding-overflow: grok grok-4.5 - | kimi kimi-k3 - | qwen qwen3-coder-plus - | opencode - - | off  # codex-quota relief valve: mid-tier coding; Grok 4.5 is a capable mid-tier coder but AA measures a high hallucination rate — verify every factual claim it ships. qwen3-coder-plus = 2025-09-23 snapshot alias (Qwen 3.6 Plus exists; re-evaluate before swapping). kimi/qwen model fields are CLI aliases — adjust to your login. opencode "-" model = its own configured default.
-arbitrate:       off - -                        # opinion panel is OPT-IN: it costs one call per voter per round.
+hardest-coding: claude claude-fable-5-1 xhigh | codex gpt-5.6-sol xhigh # ordered on coding capability: the leading Claude tier wins both coding components; Sol remains the established Codex-harness value fallback
+bulk-mechanical: codex gpt-5.6-sol high | gemini "Gemini 3.7 Flash (High)" - | claude claude-sonnet-5 high # ordered on endurance per dollar: Sol dominates Terra within Codex; current Flash is the faster, cheaper middle fallback; Sonnet preserves subscription quota
+triage: codex gpt-5.6-luna high | gemini "Gemini 3.7 Flash (Low)" - | claude claude-haiku-4-5 - # ordered on cost per task at usable intelligence: Luna high buys a meaningful quality lift cheaply; Flash and Haiku are low-cost cross-vendor fallbacks
+hard-judgment: claude claude-fable-5-1 xhigh | codex gpt-5.6-sol max | grok grok-4.6 - # ordered on agentic knowledge work: Fable leads the Claude field; Sol stays ahead of Grok because Grok effort is ignored and its reproduced row is unknown
+taste-final: claude claude-fable-5-1 high | codex gpt-5.6-sol max # ordered on prose and polish: Fable leads Opus on intelligence and factual breadth; Sol is the cross-vendor fallback
+consult: codex gpt-5.6-sol max | claude claude-fable-5-1 high | grok grok-4.6 - | gemini "Gemini 3.7 Flash (High)" - # direct named-model chain uses the strongest current Claude and Flash slots; keep --vendor to prevent fallback
+ui-draft: codex gpt-5.6-sol xhigh | claude claude-fable-5-1 high # ordered for drafts with a design system or reference images: Sol leads measured multimodal and coding evidence; Fable follows for polish
+long-context: gemini "Gemini 3.7 Flash (Medium)" - | codex gpt-5.6-terra max | claude claude-opus-5 medium # ordered on long-context reasoning, then cost and throughput: Flash leads; Terra matches its long-context result; Opus is the cheaper Claude fallback
+fast-agentic: gemini "Gemini 3.7 Flash (Medium)" - | codex gpt-5.6-luna high # ordered on interactive tool-loop latency: Flash gives up little agentic quality for far faster first output; Luna high is the low-latency Codex fallback
+live-search: grok grok-4.6 - | off # native X and web search lane; no real substitute
+coding-overflow: grok grok-4.6 - | gemini "Gemini 3.7 Flash (High)" - | kimi kimi-k3 - | qwen qwen3-coder-plus - | opencode - - | off # coding relief ordered by capability and value: Grok has the lowest frontier hallucination rate; Flash is the cheapest strong coder here; revisit the best-value Qwen tier when its CLI alias can be verified
+arbitrate: off - - # opinion panel remains opt-in because each voter and round consumes quota
                                                 # Enable: `arbitrate: vote codex,claude,grok -` (any 1-4 of codex/claude/grok/gemini)
                                                 # Debate round (each voter rebuts the others): set the effort field to 2.
                                                 # Custom gate: `arbitrate: exec /path/to/script -`
-# Claude Fable 5 (claude-fable-5) is deliberately absent from the defaults: the top Claude tier
-# is usually the MAIN LOOP itself, not a dispatched worker, and it prices at twice Opus 5.
-# This is a cost / guardrail / main-loop policy choice, NOT a capability verdict — AA and
-# Epoch AI disagree on which of the two leads general intelligence and call it effectively a
-# tie, while Opus 5 leads clearly on agentic knowledge work at a lower cost per task. Fable 5
-# does keep the lead on factual breadth, so name it explicitly for recall-heavy consults. If you want
-# to route to it anyway, pick it in the configurator or set e.g.
-#   taste-final: claude claude-fable-5 high
-# in ~/.omnilane/routing.local.yaml.
+# Claude Fable 5.1 is in the judgment, taste, and hardest-coding defaults because
+# it leads Opus 5 on every Artificial Analysis axis at the same effort.
+# It is not in bulk or triage: it prices at twice Opus 5 per token and consumes
+# the most subscription quota per turn. Opus 5 remains the lower-hallucination,
+# lower-price Claude choice and can return to any lane via
+# ~/.omnilane/routing.local.yaml, for example:
+# hard-judgment: claude claude-opus-5 xhigh

package/scripts/configure.sh

@@ -151,10 +151,10 @@ esac
 # Dynamic/API catalogs stay curated — "c" always accepts an exact model ID.
 CODEX_MODELS=("gpt-5.6" "gpt-5.6-sol" "gpt-5.6-terra" "gpt-5.6-luna" "gpt-5.5" "gpt-5.4" "gpt-5.4-mini" "gpt-5.3-codex-spark")
 CODEX_EFFORTS=("xhigh" "max" "ultra" "high" "medium" "low" "minimal" "none")
-CLAUDE_MODELS=("default" "best" "fable" "opus" "sonnet" "haiku" "opus[1m]" "sonnet[1m]" "opusplan" "claude-fable-5" "claude-opus-5" "claude-sonnet-5" "claude-opus-4-8" "claude-opus-4-7" "claude-opus-4-6" "claude-opus-4-5-20251101" "claude-sonnet-4-6" "claude-sonnet-4-5-20250929" "claude-haiku-4-5" "claude-haiku-4-5-20251001")
+CLAUDE_MODELS=("default" "best" "fable" "opus" "sonnet" "haiku" "opus[1m]" "sonnet[1m]" "opusplan" "claude-fable-5" "claude-fable-5-1" "claude-opus-5" "claude-sonnet-5" "claude-opus-4-8" "claude-opus-4-7" "claude-opus-4-6" "claude-opus-4-5-20251101" "claude-sonnet-4-6" "claude-sonnet-4-5-20250929" "claude-haiku-4-5" "claude-haiku-4-5-20251001")
 CLAUDE_EFFORTS=("max" "xhigh" "high" "medium" "low" "-")
-GEMINI_MODELS=("gemini-3.6-flash-high" "gemini-3.6-flash-medium" "gemini-3.6-flash-low" "gemini-3.5-flash-high" "gemini-3.5-flash-medium" "gemini-3.5-flash-low" "gemini-3.1-pro-high" "gemini-3.1-pro-low" "claude-sonnet-4-6" "claude-opus-4-6-thinking" "gpt-oss-120b-medium")
-GROK_MODELS=("grok-4.5" "headroom-grok-build" "grok-4.3-official")
+GEMINI_MODELS=("gemini-3.7-flash-high" "gemini-3.7-flash-medium" "gemini-3.7-flash-low" "gemini-3.6-flash-high" "gemini-3.6-flash-medium" "gemini-3.6-flash-low" "gemini-3.1-pro-high" "gemini-3.1-pro-low" "claude-sonnet-4-6" "claude-opus-4-6-thinking" "gpt-oss-120b-medium")
+GROK_MODELS=("grok-4.6" "headroom-grok-build" "grok-4.3-official")
 KIMI_MODELS=("kimi-k3" "kimi-k2.7-code" "kimi-k2.5")
 QWEN_MODELS=("qwen3.7-max" "qwen3.7-plus" "qwen3.6-plus" "qwen3.5-plus" "qwen3-max-2026-01-23" "qwen3-coder-next" "qwen3-coder-plus" "qwen3-coder-flash")
 # OpenCode models use provider/model form; OpenRouter models use catalog slugs.

package/scripts/doctor.sh

@@ -31,6 +31,7 @@ done
 REPO="${OMNILANE_DOCTOR_REPO:-$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)}"
 OMNILANE_HOME="${OMNILANE_HOME:-$HOME/.omnilane}"
 PROBE_SCRIPT="${OMNILANE_PROVIDER_PROBE_SCRIPT:-$REPO/scripts/provider-probe.sh}"
+GOAL_LOOP="${OMNILANE_DOCTOR_GOAL_LOOP:-$REPO/scripts/lib/goal-loop.sh}"
 PASS_COUNT=0
 WARN_COUNT=0
 FAIL_COUNT=0
@@ -123,6 +124,31 @@ else
   report PASS state "$OMNILANE_HOME is accessible"
 fi
 
+goals_store="$OMNILANE_HOME/goals"
+if [[ ! -f "$GOAL_LOOP" || -L "$GOAL_LOOP" ]]; then
+  report FAIL goal-orchestrator "$GOAL_LOOP must be a regular file"
+elif ! /bin/bash -n "$GOAL_LOOP" 2>/dev/null; then
+  report FAIL goal-orchestrator "$GOAL_LOOP has invalid bash syntax"
+elif [[ -L "$goals_store" || ( -e "$goals_store" && ! -d "$goals_store" ) ]]; then
+  report FAIL goal-orchestrator "$goals_store must be a real directory, not a symlink or file"
+elif [[ -d "$goals_store" ]]; then
+  goals_mode="$(stat -f '%Lp' "$goals_store" 2>/dev/null || true)"
+  if [[ ! "$goals_mode" =~ ^[0-7]{3,4}$ ]]; then
+    goals_mode="$(stat -c '%a' "$goals_store" 2>/dev/null || true)"
+  fi
+  if [[ "$goals_mode" =~ ^0?700$ ]]; then
+    report PASS goal-orchestrator "goal-loop.sh parses; goals store mode is $goals_mode"
+  elif [[ -n "$goals_mode" ]]; then
+    report WARN goal-orchestrator \
+      "goal-loop.sh parses; goals store mode is $goals_mode, owner-only 700 is safer"
+  else
+    report WARN goal-orchestrator \
+      "goal-loop.sh parses; could not determine goals store permissions"
+  fi
+else
+  report PASS goal-orchestrator "goal-loop.sh parses; goals store is not present yet"
+fi
+
 completion_plugin_state() {
   local config_dir state
   config_dir="${CLAUDE_CONFIG_DIR:-${HOME:-}/.claude}"