omnilane 0.8.2 → 0.9.0

package/CHANGELOG.md

@@ -6,6 +6,61 @@ semantic version tags.
 
 ## [Unreleased]
 
+## [0.9.0] - 2026-07-19
+
+### Added
+
+- `omnilane completion fish` prints a Fish completion script
+  (`completions/omnilane.fish`): static command, option, and enum completion for
+  every public command, including the direct-API vendors and `jobs cancel`/`rm`.
+  Load it with `omnilane completion fish | source`. Dynamic lane-name and job-id
+  completion remain Bash/Zsh-only for now.
+- Five OpenAI-compatible direct-API vendors — `deepseek`, `zai` (GLM),
+  `mistral`, `groq`, and `cerebras` — join `openrouter` as CLI-free inference
+  lanes (curl + a `<VENDOR>_API_KEY`, advise mode only). A single vendor registry
+  in `lib/common.sh` (endpoint, key env, and default model per vendor) now drives
+  every vendor-name check in `dispatch.sh` and `jobs.sh`, so adding another
+  OpenAI-compatible vendor is a one-line registry entry plus a thin runner
+  wrapper. `configure` lists them with suggested models. See
+  `docs/model-capabilities-2026-07.md` for the July 2026 model/benchmark snapshot
+  behind the routing defaults.
+- `omnilane doctor` now reports vendor-CLI availability: which of codex, claude,
+  grok, gemini (the `agy` CLI), kimi, qwen, and opencode are reachable, plus
+  the direct-API vendors openrouter, deepseek, zai, mistral, groq, and cerebras
+  (curl + each `<VENDOR>_API_KEY`). Shown in human and `--json` output;
+  the probe mirrors the runners' `*_BIN` overrides and `local.sh` in an isolated
+  subshell. It is a `PASS`/`WARN` check and never fails the report.
+- MCP server read-only introspection tools: `explain`, `validate`, `dry_run`,
+  `jobs_list`, `jobs_stats`, `jobs_audit`, and `doctor` join the existing
+  `route`, `list_lanes`, `jobs_status`, and `jobs_result`, so the MCP surface
+  mirrors the CLI's full offline read-only surface. All refuse provider calls and
+  create no job state; `dry_run` requires an explicit `workdir` only when `mode`
+  is `work`. `jobs_stats`/`jobs_audit` accept an optional `last` and `json`.
+- `configure set|get|unset|list|diff` non-interactive subcommands: script or
+  inspect `routing.local.yaml` without a tty. `set` validates the lane, refuses
+  unsafe specs, and rolls back on a structural FAIL; `diff` shows how local
+  overrides change the effective table versus the defaults. The interactive menu
+  is unchanged.
+- `jobs list` and `jobs stats` filters: `--lane` and `--vendor` (and, for
+  `list`, `--status running|done`) narrow the output in both human and `--json`
+  modes. `list` scans all jobs for up to 20 matches; `stats` filters within its
+  `--last` sample and recomputes the success rate and lane/vendor aggregates.
+  Invalid filter values exit 2; unfiltered behavior is unchanged.
+- `jobs cancel ID`: stop a running background job. Signals the worker's process
+  group (SIGTERM, then SIGKILL after a grace period) so the vendor CLI child
+  dies too, and leaves a terminal exit recorded — 143 on graceful stop, 137 if
+  force-killed. Idempotent on already-finished, dead, or never-backgrounded jobs.
+- `jobs rm ID`: delete one job's stored directory (task, output, exit, metadata).
+  Refuses a job whose worker is still alive — cancel it first — complementing
+  `jobs prune`, which only bulk-deletes completed jobs by count or age.
+
+## [0.8.3] - 2026-07-18
+
+### Added
+
+- Dependency-free MCP stdio server (`omnilane mcp`) exposing routing, lane
+  discovery, and background-job status/results to any MCP-capable host.
+
 ## [0.8.2] - 2026-07-18
 
 ### Added
@@ -252,7 +307,9 @@ semantic version tags.
 - Initial shared routing table, cross-vendor dispatcher, runners, installer,
   and baseline lint fixes.
 
-[Unreleased]: https://github.com/Seraphim0916/omnilane/compare/v0.8.2...HEAD
+[Unreleased]: https://github.com/Seraphim0916/omnilane/compare/v0.9.0...HEAD
+[0.9.0]: https://github.com/Seraphim0916/omnilane/compare/v0.8.3...v0.9.0
+[0.8.3]: https://github.com/Seraphim0916/omnilane/compare/v0.8.2...v0.8.3
 [0.8.2]: https://github.com/Seraphim0916/omnilane/compare/v0.8.1...v0.8.2
 
 [0.8.1]: https://github.com/Seraphim0916/omnilane/compare/v0.8.0...v0.8.1

package/README.ja.md

@@ -5,8 +5,9 @@
 ### ルーティングテーブルは一枚、ハーネスは全部。
 
 *メインループはもう、どのモデルを使うか迷わない。*<br/>
-すべてのサブタスクを、その作業が本当に得意なモデルへ——<br/>
-**Claude Code · Codex · Grok Build · Antigravity** を横断、いまのサブスクリプションのままで。
+**Claude Code · Codex · Grok Build · Antigravity** のどれから操縦しても、すべてのサブタスクを<br/>
+その作業が本当に得意なモデルへ——Codex、Claude、Grok、Gemini、Kimi、Qwen、OpenCode、<br/>
+さらに OpenRouter 経由の任意のホスト型モデル——いまのサブスクリプションのまま、または API キー一つで。
 
 <img src="docs/hero.ja.png" alt="omnilane が各サブタスクを Claude Code・Codex・Grok・Antigravity の最適なモデルへ振り分ける" width="820"/>
 
@@ -20,6 +21,22 @@
 
 ---
 
+## v0.9.0 の新機能
+
+- **OpenAI 互換の direct-API ベンダーを 5 つ追加** — `deepseek`、`zai`(GLM)、
+  `mistral`、`groq`、`cerebras` が `openrouter` と同じく CLI 不要のレーンに
+  (curl と `<VENDOR>_API_KEY` だけ)。`lib/common.sh` のレジストリに 1 行で
+  追加でき、モデル能力の比較は [`docs/model-capabilities-2026-07.md`](docs/model-capabilities-2026-07.md) を参照。
+- **Fish シェル補完** — `omnilane completion fish | source`。
+
+## v0.8.3 の新機能
+
+- **MCP サーバー** — `omnilane mcp` は依存ゼロの stdio MCP サーバーを起動し、
+  MCP 対応ホスト(Claude Code、Codex、Gemini CLI、Cursor、OpenCode など)が
+  スキルなしで omnilane を発見・呼び出しできます:ツールは `route`、
+  `jobs_status`、`jobs_result`、`list_lanes`。`route` は読み取り専用の
+  advise がデフォルトで、work モードには明示的な workdir が必須です。
+
 ## v0.8.2 の新機能
 
 - **`openrouter` vendor** — `curl` と `OPENROUTER_API_KEY` だけで
@@ -114,8 +131,9 @@ omnilane ui start     # 任意:ブラウザでジョブをライブ表示
 ## 🧭 仕組み
 
 omnilane は、**どの** agentic CLI のメインループでも、サブタスクをレーンに
-分類し、各レーンをその作業が最も得意なベンダー CLI へヘッドレスで
-ディスパッチさせる仕組みです。既存のサブスクリプションログインをそのまま使います:
+分類し、各レーンをその作業が最も得意なベンダーへヘッドレスで
+ディスパッチさせる仕組みです。既存のサブスクリプションログインをそのまま
+使います(`openrouter` vendor は例外:CLI 不要、API キー一つで直接接続):
 
 ```mermaid
 flowchart LR
@@ -138,6 +156,8 @@ flowchart LR
   指定ベンダーに固定し、フォールバックしません。
 - **`skills/omnilane/SKILL.md`** — 四つのハーネス共通のスキル:
   自分のモデルを特定し、自分のレーンは自前で実行、残りはディスパッチ。
+- **`omnilane mcp`** — 同じルーティングを MCP stdio サーバーとして提供。
+  スキルではなく MCP で統合するホスト向け。
 
 <div align="center">
 
@@ -243,9 +263,11 @@ omnilane ui stop     # 正常に停止
 
 ## 📦 インストール
 
-前提:ルーティングしたいベンダー CLI(`codex`、`claude`、`grok`、`agy`)が
-ログイン済みで `PATH` 上にあること——**持っている分だけで OK**、
-足りないレーンは自動的に降格します。
+前提:ルーティングしたいベンダー CLI(`codex`、`claude`、`grok`、`agy`、
+任意で `kimi`、`qwen`、`opencode`)がログイン済みで `PATH` 上にあること——
+**持っている分だけで OK**、足りないレーンは自動的に降格します。
+`openrouter` vendor は例外で、CLI は不要——`curl` と環境変数の
+`OPENROUTER_API_KEY` だけで動きます。
 
 最速:`./install.sh` — 本機の CLI を検出してスキルを接続し、残りのプラグイン
 コマンドを表示、実効ルーティングを出力し、最後に対話式設定メニューを
@@ -271,6 +293,31 @@ CLI バージョンにより異なる場合あり)へマーカー付きの可逆
 - **Antigravity**:`agy plugin install <このリポジトリ>`(先に
   `agy plugin validate` で確認)
 
+### MCP サーバー
+
+`omnilane mcp` は依存ゼロでローカル実行される MCP stdio サーバーを起動し、
+MCP 対応ホストがスキルの導入やルーティングリマインダーなしで omnilane を
+発見・呼び出しできるようにします。ホスト側の設定でインストール済み CLI を
+指定してください:
+
+```json
+{
+  "mcpServers": {
+    "omnilane": {
+      "command": "omnilane",
+      "args": ["mcp"]
+    }
+  }
+}
+```
+
+サーバーは `route` に加えて、読み取り専用の照会ツール `list_lanes`、`explain`、`validate`、`dry_run`、`jobs_list`、`jobs_status`、`jobs_result`、`jobs_stats`、`jobs_audit`、`doctor` を公開します。
+`route` のデフォルトは読み取り専用の `advise` モードで、`work` を選ぶ呼び
+出しには明示的な `workdir` が必須です。
+
+実行時に必要なのは Node.js のみ(npm パッケージは不要)。npm 派なら
+`npm install -g omnilane` で MCP サーバーごと CLI をインストールできます。
+
 ## ⚙️ カスタマイズ
 
 三層、すべて任意:
@@ -295,6 +342,8 @@ scripts/dispatch.sh --list     # 実効テーブル(フォールバック解決
 ```
 eval "$(omnilane completion bash)"             # 現在の Bash で補完を有効化
 source <(omnilane completion zsh)               # 現在の Zsh で補完を有効化
+omnilane completion fish | source              # 現在の Fish で補完を有効化
+omnilane mcp                                   # MCP stdio サーバー(Node.js が必要)
 omnilane release-audit [--target VERSION] [--json] # オフライン・読み取り専用のリリースゲート
 omnilane ui start                              # ローカル Live UI を起動または再利用し、URL を表示
 omnilane ui status                             # Live UI の稼働状態を表示
@@ -308,11 +357,15 @@ dispatch.sh [--json] --list [--json]
 dispatch.sh [--json] --explain LANE [--json]       # 候補ごとの決定理由をオフライン表示
 dispatch.sh [--json] --validate [--json]           # プロバイダーを呼ばず実効ルーティングを検証
 jobs.sh [--json] {list | status ID | result ID}    # JSON は本文を返さずメタデータのみ
+jobs.sh [--json] list [--lane L] [--vendor V] [--status running|done]  # 一覧を絞り込み
 jobs.sh wait ID [--timeout N]                     # ジョブ終了値。124 はタイムアウト、125 はワーカー消失
-jobs.sh [--json] stats [--last N]                  # ローカル成功率とルーティング集計
+jobs.sh cancel ID                                 # 実行中ジョブを停止:グループに SIGTERM、その後 SIGKILL
+jobs.sh rm ID                                     # 完了/停止ジョブを1件削除(実行中は拒否)
+jobs.sh [--json] stats [--last N] [--lane L] [--vendor V]  # ローカル成功率とルーティング集計
 jobs.sh audit [--last N] [--json]                  # 読み取り専用のジョブ整合性・プライバシー検査
 jobs.sh prune [--keep N] [--apply]                # 既定はプレビューのみ。完了ジョブだけを対象
 configure.sh                                        # 対話式レーンメニュー
+configure.sh set|get|unset|list|diff LANE [SPEC]    # routing.local.yaml を非対話で編集/確認
 ```
 
 終了コード:`2` 使い方エラー(無効なベンダー、または指定ベンダーがレーンに
@@ -325,9 +378,13 @@ configure.sh                                        # 対話式レーンメニ
 ## 🎭 モード
 
 - **advise(デフォルト)** — 読み取り専用ワーカー。Codex は read-only
-  サンドボックス、Claude は Read/Glob/Grep のみ、Grok は plan モード。
+  サンドボックス、Claude は Read/Glob/Grep のみ、Grok は plan モード、
+  Kimi と OpenCode はそれぞれの読み取り専用 plan モードに固定、
+  OpenRouter は設計上 advise 専用(純推論)。
 - **work** — 指定した `--workdir` 内でのみファイル編集可。Codex は
   workspace-write、Claude は編集自動承認、Gemini は accept-edits モード。
+  `openrouter` vendor は work モードを明確に拒否します——編集はエージェント型
+  CLI ベンダーへ。
 
 ## 🔒 安全機構
 
@@ -381,11 +438,15 @@ configure.sh                                        # 対話式レーンメニ
 
 ## 🌱 ステータス
 
-v0.5.1 は Git 管理外でも Codex `work` を利用可能に保ち、process group の
-クリーンアップで停止を制限し、公開バージョン情報を同期します。v0.5.0 の
-インストーラー、ライフサイクル、ジョブストア、期限、診断、リリース CI の
-強化も維持します。Grok/Antigravity のコマンドシェル挙動は CLI バージョンで
-変わる可能性があります。issue と PR を歓迎します。
+omnilane は 13 のディスパッチベンダーを備えます——4 つのハーネスネイティブ
+(codex、claude、grok、gemini)、3 つの集約/オーバーフロー CLI(kimi、qwen、
+opencode)、そして CLI 不要の OpenAI 互換 direct-API ベンダー 6 つ(openrouter、
+deepseek、zai、mistral、groq、cerebras)——すべて統一 runner 契約と契約テスト
+付き。Claude Code の `SessionStart` 自動リマインダーと MCP stdio サーバー
+(`omnilane mcp`)も同梱。direct-API と集約系の runner はフェイクバイナリで
+契約テスト済みです。実モデルでの報告を歓迎します。Grok/Antigravity の
+コマンドシェル挙動は CLI バージョンで変わる可能性があります。issue と PR を
+歓迎します。
 
 プロジェクト文書：[コントリビューション](CONTRIBUTING.md) ·
 [セキュリティ](SECURITY.md) · [変更履歴](CHANGELOG.md)

package/README.ko.md

@@ -5,8 +5,9 @@
 ### 라우팅 테이블 하나로, 모든 하네스를.
 
 *메인 루프가 더 이상 어떤 모델을 쓸지 고민하지 않습니다.*<br/>
-모든 서브태스크를 그 일을 정말 잘하는 모델에게——<br/>
-**Claude Code · Codex · Grok Build · Antigravity** 를 가로질러, 이미 내고 있는 구독 그대로.
+**Claude Code · Codex · Grok Build · Antigravity** 어디서 운전하든, 모든 서브태스크를<br/>
+그 일을 정말 잘하는 모델에게——Codex, Claude, Grok, Gemini, Kimi, Qwen, OpenCode,<br/>
+그리고 OpenRouter 를 통한 모든 호스팅 모델까지——이미 내고 있는 구독 그대로, 또는 API 키 하나로.
 
 <img src="docs/hero.ko.png" alt="omnilane 이 각 서브태스크를 Claude Code·Codex·Grok·Antigravity 중 가장 잘하는 모델로 보냅니다" width="820"/>
 
@@ -20,6 +21,22 @@
 
 ---
 
+## v0.9.0 새 기능
+
+- **OpenAI 호환 direct-API 벤더 5개 추가** — `deepseek`, `zai` (GLM), `mistral`,
+  `groq`, `cerebras` 가 `openrouter` 처럼 CLI 없는 레인으로 추가(curl 과
+  `<VENDOR>_API_KEY` 만 필요). `lib/common.sh` 레지스트리에 한 줄로 추가되며,
+  모델 능력 비교는 [`docs/model-capabilities-2026-07.md`](docs/model-capabilities-2026-07.md) 참고.
+- **Fish 셸 자동완성** — `omnilane completion fish | source`.
+
+## v0.8.3 새 기능
+
+- **MCP 서버** — `omnilane mcp` 는 의존성 없는 stdio MCP 서버를 시작하여,
+  MCP 지원 호스트(Claude Code, Codex, Gemini CLI, Cursor, OpenCode 등)가
+  스킬 설치 없이 omnilane 을 발견하고 호출할 수 있습니다: 도구는 `route`,
+  `jobs_status`, `jobs_result`, `list_lanes`. `route` 는 읽기 전용 advise 가
+  기본값이며, work 모드는 명시적 workdir 이 필요합니다.
+
 ## v0.8.2 새 기능
 
 - **`openrouter` vendor** — `curl`과 `OPENROUTER_API_KEY`만으로
@@ -114,8 +131,9 @@ omnilane ui start     # 선택: 브라우저에서 잡을 실시간 확인
 ## 🧭 동작 방식
 
 omnilane 은 **어떤** agentic CLI 든 메인 루프가 서브태스크를 레인으로 분류하고,
-각 레인을 그 작업에 가장 강한 벤더 CLI 로 헤드리스 디스패치하게 해 줍니다.
-기존 구독 로그인을 그대로 사용합니다:
+각 레인을 그 작업에 가장 강한 벤더로 헤드리스 디스패치하게 해 줍니다.
+기존 구독 로그인을 그대로 사용합니다(`openrouter` vendor 는 예외: CLI 없이
+API 키 하나로 직접 연결):
 
 ```mermaid
 flowchart LR
@@ -137,6 +155,8 @@ flowchart LR
   고정하며 폴백하지 않습니다.
 - **`skills/omnilane/SKILL.md`** — 네 하네스 공용 스킬: 자기 모델을 파악하고,
   자기 레인은 직접 수행, 나머지는 디스패치.
+- **`omnilane mcp`** — 같은 라우팅을 MCP stdio 서버로 제공.
+  스킬 대신 MCP 로 통합하는 호스트용.
 
 <div align="center">
 
@@ -240,8 +260,10 @@ omnilane ui stop     # 정상 중지
 
 ## 📦 설치
 
-전제: 라우팅할 벤더 CLI(`codex`, `claude`, `grok`, `agy`)가 로그인된 채
-`PATH` 에 있을 것——**가진 것만 있으면 됩니다**, 없는 레인은 자동 강등.
+전제: 라우팅할 벤더 CLI(`codex`, `claude`, `grok`, `agy`, 선택적으로
+`kimi`, `qwen`, `opencode`)가 로그인된 채 `PATH` 에 있을 것——**가진 것만
+있으면 됩니다**, 없는 레인은 자동 강등. `openrouter` vendor 는 예외로 CLI 가
+필요 없습니다——`curl` 과 환경 변수의 `OPENROUTER_API_KEY` 만 있으면 됩니다.
 
 가장 빠른 방법: `./install.sh` — 로컬 CLI 를 감지해 스킬을 연결하고, 나머지
 플러그인 명령을 안내하며, 실효 라우팅을 출력한 뒤 대화형 설정 메뉴를
@@ -265,6 +287,30 @@ omnilane ui stop     # 정상 중지
 - **Antigravity**: `agy plugin install <이 저장소>`(먼저
   `agy plugin validate` 로 확인)
 
+### MCP 서버
+
+`omnilane mcp` 는 의존성 없이 로컬에서 실행되는 MCP stdio 서버를 시작하여,
+MCP 지원 호스트가 스킬 설치나 라우팅 리마인더 없이 omnilane 을 발견하고
+호출할 수 있게 합니다. 호스트 설정에서 설치된 CLI 를 지정하면 됩니다:
+
+```json
+{
+  "mcpServers": {
+    "omnilane": {
+      "command": "omnilane",
+      "args": ["mcp"]
+    }
+  }
+}
+```
+
+서버는 `route` 와 함께 읽기 전용 조회 도구 `list_lanes`, `explain`, `validate`, `dry_run`, `jobs_list`, `jobs_status`, `jobs_result`, `jobs_stats`, `jobs_audit`, `doctor` 를 제공합니다.
+`route` 의 기본값은 읽기 전용 `advise` 모드이며, `work` 를 선택한 호출은
+명시적 `workdir` 을 함께 제공해야 합니다.
+
+실행에 필요한 것은 Node.js 뿐입니다(npm 패키지 없음). npm 을 선호하면
+`npm install -g omnilane` 으로 MCP 서버가 포함된 CLI 를 설치할 수 있습니다.
+
 ## ⚙️ 사용자 설정
 
 세 계층, 모두 선택 사항:
@@ -289,6 +335,8 @@ scripts/dispatch.sh --list     # 실효 테이블(폴백 해석 주석 포함)
 ```
 eval "$(omnilane completion bash)"             # 현재 Bash 에서 완성 활성화
 source <(omnilane completion zsh)               # 현재 Zsh 에서 완성 활성화
+omnilane completion fish | source              # 현재 Fish 에서 완성 활성화
+omnilane mcp                                   # MCP stdio 서버(Node.js 필요)
 omnilane release-audit [--target VERSION] [--json] # 오프라인 읽기 전용 릴리스 게이트
 omnilane ui start                              # 로컬 Live UI 를 시작하거나 재사용하고 URL 표시
 omnilane ui status                             # Live UI 실행 상태 표시
@@ -302,11 +350,15 @@ dispatch.sh [--json] --list [--json]
 dispatch.sh [--json] --explain LANE [--json]       # 후보별 라우팅 결정을 오프라인 설명
 dispatch.sh [--json] --validate [--json]           # 공급자 호출 없이 실효 라우팅 검사
 jobs.sh [--json] {list | status ID | result ID}    # JSON은 본문 없이 메타데이터만 반환
+jobs.sh [--json] list [--lane L] [--vendor V] [--status running|done]  # 목록 필터
 jobs.sh wait ID [--timeout N]                     # 작업 종료값, 124 시간 초과, 125 작업자 소실
-jobs.sh [--json] stats [--last N]                  # 로컬 성공률과 라우팅 집계
+jobs.sh cancel ID                                 # 실행 중 작업 중지: 그룹 SIGTERM 후 SIGKILL
+jobs.sh rm ID                                     # 완료/종료 작업 1건 삭제(실행 중이면 거부)
+jobs.sh [--json] stats [--last N] [--lane L] [--vendor V]  # 로컬 성공률과 라우팅 집계
 jobs.sh audit [--last N] [--json]                  # 읽기 전용 작업 무결성/개인정보 검사
 jobs.sh prune [--keep N] [--apply]                # 기본은 미리보기이며 완료된 작업만 정리
 configure.sh                                        # 대화형 레인 메뉴
+configure.sh set|get|unset|list|diff LANE [SPEC]    # routing.local.yaml 비대화식 편집/확인
 ```
 
 종료 코드: `2` 사용법 오류(잘못된 벤더 또는 지정 벤더가 레인에 없는 경우 포함),
@@ -318,9 +370,12 @@ CLI 를 사용할 수 없음, `5` 1라운드 성공 투표자 부족, `6` 2라
 ## 🎭 모드
 
 - **advise(기본)** — 읽기 전용 워커. Codex 는 read-only 샌드박스,
-  Claude 는 Read/Glob/Grep 만, Grok 은 plan 모드.
+  Claude 는 Read/Glob/Grep 만, Grok 은 plan 모드, Kimi 와 OpenCode 는 각자의
+  읽기 전용 plan 모드 고정, OpenRouter 는 설계상 advise 전용(순수 추론).
 - **work** — 지정한 `--workdir` 안에서만 파일 수정 허용. Codex 는
   workspace-write, Claude 는 편집 자동 승인, Gemini 는 accept-edits 모드.
+  `openrouter` vendor 는 work 모드를 명확한 오류로 거부합니다——파일 편집은
+  에이전트형 CLI 벤더로 보내세요.
 
 ## 🔒 안전 장치
 
@@ -371,11 +426,14 @@ CLI 를 사용할 수 없음, `5` 1라운드 성공 투표자 부족, `6` 2라
 
 ## 🌱 상태
 
-v0.5.1 은 Git 저장소 밖에서도 Codex `work` 를 사용할 수 있게 유지하고 process
-group 정리로 멈춤을 제한하며 모든 공개 버전 정보를 동기화합니다. v0.5.0 의 설치
-프로그램, 디스패치 수명주기, 잡 저장소, 전체 기한, 진단, 릴리스 CI 강화도 그대로
-유지합니다. Grok/Antigravity 커맨드 셸 동작은 CLI 버전에 따라 달라질 수 있습니다.
-issue 와 PR 환영합니다.
+omnilane 은 13개의 디스패치 벤더를 갖춥니다——4개의 하네스 네이티브(codex,
+claude, grok, gemini), 3개의 집합/오버플로 CLI(kimi, qwen, opencode), 그리고
+CLI 가 필요 없는 OpenAI 호환 direct-API 벤더 6개(openrouter, deepseek, zai,
+mistral, groq, cerebras)——모두 통일 runner 계약과 계약 테스트를 갖추었고,
+Claude Code `SessionStart` 자동 리마인더와 MCP stdio 서버(`omnilane mcp`)도 포함합니다.
+kimi, qwen, opencode, openrouter 의 runner 는 페이크 바이너리로 계약 테스트를
+마쳤습니다. 실제 모델 사용 보고를 환영합니다. Grok/Antigravity 커맨드 셸
+동작은 CLI 버전에 따라 달라질 수 있습니다. issue 와 PR 환영합니다.
 
 프로젝트 문서: [기여 가이드](CONTRIBUTING.md) · [보안 정책](SECURITY.md) ·
 [변경 기록](CHANGELOG.md)

package/README.md

@@ -5,8 +5,9 @@
 ### One routing table, every harness.
 
 *Your main loop stops guessing which model to use.*<br/>
-Every subtask goes to the model that is actually best at it — across<br/>
-**Claude Code · Codex · Grok Build · Antigravity**, on the subscriptions you already pay for.
+Drive it from **Claude Code · Codex · Grok Build · Antigravity**, and every subtask goes<br/>
+to the model that is actually best at it — Codex, Claude, Grok, Gemini, Kimi, Qwen, OpenCode,<br/>
+or any hosted model via OpenRouter — on the subscriptions you already pay for, or a single API key.
 
 <img src="docs/hero.png" alt="omnilane routes each subtask to the best model across Claude Code, Codex, Grok and Antigravity" width="820"/>
 
@@ -20,6 +21,22 @@ Every subtask goes to the model that is actually best at it — across<br/>
 
 ---
 
+## What's new in v0.9.0
+
+- **Five OpenAI-compatible direct-API vendors** — `deepseek`, `zai` (GLM),
+  `mistral`, `groq`, and `cerebras` join `openrouter` as CLI-free lanes (curl +
+  a `<VENDOR>_API_KEY`). A one-line `lib/common.sh` registry entry adds each;
+  see [`docs/model-capabilities-2026-07.md`](docs/model-capabilities-2026-07.md).
+- **Fish shell completion** — `omnilane completion fish | source`.
+
+## What's new in v0.8.3
+
+- **MCP server** — `omnilane mcp` starts a zero-dependency stdio MCP server,
+  so any MCP-capable host (Claude Code, Codex, Gemini CLI, Cursor, OpenCode…)
+  can discover and call omnilane without installing the skill: tools `route`,
+  `jobs_status`, `jobs_result`, and `list_lanes`. `route` defaults to
+  read-only advise mode; work mode requires an explicit workdir.
+
 ## What's new in v0.8.2
 
 - **`openrouter` vendor** — dispatch straight to the OpenRouter API with
@@ -28,6 +45,12 @@ Every subtask goes to the model that is actually best at it — across<br/>
   Advise/consult only (it cannot edit files; work mode fails with guidance)
   and the model slug is mandatory, e.g.
   `dispatch.sh --vendor openrouter --model anthropic/claude-sonnet-5 consult "..."`.
+- **`deepseek`, `zai`, `mistral`, `groq`, `cerebras` vendors** — the same
+  CLI-free direct-API path as `openrouter`, for OpenAI-compatible providers:
+  DeepSeek, Z.ai GLM, Mistral, Groq, and Cerebras. Each needs only `curl` and
+  its `<VENDOR>_API_KEY`; advise/consult only. A one-line `lib/common.sh`
+  registry entry defines each endpoint, key env, and default model. See
+  [`docs/model-capabilities-2026-07.md`](docs/model-capabilities-2026-07.md).
 - **`opencode` vendor** — headless dispatch through the OpenCode
   multi-provider aggregator CLI (`opencode run`). Advise mode pins OpenCode's
   built-in read-only `plan` agent; work mode uses `--auto`. Joins the default
@@ -115,8 +138,9 @@ omnilane ui start     # optional: watch jobs live in your browser
 ## 🧭 How it works
 
 omnilane lets the main loop of **any** agentic CLI classify subtasks into
-lanes and dispatch each lane to the best vendor's CLI — headlessly, using
-your existing subscription logins:
+lanes and dispatch each lane to the best vendor — headlessly, using your
+existing subscription logins (or, for the `openrouter` vendor, a direct API
+key with no extra CLI at all):
 
 ```mermaid
 flowchart LR
@@ -139,6 +163,8 @@ flowchart LR
   vendor without fallback.
 - **`skills/omnilane/SKILL.md`** — a single skill every harness can load:
   identify your own model, self-execute your lane, dispatch the rest.
+- **`omnilane mcp`** — the same routing surface as an MCP stdio server,
+  for hosts that integrate via MCP instead of skills.
 
 <div align="center">
 
@@ -244,8 +270,10 @@ Core routing does not need Python; only this UI requires Python 3.9 or newer.
 ## 📦 Install
 
 Requirements: the vendor CLIs you want to route to, logged in (`codex`,
-`claude`, `grok`, `agy`) and on `PATH` — install only the ones you have; the
-rest of the table degrades automatically.
+`claude`, `grok`, `agy`, and optionally `kimi`, `qwen`, `opencode`) and on
+`PATH` — install only the ones you have; the rest of the table degrades
+automatically. The `openrouter` vendor is the exception: it needs no CLI,
+only `curl` and an `OPENROUTER_API_KEY` in your environment.
 
 Quickest: `./install.sh` — symlinks the skill for the CLIs it finds, prints
 the plugin commands for the rest, shows your effective routing, and offers the
@@ -272,6 +300,32 @@ Rollback the installer-owned links and marked reminders with
 - **Antigravity**: `agy plugin install <this repo>` (check first with
   `agy plugin validate <this repo>`)
 
+### MCP server
+
+`omnilane mcp` starts a zero-dependency, local MCP stdio server so any
+MCP-capable host can discover and call omnilane without installing the skill or
+adding a routing reminder. Configure the host to launch the installed CLI:
+
+```json
+{
+  "mcpServers": {
+    "omnilane": {
+      "command": "omnilane",
+      "args": ["mcp"]
+    }
+  }
+}
+```
+
+The server exposes `route` plus read-only introspection: `list_lanes`,
+`explain`, `validate`, `dry_run`, `jobs_list`, `jobs_status`, `jobs_result`,
+`jobs_stats`, `jobs_audit`, and `doctor`. `route` defaults to read-only `advise` mode. Calls that select `work` must also
+provide an explicit `workdir`.
+
+Node.js is the only runtime requirement (no npm packages). If you prefer
+npm, `npm install -g omnilane` installs the CLI with the MCP server
+included.
+
 ## ⚙️ Configure
 
 Three layers, all optional:
@@ -281,6 +335,9 @@ Three layers, all optional:
    future models), and writes the result to `~/.omnilane/routing.local.yaml`.
    It intentionally skips the multi-vendor `consult` lane; edit that one by
    hand if needed. `install.sh` offers to run the menu at the end of a normal install.
+   For scripting, `configure set|get|unset|list|diff LANE [SPEC]` edits or inspects the same file
+   without a tty — `set` validates the lane and rejects an unsafe or structurally
+   invalid spec, rolling back on failure.
 2. **`~/.omnilane/routing.local.yaml`** — hand-edited overrides, same format
    as `routing.yaml`; local lines win. See `routing.local.yaml.example`.
 3. **`~/.omnilane/local.sh`** — per-machine binaries, proxies, auth wrappers;
@@ -299,6 +356,8 @@ omnilane list | route … | jobs … | configure   # global wrapper, works anywh
                                                # (install.sh links it into ~/.local/bin)
 eval "$(omnilane completion bash)"             # enable Bash completion for this shell
 source <(omnilane completion zsh)               # enable Zsh completion for this shell
+omnilane completion fish | source              # enable Fish completion for this shell
+omnilane mcp                                   # MCP stdio server (needs Node.js)
 omnilane release-audit [--target VERSION] [--json] # offline, read-only release gate
 omnilane ui start                              # start/reuse the local Live UI; print its URL
 omnilane ui status                             # report whether the Live UI is running
@@ -312,11 +371,15 @@ dispatch.sh [--json] --list [--json]
 dispatch.sh [--json] --explain LANE [--json]       # offline candidate-by-candidate decision trace
 dispatch.sh [--json] --validate [--json]           # lint effective routing; no provider calls
 jobs.sh [--json] {list | status ID | result ID}    # JSON result reports metadata, never bodies
+jobs.sh [--json] list [--lane L] [--vendor V] [--status running|done]  # filter the listing
 jobs.sh wait ID [--timeout N]                     # job exit; 124 timeout; 125 dead worker
-jobs.sh [--json] stats [--last N]                  # local success and routing aggregates
+jobs.sh cancel ID                                 # stop a running job: group SIGTERM, then SIGKILL
+jobs.sh rm ID                                     # delete one finished/dead job (refuses a running job)
+jobs.sh [--json] stats [--last N] [--lane L] [--vendor V]  # local success and routing aggregates
 jobs.sh audit [--last N] [--json]                  # read-only job integrity/privacy check
 jobs.sh prune [--keep N] [--apply]                # preview by default; completed jobs only
 configure.sh                                        # interactive lane menu
+configure.sh set|get|unset|list|diff LANE [SPEC]    # script/inspect routing.local.yaml, no tty
 ```
 
 **Big decisions can get a panel, not a person.** The `arbitrate` lane ships
@@ -343,11 +406,13 @@ code passes through.
 ## 🎭 Modes
 
 - **advise** (default) — read-only worker. Codex runs in a read-only sandbox;
-  Claude gets only Read/Glob/Grep; Grok runs in plan mode. Use for reviews,
-  questions, second opinions.
+  Claude gets only Read/Glob/Grep; Grok runs in plan mode; Kimi and OpenCode
+  pin their read-only plan modes; OpenRouter is advise-only by design (pure
+  inference). Use for reviews, questions, second opinions.
 - **work** — the worker may edit files, only inside the `--workdir` you name.
   Codex gets a workspace-write sandbox; Claude auto-accepts edits; Gemini runs
-  in accept-edits mode.
+  in accept-edits mode. The `openrouter` vendor refuses work mode with a clear
+  error — route edits to an agentic CLI vendor instead.
 
 ## 🔒 Safety rails
 
@@ -404,11 +469,15 @@ configurator and `routing.local.yaml` exist so you can disagree.
 
 ## 🌱 Status
 
-v0.5.1 keeps Codex `work` usable outside Git while bounding stalls with
-process-group cleanup, and synchronizes every public version surface. It builds
-on v0.5.0's installer, lifecycle, job-store, deadline, diagnostics, and release
-CI hardening. Grok/Antigravity command-shell behavior may still vary across CLI
-versions. Issues and PRs welcome.
+omnilane now spans thirteen dispatch vendors — four harness natives (codex,
+claude, grok, gemini), three aggregator/overflow CLIs (kimi, qwen, opencode),
+and six CLI-free OpenAI-compatible direct-API vendors (openrouter, deepseek,
+zai, mistral, groq, cerebras) — on the uniform runner contract with
+contract tests, plus the Claude Code `SessionStart`
+auto-reminder and an MCP stdio server surface (`omnilane mcp`). The direct-API
+and aggregator runners are contract-tested against fake binaries;
+real-model reports welcome. Grok/Antigravity command-shell behavior may still
+vary across CLI versions. Issues and PRs welcome.
 
 Project policies: [Contributing](CONTRIBUTING.md) · [Security](SECURITY.md) ·
 [Changelog](CHANGELOG.md)