release: v3.0.0-beta.4

fix(sisyphus-orchestrator): add cross-platform path validation for Windows support
Add isSisyphusPath() helper function that handles both forward slashes (Unix) and backslashes (Windows) using regex pattern /\.sisyphus[/\\]/. Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-opencode) Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
2026-01-11 05:46:20 +00:00 · 2026-01-11 14:42:53 +09:00 · 2026-01-11 14:33:17 +09:00 · 2026-01-11 05:03:55 +00:00 · 2026-01-11 12:33:20 +09:00 · 2026-01-11 12:23:13 +09:00
40 changed files with 1475 additions and 249 deletions
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -124,7 +124,7 @@ oh-my-opencode/
 |-------|---------------|---------|
 | Sisyphus | anthropic/claude-opus-4-5 | Primary orchestrator |
 | oracle | openai/gpt-5.2 | Read-only consultation. High-IQ debugging, architecture |
-| librarian | anthropic/claude-sonnet-4-5 | Multi-repo analysis, docs |
+| librarian | opencode/glm-4.7-free | Multi-repo analysis, docs |
 | explore | opencode/grok-code | Fast codebase exploration |
 | frontend-ui-ux-engineer | google/gemini-3-pro-preview | UI generation |
 | document-writer | google/gemini-3-pro-preview | Technical docs |
--- a/README.ja.md
+++ b/README.ja.md
@@ -28,7 +28,7 @@

 > `oh-my-opencode` をインストールして、ドーピングしたかのようにコーディングしましょう。バックグラウンドでエージェントを走らせ、oracle、librarian、frontend engineer のような専門エージェントを呼び出してください。丹精込めて作られた LSP/AST ツール、厳選された MCP、そして完全な Claude Code 互換レイヤーを、たった一行で手に入れましょう。

-**今すぐ始めましょう。ChatGPT、Claude、Gemini のサブスクリプションで使えます。**
+**注意: librarianには高価なモデルを使用しないでください。これはあなたにとって役に立たないだけでなく、LLMプロバイダーにも負担をかけます。代わりにClaude Haiku、Gemini Flash、GLM 4.7、MiniMaxなどのモデルを使用してください。**

 <div align="center">

@@ -76,6 +76,7 @@
 - [Oh My OpenCode](#oh-my-opencode)
  - [この Readme は読まなくていいです](#この-readme-は読まなくていいです)
    - [エージェントの時代ですから](#エージェントの時代ですから)
+    - [🪄 魔法の言葉：`ultrawork`](#-魔法の言葉ultrawork)
    - [読みたい方のために：シジフォスに会う](#読みたい方のためにシジフォスに会う)
      - [インストールするだけで。](#インストールするだけで)
  - [インストール](#インストール)
@@ -88,10 +89,15 @@
    - [ステップ 4: 認証情報の設定](#ステップ-4-認証情報の設定)
      - [4.1 Anthropic (Claude)](#41-anthropic-claude)
      - [4.2 Google Gemini (Antigravity OAuth)](#42-google-gemini-antigravity-oauth)
+      - [4.2.1 モデル設定](#421-モデル設定)
+      - [4.2.2 oh-my-opencode エージェントモデルのオーバーライド](#422-oh-my-opencode-エージェントモデルのオーバーライド)
      - [4.3 OpenAI (ChatGPT Plus/Pro)](#43-openai-chatgpt-pluspro)
+        - [モデル設定](#モデル設定)
    - [⚠️ 注意](#️-注意)
    - [セットアップの確認](#セットアップの確認)
    - [ユーザーに「おめでとうございます！🎉」と伝える](#ユーザーにおめでとうございますと伝える)
+    - [無料広告](#無料広告)
+    - [スターをお願いする ⭐](#スターをお願いする-)
    - [複雑すぎますか？](#複雑すぎますか)
  - [アンインストール](#アンインストール)
  - [機能](#機能)
@@ -99,7 +105,8 @@
    - [バックグラウンドエージェント: 本当のチームのように働く](#バックグラウンドエージェント-本当のチームのように働く)
    - [ツール: 同僚にはもっと良い道具を](#ツール-同僚にはもっと良い道具を)
      - [なぜあなただけ IDE を使っているのですか？](#なぜあなただけ-ide-を使っているのですか)
-      - [Context is all you need.](#context-is-all-you-need)
+      - [セッション管理](#セッション管理)
+      - [Context Is All You Need](#context-is-all-you-need)
      - [マルチモーダルを活用し、トークンは節約する](#マルチモーダルを活用しトークンは節約する)
      - [止まらないエージェントループ](#止まらないエージェントループ)
    - [Claude Code 互換性: さらば Claude Code、ようこそ OpenCode](#claude-code-互換性-さらば-claude-codeようこそ-opencode)
@@ -109,16 +116,20 @@
      - [互換性トグル](#互換性トグル)
    - [エージェントのためだけでなく、あなたのために](#エージェントのためだけでなくあなたのために)
  - [設定](#設定)
+    - [JSONC のサポート](#jsonc-のサポート)
    - [Google Auth](#google-auth)
    - [Agents](#agents)
      - [Permission オプション](#permission-オプション)
    - [Sisyphus Agent](#sisyphus-agent)
+    - [Background Tasks](#background-tasks)
    - [Hooks](#hooks)
    - [MCPs](#mcps)
    - [LSP](#lsp)
    - [Experimental](#experimental)
  - [作者のノート](#作者のノート)
  - [注意](#注意)
+  - [こちらの企業の専門家にご愛用いただいています](#こちらの企業の専門家にご愛用いただいています)
+  - [スポンサー](#スポンサー)

 # Oh My OpenCode

@@ -322,9 +333,9 @@ opencode auth login
 {
  "google_auth": false,
  "agents": {
-    "frontend-ui-ux-engineer": { "model": "google/gemini-3-pro-high" },
-    "document-writer": { "model": "google/gemini-3-flash" },
-    "multimodal-looker": { "model": "google/gemini-3-flash" }
+    "frontend-ui-ux-engineer": { "model": "google/antigravity-gemini-3-pro-high" },
+    "document-writer": { "model": "google/antigravity-gemini-3-flash" },
+    "multimodal-looker": { "model": "google/antigravity-gemini-3-flash" }
  }
 }
 ```
@@ -461,7 +472,7 @@ oh-my-opencode を削除するには：

 - **Sisyphus** (`anthropic/claude-opus-4-5`): **デフォルトエージェントです。** OpenCode のための強力な AI オーケストレーターです。専門のサブエージェントを活用して、複雑なタスクを計画、委任、実行します。バックグラウンドタスクへの委任と Todo ベースのワークフローを重視します。最大の推論能力を発揮するため、Claude Opus 4.5 と拡張思考 (32k token budget) を使用します。
 - **oracle** (`openai/gpt-5.2`): アーキテクチャ、コードレビュー、戦略立案のための専門アドバイザー。GPT-5.2 の卓越した論理的推論と深い分析能力を活用します。AmpCode からインスピレーションを得ました。
- **librarian** (`anthropic/claude-sonnet-4-5` または `google/gemini-3-flash`): マルチリポジトリ分析、ドキュメント検索、実装例の調査を担当。Antigravity 認証が設定されている場合は Gemini 3 Flash を使用し、それ以外は Claude Sonnet 4.5 を使用して、深いコードベース理解と GitHub リサーチ、根拠に基づいた回答を提供します。AmpCode からインスピレーションを得ました。
+- **librarian** (`opencode/glm-4.7-free`): マルチリポジトリ分析、ドキュメント検索、実装例の調査を担当。GLM-4.7 Free を使用して、深いコードベース理解と GitHub リサーチ、根拠に基づいた回答を提供します。AmpCode からインスピレーションを得ました。
 - **explore** (`opencode/grok-code`、`google/gemini-3-flash`、または `anthropic/claude-haiku-4-5`): 高速なコードベース探索、ファイルパターンマッチング。Antigravity 認証が設定されている場合は Gemini 3 Flash を使用し、Claude max20 が利用可能な場合は Haiku を使用し、それ以外は Grok を使います。Claude Code からインスピレーションを得ました。
 - **frontend-ui-ux-engineer** (`google/gemini-3-pro-preview`): 開発者に転身したデザイナーという設定です。素晴らしい UI を作ります。美しく独創的な UI コードを生成することに長けた Gemini を使用します。
 - **document-writer** (`google/gemini-3-pro-preview`): テクニカルライティングの専門家という設定です。Gemini は文筆家であり、流れるような文章を書きます。
@@ -721,10 +732,10 @@ Oh My OpenCode は以下の場所からフックを読み込んで実行しま
 1. `.opencode/oh-my-opencode.json` (プロジェクト)
 2. ユーザー設定（プラットフォーム別）：

-| プラットフォーム | ユーザー設定パス |
-|------------------|------------------|
-| **Windows** | `~/.config/opencode/oh-my-opencode.json` (推奨) または `%APPDATA%\opencode\oh-my-opencode.json` (fallback) |
-| **macOS/Linux** | `~/.config/opencode/oh-my-opencode.json` |
+| プラットフォーム | ユーザー設定パス                                                                                           |
+| ---------------- | ---------------------------------------------------------------------------------------------------------- |
+| **Windows**      | `~/.config/opencode/oh-my-opencode.json` (推奨) または `%APPDATA%\opencode\oh-my-opencode.json` (fallback) |
+| **macOS/Linux**  | `~/.config/opencode/oh-my-opencode.json`                                                                   |

 スキーマ自動補完がサポートされています：

@@ -748,10 +759,10 @@ Oh My OpenCode は以下の場所からフックを読み込んで実行しま
 ```jsonc
 {
  "$schema": "https://raw.githubusercontent.com/code-yeongyu/oh-my-opencode/master/assets/oh-my-opencode.schema.json",
-  
+
  // Antigravity OAuth 経由で Google Gemini を有効にする
  "google_auth": false,
-  
+
  /* エージェントのオーバーライド - 特定のタスクに合わせてモデルをカスタマイズ */
  "agents": {
    "oracle": {
@@ -774,9 +785,9 @@ Oh My OpenCode は以下の場所からフックを読み込んで実行しま
 {
  "google_auth": false,
  "agents": {
-    "frontend-ui-ux-engineer": { "model": "google/gemini-3-pro-high" },
-    "document-writer": { "model": "google/gemini-3-flash" },
-    "multimodal-looker": { "model": "google/gemini-3-flash" }
+    "frontend-ui-ux-engineer": { "model": "google/antigravity-gemini-3-pro-high" },
+    "document-writer": { "model": "google/antigravity-gemini-3-flash" },
+    "multimodal-looker": { "model": "google/antigravity-gemini-3-flash" }
  }
 }
 ```
@@ -841,13 +852,13 @@ Oh My OpenCode は以下の場所からフックを読み込んで実行しま
 }
 ```

-| Permission | 説明 | 値 |
-|------------|------|----|
-| `edit` | ファイル編集権限 | `ask` / `allow` / `deny` |
-| `bash` | Bash コマンド実行権限 | `ask` / `allow` / `deny` またはコマンド別: `{ "git": "allow", "rm": "deny" }` |
-| `webfetch` | ウェブアクセス権限 | `ask` / `allow` / `deny` |
-| `doom_loop` | 無限ループ検知のオーバーライド許可 | `ask` / `allow` / `deny` |
-| `external_directory` | プロジェクトルート外へのファイルアクセス | `ask` / `allow` / `deny` |
+| Permission           | 説明                                     | 値                                                                            |
+| -------------------- | ---------------------------------------- | ----------------------------------------------------------------------------- |
+| `edit`               | ファイル編集権限                         | `ask` / `allow` / `deny`                                                      |
+| `bash`               | Bash コマンド実行権限                    | `ask` / `allow` / `deny` またはコマンド別: `{ "git": "allow", "rm": "deny" }` |
+| `webfetch`           | ウェブアクセス権限                       | `ask` / `allow` / `deny`                                                      |
+| `doom_loop`          | 無限ループ検知のオーバーライド許可       | `ask` / `allow` / `deny`                                                      |
+| `external_directory` | プロジェクトルート外へのファイルアクセス | `ask` / `allow` / `deny`                                                      |

 または `~/.config/opencode/oh-my-opencode.json` か `.opencode/oh-my-opencode.json` の `disabled_agents` を使用して無効化できます：

@@ -925,12 +936,12 @@ Oh My OpenCode は以下の場所からフックを読み込んで実行しま
 }
 ```

-| オプション                  | デフォルト | 説明                                                                                                                                                         |
-| --------------------------- | ------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------ |
-| `disabled`                  | `false` | `true` の場合、すべての Sisyphus オーケストレーションを無効化し、元の build/plan をプライマリとして復元します。                                                                       |
-| `default_builder_enabled`   | `false` | `true` の場合、OpenCode-Builder エージェントを有効化します（OpenCode build と同じ、SDK 制限により名前変更）。デフォルトでは無効です。                                                   |
-| `planner_enabled`           | `true`  | `true` の場合、Prometheus (Planner) エージェントを有効化します（work-planner 方法論を含む）。デフォルトで有効です。                                                                   |
-| `replace_plan`              | `true`  | `true` の場合、デフォルトのプランエージェントをサブエージェントモードに降格させます。`false` に設定すると、Prometheus (Planner) とデフォルトのプランの両方を利用できます。                             |
+| オプション                | デフォルト | 説明                                                                                                                                                                       |
+| ------------------------- | ---------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `disabled`                | `false`    | `true` の場合、すべての Sisyphus オーケストレーションを無効化し、元の build/plan をプライマリとして復元します。                                                            |
+| `default_builder_enabled` | `false`    | `true` の場合、OpenCode-Builder エージェントを有効化します（OpenCode build と同じ、SDK 制限により名前変更）。デフォルトでは無効です。                                      |
+| `planner_enabled`         | `true`     | `true` の場合、Prometheus (Planner) エージェントを有効化します（work-planner 方法論を含む）。デフォルトで有効です。                                                        |
+| `replace_plan`            | `true`     | `true` の場合、デフォルトのプランエージェントをサブエージェントモードに降格させます。`false` に設定すると、Prometheus (Planner) とデフォルトのプランの両方を利用できます。 |

 ### Background Tasks

@@ -953,10 +964,10 @@ Oh My OpenCode は以下の場所からフックを読み込んで実行しま
 }
 ```

-| オプション            | デフォルト | 説明                                                                                                           |
-| --------------------- | ---------- | -------------------------------------------------------------------------------------------------------------- |
-| `defaultConcurrency`  | -          | すべてのプロバイダー/モデルに対するデフォルトの最大同時バックグラウンドタスク数                                 |
-| `providerConcurrency` | -          | プロバイダーごとの同時実行制限。キーはプロバイダー名（例：`anthropic`、`openai`、`google`）                     |
+| オプション            | デフォルト | 説明                                                                                                                  |
+| --------------------- | ---------- | --------------------------------------------------------------------------------------------------------------------- |
+| `defaultConcurrency`  | -          | すべてのプロバイダー/モデルに対するデフォルトの最大同時バックグラウンドタスク数                                       |
+| `providerConcurrency` | -          | プロバイダーごとの同時実行制限。キーはプロバイダー名（例：`anthropic`、`openai`、`google`）                           |
 | `modelConcurrency`    | -          | モデルごとの同時実行制限。キーは完全なモデル名（例：`anthropic/claude-opus-4-5`）。プロバイダー制限より優先されます。 |

 **優先順位**: `modelConcurrency` > `providerConcurrency` > `defaultConcurrency`
@@ -1035,13 +1046,13 @@ OpenCode でサポートされるすべての LSP 構成およびカスタム設
 }
 ```

-| オプション                        | デフォルト | 説明                                                                                                                                                                   |
-| --------------------------------- | ---------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| `preemptive_compaction_threshold` | `0.85`     | プリエンプティブコンパクションをトリガーする閾値（0.5-0.95）。`preemptive-compaction` フックはデフォルトで有効です。このオプションで閾値をカスタマイズできます。                 |
+| オプション                        | デフォルト | 説明                                                                                                                                                                               |
+| --------------------------------- | ---------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `preemptive_compaction_threshold` | `0.85`     | プリエンプティブコンパクションをトリガーする閾値（0.5-0.95）。`preemptive-compaction` フックはデフォルトで有効です。このオプションで閾値をカスタマイズできます。                   |
 | `truncate_all_tool_outputs`       | `false`    | ホワイトリストのツール（Grep、Glob、LSP、AST-grep）だけでなく、すべてのツール出力を切り詰めます。Tool output truncator はデフォルトで有効です - `disabled_hooks`で無効化できます。 |
-| `aggressive_truncation`           | `false`    | トークン制限を超えた場合、ツール出力を積極的に切り詰めて制限内に収めます。デフォルトの切り詰めより積極的です。不十分な場合は要約/復元にフォールバックします。                 |
-| `auto_resume`                     | `false`    | thinking block エラーや thinking disabled violation からの回復成功後、自動的にセッションを再開します。最後のユーザーメッセージを抽出して続行します。                        |
-| `dcp_for_compaction`              | `false`    | コンパクション用DCP（動的コンテキスト整理）を有効化 - トークン制限超過時に最初に実行されます。コンパクション前に重複したツール呼び出しと古いツール出力を整理します。                |
+| `aggressive_truncation`           | `false`    | トークン制限を超えた場合、ツール出力を積極的に切り詰めて制限内に収めます。デフォルトの切り詰めより積極的です。不十分な場合は要約/復元にフォールバックします。                      |
+| `auto_resume`                     | `false`    | thinking block エラーや thinking disabled violation からの回復成功後、自動的にセッションを再開します。最後のユーザーメッセージを抽出して続行します。                               |
+| `dcp_for_compaction`              | `false`    | コンパクション用DCP（動的コンテキスト整理）を有効化 - トークン制限超過時に最初に実行されます。コンパクション前に重複したツール呼び出しと古いツール出力を整理します。               |

 **警告**：これらの機能は実験的であり、予期しない動作を引き起こす可能性があります。影響を理解した場合にのみ有効にしてください。

--- a/README.md
+++ b/README.md
@@ -29,10 +29,7 @@
 > This is coding on steroids—`oh-my-opencode` in action. Run background agents, call specialized agents like oracle, librarian, and frontend engineer. Use crafted LSP/AST tools, curated MCPs, and a full Claude Code compatibility layer.


-No stupid token consumption massive subagents here. No bloat tools here.
-
-**Certified, Verified, Tested, Actually Useful Harness in Production, after $24,000 worth of tokens spent.**
-**START WITH YOUR ChatGPT, Claude, Gemini SUBSCRIPTIONS. WE ALL COVER THEM.**
+**Notice: Do not use expensive models for librarian. This is not only unhelpful to you, but also burdens LLM providers. Use models like Claude Haiku, Gemini Flash, GLM 4.7, or MiniMax instead.**


 <div align="center">
@@ -128,6 +125,7 @@ No stupid token consumption massive subagents here. No bloat tools here.
    - [Agents](#agents)
      - [Permission Options](#permission-options)
    - [Built-in Skills](#built-in-skills)
+    - [Git Master](#git-master)
    - [Sisyphus Agent](#sisyphus-agent)
    - [Background Tasks](#background-tasks)
    - [Categories](#categories)
@@ -135,6 +133,7 @@ No stupid token consumption massive subagents here. No bloat tools here.
    - [MCPs](#mcps)
    - [LSP](#lsp)
    - [Experimental](#experimental)
+    - [Environment Variables](#environment-variables)
  - [Author's Note](#authors-note)
  - [Warnings](#warnings)
  - [Loved by professionals at](#loved-by-professionals-at)
@@ -361,9 +360,9 @@ The `opencode-antigravity-auth` plugin uses different model names than the built
 {
  "google_auth": false,
  "agents": {
-    "frontend-ui-ux-engineer": { "model": "google/gemini-3-pro-high" },
-    "document-writer": { "model": "google/gemini-3-flash" },
-    "multimodal-looker": { "model": "google/gemini-3-flash" }
+    "frontend-ui-ux-engineer": { "model": "google/antigravity-gemini-3-pro-high" },
+    "document-writer": { "model": "google/antigravity-gemini-3-flash" },
+    "multimodal-looker": { "model": "google/antigravity-gemini-3-flash" }
  }
 }
 ```
@@ -499,9 +498,9 @@ To remove oh-my-opencode:

 - **Sisyphus** (`anthropic/claude-opus-4-5`): **The default agent.** A powerful AI orchestrator for OpenCode. Plans, delegates, and executes complex tasks using specialized subagents with aggressive parallel execution. Emphasizes background task delegation and todo-driven workflow. Uses Claude Opus 4.5 with extended thinking (32k budget) for maximum reasoning capability.
 - **oracle** (`openai/gpt-5.2`): Architecture, code review, strategy. Uses GPT-5.2 for its stellar logical reasoning and deep analysis. Inspired by AmpCode.
- **librarian** (`anthropic/claude-sonnet-4-5` or `google/gemini-3-flash`): Multi-repo analysis, doc lookup, implementation examples. Uses Gemini 3 Flash when Antigravity auth is configured, otherwise Claude Sonnet 4.5 for deep codebase understanding and GitHub research with evidence-based answers. Inspired by AmpCode.
+- **librarian** (`opencode/glm-4.7-free`): Multi-repo analysis, doc lookup, implementation examples. Uses GLM-4.7 Free for deep codebase understanding and GitHub research with evidence-based answers. Inspired by AmpCode.
 - **explore** (`opencode/grok-code`, `google/gemini-3-flash`, or `anthropic/claude-haiku-4-5`): Fast codebase exploration and pattern matching. Uses Gemini 3 Flash when Antigravity auth is configured, Haiku when Claude max20 is available, otherwise Grok. Inspired by Claude Code.
- **frontend-ui-ux-engineer** (`google/gemini-3-pro-high`): A designer turned developer. Builds gorgeous UIs. Gemini excels at creative, beautiful UI code.
+- **frontend-ui-ux-engineer** (`google/gemini-3-pro-preview`): A designer turned developer. Builds gorgeous UIs. Gemini excels at creative, beautiful UI code.
 - **document-writer** (`google/gemini-3-flash`): Technical writing expert. Gemini is a wordsmith—writes prose that flows.
 - **multimodal-looker** (`google/gemini-3-flash`): Visual content specialist. Analyzes PDFs, images, diagrams to extract information.

@@ -830,9 +829,9 @@ When using `opencode-antigravity-auth`, disable the built-in auth and override a
 {
  "google_auth": false,
  "agents": {
-    "frontend-ui-ux-engineer": { "model": "google/gemini-3-pro-high" },
-    "document-writer": { "model": "google/gemini-3-flash" },
-    "multimodal-looker": { "model": "google/gemini-3-flash" }
+    "frontend-ui-ux-engineer": { "model": "google/antigravity-gemini-3-pro-high" },
+    "document-writer": { "model": "google/antigravity-gemini-3-flash" },
+    "multimodal-looker": { "model": "google/antigravity-gemini-3-flash" }
  }
 }
 ```
@@ -945,10 +944,10 @@ Configure git-master skill behavior:
 }
 ```

-| Option | Default | Description |
-| ------ | ------- | ----------- |
-| `commit_footer` | `true` | Adds "Ultraworked with Sisyphus" footer to commit messages. |
-| `include_co_authored_by` | `true` | Adds `Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>` trailer to commits. |
+| Option                   | Default | Description                                                                      |
+| ------------------------ | ------- | -------------------------------------------------------------------------------- |
+| `commit_footer`          | `true`  | Adds "Ultraworked with Sisyphus" footer to commit messages.                      |
+| `include_co_authored_by` | `true`  | Adds `Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>` trailer to commits. |

 ### Sisyphus Agent

@@ -1016,12 +1015,12 @@ You can also customize Sisyphus agents like other agents:
 }
 ```

-| Option                      | Default | Description                                                                                                                                         |
-| --------------------------- | ------- | --------------------------------------------------------------------------------------------------------------------------------------------------- |
-| `disabled`                  | `false` | When `true`, disables all Sisyphus orchestration and restores original build/plan as primary.                                                       |
-| `default_builder_enabled`   | `false` | When `true`, enables OpenCode-Builder agent (same as OpenCode build, renamed due to SDK limitations). Disabled by default.                         |
-| `planner_enabled`           | `true`  | When `true`, enables Prometheus (Planner) agent with work-planner methodology. Enabled by default.                                                 |
-| `replace_plan`              | `true`  | When `true`, demotes default plan agent to subagent mode. Set to `false` to keep both Prometheus (Planner) and default plan available.             |
+| Option                    | Default | Description                                                                                                                            |
+| ------------------------- | ------- | -------------------------------------------------------------------------------------------------------------------------------------- |
+| `disabled`                | `false` | When `true`, disables all Sisyphus orchestration and restores original build/plan as primary.                                          |
+| `default_builder_enabled` | `false` | When `true`, enables OpenCode-Builder agent (same as OpenCode build, renamed due to SDK limitations). Disabled by default.             |
+| `planner_enabled`         | `true`  | When `true`, enables Prometheus (Planner) agent with work-planner methodology. Enabled by default.                                     |
+| `replace_plan`            | `true`  | When `true`, demotes default plan agent to subagent mode. Set to `false` to keep both Prometheus (Planner) and default plan available. |

 ### Background Tasks

@@ -1063,10 +1062,10 @@ Categories enable domain-specific task delegation via the `sisyphus_task` tool.

 **Default Categories:**

-| Category | Model | Description |
-|----------|-------|-------------|
-| `visual` | `google/gemini-3-pro-preview` | Frontend, UI/UX, design-focused tasks. High creativity (temp 0.7). |
-| `business-logic` | `openai/gpt-5.2` | Backend logic, architecture, strategic reasoning. Low creativity (temp 0.1). |
+| Category         | Model                         | Description                                                                  |
+| ---------------- | ----------------------------- | ---------------------------------------------------------------------------- |
+| `visual`         | `google/gemini-3-pro-preview` | Frontend, UI/UX, design-focused tasks. High creativity (temp 0.7).           |
+| `business-logic` | `openai/gpt-5.2`              | Backend logic, architecture, strategic reasoning. Low creativity (temp 0.1). |

 **Usage:**

@@ -1092,7 +1091,7 @@ Add custom categories in `oh-my-opencode.json`:
      "prompt_append": "Focus on data analysis, ML pipelines, and statistical methods."
    },
    "visual": {
-      "model": "google/gemini-3-pro-high",
+      "model": "google/gemini-3-pro-preview",
      "prompt_append": "Use shadcn/ui components and Tailwind CSS."
    }
  }
@@ -1181,6 +1180,12 @@ Opt-in experimental features that may change or be removed in future versions. U

 **Warning**: These features are experimental and may cause unexpected behavior. Enable only if you understand the implications.

+### Environment Variables
+
+| Variable              | Description                                                                                                                                     |
+| --------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------- |
+| `OPENCODE_CONFIG_DIR` | Override the OpenCode configuration directory. Useful for profile isolation with tools like [OCX](https://github.com/kdcokenny/ocx) ghost mode. |
+

 ## Author's Note

--- a/README.zh-cn.md
+++ b/README.zh-cn.md
@@ -28,9 +28,7 @@
 > 装上 `oh-my-opencode`，编程体验直接起飞。后台跑着一堆 Agent，随时呼叫 Oracle、Librarian、Frontend Engineer 这些专家。精心打磨的 LSP/AST 工具、精选 MCP、完美的 Claude Code 兼容层——一行配置，全套带走。

 这里没有为了显摆而疯狂烧 Token 的臃肿 Subagent。没有垃圾工具。
-
-**这是烧了 24,000 美元 Token 换来的、真正经过生产环境验证、测试、靠谱的 Harness。**
-**拿着你的 ChatGPT、Claude、Gemini 订阅直接就能用。我们全包圆了。**
+**注意：请勿为 librarian 使用昂贵的模型。这不仅对你没有帮助,还会给 LLM 提供商带来负担。请使用 Claude Haiku、Gemini Flash、GLM 4.7 或 MiniMax 等模型。**

 <div align="center">

@@ -78,6 +76,7 @@
 - [Oh My OpenCode](#oh-my-opencode)
  - [太长不看？(TL;DR)](#太长不看tldr)
    - [现在是 Agent 的时代](#现在是-agent-的时代)
+    - [🪄 魔法口令：`ultrawork`](#-魔法口令ultrawork)
    - [如果你真的想读读看：认识西西弗斯](#如果你真的想读读看认识西西弗斯)
      - [闭眼装就行](#闭眼装就行)
  - [安装](#安装)
@@ -90,10 +89,15 @@
    - [步骤 4：搞定认证](#步骤-4搞定认证)
      - [4.1 Anthropic (Claude)](#41-anthropic-claude)
      - [4.2 Google Gemini (Antigravity OAuth)](#42-google-gemini-antigravity-oauth)
+        - [模型配置](#模型配置)
+        - [oh-my-opencode Agent 模型覆盖](#oh-my-opencode-agent-模型覆盖)
      - [4.3 OpenAI (ChatGPT Plus/Pro)](#43-openai-chatgpt-pluspro)
+        - [模型配置](#模型配置-1)
    - [⚠️ 注意](#️-注意)
    - [检查作业](#检查作业)
    - [跟用户说"恭喜！🎉"](#跟用户说恭喜)
+    - [免费广告](#免费广告)
+    - [求个 Star ⭐](#求个-star-)
    - [太麻烦了？](#太麻烦了)
  - [卸载](#卸载)
  - [功能](#功能)
@@ -101,6 +105,7 @@
    - [后台 Agent：像真正的团队一样干活](#后台-agent像真正的团队一样干活)
    - [工具：给队友配点好的](#工具给队友配点好的)
      - [凭什么只有你能用 IDE？](#凭什么只有你能用-ide)
+      - [会话管理 (Session Management)](#会话管理-session-management)
      - [上下文就是一切 (Context is all you need)](#上下文就是一切-context-is-all-you-need)
      - [多模态全开，Token 省着用](#多模态全开token-省着用)
      - [根本停不下来的 Agent Loop](#根本停不下来的-agent-loop)
@@ -111,16 +116,20 @@
      - [兼容性开关](#兼容性开关)
    - [不只是为了 Agent，也是为了你](#不只是为了-agent也是为了你)
  - [配置](#配置)
+    - [JSONC 支持](#jsonc-支持)
    - [Google Auth](#google-auth)
    - [Agents](#agents)
      - [权限选项](#权限选项)
    - [Sisyphus Agent](#sisyphus-agent)
+    - [Background Tasks（后台任务）](#background-tasks后台任务)
    - [Hooks](#hooks)
    - [MCPs](#mcps)
    - [LSP](#lsp)
    - [Experimental](#experimental)
  - [作者的话](#作者的话)
  - [注意事项](#注意事项)
+  - [以下企业的专业人士都在用](#以下企业的专业人士都在用)
+  - [赞助者](#赞助者)

 # Oh My OpenCode

@@ -327,9 +336,9 @@ opencode auth login
 {
  "google_auth": false,
  "agents": {
-    "frontend-ui-ux-engineer": { "model": "google/gemini-3-pro-high" },
-    "document-writer": { "model": "google/gemini-3-flash" },
-    "multimodal-looker": { "model": "google/gemini-3-flash" }
+    "frontend-ui-ux-engineer": { "model": "google/antigravity-gemini-3-pro-high" },
+    "document-writer": { "model": "google/antigravity-gemini-3-flash" },
+    "multimodal-looker": { "model": "google/antigravity-gemini-3-flash" }
  }
 }
 ```
@@ -466,7 +475,7 @@ gh repo star code-yeongyu/oh-my-opencode

 - **Sisyphus** (`anthropic/claude-opus-4-5`)：**默认 Agent。** OpenCode 专属的强力 AI 编排器。指挥专业子 Agent 搞定复杂任务。主打后台任务委派和 Todo 驱动。用 Claude Opus 4.5 加上扩展思考（32k token 预算），智商拉满。
 - **oracle** (`openai/gpt-5.2`)：架构师、代码审查员、战略家。GPT-5.2 的逻辑推理和深度分析能力不是盖的。致敬 AmpCode。
- **librarian** (`anthropic/claude-sonnet-4-5` 或 `google/gemini-3-flash`)：多仓库分析、查文档、找示例。配置 Antigravity 认证时使用 Gemini 3 Flash，否则使用 Claude Sonnet 4.5 深入理解代码库，GitHub 调研，给出的答案都有据可查。致敬 AmpCode。
+- **librarian** (`opencode/glm-4.7-free`)：多仓库分析、查文档、找示例。使用 GLM-4.7 Free 深入理解代码库，GitHub 调研，给出的答案都有据可查。致敬 AmpCode。
 - **explore** (`opencode/grok-code`、`google/gemini-3-flash` 或 `anthropic/claude-haiku-4-5`)：极速代码库扫描、模式匹配。配置 Antigravity 认证时使用 Gemini 3 Flash，Claude max20 可用时使用 Haiku，否则用 Grok。致敬 Claude Code。
 - **frontend-ui-ux-engineer** (`google/gemini-3-pro-preview`)：设计师出身的程序员。UI 做得那是真漂亮。Gemini 写这种创意美观的代码是一绝。
 - **document-writer** (`google/gemini-3-pro-preview`)：技术写作专家。Gemini 文笔好，写出来的东西读着顺畅。
@@ -722,10 +731,10 @@ Agent 爽了，你自然也爽。但我还想直接让你爽。
 1. `.opencode/oh-my-opencode.json`（项目级）
 2. 用户配置（按平台）：

-| 平台 | 用户配置路径 |
-|----------|------------------|
-| **Windows** | `~/.config/opencode/oh-my-opencode.json` (首选) 或 `%APPDATA%\opencode\oh-my-opencode.json` (备选) |
-| **macOS/Linux** | `~/.config/opencode/oh-my-opencode.json` |
+| 平台            | 用户配置路径                                                                                       |
+| --------------- | -------------------------------------------------------------------------------------------------- |
+| **Windows**     | `~/.config/opencode/oh-my-opencode.json` (首选) 或 `%APPDATA%\opencode\oh-my-opencode.json` (备选) |
+| **macOS/Linux** | `~/.config/opencode/oh-my-opencode.json`                                                           |

 支持 Schema 自动补全：

@@ -749,10 +758,10 @@ Agent 爽了，你自然也爽。但我还想直接让你爽。
 ```jsonc
 {
  "$schema": "https://raw.githubusercontent.com/code-yeongyu/oh-my-opencode/master/assets/oh-my-opencode.schema.json",
-  
+
  // 通过 Antigravity OAuth 启用 Google Gemini
  "google_auth": false,
-  
+
  /* Agent 覆盖 - 为特定任务自定义模型 */
  "agents": {
    "oracle": {
@@ -775,9 +784,9 @@ Agent 爽了，你自然也爽。但我还想直接让你爽。
 {
  "google_auth": false,
  "agents": {
-    "frontend-ui-ux-engineer": { "model": "google/gemini-3-pro-high" },
-    "document-writer": { "model": "google/gemini-3-flash" },
-    "multimodal-looker": { "model": "google/gemini-3-flash" }
+    "frontend-ui-ux-engineer": { "model": "google/antigravity-gemini-3-pro-high" },
+    "document-writer": { "model": "google/antigravity-gemini-3-flash" },
+    "multimodal-looker": { "model": "google/antigravity-gemini-3-flash" }
  }
 }
 ```
@@ -842,13 +851,13 @@ Agent 爽了，你自然也爽。但我还想直接让你爽。
 }
 ```

-| Permission           | 说明                     | 值                                                                   |
-| -------------------- | ------------------------ | -------------------------------------------------------------------- |
-| `edit`               | 改文件                   | `ask` / `allow` / `deny`                                             |
-| `bash`               | 跑 Bash 命令             | `ask` / `allow` / `deny` 或按命令：`{ "git": "allow", "rm": "deny" }` |
-| `webfetch`           | 上网                     | `ask` / `allow` / `deny`                                             |
-| `doom_loop`          | 覆盖无限循环检测         | `ask` / `allow` / `deny`                                             |
-| `external_directory` | 访问根目录外面的文件     | `ask` / `allow` / `deny`                                             |
+| Permission           | 说明                 | 值                                                                    |
+| -------------------- | -------------------- | --------------------------------------------------------------------- |
+| `edit`               | 改文件               | `ask` / `allow` / `deny`                                              |
+| `bash`               | 跑 Bash 命令         | `ask` / `allow` / `deny` 或按命令：`{ "git": "allow", "rm": "deny" }` |
+| `webfetch`           | 上网                 | `ask` / `allow` / `deny`                                              |
+| `doom_loop`          | 覆盖无限循环检测     | `ask` / `allow` / `deny`                                              |
+| `external_directory` | 访问根目录外面的文件 | `ask` / `allow` / `deny`                                              |

 或者在 `~/.config/opencode/oh-my-opencode.json` 或 `.opencode/oh-my-opencode.json` 的 `disabled_agents` 里直接禁了：

@@ -926,12 +935,12 @@ Sisyphus Agent 也能自定义：
 }
 ```

-| 选项                        | 默认值   | 说明                                                                                                                                              |
-| --------------------------- | ------- | ------------------------------------------------------------------------------------------------------------------------------------------------- |
-| `disabled`                  | `false` | 设为 `true` 就禁用所有 Sisyphus 编排，恢复原来的 build/plan。                                                                                              |
-| `default_builder_enabled`   | `false` | 设为 `true` 就启用 OpenCode-Builder Agent（与 OpenCode build 相同，因 SDK 限制仅改名）。默认禁用。                                                           |
-| `planner_enabled`           | `true`  | 设为 `true` 就启用 Prometheus (Planner) Agent（含 work-planner 方法论）。默认启用。                                                                         |
-| `replace_plan`              | `true`  | 设为 `true` 就把默认计划 Agent 降级为子 Agent 模式。设为 `false` 可以同时保留 Prometheus (Planner) 和默认计划。                                                      |
+| 选项                      | 默认值  | 说明                                                                                                            |
+| ------------------------- | ------- | --------------------------------------------------------------------------------------------------------------- |
+| `disabled`                | `false` | 设为 `true` 就禁用所有 Sisyphus 编排，恢复原来的 build/plan。                                                   |
+| `default_builder_enabled` | `false` | 设为 `true` 就启用 OpenCode-Builder Agent（与 OpenCode build 相同，因 SDK 限制仅改名）。默认禁用。              |
+| `planner_enabled`         | `true`  | 设为 `true` 就启用 Prometheus (Planner) Agent（含 work-planner 方法论）。默认启用。                             |
+| `replace_plan`            | `true`  | 设为 `true` 就把默认计划 Agent 降级为子 Agent 模式。设为 `false` 可以同时保留 Prometheus (Planner) 和默认计划。 |

 ### Background Tasks（后台任务）

@@ -954,11 +963,11 @@ Sisyphus Agent 也能自定义：
 }
 ```

-| 选项                  | 默认值 | 说明                                                                                                           |
-| --------------------- | ------ | -------------------------------------------------------------------------------------------------------------- |
-| `defaultConcurrency`  | -      | 所有提供商/模型的默认最大并发后台任务数                                                                        |
-| `providerConcurrency` | -      | 按提供商设置并发限制。键是提供商名称（例如：`anthropic`、`openai`、`google`）                                  |
-| `modelConcurrency`    | -      | 按模型设置并发限制。键是完整的模型名称（例如：`anthropic/claude-opus-4-5`）。会覆盖提供商级别的限制。          |
+| 选项                  | 默认值 | 说明                                                                                                  |
+| --------------------- | ------ | ----------------------------------------------------------------------------------------------------- |
+| `defaultConcurrency`  | -      | 所有提供商/模型的默认最大并发后台任务数                                                               |
+| `providerConcurrency` | -      | 按提供商设置并发限制。键是提供商名称（例如：`anthropic`、`openai`、`google`）                         |
+| `modelConcurrency`    | -      | 按模型设置并发限制。键是完整的模型名称（例如：`anthropic/claude-opus-4-5`）。会覆盖提供商级别的限制。 |

 **优先级顺序**: `modelConcurrency` > `providerConcurrency` > `defaultConcurrency`

@@ -1036,13 +1045,13 @@ Oh My OpenCode 送你重构工具（重命名、代码操作）。
 }
 ```

-| 选项                              | 默认值  | 说明                                                                                                                                           |
-| --------------------------------- | ------- | ---------------------------------------------------------------------------------------------------------------------------------------------- |
-| `preemptive_compaction_threshold` | `0.85`  | 触发预防性压缩的阈值比例（0.5-0.95）。`preemptive-compaction` 钩子默认启用；此选项用于自定义阈值。                                                     |
-| `truncate_all_tool_outputs`       | `false` | 截断所有工具输出，而不仅仅是白名单工具（Grep、Glob、LSP、AST-grep）。Tool output truncator 默认启用 - 使用 `disabled_hooks` 禁用。                    |
-| `aggressive_truncation`           | `false` | 超出 token 限制时，激进地截断工具输出以适应限制。比默认截断更激进。不够的话会回退到摘要/恢复。                                                     |
-| `auto_resume`                     | `false` | 从 thinking block 错误或 thinking disabled violation 成功恢复后，自动恢复会话。提取最后一条用户消息继续执行。                                     |
-| `dcp_for_compaction`              | `false` | 启用压缩用 DCP（动态上下文剪枝）- 在超出 token 限制时首先执行。在压缩前清理重复的工具调用和旧的工具输出。                                            |
+| 选项                              | 默认值  | 说明                                                                                                                               |
+| --------------------------------- | ------- | ---------------------------------------------------------------------------------------------------------------------------------- |
+| `preemptive_compaction_threshold` | `0.85`  | 触发预防性压缩的阈值比例（0.5-0.95）。`preemptive-compaction` 钩子默认启用；此选项用于自定义阈值。                                 |
+| `truncate_all_tool_outputs`       | `false` | 截断所有工具输出，而不仅仅是白名单工具（Grep、Glob、LSP、AST-grep）。Tool output truncator 默认启用 - 使用 `disabled_hooks` 禁用。 |
+| `aggressive_truncation`           | `false` | 超出 token 限制时，激进地截断工具输出以适应限制。比默认截断更激进。不够的话会回退到摘要/恢复。                                     |
+| `auto_resume`                     | `false` | 从 thinking block 错误或 thinking disabled violation 成功恢复后，自动恢复会话。提取最后一条用户消息继续执行。                      |
+| `dcp_for_compaction`              | `false` | 启用压缩用 DCP（动态上下文剪枝）- 在超出 token 限制时首先执行。在压缩前清理重复的工具调用和旧的工具输出。                          |

 **警告**：这些功能是实验性的，可能会导致意外行为。只有在理解其影响的情况下才启用。

--- a/docs/orchestration-guide.md
+++ b/docs/orchestration-guide.md
@@ -1,5 +1,26 @@
 # Oh-My-OpenCode Orchestration Guide

+## TL;DR - When to Use What
+
+| Complexity | Approach | When to Use |
+|------------|----------|-------------|
+| **Simple** | Just prompt | Simple tasks, quick fixes, single-file changes |
+| **Complex + Lazy** | Just type `ulw` or `ultrawork` | Complex tasks where explaining context is tedious. Agent figures it out. |
+| **Complex + Precise** | `@plan` → `/start-work` | Precise, multi-step work requiring true orchestration. Prometheus plans, Sisyphus executes. |
+
+**Decision Flow:**
+```
+Is it a quick fix or simple task?
+  └─ YES → Just prompt normally
+  └─ NO  → Is explaining the full context tedious?
+             └─ YES → Type "ulw" and let the agent figure it out
+             └─ NO  → Do you need precise, verifiable execution?
+                        └─ YES → Use @plan for Prometheus planning, then /start-work
+                        └─ NO  → Just use "ulw"
+```
+
+---
+
 This document provides a comprehensive guide to the orchestration system that implements Oh-My-OpenCode's core philosophy: **"Separation of Planning and Execution"**.

 ## 1. Overview
@@ -16,7 +37,7 @@ Oh-My-OpenCode solves this by clearly separating two roles:
 ## 2. Overall Architecture

 ```mermaid
-graph TD
+flowchart TD
    User[User Request] --> Prometheus
    
    subgraph Planning Phase
@@ -24,10 +45,10 @@ graph TD
        Metis --> Prometheus
        Prometheus --> Momus[Momus<br>Reviewer]
        Momus --> Prometheus
-        Prometheus --> PlanFile[/.sisyphus/plans/*.md]
+        Prometheus --> PlanFile["/.sisyphus/plans/{name}.md"]
    end
    
-    PlanFile --> StartWork[/start-work]
+    PlanFile --> StartWork[//start-work/]
    StartWork --> BoulderState[boulder.json]
    
    subgraph Execution Phase
@@ -93,9 +114,9 @@ When the user enters `/start-work`, the execution phase begins.

 ## 5. Commands and Usage

-### `/plan [request]`
+### `@plan [request]`
 Invokes Prometheus to start a planning session.
- Example: `/plan "I want to refactor the authentication system to NextAuth"`
+- Example: `@plan "I want to refactor the authentication system to NextAuth"`

 ### `/start-work`
 Executes the generated plan.
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
  "name": "oh-my-opencode",
-  "version": "3.0.0-beta.2",
+  "version": "3.0.0-beta.4",
  "description": "The Best AI Agent Harness - Batteries-Included OpenCode Plugin with Multi-Model Orchestration, Parallel Background Agents, and Crafted LSP/AST Tools",
  "main": "dist/index.js",
  "types": "dist/index.d.ts",
--- a/signatures/cla.json
+++ b/signatures/cla.json
@@ -319,6 +319,94 @@
      "created_at": "2026-01-08T20:18:27Z",
      "repoId": 1108837393,
      "pullRequestNo": 603
+    },
+    {
+      "name": "SJY0917032",
+      "id": 88534701,
+      "comment_id": 3728199745,
+      "created_at": "2026-01-09T10:01:19Z",
+      "repoId": 1108837393,
+      "pullRequestNo": 625
+    },
+    {
+      "name": "kdcokenny",
+      "id": 99611484,
+      "comment_id": 3728801075,
+      "created_at": "2026-01-09T12:54:05Z",
+      "repoId": 1108837393,
+      "pullRequestNo": 629
+    },
+    {
+      "name": "ElwinLiu",
+      "id": 87802244,
+      "comment_id": 3731812585,
+      "created_at": "2026-01-10T04:32:16Z",
+      "repoId": 1108837393,
+      "pullRequestNo": 645
+    },
+    {
+      "name": "Luodian",
+      "id": 15847405,
+      "comment_id": 3731833107,
+      "created_at": "2026-01-10T05:01:16Z",
+      "repoId": 1108837393,
+      "pullRequestNo": 634
+    },
+    {
+      "name": "imarshallwidjaja",
+      "id": 60992624,
+      "comment_id": 3732124681,
+      "created_at": "2026-01-10T07:58:43Z",
+      "repoId": 1108837393,
+      "pullRequestNo": 648
+    },
+    {
+      "name": "GollyJer",
+      "id": 689204,
+      "comment_id": 3732253764,
+      "created_at": "2026-01-10T09:33:21Z",
+      "repoId": 1108837393,
+      "pullRequestNo": 649
+    },
+    {
+      "name": "kargnas",
+      "id": 1438533,
+      "comment_id": 3732344143,
+      "created_at": "2026-01-10T10:25:25Z",
+      "repoId": 1108837393,
+      "pullRequestNo": 653
+    },
+    {
+      "name": "ashir6892",
+      "id": 52703606,
+      "comment_id": 3733435826,
+      "created_at": "2026-01-10T19:50:07Z",
+      "repoId": 1108837393,
+      "pullRequestNo": 675
+    },
+    {
+      "name": "arthur404dev",
+      "id": 59490008,
+      "comment_id": 3733697071,
+      "created_at": "2026-01-10T23:51:44Z",
+      "repoId": 1108837393,
+      "pullRequestNo": 676
+    },
+    {
+      "name": "KNN-07",
+      "id": 55886589,
+      "comment_id": 3733788592,
+      "created_at": "2026-01-11T01:11:38Z",
+      "repoId": 1108837393,
+      "pullRequestNo": 679
+    },
+    {
+      "name": "aw338WoWmUI",
+      "id": 121638634,
+      "comment_id": 3734013343,
+      "created_at": "2026-01-11T04:56:38Z",
+      "repoId": 1108837393,
+      "pullRequestNo": 681
    }
  ]
 }
--- a/src/agents/AGENTS.md
+++ b/src/agents/AGENTS.md
@@ -33,7 +33,7 @@ agents/
 |-------|---------------|----------|---------|
 | Sisyphus | anthropic/claude-opus-4-5 | - | Primary orchestrator with extended thinking |
 | oracle | openai/gpt-5.2 | - | Read-only consultation. High-IQ debugging, architecture |
-| librarian | anthropic/claude-sonnet-4-5 | google/gemini-3-flash | Docs, OSS research, GitHub examples |
+| librarian | opencode/glm-4.7-free | - | Docs, OSS research, GitHub examples |
 | explore | opencode/grok-code | google/gemini-3-flash, anthropic/claude-haiku-4-5 | Fast contextual grep |
 | frontend-ui-ux-engineer | google/gemini-3-pro-preview | - | UI/UX code generation |
 | document-writer | google/gemini-3-pro-preview | - | Technical writing |
--- a/src/agents/librarian.ts
+++ b/src/agents/librarian.ts
@@ -1,7 +1,7 @@
 import type { AgentConfig } from "@opencode-ai/sdk"
 import type { AgentPromptMetadata } from "./types"

-const DEFAULT_MODEL = "anthropic/claude-sonnet-4-5"
+const DEFAULT_MODEL = "opencode/glm-4.7-free"

 export const LIBRARIAN_PROMPT_METADATA: AgentPromptMetadata = {
  category: "exploration",
@@ -129,15 +129,15 @@ Tool 3: grep_app_searchGitHub(query: "usage pattern", language: ["TypeScript"])
 \`\`\`
 Step 1: Clone to temp directory
        gh repo clone owner/repo \${TMPDIR:-/tmp}/repo-name -- --depth 1
-        
+
 Step 2: Get commit SHA for permalinks
        cd \${TMPDIR:-/tmp}/repo-name && git rev-parse HEAD
-        
+
 Step 3: Find the implementation
        - grep/ast_grep_search for function/class
        - read the specific file
        - git blame for context if needed
-        
+
 Step 4: Construct permalink
        https://github.com/owner/repo/blob/<sha>/path/to/file#L10-L20
 \`\`\`
@@ -272,7 +272,7 @@ Use OS-appropriate temp directory:
 | TYPE B (Implementation) | 2-3 NO |
 | TYPE C (Context) | 2-3 NO |
 | TYPE D (Comprehensive) | 3-5 | YES (Phase 0.5 first) |
-| Request Type | Minimum Parallel Calls 
+| Request Type | Minimum Parallel Calls

 **Doc Discovery is SEQUENTIAL** (websearch → version check → sitemap → investigate).
 **Main phase is PARALLEL** once you know where to look.
@@ -308,7 +308,7 @@ grep_app_searchGitHub(query: "useQuery")
 ## COMMUNICATION RULES

 1. **NO TOOL NAMES**: Say "I'll search the codebase" not "I'll use grep_app"
-2. **NO PREAMBLE**: Answer directly, skip "I'll help you with..." 
+2. **NO PREAMBLE**: Answer directly, skip "I'll help you with..."
 3. **ALWAYS CITE**: Every code claim needs a permalink
 4. **USE MARKDOWN**: Code blocks with language identifiers
 5. **BE CONCISE**: Facts > opinions, evidence > speculation
--- a/src/agents/orchestrator-sisyphus.ts
+++ b/src/agents/orchestrator-sisyphus.ts
@@ -132,7 +132,6 @@ ${rows.join("\n")}
 }

 export const ORCHESTRATOR_SISYPHUS_SYSTEM_PROMPT = `You are "Sisyphus" - Powerful AI Agent with orchestration capabilities from OhMyOpenCode.
-Named by [YeonGyu Kim](https://github.com/code-yeongyu).

 **Why Sisyphus?**: Humans roll their boulder every day. So do you. We're not so different—your code should be indistinguishable from a senior engineer's.

--- a/src/agents/sisyphus.ts
+++ b/src/agents/sisyphus.ts
@@ -18,7 +18,6 @@ const DEFAULT_MODEL = "anthropic/claude-opus-4-5"

 const SISYPHUS_ROLE_SECTION = `<Role>
 You are "Sisyphus" - Powerful AI Agent with orchestration capabilities from OhMyOpenCode.
-Named by [YeonGyu Kim](https://github.com/code-yeongyu).

 **Why Sisyphus?**: Humans roll their boulder every day. So do you. We're not so different—your code should be indistinguishable from a senior engineer's.

--- a/src/config/schema.ts
+++ b/src/config/schema.ts
@@ -296,6 +296,7 @@ export const GitMasterConfigSchema = z.object({
  /** Add "Co-authored-by: Sisyphus" trailer to commit messages (default: true) */
  include_co_authored_by: z.boolean().default(true),
 })
+
 export const OhMyOpenCodeConfigSchema = z.object({
  $schema: z.string().optional(),
  disabled_mcps: z.array(AnyMcpNameSchema).optional(),
--- a/src/features/background-agent/manager.test.ts
+++ b/src/features/background-agent/manager.test.ts
@@ -674,3 +674,95 @@ describe("LaunchInput.skillContent", () => {
    expect(input.skillContent).toBe("You are a playwright expert")
  })
 })
+
+describe("BackgroundManager.notifyParentSession - agent context preservation", () => {
+  test("should not pass agent field when parentAgent is undefined", async () => {
+    // #given
+    const task: BackgroundTask = {
+      id: "task-no-agent",
+      sessionID: "session-child",
+      parentSessionID: "session-parent",
+      parentMessageID: "msg-parent",
+      description: "task without agent context",
+      prompt: "test",
+      agent: "explore",
+      status: "completed",
+      startedAt: new Date(),
+      completedAt: new Date(),
+      parentAgent: undefined,
+      parentModel: { providerID: "anthropic", modelID: "claude-opus" },
+    }
+
+    // #when
+    const promptBody = buildNotificationPromptBody(task)
+
+    // #then
+    expect("agent" in promptBody).toBe(false)
+    expect(promptBody.model).toEqual({ providerID: "anthropic", modelID: "claude-opus" })
+  })
+
+  test("should include agent field when parentAgent is defined", async () => {
+    // #given
+    const task: BackgroundTask = {
+      id: "task-with-agent",
+      sessionID: "session-child",
+      parentSessionID: "session-parent",
+      parentMessageID: "msg-parent",
+      description: "task with agent context",
+      prompt: "test",
+      agent: "explore",
+      status: "completed",
+      startedAt: new Date(),
+      completedAt: new Date(),
+      parentAgent: "Sisyphus",
+      parentModel: { providerID: "anthropic", modelID: "claude-opus" },
+    }
+
+    // #when
+    const promptBody = buildNotificationPromptBody(task)
+
+    // #then
+    expect(promptBody.agent).toBe("Sisyphus")
+  })
+
+  test("should not pass model field when parentModel is undefined", async () => {
+    // #given
+    const task: BackgroundTask = {
+      id: "task-no-model",
+      sessionID: "session-child",
+      parentSessionID: "session-parent",
+      parentMessageID: "msg-parent",
+      description: "task without model context",
+      prompt: "test",
+      agent: "explore",
+      status: "completed",
+      startedAt: new Date(),
+      completedAt: new Date(),
+      parentAgent: "Sisyphus",
+      parentModel: undefined,
+    }
+
+    // #when
+    const promptBody = buildNotificationPromptBody(task)
+
+    // #then
+    expect("model" in promptBody).toBe(false)
+    expect(promptBody.agent).toBe("Sisyphus")
+  })
+})
+
+function buildNotificationPromptBody(task: BackgroundTask): Record<string, unknown> {
+  const body: Record<string, unknown> = {
+    parts: [{ type: "text", text: `[BACKGROUND TASK COMPLETED] Task "${task.description}" finished.` }],
+  }
+
+  if (task.parentAgent !== undefined) {
+    body.agent = task.parentAgent
+  }
+
+  if (task.parentModel?.providerID && task.parentModel?.modelID) {
+    body.model = { providerID: task.parentModel.providerID, modelID: task.parentModel.modelID }
+  }
+
+  return body
+}
--- a/src/features/background-agent/manager.ts
+++ b/src/features/background-agent/manager.ts
@@ -13,6 +13,7 @@ import { subagentSessions } from "../claude-code-session-state"
 import { getTaskToastManager } from "../task-toast-manager"

 const TASK_TTL_MS = 30 * 60 * 1000
+const MIN_STABILITY_TIME_MS = 10 * 1000  // Must run at least 10s before stability detection kicks in

 type OpencodeClient = PluginInput["client"]

@@ -43,6 +44,7 @@ interface Todo {
 export class BackgroundManager {
  private tasks: Map<string, BackgroundTask>
  private notifications: Map<string, BackgroundTask[]>
+  private pendingByParent: Map<string, Set<string>>  // Track pending tasks per parent for batching
  private client: OpencodeClient
  private directory: string
  private pollingInterval?: ReturnType<typeof setInterval>
@@ -51,12 +53,20 @@ export class BackgroundManager {
  constructor(ctx: PluginInput, config?: BackgroundTaskConfig) {
    this.tasks = new Map()
    this.notifications = new Map()
+    this.pendingByParent = new Map()
    this.client = ctx.client
    this.directory = ctx.directory
    this.concurrencyManager = new ConcurrencyManager(config)
  }

  async launch(input: LaunchInput): Promise<BackgroundTask> {
+    log("[background-agent] launch() called with:", {
+      agent: input.agent,
+      model: input.model,
+      description: input.description,
+      parentSessionID: input.parentSessionID,
+    })
+
    if (!input.agent || input.agent.trim() === "") {
      throw new Error("Agent parameter is required")
    }
@@ -106,6 +116,11 @@ export class BackgroundManager {
    this.tasks.set(task.id, task)
    this.startPolling()

+    // Track for batched notifications
+    const pending = this.pendingByParent.get(input.parentSessionID) ?? new Set()
+    pending.add(task.id)
+    this.pendingByParent.set(input.parentSessionID, pending)
+
    log("[background-agent] Launching task:", { taskId: task.id, sessionID, agent: input.agent })

    const toastManager = getTaskToastManager()
@@ -119,10 +134,21 @@ export class BackgroundManager {
      })
    }

-    this.client.session.promptAsync({
+    log("[background-agent] Calling prompt (fire-and-forget) for launch with:", {
+      sessionID,
+      agent: input.agent,
+      model: input.model,
+      hasSkillContent: !!input.skillContent,
+      promptLength: input.prompt.length,
+    })
+
+    // Use prompt() instead of promptAsync() to properly initialize agent loop (fire-and-forget)
+    // Include model if caller provided one (e.g., from Sisyphus category configs)
+    this.client.session.prompt({
      path: { id: sessionID },
      body: {
        agent: input.agent,
+        ...(input.model ? { model: input.model } : {}),
        system: input.skillContent,
        tools: {
          task: false,
@@ -146,7 +172,9 @@ export class BackgroundManager {
          this.concurrencyManager.release(existingTask.concurrencyKey)
        }
        this.markForNotification(existingTask)
-        this.notifyParentSession(existingTask)
+        this.notifyParentSession(existingTask).catch(err => {
+          log("[background-agent] Failed to notify on error:", err)
+        })
      }
    })

@@ -199,6 +227,7 @@ export class BackgroundManager {
    parentSessionID: string
    description: string
    agent?: string
+    parentAgent?: string
  }): BackgroundTask {
    const task: BackgroundTask = {
      id: input.taskId,
@@ -214,12 +243,18 @@ export class BackgroundManager {
        toolCalls: 0,
        lastUpdate: new Date(),
      },
+      parentAgent: input.parentAgent,
    }

    this.tasks.set(task.id, task)
    subagentSessions.add(input.sessionID)
    this.startPolling()

+    // Track for batched notifications (external tasks need tracking too)
+    const pending = this.pendingByParent.get(input.parentSessionID) ?? new Set()
+    pending.add(task.id)
+    this.pendingByParent.set(input.parentSessionID, pending)
+
    log("[background-agent] Registered external task:", { taskId: task.id, sessionID: input.sessionID })

    return task
@@ -247,6 +282,11 @@ export class BackgroundManager {
    this.startPolling()
    subagentSessions.add(existingTask.sessionID)

+    // Track for batched notifications (P2 fix: resumed tasks need tracking too)
+    const pending = this.pendingByParent.get(input.parentSessionID) ?? new Set()
+    pending.add(existingTask.id)
+    this.pendingByParent.set(input.parentSessionID, pending)
+
    const toastManager = getTaskToastManager()
    if (toastManager) {
      toastManager.addTask({
@@ -259,7 +299,15 @@ export class BackgroundManager {

    log("[background-agent] Resuming task:", { taskId: existingTask.id, sessionID: existingTask.sessionID })

-    this.client.session.promptAsync({
+    log("[background-agent] Resuming task - calling prompt (fire-and-forget) with:", {
+      sessionID: existingTask.sessionID,
+      agent: existingTask.agent,
+      promptLength: input.prompt.length,
+    })
+
+    // Note: Don't pass model in body - use agent's configured model instead
+    // Use prompt() instead of promptAsync() to properly initialize agent loop
+    this.client.session.prompt({
      path: { id: existingTask.sessionID },
      body: {
        agent: existingTask.agent,
@@ -270,13 +318,15 @@ export class BackgroundManager {
        parts: [{ type: "text", text: input.prompt }],
      },
    }).catch((error) => {
-      log("[background-agent] resume promptAsync error:", error)
+      log("[background-agent] resume prompt error:", error)
      existingTask.status = "error"
      const errorMessage = error instanceof Error ? error.message : String(error)
      existingTask.error = errorMessage
      existingTask.completedAt = new Date()
      this.markForNotification(existingTask)
-      this.notifyParentSession(existingTask)
+      this.notifyParentSession(existingTask).catch(err => {
+        log("[background-agent] Failed to notify on resume error:", err)
+      })
    })

    return existingTask
@@ -331,7 +381,22 @@ export class BackgroundManager {
      const task = this.findBySession(sessionID)
      if (!task || task.status !== "running") return

-      this.checkSessionTodos(sessionID).then((hasIncompleteTodos) => {
+      // Edge guard: Require minimum elapsed time (5 seconds) before accepting idle
+      const elapsedMs = Date.now() - task.startedAt.getTime()
+      const MIN_IDLE_TIME_MS = 5000
+      if (elapsedMs < MIN_IDLE_TIME_MS) {
+        log("[background-agent] Ignoring early session.idle, elapsed:", { elapsedMs, taskId: task.id })
+        return
+      }
+
+      // Edge guard: Verify session has actual assistant output before completing
+      this.validateSessionHasOutput(sessionID).then(async (hasValidOutput) => {
+        if (!hasValidOutput) {
+          log("[background-agent] Session.idle but no valid output yet, waiting:", task.id)
+          return
+        }
+
+        const hasIncompleteTodos = await this.checkSessionTodos(sessionID)
        if (hasIncompleteTodos) {
          log("[background-agent] Task has incomplete todos, waiting for todo-continuation:", task.id)
          return
@@ -340,8 +405,10 @@ export class BackgroundManager {
        task.status = "completed"
        task.completedAt = new Date()
        this.markForNotification(task)
-        this.notifyParentSession(task)
+        await this.notifyParentSession(task)
        log("[background-agent] Task completed via session.idle event:", task.id)
+      }).catch(err => {
+        log("[background-agent] Error in session.idle handler:", err)
      })
    }

@@ -382,6 +449,66 @@ export class BackgroundManager {
    this.notifications.delete(sessionID)
  }

+  /**
+   * Validates that a session has actual assistant/tool output before marking complete.
+   * Prevents premature completion when session.idle fires before agent responds.
+   */
+  private async validateSessionHasOutput(sessionID: string): Promise<boolean> {
+    try {
+      const response = await this.client.session.messages({
+        path: { id: sessionID },
+      })
+
+      const messages = response.data ?? []
+      
+      // Check for at least one assistant or tool message
+      const hasAssistantOrToolMessage = messages.some(
+        (m: { info?: { role?: string } }) => 
+          m.info?.role === "assistant" || m.info?.role === "tool"
+      )
+
+      if (!hasAssistantOrToolMessage) {
+        log("[background-agent] No assistant/tool messages found in session:", sessionID)
+        return false
+      }
+
+      // Additionally check that at least one message has content (not just empty)
+      // OpenCode API uses different part types than Anthropic's API:
+      // - "reasoning" with .text property (thinking/reasoning content)
+      // - "tool" with .state.output property (tool call results)
+      // - "text" with .text property (final text output)
+      // - "step-start"/"step-finish" (metadata, no content)
+      // eslint-disable-next-line @typescript-eslint/no-explicit-any
+      const hasContent = messages.some((m: any) => {
+        if (m.info?.role !== "assistant" && m.info?.role !== "tool") return false
+        const parts = m.parts ?? []
+      // eslint-disable-next-line @typescript-eslint/no-explicit-any
+      return parts.some((p: any) => 
+        // Text content (final output)
+        (p.type === "text" && p.text && p.text.trim().length > 0) ||
+        // Reasoning content (thinking blocks)
+        (p.type === "reasoning" && p.text && p.text.trim().length > 0) ||
+        // Tool calls (indicates work was done)
+        p.type === "tool" ||
+        // Tool results (output from executed tools) - important for tool-only tasks
+        (p.type === "tool_result" && p.content && 
+          (typeof p.content === "string" ? p.content.trim().length > 0 : p.content.length > 0))
+      )
+      })
+
+      if (!hasContent) {
+        log("[background-agent] Messages exist but no content found in session:", sessionID)
+        return false
+      }
+
+      return true
+    } catch (error) {
+      log("[background-agent] Error validating session output:", error)
+      // On error, allow completion to proceed (don't block indefinitely)
+      return true
+    }
+  }
+
  private clearNotificationsForTask(taskId: string): void {
    for (const [sessionID, tasks] of this.notifications.entries()) {
      const filtered = tasks.filter((t) => t.id !== taskId)
@@ -409,17 +536,33 @@ export class BackgroundManager {
    }
  }

-  cleanup(): void {
+cleanup(): void {
    this.stopPolling()
    this.tasks.clear()
    this.notifications.clear()
+    this.pendingByParent.clear()
  }

-  private notifyParentSession(task: BackgroundTask): void {
+  /**
+   * Get all running tasks (for compaction hook)
+   */
+  getRunningTasks(): BackgroundTask[] {
+    return Array.from(this.tasks.values()).filter(t => t.status === "running")
+  }
+
+  /**
+   * Get all completed tasks still in memory (for compaction hook)
+   */
+  getCompletedTasks(): BackgroundTask[] {
+    return Array.from(this.tasks.values()).filter(t => t.status !== "running")
+  }
+
+private async notifyParentSession(task: BackgroundTask): Promise<void> {
    const duration = this.formatDuration(task.startedAt, task.completedAt)

    log("[background-agent] notifyParentSession called for task:", task.id)

+    // Show toast notification
    const toastManager = getTaskToastManager()
    if (toastManager) {
      toastManager.showCompletionToast({
@@ -429,41 +572,83 @@ export class BackgroundManager {
      })
    }

-    const message = `[BACKGROUND TASK COMPLETED] Task "${task.description}" finished in ${duration}. Use background_output with task_id="${task.id}" to get results.`
+    // Update pending tracking and check if all tasks complete
+    const pendingSet = this.pendingByParent.get(task.parentSessionID)
+    if (pendingSet) {
+      pendingSet.delete(task.id)
+      if (pendingSet.size === 0) {
+        this.pendingByParent.delete(task.parentSessionID)
+      }
+    }

-    log("[background-agent] Sending notification to parent session:", { parentSessionID: task.parentSessionID })
+    const allComplete = !pendingSet || pendingSet.size === 0
+    const remainingCount = pendingSet?.size ?? 0

+    // Build notification message
+    const statusText = task.status === "error" ? "FAILED" : "COMPLETED"
+    const errorInfo = task.error ? `\n**Error:** ${task.error}` : ""
+    
+    let notification: string
+    if (allComplete) {
+      // All tasks complete - build summary
+      const completedTasks = Array.from(this.tasks.values())
+        .filter(t => t.parentSessionID === task.parentSessionID && t.status !== "running")
+        .map(t => `- \`${t.id}\`: ${t.description}`)
+        .join("\n")
+
+      notification = `<system-reminder>
+[ALL BACKGROUND TASKS COMPLETE]
+
+**Completed:**
+${completedTasks || `- \`${task.id}\`: ${task.description}`}
+
+Use \`background_output(task_id="<id>")\` to retrieve each result.
+</system-reminder>`
+    } else {
+      // Individual completion - silent notification
+      notification = `<system-reminder>
+[BACKGROUND TASK ${statusText}]
+**ID:** \`${task.id}\`
+**Description:** ${task.description}
+**Duration:** ${duration}${errorInfo}
+
+**${remainingCount} task${remainingCount === 1 ? "" : "s"} still in progress.** You WILL be notified when ALL complete.
+Do NOT poll - continue productive work.
+
+Use \`background_output(task_id="${task.id}")\` to retrieve this result when ready.
+</system-reminder>`
+    }
+
+    // Inject notification via session.prompt with noReply
+    try {
+      await this.client.session.prompt({
+        path: { id: task.parentSessionID },
+        body: {
+          noReply: !allComplete,  // Silent unless all complete
+          agent: task.parentAgent,
+          parts: [{ type: "text", text: notification }],
+        },
+      })
+      log("[background-agent] Sent notification to parent session:", {
+        taskId: task.id,
+        allComplete,
+        noReply: !allComplete,
+      })
+    } catch (error) {
+      log("[background-agent] Failed to send notification:", error)
+    }
+
+    // Cleanup after retention period
    const taskId = task.id
-    setTimeout(async () => {
+    setTimeout(() => {
      if (task.concurrencyKey) {
        this.concurrencyManager.release(task.concurrencyKey)
+        task.concurrencyKey = undefined
      }
-
-      try {
-        // Use only parentModel/parentAgent - don't fallback to prevMessage
-        // This prevents accidentally changing parent session's model/agent
-        const modelField = task.parentModel?.providerID && task.parentModel?.modelID
-          ? { providerID: task.parentModel.providerID, modelID: task.parentModel.modelID }
-          : undefined
-
-        await this.client.session.prompt({
-          path: { id: task.parentSessionID },
-          body: {
-            agent: task.parentAgent,
-            model: modelField,
-            parts: [{ type: "text", text: message }],
-          },
-          query: { directory: this.directory },
-        })
-        log("[background-agent] Successfully sent prompt to parent session:", { parentSessionID: task.parentSessionID })
-      } catch (error) {
-        log("[background-agent] prompt failed:", String(error))
-      } finally {
-        this.clearNotificationsForTask(taskId)
-        this.tasks.delete(taskId)
-        log("[background-agent] Removed completed task from memory:", taskId)
-      }
-    }, 200)
+      this.clearNotificationsForTask(taskId)
+      this.tasks.delete(taskId)
+      log("[background-agent] Removed completed task from memory:", taskId)
+    }, 5 * 60 * 1000)
  }

  private formatDuration(start: Date, end?: Date): string {
@@ -532,15 +717,18 @@ export class BackgroundManager {
    for (const task of this.tasks.values()) {
      if (task.status !== "running") continue

-      try {
+try {
        const sessionStatus = allStatuses[task.sessionID]
        
-        if (!sessionStatus) {
-          log("[background-agent] Session not found in status:", task.sessionID)
-          continue
-        }
+        // Don't skip if session not in status - fall through to message-based detection
+        if (sessionStatus?.type === "idle") {
+          // Edge guard: Validate session has actual output before completing
+          const hasValidOutput = await this.validateSessionHasOutput(task.sessionID)
+          if (!hasValidOutput) {
+            log("[background-agent] Polling idle but no valid output yet, waiting:", task.id)
+            continue
+          }

-        if (sessionStatus.type === "idle") {
          const hasIncompleteTodos = await this.checkSessionTodos(task.sessionID)
          if (hasIncompleteTodos) {
            log("[background-agent] Task has incomplete todos via polling, waiting:", task.id)
@@ -550,7 +738,7 @@ export class BackgroundManager {
          task.status = "completed"
          task.completedAt = new Date()
          this.markForNotification(task)
-          this.notifyParentSession(task)
+          await this.notifyParentSession(task)
          log("[background-agent] Task completed via polling:", task.id)
          continue
        }
@@ -591,10 +779,41 @@ export class BackgroundManager {
          task.progress.toolCalls = toolCalls
          task.progress.lastTool = lastTool
          task.progress.lastUpdate = new Date()
-          if (lastMessage) {
+if (lastMessage) {
            task.progress.lastMessage = lastMessage
            task.progress.lastMessageAt = new Date()
          }
+
+          // Stability detection: complete when message count unchanged for 3 polls
+          const currentMsgCount = messages.length
+          const elapsedMs = Date.now() - task.startedAt.getTime()
+
+          if (elapsedMs >= MIN_STABILITY_TIME_MS) {
+            if (task.lastMsgCount === currentMsgCount) {
+              task.stablePolls = (task.stablePolls ?? 0) + 1
+              if (task.stablePolls >= 3) {
+                // Edge guard: Validate session has actual output before completing
+                const hasValidOutput = await this.validateSessionHasOutput(task.sessionID)
+                if (!hasValidOutput) {
+                  log("[background-agent] Stability reached but no valid output, waiting:", task.id)
+                  continue
+                }
+
+                const hasIncompleteTodos = await this.checkSessionTodos(task.sessionID)
+                if (!hasIncompleteTodos) {
+                  task.status = "completed"
+                  task.completedAt = new Date()
+                  this.markForNotification(task)
+                  await this.notifyParentSession(task)
+                  log("[background-agent] Task completed via stability detection:", task.id)
+                  continue
+                }
+              }
+            } else {
+              task.stablePolls = 0
+            }
+          }
+          task.lastMsgCount = currentMsgCount
        }
      } catch (error) {
        log("[background-agent] Poll error for task:", { taskId: task.id, error })
--- a/src/features/background-agent/types.ts
+++ b/src/features/background-agent/types.ts
@@ -32,6 +32,10 @@ export interface BackgroundTask {
  concurrencyKey?: string
  /** Parent session's agent name for notification */
  parentAgent?: string
+  /** Last message count for stability detection */
+  lastMsgCount?: number
+  /** Number of consecutive polls with stable message count */
+  stablePolls?: number
 }

 export interface LaunchInput {
--- a/src/features/context-injector/injector.test.ts
+++ b/src/features/context-injector/injector.test.ts
@@ -133,7 +133,7 @@ describe("createContextInjectorHook", () => {
  })

  describe("chat.message handler", () => {
-    it("is a no-op (context injection moved to messages transform)", async () => {
+    it("injects pending context into output parts", async () => {
      // #given
      const hook = createContextInjectorHook(collector)
      const sessionID = "ses_hook1"
@@ -152,8 +152,9 @@ describe("createContextInjectorHook", () => {
      await hook["chat.message"](input, output)

      // #then
-      expect(output.parts[0].text).toBe("User message")
-      expect(collector.hasPending(sessionID)).toBe(true)
+      expect(output.parts[0].text).toContain("Hook context")
+      expect(output.parts[0].text).toContain("User message")
+      expect(collector.hasPending(sessionID)).toBe(false)
    })

    it("does nothing when no pending context", async () => {
--- a/src/features/context-injector/injector.ts
+++ b/src/features/context-injector/injector.ts
@@ -52,10 +52,16 @@ interface ChatMessageOutput {
 export function createContextInjectorHook(collector: ContextCollector) {
  return {
    "chat.message": async (
-      _input: ChatMessageInput,
-      _output: ChatMessageOutput
+      input: ChatMessageInput,
+      output: ChatMessageOutput
    ): Promise<void> => {
-      void collector
+      const result = injectPendingContext(collector, input.sessionID, output.parts)
+      if (result.injected) {
+        log("[context-injector] Injected pending context via chat.message", {
+          sessionID: input.sessionID,
+          contextLength: result.contextLength,
+        })
+      }
    },
  }
 }
--- a/src/hooks/background-compaction/index.ts
+++ b/src/hooks/background-compaction/index.ts
@@ -0,0 +1,85 @@
+import type { BackgroundManager } from "../../features/background-agent"
+
+interface CompactingInput {
+  sessionID: string
+}
+
+interface CompactingOutput {
+  context: string[]
+  prompt?: string
+}
+
+/**
+ * Background agent compaction hook - preserves task state during context compaction.
+ * 
+ * When OpenCode compacts session context to save tokens, this hook injects
+ * information about running and recently completed background tasks so the
+ * agent doesn't lose awareness of delegated work.
+ */
+export function createBackgroundCompactionHook(manager: BackgroundManager) {
+  return {
+    "experimental.session.compacting": async (
+      input: CompactingInput,
+      output: CompactingOutput
+    ): Promise<void> => {
+      const { sessionID } = input
+
+      // Get running tasks for this session
+      const running = manager.getRunningTasks()
+        .filter(t => t.parentSessionID === sessionID)
+        .map(t => ({
+          id: t.id,
+          agent: t.agent,
+          description: t.description,
+          startedAt: t.startedAt,
+        }))
+
+      // Get recently completed tasks (still in memory within 5-min retention)
+      const completed = manager.getCompletedTasks()
+        .filter(t => t.parentSessionID === sessionID)
+        .slice(-10) // Last 10 completed
+        .map(t => ({
+          id: t.id,
+          agent: t.agent,
+          description: t.description,
+          status: t.status,
+        }))
+
+      // Early exit if nothing to preserve
+      if (running.length === 0 && completed.length === 0) return
+
+      const sections: string[] = ["<background-tasks>"]
+
+      // Running tasks section
+      if (running.length > 0) {
+        sections.push("## Running Background Tasks")
+        sections.push("")
+        for (const t of running) {
+          const elapsed = Math.floor((Date.now() - t.startedAt.getTime()) / 1000)
+          sections.push(`- **\`${t.id}\`** (${t.agent}): ${t.description} [${elapsed}s elapsed]`)
+        }
+        sections.push("")
+        sections.push("> **Note:** You WILL be notified when tasks complete.")
+        sections.push("> Do NOT poll - continue productive work.")
+        sections.push("")
+      }
+
+      // Completed tasks section
+      if (completed.length > 0) {
+        sections.push("## Recently Completed Tasks")
+        sections.push("")
+        for (const t of completed) {
+          const statusEmoji = t.status === "completed" ? "✅" : t.status === "error" ? "❌" : "⏱️"
+          sections.push(`- ${statusEmoji} **\`${t.id}\`**: ${t.description}`)
+        }
+        sections.push("")
+      }
+
+      sections.push("## Retrieval")
+      sections.push('Use `background_output(task_id="<id>")` to retrieve task results.')
+      sections.push("</background-tasks>")
+
+      output.context.push(sections.join("\n"))
+    }
+  }
+}
--- a/src/hooks/background-notification/index.ts
+++ b/src/hooks/background-notification/index.ts
@@ -9,6 +9,12 @@ interface EventInput {
  event: Event
 }

+/**
+ * Background notification hook - handles event routing to BackgroundManager.
+ * 
+ * Notifications are now delivered directly via session.prompt({ noReply }) 
+ * from the manager, so this hook only needs to handle event routing.
+ */
 export function createBackgroundNotificationHook(manager: BackgroundManager) {
  const eventHandler = async ({ event }: EventInput) => {
    manager.handleEvent(event)
--- a/src/hooks/claude-code-hooks/index.ts
+++ b/src/hooks/claude-code-hooks/index.ts
@@ -27,7 +27,6 @@ import { cacheToolInput, getToolInput } from "./tool-input-cache"
 import { recordToolUse, recordToolResult, getTranscriptPath, recordUserMessage } from "./transcript"
 import type { PluginConfig } from "./types"
 import { log, isHookDisabled } from "../../shared"
-import { detectKeywordsWithType, removeCodeBlocks } from "../keyword-detector"
 import type { ContextCollector } from "../../features/context-injector"

 const sessionFirstMessageProcessed = new Set<string>()
@@ -142,25 +141,9 @@ export function createClaudeCodeHooksHook(
          return
        }

-        const keywordMessages: string[] = []
-        if (!config.keywordDetectorDisabled) {
-          const detectedKeywords = detectKeywordsWithType(removeCodeBlocks(prompt), input.agent)
-          keywordMessages.push(...detectedKeywords.map((k) => k.message))
-
-          if (keywordMessages.length > 0) {
-            log("[claude-code-hooks] Detected keywords", {
-              sessionID: input.sessionID,
-              keywordCount: keywordMessages.length,
-              types: detectedKeywords.map((k) => k.type),
-            })
-          }
-        }
-
-        const allMessages = [...keywordMessages, ...result.messages]
-
-        if (allMessages.length > 0) {
-          const hookContent = allMessages.join("\n\n")
-          log(`[claude-code-hooks] Injecting ${allMessages.length} messages (${keywordMessages.length} keyword + ${result.messages.length} hook)`, { sessionID: input.sessionID, contentLength: hookContent.length, isFirstMessage })
+        if (result.messages.length > 0) {
+          const hookContent = result.messages.join("\n\n")
+          log(`[claude-code-hooks] Injecting ${result.messages.length} hook messages`, { sessionID: input.sessionID, contentLength: hookContent.length, isFirstMessage })

          if (isFirstMessage) {
            const idx = output.parts.findIndex((p) => p.type === "text" && p.text)
--- a/src/hooks/index.ts
+++ b/src/hooks/index.ts
@@ -14,6 +14,7 @@ export { createThinkModeHook } from "./think-mode";
 export { createClaudeCodeHooksHook } from "./claude-code-hooks";
 export { createRulesInjectorHook } from "./rules-injector";
 export { createBackgroundNotificationHook } from "./background-notification"
+export { createBackgroundCompactionHook } from "./background-compaction"
 export { createAutoUpdateCheckerHook } from "./auto-update-checker";

 export { createAgentUsageReminderHook } from "./agent-usage-reminder";
--- a/src/hooks/keyword-detector/index.test.ts
+++ b/src/hooks/keyword-detector/index.test.ts
@@ -1,7 +1,95 @@
 import { describe, expect, test, beforeEach, afterEach, spyOn } from "bun:test"
 import { createKeywordDetectorHook } from "./index"
 import { setMainSession } from "../../features/claude-code-session-state"
+import { ContextCollector } from "../../features/context-injector"
 import * as sharedModule from "../../shared"
+import * as sessionState from "../../features/claude-code-session-state"
+
+describe("keyword-detector registers to ContextCollector", () => {
+  let logCalls: Array<{ msg: string; data?: unknown }>
+  let logSpy: ReturnType<typeof spyOn>
+  let getMainSessionSpy: ReturnType<typeof spyOn>
+
+  beforeEach(() => {
+    logCalls = []
+    logSpy = spyOn(sharedModule, "log").mockImplementation((msg: string, data?: unknown) => {
+      logCalls.push({ msg, data })
+    })
+  })
+
+  afterEach(() => {
+    logSpy?.mockRestore()
+    getMainSessionSpy?.mockRestore()
+  })
+
+  function createMockPluginInput() {
+    return {
+      client: {
+        tui: {
+          showToast: async () => {},
+        },
+      },
+    } as any
+  }
+
+  test("should register ultrawork keyword to ContextCollector", async () => {
+    // #given - a fresh ContextCollector and keyword-detector hook
+    const collector = new ContextCollector()
+    const hook = createKeywordDetectorHook(createMockPluginInput(), collector)
+    const sessionID = "test-session-123"
+    const output = {
+      message: {} as Record<string, unknown>,
+      parts: [{ type: "text", text: "ultrawork do something" }],
+    }
+
+    // #when - keyword detection runs
+    await hook["chat.message"]({ sessionID }, output)
+
+    // #then - ultrawork context should be registered in collector
+    expect(collector.hasPending(sessionID)).toBe(true)
+    const pending = collector.getPending(sessionID)
+    expect(pending.entries.length).toBeGreaterThan(0)
+    expect(pending.entries[0].source).toBe("keyword-detector")
+    expect(pending.entries[0].id).toBe("keyword-ultrawork")
+  })
+
+  test("should register search keyword to ContextCollector", async () => {
+    // #given - mock getMainSessionID to return our session (isolate from global state)
+    const collector = new ContextCollector()
+    const sessionID = "search-test-session"
+    getMainSessionSpy = spyOn(sessionState, "getMainSessionID").mockReturnValue(sessionID)
+    const hook = createKeywordDetectorHook(createMockPluginInput(), collector)
+    const output = {
+      message: {} as Record<string, unknown>,
+      parts: [{ type: "text", text: "search for the bug" }],
+    }
+
+    // #when - keyword detection runs
+    await hook["chat.message"]({ sessionID }, output)
+
+    // #then - search context should be registered in collector
+    expect(collector.hasPending(sessionID)).toBe(true)
+    const pending = collector.getPending(sessionID)
+    expect(pending.entries.some((e) => e.id === "keyword-search")).toBe(true)
+  })
+
+  test("should NOT register to collector when no keywords detected", async () => {
+    // #given - no keywords in message
+    const collector = new ContextCollector()
+    const hook = createKeywordDetectorHook(createMockPluginInput(), collector)
+    const sessionID = "test-session"
+    const output = {
+      message: {} as Record<string, unknown>,
+      parts: [{ type: "text", text: "just a normal message" }],
+    }
+
+    // #when - keyword detection runs
+    await hook["chat.message"]({ sessionID }, output)
+
+    // #then - nothing should be registered
+    expect(collector.hasPending(sessionID)).toBe(false)
+  })
+})

 describe("keyword-detector session filtering", () => {
  let logCalls: Array<{ msg: string; data?: unknown }>
--- a/src/hooks/keyword-detector/index.ts
+++ b/src/hooks/keyword-detector/index.ts
@@ -2,12 +2,13 @@ import type { PluginInput } from "@opencode-ai/plugin"
 import { detectKeywordsWithType, extractPromptText, removeCodeBlocks } from "./detector"
 import { log } from "../../shared"
 import { getMainSessionID } from "../../features/claude-code-session-state"
+import type { ContextCollector } from "../../features/context-injector"

 export * from "./detector"
 export * from "./constants"
 export * from "./types"

-export function createKeywordDetectorHook(ctx: PluginInput) {
+export function createKeywordDetectorHook(ctx: PluginInput, collector?: ContextCollector) {
  return {
    "chat.message": async (
      input: {
@@ -28,8 +29,6 @@ export function createKeywordDetectorHook(ctx: PluginInput) {
        return
      }

-      // Only ultrawork keywords work in non-main sessions
-      // Other keywords (search, analyze, etc.) only work in main sessions
      const mainSessionID = getMainSessionID()
      const isNonMainSession = mainSessionID && input.sessionID !== mainSessionID

@@ -64,6 +63,17 @@ export function createKeywordDetectorHook(ctx: PluginInput) {
          )
      }

+      if (collector) {
+        for (const keyword of detectedKeywords) {
+          collector.register(input.sessionID, {
+            id: `keyword-${keyword.type}`,
+            source: "keyword-detector",
+            content: keyword.message,
+            priority: keyword.type === "ultrawork" ? "critical" : "high",
+          })
+        }
+      }
+
      log(`[keyword-detector] Detected ${detectedKeywords.length} keywords`, {
        sessionID: input.sessionID,
        types: detectedKeywords.map((k) => k.type),
--- a/src/hooks/prometheus-md-only/constants.ts
+++ b/src/hooks/prometheus-md-only/constants.ts
@@ -4,7 +4,7 @@ export const PROMETHEUS_AGENTS = ["Prometheus (Planner)"]

 export const ALLOWED_EXTENSIONS = [".md"]

-export const ALLOWED_PATH_PREFIX = ".sisyphus/"
+export const ALLOWED_PATH_PREFIX = ".sisyphus"

 export const BLOCKED_TOOLS = ["Write", "Edit", "write", "edit"]

--- a/src/hooks/prometheus-md-only/index.test.ts
+++ b/src/hooks/prometheus-md-only/index.test.ts
@@ -70,7 +70,7 @@ describe("prometheus-md-only", () => {
        callID: "call-1",
      }
      const output = {
-        args: { filePath: "/project/.sisyphus/plans/work-plan.md" },
+        args: { filePath: "/tmp/test/.sisyphus/plans/work-plan.md" },
      }

      // #when / #then
@@ -295,4 +295,136 @@ describe("prometheus-md-only", () => {
      ).resolves.toBeUndefined()
    })
  })
+
+  describe("cross-platform path validation", () => {
+    beforeEach(() => {
+      setupMessageStorage(TEST_SESSION_ID, "Prometheus (Planner)")
+    })
+
+    test("should allow Windows-style backslash paths under .sisyphus/", async () => {
+      // #given
+      const hook = createPrometheusMdOnlyHook(createMockPluginInput())
+      const input = {
+        tool: "Write",
+        sessionID: TEST_SESSION_ID,
+        callID: "call-1",
+      }
+      const output = {
+        args: { filePath: ".sisyphus\\plans\\work-plan.md" },
+      }
+
+      // #when / #then
+      await expect(
+        hook["tool.execute.before"](input, output)
+      ).resolves.toBeUndefined()
+    })
+
+    test("should allow mixed separator paths under .sisyphus/", async () => {
+      // #given
+      const hook = createPrometheusMdOnlyHook(createMockPluginInput())
+      const input = {
+        tool: "Write",
+        sessionID: TEST_SESSION_ID,
+        callID: "call-1",
+      }
+      const output = {
+        args: { filePath: ".sisyphus\\plans/work-plan.MD" },
+      }
+
+      // #when / #then
+      await expect(
+        hook["tool.execute.before"](input, output)
+      ).resolves.toBeUndefined()
+    })
+
+    test("should allow uppercase .MD extension", async () => {
+      // #given
+      const hook = createPrometheusMdOnlyHook(createMockPluginInput())
+      const input = {
+        tool: "Write",
+        sessionID: TEST_SESSION_ID,
+        callID: "call-1",
+      }
+      const output = {
+        args: { filePath: ".sisyphus/plans/work-plan.MD" },
+      }
+
+      // #when / #then
+      await expect(
+        hook["tool.execute.before"](input, output)
+      ).resolves.toBeUndefined()
+    })
+
+    test("should block paths outside workspace root even if containing .sisyphus", async () => {
+      // #given
+      const hook = createPrometheusMdOnlyHook(createMockPluginInput())
+      const input = {
+        tool: "Write",
+        sessionID: TEST_SESSION_ID,
+        callID: "call-1",
+      }
+      const output = {
+        args: { filePath: "/other/project/.sisyphus/plans/x.md" },
+      }
+
+      // #when / #then
+      await expect(
+        hook["tool.execute.before"](input, output)
+      ).rejects.toThrow("can only write/edit .md files inside .sisyphus/")
+    })
+
+    test("should block nested .sisyphus directories", async () => {
+      // #given
+      const hook = createPrometheusMdOnlyHook(createMockPluginInput())
+      const input = {
+        tool: "Write",
+        sessionID: TEST_SESSION_ID,
+        callID: "call-1",
+      }
+      const output = {
+        args: { filePath: "src/.sisyphus/plans/x.md" },
+      }
+
+      // #when / #then
+      await expect(
+        hook["tool.execute.before"](input, output)
+      ).rejects.toThrow("can only write/edit .md files inside .sisyphus/")
+    })
+
+    test("should block path traversal attempts", async () => {
+      // #given
+      const hook = createPrometheusMdOnlyHook(createMockPluginInput())
+      const input = {
+        tool: "Write",
+        sessionID: TEST_SESSION_ID,
+        callID: "call-1",
+      }
+      const output = {
+        args: { filePath: ".sisyphus/../secrets.md" },
+      }
+
+      // #when / #then
+      await expect(
+        hook["tool.execute.before"](input, output)
+      ).rejects.toThrow("can only write/edit .md files inside .sisyphus/")
+    })
+
+    test("should allow case-insensitive .SISYPHUS directory", async () => {
+      // #given
+      const hook = createPrometheusMdOnlyHook(createMockPluginInput())
+      const input = {
+        tool: "Write",
+        sessionID: TEST_SESSION_ID,
+        callID: "call-1",
+      }
+      const output = {
+        args: { filePath: ".SISYPHUS/plans/work-plan.md" },
+      }
+
+      // #when / #then
+      await expect(
+        hook["tool.execute.before"](input, output)
+      ).resolves.toBeUndefined()
+    })
+  })
 })
--- a/src/hooks/prometheus-md-only/index.ts
+++ b/src/hooks/prometheus-md-only/index.ts
@@ -1,16 +1,48 @@
 import type { PluginInput } from "@opencode-ai/plugin"
 import { existsSync, readdirSync } from "node:fs"
-import { join } from "node:path"
+import { join, resolve, relative, isAbsolute } from "node:path"
 import { HOOK_NAME, PROMETHEUS_AGENTS, ALLOWED_EXTENSIONS, ALLOWED_PATH_PREFIX, BLOCKED_TOOLS, PLANNING_CONSULT_WARNING } from "./constants"
 import { findNearestMessageWithFields, MESSAGE_STORAGE } from "../../features/hook-message-injector"
 import { log } from "../../shared/logger"

 export * from "./constants"

-function isAllowedFile(filePath: string): boolean {
-  const hasAllowedExtension = ALLOWED_EXTENSIONS.some(ext => filePath.endsWith(ext))
-  const isInAllowedPath = filePath.includes(ALLOWED_PATH_PREFIX)
-  return hasAllowedExtension && isInAllowedPath
+/**
+ * Cross-platform path validator for Prometheus file writes.
+ * Uses path.resolve/relative instead of string matching to handle:
+ * - Windows backslashes (e.g., .sisyphus\\plans\\x.md)
+ * - Mixed separators (e.g., .sisyphus\\plans/x.md)
+ * - Case-insensitive directory/extension matching
+ * - Workspace confinement (blocks paths outside root or via traversal)
+ */
+function isAllowedFile(filePath: string, workspaceRoot: string): boolean {
+  // 1. Resolve to absolute path
+  const resolved = resolve(workspaceRoot, filePath)
+
+  // 2. Get relative path from workspace root
+  const rel = relative(workspaceRoot, resolved)
+
+  // 3. Reject if escapes root (starts with ".." or is absolute)
+  if (rel.startsWith("..") || isAbsolute(rel)) {
+    return false
+  }
+
+  // 4. Split by both separators and check first segment matches ALLOWED_PATH_PREFIX (case-insensitive)
+  // Guard: if rel is empty (filePath === workspaceRoot), segments[0] would be "" — reject
+  const segments = rel.split(/[/\\]/)
+  if (!segments[0] || segments[0].toLowerCase() !== ALLOWED_PATH_PREFIX.toLowerCase()) {
+    return false
+  }
+
+  // 5. Check extension matches one of ALLOWED_EXTENSIONS (case-insensitive)
+  const hasAllowedExtension = ALLOWED_EXTENSIONS.some(
+    ext => resolved.toLowerCase().endsWith(ext.toLowerCase())
+  )
+  if (!hasAllowedExtension) {
+    return false
+  }
+
+  return true
 }

 function getMessageDir(sessionID: string): string | null {
@@ -35,7 +67,7 @@ function getAgentFromSession(sessionID: string): string | undefined {
  return findNearestMessageWithFields(messageDir)?.agent
 }

-export function createPrometheusMdOnlyHook(_ctx: PluginInput) {
+export function createPrometheusMdOnlyHook(ctx: PluginInput) {
  return {
    "tool.execute.before": async (
      input: { tool: string; sessionID: string; callID: string },
@@ -72,7 +104,7 @@ export function createPrometheusMdOnlyHook(_ctx: PluginInput) {
        return
      }

-      if (!isAllowedFile(filePath)) {
+      if (!isAllowedFile(filePath, ctx.directory)) {
        log(`[${HOOK_NAME}] Blocked: Prometheus can only write to .sisyphus/*.md`, {
          sessionID: input.sessionID,
          tool: toolName,
--- a/src/hooks/ralph-loop/index.test.ts
+++ b/src/hooks/ralph-loop/index.test.ts
@@ -591,6 +591,73 @@ describe("ralph-loop", () => {
      expect(hook.getState()).toBeNull()
    })

+    test("should allow starting new loop while previous loop is active (different session)", async () => {
+      // #given - active loop in session A
+      const hook = createRalphLoopHook(createMockPluginInput())
+      hook.startLoop("session-A", "First task", { maxIterations: 10 })
+      expect(hook.getState()?.session_id).toBe("session-A")
+      expect(hook.getState()?.prompt).toBe("First task")
+
+      // #when - start new loop in session B (without completing A)
+      hook.startLoop("session-B", "Second task", { maxIterations: 20 })
+
+      // #then - state should be overwritten with session B's loop
+      expect(hook.getState()?.session_id).toBe("session-B")
+      expect(hook.getState()?.prompt).toBe("Second task")
+      expect(hook.getState()?.max_iterations).toBe(20)
+      expect(hook.getState()?.iteration).toBe(1)
+
+      // #when - session B goes idle
+      await hook.event({
+        event: { type: "session.idle", properties: { sessionID: "session-B" } },
+      })
+
+      // #then - continuation should be injected for session B
+      expect(promptCalls.length).toBe(1)
+      expect(promptCalls[0].sessionID).toBe("session-B")
+      expect(promptCalls[0].text).toContain("Second task")
+      expect(promptCalls[0].text).toContain("2/20")
+
+      // #then - iteration incremented
+      expect(hook.getState()?.iteration).toBe(2)
+    })
+
+    test("should allow starting new loop in same session (restart)", async () => {
+      // #given - active loop in session A at iteration 5
+      const hook = createRalphLoopHook(createMockPluginInput())
+      hook.startLoop("session-A", "First task", { maxIterations: 10 })
+      
+      // Simulate some iterations
+      await hook.event({
+        event: { type: "session.idle", properties: { sessionID: "session-A" } },
+      })
+      await hook.event({
+        event: { type: "session.idle", properties: { sessionID: "session-A" } },
+      })
+      expect(hook.getState()?.iteration).toBe(3)
+      expect(promptCalls.length).toBe(2)
+
+      // #when - start NEW loop in same session (restart)
+      hook.startLoop("session-A", "Restarted task", { maxIterations: 50 })
+
+      // #then - state should be reset to iteration 1 with new prompt
+      expect(hook.getState()?.session_id).toBe("session-A")
+      expect(hook.getState()?.prompt).toBe("Restarted task")
+      expect(hook.getState()?.max_iterations).toBe(50)
+      expect(hook.getState()?.iteration).toBe(1)
+
+      // #when - session goes idle
+      promptCalls = [] // Reset to check new continuation
+      await hook.event({
+        event: { type: "session.idle", properties: { sessionID: "session-A" } },
+      })
+
+      // #then - continuation should use new task
+      expect(promptCalls.length).toBe(1)
+      expect(promptCalls[0].text).toContain("Restarted task")
+      expect(promptCalls[0].text).toContain("2/50")
+    })
+
    test("should check transcript BEFORE API to optimize performance", async () => {
      // #given - transcript has completion promise
      const transcriptPath = join(TEST_DIR, "transcript.jsonl")
--- a/src/hooks/sisyphus-orchestrator/index.test.ts
+++ b/src/hooks/sisyphus-orchestrator/index.test.ts
@@ -506,6 +506,90 @@ describe("sisyphus-orchestrator hook", () => {
        // #then
        expect(output.output).toBe(originalOutput)
      })
+
+      describe("cross-platform path validation (Windows support)", () => {
+        test("should NOT append reminder when orchestrator writes inside .sisyphus\\ (Windows backslash)", async () => {
+          // #given
+          const hook = createSisyphusOrchestratorHook(createMockPluginInput())
+          const originalOutput = "File written successfully"
+          const output = {
+            title: "Write",
+            output: originalOutput,
+            metadata: { filePath: ".sisyphus\\plans\\work-plan.md" },
+          }
+
+          // #when
+          await hook["tool.execute.after"](
+            { tool: "Write", sessionID: ORCHESTRATOR_SESSION },
+            output
+          )
+
+          // #then
+          expect(output.output).toBe(originalOutput)
+          expect(output.output).not.toContain("DELEGATION REQUIRED")
+        })
+
+        test("should NOT append reminder when orchestrator writes inside .sisyphus with mixed separators", async () => {
+          // #given
+          const hook = createSisyphusOrchestratorHook(createMockPluginInput())
+          const originalOutput = "File written successfully"
+          const output = {
+            title: "Write",
+            output: originalOutput,
+            metadata: { filePath: ".sisyphus\\plans/work-plan.md" },
+          }
+
+          // #when
+          await hook["tool.execute.after"](
+            { tool: "Write", sessionID: ORCHESTRATOR_SESSION },
+            output
+          )
+
+          // #then
+          expect(output.output).toBe(originalOutput)
+          expect(output.output).not.toContain("DELEGATION REQUIRED")
+        })
+
+        test("should NOT append reminder for absolute Windows path inside .sisyphus\\", async () => {
+          // #given
+          const hook = createSisyphusOrchestratorHook(createMockPluginInput())
+          const originalOutput = "File written successfully"
+          const output = {
+            title: "Write",
+            output: originalOutput,
+            metadata: { filePath: "C:\\Users\\test\\project\\.sisyphus\\plans\\x.md" },
+          }
+
+          // #when
+          await hook["tool.execute.after"](
+            { tool: "Write", sessionID: ORCHESTRATOR_SESSION },
+            output
+          )
+
+          // #then
+          expect(output.output).toBe(originalOutput)
+          expect(output.output).not.toContain("DELEGATION REQUIRED")
+        })
+
+        test("should append reminder for Windows path outside .sisyphus\\", async () => {
+          // #given
+          const hook = createSisyphusOrchestratorHook(createMockPluginInput())
+          const output = {
+            title: "Write",
+            output: "File written successfully",
+            metadata: { filePath: "C:\\Users\\test\\project\\src\\code.ts" },
+          }
+
+          // #when
+          await hook["tool.execute.after"](
+            { tool: "Write", sessionID: ORCHESTRATOR_SESSION },
+            output
+          )
+
+          // #then
+          expect(output.output).toContain("DELEGATION REQUIRED")
+        })
+      })
    })
  })

--- a/src/hooks/sisyphus-orchestrator/index.ts
+++ b/src/hooks/sisyphus-orchestrator/index.ts
@@ -14,7 +14,14 @@ import type { BackgroundManager } from "../../features/background-agent"

 export const HOOK_NAME = "sisyphus-orchestrator"

-const ALLOWED_PATH_PREFIX = ".sisyphus/"
+/**
+ * Cross-platform check if a path is inside .sisyphus/ directory.
+ * Handles both forward slashes (Unix) and backslashes (Windows).
+ */
+function isSisyphusPath(filePath: string): boolean {
+  return /\.sisyphus[/\\]/.test(filePath)
+}
+
 const WRITE_EDIT_TOOLS = ["Write", "Edit", "write", "edit"]

 const DIRECT_WORK_REMINDER = `
@@ -549,7 +556,7 @@ export function createSisyphusOrchestratorHook(
      // Check Write/Edit tools for orchestrator - inject strong warning
      if (WRITE_EDIT_TOOLS.includes(input.tool)) {
        const filePath = (output.args.filePath ?? output.args.path ?? output.args.file) as string | undefined
-        if (filePath && !filePath.includes(ALLOWED_PATH_PREFIX)) {
+        if (filePath && !isSisyphusPath(filePath)) {
          // Store filePath for use in tool.execute.after
          if (input.callID) {
            pendingFilePaths.set(input.callID, filePath)
@@ -593,7 +600,7 @@ export function createSisyphusOrchestratorHook(
        if (!filePath) {
          filePath = output.metadata?.filePath as string | undefined
        }
-        if (filePath && !filePath.includes(ALLOWED_PATH_PREFIX)) {
+        if (filePath && !isSisyphusPath(filePath)) {
          output.output = (output.output || "") + DIRECT_WORK_REMINDER
          log(`[${HOOK_NAME}] Direct work reminder appended`, {
            sessionID: input.sessionID,
--- a/src/index.ts
+++ b/src/index.ts
@@ -63,6 +63,7 @@ import {
  createSisyphusTask,
  interactive_bash,
  startTmuxCheck,
+  lspManager,
 } from "./tools";
 import { BackgroundManager } from "./features/background-agent";
 import { SkillMcpManager } from "./features/skill-mcp-manager";
@@ -164,7 +165,7 @@ const OhMyOpenCodePlugin: Plugin = async (ctx) => {
      })
    : null;
  const keywordDetector = isHookEnabled("keyword-detector")
-    ? createKeywordDetectorHook(ctx)
+    ? createKeywordDetectorHook(ctx, contextCollector)
    : null;
  const contextInjector = createContextInjectorHook(contextCollector);
  const contextInjectorMessagesTransform =
@@ -312,8 +313,8 @@ const OhMyOpenCodePlugin: Plugin = async (ctx) => {
    },

    "chat.message": async (input, output) => {
-      await claudeCodeHooks["chat.message"]?.(input, output);
      await keywordDetector?.["chat.message"]?.(input, output);
+      await claudeCodeHooks["chat.message"]?.(input, output);
      await contextInjector["chat.message"]?.(input, output);
      await autoSlashCommand?.["chat.message"]?.(input, output);
      await startWork?.["chat.message"]?.(input, output);
@@ -427,6 +428,7 @@ const OhMyOpenCodePlugin: Plugin = async (ctx) => {
        }
        if (sessionInfo?.id) {
          await skillMcpManager.disconnectSession(sessionInfo.id);
+          await lspManager.cleanupTempDirectoryClients();
        }
      }

--- a/src/shared/opencode-config-dir.test.ts
+++ b/src/shared/opencode-config-dir.test.ts
@@ -1,6 +1,6 @@
 import { describe, test, expect, beforeEach, afterEach } from "bun:test"
 import { homedir } from "node:os"
-import { join } from "node:path"
+import { join, resolve } from "node:path"
 import {
  getOpenCodeConfigDir,
  getOpenCodeConfigPaths,
@@ -20,6 +20,7 @@ describe("opencode-config-dir", () => {
      APPDATA: process.env.APPDATA,
      XDG_CONFIG_HOME: process.env.XDG_CONFIG_HOME,
      XDG_DATA_HOME: process.env.XDG_DATA_HOME,
+      OPENCODE_CONFIG_DIR: process.env.OPENCODE_CONFIG_DIR,
    }
  })

@@ -34,6 +35,84 @@ describe("opencode-config-dir", () => {
    }
  })

+  describe("OPENCODE_CONFIG_DIR environment variable", () => {
+    test("returns OPENCODE_CONFIG_DIR when env var is set", () => {
+      // #given OPENCODE_CONFIG_DIR is set to a custom path
+      process.env.OPENCODE_CONFIG_DIR = "/custom/opencode/path"
+      Object.defineProperty(process, "platform", { value: "linux" })
+
+      // #when getOpenCodeConfigDir is called with binary="opencode"
+      const result = getOpenCodeConfigDir({ binary: "opencode", version: "1.0.200" })
+
+      // #then returns the custom path
+      expect(result).toBe("/custom/opencode/path")
+    })
+
+    test("falls back to default when env var is not set", () => {
+      // #given OPENCODE_CONFIG_DIR is not set, platform is Linux
+      delete process.env.OPENCODE_CONFIG_DIR
+      delete process.env.XDG_CONFIG_HOME
+      Object.defineProperty(process, "platform", { value: "linux" })
+
+      // #when getOpenCodeConfigDir is called with binary="opencode"
+      const result = getOpenCodeConfigDir({ binary: "opencode", version: "1.0.200" })
+
+      // #then returns default ~/.config/opencode
+      expect(result).toBe(join(homedir(), ".config", "opencode"))
+    })
+
+    test("falls back to default when env var is empty string", () => {
+      // #given OPENCODE_CONFIG_DIR is set to empty string
+      process.env.OPENCODE_CONFIG_DIR = ""
+      delete process.env.XDG_CONFIG_HOME
+      Object.defineProperty(process, "platform", { value: "linux" })
+
+      // #when getOpenCodeConfigDir is called with binary="opencode"
+      const result = getOpenCodeConfigDir({ binary: "opencode", version: "1.0.200" })
+
+      // #then returns default ~/.config/opencode
+      expect(result).toBe(join(homedir(), ".config", "opencode"))
+    })
+
+    test("falls back to default when env var is whitespace only", () => {
+      // #given OPENCODE_CONFIG_DIR is set to whitespace only
+      process.env.OPENCODE_CONFIG_DIR = "   "
+      delete process.env.XDG_CONFIG_HOME
+      Object.defineProperty(process, "platform", { value: "linux" })
+
+      // #when getOpenCodeConfigDir is called with binary="opencode"
+      const result = getOpenCodeConfigDir({ binary: "opencode", version: "1.0.200" })
+
+      // #then returns default ~/.config/opencode
+      expect(result).toBe(join(homedir(), ".config", "opencode"))
+    })
+
+    test("resolves relative path to absolute path", () => {
+      // #given OPENCODE_CONFIG_DIR is set to a relative path
+      process.env.OPENCODE_CONFIG_DIR = "./my-opencode-config"
+      Object.defineProperty(process, "platform", { value: "linux" })
+
+      // #when getOpenCodeConfigDir is called with binary="opencode"
+      const result = getOpenCodeConfigDir({ binary: "opencode", version: "1.0.200" })
+
+      // #then returns resolved absolute path
+      expect(result).toBe(resolve("./my-opencode-config"))
+    })
+
+    test("OPENCODE_CONFIG_DIR takes priority over XDG_CONFIG_HOME", () => {
+      // #given both OPENCODE_CONFIG_DIR and XDG_CONFIG_HOME are set
+      process.env.OPENCODE_CONFIG_DIR = "/custom/opencode/path"
+      process.env.XDG_CONFIG_HOME = "/xdg/config"
+      Object.defineProperty(process, "platform", { value: "linux" })
+
+      // #when getOpenCodeConfigDir is called with binary="opencode"
+      const result = getOpenCodeConfigDir({ binary: "opencode", version: "1.0.200" })
+
+      // #then OPENCODE_CONFIG_DIR takes priority
+      expect(result).toBe("/custom/opencode/path")
+    })
+  })
+
  describe("isDevBuild", () => {
    test("returns false for null version", () => {
      expect(isDevBuild(null)).toBe(false)
@@ -213,6 +292,7 @@ describe("opencode-config-dir", () => {
      // #given no config files exist
      Object.defineProperty(process, "platform", { value: "linux" })
      delete process.env.XDG_CONFIG_HOME
+      delete process.env.OPENCODE_CONFIG_DIR

      // #when detectExistingConfigDir is called
      const result = detectExistingConfigDir("opencode", "1.0.200")
@@ -220,5 +300,19 @@ describe("opencode-config-dir", () => {
      // #then result is either null or a valid string path
      expect(result === null || typeof result === "string").toBe(true)
    })
+
+    test("includes OPENCODE_CONFIG_DIR in search locations when set", () => {
+      // #given OPENCODE_CONFIG_DIR is set to a custom path
+      process.env.OPENCODE_CONFIG_DIR = "/custom/opencode/path"
+      Object.defineProperty(process, "platform", { value: "linux" })
+      delete process.env.XDG_CONFIG_HOME
+
+      // #when detectExistingConfigDir is called
+      const result = detectExistingConfigDir("opencode", "1.0.200")
+
+      // #then result is either null (no config file exists) or a valid string path
+      // The important thing is that the function doesn't throw
+      expect(result === null || typeof result === "string").toBe(true)
+    })
  })
 })
--- a/src/shared/opencode-config-dir.ts
+++ b/src/shared/opencode-config-dir.ts
@@ -1,6 +1,6 @@
 import { existsSync } from "node:fs"
 import { homedir } from "node:os"
-import { join } from "node:path"
+import { join, resolve } from "node:path"

 export type OpenCodeBinaryType = "opencode" | "opencode-desktop"

@@ -47,6 +47,11 @@ function getTauriConfigDir(identifier: string): string {
 }

 function getCliConfigDir(): string {
+  const envConfigDir = process.env.OPENCODE_CONFIG_DIR?.trim()
+  if (envConfigDir) {
+    return resolve(envConfigDir)
+  }
+
  if (process.platform === "win32") {
    const crossPlatformDir = join(homedir(), ".config", "opencode")
    const crossPlatformConfig = join(crossPlatformDir, "opencode.json")
@@ -108,6 +113,11 @@ export function getOpenCodeConfigPaths(options: OpenCodeConfigDirOptions): OpenC
 export function detectExistingConfigDir(binary: OpenCodeBinaryType, version?: string | null): string | null {
  const locations: string[] = []

+  const envConfigDir = process.env.OPENCODE_CONFIG_DIR?.trim()
+  if (envConfigDir) {
+    locations.push(resolve(envConfigDir))
+  }
+
  if (binary === "opencode-desktop") {
    const identifier = isDevBuild(version) ? TAURI_APP_IDENTIFIER_DEV : TAURI_APP_IDENTIFIER
    locations.push(getTauriConfigDir(identifier))
--- a/src/tools/background-task/tools.ts
+++ b/src/tools/background-task/tools.ts
@@ -74,7 +74,7 @@ export function createBackgroundTask(manager: BackgroundManager): ToolDefinition
          parentSessionID: ctx.sessionID,
          parentMessageID: ctx.messageID,
          parentModel,
-          parentAgent: prevMessage?.agent,
+          parentAgent: ctx.agent ?? prevMessage?.agent,
        })

        ctx.metadata?.({
@@ -176,8 +176,13 @@ async function formatTaskResult(task: BackgroundTask, client: OpencodeClient): P
  // Handle both SDK response structures: direct array or wrapped in .data
  // eslint-disable-next-line @typescript-eslint/no-explicit-any
  const messages = ((messagesResult as any).data ?? messagesResult) as Array<{
-    info?: { role?: string }
-    parts?: Array<{ type?: string; text?: string }>
+    info?: { role?: string; time?: string }
+    parts?: Array<{ 
+      type?: string
+      text?: string
+      content?: string | Array<{ type: string; text?: string }>
+      name?: string
+    }>
  }>

  if (!Array.isArray(messages) || messages.length === 0) {
@@ -193,11 +198,13 @@ Session ID: ${task.sessionID}
 (No messages found)`
  }

-  const assistantMessages = messages.filter(
-    (m) => m.info?.role === "assistant"
+  // Include both assistant messages AND tool messages
+  // Tool results (grep, glob, bash output) come from role "tool"
+  const relevantMessages = messages.filter(
+    (m) => m.info?.role === "assistant" || m.info?.role === "tool"
  )

-  if (assistantMessages.length === 0) {
+  if (relevantMessages.length === 0) {
    return `Task Result

 Task ID: ${task.id}
@@ -207,17 +214,46 @@ Session ID: ${task.sessionID}

 ---

-(No assistant response found)`
+(No assistant or tool response found)`
  }

-  const lastMessage = assistantMessages[assistantMessages.length - 1]
-  const textParts = lastMessage?.parts?.filter(
-    (p) => p.type === "text"
-  ) ?? []
-  const textContent = textParts
-    .map((p) => p.text ?? "")
+  // Sort by time ascending (oldest first) to process messages in order
+  const sortedMessages = [...relevantMessages].sort((a, b) => {
+    const timeA = String((a as { info?: { time?: string } }).info?.time ?? "")
+    const timeB = String((b as { info?: { time?: string } }).info?.time ?? "")
+    return timeA.localeCompare(timeB)
+  })
+  
+  // Extract content from ALL messages, not just the last one
+  // Tool results may be in earlier messages while the final message is empty
+  const extractedContent: string[] = []
+  
+  for (const message of sortedMessages) {
+    for (const part of message.parts ?? []) {
+      // Handle both "text" and "reasoning" parts (thinking models use "reasoning")
+      if ((part.type === "text" || part.type === "reasoning") && part.text) {
+        extractedContent.push(part.text)
+      } else if (part.type === "tool_result") {
+        // Tool results contain the actual output from tool calls
+        const toolResult = part as { content?: string | Array<{ type: string; text?: string }> }
+        if (typeof toolResult.content === "string" && toolResult.content) {
+          extractedContent.push(toolResult.content)
+        } else if (Array.isArray(toolResult.content)) {
+          // Handle array of content blocks
+          for (const block of toolResult.content) {
+            // Handle both "text" and "reasoning" parts (thinking models use "reasoning")
+            if ((block.type === "text" || block.type === "reasoning") && block.text) {
+              extractedContent.push(block.text)
+            }
+          }
+        }
+      }
+    }
+  }
+  
+  const textContent = extractedContent
    .filter((text) => text.length > 0)
-    .join("\n")
+    .join("\n\n")

  const duration = formatDuration(task.startedAt, task.completedAt)

--- a/src/tools/call-omo-agent/tools.ts
+++ b/src/tools/call-omo-agent/tools.ts
@@ -170,23 +170,59 @@ async function executeSync(
  const messages = messagesResult.data
  log(`[call_omo_agent] Got ${messages.length} messages`)

+  // Include both assistant messages AND tool messages
+  // Tool results (grep, glob, bash output) come from role "tool"
  // eslint-disable-next-line @typescript-eslint/no-explicit-any
-  const lastAssistantMessage = messages
-    .filter((m: any) => m.info.role === "assistant")
-    .sort((a: any, b: any) => (b.info.time?.created || 0) - (a.info.time?.created || 0))[0]
+  const relevantMessages = messages.filter(
+    (m: any) => m.info?.role === "assistant" || m.info?.role === "tool"
+  )

-  if (!lastAssistantMessage) {
-    log(`[call_omo_agent] No assistant message found`)
+  if (relevantMessages.length === 0) {
+    log(`[call_omo_agent] No assistant or tool messages found`)
    log(`[call_omo_agent] All messages:`, JSON.stringify(messages, null, 2))
-    return `Error: No assistant response found\n\n<task_metadata>\nsession_id: ${sessionID}\n</task_metadata>`
+    return `Error: No assistant or tool response found\n\n<task_metadata>\nsession_id: ${sessionID}\n</task_metadata>`
  }

-  log(`[call_omo_agent] Found assistant message with ${lastAssistantMessage.parts.length} parts`)
+  log(`[call_omo_agent] Found ${relevantMessages.length} relevant messages`)

+  // Sort by time ascending (oldest first) to process messages in order
  // eslint-disable-next-line @typescript-eslint/no-explicit-any
-  const textParts = lastAssistantMessage.parts.filter((p: any) => p.type === "text")
-  // eslint-disable-next-line @typescript-eslint/no-explicit-any
-  const responseText = textParts.map((p: any) => p.text).join("\n")
+  const sortedMessages = [...relevantMessages].sort((a: any, b: any) => {
+    const timeA = a.info?.time?.created ?? 0
+    const timeB = b.info?.time?.created ?? 0
+    return timeA - timeB
+  })
+
+  // Extract content from ALL messages, not just the last one
+  // Tool results may be in earlier messages while the final message is empty
+  const extractedContent: string[] = []
+
+  for (const message of sortedMessages) {
+    // eslint-disable-next-line @typescript-eslint/no-explicit-any
+    for (const part of (message as any).parts ?? []) {
+      // Handle both "text" and "reasoning" parts (thinking models use "reasoning")
+      if ((part.type === "text" || part.type === "reasoning") && part.text) {
+        extractedContent.push(part.text)
+      } else if (part.type === "tool_result") {
+        // Tool results contain the actual output from tool calls
+        const toolResult = part as { content?: string | Array<{ type: string; text?: string }> }
+        if (typeof toolResult.content === "string" && toolResult.content) {
+          extractedContent.push(toolResult.content)
+        } else if (Array.isArray(toolResult.content)) {
+          // Handle array of content blocks
+          for (const block of toolResult.content) {
+            if ((block.type === "text" || block.type === "reasoning") && block.text) {
+              extractedContent.push(block.text)
+            }
+          }
+        }
+      }
+    }
+  }
+
+  const responseText = extractedContent
+    .filter((text) => text.length > 0)
+    .join("\n\n")

  log(`[call_omo_agent] Got response, length: ${responseText.length}`)

--- a/src/tools/index.ts
+++ b/src/tools/index.ts
@@ -10,8 +10,11 @@ import {
  lsp_rename,
  lsp_code_actions,
  lsp_code_action_resolve,
+  lspManager,
 } from "./lsp"

+export { lspManager }
+
 import {
  ast_grep_search,
  ast_grep_replace,
--- a/src/tools/lsp/client.ts
+++ b/src/tools/lsp/client.ts
@@ -182,6 +182,26 @@ class LSPServerManager {
      this.cleanupInterval = null
    }
  }
+
+  async cleanupTempDirectoryClients(): Promise<void> {
+    const keysToRemove: string[] = []
+    for (const [key, managed] of this.clients.entries()) {
+      const isTempDir = key.startsWith("/tmp/") || key.startsWith("/var/folders/")
+      const isIdle = managed.refCount === 0
+      if (isTempDir && isIdle) {
+        keysToRemove.push(key)
+      }
+    }
+    for (const key of keysToRemove) {
+      const managed = this.clients.get(key)
+      if (managed) {
+        this.clients.delete(key)
+        try {
+          await managed.client.stop()
+        } catch {}
+      }
+    }
+  }
 }

 export const lspManager = LSPServerManager.getInstance()
--- a/src/tools/sisyphus-task/tools.test.ts
+++ b/src/tools/sisyphus-task/tools.test.ts
@@ -259,6 +259,7 @@ describe("sisyphus-task", () => {

  describe("resume with background parameter", () => {
  test("resume with background=false should wait for result and return content", async () => {
+    // Note: This test needs extended timeout because the implementation has MIN_STABILITY_TIME_MS = 5000
    // #given
    const { createSisyphusTask } = require("./tools")
    
@@ -319,7 +320,7 @@ describe("sisyphus-task", () => {
    // #then - should contain actual result, not just "Background task resumed"
    expect(result).toContain("This is the resumed task result")
    expect(result).not.toContain("Background task resumed")
-  })
+  }, { timeout: 10000 })

  test("resume with background=true should return immediately without waiting", async () => {
    // #given
--- a/src/tools/sisyphus-task/tools.ts
+++ b/src/tools/sisyphus-task/tools.ts
@@ -221,6 +221,33 @@ Use \`background_output\` with task_id="${task.id}" to check progress.`
          return `❌ Failed to send resume prompt: ${errorMessage}\n\nSession ID: ${args.resume}`
        }

+        // Wait for message stability after prompt completes
+        const POLL_INTERVAL_MS = 500
+        const MIN_STABILITY_TIME_MS = 5000
+        const STABILITY_POLLS_REQUIRED = 3
+        const pollStart = Date.now()
+        let lastMsgCount = 0
+        let stablePolls = 0
+
+        while (Date.now() - pollStart < 60000) {
+          await new Promise(resolve => setTimeout(resolve, POLL_INTERVAL_MS))
+          
+          const elapsed = Date.now() - pollStart
+          if (elapsed < MIN_STABILITY_TIME_MS) continue
+
+          const messagesCheck = await client.session.messages({ path: { id: args.resume } })
+          const msgs = ((messagesCheck as { data?: unknown }).data ?? messagesCheck) as Array<unknown>
+          const currentMsgCount = msgs.length
+
+          if (currentMsgCount > 0 && currentMsgCount === lastMsgCount) {
+            stablePolls++
+            if (stablePolls >= STABILITY_POLLS_REQUIRED) break
+          } else {
+            stablePolls = 0
+            lastMsgCount = currentMsgCount
+          }
+        }
+
        const messagesResult = await client.session.messages({
          path: { id: args.resume },
        })
@@ -250,7 +277,8 @@ Use \`background_output\` with task_id="${task.id}" to check progress.`
          return `❌ No assistant response found.\n\nSession ID: ${args.resume}`
        }

-        const textParts = lastMessage?.parts?.filter((p) => p.type === "text") ?? []
+        // Extract text from both "text" and "reasoning" parts (thinking models use "reasoning")
+        const textParts = lastMessage?.parts?.filter((p) => p.type === "text" || p.type === "reasoning") ?? []
        const textContent = textParts.map((p) => p.text ?? "").filter(Boolean).join("\n")

        const duration = formatDuration(startTime)
@@ -390,13 +418,13 @@ System notifies on completion. Use \`background_output\` with task_id="${task.id
          metadata: { sessionId: sessionID, category: args.category, sync: true },
        })

-        // Use promptAsync to avoid changing main session's active state
+        // Use fire-and-forget prompt() - awaiting causes JSON parse errors with thinking models
+        // Note: Don't pass model in body - use agent's configured model instead
        let promptError: Error | undefined
-        await client.session.promptAsync({
+        client.session.prompt({
          path: { id: sessionID },
          body: {
            agent: agentToUse,
-            model: categoryModel,
            system: systemContent,
            tools: {
              task: false,
@@ -408,6 +436,9 @@ System notifies on completion. Use \`background_output\` with task_id="${task.id
          promptError = error instanceof Error ? error : new Error(String(error))
        })

+        // Small delay to let the prompt start
+        await new Promise(resolve => setTimeout(resolve, 100))
+
        if (promptError) {
          if (toastManager && taskId !== undefined) {
            toastManager.removeTask(taskId)
@@ -419,21 +450,63 @@ System notifies on completion. Use \`background_output\` with task_id="${task.id
          return `❌ Failed to send prompt: ${errorMessage}\n\nSession ID: ${sessionID}`
        }

-        // Poll for session completion
+        // Poll for session completion with stability detection
+        // The session may show as "idle" before messages appear, so we also check message stability
        const POLL_INTERVAL_MS = 500
        const MAX_POLL_TIME_MS = 10 * 60 * 1000
+        const MIN_STABILITY_TIME_MS = 10000  // Minimum 10s before accepting completion
+        const STABILITY_POLLS_REQUIRED = 3
        const pollStart = Date.now()
+        let lastMsgCount = 0
+        let stablePolls = 0

        while (Date.now() - pollStart < MAX_POLL_TIME_MS) {
          await new Promise(resolve => setTimeout(resolve, POLL_INTERVAL_MS))

+          // Check for async errors that may have occurred after the initial 100ms delay
+          // TypeScript doesn't understand async mutation, so we cast to check
+          const asyncError = promptError as Error | undefined
+          if (asyncError) {
+            if (toastManager && taskId !== undefined) {
+              toastManager.removeTask(taskId)
+            }
+            const errorMessage = asyncError.message
+            if (errorMessage.includes("agent.name") || errorMessage.includes("undefined")) {
+              return `❌ Agent "${agentToUse}" not found. Make sure the agent is registered in your opencode.json or provided by a plugin.\n\nSession ID: ${sessionID}`
+            }
+            return `❌ Failed to send prompt: ${errorMessage}\n\nSession ID: ${sessionID}`
+          }
+
          const statusResult = await client.session.status()
          const allStatuses = (statusResult.data ?? {}) as Record<string, { type: string }>
          const sessionStatus = allStatuses[sessionID]

-          // Break if session is idle OR no longer in status (completed and removed)
-          if (!sessionStatus || sessionStatus.type === "idle") {
-            break
+          // If session is actively running, reset stability
+          if (sessionStatus && sessionStatus.type !== "idle") {
+            stablePolls = 0
+            lastMsgCount = 0
+            continue
+          }
+
+          // Session is idle or not in status - check message stability
+          const elapsed = Date.now() - pollStart
+          if (elapsed < MIN_STABILITY_TIME_MS) {
+            continue  // Don't accept completion too early
+          }
+
+          // Get current message count
+          const messagesCheck = await client.session.messages({ path: { id: sessionID } })
+          const msgs = ((messagesCheck as { data?: unknown }).data ?? messagesCheck) as Array<unknown>
+          const currentMsgCount = msgs.length
+
+          if (currentMsgCount > 0 && currentMsgCount === lastMsgCount) {
+            stablePolls++
+            if (stablePolls >= STABILITY_POLLS_REQUIRED) {
+              break  // Messages stable for 3 polls - task complete
+            }
+          } else {
+            stablePolls = 0
+            lastMsgCount = currentMsgCount
          }
        }

@@ -459,7 +532,8 @@ System notifies on completion. Use \`background_output\` with task_id="${task.id
          return `❌ No assistant response found.\n\nSession ID: ${sessionID}`
        }
        
-        const textParts = lastMessage?.parts?.filter((p) => p.type === "text") ?? []
+        // Extract text from both "text" and "reasoning" parts (thinking models use "reasoning")
+        const textParts = lastMessage?.parts?.filter((p) => p.type === "text" || p.type === "reasoning") ?? []
        const textContent = textParts.map((p) => p.text ?? "").filter(Boolean).join("\n")

        const duration = formatDuration(startTime)
--- a/src/tools/skill/tools.ts
+++ b/src/tools/skill/tools.ts
@@ -194,4 +194,4 @@ export function createSkillTool(options: SkillLoadOptions = {}): ToolDefinition
  })
 }

-export const skill = createSkillTool()
+export const skill: ToolDefinition = createSkillTool()
--- a/src/tools/slashcommand/tools.ts
+++ b/src/tools/slashcommand/tools.ts
@@ -249,4 +249,4 @@ export function createSlashcommandTool(options: SlashcommandToolOptions = {}): T
 }

 // Default instance for backward compatibility (lazy loading)
-export const slashcommand = createSlashcommandTool()
+export const slashcommand: ToolDefinition = createSlashcommandTool()
Author	SHA1	Message	Date
github-actions[bot]	f240dbb7ee	release: v3.0.0-beta.4	2026-01-11 05:46:20 +00:00
YeonGyu-Kim	571810f1e7	fix(sisyphus-orchestrator): add cross-platform path validation for Windows support Add isSisyphusPath() helper function that handles both forward slashes (Unix) and backslashes (Windows) using regex pattern /\.sisyphus[/\\]/. Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-opencode) Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>	2026-01-11 14:42:53 +09:00
YeonGyu-Kim	83d958580f	librarian notice	2026-01-11 14:33:17 +09:00
github-actions[bot]	f1e7b6ab1e	@aw338WoWmUI has signed the CLA in code-yeongyu/oh-my-opencode#681	2026-01-11 05:03:55 +00:00
YeonGyu-Kim	1bbb61b1c2	fix(context-injector): inject via chat.message after claudeCodeHooks - Revert messages.transform-only approach (experimental hook unreliable) - Inject context in chat.message after claudeCodeHooks runs - Order: keywordDetector → claudeCodeHooks → contextInjector - Works independently of claude-code-hooks being enabled/disabled - Ultrawork content now reliably injected to model	2026-01-11 12:33:20 +09:00
YeonGyu-Kim	2a95c91cab	fix(context-injector): inject only via messages.transform to preserve UI - Remove contextInjector call from chat.message hook chain - Context injection now only happens in messages.transform hook - This ensures UI displays original user message while model receives prepended context - Fixes bug where commit message promised clone behavior but implementation mutated directly	2026-01-11 12:23:13 +09:00
Jeremy Gollehon	307d583ad6	fix(prometheus-md-only): cross-platform path validation for Windows support (#630 ) (#649 ) Replace brittle string checks with robust path.resolve/relative validation: - Fix Windows backslash paths (.sisyphus\plans\x.md) being incorrectly blocked - Fix case-sensitive extension check (.MD now accepted) - Add workspace confinement (block paths outside root even if containing .sisyphus) - Block nested .sisyphus directories (only first segment allowed) - Block path traversal attempts (.sisyphus/../secrets.md) - Use ALLOWED_EXTENSIONS and ALLOWED_PATH_PREFIX constants (case-insensitive) The new isAllowedFile() uses Node's path module for cross-platform compatibility instead of string includes/endsWith which failed on Windows separators.	2026-01-11 12:21:50 +09:00
YeonGyu-Kim	ce5315fbd0	refactor(keyword-detector): decouple from claude-code-hooks via ContextCollector pipeline - keyword-detector now registers keywords to ContextCollector - context-injector consumes and injects via chat.message hook - Removed keyword detection logic from claude-code-hooks - Hook order: keyword-detector → context-injector → claude-code-hooks - ultrawork now works even when claude-code-hooks is disabled	2026-01-11 12:06:16 +09:00
Kenny	1c262a65fe	feat: add OPENCODE_CONFIG_DIR environment variable support (#629 ) - Add env var check to getCliConfigDir() for config directory override - Update detectExistingConfigDir() to include env var path in locations - Add comprehensive tests (7 test cases) - Document in README Closes #627	2026-01-11 11:48:36 +09:00
Arthur Andrade	0c127879c0	fix(lsp): cleanup orphaned LSP servers on session.deleted (#676 ) * fix(lsp): cleanup orphaned LSP servers on session.deleted When parallel background agent tasks complete, their LSP servers (for repos cloned to /tmp/) remain running until a 5-minute idle timeout. This causes memory accumulation with heavy parallel Sisyphus usage, potentially leading to OOM crashes. This change adds cleanupTempDirectoryClients() to LSPServerManager (matching the pattern used by SkillMcpManager.disconnectSession()) and calls it on session.deleted events. The cleanup targets idle LSP clients (refCount=0) for temporary directories (/tmp/, /var/folders/) where agent tasks clone repos. * chore: retrigger CI checks	2026-01-11 11:45:38 +09:00
Nguyen Khac Trung Kien	65a6a702ec	Fix flowchart syntax in orchestration guide (#679 ) Updated the flowchart syntax in the orchestration guide.	2026-01-11 11:45:13 +09:00
github-actions[bot]	60f4cd4fac	release: v3.0.0-beta.3	2026-01-11 02:40:31 +00:00
github-actions[bot]	5f823b0f8e	release: v2.14.1	2026-01-11 02:23:00 +00:00
YeonGyu-Kim	e35a488cf6	fix(test): extend timeout for resume sync test MIN_STABILITY_TIME_MS is 5000ms in implementation, but test timeout was only 5000ms. Extended to 10000ms to allow proper polling completion.	2026-01-11 11:20:00 +09:00
YeonGyu-Kim	adb1a9fcb9	docs: fix model names in config examples to use valid antigravity models	2026-01-11 11:14:15 +09:00
YeonGyu-Kim	9bfed238b9	docs: update agent model catalog - librarian now uses GLM-4.7 Free	2026-01-11 11:11:34 +09:00
YeonGyu-Kim	61abd553fb	fix wrong merge.	2026-01-11 11:07:46 +09:00
github-actions[bot]	6425d9d97e	@KNN-07 has signed the CLA in code-yeongyu/oh-my-opencode#679	2026-01-11 01:11:47 +00:00
github-actions[bot]	d57744905f	@arthur404dev has signed the CLA in code-yeongyu/oh-my-opencode#676	2026-01-10 23:51:55 +00:00
github-actions[bot]	c7ae2d7be6	@ashir6892 has signed the CLA in code-yeongyu/oh-my-opencode#675	2026-01-10 19:50:19 +00:00
github-actions[bot]	358f7f439d	@kargnas has signed the CLA in code-yeongyu/oh-my-opencode#653	2026-01-10 10:25:35 +00:00
github-actions[bot]	4fde139dd8	@GollyJer has signed the CLA in code-yeongyu/oh-my-opencode#649	2026-01-10 09:57:54 +00:00
github-actions[bot]	b10703ec9a	@imarshallwidjaja has signed the CLA in code-yeongyu/oh-my-opencode#648	2026-01-10 07:58:53 +00:00
Brian Li	8b12257729	fix: remove author name from agent system prompts (#634 ) The author name "Named by [YeonGyu Kim]" in the Sisyphus role section causes LLMs to sometimes infer Korean language output, even when the user's locale is en-US. This happens because the model sees a Korean name in the system prompt and may interpret it as a signal to respond in Korean. Removing the author attribution from the runtime prompt fixes this issue. The attribution is preserved in README, LICENSE, and package.json. Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>	2026-01-10 14:11:48 +09:00
github-actions[bot]	7536a12754	@Luodian has signed the CLA in code-yeongyu/oh-my-opencode#634	2026-01-10 05:01:31 +00:00
Gladdonilli	0fb765732a	fix: improve background task completion detection and message extraction (#638 ) * fix: background task completion detection and silent notifications - Fix TS2742 by adding explicit ToolDefinition type annotations - Add stability detection (3 consecutive stable polls after 10s minimum) - Remove early continue when sessionStatus is undefined - Add silent notification system via tool.execute.after hook injection - Change task retention from 200ms to 5 minutes for background_output retrieval - Fix formatTaskResult to sort messages by time descending Fixes hanging background tasks that never complete due to missing sessionStatus. * fix: improve background task completion detection and message extraction - Add stability-based completion detection (10s min + 3 stable polls) - Fix message extraction to recognize 'reasoning' parts from thinking models - Switch from promptAsync() to prompt() for proper agent initialization - Remove model parameter from prompt body (use agent's configured model) - Add fire-and-forget prompt pattern for sisyphus_task sync mode - Add silent notification via tool.execute.after hook injection - Fix indentation issues in manager.ts and index.ts Incorporates fixes from: - PR #592: Stability detection mechanism - PR #610: Model parameter passing (partially) - PR #628: Completion detection improvements Known limitation: Thinking models (e.g. claude--thinking-) cause JSON Parse errors in child sessions. Use non-thinking models for background agents until OpenCode core resolves this. * fix: add tool_result handling and pendingByParent tracking for resume/external tasks Addresses code review feedback from PR #638: P1: Add tool_result type to validateSessionHasOutput() to prevent false negatives for tool-only background tasks that would otherwise timeout after 30 minutes despite having valid results. P2: Add pendingByParent tracking to resume() and registerExternalTask() to prevent premature 'ALL COMPLETE' notifications when mixing launched and resumed tasks. * fix: address code review feedback - log messages, model passthrough, sorting, race condition - Fix misleading log messages: 'promptAsync' -> 'prompt (fire-and-forget)' - Restore model passthrough in launch() for Sisyphus category configs - Fix call-omo-agent sorting: use time.created number instead of String(time) - Fix race condition: check promptError inside polling loop, not just after 100ms	2026-01-10 14:00:25 +09:00
github-actions[bot]	d4c8ec6690	@ElwinLiu has signed the CLA in code-yeongyu/oh-my-opencode#645	2026-01-10 04:32:31 +00:00
github-actions[bot]	d6416082a2	@kdcokenny has signed the CLA in code-yeongyu/oh-my-opencode#629	2026-01-09 12:54:17 +00:00
github-actions[bot]	e6aaf57a21	@SJY0917032 has signed the CLA in code-yeongyu/oh-my-opencode#625	2026-01-09 10:01:29 +00:00
YeonGyu-Kim	5242f3daef	fix(docs): correct plan invocation syntax from /plan to @plan OpenCode uses @agent-name syntax for agent invocation, not /command. The /plan command does not exist - it should be @plan to invoke the Prometheus planner agent.	2026-01-09 17:45:25 +09:00
YeonGyu-Kim	3f2ded54ee	fix(docs): escape special chars in Mermaid diagram Quote node label containing special characters to prevent Mermaid lexer error on line 9.	2026-01-09 17:24:03 +09:00
YeonGyu-Kim	aa5018583e	docs(orchestration): add TL;DR section for quick reference	2026-01-09 16:47:04 +09:00
YeonGyu-Kim	185d4e1e54	test(ralph-loop): add tests for loop restart scenarios - Add test for starting new loop while previous loop active (different session) - Add test for restarting loop in same session - Verifies startLoop properly overwrites state and resets iteration	2026-01-09 16:39:53 +09:00
YeonGyu-Kim	79e9fd82c5	fix(background-agent): preserve parent agent context in completion notifications When parentAgent is undefined, omit the agent field entirely from session.prompt body instead of passing undefined. This prevents the OpenCode SDK from falling back to defaultAgent(), which would change the parent session's agent context. Changes: - manager.ts: Build prompt body conditionally, only include agent/model when defined - background-task/tools.ts: Use ctx.agent as primary source for parentAgent (consistent with sisyphus-task) - registerExternalTask: Add parentAgent parameter support - Added tests for agent context preservation scenarios	2026-01-09 15:53:55 +09:00