omnilane 0.44.0 → 0.45.0

package/.claude-plugin/marketplace.json

@@ -6,14 +6,14 @@
   },
   "metadata": {
     "description": "One routing table, native-first or CLI delegation, supervised completion workflows.",
-    "version": "0.44.0"
+    "version": "0.45.0"
   },
   "plugins": [
     {
       "name": "omnilane",
       "source": "./",
       "description": "Route each subtask to the model that fits, then delegate through a compatible caller-owned native agent or vendor CLI.",
-      "version": "0.44.0",
+      "version": "0.45.0",
       "category": "development",
       "keywords": [
         "routing",

package/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "omnilane",
-  "version": "0.44.0",
+  "version": "0.45.0",
   "description": "One routing table, every harness: classify subtasks into lanes and delegate through compatible caller-owned native agents or vendor CLIs with exact-AA downward policy and supervised jobs.",
   "author": {
     "name": "Seraphim0916"

package/CHANGELOG.md

@@ -6,6 +6,88 @@ semantic version tags.
 
 ## [Unreleased]
 
+## [0.45.0] - 2026-09-22
+
+The score registry moves to Artificial Analysis Intelligence Index v4.3.2, the
+lanes are re-allocated on it, and Grok 4.7 joins. **Every host must run
+`omnilane resign` once after upgrading**: the transport overlay is bound to the
+registry snapshot, and until it is rebuilt a model caller is refused on every
+lane with `transport overlay snapshot mismatch`.
+
+### Changed
+
+- The registry is re-scored from AA v4.3.2 (snapshot `aa-v4.3.2-2026-09-22-v1`).
+  v4.2 and v4.3.2 are different scales and the gaps are uneven (Fable 5.1 max
+  57 → 53, Astra xhigh 54 → 52, Opus 5 medium 50 → 45, Grok 4.6 high 51 → 44,
+  Sol high 48 → 42), so ceilings change: Fable 5.1 max, Fable 5.1 xhigh and
+  Astra max now tie at 53 and may dispatch to each other, and a mid-effort
+  controller that reached nothing in `hard-judgment` under v4.2 now does.
+  25 rows that were estimates are measured; 23 remain estimated. Each row
+  carries `score_raw`, and `schema_notes.score_rounding` records the half-up
+  rule.
+- `routing.yaml` is rewritten rather than amended. Each lane names the
+  measurements that match its kind of work and lists candidates best-first on
+  them, and each chain steps down through the score range so that whatever a
+  controller's ceiling, the first candidate it can reach is the best one it can
+  reach. Eight lanes change their first choice:
+  - `hardest-coding`: GPT-6 Astra (xhigh), on Terminal-Bench 4.0 (AA moved to it
+    in v4.3; the frontier is saturated on 2.1) and the lowest hallucination rate
+    of the top rows. Fable (xhigh) follows; it beats Fable (max) on 4.0 for about
+    two thirds of the cost, so `max` is in no chain and stays explicit-only.
+  - `bulk-mechanical` and `fast-agentic`: Astra (low). It costs what Sol (high)
+    costs and is ahead on coding, speed and hallucination; it is the fastest row
+    that still scores near the top on AutomationBench. Claude Haiku leaves
+    `fast-agentic`, where it scores close to nothing.
+  - `taste-final`: Claude Opus 5 (max), on AA's expert grading of finished
+    documents. Grok 4.7 is graded above every Astra row there and precedes Astra.
+  - `long-context`: Claude Opus 5 (high, then medium, then low). AA-LCR is
+    saturated; on the harder long-context figure Claude leads every other family
+    by a wide margin at every effort.
+  - `ui-draft`: Astra (high), on MMMU-Pro then Terminal-Bench 4.0.
+  - `live-search` and `coding-overflow`: Grok 4.7.
+  `hard-judgment` keeps Fable (xhigh) then Astra (xhigh) and gains Opus rungs,
+  which hold their graded analytical quality down the effort ladder where Astra
+  does not. The figures behind every ordering, generated from the tracked
+  extract, are in `docs/model-capabilities-2026-09.md`.
+
+### Added
+
+- Grok 4.7: registry rows at xhigh and high, a catalog entry in
+  `omnilane configure`, and a place ahead of Grok 4.6 in every lane Grok serves.
+  A lane skips a candidate whose transport this host has not proven, so Grok 4.6
+  keeps serving until `omnilane resign` has probed 4.7.
+- `scripts/aa_rebaseline.py lanes` prints, per lane, each candidate with the
+  measurements that lane is ordered on.
+- Claude Sonnet 5 at xhigh, high, medium and low. AA now scores them; they were
+  `unknown_configs`.
+- `scripts/aa_rebaseline.py` (`fetch`, `build`, `report`, `matrix`) regenerates
+  the registry from a saved extract of one AA page, byte-reproducibly. The
+  extract and per-vendor evidence for this snapshot are tracked under
+  `docs/reports/`.
+
+### Fixed
+
+- `omnilane resign` now treats a re-scored registry as drift. It compared only
+  executables and runner scripts, so on a host where no CLI had moved it
+  reported "nothing to re-sign" and left an overlay that dispatch refuses for
+  good. With the executables unchanged it reuses the existing probe evidence,
+  probes only the rows that have none, and rebuilds the overlay for the new
+  snapshot; a vendor whose CLI also moved is still probed in full.
+- `claude claude-sonnet-5 high` was a fallback that could never be dispatched:
+  the v4.2 registry had only a non-reasoning row at that effort, which Claude's
+  `--effort` cannot select. It now resolves to the adaptive row once probed.
+
+### Removed
+
+- `codex/gpt-6-astra-non-reasoning`. AA no longer lists it and Astra rejects
+  effort `none`; it moves to `unknown_configs` and its old score is not kept.
+
+### Known gap until `omnilane resign`
+
+- `--vendor grok` is refused on every lane (`runtime-mapping-unverified`): with
+  `--vendor` a lane takes that vendor's first segment without falling through,
+  and it is now 4.7. Add `--model grok-4.6` until the host has proven 4.7.
+
 ## [0.44.0] - 2026-09-21
 
 The daily `omnilane resign` now also covers a CLI you patch and re-sign
@@ -1242,7 +1324,8 @@ work to the wrong model, and records the evidence behind the shipped defaults.
 - Initial shared routing table, cross-vendor dispatcher, runners, installer,
   and baseline lint fixes.
 
-[Unreleased]: https://github.com/Seraphim0916/omnilane/compare/v0.44.0...HEAD
+[Unreleased]: https://github.com/Seraphim0916/omnilane/compare/v0.45.0...HEAD
+[0.45.0]: https://github.com/Seraphim0916/omnilane/compare/v0.44.0...v0.45.0
 [0.44.0]: https://github.com/Seraphim0916/omnilane/compare/v0.43.1...v0.44.0
 [0.43.1]: https://github.com/Seraphim0916/omnilane/compare/v0.43.0...v0.43.1
 [0.43.0]: https://github.com/Seraphim0916/omnilane/compare/v0.42.9...v0.43.0

package/README.ja.md

@@ -220,17 +220,17 @@ flowchart LR
 
 | レーン | 第一候補 | バックアップ | 用途 |
 |---|---|---|---|
-| 🔥 hardest-coding | Claude Fable 5.1 (max) | GPT-6 Astra (xhigh) → Grok 4.6 → Gemini 3.8 Flash (High) | 最難関の実装、深い根本原因調査、正確性が重要な修正 |
-| 🏗️ bulk-mechanical | GPT-5.6 Sol (high) | Gemini 3.8 Flash (High) → Claude Sonnet 5 (high) | リファクタリング、移行、テスト、大規模レビュー——機械的な持久作業 |
-| 🧹 triage | GPT-5.6 Luna (high) | Gemini 3.8 Flash (Low) → Claude Haiku 4.5 | 大量スキャン、一次選別 |
-| ⚖️ hard-judgment | Claude Fable 5.1 (xhigh) | GPT-6 Astra (xhigh) → Grok 4.6 | アーキテクチャ判断、深い推論、セカンドオピニオン |
-| ✒️ taste-final | Claude Fable 5.1 (xhigh) | GPT-6 Astra (xhigh) → Grok 4.6 → Gemini 3.8 Flash (High) | ユーザー向け文章、プロンプト／文書の仕上げ、文体判断 |
-| 💬 consult | GPT-6 Astra (xhigh) | Claude Fable 5.1 (xhigh) → Grok 4.6 → Gemini 3.8 Flash (Medium) | 指名モデルへの直接相談。フォールバック防止のため `--vendor` を維持 |
-| 🎨 ui-draft | GPT-5.6 Sol (high) | Claude Fable 5.1 (xhigh) → Gemini 3.8 Flash (High) | デザインシステム／参照画像がある場合だけの UI ドラフト |
-| 📚 long-context | Gemini 3.8 Flash (Medium) | GPT-5.6 Terra (max) → Claude Opus 5 (medium) | 長文書の抽出と統合。AA-LCR、コスト、スループット順 |
-| ⚡ fast-agentic | Gemini 3.8 Flash (Low) | GPT-5.6 Luna (high) → Claude Haiku 4.5 | 高速なマルチステップ agentic ループ、マルチモーダル確認 |
-| 📡 live-search | Grok 4.6 | Gemini 3.8 Flash (High) → Claude Sonnet 5 (high) | リアルタイム X／Web 検索とソーシャル文脈 |
-| 🚰 coding-overflow | Grok 4.6 | Gemini 3.8 Flash (High) → Kimi K3 → Qwen3 Coder Plus → OpenCode | Codex クォータ不足時の中級コーディング逃がし弁 |
+| 🔥 hardest-coding | GPT-6 Astra (xhigh) | Claude Fable 5.1 (xhigh) → GPT-6 Astra (high) → Claude Fable 5.1 (high) → GPT-6 Astra (medium) → Claude Opus 5 (high) → GPT-6 Astra (low) → GPT-5.6 Sol (xhigh) → Grok 4.7 → Grok 4.6 → Gemini 3.8 Flash (High) | 最難関の実装、深い根本原因調査、正確性が重要な修正 |
+| 🏗️ bulk-mechanical | GPT-6 Astra (low) | GPT-5.6 Sol (high) → Gemini 3.8 Flash (High) → Claude Opus 5 (medium) | リファクタリング、移行、テスト、大規模レビュー——機械的な持久作業 |
+| 🧹 triage | GPT-5.6 Luna (high) | Gemini 3.8 Flash (Low) → Claude Sonnet 5 (low) → Claude Haiku 4.5 | 大量スキャン、一次選別 |
+| ⚖️ hard-judgment | Claude Fable 5.1 (xhigh) | GPT-6 Astra (xhigh) → Claude Opus 5 (max) → GPT-6 Astra (high) → Claude Opus 5 (xhigh) → Claude Opus 5 (high) → Grok 4.7 → Grok 4.6 → Gemini 3.8 Flash (High) | アーキテクチャ判断、深い推論、セカンドオピニオン |
+| ✒️ taste-final | Claude Opus 5 (max) | Claude Fable 5.1 (xhigh) → Claude Opus 5 (xhigh) → Grok 4.7 → GPT-6 Astra (xhigh) → Claude Opus 5 (high) → Grok 4.6 → Gemini 3.8 Flash (High) | ユーザー向け文章、プロンプト／文書の仕上げ、文体判断 |
+| 💬 consult | GPT-6 Astra (xhigh) | Claude Fable 5.1 (xhigh) → Grok 4.7 → Grok 4.6 → Gemini 3.8 Flash (High) | 指名モデルへの直接相談。フォールバック防止のため `--vendor` を維持 |
+| 🎨 ui-draft | GPT-6 Astra (high) | Claude Opus 5 (high) → GPT-6 Astra (low) → Gemini 3.8 Flash (High) | デザインシステム／参照画像がある場合だけの UI ドラフト |
+| 📚 long-context | Claude Opus 5 (high) | Claude Opus 5 (medium) → Claude Opus 5 (low) → GPT-5.6 Terra (max) → Gemini 3.8 Flash (High) | 長文書の抽出と統合。AA-LCR、コスト、スループット順 |
+| ⚡ fast-agentic | GPT-6 Astra (low) | Gemini 3.8 Flash (Medium) → GPT-5.6 Sol (medium) → Claude Opus 5 (low) | 高速なマルチステップ agentic ループ、マルチモーダル確認 |
+| 📡 live-search | Grok 4.7 | Grok 4.6 → Gemini 3.8 Flash (High) → Claude Opus 5 (medium) | リアルタイム X／Web 検索とソーシャル文脈 |
+| 🚰 coding-overflow | Grok 4.7 | Grok 4.6 → Gemini 3.8 Flash (High) → Kimi K3 → Qwen3 Coder Plus → OpenCode | Codex クォータ不足時の中級コーディング逃がし弁 |
 | 🗳️ arbitrate | off (opt-in vote panel) | — | 重大判断用の内蔵意見パネル。デフォルト無効、`routing.local.yaml` で有効化し、投票者・ラウンドごとに 1 コール |
 
 **バックアップ**はチェーンの次の候補——第一候補のベンダー CLI が未インストールの
@@ -267,7 +267,7 @@ flowchart LR
 - **Claude Code · Opus 5**——自分で実行：hard-judgment(これがデフォルトのレーン)。低いハルシネーション率や価格を優先するときはローカル override で taste-final も担当。最難関コーディング → Fable 5.1 または Sol、bulk → Sol high、long-context／高速ループ → Gemini 3.7 Flash、live-search → Grok。
 - **Codex · Sol**——自分で実行：hardest-coding、bulk-mechanical、hard-judgment、ui-draft。ディスパッチ：taste-final → Claude、long-context／高速ループ → Gemini 3.7 Flash、live-search → Grok。
 - **Codex · Terra**——long-context の Codex フォールバックを自分で実行。bulk-mechanical のデフォルトは Sol high に移動。最難関は Sol xhigh、taste → Claude、高速ループ → Gemini 3.7 Flash、live-search → Grok。
-- **Grok Build · Grok 4.6**——live-search と coding-overflow を自分で実行し、hardest-coding・hard-judgment・taste-final のフォールバックも兼任。第一候補が使えるときは難しいコーディング／判断／文章を Codex、Claude、Gemini へ送り、API シグネチャと引用事実は検証します。
+- **Grok Build · Grok 4.7 / 4.6**——live-search と coding-overflow を自分で実行し、hardest-coding・hard-judgment・taste-final のフォールバックも兼任。第一候補が使えるときは難しいコーディング／判断／文章を Codex、Claude、Gemini へ送り、API シグネチャと引用事実は検証します。
 - **Antigravity · Gemini 3.7 Flash**——Medium の long-context／高速ループ、High の bulk／overflow、Low の triage を自分で実行し、High で hardest-coding・taste-final・ui-draft・live-search のフォールバックも兼任。第一候補が使えるときは最難関のコーディング／判断／文章を Codex、Claude へ。
 
 </details>
@@ -702,6 +702,15 @@ work の別名ではありません。サービス管理など、work の境界
 
 ## 📜 リリース履歴
 
+## v0.45.0 の新機能
+
+- **アップグレード後に `omnilane resign` を一度実行してください。** スコアのレジストリが新しいスナップショットに移り、トランスポートオーバーレイはスナップショットに紐づいています。オーバーレイを再構築するまで、モデル呼び出し元はすべてのレーンで `transport overlay snapshot mismatch` として拒否されます。
+- **スコアは Artificial Analysis Intelligence Index v4.3.2 に準拠。** v4.2 と v4.3.2 は別の尺度で、差も一様ではありません（Fable 5.1 max 57 → 53、Grok 4.6 high 51 → 44、Sol high 48 → 42）。そのため 1 行を追加するのではなく全行を再採点しました。上限も変わります。Fable 5.1 max・Fable 5.1 xhigh・Astra max は 53 で並び、これまで `hard-judgment` で何にも届かなかった中程度 effort のコントローラーも届くようになりました。`scripts/aa_rebaseline.py` は保存した AA の抽出ファイルからレジストリを再構築するので、次の指数改訂は再実行で済みます。
+- **レーン表は手直しではなく書き直しです。** 各レーンはまず「この種の作業に合う測定」を定め、候補をその測定で強い順に並べます。`hardest-coding` は高難度のターミナルコーディング（Terminal-Bench 4.0、最上位モデルでもまだ飽和していないコーディング評価）、`hard-judgment` は専門家レベルの推論と文書の分析品質の採点、`taste-final` は文書の提示品質の採点、`fast-agentic` は自動化の正確さと 1 ステップあたりの時間、`long-context` はより難しい方の長文脈評価、`ui-draft` は画像理解を見ます。8 つのレーンで第一候補が変わりました。`hardest-coding` は Astra (xhigh)、`bulk-mechanical` と `fast-agentic` は Astra (low)、`ui-draft` は Astra (high)、`taste-final` は Opus 5 (max)、`long-context` は Opus 5 (high)、`coding-overflow` は `live-search` と同じく Grok 4.7 が先頭です。どのチェーンもスコアの範囲を順に下っていくので、コントローラーの上限がどこにあっても、最初に届く候補が届く中で最良の候補になります。Fable max と Astra max はどのチェーンにも入っていません。xhigh が同等かそれ以上で、コストは大幅に小さいためです。数値は `docs/model-capabilities-2026-09.md` にあります。
+- **Grok が担当するすべてのレーンに Grok 4.7 を追加**し、Grok 4.6 の前に置きました。4.6 はその後ろに残ります。レーンはこのホストで未検証の候補を飛ばすため、`resign` が 4.7 をプローブするまでは 4.6 が処理します。`--vendor grok` を付けた場合は後ろへ進みません。それまでは `--model grok-4.6` を指定してください。
+- **機能していなかったフォールバックを修正。** `claude claude-sonnet-5 high` は旧レジストリでは決してディスパッチできませんでした。プローブ後は解決されます。
+- アップグレード: `npm i -g omnilane@0.45.0`、続けて `omnilane resign`。Grok 4.7 を使うには、再署名時に `grok` CLI がログイン済みである必要があります。
+
 ## v0.44.0 の新機能
 
 - **自分でパッチした CLI も無人で再署名されます。** 更新のたびにベンダー CLI をパッチして adhoc で再署名するローカルの手順がある場合、これまでは署名者チェックがそのような更新をすべて `--approve` 待ちで止めていました。ベンダーごとに一度 `omnilane resign --trust-adhoc VENDOR` を実行すれば、同じインストールディレクトリ内の adhoc な更新は同一署名者の更新と同じく毎日の `omnilane resign` で自動的に再署名されます。未署名の実行ファイル、別ディレクトリの adhoc、ほかのベンダーは引き続き停止して確認を求めます。信頼はオーバーレイに記録され、以後どのベンダーの再署名でも保持され、モデルが実行しない運用者の操作です。4 ベンダーすべてに対応。

package/README.ko.md

@@ -218,17 +218,17 @@ flowchart LR
 
 | 레인 | 1순위 모델 | 백업 | 용도 |
 |---|---|---|---|
-| 🔥 hardest-coding | Claude Fable 5.1 (max) | GPT-6 Astra (xhigh) → Grok 4.6 → Gemini 3.8 Flash (High) | 가장 어려운 구현, 근본 원인 디버깅, 정확성이 핵심인 수정 |
-| 🏗️ bulk-mechanical | GPT-5.6 Sol (high) | Gemini 3.8 Flash (High) → Claude Sonnet 5 (high) | 리팩터링, 마이그레이션, 테스트, 대량 스윕——기계적 지구력 작업 |
-| 🧹 triage | GPT-5.6 Luna (high) | Gemini 3.8 Flash (Low) → Claude Haiku 4.5 | 대량 스캔과 1차 선별 |
-| ⚖️ hard-judgment | Claude Fable 5.1 (xhigh) | GPT-6 Astra (xhigh) → Grok 4.6 | 아키텍처 판정, 심층 추론, 2차 의견 |
-| ✒️ taste-final | Claude Fable 5.1 (xhigh) | GPT-6 Astra (xhigh) → Grok 4.6 → Gemini 3.8 Flash (High) | 사용자 대상 문장, 프롬프트／문서 다듬기, 스타일 판정 |
-| 💬 consult | GPT-6 Astra (xhigh) | Claude Fable 5.1 (xhigh) → Grok 4.6 → Gemini 3.8 Flash (Medium) | 지정 모델 직접 상담. 폴백 방지를 위해 `--vendor` 유지 |
-| 🎨 ui-draft | GPT-5.6 Sol (high) | Claude Fable 5.1 (xhigh) → Gemini 3.8 Flash (High) | 디자인 시스템／참조 이미지가 있을 때만 UI 초안 |
-| 📚 long-context | Gemini 3.8 Flash (Medium) | GPT-5.6 Terra (max) → Claude Opus 5 (medium) | 장문 추출과 종합. AA-LCR, 비용, 처리량 순 |
-| ⚡ fast-agentic | Gemini 3.8 Flash (Low) | GPT-5.6 Luna (high) → Claude Haiku 4.5 | 빠른 멀티스텝 agentic 루프, 멀티모달 확인 |
-| 📡 live-search | Grok 4.6 | Gemini 3.8 Flash (High) → Claude Sonnet 5 (high) | 실시간 X／웹 검색과 소셜 맥락 |
-| 🚰 coding-overflow | Grok 4.6 | Gemini 3.8 Flash (High) → Kimi K3 → Qwen3 Coder Plus → OpenCode | Codex 쿼터 소진 시 중급 코딩 안전 밸브 |
+| 🔥 hardest-coding | GPT-6 Astra (xhigh) | Claude Fable 5.1 (xhigh) → GPT-6 Astra (high) → Claude Fable 5.1 (high) → GPT-6 Astra (medium) → Claude Opus 5 (high) → GPT-6 Astra (low) → GPT-5.6 Sol (xhigh) → Grok 4.7 → Grok 4.6 → Gemini 3.8 Flash (High) | 가장 어려운 구현, 근본 원인 디버깅, 정확성이 핵심인 수정 |
+| 🏗️ bulk-mechanical | GPT-6 Astra (low) | GPT-5.6 Sol (high) → Gemini 3.8 Flash (High) → Claude Opus 5 (medium) | 리팩터링, 마이그레이션, 테스트, 대량 스윕——기계적 지구력 작업 |
+| 🧹 triage | GPT-5.6 Luna (high) | Gemini 3.8 Flash (Low) → Claude Sonnet 5 (low) → Claude Haiku 4.5 | 대량 스캔과 1차 선별 |
+| ⚖️ hard-judgment | Claude Fable 5.1 (xhigh) | GPT-6 Astra (xhigh) → Claude Opus 5 (max) → GPT-6 Astra (high) → Claude Opus 5 (xhigh) → Claude Opus 5 (high) → Grok 4.7 → Grok 4.6 → Gemini 3.8 Flash (High) | 아키텍처 판정, 심층 추론, 2차 의견 |
+| ✒️ taste-final | Claude Opus 5 (max) | Claude Fable 5.1 (xhigh) → Claude Opus 5 (xhigh) → Grok 4.7 → GPT-6 Astra (xhigh) → Claude Opus 5 (high) → Grok 4.6 → Gemini 3.8 Flash (High) | 사용자 대상 문장, 프롬프트／문서 다듬기, 스타일 판정 |
+| 💬 consult | GPT-6 Astra (xhigh) | Claude Fable 5.1 (xhigh) → Grok 4.7 → Grok 4.6 → Gemini 3.8 Flash (High) | 지정 모델 직접 상담. 폴백 방지를 위해 `--vendor` 유지 |
+| 🎨 ui-draft | GPT-6 Astra (high) | Claude Opus 5 (high) → GPT-6 Astra (low) → Gemini 3.8 Flash (High) | 디자인 시스템／참조 이미지가 있을 때만 UI 초안 |
+| 📚 long-context | Claude Opus 5 (high) | Claude Opus 5 (medium) → Claude Opus 5 (low) → GPT-5.6 Terra (max) → Gemini 3.8 Flash (High) | 장문 추출과 종합. AA-LCR, 비용, 처리량 순 |
+| ⚡ fast-agentic | GPT-6 Astra (low) | Gemini 3.8 Flash (Medium) → GPT-5.6 Sol (medium) → Claude Opus 5 (low) | 빠른 멀티스텝 agentic 루프, 멀티모달 확인 |
+| 📡 live-search | Grok 4.7 | Grok 4.6 → Gemini 3.8 Flash (High) → Claude Opus 5 (medium) | 실시간 X／웹 검색과 소셜 맥락 |
+| 🚰 coding-overflow | Grok 4.7 | Grok 4.6 → Gemini 3.8 Flash (High) → Kimi K3 → Qwen3 Coder Plus → OpenCode | Codex 쿼터 소진 시 중급 코딩 안전 밸브 |
 | 🗳️ arbitrate | off (opt-in vote panel) | — | 중대한 판단을 위한 내장 의견 패널. 기본 비활성, `routing.local.yaml` 에서 활성화하며 투표자·라운드당 1회 호출 |
 
 **백업**은 체인의 다음 후보입니다——1순위 벤더 CLI 가 설치되지 않았을 때
@@ -265,7 +265,7 @@ flowchart LR
 - **Claude Code · Opus 5**——직접 실행: hard-judgment(기본 레인). 더 낮은 환각률이나 가격이 중요할 때는 로컬 오버라이드로 taste-final 도 맡을 수 있습니다. 최고난도 코딩 → Fable 5.1 또는 Sol, bulk → Sol high, long-context／빠른 루프 → Gemini 3.7 Flash, live-search → Grok.
 - **Codex · Sol**——직접 실행: hardest-coding, bulk-mechanical, hard-judgment, ui-draft. 디스패치: taste-final → Claude, long-context／빠른 루프 → Gemini 3.7 Flash, live-search → Grok.
 - **Codex · Terra**——long-context 의 Codex 폴백을 직접 실행. bulk-mechanical 기본값은 Sol high 로 이동했습니다. 최고난도는 Sol xhigh, taste → Claude, 빠른 루프 → Gemini 3.7 Flash, live-search → Grok.
-- **Grok Build · Grok 4.6**——live-search 와 coding-overflow 를 직접 실행하며, hardest-coding, hard-judgment, taste-final 의 폴백도 겸합니다. 1순위 후보를 쓸 수 있으면 어려운 코딩／판단／문장은 Codex, Claude, Gemini 로 보내고 API 시그니처와 인용 사실을 검증합니다.
+- **Grok Build · Grok 4.7 / 4.6**——live-search 와 coding-overflow 를 직접 실행하며, hardest-coding, hard-judgment, taste-final 의 폴백도 겸합니다. 1순위 후보를 쓸 수 있으면 어려운 코딩／판단／문장은 Codex, Claude, Gemini 로 보내고 API 시그니처와 인용 사실을 검증합니다.
 - **Antigravity · Gemini 3.7 Flash**——Medium 의 long-context／빠른 루프, High 의 bulk／overflow, Low 의 triage 를 직접 실행하며, High 로 hardest-coding, taste-final, ui-draft, live-search 의 폴백도 겸합니다. 1순위 후보를 쓸 수 있으면 최고난도 코딩／판단／문장은 Codex, Claude 로 보냅니다.
 
 </details>
@@ -687,6 +687,15 @@ doctor가 파일과 벤더를 지목하며, 재서명 절차는 디스패치 스
 
 ## 📜 릴리스 기록
 
+## v0.45.0 새 기능
+
+- **업그레이드 후 `omnilane resign` 을 한 번 실행하세요.** 점수 레지스트리가 새 스냅샷으로 바뀌었고 트랜스포트 오버레이는 스냅샷에 묶여 있습니다. 오버레이를 다시 만들기 전까지 모델 호출자는 모든 레인에서 `transport overlay snapshot mismatch` 로 거부됩니다.
+- **점수는 Artificial Analysis Intelligence Index v4.3.2 를 따릅니다.** v4.2 와 v4.3.2 는 서로 다른 척도이고 차이도 고르지 않습니다(Fable 5.1 max 57 → 53, Grok 4.6 high 51 → 44, Sol high 48 → 42). 그래서 한 행을 추가하지 않고 모든 행을 다시 채점했습니다. 상한도 함께 바뀝니다. Fable 5.1 max, Fable 5.1 xhigh, Astra max 가 53 으로 같아졌고, 이전에는 `hard-judgment` 에서 아무것도 닿지 않던 중간 effort 컨트롤러도 이제 닿습니다. `scripts/aa_rebaseline.py` 는 저장해 둔 AA 추출 파일로 레지스트리를 다시 만들기 때문에 다음 지수 개정은 재실행으로 끝납니다.
+- **레인 표는 손질이 아니라 다시 썼습니다.** 각 레인은 먼저 "이런 종류의 작업에 맞는 측정"을 정하고, 후보를 그 측정에서 강한 순서로 나열합니다. `hardest-coding` 은 고난도 터미널 코딩(Terminal-Bench 4.0, 최상위 모델도 아직 포화시키지 못한 코딩 평가), `hard-judgment` 는 전문가 수준의 추론과 문서의 분석 품질 채점, `taste-final` 은 문서의 표현 품질 채점, `fast-agentic` 은 자동화 정확도와 단계당 시간, `long-context` 는 더 어려운 쪽의 긴 컨텍스트 평가, `ui-draft` 는 이미지 이해를 봅니다. 여덟 레인의 첫 번째 선택이 바뀌었습니다. `hardest-coding` 은 Astra (xhigh), `bulk-mechanical` 과 `fast-agentic` 은 Astra (low), `ui-draft` 는 Astra (high), `taste-final` 은 Opus 5 (max), `long-context` 는 Opus 5 (high), `coding-overflow` 는 `live-search` 와 마찬가지로 Grok 4.7 이 맨 앞입니다. 모든 체인은 점수 범위를 차례로 내려가므로, 컨트롤러의 상한이 어디에 있든 처음 닿는 후보가 닿을 수 있는 최선의 후보입니다. Fable max 와 Astra max 는 어떤 체인에도 없습니다. xhigh 가 같거나 앞서면서 비용은 훨씬 적기 때문입니다. 수치는 `docs/model-capabilities-2026-09.md` 에 있습니다.
+- **Grok 이 맡는 모든 레인에 Grok 4.7 을 추가**해 Grok 4.6 앞에 두었고, 4.6 은 그 뒤에 남습니다. 레인은 이 호스트에서 검증되지 않은 후보를 건너뛰므로 `resign` 이 4.7 을 프로브하기 전까지는 4.6 이 처리합니다. `--vendor grok` 을 붙이면 뒤로 넘어가지 않습니다. 그 전까지는 `--model grok-4.6` 을 지정하세요.
+- **동작하지 않던 폴백 수정.** `claude claude-sonnet-5 high` 는 이전 레지스트리에서는 결코 디스패치될 수 없었습니다. 이제 프로브 후에는 해석됩니다.
+- 업그레이드: `npm i -g omnilane@0.45.0`, 이어서 `omnilane resign`. Grok 4.7 을 쓰려면 재서명 시 `grok` CLI 가 로그인 상태여야 합니다.
+
 ## v0.44.0 새 기능
 
 - **직접 패치한 CLI 도 무인으로 재서명됩니다.** 업데이트 때마다 벤더 CLI 를 패치하고 adhoc 으로 다시 서명하는 로컬 단계가 있다면, 지금까지는 서명자 검사가 그런 업데이트를 모두 `--approve` 대기로 막았습니다. 벤더별로 한 번 `omnilane resign --trust-adhoc VENDOR` 를 실행하면 같은 설치 디렉터리 안의 adhoc 업데이트는 같은 서명자의 업데이트처럼 매일의 `omnilane resign` 으로 자동 재서명됩니다. 서명되지 않은 실행 파일, 다른 디렉터리의 adhoc, 다른 벤더는 여전히 멈추고 확인을 기다립니다. 신뢰는 오버레이에 기록되고 이후 어떤 벤더의 재서명에서도 유지되며, 모델이 실행하지 않는 운영자 작업입니다. 네 벤더 모두 지원.

package/README.md

@@ -256,17 +256,17 @@ actually resolves.
 
 | Lane | First choice | Backup | When |
 |---|---|---|---|
-| 🔥 hardest-coding | Claude Fable 5.1 (max) | GPT-6 Astra (xhigh) → Grok 4.6 → Gemini 3.8 Flash (High) | Hardest implementation, deep root-cause debug, correctness-critical edits |
-| 🏗️ bulk-mechanical | GPT-5.6 Sol (high) | Gemini 3.8 Flash (High) → Claude Sonnet 5 (high) | Refactors, migrations, tests, review sweeps — mechanical endurance |
-| 🧹 triage | GPT-5.6 Luna (high) | Gemini 3.8 Flash (Low) → Claude Haiku 4.5 | High-volume scans, first-pass filtering |
-| ⚖️ hard-judgment | Claude Fable 5.1 (xhigh) | GPT-6 Astra (xhigh) → Grok 4.6 | Architecture arbitration, deep reasoning, second opinions |
-| ✒️ taste-final | Claude Fable 5.1 (xhigh) | GPT-6 Astra (xhigh) → Grok 4.6 → Gemini 3.8 Flash (High) | User-facing prose and style arbitration; benchmarks do not prove visual or editorial taste |
-| 💬 consult | GPT-6 Astra (xhigh) | Claude Fable 5.1 (xhigh) → Grok 4.6 → Gemini 3.8 Flash (Medium) | Direct named-model consultation; keep `--vendor` to prevent fallback |
-| 🎨 ui-draft | GPT-5.6 Sol (high) | Claude Fable 5.1 (xhigh) → Gemini 3.8 Flash (High) | UI drafts only with a design system or reference images; no aesthetic benchmark claim |
-| 📚 long-context | Gemini 3.8 Flash (Medium) | GPT-5.6 Terra (max) → Claude Opus 5 (medium) | Long-document synthesis; context size alone does not prove task quality |
-| ⚡ fast-agentic | Gemini 3.8 Flash (Low) | GPT-5.6 Luna (high) → Claude Haiku 4.5 | Fast multi-step tool loops and multimodal checks |
-| 📡 live-search | Grok 4.6 | Gemini 3.8 Flash (High) → Claude Sonnet 5 (high) | Realtime X/web search; backups provide generic web search, not equivalent X context |
-| 🚰 coding-overflow | Grok 4.6 | Gemini 3.8 Flash (High) → Kimi K3 → Qwen3 Coder Plus → OpenCode | Explicit Codex-quota relief; provider failure does not auto-retry another vendor |
+| 🔥 hardest-coding | GPT-6 Astra (xhigh) | Claude Fable 5.1 (xhigh) → GPT-6 Astra (high) → Claude Fable 5.1 (high) → GPT-6 Astra (medium) → Claude Opus 5 (high) → GPT-6 Astra (low) → GPT-5.6 Sol (xhigh) → Grok 4.7 → Grok 4.6 → Gemini 3.8 Flash (High) | Hardest implementation, deep root-cause debug, correctness-critical edits |
+| 🏗️ bulk-mechanical | GPT-6 Astra (low) | GPT-5.6 Sol (high) → Gemini 3.8 Flash (High) → Claude Opus 5 (medium) | Refactors, migrations, tests, review sweeps — mechanical endurance |
+| 🧹 triage | GPT-5.6 Luna (high) | Gemini 3.8 Flash (Low) → Claude Sonnet 5 (low) → Claude Haiku 4.5 | High-volume scans, first-pass filtering |
+| ⚖️ hard-judgment | Claude Fable 5.1 (xhigh) | GPT-6 Astra (xhigh) → Claude Opus 5 (max) → GPT-6 Astra (high) → Claude Opus 5 (xhigh) → Claude Opus 5 (high) → Grok 4.7 → Grok 4.6 → Gemini 3.8 Flash (High) | Architecture arbitration, deep reasoning, second opinions |
+| ✒️ taste-final | Claude Opus 5 (max) | Claude Fable 5.1 (xhigh) → Claude Opus 5 (xhigh) → Grok 4.7 → GPT-6 Astra (xhigh) → Claude Opus 5 (high) → Grok 4.6 → Gemini 3.8 Flash (High) | User-facing prose and style arbitration; benchmarks do not prove visual or editorial taste |
+| 💬 consult | GPT-6 Astra (xhigh) | Claude Fable 5.1 (xhigh) → Grok 4.7 → Grok 4.6 → Gemini 3.8 Flash (High) | Direct named-model consultation; keep `--vendor` to prevent fallback |
+| 🎨 ui-draft | GPT-6 Astra (high) | Claude Opus 5 (high) → GPT-6 Astra (low) → Gemini 3.8 Flash (High) | UI drafts only with a design system or reference images; no aesthetic benchmark claim |
+| 📚 long-context | Claude Opus 5 (high) | Claude Opus 5 (medium) → Claude Opus 5 (low) → GPT-5.6 Terra (max) → Gemini 3.8 Flash (High) | Long-document synthesis; context size alone does not prove task quality |
+| ⚡ fast-agentic | GPT-6 Astra (low) | Gemini 3.8 Flash (Medium) → GPT-5.6 Sol (medium) → Claude Opus 5 (low) | Fast multi-step tool loops and multimodal checks |
+| 📡 live-search | Grok 4.7 | Grok 4.6 → Gemini 3.8 Flash (High) → Claude Opus 5 (medium) | Realtime X/web search; backups provide generic web search, not equivalent X context |
+| 🚰 coding-overflow | Grok 4.7 | Grok 4.6 → Gemini 3.8 Flash (High) → Kimi K3 → Qwen3 Coder Plus → OpenCode | Explicit Codex-quota relief; provider failure does not auto-retry another vendor |
 | 🗳️ 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 |
 
 The **backup** is the next candidate in the lane's `routing.yaml` chain — what
@@ -358,7 +358,7 @@ The commander orchestrates and validates; workers do not delegate again.
 - **Codex · Sol** — delegate bulk-mechanical and constrained ui-draft at high. Escalate hardest coding and judgment to Fable/Astra; route long/fast work to Gemini 3.8 Flash and live search to Grok.
 - **Codex · Astra** — prompt-level controller backup and independent reviewer. Use xhigh by default for hardest coding/judgment and consult/taste; explicitly select `--vendor codex --effort max` when needed. Explicit model/effort always win.
 - **Codex · Terra** — delegate the Codex long-context fallback at max. Bulk stays on Sol high; escalate hard work to Fable/Astra.
-- **Grok Build · Grok 4.6** — delegate live-search and coding-overflow, plus fallback duty in hardest-coding, hard-judgment, and taste-final. Dispatch primary hard coding/judgment/taste work to Codex/Claude/Gemini when available; verify API signatures and cited facts.
+- **Grok Build · Grok 4.7 / 4.6** — delegate live-search and coding-overflow, plus fallback duty in hardest-coding, hard-judgment, and taste-final. Dispatch primary hard coding/judgment/taste work to Codex/Claude/Gemini when available; verify API signatures and cited facts.
 - **Antigravity · Gemini 3.8 Flash** — delegate long-context Medium, fast-agentic/triage Low, and bulk/overflow/web fallbacks High. Do not infer visual taste or controller authority from agent/coding benchmarks.
 
 </details>
@@ -850,6 +850,43 @@ working notes, including per-benchmark caveats, live in
 
 ## 📜 Release history
 
+## What's new in v0.45.0
+
+- **Run `omnilane resign` once after upgrading.** The score registry moved to a
+  new snapshot and the transport overlay is bound to it; until the overlay is
+  rebuilt, a model caller is refused on every lane with
+  `transport overlay snapshot mismatch`.
+- **Scores follow Artificial Analysis Intelligence Index v4.3.2.** v4.2 and
+  v4.3.2 are different scales with uneven gaps (Fable 5.1 max 57 → 53, Grok 4.6
+  high 51 → 44, Sol high 48 → 42), so every row was re-scored rather than one
+  added. Ceilings move with it: Fable 5.1 max, Fable 5.1 xhigh and Astra max now
+  tie at 53, and a mid-effort controller that reached nothing in `hard-judgment`
+  before now does. `scripts/aa_rebaseline.py` rebuilds the registry from a saved
+  AA extract, so the next index revision is a re-run.
+- **The lane table is rewritten, not patched.** Each lane now names the
+  measurements that match its kind of work and lists candidates best-first on
+  them: hard terminal coding (Terminal-Bench 4.0, the coding benchmark the
+  frontier has not saturated) for `hardest-coding`, expert-level reasoning and
+  graded analytical quality for `hard-judgment`, graded document presentation for
+  `taste-final`, automation accuracy per minute for `fast-agentic`, the harder
+  long-context benchmark for `long-context`, visual understanding for `ui-draft`.
+  Eight lanes change their first choice: Astra (xhigh) leads `hardest-coding`,
+  Astra (low) leads `bulk-mechanical` and `fast-agentic`, Astra (high) leads
+  `ui-draft`, Opus 5 (max) leads `taste-final`, Opus 5 (high) leads
+  `long-context`, and Grok 4.7 leads `coding-overflow` as well as `live-search`.
+  Every chain also steps down through the score range, so whatever a
+  controller's ceiling, the first candidate it can reach is the best one it can
+  reach. Fable max and Astra max are in no chain: xhigh is level or ahead for
+  markedly less. The figures are in `docs/model-capabilities-2026-09.md`.
+- **Grok 4.7 joins every lane Grok serves**, ahead of Grok 4.6, which stays
+  behind it: a lane skips a candidate this host has not proven, so 4.6 keeps
+  serving until `resign` has probed 4.7. With `--vendor grok` there is no
+  fall-through; until then ask for `--model grok-4.6`.
+- **A dead fallback fixed.** `claude claude-sonnet-5 high` could never be
+  dispatched under the old registry; it now resolves once probed.
+- Upgrade: `npm i -g omnilane@0.45.0`, then `omnilane resign`. For Grok 4.7 the
+  `grok` CLI must be logged in when you re-sign.
+
 ## What's new in v0.44.0
 
 - **A CLI you patch yourself is re-signed unattended too.** If a local step

package/README.zh-CN.md

@@ -211,17 +211,17 @@ flowchart LR
 
 | 通道 | 首选模型 | 备选模型 | 用途 |
 |---|---|---|---|
-| 🔥 hardest-coding | Claude Fable 5.1 (max) | GPT-6 Astra (xhigh) → Grok 4.6 → Gemini 3.8 Flash (High) | 最难的实现、深度调试、正确性关键的修改 |
-| 🏗️ bulk-mechanical | GPT-5.6 Sol (high) | Gemini 3.8 Flash (High) → Claude Sonnet 5 (high) | 重构、迁移、测试、大范围扫描——机械耐力活 |
-| 🧹 triage | GPT-5.6 Luna (high) | Gemini 3.8 Flash (Low) → Claude Haiku 4.5 | 大量扫描、第一轮筛选 |
-| ⚖️ hard-judgment | Claude Fable 5.1 (xhigh) | GPT-6 Astra (xhigh) → Grok 4.6 | 架构裁决、深度推理、第二意见 |
-| ✒️ taste-final | Claude Fable 5.1 (xhigh) | GPT-6 Astra (xhigh) → Grok 4.6 → Gemini 3.8 Flash (High) | 对外文字、提示词／文档润色、风格裁决 |
-| 💬 consult | GPT-6 Astra (xhigh) | Claude Fable 5.1 (xhigh) → Grok 4.6 → Gemini 3.8 Flash (Medium) | 直接指定模型咨询；保留 `--vendor` 避免降级 |
-| 🎨 ui-draft | GPT-5.6 Sol (high) | Claude Fable 5.1 (xhigh) → Gemini 3.8 Flash (High) | 仅在提供设计系统／参考图时生成 UI 草稿 |
-| 📚 long-context | Gemini 3.8 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 |
-| ⚡ fast-agentic | Gemini 3.8 Flash (Low) | GPT-5.6 Luna (high) → Claude Haiku 4.5 | 高速多步骤工具循环、多模态检查 |
-| 📡 live-search | Grok 4.6 | Gemini 3.8 Flash (High) → Claude Sonnet 5 (high) | 实时 X／网页搜索与社交上下文 |
-| 🚰 coding-overflow | Grok 4.6 | Gemini 3.8 Flash (High) → Kimi K3 → Qwen3 Coder Plus → OpenCode | Codex 配额耗尽时的中量级编码安全阀 |
+| 🔥 hardest-coding | GPT-6 Astra (xhigh) | Claude Fable 5.1 (xhigh) → GPT-6 Astra (high) → Claude Fable 5.1 (high) → GPT-6 Astra (medium) → Claude Opus 5 (high) → GPT-6 Astra (low) → GPT-5.6 Sol (xhigh) → Grok 4.7 → Grok 4.6 → Gemini 3.8 Flash (High) | 最难的实现、深度调试、正确性关键的修改 |
+| 🏗️ bulk-mechanical | GPT-6 Astra (low) | GPT-5.6 Sol (high) → Gemini 3.8 Flash (High) → Claude Opus 5 (medium) | 重构、迁移、测试、大范围扫描——机械耐力活 |
+| 🧹 triage | GPT-5.6 Luna (high) | Gemini 3.8 Flash (Low) → Claude Sonnet 5 (low) → Claude Haiku 4.5 | 大量扫描、第一轮筛选 |
+| ⚖️ hard-judgment | Claude Fable 5.1 (xhigh) | GPT-6 Astra (xhigh) → Claude Opus 5 (max) → GPT-6 Astra (high) → Claude Opus 5 (xhigh) → Claude Opus 5 (high) → Grok 4.7 → Grok 4.6 → Gemini 3.8 Flash (High) | 架构裁决、深度推理、第二意见 |
+| ✒️ taste-final | Claude Opus 5 (max) | Claude Fable 5.1 (xhigh) → Claude Opus 5 (xhigh) → Grok 4.7 → GPT-6 Astra (xhigh) → Claude Opus 5 (high) → Grok 4.6 → Gemini 3.8 Flash (High) | 对外文字、提示词／文档润色、风格裁决 |
+| 💬 consult | GPT-6 Astra (xhigh) | Claude Fable 5.1 (xhigh) → Grok 4.7 → Grok 4.6 → Gemini 3.8 Flash (High) | 直接指定模型咨询；保留 `--vendor` 避免降级 |
+| 🎨 ui-draft | GPT-6 Astra (high) | Claude Opus 5 (high) → GPT-6 Astra (low) → Gemini 3.8 Flash (High) | 仅在提供设计系统／参考图时生成 UI 草稿 |
+| 📚 long-context | Claude Opus 5 (high) | Claude Opus 5 (medium) → Claude Opus 5 (low) → GPT-5.6 Terra (max) → Gemini 3.8 Flash (High) | 长文档提取与综合，按 AA-LCR、成本和吞吐排序 |
+| ⚡ fast-agentic | GPT-6 Astra (low) | Gemini 3.8 Flash (Medium) → GPT-5.6 Sol (medium) → Claude Opus 5 (low) | 高速多步骤工具循环、多模态检查 |
+| 📡 live-search | Grok 4.7 | Grok 4.6 → Gemini 3.8 Flash (High) → Claude Opus 5 (medium) | 实时 X／网页搜索与社交上下文 |
+| 🚰 coding-overflow | Grok 4.7 | Grok 4.6 → Gemini 3.8 Flash (High) → Kimi K3 → Qwen3 Coder Plus → OpenCode | Codex 配额耗尽时的中量级编码安全阀 |
 | 🗳️ arbitrate | off (opt-in vote panel) | — | 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 |
 
 **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
@@ -252,7 +252,7 @@ flowchart LR
 - **Claude Code · Opus 5**——自己执行：hard-judgment(默认车道)。需要更低幻觉率或价格时，用本地覆写让它接手 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。
 - **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。
 - **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。
-- **Grok Build · Grok 4.6**——自己执行 live-search、coding-overflow，并兼任 hardest-coding、hard-judgment、taste-final 的备用。首选可用时，最难的编码／判断／文字交给 Codex、Claude、Gemini；仍需验证 API 签名和引用事实。
+- **Grok Build · Grok 4.7 / 4.6**——自己执行 live-search、coding-overflow，并兼任 hardest-coding、hard-judgment、taste-final 的备用。首选可用时，最难的编码／判断／文字交给 Codex、Claude、Gemini；仍需验证 API 签名和引用事实。
 - **Antigravity · Gemini 3.7 Flash**——自己执行：Medium 的 long-context／高速循环、High 的 bulk／overflow、Low 的 triage，并以 High 兼任 hardest-coding、taste-final、ui-draft、live-search 的备用。首选可用时，最难编码／判断／文字交给 Codex、Claude。
 
 </details>
@@ -682,6 +682,15 @@ codex 记在 session rollout，agy 写进 `cli.log`。这是 CLI 自己抄的订
 
 ## 📜 版本历程
 
+## v0.45.0 新功能
+
+- **升级后请运行一次 `omnilane resign`。** 分数策略文件换了新快照，而传输覆盖文件绑定的是快照；覆盖文件重建之前，模型身份的派工在每条车道都会被拒，消息是 `transport overlay snapshot mismatch`。
+- **分数改用 Artificial Analysis Intelligence Index v4.3.2。** v4.2 与 v4.3.2 是不同的量尺，各家差距也不一样（Fable 5.1 max 57 → 53、Grok 4.6 high 51 → 44、Sol high 48 → 42），所以是整份重新计分，而不是只加一行。上限随之变化：Fable 5.1 max、Fable 5.1 xhigh、Astra max 现在同为 53 分；以前在 `hard-judgment` 什么都派不到的中档主控，现在派得到了。`scripts/aa_rebaseline.py` 会从保存下来的 AA 抽取文件重建策略文件，下次指数改版重跑即可。
+- **车道表是重写，不是修补。** 每条车道先说明“这类工作该看哪些测量”，候选再按那些测量由强到弱排列：`hardest-coding` 看高难度终端编码（Terminal-Bench 4.0，顶尖模型还没有做满的那个编码评测），`hard-judgment` 看专家级推理与文档的分析质量评分，`taste-final` 看文档呈现质量评分，`fast-agentic` 看自动化正确率与每步耗时，`long-context` 看较难的那个长上下文评测，`ui-draft` 看图像理解。有八条车道换了首选：`hardest-coding` 由 Astra (xhigh) 领先，`bulk-mechanical` 与 `fast-agentic` 由 Astra (low) 领先，`ui-draft` 由 Astra (high) 领先，`taste-final` 由 Opus 5 (max) 领先，`long-context` 由 Opus 5 (high) 领先，`coding-overflow` 和 `live-search` 一样由 Grok 4.7 领先。每条车道链同时沿着分数往下排，所以无论主控的上限在哪，它第一个派得到的就是它派得到的最好的。Fable max 与 Astra max 不在任何车道链里：xhigh 与它们持平或更好，花费少得多。数字都在 `docs/model-capabilities-2026-09.md`。
+- **Grok 服务的每条车道都加入 Grok 4.7**，排在 Grok 4.6 前面，4.6 留在后面：车道会跳过本机尚未验证的候选，所以在 `resign` 探测过 4.7 之前仍由 4.6 服务。带 `--vendor grok` 时不会往下找；在那之前请指定 `--model grok-4.6`。
+- **修掉一个失效的备用。** `claude claude-sonnet-5 high` 在旧策略文件下永远派不出去；现在探测通过后即可解析。
+- 升级：`npm i -g omnilane@0.45.0`，然后 `omnilane resign`。要用 Grok 4.7，重签时 `grok` CLI 必须处于登录状态。
+
 ## v0.44.0 新功能
 
 - **你自己修补过的 CLI 也能无人值守重新签署。** 如果本机有个步骤会在每次更新后修补某家 CLI、再用 adhoc 重新签署，以前签署者检查会把每一次这种更新都拦下来等 `--approve`。现在对该家运行一次 `omnilane resign --trust-adhoc VENDOR`，同一个安装目录里的 adhoc 新版就会和同签署者的更新一样，走每日 `omnilane resign` 自动重签。未签名的可执行文件、换了目录的 adhoc、其他厂商，仍然会停下来等你。信任记录在覆盖文件上，之后任何一家重签都会保留，而且是操作者动作，模型不会执行。四家都适用。

package/README.zh-TW.md

@@ -211,17 +211,17 @@ flowchart LR
 
 | 通道 | 首選模型 | 備選模型 | 用途 |
 |---|---|---|---|
-| 🔥 hardest-coding | Claude Fable 5.1（max） | GPT-6 Astra（xhigh）→ Grok 4.6 → Gemini 3.8 Flash（High） | 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 3.8 Flash（High）→ Claude Sonnet 5（high） | 重構、搬遷、測試、大面積掃描等耐力工作 |
-| 🧹 triage | GPT-5.6 Luna（high） | Gemini 3.8 Flash（Low）→ Claude Haiku 4.5 | 大量掃描、第一輪篩選 |
-| ⚖️ hard-judgment | Claude Fable 5.1（xhigh） | GPT-6 Astra（xhigh）→ Grok 4.6 | 架構裁決、深度推理、第二意見 |
-| ✒️ taste-final | Claude Fable 5.1（xhigh） | GPT-6 Astra（xhigh）→ Grok 4.6 → Gemini 3.8 Flash（High） | 對外文字與風格裁決；評測不等於審美證明 |
-| 💬 consult | GPT-6 Astra（xhigh） | Claude Fable 5.1（xhigh）→ Grok 4.6 → Gemini 3.8 Flash（Medium） | 直接點名模型諮詢；保留 `--vendor` 避免降級 |
-| 🎨 ui-draft | GPT-5.6 Sol（high） | Claude Fable 5.1（xhigh）→ Gemini 3.8 Flash（High） | 只有附設計系統／參考圖時做 UI 草稿；不把評測誇大成審美證明 |
-| 📚 long-context | Gemini 3.8 Flash（Medium） | GPT-5.6 Terra（max）→ Claude Opus 5（medium） | 長文件整合；上下文容量本身不證明任務品質 |
-| ⚡ fast-agentic | Gemini 3.8 Flash（Low） | GPT-5.6 Luna（high）→ Claude Haiku 4.5 | 高速多步驟工具迴圈、多模態檢查 |
-| 📡 live-search | Grok 4.6 | Gemini 3.8 Flash（High）→ Claude Sonnet 5（high） | 即時 X／網頁搜尋；備援只有一般網搜，不等同 X 脈絡 |
-| 🚰 coding-overflow | Grok 4.6 | Gemini 3.8 Flash（High）→ Kimi K3 → Qwen3 Coder Plus → OpenCode | 顯式 Codex 額度卸載；供應商失敗後不自動跨家重試 |
+| 🔥 hardest-coding | GPT-6 Astra（xhigh） | Claude Fable 5.1（xhigh）→ GPT-6 Astra（high）→ Claude Fable 5.1（high）→ GPT-6 Astra（medium）→ Claude Opus 5（high）→ GPT-6 Astra（low）→ GPT-5.6 Sol（xhigh）→ Grok 4.7 → Grok 4.6 → Gemini 3.8 Flash（High） | 最難的實作、深度除錯、正確性攸關的修改 |
+| 🏗️ bulk-mechanical | GPT-6 Astra（low） | GPT-5.6 Sol（high）→ Gemini 3.8 Flash（High）→ Claude Opus 5（medium） | 重構、搬遷、測試、大面積掃描等耐力工作 |
+| 🧹 triage | GPT-5.6 Luna（high） | Gemini 3.8 Flash（Low）→ Claude Sonnet 5（low）→ Claude Haiku 4.5 | 大量掃描、第一輪篩選 |
+| ⚖️ hard-judgment | Claude Fable 5.1（xhigh） | GPT-6 Astra（xhigh）→ Claude Opus 5（max）→ GPT-6 Astra（high）→ Claude Opus 5（xhigh）→ Claude Opus 5（high）→ Grok 4.7 → Grok 4.6 → Gemini 3.8 Flash（High） | 架構裁決、深度推理、第二意見 |
+| ✒️ taste-final | Claude Opus 5（max） | Claude Fable 5.1（xhigh）→ Claude Opus 5（xhigh）→ Grok 4.7 → GPT-6 Astra（xhigh）→ Claude Opus 5（high）→ Grok 4.6 → Gemini 3.8 Flash（High） | 對外文字與風格裁決；評測不等於審美證明 |
+| 💬 consult | GPT-6 Astra（xhigh） | Claude Fable 5.1（xhigh）→ Grok 4.7 → Grok 4.6 → Gemini 3.8 Flash（High） | 直接點名模型諮詢；保留 `--vendor` 避免降級 |
+| 🎨 ui-draft | GPT-6 Astra（high） | Claude Opus 5（high）→ GPT-6 Astra（low）→ Gemini 3.8 Flash（High） | 只有附設計系統／參考圖時做 UI 草稿；不把評測誇大成審美證明 |
+| 📚 long-context | Claude Opus 5（high） | Claude Opus 5（medium）→ Claude Opus 5（low）→ GPT-5.6 Terra（max）→ Gemini 3.8 Flash（High） | 長文件整合；上下文容量本身不證明任務品質 |
+| ⚡ fast-agentic | GPT-6 Astra（low） | Gemini 3.8 Flash（Medium）→ GPT-5.6 Sol（medium）→ Claude Opus 5（low） | 高速多步驟工具迴圈、多模態檢查 |
+| 📡 live-search | Grok 4.7 | Grok 4.6 → Gemini 3.8 Flash（High）→ Claude Opus 5（medium） | 即時 X／網頁搜尋；備援只有一般網搜，不等同 X 脈絡 |
+| 🚰 coding-overflow | Grok 4.7 | Grok 4.6 → Gemini 3.8 Flash（High）→ Kimi K3 → Qwen3 Coder Plus → OpenCode | 顯式 Codex 額度卸載；供應商失敗後不自動跨家重試 |
 | 🗳️ 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
@@ -282,7 +282,7 @@ omnilane jobs --json status JOB_ID
 - **Codex · Sol**——bulk-mechanical 與有參考限制的 ui-draft 委派並使用 high；最難編碼／判斷升級 Fable 或 Astra，長文／高速工作交 Gemini 3.8 Flash，即時搜尋交 Grok。
 - **Codex · Astra**——提示詞層主控備位與獨立複核者；最難編碼／判斷與 consult／taste 預設用 xhigh，需要時可明確指定 `--vendor codex --effort max`；顯式 model／effort 永遠優先。
 - **Codex · Terra**——用 max 接 Codex 的 long-context 備援；bulk 留給 Sol high，困難工作升級 Fable／Astra。
-- **Grok Build · Grok 4.6**——委派 live-search、coding-overflow，並兼任 hardest-coding、hard-judgment、taste-final 的備援。首選人手在的話，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。
+- **Grok Build · Grok 4.7／4.6**——委派 live-search、coding-overflow，並兼任 hardest-coding、hard-judgment、taste-final 的備援。首選人手在的話，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 3.8 Flash**——long-context 用 Medium，fast-agentic／triage 用 Low，bulk／overflow／網搜備援用 High。不要把代理／編碼評測推論成審美或主控權。
 
 </details>
@@ -691,6 +691,15 @@ codex 記在 session rollout，agy 寫進 `cli.log`。這是 CLI 自己抄的訂
 
 ## 📜 版本歷程
 
+## v0.45.0 新功能
+
+- **升級後請跑一次 `omnilane resign`。** 分數政策檔換了新快照，而傳輸覆蓋檔綁的是快照；覆蓋檔重建之前，模型身分的派工在每條車道都會被拒，訊息是 `transport overlay snapshot mismatch`。
+- **分數改採 Artificial Analysis Intelligence Index v4.3.2。** v4.2 與 v4.3.2 是不同的量尺，各家落差也不一樣（Fable 5.1 max 57 → 53、Grok 4.6 high 51 → 44、Sol high 48 → 42），所以是整份重新計分，不是只加一列。上限跟著變：Fable 5.1 max、Fable 5.1 xhigh、Astra max 現在同為 53 分；以前在 `hard-judgment` 什麼都派不到的中階主控，現在派得到了。`scripts/aa_rebaseline.py` 會從存下來的 AA 抽取檔重建政策檔，下次指數改版重跑即可。
+- **車道表是重寫，不是修補。** 每條車道先講明「這類工作該看哪些量測」，候選再依那些量測由強到弱排：`hardest-coding` 看高難度終端編碼（Terminal-Bench 4.0，頂尖模型還沒把它做滿的那個編碼評測），`hard-judgment` 看專家級推理與文件的分析品質評分，`taste-final` 看文件呈現品質評分，`fast-agentic` 看自動化正確率與每步耗時，`long-context` 看較難的那個長上下文評測，`ui-draft` 看圖像理解。有八條車道換了首選：`hardest-coding` 由 Astra（xhigh）領先，`bulk-mechanical` 與 `fast-agentic` 由 Astra（low）領先，`ui-draft` 由 Astra（high）領先，`taste-final` 由 Opus 5（max）領先，`long-context` 由 Opus 5（high）領先，`coding-overflow` 和 `live-search` 一樣由 Grok 4.7 領先。每條車道鏈同時沿著分數往下排，所以不論主控的上限在哪，它第一個派得到的就是它派得到的最好的。Fable max 與 Astra max 不在任何車道鏈裡：xhigh 與它們持平或更好，花費少得多。數字都在 `docs/model-capabilities-2026-09.md`。
+- **Grok 服務的每條車道都加入 Grok 4.7**，排在 Grok 4.6 前面，4.6 留在後面：車道會跳過本機尚未驗證的候選，所以在 `resign` 探測過 4.7 之前仍由 4.6 服務。帶 `--vendor grok` 時不會往下找；在那之前請指定 `--model grok-4.6`。
+- **修掉一個失效的備援。** `claude claude-sonnet-5 high` 在舊政策檔下永遠派不出去；現在探測通過後就能解析。
+- 升級：`npm i -g omnilane@0.45.0`，接著 `omnilane resign`。要用 Grok 4.7，重簽時 `grok` CLI 必須是登入狀態。
+
 ## v0.44.0 新功能
 
 - **你自己修補過的 CLI 也能無人值守重簽。** 如果本機有個步驟會在每次更新後修補某家 CLI、再用 adhoc 重新簽署，以前簽署者檢查會把每一次這種更新都擋下來等 `--approve`。現在對該家跑一次 `omnilane resign --trust-adhoc VENDOR`，同一個安裝目錄裡的 adhoc 新版就會跟同簽署者的更新一樣，走每日 `omnilane resign` 自動重簽。未簽章的執行檔、換了目錄的 adhoc、其他廠商，仍然會停下來等你。信任記在覆蓋檔上，之後任何一家重簽都會保留，而且是操作者動作，模型不會執行。四家都適用。

package/VERSION

@@ -1 +1 @@
-0.44.0
+0.45.0