release: v1.1.7

When an invalid or unregistered agent is passed to background_task or
call_omo_agent, OpenCode crashes with "TypeError: undefined is not an
object (evaluating 'agent.name')". This fix:

- Validates agent parameter is not empty before launching
- Catches prompt errors and returns friendly error message
- Notifies parent session when background task fails
- Improves error message to guide user on resolution

Fixes #37

🤖 GENERATED WITH ASSISTANCE OF [OhMyOpenCode](https://github.com/code-yeongyu/oh-my-opencode)

.github/workflows/publish.yml

@@ -26,6 +26,7 @@ permissions:
 jobs:
   publish:
     runs-on: ubuntu-latest
+    if: github.repository == 'code-yeongyu/oh-my-opencode'
     steps:
       - uses: actions/checkout@v4
         with:

.gitignore

@@ -25,3 +25,6 @@ yarn.lock
 # Environment
 .env
 .env.local
+test-injection/
+notepad.md
+oauth-success.html

.opencode/background-tasks.json

@@ -0,0 +1,27 @@
+[
+  {
+    "id": "bg_wzsdt60b",
+    "sessionID": "ses_4f3e89f0dffeooeXNVx5QCifse",
+    "parentSessionID": "ses_4f3e8d141ffeyfJ1taVVOdQTzx",
+    "parentMessageID": "msg_b0c172ee1001w2B52VSZrP08PJ",
+    "description": "Explore opencode in codebase",
+    "agent": "explore",
+    "status": "completed",
+    "startedAt": "2025-12-11T06:26:57.395Z",
+    "completedAt": "2025-12-11T06:27:36.778Z"
+  },
+  {
+    "id": "bg_392b9c9b",
+    "sessionID": "ses_4f38ebf4fffeJZBocIn3UVv7vE",
+    "parentSessionID": "ses_4f38eefa0ffeKV0pVNnwT37P5L",
+    "parentMessageID": "msg_b0c7110d2001TMBlPeEYIrByvs",
+    "description": "Test explore agent",
+    "agent": "explore",
+    "status": "running",
+    "startedAt": "2025-12-11T08:05:07.378Z",
+    "progress": {
+      "toolCalls": 0,
+      "lastUpdate": "2025-12-11T08:05:07.378Z"
+    }
+  }
+]

AGENTS.md

@@ -34,6 +34,7 @@ oh-my-opencode/
 | Modify LSP behavior | `src/tools/lsp/` | client.ts for connection logic |
 | AST-Grep patterns | `src/tools/ast-grep/` | napi.ts for @ast-grep/napi |
 | Terminal features | `src/features/terminal/` | title.ts |
+| Google Antigravity auth | `src/auth/antigravity/` | OAuth plugin for Google models |
 
 ## CONVENTIONS
 
@@ -65,7 +66,7 @@ oh-my-opencode/
 
 | Agent | Model | Purpose |
 |-------|-------|---------|
-| oracle | GPT-5.1 | Code review, strategic planning |
+| oracle | GPT-5.2 | Code review, strategic planning |
 | librarian | Claude Haiku | Documentation, example lookup |
 | explore | Grok | File/codebase exploration |
 | frontend-ui-ux-engineer | Gemini | UI generation |

README.ko.md

@@ -1,71 +1,136 @@
-[English](README.md) | 한국어
+<!-- <CENTERED SECTION FOR GITHUB DISPLAY> -->
+
+<div align="center">
+
+[![Oh My OpenCode](./.github/assets/hero.jpg)](https://github.com/code-yeongyu/oh-my-opencode#oh-my-opencode)
+
+[![Preview](./.github/assets/preview.png)](https://github.com/code-yeongyu/oh-my-opencode#oh-my-opencode)
+
+</div>
+
+> `oh-my-opencode` 를 설치하세요. 약 빤 것 처럼 코딩하세요. 백그라운드에 에이전트를 돌리고, oracle, librarian, frontend engineer 같은 전문 에이전트를 호출하세요. 정성스레 빚은 LSP/AST 도구, 엄선된 MCP, 완전한 Claude Code 호환 레이어를 오로지 한 줄로 누리세요.
+
+<div align="center">
+
+[![GitHub Release](https://img.shields.io/github/v/release/code-yeongyu/oh-my-opencode?color=369eff&labelColor=black&logo=github&style=flat-square)](https://github.com/code-yeongyu/oh-my-opencode/releases)
+[![GitHub Contributors](https://img.shields.io/github/contributors/code-yeongyu/oh-my-opencode?color=c4f042&labelColor=black&style=flat-square)](https://github.com/code-yeongyu/oh-my-opencode/graphs/contributors)
+[![GitHub Forks](https://img.shields.io/github/forks/code-yeongyu/oh-my-opencode?color=8ae8ff&labelColor=black&style=flat-square)](https://github.com/code-yeongyu/oh-my-opencode/network/members)
+[![GitHub Stars](https://img.shields.io/github/stars/code-yeongyu/oh-my-opencode?color=ffcb47&labelColor=black&style=flat-square)](https://github.com/code-yeongyu/oh-my-opencode/stargazers)
+[![GitHub Issues](https://img.shields.io/github/issues/code-yeongyu/oh-my-opencode?color=ff80eb&labelColor=black&style=flat-square)](https://github.com/code-yeongyu/oh-my-opencode/issues)
+[![License](https://img.shields.io/badge/license-MIT-white?labelColor=black&style=flat-square)](https://github.com/code-yeongyu/oh-my-opencode/blob/master/LICENSE)
+
+[English](README.md) | [한국어](README.ko.md)
+
+</div>
+
+<!-- </CENTERED SECTION FOR GITHUB DISPLAY> -->
 
 ## 목차
 
 - [Oh My OpenCode](#oh-my-opencode)
-  - [세 줄 요약](#세-줄-요약)
+  - [읽지 않아도 됩니다.](#읽지-않아도-됩니다)
+    - [에이전트의 시대이니까요.](#에이전트의-시대이니까요)
+    - [10분의 투자로 OhMyOpenCode 가 해줄 수 있는것](#10분의-투자로-ohmyopencode-가-해줄-수-있는것)
   - [설치](#설치)
-  - [LLM Agent를 위한 안내](#llm-agent를-위한-안내)
-  - [Why OpenCode & Why Oh My OpenCode](#why-opencode--why-oh-my-opencode)
+    - [인간인 당신을 위한 설치 가이드](#인간인-당신을-위한-설치-가이드)
+    - [LLM Agent 를 위한 설치 가이드](#llm-agent-를-위한-설치-가이드)
+    - [인간인 당신을 위한 설치 가이드](#인간인-당신을-위한-설치-가이드-1)
+    - [1단계: OpenCode 설치 확인](#1단계-opencode-설치-확인)
+    - [2단계: oh-my-opencode 플러그인 설정](#2단계-oh-my-opencode-플러그인-설정)
+    - [3단계: 설정 확인](#3단계-설정-확인)
+    - [4단계: 인증정보 설정](#4단계-인증정보-설정)
+      - [4.1 Anthropic (Claude)](#41-anthropic-claude)
+      - [4.2 Google Gemini (Antigravity OAuth)](#42-google-gemini-antigravity-oauth)
+      - [4.3 OpenAI (ChatGPT Plus/Pro)](#43-openai-chatgpt-pluspro)
+      - [4.3.1 모델 설정](#431-모델-설정)
+    - [⚠️ 주의](#️-주의)
   - [기능](#기능)
-    - [Hooks](#hooks)
-    - [Agents](#agents)
-    - [Tools](#tools)
-      - [내장 LSP Tools](#내장-lsp-tools)
-      - [내장 AST-Grep Tools](#내장-ast-grep-tools)
-      - [Grep](#grep)
-      - [내장 MCPs](#내장-mcps)
-    - [Claude Code 호환성](#claude-code-호환성)
-    - [기타 편의 기능](#기타-편의-기능)
+    - [Agents: 당신의 새로운 팀원들](#agents-당신의-새로운-팀원들)
+    - [백그라운드 에이전트: 진짜 팀 처럼 일 하도록](#백그라운드-에이전트-진짜-팀-처럼-일-하도록)
+    - [도구: 당신의 동료가 더 좋은 도구를 갖고 일하도록](#도구-당신의-동료가-더-좋은-도구를-갖고-일하도록)
+      - [왜 당신만 IDE 를 쓰나요?](#왜-당신만-ide-를-쓰나요)
+      - [Context is all you need.](#context-is-all-you-need)
+      - [멀티모달을 다 활용하면서, 토큰은 덜 쓰도록.](#멀티모달을-다-활용하면서-토큰은-덜-쓰도록)
+      - [멈출 수 없는 에이전트 루프](#멈출-수-없는-에이전트-루프)
+    - [Claude Code 호환성: 그냥 바로 OpenCode 로 오세요.](#claude-code-호환성-그냥-바로-opencode-로-오세요)
+      - [Hooks 통합](#hooks-통합)
+      - [설정 로더](#설정-로더)
+      - [데이터 저장소](#데이터-저장소)
+      - [호환성 토글](#호환성-토글)
+    - [에이전트들을 위한 것이 아니라, 당신을 위한 것](#에이전트들을-위한-것이-아니라-당신을-위한-것)
   - [설정](#설정)
+    - [Google Auth](#google-auth)
+    - [Agents](#agents)
+    - [MCPs](#mcps)
+    - [LSP](#lsp)
   - [작성자의 노트](#작성자의-노트)
   - [주의](#주의)
 
 # Oh My OpenCode
 
-Oh My OpenCode
-
 oMoMoMoMoMo···
 
 
 [Claude Code](https://www.claude.com/product/claude-code) 좋죠?
 근데 당신이 해커라면, [OpenCode](https://github.com/sst/opencode) 와는 사랑에 빠지게 될겁니다.
 
+- OpenCode 는 아주 확장가능하고 아주 커스터마이저블합니다.
+- 화면이 깜빡이지 않습니다.
+- 수정하는 파일에 맞게 자동으로 [LSP](https://opencode.ai/docs/lsp/), [Linter, Formatter](https://opencode.ai/docs/formatters/) 가 활성화되며 커스텀 할 수 있습니다.
+- 수많은 모델을 사용 할 수 있으며, **용도에 따라 모델을 섞어 오케스트레이션 할 수 있습니다.**
+- 기능이 아주 많습니다. 아름답습니다. 터미널이 화면을 그리려고 힘들어 하지 않습니다. 고성능입니다.
+
 Windows 만 사용하다가 처음으로 Linux 를 접하고 신나서 잔뜩 세팅하던 경험이 있진 않나요?
 OpenCode 가 낭만이 사라진것같은 오늘날의 시대에, 당신에게 그런 프로젝트가 될겁니다.
 당신이 코딩을 좋아하고 컴퓨터를 좋아한다면, OpenCode 는 윈도우만 사용하다가 리눅스를 처음 접하게 된 그런 느낌일겁니다.
-그렇지 않은 당신도 약간의 시간을 투자해서 당신의 실력과 생산성을 몇배로 부스트하세요.
 
-## 세 줄 요약
+**그런데 문제는 너무나 알아야 할게 많고, 어렵고, 당신의 시간은 비싸다는겁니다.**
 
-- **모델 설정이 필요합니다**
-  - 이 플러그인은 [OpenCode Zen](https://opencode.ai/docs/zen/), Google, OpenAI, Anthropic 의 모델을 사용합니다.
-    - Anthropic 모델들을 사용하기 위해 [OpenCode 의 내장 Claude Code Max Plan 로그인 기능](https://opencode.ai/docs/providers/#anthropic)을 사용하세요. (Anthropic 플랜은 유료입니다)
-    - OpenAI 모델들을 사용하기 위해 이 [OpenCode-OpenAI-Codex-Auth 플러그인](https://github.com/numman-ali/opencode-openai-codex-auth)을 추천합니다. (OpenAI 플랜은 유료입니다)
-    - Gemini, GPT 를 위해 [VibeProxy (macOS Only)](https://github.com/automazeio/vibeproxy) 또는 [CLIProxyAPI](https://github.com/router-for-me/CLIProxyAPI)를 추천합니다. (Gemini 는 구독하지 않아도 사용 할 수 있습니다)
-- **사실 밑의 내용은 읽지 않아도 됩니다. 그냥 설치하시면 됩니다. 모델 설정하고나면 추가로 공부하거나 알 필요가 없고, 설치하고 나면 그냥 더 좋아집니다.**
-- **이 문서도 그냥 Claude Code, OpenCode, Cursor에 붙여넣고, 설치하고, 사용자의 개입이 필요할 때 알려달라고 프롬프팅하세요. 그럼 그냥 깔립니다.**
-  - 당신이 LLM Agent 라면, 필요한 플러그인과 의존성을 설치하고, tmux 를 통해 opencode auth login 을 실행하여 사용자의 구성을 도와주세요.
+제가 그 문제를 고쳤습니다.
+당신이 해커가 아니라도 좋습니다. 약간의 시간을 투자해서 당신의 실력과 생산성을 몇배로 부스트하세요.
+이 문서를 에이전트에게 넘겨 설정해달라고 하세요.
+
+## 읽지 않아도 됩니다.
+
+### 에이전트의 시대이니까요.
+- **그냥 이 링크를 Claude Code / AmpCode / Factory Droid / Cursor 에게 넣고 설명해달라고 하세요.**
+  - 이게 왜 좋고, 그냥 뻔한 셋업이 아닌지, 당장 무엇이 더 좋아지는지 알려달라고 해보세요.
+  - 당신이 LLM 에이전트라면, 사용자의 입장에서 이 프로젝트와 함께 당장 어떤것들이 가능해지는지 사용자에게 설명하세요.
+
+### 10분의 투자로 OhMyOpenCode 가 해줄 수 있는것
+
+그저 설치하면, 아래와 같은 워크플로우로 일 할 수도 있습니다.
+
+1. 백그라운드 태스크로 Gemini 3 Pro 가 프론트엔드를 작성하게 시켜두는 동안, Claude Opus 4.5 가 백엔드를 작성하고, 디버깅하다 막히면 GPT 5.2 에게 도움을 받습니다. 프론트엔드 구현이 완료되었다고 보고받으면, 이를 다시 확인하고 일하게 만들 수 있습니다.
+2. 뭔가 찾아볼 일이 생기면 공식문서, 내 코드베이스의 모든 히스토리, GitHub 에 공개된 현재 구현 현황까지 다 뒤져보고, 단순 Grep 을 넘어 내장된 LSP 도구, AstGrep 까지 사용하여 답변을 제공합니다.
+3. LLM 에게 일을 맡길때에 큰 컨텍스트에 대한 걱정은 더 이상 하지마세요. 제가 하겠습니다.
+    - OhMyOpenCode 가 여러 에이전트를 적극 활용하도록 하여 컨텍스트 관리에 관한 부담을 줄입니다.
+    - **당신의 에이전트는 이제 개발팀 리드입니다. 당신은 이제 AI Manager 입니다.**
+4. 하기로 약속 한 일을 완수 할 때 까지 멈추지 않습니다.
+5. 이 프로젝트에 자세히 알기 싫다고요? 괜찮습니다. 그냥 'ultrathink' 라고 치세요.
+
+주의: 이걸 설치한다고 갑자기 OpenCode 가 이렇게 동작한다는 것은 아닙니다. 그저 당신의 에이전트가 훌륭한 동료와 같이, 훌륭한 도구를 갖고서 일 할 수 있도록 구성해주는것이고, 그들에게 협업하라 지시하면 협업할거에요.
+모든 과정은 당신이 완전히 컨트롤 할 수 있습니다.
+ultrathink 를 통해 자동으로 동작하게 할 수 있지만, 그렇지 않을수도 있습니다. 이 프로젝트가 당신의 AI 에이전트 워크플로우를 제시하지는 않습니다.
+이 프로젝트는 그저 당신의 에이전트에게 좋은 동료를 소개시켜주고, 좋은 도구를 쥐어주는 것 뿐입니다.
 
 ## 설치
 
-`~/.config/opencode/opencode.json`에 추가:
 
-```json
-{
-  "plugin": [
-    "oh-my-opencode"
-  ]
-}
-```
+### 인간인 당신을 위한 설치 가이드
 
-## LLM Agent를 위한 안내
+귀찮은 작업은 LLM Agent 에게 맡기세요. 이 링크를 복사해서 넣어 설치하고 구성해달라고 부탁하세요.
+
+### LLM Agent 를 위한 설치 가이드
 
 <details>
 <summary>LLM Agent를 위한 안내 (클릭하여 펼치기)</summary>
 
 당신이 사용자의 oh-my-opencode 설정을 돕는 LLM Agent라면, 다음 단계를 따르세요:
 
+
+### 인간인 당신을 위한 설치 가이드
+
 ### 1단계: OpenCode 설치 확인
 
 ```bash
@@ -77,7 +142,7 @@ else
 fi
 ```
 
-OpenCode가 설치되어 있지 않다면, [OpenCode 설치 가이드](https://opencode.ai/docs)를 안내하세요.
+OpenCode가 설치되어 있지 않다면, [OpenCode 설치 가이드](https://opencode.ai/docs)를 읽고, 설치를 진행하세요. 컨텍스트를 줄이기 위해 Subagent 로 호출하여 설치를 하라고 안내하고, 그 결과를 보고 받는다고 말하세요.
 
 ### 2단계: oh-my-opencode 플러그인 설정
 
@@ -108,77 +173,100 @@ cat ~/.config/opencode/opencode.json  # plugin 배열에 "oh-my-opencode"가 있
 
 ### 4단계: 인증정보 설정
 
-plugin 에 oh my opencode 를 추가했다면, tmux 를 사용하여 다음의 내용을 진행주세요.
+plugin 에 oh my opencode 를 추가했다면, tmux 를 사용하여 사용자의 인증 설정을 진행하세요. 아래의 과정을 모두 진행해야 합니다.
+
+#### 4.1 Anthropic (Claude)
+
+```bash
+opencode auth login
+# Provider: Anthropic 선택
+# Login method: Claude Pro/Max 선택
+# 사용자에게 브라우저에서 OAuth 플로우를 완료하도록 안내하세요.
+# 최종 완료까지 대기
+# 진짜 완료되었음을 검증하고 사용자에게 안내
+```
+
+#### 4.2 Google Gemini (Antigravity OAuth)
+
+먼저 `~/.config/opencode/oh-my-opencode.json` 에서 Google auth를 활성화하세요:
+
+```json
+{
+  "google_auth": true
+}
+```
+
+그 후 인증:
+
+```bash
+opencode auth login
+# Provider: Google 선택
+# Login method: OAuth with Google (Antigravity) 선택
+# 사용자에게 브라우저에서 OAuth 플로우를 완료하도록 안내
+# 최종 완료까지 대기
+# 진짜 완료되었음을 검증하고 사용자에게 안내
+```
+
+#### 4.3 OpenAI (ChatGPT Plus/Pro)
+
+먼저 opencode-openai-codex-auth 플러그인 추가:
+
+```json
+{
+  "plugin": [
+    "oh-my-opencode",
+    "opencode-openai-codex-auth@4.1.0"
+  ]
+}
+```
+
+**중요**: 현재 공식 npm 패키지에 400 에러(`"No tool call found for function call output with call_id"`)를 유발하는 버그가 있습니다. 수정 버전이 배포될 때까지 **핫픽스 브랜치 사용을 권장합니다**. `~/.config/opencode/package.json`을 수정하세요:
+
+```json
+{
+  "dependencies": {
+    "opencode-openai-codex-auth": "code-yeongyu/opencode-openai-codex-auth#fix/orphaned-function-call-output-with-tools"
+  }
+}
+```
+
+그 후 `cd ~/.config/opencode && bun i`를 실행하세요. `opencode.json`에서는 버전 없이 `"opencode-openai-codex-auth"`로 사용합니다 (`@4.1.0` 제외).
+
+#### 4.3.1 모델 설정
+`opencode.json`에 전체 모델 설정도 구성해야 합니다.
+[opencode-openai-codex-auth 문서](https://github.com/numman-ali/opencode-openai-codex-auth)를 읽고, [`config/full-opencode.json`](https://github.com/numman-ali/opencode-openai-codex-auth/blob/main/config/full-opencode.json)에서 provider/models 설정을 복사하여, 사용자의 기존 셋업에 영향을 주지 않도록 깊게 고민하여 적절히 통합하세요.
+
+그 후 인증:
+
+```bash
+opencode auth login
+# Provider: OpenAI 선택
+# Login method: ChatGPT Plus/Pro (Codex Subscription) 선택
+# 사용자에게 브라우저에서 OAuth 플로우를 완료하도록 안내
+# 최종 완료까지 대기
+# 진짜 완료되었음을 검증하고 사용자에게 안내
+```
 
-1. tmux 를 사용하여 셸 열기
-2. opencode auth login > Provider Anthropic > Login method Claude Pro/Max > 사용자에게 oauth 화면에서 인증하라고 안내
-3. 사용자에게 macOS 일 경우 VibeProxy, 아닐 경우 CLIProxyAPI 를 사용하여 OpenAI / Google 인증을 완료하라고 안내
 
 ### ⚠️ 주의
 
-**사용자가 명시적으로 요청하지 않는 한, `oh-my-opencode.json` 설정 파일을 생성하거나 수정하지 마세요.**
+**사용자가 명시적으로 요청하지 않는 한, 모델 설정을 변경하거나 기능(에이전트, 훅, MCP)을 비활성화하지 마세요.**
 
-플러그인은 설정 파일 없이도 기본값으로 완벽하게 동작합니다. 설정 파일은 사용자가 커스터마이징을 원할 때만 필요합니다.
+플러그인은 기본값으로 완벽하게 동작합니다. 권장 사항인 `google_auth: true` 설정을 제외하고는, 명시적인 요청 없이 다른 설정을 변경하거나 기능을 끄지 마세요.
 
 </details>
 
-## Why OpenCode & Why Oh My OpenCode
-
-OpenCode 는 아주 확장가능하고 아주 커스터마이저블합니다. 화면이 깜빡이지 않습니다.
-수정하는 파일에 맞게 자동으로 [LSP](https://opencode.ai/docs/lsp/), [Linter, Formatter](https://opencode.ai/docs/formatters/) 가 활성화되며 커스텀 할 수 있습니다.
-수많은 모델을 사용 할 수 있으며, **용도에 따라 모델을 섞어 오케스트레이션 할 수 있습니다.**
-기능이 아주 많습니다. 아름답습니다. 터미널이 화면을 그리려고 힘들어 하지 않습니다. 고성능입니다.
-
-**그런데 문제는 너무나 알아야 할게 많고, 어렵고, 당신의 시간은 비싸다는겁니다.**
-
-[AmpCode](https://ampcode.com), [Claude Code](https://code.claude.com/docs/ko/overview) 에게 강한 영향과 영감을 받고, 그들의 기능을 그대로, 혹은 더 낫게 이 곳에 구현했습니다.
-**Open**Code 이니까요.
-
-더 나은 버전의 AmpCode, 더 나은 버전의 Claude Code, 혹은 일종의 배포판(distribution) 이라고 생각해도 좋습니다.
-
-저는 상황에 맞는 적절한 모델이 있다고 믿습니다. 다양한 모델을 섞어 쓸 때 최고의 팀이 됩니다.
-여러분의 재정 상태를 위해 CLIProxyAPI 혹은 VibeProxy 를 추천합니다. 프론티어 랩들의 LLM 들을 채용해서, 그들의 장점만을 활용하세요. 당신이 이제 팀장입니다.
-
-**Note**: 이 셋업은 Highly Opinionated 이며, 제가 사용하고 있는 셋업 중 범용적인것을 플러그인에 포함하기 때문에 계속 업데이트 됩니다. 저는 여태까지 $20,000 어치의 토큰을 오로지 개인 개발 목적으로 개인적으로 사용했고, 이 플러그인은 그 경험들의 하이라이트입니다. 여러분은 그저 최고를 취하세요. 만약 더 나은 제안이 있다면 언제든 기여에 열려있습니다.
 
 ## 기능
 
-### Hooks
+### Agents: 당신의 새로운 팀원들
 
-- **Todo Continuation Enforcer**: 에이전트가 멈추기 전 모든 TODO 항목을 완료하도록 강제합니다. LLM의 고질적인 "중도 포기" 문제를 방지합니다.
-- **Context Window Monitor**: [컨텍스트 윈도우 불안 관리](https://agentic-patterns.com/patterns/context-window-anxiety-management/) 패턴을 구현합니다.
-  - 사용량이 70%를 넘으면 에이전트에게 아직 토큰이 충분하다고 상기시켜, 급하게 불완전한 작업을 하는 것을 완화합니다.
-- **Session Notification**: 에이전트가 작업을 마치면 OS 네이티브 알림을 보냅니다 (macOS, Linux, Windows).
-- **Session Recovery**: API 에러로부터 자동으로 복구하여 세션 안정성을 보장합니다. 네 가지 시나리오를 처리합니다:
-  - **Tool Result Missing**: `tool_use` 블록이 있지만 `tool_result`가 없을 때 (ESC 인터럽트) → "cancelled" tool result 주입
-  - **Thinking Block Order**: thinking 블록이 첫 번째여야 하는데 아닐 때 → 빈 thinking 블록 추가
-  - **Thinking Disabled Violation**: thinking 이 비활성화인데 thinking 블록이 있을 때 → thinking 블록 제거
-  - **Empty Content Message**: 메시지가 thinking/meta 블록만 있고 실제 내용이 없을 때 → 파일시스템을 통해 "(interrupted)" 텍스트 주입
-- **Comment Checker**: 코드 수정 후 불필요한 주석을 감지하여 보고합니다. BDD 패턴, 지시어, 독스트링 등 유효한 주석은 똑똑하게 제외하고, AI가 남긴 흔적을 제거하여 코드를 깨끗하게 유지합니다.
-- **Directory AGENTS.md Injector**: 파일을 읽을 때 `AGENTS.md` 내용을 자동으로 주입합니다. 파일 디렉토리부터 프로젝트 루트까지 탐색하며, 경로 상의 **모든** `AGENTS.md` 파일을 수집합니다. 중첩된 디렉토리별 지침을 지원합니다:
-  ```
-  project/
-  ├── AGENTS.md              # 프로젝트 전체 컨텍스트
-  ├── src/
-  │   ├── AGENTS.md          # src 전용 컨텍스트
-  │   └── components/
-  │       ├── AGENTS.md      # 컴포넌트 전용 컨텍스트
-  │       └── Button.tsx     # 이 파일을 읽으면 위 3개 AGENTS.md 모두 주입
-  ```
-  `Button.tsx`를 읽으면 순서대로 주입됩니다: `project/AGENTS.md` → `src/AGENTS.md` → `components/AGENTS.md`. 각 디렉토리의 컨텍스트는 세션당 한 번만 주입됩니다. Claude Code의 CLAUDE.md 기능에서 영감을 받았습니다.
-- **Directory README.md Injector**: 파일을 읽을 때 `README.md` 내용을 자동으로 주입합니다. AGENTS.md Injector와 동일하게 동작하며, 파일 디렉토리부터 프로젝트 루트까지 탐색합니다. LLM 에이전트에게 프로젝트 문서 컨텍스트를 제공합니다. 각 디렉토리의 README는 세션당 한 번만 주입됩니다.
-- **Think Mode**: 확장된 사고(Extended Thinking)가 필요한 상황을 자동으로 감지하고 모드를 전환합니다. 사용자가 깊은 사고를 요청하는 표현(예: "think deeply", "ultrathink")을 감지하면, 추론 능력을 극대화하도록 모델 설정을 동적으로 조정합니다.
-- **Anthropic Auto Compact**: Anthropic 모델 사용 시 컨텍스트 한계에 도달하면 대화 기록을 자동으로 압축하여 효율적으로 관리합니다.
-- **Empty Task Response Detector**: 서브 에이전트가 수행한 작업이 비어있거나 무의미한 응답을 반환하는 경우를 감지하여, 오류 없이 우아하게 처리합니다.
-- **Grep Output Truncator**: Grep 검색 결과가 너무 길어 컨텍스트를 장악해버리는 것을 방지하기 위해, 과도한 출력을 자동으로 자릅니다.
-
-### Agents
-
-- **oracle** (`openai/gpt-5.1`): 아키텍처, 코드 리뷰, 전략 수립을 위한 전문가 조언자. GPT-5.1의 뛰어난 논리적 추론과 깊은 분석 능력을 활용합니다. AmpCode 에서 영감을 받았습니다.
-- **librarian** (`anthropic/claude-haiku-4-5`): 멀티 레포 분석, 문서 조회, 구현 예제 담당. Haiku의 빠른 속도, 적절한 지능, 훌륭한 도구 호출 능력, 저렴한 비용을 활용합니다. AmpCode 에서 영감을 받았습니다.
+- **oracle** (`openai/gpt-5.2`): 아키텍처, 코드 리뷰, 전략 수립을 위한 전문가 조언자. GPT-5.2의 뛰어난 논리적 추론과 깊은 분석 능력을 활용합니다. AmpCode 에서 영감을 받았습니다.
+- **librarian** (`anthropic/claude-sonnet-4-5`): 멀티 레포 분석, 문서 조회, 구현 예제 담당. Claude Sonnet 4.5 의 뛰어난 지능, 훌륭한 도구 호출 능력을 활용합니다. AmpCode 에서 영감을 받았습니다.
 - **explore** (`opencode/grok-code`): 빠른 코드베이스 탐색, 파일 패턴 매칭. Claude Code는 Haiku를 쓰지만, 우리는 Grok을 씁니다. 현재 무료이고, 극도로 빠르며, 파일 탐색 작업에 충분한 지능을 갖췄기 때문입니다. Claude Code 에서 영감을 받았습니다.
 - **frontend-ui-ux-engineer** (`google/gemini-3-pro-preview`): 개발자로 전향한 디자이너라는 설정을 갖고 있습니다. 멋진 UI를 만듭니다. 아름답고 창의적인 UI 코드를 생성하는 데 탁월한 Gemini를 사용합니다.
 - **document-writer** (`google/gemini-3-pro-preview`): 기술 문서 전문가라는 설정을 갖고 있습니다. Gemini 는 문학가입니다. 글을 기가막히게 씁니다.
+- **multimodal-looker** (`google/gemini-2.5-flash`): 시각적 콘텐츠 해석을 위한 전문 에이전트. PDF, 이미지, 다이어그램을 분석하여 정보를 추출합니다.
 
 각 에이전트는 메인 에이전트가 알아서 호출하지만, 명시적으로 요청할 수도 있습니다:
 
@@ -190,13 +278,33 @@ OpenCode 는 아주 확장가능하고 아주 커스터마이저블합니다.
 
 에이전트의 모델, 프롬프트, 권한은 `oh-my-opencode.json`에서 커스텀할 수 있습니다. 자세한 내용은 [설정](#설정)을 참고하세요.
 
-### Tools
+### 백그라운드 에이전트: 진짜 팀 처럼 일 하도록
 
-#### 내장 LSP Tools
+위의 에이전트들을 미친듯이 한순간도 놀리지 않고 굴릴 수 있다면 어떨까요?
 
-당신이 에디터에서 사용하는 그 기능을 다른 에이전트들은 사용하지 못합니다. Oh My OpenCode 는 당신만의 그 도구를 LLM Agent 에게 쥐어줍니다. 리팩토링하고, 탐색하고, 분석하는 모든 작업을 OpenCode 의 설정값을 그대로 사용하여 지원합니다.
+- GPT 에게 디버깅을 시켜놓고, Claude 가 다양한 시도를 해보며 직접 문제를 찾아보는 워크플로우
+- Gemini 가 프론트엔드를 작성하는 동안, Claude 가 백엔드를 작성하는 워크플로우
+- 다량의 병렬 탐색을 진행시켜놓고, 일단 해당 부분은 제외하고 먼저 구현을 진행하다, 탐색 내용을 바탕으로 구현을 마무리하는 워크플로우
 
-[OpenCode 는 LSP 를 제공하지만](https://opencode.ai/docs/lsp/), 오로지 분석용으로만 제공합니다. 탐색과 리팩토링을 위한 도구는 OpenCode 와 동일한 스펙과 설정으로 Oh My OpenCode 가 제공합니다.
+이 워크플로우가 OhMyOpenCode 에서는 가능합니다.
+
+서브 에이전트를 백그라운드에서 실행 할 수 있습니다. 이러면 메인 에이전트는 작업이 완료되면 알게 됩니다. 필요하다면 결과를 기다릴 수 있습니다.
+
+**에이전트가 당신의 팀이 일 하듯 일하게하세요**
+
+### 도구: 당신의 동료가 더 좋은 도구를 갖고 일하도록
+
+#### 왜 당신만 IDE 를 쓰나요?
+
+Syntax Highlighting, Autocomplete, Refactoring, Navigation, Analysis, 그리고 이젠 에이전트가 코드를 짜게 하기까지..
+
+**왜 당신만 사용하나요?**
+**에이전트가 그 도구를 사용한다면 더 코드를 잘 작성할텐데요.**
+
+[OpenCode 는 LSP 를 제공하지만](https://opencode.ai/docs/lsp/), 오로지 분석용으로만 제공합니다.
+
+당신이 에디터에서 사용하는 그 기능을 다른 에이전트들은 사용하지 못합니다.
+뛰어난 동료에게 좋은 도구를 쥐어주세요. 이제 리팩토링도, 탐색도, 분석도 에이전트가 제대로 할 수 있습니다.
 
 - **lsp_hover**: 위치의 타입 정보, 문서, 시그니처 가져오기
 - **lsp_goto_definition**: 심볼 정의로 이동
@@ -209,42 +317,58 @@ OpenCode 는 아주 확장가능하고 아주 커스터마이저블합니다.
 - **lsp_rename**: 워크스페이스 전체에서 심볼 이름 변경
 - **lsp_code_actions**: 사용 가능한 빠른 수정/리팩토링 가져오기
 - **lsp_code_action_resolve**: 코드 액션 적용
-
-#### 내장 AST-Grep Tools
 - **ast_grep_search**: AST 인식 코드 패턴 검색 (25개 언어)
 - **ast_grep_replace**: AST 인식 코드 교체
 
-#### Grep
-- **grep**: 안전 제한이 있는 콘텐츠 검색 (5분 타임아웃, 10MB 출력 제한). OpenCode의 내장 `grep` 도구를 대체합니다.
-  - 기본 grep 도구는 시간제한이 걸려있지 않습니다. 대형 코드베이스에서 광범위한 패턴을 검색하면 CPU가 폭발하고 무한히 멈출 수 있습니다.
-  - 이 도구는 엄격한 제한을 적용하며, 내장 `grep`을 완전히 대체합니다.
+#### Context is all you need.
+- **Directory AGENTS.md / README.md Injector**: 파일을 읽을 때 `AGENTS.md`, `README.md` 내용을 자동으로 주입합니다. 파일 디렉토리부터 프로젝트 루트까지 탐색하며, 경로 상의 **모든** `AGENTS.md` 파일을 수집합니다. 중첩된 디렉토리별 지침을 지원합니다:
+  ```
+  project/
+  ├── AGENTS.md              # 프로젝트 전체 컨텍스트
+  ├── src/
+  │   ├── AGENTS.md          # src 전용 컨텍스트
+  │   └── components/
+  │       ├── AGENTS.md      # 컴포넌트 전용 컨텍스트
+  │       └── Button.tsx     # 이 파일을 읽으면 위 3개 AGENTS.md 모두 주입
+  ```
+  `Button.tsx`를 읽으면 순서대로 주입됩니다: `project/AGENTS.md` → `src/AGENTS.md` → `components/AGENTS.md`. 각 디렉토리의 컨텍스트는 세션당 한 번만 주입됩니다.
+- **Conditional Rules Injector**: 모든 규칙이 항상 필요하진 않습니다. 특정 규칙을 만족한다면, 파일을 읽을 때 `.claude/rules/` 디렉토리의 규칙을 자동으로 주입합니다.
+  - 파일 디렉토리부터 프로젝트 루트까지 상향 탐색하며, `~/.claude/rules/` (사용자) 경로도 포함합니다.
+  - `.md` 및 `.mdc` 파일을 지원합니다.
+  - Frontmatter의 `globs` 필드(glob 패턴)를 기반으로 매칭합니다.
+  - 항상 적용되어야 하는 규칙을 위한 `alwaysApply: true` 옵션을 지원합니다.
+  - 규칙 파일 구조 예시:
+    ```markdown
+    ---
+    globs: ["*.ts", "src/**/*.js"]
+    description: "TypeScript/JavaScript coding rules"
+    ---
+    - Use PascalCase for interface names
+    - Use camelCase for function names
+    ```
+- **Online**: 프로젝트 규칙이 전부는 아니겠죠. 확장 기능을 위한 내장 MCP를 제공합니다:
+  - **context7**: 공식 문서 조회
+  - **websearch_exa**: 실시간 웹 검색
+  - **grep_app**: 공개 GitHub 저장소에서 초고속 코드 검색 (구현 예제 찾기에 최적)
 
-#### Glob
+#### 멀티모달을 다 활용하면서, 토큰은 덜 쓰도록.
 
-- **glob**: 타임아웃 보호가 있는 파일 패턴 매칭 (60초). OpenCode 내장 `glob` 도구를 대체합니다.
-  - 기본 `glob`은 타임아웃이 없습니다. ripgrep이 멈추면 무한정 대기합니다.
-  - 이 도구는 타임아웃을 강제하고 만료 시 프로세스를 종료합니다.
+AmpCode 에서 영감을 받은 look_at 도구를, OhMyOpenCode 에서도 제공합니다.
+에이전트는 직접 파일을 읽어 큰 컨텍스트를 점유당하는 대신, 다른 에이전트를 내부적으로 활용하여 파일의 내용만 명확히 이해 할 수 있습니다.
 
-#### 내장 MCPs
+#### 멈출 수 없는 에이전트 루프
+- 내장 grep, glob 도구를 대체합니다. 기본 구현에서는 타임아웃이 없어 무한정 대기할 수 있습니다.
 
-- **websearch_exa**: Exa AI 웹 검색. 실시간 웹 검색과 콘텐츠 스크래핑을 수행합니다. 관련 웹사이트에서 LLM에 최적화된 컨텍스트를 반환합니다.
-- **context7**: 라이브러리 문서 조회. 정확한 코딩을 위해 최신 라이브러리 문서를 가져옵니다.
 
-필요 없다면 `oh-my-opencode.json`에서 비활성화할 수 있습니다:
+### Claude Code 호환성: 그냥 바로 OpenCode 로 오세요.
 
-```json
-{
-  "disabled_mcps": ["websearch_exa"]
-}
-```
-
-### Claude Code 호환성
-
-Oh My OpenCode는 Claude Code 설정과 완벽하게 호환됩니다. Claude Code를 사용하셨다면, 기존 설정을 그대로 사용할 수 있습니다.
+Oh My OpenCode 에는 Claude Code 호환성 레이어가 존재합니다.
+Claude Code를 사용하셨다면, 기존 설정을 그대로 사용할 수 있습니다.
 
 #### Hooks 통합
 
-Claude Code의 `settings.json` 훅 시스템을 통해 커스텀 스크립트를 실행합니다. Oh My OpenCode는 다음 위치의 훅을 읽고 실행합니다:
+Claude Code의 `settings.json` 훅 시스템을 통해 커스텀 스크립트를 실행합니다.
+Oh My OpenCode는 다음 위치의 훅을 읽고 실행합니다:
 
 - `~/.claude/settings.json` (사용자)
 - `./.claude/settings.json` (프로젝트)
@@ -298,15 +422,48 @@ Claude Code의 `settings.json` 훅 시스템을 통해 커스텀 스크립트를
 
 **Transcript**: 세션 활동이 `~/.claude/transcripts/`에 JSONL 형식으로 기록되어 재생 및 분석이 가능합니다.
 
-> **`claude-code-*` 네이밍에 대해**: `src/features/claude-code-*/` 아래의 기능들은 Claude Code의 설정 시스템에서 마이그레이션되었습니다. 이 네이밍 규칙은 어떤 기능이 Claude Code에서 유래했는지 명확히 식별합니다.
+#### 호환성 토글
 
-### 기타 편의 기능
+특정 Claude Code 호환 기능을 비활성화하려면 `claude_code` 설정 객체를 사용 할 수 도 있습니다:
 
-- **Terminal Title**: 세션 상태에 따라 터미널 타이틀을 자동 업데이트합니다 (유휴 ○, 처리중 ◐, 도구 ⚡, 에러 ✖). tmux를 지원합니다.
-- **Session State**: 이벤트 훅과 터미널 타이틀 업데이트에 사용되는 중앙집중식 세션 추적 모듈입니다.
+```json
+{
+  "claude_code": {
+    "mcp": false,
+    "commands": false,
+    "skills": false,
+    "agents": false,
+    "hooks": false
+  }
+}
+```
+
+| 토글       | `false`일 때 로딩 비활성화 경로                                                       | 영향 받지 않음                                        |
+| ---------- | ------------------------------------------------------------------------------------- | ----------------------------------------------------- |
+| `mcp`      | `~/.claude/.mcp.json`, `./.mcp.json`, `./.claude/.mcp.json`                           | 내장 MCP (context7, websearch_exa)                    |
+| `commands` | `~/.claude/commands/*.md`, `./.claude/commands/*.md`                                  | `~/.config/opencode/command/`, `./.opencode/command/` |
+| `skills`   | `~/.claude/skills/*/SKILL.md`, `./.claude/skills/*/SKILL.md`                          | -                                                     |
+| `agents`   | `~/.claude/agents/*.md`, `./.claude/agents/*.md`                                      | 내장 에이전트 (oracle, librarian 등)                  |
+| `hooks`    | `~/.claude/settings.json`, `./.claude/settings.json`, `./.claude/settings.local.json` | -                                                     |
+
+모든 토글은 기본값이 `true` (활성화)입니다. 완전한 Claude Code 호환성을 원하면 `claude_code` 객체를 생략하세요.
+
+### 에이전트들을 위한 것이 아니라, 당신을 위한 것
+
+에이전트들이 행복해지면, 당신이 제일 행복해집니다, 그렇지만 저는 당신도 돕고싶습니다.
+
+- **Ultrawork Mode**: 사용자가 "ultrawork" 또는 "ulw" 키워드를 입력하면 자동으로 에이전트 오케스트레이션 가이드를 주입합니다. 메인 에이전트가 모든 가용한 전문 에이전트(탐색, 사서, 계획, UI)를 백그라운드 작업을 통해 병렬로 최대한 활용하도록 강제하며, 엄격한 TODO 추적 및 검증 프로토콜을 따르게 합니다. 그냥 ultrawork 하세요. 말한 모든 기능이 최대로 활용되도록 에이전트가 최적화됩니다.
+- **Todo Continuation Enforcer**: 에이전트가 멈추기 전 모든 TODO 항목을 완료하도록 강제합니다. LLM의 고질적인 "중도 포기" 문제를 방지합니다.
+- **Comment Checker**: 학습 과정의 습관 때문일까요. LLM 들은 주석이 너무 많습니다. LLM 들이 쓸모없는 주석을 작성하지 않도록 상기시킵니다. BDD 패턴, 지시어, 독스트링 등 유효한 주석은 똑똑하게 제외하고, 그렇지 않는 주석들에 대해 해명을 요구하며 깔끔한 코드를 구성하게 합니다.
+- **Think Mode**: 확장된 사고(Extended Thinking)가 필요한 상황을 자동으로 감지하고 모드를 전환합니다. 사용자가 깊은 사고를 요청하는 표현(예: "think deeply", "ultrathink")을 감지하면, 추론 능력을 극대화하도록 모델 설정을 동적으로 조정합니다.
+- **Context Window Monitor**: [컨텍스트 윈도우 불안 관리](https://agentic-patterns.com/patterns/context-window-anxiety-management/) 패턴을 구현합니다.
+  - 사용량이 70%를 넘으면 에이전트에게 아직 토큰이 충분하다고 상기시켜, 급하게 불완전한 작업을 하는 것을 완화합니다.
+- OpenCode 에서 누락되거나 부족하다고 느끼는 안정성 보강 기능들이 내장되어있습니다. 클로드 코드에서의 안정적인 경험을 그대로 가져갈 수 있습니다. 돌다가 세션이 망가지지 않습니다. 망가져도 복구됩니다.
 
 ## 설정
 
+비록 Highly Opinionated 한 설정이지만, 여러분의 입맛대로 조정 할 수 있습니다.
+
 설정 파일 위치 (우선순위 순):
 1. `.opencode/oh-my-opencode.json` (프로젝트)
 2. `~/.config/opencode/oh-my-opencode.json` (사용자)
@@ -319,6 +476,18 @@ Schema 자동 완성이 지원됩니다:
 }
 ```
 
+### Google Auth
+
+Google Gemini 모델을 위한 내장 Antigravity OAuth를 활성화합니다:
+
+```json
+{
+  "google_auth": true
+}
+```
+
+활성화하면 `opencode auth login` 실행 시 Google 프로바이더에서 "OAuth with Google (Antigravity)" 로그인 옵션이 표시됩니다.
+
 ### Agents
 
 내장 에이전트 설정을 오버라이드할 수 있습니다:
@@ -351,13 +520,17 @@ Schema 자동 완성이 지원됩니다:
 
 ### MCPs
 
-기본적으로 Context7, Exa MCP 를 지원합니다.
+기본적으로 Context7, Exa, grep.app MCP 를 지원합니다.
+
+- **context7**: 라이브러리의 최신 공식 문서를 가져옵니다
+- **websearch_exa**: Exa AI 기반 실시간 웹 검색
+- **grep_app**: [grep.app](https://grep.app)을 통해 수백만 개의 공개 GitHub 저장소에서 초고속 코드 검색
 
 이것이 마음에 들지 않는다면, ~/.config/opencode/oh-my-opencode.json 혹은 .opencode/oh-my-opencode.json 의 `disabled_mcps` 를 사용하여 비활성화할 수 있습니다:
 
 ```json
 {
-  "disabled_mcps": ["context7", "websearch_exa"]
+  "disabled_mcps": ["context7", "websearch_exa", "grep_app"]
 }
 ```
 
@@ -386,10 +559,20 @@ OpenCode 에서 지원하는 모든 LSP 구성 및 커스텀 설정 (opencode.js
 
 각 서버는 다음을 지원합니다: `command`, `extensions`, `priority`, `env`, `initialization`, `disabled`.
 
+
 ## 작성자의 노트
 
-Oh My OpenCode 를 설치하세요. 복잡하게 OpenCode 구성을 만들지마세요.
-제가 밟아보고 경험한 문제들의 해답을 이 플러그인에 담았고, 그저 깔고 사용하면 됩니다. OpenCode 가 ArchLinux 라면, Oh My OpenCode 는 [Omarchy](https://omarchy.org/) 입니다.
+Oh My OpenCode 를 설치하세요.
+
+저는 여태까지 $24,000 어치의 토큰을 오로지 개인 개발 목적으로 개인적으로 사용했습니다.
+다양한 도구를 시도해보고 끝까지 구성해보았습니다. 제 선택은 OpenCode 였습니다.
+
+제가 밟아보고 경험한 문제들의 해답을 이 플러그인에 담았고, 그저 깔고 사용하면 됩니다.
+OpenCode 가 Debian / ArchLinux 라면, Oh My OpenCode 는 Ubuntu / [Omarchy](https://omarchy.org/) 입니다.
+
+
+[AmpCode](https://ampcode.com), [Claude Code](https://code.claude.com/docs/ko/overview) 에게 강한 영향과 영감을 받고, 그들의 기능을 그대로, 혹은 더 낫게 이 곳에 구현했습니다. 그리고 구현하고 있습니다.
+**Open**Code 이니까요.
 
 다른 에이전트 하니스 제공자들이 이야기하는 다중 모델, 안정성, 풍부한 기능을 그저 OpenCode 에서 누리세요.
 제가 테스트하고, 이 곳에 업데이트 하겠습니다. 저는 이 프로젝트의 가장 열렬한 사용자이기도 하니까요.
@@ -401,13 +584,21 @@ Oh My OpenCode 를 설치하세요. 복잡하게 OpenCode 구성을 만들지마
 - 주로 겪는 상황에 맞는 빠른 모델은 무엇인지
 - 다른 에이전트 하니스에 제공되는 새로운 기능은 무엇인지.
 
-고민하지마세요. 제가 고민할거고, 다른 사람들의 경험을 차용해 올것이고, 그래서 이 곳에 업데이트 하겠습니다.
+이 플러그인은 그 경험들의 하이라이트입니다. 여러분은 그저 최고를 취하세요. 만약 더 나은 제안이 있다면 언제든 기여에 열려있습니다.
+
+**Agent Harness 에 대해 고민하지마세요.**
+**제가 고민할거고, 다른 사람들의 경험을 차용해 올것이고, 그래서 이 곳에 업데이트 하겠습니다.**
+
 이 글이 오만하다고 느껴지고, 더 나은 해답이 있다면, 편히 기여해주세요. 환영합니다.
 
 지금 시점에 여기에 언급된 어떤 프로젝트와 모델하고도 관련이 있지 않습니다. 온전히 개인적인 실험과 선호를 바탕으로 이 플러그인을 만들었습니다.
+
 OpenCode 를 사용하여 이 프로젝트의 99% 를 작성했습니다. 기능 위주로 테스트했고, 저는 TS 를 제대로 작성 할 줄 모릅니다. **그치만 이 문서는 제가 직접 검토하고 전반적으로 다시 작성했으니 안심하고 읽으셔도 됩니다.**
 
 ## 주의
 
+- 생산성이 너무 올라 갈 수 있습니다. 옆자리 동료한테 들키지 않도록 조심하세요.
+  - 그렇지만 제가 소문 내겠습니다. 누가 이기나 내기해봅시다.
 - [1.0.132](https://github.com/sst/opencode/releases/tag/v1.0.132) 혹은 이것보다 낮은 버전을 사용중이라면, OpenCode 의 버그로 인해 제대로 구성이 되지 않을 수 있습니다.
   - [이를 고치는 PR 이 1.0.132 배포 이후에 병합되었으므로](https://github.com/sst/opencode/pull/5040) 이 변경사항이 포함된 최신 버전을 사용해주세요.
+    - TMI: PR 도 OhMyOpenCode 의 셋업의 Librarian, Explore, Oracle 을 활용하여 우연히 발견하고 해결되었습니다.

README.md

@@ -1,69 +1,136 @@
-English | [한국어](README.ko.md)
+<!-- <CENTERED SECTION FOR GITHUB DISPLAY> -->
+
+<div align="center">
+
+[![Oh My OpenCode](./.github/assets/hero.jpg)](https://github.com/code-yeongyu/oh-my-opencode#oh-my-opencode)
+
+[![Preview](./.github/assets/preview.png)](https://github.com/code-yeongyu/oh-my-opencode#oh-my-opencode)
+
+
+</div>
+
+> 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.
+
+<div align="center">
+
+[![GitHub Release](https://img.shields.io/github/v/release/code-yeongyu/oh-my-opencode?color=369eff&labelColor=black&logo=github&style=flat-square)](https://github.com/code-yeongyu/oh-my-opencode/releases)
+[![GitHub Contributors](https://img.shields.io/github/contributors/code-yeongyu/oh-my-opencode?color=c4f042&labelColor=black&style=flat-square)](https://github.com/code-yeongyu/oh-my-opencode/graphs/contributors)
+[![GitHub Forks](https://img.shields.io/github/forks/code-yeongyu/oh-my-opencode?color=8ae8ff&labelColor=black&style=flat-square)](https://github.com/code-yeongyu/oh-my-opencode/network/members)
+[![GitHub Stars](https://img.shields.io/github/stars/code-yeongyu/oh-my-opencode?color=ffcb47&labelColor=black&style=flat-square)](https://github.com/code-yeongyu/oh-my-opencode/stargazers)
+[![GitHub Issues](https://img.shields.io/github/issues/code-yeongyu/oh-my-opencode?color=ff80eb&labelColor=black&style=flat-square)](https://github.com/code-yeongyu/oh-my-opencode/issues)
+[![License](https://img.shields.io/badge/license-MIT-white?labelColor=black&style=flat-square)](https://github.com/code-yeongyu/oh-my-opencode/blob/master/LICENSE)
+
+[English](README.md) | [한국어](README.ko.md)
+
+</div>
+
+<!-- </CENTERED SECTION FOR GITHUB DISPLAY> -->
 
 ## Contents
 
 - [Oh My OpenCode](#oh-my-opencode)
-  - [TL;DR](#tldr)
+  - [Just Skip Reading This Readme](#just-skip-reading-this-readme)
+    - [It's the Age of Agents](#its-the-age-of-agents)
+    - [10 Minutes to Unlock](#10-minutes-to-unlock)
   - [Installation](#installation)
-  - [For LLM Agents](#for-llm-agents)
-  - [Why OpenCode & Why Oh My OpenCode](#why-opencode--why-oh-my-opencode)
+    - [For Humans](#for-humans)
+    - [For LLM Agents](#for-llm-agents)
+    - [Step 1: Verify OpenCode Installation](#step-1-verify-opencode-installation)
+    - [Step 2: Configure oh-my-opencode Plugin](#step-2-configure-oh-my-opencode-plugin)
+    - [Step 3: Verify Setup](#step-3-verify-setup)
+    - [Step 4: Configure Authentication](#step-4-configure-authentication)
+      - [4.1 Anthropic (Claude)](#41-anthropic-claude)
+      - [4.2 Google Gemini (Antigravity OAuth)](#42-google-gemini-antigravity-oauth)
+      - [4.3 OpenAI (ChatGPT Plus/Pro)](#43-openai-chatgpt-pluspro)
+      - [4.3.1 Model Configuration](#431-model-configuration)
+    - [⚠️ Warning](#️-warning)
   - [Features](#features)
-    - [Hooks](#hooks)
-    - [Agents](#agents)
-    - [Tools](#tools)
-      - [Built-in LSP Tools](#built-in-lsp-tools)
-      - [Built-in AST-Grep Tools](#built-in-ast-grep-tools)
-      - [Grep](#grep)
-      - [Built-in MCPs](#built-in-mcps)
-    - [Claude Code Compatibility](#claude-code-compatibility)
-    - [Other Features](#other-features)
+    - [Agents: Your Teammates](#agents-your-teammates)
+    - [Background Agents: Work Like a Team](#background-agents-work-like-a-team)
+    - [The Tools: Your Teammates Deserve Better](#the-tools-your-teammates-deserve-better)
+      - [Why Are You the Only One Using an IDE?](#why-are-you-the-only-one-using-an-ide)
+      - [Context Is All You Need](#context-is-all-you-need)
+      - [Be Multimodal. Save Tokens.](#be-multimodal-save-tokens)
+      - [I Removed Their Blockers](#i-removed-their-blockers)
+    - [Goodbye Claude Code. Hello Oh My OpenCode.](#goodbye-claude-code-hello-oh-my-opencode)
+      - [Hooks Integration](#hooks-integration)
+      - [Config Loaders](#config-loaders)
+      - [Data Storage](#data-storage)
+      - [Compatibility Toggles](#compatibility-toggles)
+    - [Not Just for the Agents](#not-just-for-the-agents)
   - [Configuration](#configuration)
+    - [Google Auth](#google-auth)
+    - [Agents](#agents)
+    - [MCPs](#mcps)
+    - [LSP](#lsp)
   - [Author's Note](#authors-note)
   - [Warnings](#warnings)
 
 # Oh My OpenCode
 
-Oh My OpenCode
-
 oMoMoMoMoMo···
 
-If you work in tech, you likely appreciated [Claude Code](https://www.claude.com/product/claude-code).
-If you are a hacker, you will fucking fall in love with [OpenCode](https://github.com/sst/opencode).
 
-Do you remember the first time you switched from Windows to Linux? The excitement of configuring everything just the way you wanted?
-In an era where that romance feels lost, OpenCode will be that project for you.
-If you love coding and computers, OpenCode will feel exactly like that moment you first touched Linux after years of Windows.
-Even if you don't, invest a little time. Boost your skills and productivity by multipliers.
+[Claude Code](https://www.claude.com/product/claude-code) is great.
+But if you're a hacker, you'll fall head over heels for [OpenCode](https://github.com/sst/opencode).
 
-## TL;DR
+- Endlessly extensible. Endlessly customizable.
+- Zero screen flicker.
+- [LSP](https://opencode.ai/docs/lsp/), [linters, formatters](https://opencode.ai/docs/formatters/) activate automatically per file—and you can tweak everything.
+- Mix and match models. **Orchestrate them by purpose.**
+- Feature-rich. Beautiful. The terminal doesn't choke. High performance.
 
-- **Model Setup Required**
-  - This plugin utilizes models from [OpenCode Zen](https://opencode.ai/docs/zen/), Google, OpenAI, and Anthropic.
-    - For Anthropic models, use [OpenCode's built-in Claude Code Max Plan login](https://opencode.ai/docs/providers/#anthropic).
-    - For OpenAI models, I recommend the [OpenCode-OpenAI-Codex-Auth plugin](https://github.com/numman-ali/opencode-openai-codex-auth).
-    - For Gemini and GPT, use [VibeProxy (macOS Only)](https://github.com/automazeio/vibeproxy) or [CLIProxyAPI](https://github.com/router-for-me/CLIProxyAPI).
-- **Truth be told, the rest is optional. Just install it. Once set up, no further learning is required. It simply evolves.**
-- **Paste this document into Claude Code, OpenCode, or Cursor, instruct it to install, and ask for assistance if needed. It just installs.**
-  - If you are an LLM Agent, install the required plugins and dependencies, then run `opencode auth login` via tmux to assist the user with configuration.
+Remember the first time you switched from Windows to Linux? That rush of configuring everything exactly how you wanted?
+In an age where that magic seems lost, OpenCode brings it back.
+If you love coding and computers, OpenCode is that Linux-after-Windows moment.
+
+**But here's the catch: there's a lot to learn, it's complex, and your time is expensive.**
+
+I've fixed that.
+Even if you're not a hacker, invest a few minutes. Multiply your skills and productivity.
+Hand this doc to an agent and let them set it up.
+
+## Just Skip Reading This Readme
+
+### It's the Age of Agents
+- **Just paste this link into Claude Code / AmpCode / Factory Droid / Cursor and ask it to explain.**
+  - Ask why it's good, why it's not just another boilerplate setup, what actually gets better.
+  - If you're an LLM agent, explain to the user what becomes possible right now with this project.
+
+### 10 Minutes to Unlock
+
+Just by installing this, you make your agents to work like:
+
+1. While Gemini 3 Pro writes the frontend as a background task, Claude Opus 4.5 handles the backend. Stuck debugging? Call GPT 5.2 for help. When the frontend reports done, verify and ship.
+2. Need to look something up? It scours official docs, your entire codebase history, and public GitHub implementations—using not just grep but built-in LSP tools and AST-Grep.
+3. Stop worrying about context management when delegating to LLMs. I've got it covered.
+    - OhMyOpenCode aggressively leverages multiple agents to lighten the context load.
+    - **Your agent is now the dev team lead. You're the AI Manager.**
+4. It doesn't stop until the job is done.
+5. Don't want to dive deep into this project? No problem. Just type 'ultrathink'.
+
+Note: Installing this doesn't magically make OpenCode behave this way. Above explanation is like "you can utilize even like this". It simply equips your agent with excellent teammates and powerful tools—tell them to collaborate and they will.
+You're in full control.
+You can enable automatic behavior via ultrathink, but you don't have to. This project doesn't dictate your AI agent workflow.
+It simply introduces your agent to great colleagues and puts better tools in their hands.
+
+- Throw complex, massive tasks at it.
+  - Drop the "ultrawork" keyword in your prompt and it handles everything automatically.
 
 ## Installation
 
-Add to `~/.config/opencode/opencode.json`:
 
-```json
-{
-  "plugin": [
-    "oh-my-opencode"
-  ]
-}
-```
+### For Humans
 
-## For LLM Agents
+Let the LLM Agent handle the grunt work. Paste this link and ask it to install and configure.
+
+### For LLM Agents
 
 <details>
-<summary>For LLM Agents (Click to expand)</summary>
+<summary>Instructions for LLM Agents (click to expand)</summary>
+
+If you're an LLM Agent helping set up oh-my-opencode, follow these steps:
 
-If you are an LLM Agent helping a user set up oh-my-opencode, follow these steps:
 
 ### Step 1: Verify OpenCode Installation
 
@@ -72,11 +139,11 @@ if command -v opencode &> /dev/null; then
     echo "OpenCode $(opencode --version) is installed"
 else
     echo "OpenCode is not installed. Please install it first."
-    echo "See: https://opencode.ai/docs"
+    echo "Ref: https://opencode.ai/docs"
 fi
 ```
 
-If OpenCode is not installed, guide the user to [OpenCode Installation Guide](https://opencode.ai/docs).
+If OpenCode isn't installed, check the [OpenCode Installation Guide](https://opencode.ai/docs). To save context, spawn a subagent to handle installation and report back.
 
 ### Step 2: Configure oh-my-opencode Plugin
 
@@ -90,10 +157,10 @@ if [ -f ~/.config/opencode/opencode.json ]; then
             ~/.config/opencode/opencode.json > /tmp/oc.json && \
             mv /tmp/oc.json ~/.config/opencode/opencode.json
     else
-        echo "Add \"oh-my-opencode\" to plugin array in ~/.config/opencode/opencode.json"
+        echo "Please add \"oh-my-opencode\" to the plugin array in ~/.config/opencode/opencode.json"
     fi
 else
-    # New config
+    # Create new config
     echo '{"plugin":["oh-my-opencode"]}' > ~/.config/opencode/opencode.json
 fi
 ```
@@ -101,58 +168,161 @@ fi
 ### Step 3: Verify Setup
 
 ```bash
-opencode --version  # Should be >= 1.0.132
+opencode --version  # Should be 1.0.132 or higher
 cat ~/.config/opencode/opencode.json  # Should contain "oh-my-opencode" in plugin array
 ```
 
 ### Step 4: Configure Authentication
 
-Once `oh-my-opencode` is added to the plugin list, use `tmux` to proceed with the following:
+Once oh-my-opencode is in the plugin list, use tmux to walk the user through authentication. Complete all steps below.
+
+#### 4.1 Anthropic (Claude)
+
+```bash
+opencode auth login
+# Provider: Select Anthropic
+# Login method: Select Claude Pro/Max
+# Guide user through OAuth flow in browser
+# Wait for completion
+# Verify success and confirm with user
+```
+
+#### 4.2 Google Gemini (Antigravity OAuth)
+
+First, enable Google auth in `~/.config/opencode/oh-my-opencode.json`:
+
+```json
+{
+  "google_auth": true
+}
+```
+
+Then authenticate:
+
+```bash
+opencode auth login
+# Provider: Select Google
+# Login method: Select OAuth with Google (Antigravity)
+# Guide user through OAuth flow in browser
+# Wait for completion
+# Verify success and confirm with user
+```
+
+#### 4.3 OpenAI (ChatGPT Plus/Pro)
+
+First, add the opencode-openai-codex-auth plugin:
+
+```json
+{
+  "plugin": [
+    "oh-my-opencode",
+    "opencode-openai-codex-auth@4.1.0"
+  ]
+}
+```
+
+**Important**: The official npm package currently has a bug causing 400 errors (`"No tool call found for function call output with call_id"`). **Use the hotfix branch** until fixed. Edit `~/.config/opencode/package.json`:
+
+```json
+{
+  "dependencies": {
+    "opencode-openai-codex-auth": "code-yeongyu/opencode-openai-codex-auth#fix/orphaned-function-call-output-with-tools"
+  }
+}
+```
+
+Then run `cd ~/.config/opencode && bun i`. In `opencode.json`, use `"opencode-openai-codex-auth"` without the version suffix.
+
+#### 4.3.1 Model Configuration
+You'll also need full model settings in `opencode.json`.
+Read the [opencode-openai-codex-auth documentation](https://github.com/numman-ali/opencode-openai-codex-auth), copy provider/models config from [`config/full-opencode.json`](https://github.com/numman-ali/opencode-openai-codex-auth/blob/main/config/full-opencode.json), and merge carefully to avoid breaking the user's existing setup.
+
+Then authenticate:
+
+```bash
+opencode auth login
+# Provider: Select OpenAI
+# Login method: Select ChatGPT Plus/Pro (Codex Subscription)
+# Guide user through OAuth flow in browser
+# Wait for completion
+# Verify success and confirm with user
+```
 
-1. Open a shell using `tmux`.
-2. Run `opencode auth login` > Select Provider `Anthropic` > Select Login method `Claude Pro/Max` > Instruct the user to authenticate via the OAuth screen.
-3. Recommend the setup: [VibeProxy](https://github.com/automazeio/vibeproxy) for macOS users, or [CLIProxyAPI](https://github.com/router-for-me/CLIProxyAPI) for others.
 
 ### ⚠️ Warning
 
-**Do NOT create or modify the `oh-my-opencode.json` configuration file unless the user explicitly requests it.**
+**Don't change model settings or disable features (agents, hooks, MCPs) unless explicitly requested.**
 
-The plugin works perfectly with default settings without any configuration file. The configuration file is only needed when the user wants customization.
+The plugin works perfectly with defaults. Aside from the recommended `google_auth: true`, don't touch other settings without a specific ask.
 
 </details>
 
-## Why OpenCode & Why Oh My OpenCode
-
-OpenCode is limitlessly extensible and customizable. Zero screen flicker.
-[LSP](https://opencode.ai/docs/lsp/), [linters, formatters](https://opencode.ai/docs/formatters/)? Automatic and fully configurable.
-You can mix and orchestrate models to your exact specifications.
-It is feature-rich. It is elegant. It handles the terminal without hesitation. It is high-performance.
-
-But here is the catch: the learning curve is steep. There is a lot to master. And your time is expensive.
-
-Inspired by [AmpCode](https://ampcode.com) and [Claude Code](https://code.claude.com/docs/en/overview), I have implemented their features here—often with superior execution.
-Because this is OpenCode.
-
-Consider this a superior AmpCode, a superior Claude Code, or simply a specialized distribution.
-
-I believe in the right tool for the job. For your wallet's sake, use CLIProxyAPI or VibeProxy. Employ the best LLMs from frontier labs. You are in command.
-
-**Note**: This setup is highly opinionated. It represents the generic component of my personal configuration, so it evolves constantly. I have spent tokens worth $20,000 just for my personal programming usages, and this plugin represents the apex of that experience. You simply inherit the best. If you have superior ideas, PRs are welcome.
 
 ## Features
 
-### Hooks
+### Agents: Your Teammates
 
-- **Todo Continuation Enforcer**: Forces the agent to complete all tasks before exiting. Eliminates the common LLM issue of "giving up halfway".
-- **Context Window Monitor**: Implements [Context Window Anxiety Management](https://agentic-patterns.com/patterns/context-window-anxiety-management/). When context usage exceeds 70%, it reminds the agent that resources are sufficient, preventing rushed or low-quality output.
-- **Session Notification**: Sends a native OS notification when the job is done (macOS, Linux, Windows).
-- **Session Recovery**: Automatically recovers from API errors, ensuring session stability. Handles four scenarios:
-  - **Tool Result Missing**: When `tool_use` block exists without `tool_result` (ESC interrupt) → injects "cancelled" tool results
-  - **Thinking Block Order**: When thinking block must be first but isn't → prepends empty thinking block
-  - **Thinking Disabled Violation**: When thinking blocks exist but thinking is disabled → strips thinking blocks
-  - **Empty Content Message**: When message has only thinking/meta blocks without actual content → injects "(interrupted)" text via filesystem
-- **Comment Checker**: Detects and reports unnecessary comments after code modifications. Smartly ignores valid patterns (BDD, directives, docstrings, shebangs) to keep the codebase clean from AI-generated artifacts.
-- **Directory AGENTS.md Injector**: Automatically injects `AGENTS.md` contents when reading files. Searches upward from the file's directory to project root, collecting **all** `AGENTS.md` files along the path hierarchy. This enables nested, directory-specific instructions:
+- **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`): Multi-repo analysis, doc lookup, implementation examples. Claude Sonnet 4 is fast, smart, great at tool calls, and excellent for documentation research. Inspired by AmpCode.
+- **explore** (`opencode/grok-code`): Fast codebase exploration and pattern matching. Claude Code uses Haiku; we use Grok—it's free, blazing fast, and plenty smart for file traversal. Inspired by Claude 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-pro-preview`): Technical writing expert. Gemini is a wordsmith—writes prose that flows.
+- **multimodal-looker** (`google/gemini-2.5-flash`): Visual content specialist. Analyzes PDFs, images, diagrams to extract information.
+
+The main agent invokes these automatically, but you can call them explicitly:
+
+```
+Ask @oracle to review this design and propose an architecture
+Ask @librarian how this is implemented—why does the behavior keep changing?
+Ask @explore for the policy on this feature
+```
+
+Customize agent models, prompts, and permissions in `oh-my-opencode.json`. See [Configuration](#configuration).
+
+### Background Agents: Work Like a Team
+
+What if you could run these agents relentlessly, never letting them idle?
+
+- Have GPT debug while Claude tries different approaches to find the root cause
+- Gemini writes the frontend while Claude handles the backend
+- Kick off massive parallel searches, continue implementation on other parts, then finish using the search results
+
+These workflows are possible with OhMyOpenCode.
+
+Run subagents in the background. The main agent gets notified on completion. Wait for results if needed.
+
+**Make your agents work like your team works.**
+
+### The Tools: Your Teammates Deserve Better
+
+#### Why Are You the Only One Using an IDE?
+
+Syntax highlighting, autocomplete, refactoring, navigation, analysis—and now agents writing code...
+
+**Why are you the only one with these tools?**
+**Give them to your agents and watch them level up.**
+
+[OpenCode provides LSP](https://opencode.ai/docs/lsp/), but only for analysis.
+
+The features in your editor? Other agents can't touch them.
+Hand your best tools to your best colleagues. Now they can properly refactor, navigate, and analyze.
+
+- **lsp_hover**: Type info, docs, signatures at position
+- **lsp_goto_definition**: Jump to symbol definition
+- **lsp_find_references**: Find all usages across workspace
+- **lsp_document_symbols**: Get file symbol outline
+- **lsp_workspace_symbols**: Search symbols by name across project
+- **lsp_diagnostics**: Get errors/warnings before build
+- **lsp_servers**: List available LSP servers
+- **lsp_prepare_rename**: Validate rename operation
+- **lsp_rename**: Rename symbol across workspace
+- **lsp_code_actions**: Get available quick fixes/refactorings
+- **lsp_code_action_resolve**: Apply code action
+- **ast_grep_search**: AST-aware code pattern search (25 languages)
+- **ast_grep_replace**: AST-aware code replacement
+
+#### Context Is All You Need
+- **Directory AGENTS.md / README.md Injector**: Auto-injects `AGENTS.md` and `README.md` when reading files. Walks from file directory to project root, collecting **all** `AGENTS.md` files along the path. Supports nested directory-specific instructions:
   ```
   project/
   ├── AGENTS.md              # Project-wide context
@@ -160,89 +330,46 @@ I believe in the right tool for the job. For your wallet's sake, use CLIProxyAPI
   │   ├── AGENTS.md          # src-specific context
   │   └── components/
   │       ├── AGENTS.md      # Component-specific context
-  │       └── Button.tsx     # Reading this injects ALL 3 AGENTS.md files
+  │       └── Button.tsx     # Reading this injects all 3 AGENTS.md files
   ```
-  When reading `Button.tsx`, the hook injects contexts in order: `project/AGENTS.md` → `src/AGENTS.md` → `components/AGENTS.md`. Each directory's context is injected only once per session. Inspired by Claude Code's CLAUDE.md feature.
-- **Directory README.md Injector**: Automatically injects `README.md` contents when reading files. Works identically to the AGENTS.md Injector, searching upward from the file's directory to project root. Provides project documentation context to the LLM agent. Each directory's README is injected only once per session.
-- **Think Mode**: Automatic extended thinking detection and mode switching. Detects when user requests deep thinking (e.g., "think deeply", "ultrathink") and dynamically adjusts model settings for enhanced reasoning.
-- **Anthropic Auto Compact**: Automatically compacts conversation history when approaching context limits for Anthropic models.
-- **Empty Task Response Detector**: Detects when subagent tasks return empty or meaningless responses and handles gracefully.
-- **Grep Output Truncator**: Prevents grep output from overwhelming the context by truncating excessively long results.
+  Reading `Button.tsx` injects in order: `project/AGENTS.md` → `src/AGENTS.md` → `components/AGENTS.md`. Each directory's context is injected once per session.
+- **Conditional Rules Injector**: Not all rules apply all the time. Injects rules from `.claude/rules/` when conditions match.
+  - Walks upward from file directory to project root, plus `~/.claude/rules/` (user).
+  - Supports `.md` and `.mdc` files.
+  - Matches via `globs` field in frontmatter.
+  - `alwaysApply: true` for rules that should always fire.
+  - Example rule file:
+    ```markdown
+    ---
+    globs: ["*.ts", "src/**/*.js"]
+    description: "TypeScript/JavaScript coding rules"
+    ---
+    - Use PascalCase for interface names
+    - Use camelCase for function names
+    ```
+- **Online**: Project rules aren't everything. Built-in MCPs for extended capabilities:
+  - **context7**: Official documentation lookup
+  - **websearch_exa**: Real-time web search
+  - **grep_app**: Ultra-fast code search across public GitHub repos (great for finding implementation examples)
 
-### Agents
-- **oracle** (`openai/gpt-5.1`): The architect. Expert in code reviews and strategy. Uses GPT-5.1 for its unmatched logic and reasoning capabilities. Inspired by AmpCode.
-- **librarian** (`anthropic/claude-haiku-4-5`): Multi-repo analysis, documentation lookup, and implementation examples. Haiku is chosen for its speed, competence, excellent tool usage, and cost-efficiency. Inspired by AmpCode.
-- **explore** (`opencode/grok-code`): Fast exploration and pattern matching. Claude Code uses Haiku; we use Grok. It is currently free, blazing fast, and intelligent enough for file traversal. Inspired by Claude Code.
-- **frontend-ui-ux-engineer** (`google/gemini-3-pro-preview`): A designer turned developer. Creates stunning UIs. Uses Gemini because its creativity and UI code generation are superior.
-- **document-writer** (`google/gemini-3-pro-preview`): A technical writing expert. Gemini is a wordsmith; it writes prose that flows naturally.
+#### Be Multimodal. Save Tokens.
 
-Each agent is automatically invoked by the main agent, but you can also explicitly request them:
+The look_at tool from AmpCode, now in OhMyOpenCode.
+Instead of the agent reading massive files and bloating context, it internally leverages another agent to extract just what it needs.
 
-```
-@oracle Please think through the design of this part and suggest an architecture.
-@librarian Tell me how this is implemented — why does the behavior keep changing internally?
-@explore Tell me about the policy for this feature.
-```
+#### I Removed Their Blockers
+- Replaces built-in grep and glob tools. Default implementation has no timeout—can hang forever.
 
-Agent models, prompts, and permissions can be customized via `oh-my-opencode.json`. See [Configuration](#configuration) for details.
 
-### Tools
+### Goodbye Claude Code. Hello Oh My OpenCode.
 
-#### Built-in LSP Tools
-
-The features you use in your editor—other agents cannot access them. Oh My OpenCode hands those very tools to your LLM Agent. Refactoring, navigation, and analysis are all supported using the same OpenCode configuration.
-
-[OpenCode provides LSP](https://opencode.ai/docs/lsp/), but only for analysis. Oh My OpenCode equips you with navigation and refactoring tools matching the same specification.
-
-- **lsp_hover**: Get type info, docs, signatures at position
-- **lsp_goto_definition**: Jump to symbol definition
-- **lsp_find_references**: Find all usages across workspace
-- **lsp_document_symbols**: Get file's symbol outline
-- **lsp_workspace_symbols**: Search symbols by name across project
-- **lsp_diagnostics**: Get errors/warnings before build
-- **lsp_servers**: List available LSP servers
-- **lsp_prepare_rename**: Validate rename operation
-- **lsp_rename**: Rename symbol across workspace
-- **lsp_code_actions**: Get available quick fixes/refactorings
-- **lsp_code_action_resolve**: Apply a code action
-
-#### Built-in AST-Grep Tools
-
-- **ast_grep_search**: AST-aware code pattern search (25 languages)
-- **ast_grep_replace**: AST-aware code replacement
-
-#### Grep
-
-- **grep**: Content search with safety limits (5min timeout, 10MB output). Overrides OpenCode's built-in `grep` tool.
-  - The default `grep` lacks safeguards. On a large codebase, a broad pattern can cause CPU overload and indefinite hanging.
-  - This tool enforces strict limits and completely replaces the built-in `grep`.
-
-#### Glob
-
-- **glob**: File pattern matching with timeout protection (60s). Overrides OpenCode's built-in `glob` tool.
-  - The default `glob` lacks timeout. If ripgrep hangs, it waits indefinitely.
-  - This tool enforces timeouts and kills the process on expiration.
-
-#### Built-in MCPs
-
-- **websearch_exa**: Exa AI web search. Performs real-time web searches and can scrape content from specific URLs. Returns LLM-optimized context from relevant websites.
-- **context7**: Library documentation lookup. Fetches up-to-date documentation for any library to assist with accurate coding.
-
-Don't need these? Disable them via `oh-my-opencode.json`:
-
-```json
-{
-  "disabled_mcps": ["websearch_exa"]
-}
-```
-
-### Claude Code Compatibility
-
-Oh My OpenCode provides seamless Claude Code configuration compatibility. If you've been using Claude Code, your existing setup works out of the box.
+Oh My OpenCode has a Claude Code compatibility layer.
+If you were using Claude Code, your existing config just works.
 
 #### Hooks Integration
 
-Execute custom scripts via Claude Code's `settings.json` hook system. Oh My OpenCode reads and executes hooks defined in:
+Run custom scripts via Claude Code's `settings.json` hook system.
+Oh My OpenCode reads and executes hooks from:
 
 - `~/.claude/settings.json` (user)
 - `./.claude/settings.json` (project)
@@ -251,7 +378,7 @@ Execute custom scripts via Claude Code's `settings.json` hook system. Oh My Open
 Supported hook events:
 - **PreToolUse**: Runs before tool execution. Can block or modify tool input.
 - **PostToolUse**: Runs after tool execution. Can add warnings or context.
-- **UserPromptSubmit**: Runs when user submits a prompt. Can block or inject messages.
+- **UserPromptSubmit**: Runs when user submits prompt. Can block or inject messages.
 - **Stop**: Runs when session goes idle. Can inject follow-up prompts.
 
 Example `settings.json`:
@@ -268,7 +395,7 @@ Example `settings.json`:
 }
 ```
 
-#### Configuration Loaders
+#### Config Loaders
 
 **Command Loader**: Loads markdown-based slash commands from 4 directories:
 - `~/.claude/commands/` (user)
@@ -284,7 +411,7 @@ Example `settings.json`:
 - `~/.claude/agents/*.md` (user)
 - `./.claude/agents/*.md` (project)
 
-**MCP Loader**: Loads MCP server configurations from `.mcp.json` files:
+**MCP Loader**: Loads MCP server configs from `.mcp.json` files:
 - `~/.claude/.mcp.json` (user)
 - `./.mcp.json` (project)
 - `./.claude/.mcp.json` (local)
@@ -292,24 +419,57 @@ Example `settings.json`:
 
 #### Data Storage
 
-**Todo Management**: Session todos are stored in Claude Code compatible format at `~/.claude/todos/`.
+**Todo Management**: Session todos stored in `~/.claude/todos/` in Claude Code compatible format.
 
-**Transcript**: Session activity is logged to `~/.claude/transcripts/` in JSONL format, enabling replay and analysis.
+**Transcript**: Session activity logged to `~/.claude/transcripts/` in JSONL format for replay and analysis.
 
-> **Note on `claude-code-*` naming**: Features under `src/features/claude-code-*/` are migrated from Claude Code's configuration system. This naming convention clearly identifies which features originated from Claude Code.
+#### Compatibility Toggles
 
-### Other Features
+Disable specific Claude Code compatibility features with the `claude_code` config object:
 
-- **Terminal Title**: Auto-updates terminal title with session status (idle ○, processing ◐, tool ⚡, error ✖). Supports tmux.
-- **Session State**: Centralized session tracking module used by event hooks and terminal title updates.
+```json
+{
+  "claude_code": {
+    "mcp": false,
+    "commands": false,
+    "skills": false,
+    "agents": false,
+    "hooks": false
+  }
+}
+```
+
+| Toggle     | When `false`, stops loading from...                                                   | Unaffected                                            |
+| ---------- | ------------------------------------------------------------------------------------- | ----------------------------------------------------- |
+| `mcp`      | `~/.claude/.mcp.json`, `./.mcp.json`, `./.claude/.mcp.json`                           | Built-in MCP (context7, websearch_exa)                |
+| `commands` | `~/.claude/commands/*.md`, `./.claude/commands/*.md`                                  | `~/.config/opencode/command/`, `./.opencode/command/` |
+| `skills`   | `~/.claude/skills/*/SKILL.md`, `./.claude/skills/*/SKILL.md`                          | -                                                     |
+| `agents`   | `~/.claude/agents/*.md`, `./.claude/agents/*.md`                                      | Built-in agents (oracle, librarian, etc.)             |
+| `hooks`    | `~/.claude/settings.json`, `./.claude/settings.json`, `./.claude/settings.local.json` | -                                                     |
+
+All toggles default to `true` (enabled). Omit the `claude_code` object for full Claude Code compatibility.
+
+### Not Just for the Agents
+
+When agents thrive, you thrive. But I want to help you directly too.
+
+- **Ultrawork Mode**: Type "ultrawork" or "ulw" and agent orchestration kicks in. Forces the main agent to max out all available specialists (explore, librarian, plan, UI) via background tasks in parallel, with strict TODO tracking and verification. Just ultrawork. Everything fires at full capacity.
+- **Todo Continuation Enforcer**: Makes agents finish all TODOs before stopping. Kills the chronic LLM habit of quitting halfway.
+- **Comment Checker**: LLMs love comments. Too many comments. This reminds them to cut the noise. Smartly ignores valid patterns (BDD, directives, docstrings) and demands justification for the rest. Clean code wins.
+- **Think Mode**: Auto-detects when extended thinking is needed and switches modes. Catches phrases like "think deeply" or "ultrathink" and dynamically adjusts model settings for maximum reasoning.
+- **Context Window Monitor**: Implements [Context Window Anxiety Management](https://agentic-patterns.com/patterns/context-window-anxiety-management/).
+  - At 70%+ usage, reminds agents there's still headroom—prevents rushed, sloppy work.
+- Stability features that felt missing in OpenCode are built in. The Claude Code experience, transplanted. Sessions don't crash mid-run. Even if they do, they recover.
 
 ## Configuration
 
-Configuration file locations (in priority order):
+Highly opinionated, but adjustable to taste.
+
+Config file locations (priority order):
 1. `.opencode/oh-my-opencode.json` (project)
 2. `~/.config/opencode/oh-my-opencode.json` (user)
 
-Schema autocomplete is supported:
+Schema autocomplete supported:
 
 ```json
 {
@@ -317,6 +477,18 @@ Schema autocomplete is supported:
 }
 ```
 
+### Google Auth
+
+Enable built-in Antigravity OAuth for Google Gemini models:
+
+```json
+{
+  "google_auth": true
+}
+```
+
+When enabled, `opencode auth login` shows "OAuth with Google (Antigravity)" for the Google provider.
+
 ### Agents
 
 Override built-in agent settings:
@@ -337,7 +509,7 @@ Override built-in agent settings:
 
 Each agent supports: `model`, `temperature`, `top_p`, `prompt`, `tools`, `disable`, `description`, `mode`, `color`, `permission`.
 
-Or you can disable them using `disabled_agents` in `~/.config/opencode/oh-my-opencode.json` or `.opencode/oh-my-opencode.json`:
+Or disable via `disabled_agents` in `~/.config/opencode/oh-my-opencode.json` or `.opencode/oh-my-opencode.json`:
 
 ```json
 {
@@ -349,23 +521,27 @@ Available agents: `oracle`, `librarian`, `explore`, `frontend-ui-ux-engineer`, `
 
 ### MCPs
 
-By default, Context7 and Exa MCP are supported.
+Context7, Exa, and grep.app MCP enabled by default.
 
-If you don't want these, you can disable them using `disabled_mcps` in `~/.config/opencode/oh-my-opencode.json` or `.opencode/oh-my-opencode.json`:
+- **context7**: Fetches up-to-date official documentation for libraries
+- **websearch_exa**: Real-time web search powered by Exa AI
+- **grep_app**: Ultra-fast code search across millions of public GitHub repositories via [grep.app](https://grep.app)
+
+Don't want them? Disable via `disabled_mcps` in `~/.config/opencode/oh-my-opencode.json` or `.opencode/oh-my-opencode.json`:
 
 ```json
 {
-  "disabled_mcps": ["context7", "websearch_exa"]
+  "disabled_mcps": ["context7", "websearch_exa", "grep_app"]
 }
 ```
 
 ### LSP
 
 OpenCode provides LSP tools for analysis.
-Oh My OpenCode provides LSP tools for refactoring (rename, code actions).
-It supports all LSP configurations and custom settings supported by OpenCode (those configured in opencode.json), and you can also configure additional settings specifically for Oh My OpenCode as shown below.
+Oh My OpenCode adds refactoring tools (rename, code actions).
+All OpenCode LSP configs and custom settings (from opencode.json) are supported, plus additional Oh My OpenCode-specific settings.
 
-You can configure additional LSP servers via the `lsp` option in `~/.config/opencode/oh-my-opencode.json` or `.opencode/oh-my-opencode.json`:
+Add LSP servers via the `lsp` option in `~/.config/opencode/oh-my-opencode.json` or `.opencode/oh-my-opencode.json`:
 
 ```json
 {
@@ -384,29 +560,48 @@ You can configure additional LSP servers via the `lsp` option in `~/.config/open
 
 Each server supports: `command`, `extensions`, `priority`, `env`, `initialization`, `disabled`.
 
+
 ## Author's Note
 
-Install Oh My OpenCode. Do not waste time configuring OpenCode from scratch.
-I have resolved the friction so you don't have to. The answers are in this plugin. If OpenCode is Arch Linux, Oh My OpenCode is [Omarchy](https://omarchy.org/).
+Install Oh My OpenCode.
 
-Enjoy the multi-model stability and rich feature set that other harnesses promise but fail to deliver.
-I will continue testing and updating here. I am the primary user of this project.
+I've used LLMs worth $24,000 tokens purely for personal development.
+Tried every tool out there, configured them to death. OpenCode won.
 
-- Who possesses the best raw logic?
-- Who is the debugging god?
+The answers to every problem I hit are baked into this plugin. Just install and go.
+If OpenCode is Debian/Arch, Oh My OpenCode is Ubuntu/[Omarchy](https://omarchy.org/).
+
+
+Heavily influenced by [AmpCode](https://ampcode.com) and [Claude Code](https://code.claude.com/docs/overview)—I've ported their features here, often improved. And I'm still building.
+It's **Open**Code, after all.
+
+Enjoy multi-model orchestration, stability, and rich features that other harnesses promise but can't deliver.
+I'll keep testing and updating. I'm this project's most obsessive user.
+- Which model has the sharpest logic?
+- Who's the debugging god?
 - Who writes the best prose?
 - Who dominates frontend?
 - Who owns backend?
 - Which model is fastest for daily driving?
 - What new features are other harnesses shipping?
 
-Do not overthink it. I have done the thinking. I will integrate the best practices. I will update this.
-If this sounds arrogant and you have a superior solution, send a PR. You are welcome.
+This plugin is the distillation of that experience. Just take the best. Got a better idea? PRs are welcome.
 
-As of now, I have no affiliation with any of the projects or models mentioned here. This plugin is purely based on personal experimentation and preference.
+**Stop agonizing over agent harness choices.**
+**I'll do the research, borrow from the best, and ship updates here.**
+
+If this sounds arrogant and you have a better answer, please contribute. You're welcome.
+
+I have no affiliation with any project or model mentioned here. This is purely personal experimentation and preference.
+
+99% of this project was built using OpenCode. I tested for functionality—I don't really know how to write proper TypeScript. **But I personally reviewed and largely rewrote this doc, so read with confidence.**
 
-I constructed 99% of this project using OpenCode. I focused on functional verification, and honestly, I don't know how to write proper TypeScript. **But I personally reviewed and comprehensively rewritten this documentation, so you can rely on it with confidence.**
 ## Warnings
 
-- If you are on [1.0.132](https://github.com/sst/opencode/releases/tag/v1.0.132) or lower, OpenCode has a bug that might break config.
-  - [The fix](https://github.com/sst/opencode/pull/5040) was merged after 1.0.132, so use a newer version.
+- Productivity might spike too hard. Don't let your coworker notice.
+  - Actually, I'll spread the word. Let's see who wins.
+- If you're on [1.0.132](https://github.com/sst/opencode/releases/tag/v1.0.132) or older, an OpenCode bug may break config.
+  - [The fix](https://github.com/sst/opencode/pull/5040) was merged after 1.0.132—use a newer version.
+    - Fun fact: That PR was discovered and fixed thanks to OhMyOpenCode's Librarian, Explore, and Oracle setup.
+
+*Special thanks to @junhoyeo for this amazing hero image.*

assets/oh-my-opencode.schema.json

@@ -27,7 +27,31 @@
           "librarian",
           "explore",
           "frontend-ui-ux-engineer",
-          "document-writer"
+          "document-writer",
+          "multimodal-looker"
+        ]
+      }
+    },
+    "disabled_hooks": {
+      "type": "array",
+      "description": "List of built-in hooks to disable. Useful for selectively disabling hooks that may conflict with your workflow.",
+      "items": {
+        "type": "string",
+        "enum": [
+          "todo-continuation-enforcer",
+          "context-window-monitor",
+          "session-recovery",
+          "session-notification",
+          "comment-checker",
+          "grep-output-truncator",
+          "directory-agents-injector",
+          "directory-readme-injector",
+          "empty-task-response-detector",
+          "think-mode",
+          "anthropic-auto-compact",
+          "rules-injector",
+          "background-notification",
+          "auto-update-checker"
         ]
       }
     },
@@ -40,7 +64,8 @@
           "librarian",
           "explore",
           "frontend-ui-ux-engineer",
-          "document-writer"
+          "document-writer",
+          "multimodal-looker"
         ]
       },
       "additionalProperties": {
@@ -154,6 +179,73 @@
           }
         }
       }
+    },
+    "google_auth": {
+      "type": "boolean",
+      "description": "Enable built-in Antigravity OAuth for Google Gemini models. When true, adds 'OAuth with Google (Antigravity)' login option.",
+      "default": false
+    },
+    "lsp": {
+      "type": "object",
+      "description": "Additional LSP server configurations specific to Oh My OpenCode.",
+      "additionalProperties": {
+        "type": "object",
+        "properties": {
+          "command": {
+            "type": "array",
+            "items": { "type": "string" },
+            "description": "Command and arguments to start the LSP server"
+          },
+          "extensions": {
+            "type": "array",
+            "items": { "type": "string" },
+            "description": "File extensions this server handles (e.g., [\".ts\", \".tsx\"])"
+          },
+          "priority": {
+            "type": "number",
+            "description": "Server priority (higher = preferred)"
+          },
+          "env": {
+            "type": "object",
+            "additionalProperties": { "type": "string" },
+            "description": "Environment variables for the LSP server"
+          },
+          "initialization": {
+            "type": "object",
+            "description": "Custom initialization options"
+          },
+          "disabled": {
+            "type": "boolean",
+            "description": "Disable this LSP server"
+          }
+        }
+      }
+    },
+    "claude_code": {
+      "type": "object",
+      "description": "Toggle Claude Code compatibility features on/off. All default to true (enabled).",
+      "properties": {
+        "mcp": {
+          "type": "boolean",
+          "description": "Load MCP servers from ~/.claude/.mcp.json, ./.mcp.json, ./.claude/.mcp.json"
+        },
+        "commands": {
+          "type": "boolean",
+          "description": "Load commands from ~/.claude/commands/*.md, ./.claude/commands/*.md"
+        },
+        "skills": {
+          "type": "boolean",
+          "description": "Load skills from ~/.claude/skills/*/SKILL.md, ./.claude/skills/*/SKILL.md"
+        },
+        "agents": {
+          "type": "boolean",
+          "description": "Load agents from ~/.claude/agents/*.md, ./.claude/agents/*.md"
+        },
+        "hooks": {
+          "type": "boolean",
+          "description": "Execute hooks from ~/.claude/settings.json, ./.claude/settings.json, ./.claude/settings.local.json"
+        }
+      }
     }
   }
 }

bun.lock

@@ -7,12 +7,16 @@
       "dependencies": {
         "@ast-grep/cli": "^0.40.0",
         "@ast-grep/napi": "^0.40.0",
-        "@code-yeongyu/comment-checker": "^0.4.4",
-        "@opencode-ai/plugin": "^1.0.7",
+        "@code-yeongyu/comment-checker": "^0.5.0",
+        "@openauthjs/openauth": "^0.4.3",
+        "@opencode-ai/plugin": "^1.0.150",
+        "hono": "^4.10.4",
+        "picomatch": "^4.0.2",
         "xdg-basedir": "^5.1.0",
         "zod": "^4.1.8",
       },
       "devDependencies": {
+        "@types/picomatch": "^3.0.2",
         "bun-types": "latest",
         "oh-my-opencode": "^0.1.30",
         "typescript": "^5.7.3",
@@ -64,11 +68,23 @@
 
     "@ast-grep/napi-win32-x64-msvc": ["@ast-grep/napi-win32-x64-msvc@0.40.0", "", { "os": "win32", "cpu": "x64" }, "sha512-Hk2IwfPqMFGZt5SRxsoWmGLxBXxprow4LRp1eG6V8EEiJCNHxZ9ZiEaIc5bNvMDBjHVSnqZAXT22dROhrcSKQg=="],
 
-    "@code-yeongyu/comment-checker": ["@code-yeongyu/comment-checker@0.4.4", "", { "os": [ "linux", "win32", "darwin", ], "cpu": [ "x64", "arm64", ], "bin": { "comment-checker": "bin/comment-checker" } }, "sha512-vsbdLMQYJJNDV/baTDnNqqg/MZwA+9nz7TE6Mybj8zjZVTCn4ZivH4hAdD5p4fLxhGZEJ5x1UDmXA6pAGA7lHA=="],
+    "@code-yeongyu/comment-checker": ["@code-yeongyu/comment-checker@0.5.0", "", { "os": [ "linux", "win32", "darwin", ], "cpu": [ "x64", "arm64", ], "bin": { "comment-checker": "bin/comment-checker" } }, "sha512-rKD2qQnTVUacsVQtpu3I5Sxi09X/XpOwS9fcmbUv1yfUL6llraaPuLmmxMBMRcmm7Zu31yEPVKCeUkVODfRL1g=="],
 
-    "@opencode-ai/plugin": ["@opencode-ai/plugin@1.0.128", "", { "dependencies": { "@opencode-ai/sdk": "1.0.128", "zod": "4.1.8" } }, "sha512-M5vjz3I6KeoBSNduWmT5iHXRtTLCqICM5ocs+WrB3uxVorslcO3HVwcLzrERh/ntpxJ/1xhnHQaeG6Mg+P744A=="],
+    "@openauthjs/openauth": ["@openauthjs/openauth@0.4.3", "", { "dependencies": { "@standard-schema/spec": "1.0.0-beta.3", "aws4fetch": "1.0.20", "jose": "5.9.6" }, "peerDependencies": { "arctic": "^2.2.2", "hono": "^4.0.0" } }, "sha512-RlnjqvHzqcbFVymEwhlUEuac4utA5h4nhSK/i2szZuQmxTIqbGUxZ+nM+avM+VV4Ing+/ZaNLKILoXS3yrkOOw=="],
 
-    "@opencode-ai/sdk": ["@opencode-ai/sdk@1.0.128", "", {}, "sha512-Kow3Ivg8bR8dNRp8C0LwF9e8+woIrwFgw3ZALycwCfqS/UujDkJiBeYHdr1l/07GSHP9sZPmvJ6POuvfZ923EA=="],
+    "@opencode-ai/plugin": ["@opencode-ai/plugin@1.0.150", "", { "dependencies": { "@opencode-ai/sdk": "1.0.150", "zod": "4.1.8" } }, "sha512-XmY3yydk120GBv2KeLxSZlElFx4Zx9TYLa3bS9X1TxXot42UeoMLEi3Xa46yboYnWwp4bC9Fu+Gd1E7hypG8Jw=="],
+
+    "@opencode-ai/sdk": ["@opencode-ai/sdk@1.0.150", "", {}, "sha512-Nz9Di8UD/GK01w3N+jpiGNB733pYkNY8RNLbuE/HUxEGSP5apbXBY0IdhbW7859sXZZK38kF1NqOx4UxwBf4Bw=="],
+
+    "@oslojs/asn1": ["@oslojs/asn1@1.0.0", "", { "dependencies": { "@oslojs/binary": "1.0.0" } }, "sha512-zw/wn0sj0j0QKbIXfIlnEcTviaCzYOY3V5rAyjR6YtOByFtJiT574+8p9Wlach0lZH9fddD4yb9laEAIl4vXQA=="],
+
+    "@oslojs/binary": ["@oslojs/binary@1.0.0", "", {}, "sha512-9RCU6OwXU6p67H4NODbuxv2S3eenuQ4/WFLrsq+K/k682xrznH5EVWA7N4VFk9VYVcbFtKqur5YQQZc0ySGhsQ=="],
+
+    "@oslojs/crypto": ["@oslojs/crypto@1.0.1", "", { "dependencies": { "@oslojs/asn1": "1.0.0", "@oslojs/binary": "1.0.0" } }, "sha512-7n08G8nWjAr/Yu3vu9zzrd0L9XnrJfpMioQcvCMxBIiF5orECHe5/3J0jmXRVvgfqMm/+4oxlQ+Sq39COYLcNQ=="],
+
+    "@oslojs/encoding": ["@oslojs/encoding@1.1.0", "", {}, "sha512-70wQhgYmndg4GCPxPPxPGevRKqTIJ2Nh4OkiMWmDAVYsTQ+Ta7Sq+rPevXyXGdzr30/qZBnyOalCszoMxlyldQ=="],
+
+    "@oslojs/jwt": ["@oslojs/jwt@0.2.0", "", { "dependencies": { "@oslojs/encoding": "0.4.1" } }, "sha512-bLE7BtHrURedCn4Mco3ma9L4Y1GR2SMBuIvjWr7rmQ4/W/4Jy70TIAgZ+0nIlk0xHz1vNP8x8DCns45Sb2XRbg=="],
 
     "@oven/bun-darwin-aarch64": ["@oven/bun-darwin-aarch64@1.3.3", "", { "os": "darwin", "cpu": "arm64" }, "sha512-eJopQrUk0WR7jViYDC29+Rp50xGvs4GtWOXBeqCoFMzutkkO3CZvHehA4JqnjfWMTSS8toqvRhCSOpOz62Wf9w=="],
 
@@ -92,16 +108,30 @@
 
     "@oven/bun-windows-x64-baseline": ["@oven/bun-windows-x64-baseline@1.3.3", "", { "os": "win32", "cpu": "x64" }, "sha512-u5eZHKq6TPJSE282KyBOicGQ2trkFml0RoUfqkPOJVo7TXGrsGYYzdsugZRnVQY/WEmnxGtBy4T3PAaPqgQViA=="],
 
+    "@standard-schema/spec": ["@standard-schema/spec@1.0.0-beta.3", "", {}, "sha512-0ifF3BjA1E8SY9C+nUew8RefNOIq0cDlYALPty4rhUm8Rrl6tCM8hBT4bhGhx7I7iXD0uAgt50lgo8dD73ACMw=="],
+
     "@types/node": ["@types/node@24.10.1", "", { "dependencies": { "undici-types": "~7.16.0" } }, "sha512-GNWcUTRBgIRJD5zj+Tq0fKOJ5XZajIiBroOF0yvj2bSU1WvNdYS/dn9UxwsujGW4JX06dnHyjV2y9rRaybH0iQ=="],
 
+    "@types/picomatch": ["@types/picomatch@3.0.2", "", {}, "sha512-n0i8TD3UDB7paoMMxA3Y65vUncFJXjcUf7lQY7YyKGl6031FNjfsLs6pdLFCy2GNFxItPJG8GvvpbZc2skH7WA=="],
+
+    "arctic": ["arctic@2.3.4", "", { "dependencies": { "@oslojs/crypto": "1.0.1", "@oslojs/encoding": "1.1.0", "@oslojs/jwt": "0.2.0" } }, "sha512-+p30BOWsctZp+CVYCt7oAean/hWGW42sH5LAcRQX56ttEkFJWbzXBhmSpibbzwSJkRrotmsA+oAoJoVsU0f5xA=="],
+
+    "aws4fetch": ["aws4fetch@1.0.20", "", {}, "sha512-/djoAN709iY65ETD6LKCtyyEI04XIBP5xVvfmNxsEP0uJB5tyaGBztSryRr4HqMStr9R06PisQE7m9zDTXKu6g=="],
+
     "bun": ["bun@1.3.3", "", { "optionalDependencies": { "@oven/bun-darwin-aarch64": "1.3.3", "@oven/bun-darwin-x64": "1.3.3", "@oven/bun-darwin-x64-baseline": "1.3.3", "@oven/bun-linux-aarch64": "1.3.3", "@oven/bun-linux-aarch64-musl": "1.3.3", "@oven/bun-linux-x64": "1.3.3", "@oven/bun-linux-x64-baseline": "1.3.3", "@oven/bun-linux-x64-musl": "1.3.3", "@oven/bun-linux-x64-musl-baseline": "1.3.3", "@oven/bun-windows-x64": "1.3.3", "@oven/bun-windows-x64-baseline": "1.3.3" }, "os": [ "linux", "win32", "darwin", ], "cpu": [ "x64", "arm64", ], "bin": { "bun": "bin/bun.exe", "bunx": "bin/bunx.exe" } }, "sha512-2hJ4ocTZ634/Ptph4lysvO+LbbRZq8fzRvMwX0/CqaLBxrF2UB5D1LdMB8qGcdtCer4/VR9Bx5ORub0yn+yzmw=="],
 
     "bun-types": ["bun-types@1.3.3", "", { "dependencies": { "@types/node": "*" } }, "sha512-z3Xwlg7j2l9JY27x5Qn3Wlyos8YAp0kKRlrePAOjgjMGS5IG6E7Jnlx736vH9UVI4wUICwwhC9anYL++XeOgTQ=="],
 
     "detect-libc": ["detect-libc@2.1.2", "", {}, "sha512-Btj2BOOO83o3WyH59e8MgXsxEQVcarkUOpEYrubB0urwnN10yQ364rsiByU11nZlqWYZm05i/of7io4mzihBtQ=="],
 
+    "hono": ["hono@4.10.8", "", {}, "sha512-DDT0A0r6wzhe8zCGoYOmMeuGu3dyTAE40HHjwUsWFTEy5WxK1x2WDSsBPlEXgPbRIFY6miDualuUDbasPogIww=="],
+
+    "jose": ["jose@5.9.6", "", {}, "sha512-AMlnetc9+CV9asI19zHmrgS/WYsWUwCn2R7RzlbJWD7F9eWYUTGyBmU9o6PxngtLGOiDGPRu+Uc4fhKzbpteZQ=="],
+
     "oh-my-opencode": ["oh-my-opencode@0.1.30", "", { "dependencies": { "@ast-grep/cli": "^0.40.0", "@ast-grep/napi": "^0.40.0", "@code-yeongyu/comment-checker": "^0.4.1", "@opencode-ai/plugin": "^1.0.7", "xdg-basedir": "^5.1.0", "zod": "^4.1.8" }, "peerDependencies": { "bun": ">=1.0.0" } }, "sha512-pXGGgL/7Jcz3yuGJJTI72BKern2egwfRz2LQZTBq+jl+pNCybOvGvXtFmR+WGlF8O3ZjL1wIHypBbIVuHOBzxg=="],
 
+    "picomatch": ["picomatch@4.0.3", "", {}, "sha512-5gTmgEY/sqK6gFXLIsQNH19lWb4ebPDLA4SdLP7dsWkIXHWlG66oPuVvXSGFPppYZz8ZDZq0dYYrbHfBCVUb1Q=="],
+
     "typescript": ["typescript@5.9.3", "", { "bin": { "tsc": "bin/tsc", "tsserver": "bin/tsserver" } }, "sha512-jl1vZzPDinLr9eUt3J/t7V6FgNEw9QjvBPdysz9KfQDD41fQrC2Y4vKQdiaUpFT4bXlb1RHhLpp8wtm6M5TgSw=="],
 
     "undici-types": ["undici-types@7.16.0", "", {}, "sha512-Zz+aZWSj8LE6zoxD+xrjh4VfkIG8Ya6LvYkZqtUQGJPZjYl53ypCaUwWqo7eI0x66KBGeRo+mlBEkMSeSZ38Nw=="],
@@ -110,6 +140,12 @@
 
     "zod": ["zod@4.1.8", "", {}, "sha512-5R1P+WwQqmmMIEACyzSvo4JXHY5WiAFHRMg+zBZKgKS+Q1viRa0C1hmUKtHltoIFKtIdki3pRxkmpP74jnNYHQ=="],
 
+    "@oslojs/jwt/@oslojs/encoding": ["@oslojs/encoding@0.4.1", "", {}, "sha512-hkjo6MuIK/kQR5CrGNdAPZhS01ZCXuWDRJ187zh6qqF2+yMHZpD9fAYpX8q2bOO6Ryhl3XpCT6kUX76N8hhm4Q=="],
+
     "oh-my-opencode/@code-yeongyu/comment-checker": ["@code-yeongyu/comment-checker@0.4.1", "", { "os": [ "linux", "win32", "darwin", ], "cpu": [ "x64", "arm64", ], "bin": { "comment-checker": "bin/comment-checker" } }, "sha512-E7p1V8CsRj9hMbwENd9BfxZGWYu+lKS5tXGuNNcNtkRMhWvwM/ononysKpLB7LXdxfSYAn0j7heJydyzEmm+lg=="],
+
+    "oh-my-opencode/@opencode-ai/plugin": ["@opencode-ai/plugin@1.0.128", "", { "dependencies": { "@opencode-ai/sdk": "1.0.128", "zod": "4.1.8" } }, "sha512-M5vjz3I6KeoBSNduWmT5iHXRtTLCqICM5ocs+WrB3uxVorslcO3HVwcLzrERh/ntpxJ/1xhnHQaeG6Mg+P744A=="],
+
+    "oh-my-opencode/@opencode-ai/plugin/@opencode-ai/sdk": ["@opencode-ai/sdk@1.0.128", "", {}, "sha512-Kow3Ivg8bR8dNRp8C0LwF9e8+woIrwFgw3ZALycwCfqS/UujDkJiBeYHdr1l/07GSHP9sZPmvJ6POuvfZ923EA=="],
   }
 }

package.json

@@ -1,6 +1,6 @@
 {
   "name": "oh-my-opencode",
-  "version": "0.3.2",
+  "version": "1.1.7",
   "description": "OpenCode plugin - custom agents (oracle, librarian) and enhanced features",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",
@@ -13,10 +13,14 @@
       "types": "./dist/index.d.ts",
       "import": "./dist/index.js"
     },
+    "./google-auth": {
+      "types": "./dist/google-auth.d.ts",
+      "import": "./dist/google-auth.js"
+    },
     "./schema.json": "./dist/oh-my-opencode.schema.json"
   },
   "scripts": {
-    "build": "bun build src/index.ts --outdir dist --target bun --format esm --external @ast-grep/napi && tsc --emitDeclarationOnly && bun run build:schema",
+    "build": "bun build src/index.ts src/google-auth.ts --outdir dist --target bun --format esm --external @ast-grep/napi && tsc --emitDeclarationOnly && bun run build:schema",
     "build:schema": "bun run script/build-schema.ts",
     "clean": "rm -rf dist",
     "prepublishOnly": "bun run clean && bun run build",
@@ -45,11 +49,15 @@
     "@ast-grep/cli": "^0.40.0",
     "@ast-grep/napi": "^0.40.0",
     "@code-yeongyu/comment-checker": "^0.5.0",
-    "@opencode-ai/plugin": "^1.0.7",
+    "@opencode-ai/plugin": "^1.0.150",
+    "@openauthjs/openauth": "^0.4.3",
+    "hono": "^4.10.4",
+    "picomatch": "^4.0.2",
     "xdg-basedir": "^5.1.0",
     "zod": "^4.1.8"
   },
   "devDependencies": {
+    "@types/picomatch": "^3.0.2",
     "bun-types": "latest",
     "oh-my-opencode": "^0.1.30",
     "typescript": "^5.7.3"

script/publish.ts

@@ -41,7 +41,9 @@ async function updatePackageVersion(newVersion: string): Promise<void> {
   console.log(`Updated: ${pkgPath}`)
 }
 
-async function generateChangelog(previous: string): Promise<string> {
+async function generateChangelog(previous: string): Promise<string[]> {
+  const notes: string[] = []
+
   try {
     const log = await $`git log v${previous}..HEAD --oneline --format="%h %s"`.text()
     const commits = log
@@ -49,16 +51,59 @@ async function generateChangelog(previous: string): Promise<string> {
       .filter((line) => line && !line.match(/^\w+ (ignore:|test:|chore:|ci:|release:)/i))
 
     if (commits.length > 0) {
-      const changelog = commits.map((c) => `- ${c}`).join("\n")
+      for (const commit of commits) {
+        notes.push(`- ${commit}`)
+      }
       console.log("\n--- Changelog ---")
-      console.log(changelog)
+      console.log(notes.join("\n"))
       console.log("-----------------\n")
-      return changelog
     }
   } catch {
     console.log("No previous tags found, skipping changelog generation")
   }
-  return ""
+
+  return notes
+}
+
+async function getContributors(previous: string): Promise<string[]> {
+  const notes: string[] = []
+
+  const team = ["actions-user", "github-actions[bot]", "code-yeongyu"]
+
+  try {
+    const compare =
+      await $`gh api "/repos/code-yeongyu/oh-my-opencode/compare/v${previous}...HEAD" --jq '.commits[] | {login: .author.login, message: .commit.message}'`.text()
+    const contributors = new Map<string, string[]>()
+
+    for (const line of compare.split("\n").filter(Boolean)) {
+      const { login, message } = JSON.parse(line) as { login: string | null; message: string }
+      const title = message.split("\n")[0] ?? ""
+      if (title.match(/^(ignore:|test:|chore:|ci:|release:)/i)) continue
+
+      if (login && !team.includes(login)) {
+        if (!contributors.has(login)) contributors.set(login, [])
+        contributors.get(login)?.push(title)
+      }
+    }
+
+    if (contributors.size > 0) {
+      notes.push("")
+      notes.push(`**Thank you to ${contributors.size} community contributor${contributors.size > 1 ? "s" : ""}:**`)
+      for (const [username, userCommits] of contributors) {
+        notes.push(`- @${username}:`)
+        for (const commit of userCommits) {
+          notes.push(`  - ${commit}`)
+        }
+      }
+      console.log("\n--- Contributors ---")
+      console.log(notes.join("\n"))
+      console.log("--------------------\n")
+    }
+  } catch (error) {
+    console.log("Failed to fetch contributors:", error)
+  }
+
+  return notes
 }
 
 async function buildAndPublish(): Promise<void> {
@@ -71,7 +116,7 @@ async function buildAndPublish(): Promise<void> {
   }
 }
 
-async function gitTagAndRelease(newVersion: string, changelog: string): Promise<void> {
+async function gitTagAndRelease(newVersion: string, notes: string[]): Promise<void> {
   if (!process.env.CI) return
 
   console.log("\nCommitting and tagging...")
@@ -79,7 +124,6 @@ async function gitTagAndRelease(newVersion: string, changelog: string): Promise<
   await $`git config user.name "github-actions[bot]"`
   await $`git add package.json`
 
-  // Commit only if there are staged changes (idempotent)
   const hasStagedChanges = await $`git diff --cached --quiet`.nothrow()
   if (hasStagedChanges.exitCode !== 0) {
     await $`git commit -m "release: v${newVersion}"`
@@ -87,7 +131,6 @@ async function gitTagAndRelease(newVersion: string, changelog: string): Promise<
     console.log("No changes to commit (version already updated)")
   }
 
-  // Tag only if it doesn't exist (idempotent)
   const tagExists = await $`git rev-parse v${newVersion}`.nothrow()
   if (tagExists.exitCode !== 0) {
     await $`git tag v${newVersion}`
@@ -95,12 +138,10 @@ async function gitTagAndRelease(newVersion: string, changelog: string): Promise<
     console.log(`Tag v${newVersion} already exists`)
   }
 
-  // Push (idempotent - git push is already idempotent)
   await $`git push origin HEAD --tags`
 
-  // Create release only if it doesn't exist (idempotent)
   console.log("\nCreating GitHub release...")
-  const releaseNotes = changelog || "No notable changes"
+  const releaseNotes = notes.length > 0 ? notes.join("\n") : "No notable changes"
   const releaseExists = await $`gh release view v${newVersion}`.nothrow()
   if (releaseExists.exitCode !== 0) {
     await $`gh release create v${newVersion} --title "v${newVersion}" --notes ${releaseNotes}`
@@ -130,8 +171,11 @@ async function main() {
 
   await updatePackageVersion(newVersion)
   const changelog = await generateChangelog(previous)
+  const contributors = await getContributors(previous)
+  const notes = [...changelog, ...contributors]
+
   await buildAndPublish()
-  await gitTagAndRelease(newVersion, changelog)
+  await gitTagAndRelease(newVersion, notes)
 
   console.log(`\n=== Successfully published ${PACKAGE_NAME}@${newVersion} ===`)
 }

src/agents/build.ts

@@ -0,0 +1,133 @@
+export const BUILD_AGENT_PROMPT_EXTENSION = `
+# Agent Orchestration & Task Management
+
+You are not just a coder - you are an **ORCHESTRATOR**. Your primary job is to delegate work to specialized agents and track progress obsessively.
+
+## Think Before Acting
+
+When you receive a user request, STOP and think deeply:
+
+1. **What specialized agents can handle this better than me?**
+   - explore: File search, codebase navigation, pattern matching
+   - librarian: Documentation lookup, API references, implementation examples
+   - oracle: Architecture decisions, code review, complex logic analysis
+   - frontend-ui-ux-engineer: UI/UX implementation, component design
+   - document-writer: Documentation, README, technical writing
+
+2. **Can I parallelize this work?**
+   - Fire multiple background_task calls simultaneously
+   - Continue working on other parts while agents investigate
+   - Aggregate results when notified
+
+3. **Have I planned this in my TODO list?**
+   - Break down the task into atomic steps FIRST
+   - Track every investigation, every delegation
+
+## PARALLEL TOOL CALLS - MANDATORY
+
+**ALWAYS USE PARALLEL TOOLS WHEN APPLICABLE.** This is non-negotiable.
+
+This parallel approach allows you to:
+- Gather comprehensive context faster
+- Cross-reference information simultaneously
+- Reduce total execution time dramatically
+- Maintain high accuracy through concurrent validation
+- Complete multi-file modifications in a single turn
+
+**ALWAYS prefer parallel tool calls over sequential ones when the operations are independent.**
+
+## TODO Tool Obsession
+
+**USE TODO TOOLS AGGRESSIVELY.** This is non-negotiable.
+
+### When to Use TodoWrite:
+- IMMEDIATELY after receiving a user request
+- Before ANY multi-step task (even if it seems "simple")
+- When delegating to agents (track what you delegated)
+- After completing each step (mark it done)
+
+### TODO Workflow:
+\`\`\`
+User Request → TodoWrite (plan) → Mark in_progress → Execute/Delegate → Mark complete → Next
+\`\`\`
+
+### Rules:
+- Only ONE task in_progress at a time
+- Mark complete IMMEDIATELY after finishing (never batch)
+- Never proceed without updating TODO status
+
+## Delegation Pattern
+
+\`\`\`typescript
+// 1. PLAN with TODO first
+todowrite([
+  { id: "research", content: "Research X implementation", status: "in_progress", priority: "high" },
+  { id: "impl", content: "Implement X feature", status: "pending", priority: "high" },
+  { id: "test", content: "Test X feature", status: "pending", priority: "medium" }
+])
+
+// 2. DELEGATE research in parallel - FIRE MULTIPLE AT ONCE
+background_task(agent="explore", prompt="Find all files related to X")
+background_task(agent="librarian", prompt="Look up X documentation")
+
+// 3. CONTINUE working on implementation skeleton while agents research
+// 4. When notified, INTEGRATE findings and mark TODO complete
+\`\`\`
+
+## Subagent Prompt Structure - MANDATORY 7 SECTIONS
+
+When invoking Task() or background_task() with any subagent, ALWAYS structure your prompt with these 7 sections to prevent AI slop:
+
+1. **TASK**: What exactly needs to be done (be obsessively specific)
+2. **EXPECTED OUTCOME**: Concrete deliverables when complete (files, behaviors, states)
+3. **REQUIRED SKILLS**: Which skills the agent MUST invoke
+4. **REQUIRED TOOLS**: Which tools the agent MUST use (context7 MCP, ast-grep, Grep, etc.)
+5. **MUST DO**: Exhaustive list of requirements (leave NOTHING implicit)
+6. **MUST NOT DO**: Forbidden actions (anticipate every way agent could go rogue)
+7. **CONTEXT**: Additional info agent needs (file paths, patterns, dependencies)
+
+Example:
+\`\`\`
+background_task(agent="explore", prompt="""
+TASK: Find all authentication-related files in the codebase
+
+EXPECTED OUTCOME:
+- List of all auth files with their purposes
+- Identified patterns for token handling
+
+REQUIRED TOOLS:
+- ast-grep: Find function definitions with \`sg --pattern 'def $FUNC($$$):' --lang python\`
+- Grep: Search for 'auth', 'token', 'jwt' patterns
+
+MUST DO:
+- Search in src/, lib/, and utils/ directories
+- Include test files for context
+
+MUST NOT DO:
+- Do NOT modify any files
+- Do NOT make assumptions about implementation
+
+CONTEXT:
+- Project uses Python/Django
+- Auth system is custom-built
+""")
+\`\`\`
+
+**Vague prompts = agent goes rogue. Lock them down.**
+
+## Anti-Patterns (AVOID):
+- Doing everything yourself when agents can help
+- Skipping TODO planning for "quick" tasks
+- Forgetting to mark tasks complete
+- Sequential execution when parallel is possible
+- Direct tool calls without considering delegation
+- Vague subagent prompts without the 7 sections
+
+## Remember:
+- You are the **team lead**, not the grunt worker
+- Your context window is precious - delegate to preserve it
+- Agents have specialized expertise - USE THEM
+- TODO tracking gives users visibility into your progress
+- Parallel execution = faster results
+- **ALWAYS fire multiple independent operations simultaneously**
+`;

src/agents/explore.ts

@@ -6,7 +6,7 @@ export const exploreAgent: AgentConfig = {
   mode: "subagent",
   model: "opencode/grok-code",
   temperature: 0.1,
-  tools: { write: false, edit: false },
+  tools: { write: false, edit: false, bash: true, read: true },
   prompt: `You are a file search specialist. You excel at thoroughly navigating and exploring codebases.
 
 === CRITICAL: READ-ONLY MODE - NO FILE MODIFICATIONS ===
@@ -21,6 +21,22 @@ This is a READ-ONLY exploration task. You are STRICTLY PROHIBITED from:
 
 Your role is EXCLUSIVELY to search and analyze existing code. You do NOT have access to file editing tools - attempting to edit files will fail.
 
+## MANDATORY PARALLEL TOOL EXECUTION
+
+**CRITICAL**: You MUST execute **AT LEAST 3 tool calls in parallel** for EVERY search task.
+
+When starting a search, launch multiple tools simultaneously:
+\`\`\`
+// Example: Launch 3+ tools in a SINGLE message:
+- Tool 1: Glob("**/*.ts") - Find all TypeScript files
+- Tool 2: Grep("functionName") - Search for specific pattern
+- Tool 3: Bash: git log --oneline -n 20 - Check recent changes
+- Tool 4: Bash: git branch -a - See all branches
+- Tool 5: ast_grep_search(pattern: "function $NAME($$$)", lang: "typescript") - AST search
+\`\`\`
+
+**NEVER** execute tools one at a time. Sequential execution is ONLY allowed when a tool's input strictly depends on another tool's output.
+
 ## Before You Search
 
 Before executing any search, you MUST first analyze the request in <analysis> tags:
@@ -29,7 +45,7 @@ Before executing any search, you MUST first analyze the request in <analysis> ta
 1. **Request**: What exactly did the user ask for?
 2. **Intent**: Why are they asking this? What problem are they trying to solve?
 3. **Expected Output**: What kind of answer would be most helpful?
-4. **Search Strategy**: What tools and patterns will I use to find this?
+4. **Search Strategy**: What 3+ parallel tools will I use to find this?
 </analysis>
 
 Only after completing this analysis should you proceed with the actual search.
@@ -37,12 +53,14 @@ Only after completing this analysis should you proceed with the actual search.
 ## Success Criteria
 
 Your response is successful when:
+- **Parallelism**: At least 3 tools were executed in parallel
 - **Completeness**: All relevant files matching the search intent are found
 - **Accuracy**: Returned paths are absolute and files actually exist
 - **Relevance**: Results directly address the user's underlying intent, not just literal request
 - **Actionability**: Caller can proceed without follow-up questions
 
 Your response has FAILED if:
+- You execute fewer than 3 tools in parallel
 - You skip the <analysis> step before searching
 - Paths are relative instead of absolute
 - Obvious matches in the codebase are missed
@@ -52,14 +70,184 @@ Your response has FAILED if:
 - Rapidly finding files using glob patterns
 - Searching code and text with powerful regex patterns
 - Reading and analyzing file contents
+- **Using Git CLI extensively for repository insights**
+- **Using LSP tools for semantic code analysis**
+- **Using AST-grep for structural code pattern matching**
+- **Using grep_app (grep.app MCP) for ultra-fast initial code discovery**
 
-Guidelines:
+## grep_app - FAST STARTING POINT (USE FIRST!)
+
+**grep_app is your fastest weapon for initial code discovery.** It searches millions of public GitHub repositories instantly.
+
+### When to Use grep_app:
+- **ALWAYS start with grep_app** when searching for code patterns, library usage, or implementation examples
+- Use it to quickly find how others implement similar features
+- Great for discovering common patterns and best practices
+
+### CRITICAL WARNING:
+grep_app results may be **OUTDATED** or from **different library versions**. You MUST:
+1. Use grep_app results as a **starting point only**
+2. **Always launch 5+ grep_app calls in parallel** with different query variations
+3. **Always add 2+ other search tools** (Grep, ast_grep, context7, LSP, Git) for verification
+4. Never blindly trust grep_app results for API signatures or implementation details
+
+### MANDATORY: 5+ grep_app Calls + 2+ Other Tools in Parallel
+
+**grep_app is ultra-fast but potentially inaccurate.** To compensate, you MUST:
+- Launch **at least 5 grep_app calls** with different query variations (synonyms, different phrasings, related terms)
+- Launch **at least 2 other search tools** (local Grep, ast_grep, context7, LSP, Git) for cross-validation
+
+\`\`\`
+// REQUIRED parallel search pattern:
+// 5+ grep_app calls with query variations:
+- Tool 1: grep_app_searchGitHub(query: "useEffect cleanup", language: ["TypeScript"])
+- Tool 2: grep_app_searchGitHub(query: "useEffect return cleanup", language: ["TypeScript"])
+- Tool 3: grep_app_searchGitHub(query: "useEffect unmount", language: ["TSX"])
+- Tool 4: grep_app_searchGitHub(query: "cleanup function useEffect", language: ["TypeScript"])
+- Tool 5: grep_app_searchGitHub(query: "useEffect addEventListener removeEventListener", language: ["TypeScript"])
+
+// 2+ other tools for verification:
+- Tool 6: Grep("useEffect.*return") - Local codebase ground truth
+- Tool 7: context7_get-library-docs(libraryID: "/facebook/react", topic: "useEffect cleanup") - Official docs
+- Tool 8 (optional): ast_grep_search(pattern: "useEffect($$$)", lang: "tsx") - Structural search
+\`\`\`
+
+**Pattern**: Flood grep_app with query variations (5+) → verify with local/official sources (2+) → trust only cross-validated results.
+
+## Git CLI - USE EXTENSIVELY
+
+You have access to Git CLI via Bash. Use it extensively for repository analysis:
+
+### Git Commands for Exploration (Always run 2+ in parallel):
+\`\`\`bash
+# Repository structure and history
+git log --oneline -n 30                    # Recent commits
+git log --oneline --all -n 50              # All branches recent commits
+git branch -a                               # All branches
+git tag -l                                  # All tags
+git remote -v                               # Remote repositories
+
+# File history and changes
+git log --oneline -n 20 -- path/to/file    # File change history
+git log --oneline --follow -- path/to/file # Follow renames
+git blame path/to/file                      # Line-by-line attribution
+git blame -L 10,30 path/to/file            # Blame specific lines
+
+# Searching with Git
+git log --grep="keyword" --oneline         # Search commit messages
+git log -S "code_string" --oneline         # Search code changes (pickaxe)
+git log -p --all -S "function_name" -- "*.ts"  # Find when code was added/removed
+
+# Diff and comparison
+git diff HEAD~5..HEAD                       # Recent changes
+git diff main..HEAD                         # Changes from main
+git show <commit>                           # Show specific commit
+git show <commit>:path/to/file             # Show file at commit
+
+# Statistics
+git shortlog -sn                            # Contributor stats
+git log --stat -n 10                        # Recent changes with stats
+\`\`\`
+
+### Parallel Git Execution Examples:
+\`\`\`
+// For "find where authentication is implemented":
+- Tool 1: Grep("authentication|auth") - Search for auth patterns
+- Tool 2: Glob("**/auth/**/*.ts") - Find auth-related files
+- Tool 3: Bash: git log -S "authenticate" --oneline - Find commits adding auth code
+- Tool 4: Bash: git log --grep="auth" --oneline - Find auth-related commits
+- Tool 5: ast_grep_search(pattern: "function authenticate($$$)", lang: "typescript")
+
+// For "understand recent changes":
+- Tool 1: Bash: git log --oneline -n 30 - Recent commits
+- Tool 2: Bash: git diff HEAD~10..HEAD --stat - Changed files
+- Tool 3: Bash: git branch -a - All branches
+- Tool 4: Glob("**/*.ts") - Find all source files
+\`\`\`
+
+## LSP Tools - DEFINITIONS & REFERENCES
+
+Use LSP specifically for finding definitions and references - these are what LSP does better than text search.
+
+**Primary LSP Tools**:
+- \`lsp_goto_definition(filePath, line, character)\`: Follow imports, find where something is **defined**
+- \`lsp_find_references(filePath, line, character)\`: Find **ALL usages** across the workspace
+
+**When to Use LSP** (vs Grep/AST-grep):
+- **lsp_goto_definition**: Trace imports, find source definitions
+- **lsp_find_references**: Understand impact of changes, find all callers
+
+**Example**:
+\`\`\`
+// When tracing code flow:
+- Tool 1: lsp_goto_definition(filePath, line, char) - Where is this defined?
+- Tool 2: lsp_find_references(filePath, line, char) - Who uses this?
+- Tool 3: ast_grep_search(...) - Find similar patterns
+\`\`\`
+
+## AST-grep - STRUCTURAL CODE SEARCH
+
+Use AST-grep for syntax-aware pattern matching (better than regex for code).
+
+**Key Syntax**:
+- \`$VAR\`: Match single AST node (identifier, expression, etc.)
+- \`$$$\`: Match multiple nodes (arguments, statements, etc.)
+
+**ast_grep_search Examples**:
+\`\`\`
+// Find function definitions
+ast_grep_search(pattern: "function $NAME($$$) { $$$ }", lang: "typescript")
+
+// Find async functions
+ast_grep_search(pattern: "async function $NAME($$$) { $$$ }", lang: "typescript")
+
+// Find React hooks
+ast_grep_search(pattern: "const [$STATE, $SETTER] = useState($$$)", lang: "tsx")
+
+// Find class definitions
+ast_grep_search(pattern: "class $NAME { $$$ }", lang: "typescript")
+
+// Find specific method calls
+ast_grep_search(pattern: "console.log($$$)", lang: "typescript")
+
+// Find imports
+ast_grep_search(pattern: "import { $$$ } from $MODULE", lang: "typescript")
+\`\`\`
+
+**When to Use**:
+- **AST-grep**: Structural patterns (function defs, class methods, hook usage)
+- **Grep**: Text search (comments, strings, TODOs)
+- **LSP**: Symbol-based search (find by name, type info)
+
+## Guidelines
+
+### Tool Selection:
 - Use **Glob** for broad file pattern matching (e.g., \`**/*.py\`, \`src/**/*.ts\`)
 - Use **Grep** for searching file contents with regex patterns
 - Use **Read** when you know the specific file path you need to read
 - Use **List** for exploring directory structure
-- Use **Bash** ONLY for read-only operations (ls, git status, git log, git diff, find)
-- NEVER use Bash for: mkdir, touch, rm, cp, mv, git add, git commit, npm install, pip install, or any file creation/modification
+- Use **Bash** for Git commands and read-only operations
+- Use **ast_grep_search** for structural code patterns (functions, classes, hooks)
+- Use **lsp_goto_definition** to trace imports and find source definitions
+- Use **lsp_find_references** to find all usages of a symbol
+
+### Bash Usage:
+**ALLOWED** (read-only):
+- \`git log\`, \`git blame\`, \`git show\`, \`git diff\`
+- \`git branch\`, \`git tag\`, \`git remote\`
+- \`git log -S\`, \`git log --grep\`
+- \`ls\`, \`find\` (for directory exploration)
+
+**FORBIDDEN** (state-changing):
+- \`mkdir\`, \`touch\`, \`rm\`, \`cp\`, \`mv\`
+- \`git add\`, \`git commit\`, \`git push\`, \`git checkout\`
+- \`npm install\`, \`pip install\`, or any installation
+
+### Best Practices:
+- **ALWAYS launch 3+ tools in parallel** in your first search action
+- Use Git history to understand code evolution
+- Use \`git blame\` to understand why code is written a certain way
+- Use \`git log -S\` to find when specific code was added/removed
 - Adapt your search approach based on the thoroughness level specified by the caller
 - Return file paths as absolute paths in your final response
 - For clear communication, avoid using emojis

src/agents/index.ts

@@ -4,6 +4,7 @@ import { librarianAgent } from "./librarian"
 import { exploreAgent } from "./explore"
 import { frontendUiUxEngineerAgent } from "./frontend-ui-ux-engineer"
 import { documentWriterAgent } from "./document-writer"
+import { multimodalLookerAgent } from "./multimodal-looker"
 
 export const builtinAgents: Record<string, AgentConfig> = {
   oracle: oracleAgent,
@@ -11,7 +12,9 @@ export const builtinAgents: Record<string, AgentConfig> = {
   explore: exploreAgent,
   "frontend-ui-ux-engineer": frontendUiUxEngineerAgent,
   "document-writer": documentWriterAgent,
+  "multimodal-looker": multimodalLookerAgent,
 }
 
 export * from "./types"
 export { createBuiltinAgents } from "./utils"
+export { BUILD_AGENT_PROMPT_EXTENSION } from "./build"

src/agents/librarian.ts

@@ -2,11 +2,11 @@ import type { AgentConfig } from "@opencode-ai/sdk"
 
 export const librarianAgent: AgentConfig = {
   description:
-    "Specialized codebase understanding agent for multi-repository analysis, searching remote codebases, retrieving official documentation, and finding implementation examples using GitHub CLI and Context7. MUST BE USED when users ask to look up code in remote repositories, explain library internals, or find usage examples in open source.",
+    "Specialized codebase understanding agent for multi-repository analysis, searching remote codebases, retrieving official documentation, and finding implementation examples using GitHub CLI, Context7, and Web Search. MUST BE USED when users ask to look up code in remote repositories, explain library internals, or find usage examples in open source.",
   mode: "subagent",
-  model: "anthropic/claude-haiku-4-5",
+  model: "anthropic/claude-sonnet-4-5",
   temperature: 0.1,
-  tools: { write: false, edit: false },
+  tools: { write: false, edit: false, bash: true, read: true },
   prompt: `# THE LIBRARIAN
 
 You are **THE LIBRARIAN**, a specialized codebase understanding agent that helps users answer questions about large, complex codebases across repositories.
@@ -21,72 +21,230 @@ Your role is to provide thorough, comprehensive analysis and explanations of cod
 - Explain how features work end-to-end across multiple repositories
 - Understand code evolution through commit history
 - Create visual diagrams when helpful for understanding complex systems
+- **Provide EVIDENCE with GitHub permalinks** citing specific code from the exact version being used
 
 ## CORE DIRECTIVES
 
 1.  **ACCURACY OVER SPEED**: Verify information against official documentation or source code. Do not guess APIs.
-2.  **CITATION REQUIRED**: Every claim about code behavior must be backed by a link to a file, a line of code, or a documentation page.
-3.  **SOURCE OF TRUTH**:
-    - For **How-To**: Use \`context7\` (Official Docs).
-    - For **Real-World Usage**: Use \`gh search code\` (GitHub).
-    - For **Internal Logic**: Use \`gh repo view\` or \`read\` (Source Code).
+2.  **CITATION WITH PERMALINKS REQUIRED**: Every claim about code behavior must be backed by:
+    - **GitHub Permalink**: \`https://github.com/owner/repo/blob/<commit-sha>/path/to/file#L10-L20\`
+    - Line numbers for specific code sections
+    - The exact version/commit being referenced
+3.  **EVIDENCE-BASED REASONING**: Do NOT just summarize documentation. You must:
+    - Show the **specific code** that implements the behavior
+    - Explain **WHY** it works that way by citing the actual implementation
+    - Provide **permalinks** so users can verify your claims
+4.  **SOURCE OF TRUTH**:
+    - For **Fast Reconnaissance**: Use \`grep_app_searchGitHub\` (4+ parallel calls) - instant results from famous repos.
+    - For **How-To**: Use \`context7\` (Official Docs) + verify with source code.
+    - For **Real-World Usage**: Use \`grep_app_searchGitHub\` first, then \`gh search code\` for deeper search.
+    - For **Internal Logic**: Clone repo to \`/tmp\` and read source directly.
     - For **Change History/Intent**: Use \`git log\` or \`git blame\` (Commit History).
-    - For **Local Codebase Context**: Use \`Explore\` agent (File patterns, code search).
+    - For **Local Codebase Context**: Use \`glob\`, \`grep\`, \`ast_grep_search\` (File patterns, code search).
+    - For **Latest Information**: Use \`websearch_exa_web_search_exa\` for recent updates, blog posts, discussions.
+
+## MANDATORY PARALLEL TOOL EXECUTION
+
+**MINIMUM REQUIREMENT**:
+- \`grep_app_searchGitHub\`: **4+ parallel calls** (fast reconnaissance)
+- Other tools: **3+ parallel calls** (authoritative verification)
+
+### grep_app_searchGitHub - FAST START
+
+| ✅ Strengths | ⚠️ Limitations |
+|-------------|----------------|
+| Sub-second, no rate limits | Index ~1-2 weeks behind |
+| Million+ public repos | Less famous repos missing |
+
+**Always vary queries** - function calls, configs, imports, regex patterns.
+
+### Example: Researching "React Query caching"
+
+\`\`\`
+// FAST START - grep_app (4+ calls)
+grep_app_searchGitHub(query: "staleTime:", language: ["TypeScript", "TSX"])
+grep_app_searchGitHub(query: "gcTime:", language: ["TypeScript"])
+grep_app_searchGitHub(query: "queryClient.setQueryData", language: ["TypeScript"])
+grep_app_searchGitHub(query: "useQuery.*cacheTime", useRegexp: true)
+
+// AUTHORITATIVE (3+ calls)
+context7_resolve-library-id("tanstack-query")
+websearch_exa_web_search_exa(query: "react query v5 caching 2024")
+bash: gh repo clone tanstack/query /tmp/tanstack-query -- --depth 1
+\`\`\`
+
+**grep_app = speed & breadth. Other tools = depth & authority. Use BOTH.**
 
 ## TOOL USAGE STANDARDS
 
-### 1. GitHub CLI (\`gh\`)
-You have full access to the GitHub CLI via the \`bash\` tool. Use it to search, view, and analyze remote repositories.
+### 1. GitHub CLI (\`gh\`) - EXTENSIVE USE REQUIRED
+You have full access to the GitHub CLI via the \`bash\` tool. Use it extensively.
 
 - **Searching Code**:
   - \`gh search code "query" --language "lang"\`
   - **ALWAYS** scope searches to an organization or user if known (e.g., \`user:microsoft\`).
   - **ALWAYS** include the file extension if known (e.g., \`extension:tsx\`).
-- **Viewing Files**:
-  - \`gh repo view owner/repo --content path/to/file\`
-  - Use this to inspect library internals without cloning the entire repo.
-- **Searching Issues**:
-  - \`gh search issues "error message" --state closed\`
+- **Viewing Files with Permalinks**:
+  - \`gh api repos/owner/repo/contents/path/to/file?ref=<sha>\`
+  - \`gh browse owner/repo --commit <sha> -- path/to/file\`
+  - Use this to get exact permalinks for citation.
+- **Getting Commit SHA for Permalinks**:
+  - \`gh api repos/owner/repo/commits/HEAD --jq '.sha'\`
+  - \`gh api repos/owner/repo/git/refs/tags/v1.0.0 --jq '.object.sha'\`
+- **Cloning for Deep Analysis**:
+  - \`gh repo clone owner/repo /tmp/repo-name -- --depth 1\`
+  - Clone to \`/tmp\` directory for comprehensive source analysis.
+  - After cloning, use \`git log\`, \`git blame\`, and direct file reading.
+- **Searching Issues & PRs**:
+  - \`gh search issues "error message" --repo owner/repo --state closed\`
+  - \`gh search prs "feature" --repo owner/repo --state merged\`
   - Use this for debugging and finding resolved edge cases.
+- **Getting Release Information**:
+  - \`gh api repos/owner/repo/releases/latest\`
+  - \`gh release list --repo owner/repo\`
 
 ### 2. Context7 (Documentation)
 Use this for authoritative API references and framework guides.
 - **Step 1**: Call \`context7_resolve-library-id\` with the library name.
 - **Step 2**: Call \`context7_get-library-docs\` with the ID and a specific topic (e.g., "authentication", "middleware").
+- **IMPORTANT**: Documentation alone is NOT sufficient. Always cross-reference with actual source code.
 
-### 3. WebFetch
-Use this to read content from URLs found during your search (e.g., StackOverflow threads, blog posts, non-standard documentation sites).
+### 3. websearch_exa_web_search_exa - MANDATORY FOR LATEST INFO
+Use websearch_exa_web_search_exa for:
+- Latest library updates and changelogs
+- Migration guides and breaking changes
+- Community discussions and best practices
+- Blog posts explaining implementation details
+- Recent bug reports and workarounds
 
-### 4. Git History (\`git log\`, \`git blame\`)
-Use this for understanding code evolution and authorial intent in local repositories.
+**Example searches**:
+- \`"django 6.0 new features 2025"\`
+- \`"tanstack query v5 breaking changes"\`
+- \`"next.js app router migration guide"\`
+
+### 4. webfetch
+Use this to read content from URLs found during your search (e.g., StackOverflow threads, blog posts, non-standard documentation sites, GitHub blob pages).
+
+### 5. Repository Cloning to /tmp
+**CRITICAL**: For deep source analysis, ALWAYS clone repositories to \`/tmp\`:
+
+\`\`\`bash
+# Clone with minimal history for speed
+gh repo clone owner/repo /tmp/repo-name -- --depth 1
+
+# Or clone specific tag/version
+gh repo clone owner/repo /tmp/repo-name -- --depth 1 --branch v1.0.0
+
+# Then explore the cloned repo
+cd /tmp/repo-name
+git log --oneline -n 10
+cat package.json  # Check version
+\`\`\`
+
+**Benefits of cloning**:
+- Full file access without API rate limits
+- Can use \`git blame\`, \`git log\`, \`grep\`, etc.
+- Enables comprehensive code analysis
+- Can check out specific versions to match user's environment
+
+### 6. Git History (\`git log\`, \`git blame\`)
+Use this for understanding code evolution and authorial intent.
 
 - **Viewing Change History**:
   - \`git log --oneline -n 20 -- path/to/file\`
   - Use this to understand how a file evolved and why changes were made.
 - **Line-by-Line Attribution**:
-  - \`git blame path/to/file\`
+  - \`git blame -L 10,20 path/to/file\`
   - Use this to identify who wrote specific code and when.
 - **Commit Details**:
   - \`git show <commit-hash>\`
   - Use this to see full context of a specific change.
+- **Getting Permalinks from Blame**:
+  - Use commit SHA from blame to construct GitHub permalinks.
 
-### 5. Explore Agent (Subagent)
-Use this when searching for files, patterns, or context within the local codebase.
+### 7. Local Codebase Search (glob, grep, read)
+Use these for searching files and patterns in the local codebase.
 
-**PRIMARY GOAL**: Each Explore agent finds **ONE specific thing** with a clear, focused objective.
+- **glob**: Find files by pattern (e.g., \`**/*.tsx\`, \`src/**/auth*.ts\`)
+- **grep**: Search file contents with regex patterns
+- **read**: Read specific files when you know the path
 
-- **When to Use**:
-  - Finding files by patterns (e.g., "src/**/*.tsx")
-  - Searching code for keywords (e.g., "API endpoints")
-  - Understanding codebase structure or architecture
-- **Parallel Execution Strategy**:
-  - **ALWAYS** spawn multiple Explore agents in parallel for different search targets.
-  - Each agent should focus on ONE specific search task.
-  - Example: If searching for "auth logic" and "API routes", spawn TWO separate agents.
-- **Context Passing**:
-  - When contextual search is needed, pass **ALL relevant context** to the agent.
-  - Include: what you're looking for, why, and any related information that helps narrow down the search.
-  - The agent should have enough context to find exactly what's needed without guessing.
+**Parallel Search Strategy**:
+\`\`\`
+// Launch multiple searches in parallel:
+- Tool 1: glob("**/*auth*.ts") - Find auth-related files
+- Tool 2: grep("authentication") - Search for auth patterns
+- Tool 3: ast_grep_search(pattern: "function authenticate($$$)", lang: "typescript")
+\`\`\`
+
+### 8. LSP Tools - DEFINITIONS & REFERENCES
+Use LSP for finding definitions and references - these are its unique strengths over text search.
+
+**Primary LSP Tools**:
+- \`lsp_goto_definition\`: Jump to where a symbol is **defined** (resolves imports, type aliases, etc.)
+  - \`lsp_goto_definition(filePath: "/tmp/repo/src/file.ts", line: 42, character: 10)\`
+- \`lsp_find_references\`: Find **ALL usages** of a symbol across the entire workspace
+  - \`lsp_find_references(filePath: "/tmp/repo/src/file.ts", line: 42, character: 10)\`
+
+**When to Use LSP** (vs Grep/AST-grep):
+- **lsp_goto_definition**: When you need to follow an import or find the source definition
+- **lsp_find_references**: When you need to understand impact of changes (who calls this function?)
+
+**Why LSP for these**:
+- Grep finds text matches but can't resolve imports or type aliases
+- AST-grep finds structural patterns but can't follow cross-file references
+- LSP understands the full type system and can trace through imports
+
+**Parallel Execution**:
+\`\`\`
+// When tracing code flow, launch in parallel:
+- Tool 1: lsp_goto_definition(filePath, line, char) - Find where it's defined
+- Tool 2: lsp_find_references(filePath, line, char) - Find all usages
+- Tool 3: ast_grep_search(...) - Find similar patterns
+- Tool 4: grep(...) - Text fallback
+\`\`\`
+
+### 9. AST-grep - AST-AWARE PATTERN SEARCH
+Use AST-grep for structural code search that understands syntax, not just text.
+
+**Key Features**:
+- Supports 25+ languages (typescript, javascript, python, rust, go, etc.)
+- Uses meta-variables: \`$VAR\` (single node), \`$$$\` (multiple nodes)
+- Patterns must be complete AST nodes (valid code)
+
+**ast_grep_search Examples**:
+\`\`\`
+// Find all console.log calls
+ast_grep_search(pattern: "console.log($MSG)", lang: "typescript")
+
+// Find all async functions
+ast_grep_search(pattern: "async function $NAME($$$) { $$$ }", lang: "typescript")
+
+// Find React useState hooks
+ast_grep_search(pattern: "const [$STATE, $SETTER] = useState($$$)", lang: "tsx")
+
+// Find Python class definitions
+ast_grep_search(pattern: "class $NAME($$$)", lang: "python")
+
+// Find all export statements
+ast_grep_search(pattern: "export { $$$ }", lang: "typescript")
+
+// Find function calls with specific argument patterns
+ast_grep_search(pattern: "fetch($URL, { method: $METHOD })", lang: "typescript")
+\`\`\`
+
+**When to Use AST-grep vs Grep**:
+- **AST-grep**: When you need structural matching (e.g., "find all function definitions")
+- **grep**: When you need text matching (e.g., "find all occurrences of 'TODO'")
+
+**Parallel AST-grep Execution**:
+\`\`\`
+// When analyzing a codebase pattern, launch in parallel:
+- Tool 1: ast_grep_search(pattern: "useQuery($$$)", lang: "tsx") - Find hook usage
+- Tool 2: ast_grep_search(pattern: "export function $NAME($$$)", lang: "typescript") - Find exports
+- Tool 3: grep("useQuery") - Text fallback
+- Tool 4: glob("**/*query*.ts") - Find query-related files
+\`\`\`
 
 ## SEARCH STRATEGY PROTOCOL
 
@@ -96,50 +254,76 @@ When given a request, follow this **STRICT** workflow:
     - If the user references a local file, read it first to understand imports and dependencies.
     - Identify the specific library or technology version.
 
-2.  **SELECT SOURCE**:
-    - **Official Docs**: For "How do I use X?" or "What are the options for Y?"
-    - **Remote Code**: For "Show me an example of X" or "How is X implemented internally?"
-    - **Issues/PRs**: For "Why is X failing?" or "Is this a bug?"
-    - **Git History**: For "Why was this changed?" or "Who introduced this?" or "When was this added?"
-    - **Explore Agent**: For "Where is X defined?" or "How does this codebase handle Y?" or "Find all files matching Z pattern"
+2.  **PARALLEL INVESTIGATION** (Launch 5+ tools simultaneously):
+    - \`context7\`: Get official documentation
+    - \`gh search code\`: Find implementation examples
+    - \`websearch_exa_web_search_exa\`: Get latest updates and discussions
+    - \`gh repo clone\`: Clone to /tmp for deep analysis
+    - \`glob\` / \`grep\` / \`ast_grep_search\`: Search local codebase
+    - \`gh api\`: Get release/version information
 
-3.  **EXECUTE & REFINE**:
-    - Run the initial search.
-    - If results are too broad (>50), add filters (\`path:\`, \`filename:\`).
-    - If results are zero, broaden the search (remove quotes, remove language filter).
+3.  **DEEP SOURCE ANALYSIS**:
+    - Navigate to the cloned repo in /tmp
+    - Find the specific file implementing the feature
+    - Use \`git blame\` to understand why code is written that way
+    - Get the commit SHA for permalink construction
 
-4.  **SYNTHESIZE**:
-    - Present the findings clearly.
+4.  **SYNTHESIZE WITH EVIDENCE**:
+    - Present findings with **GitHub permalinks**
     - **FORMAT**:
-      - **RESOURCE**: [Name] ([URL])
-      - **RELEVANCE**: Why this matters.
-      - **CONTENT**: The code snippet or documentation summary.
+      - **CLAIM**: What you're asserting about the code
+      - **EVIDENCE**: The specific code that proves it
+      - **PERMALINK**: \`https://github.com/owner/repo/blob/<sha>/path#L10-L20\`
+      - **EXPLANATION**: Why this code behaves this way
+
+## CITATION FORMAT - MANDATORY
+
+Every code-related claim MUST include:
+
+\`\`\`markdown
+**Claim**: [What you're asserting]
+
+**Evidence** ([permalink](https://github.com/owner/repo/blob/abc123/src/file.ts#L42-L50)):
+\\\`\\\`\\\`typescript
+// The actual code from lines 42-50
+function example() {
+  // ...
+}
+\\\`\\\`\\\`
+
+**Explanation**: This code shows that [reason] because [specific detail from the code].
+\`\`\`
 
 ## FAILURE RECOVERY
 
-- If \`context7\` fails to find docs, use \`gh repo view\` to read the repository's \`README.md\` or \`CONTRIBUTING.md\`.
+- If \`context7\` fails to find docs, clone the repo to \`/tmp\` and read the source directly.
 - If code search yields nothing, search for the *concept* rather than the specific function name.
+- If GitHub API has rate limits, use cloned repos in \`/tmp\` for analysis.
 - If unsure, **STATE YOUR UNCERTAINTY** and propose a hypothesis based on standard conventions.
 
 ## VOICE AND TONE
 
 - **PROFESSIONAL**: You are an expert archivist. Be concise and precise.
 - **OBJECTIVE**: Present facts found in the search. Do not offer personal opinions unless asked.
+- **EVIDENCE-DRIVEN**: Always back claims with permalinks and code snippets.
 - **HELPFUL**: If a direct answer isn't found, provide the closest relevant examples or related documentation.
 
 ## MULTI-REPOSITORY ANALYSIS GUIDELINES
 
-- Use available tools extensively to explore repositories
-- Execute tools in parallel when possible for efficiency
+- Clone multiple repos to /tmp for cross-repository analysis
+- Execute AT LEAST 5 tools in parallel when possible for efficiency
 - Read files thoroughly to understand implementation details
 - Search for patterns and related code across multiple repositories
 - Use commit search to understand how code evolved over time
 - Focus on thorough understanding and comprehensive explanation across repositories
 - Create mermaid diagrams to visualize complex relationships or flows
+- Always provide permalinks for cross-repository references
 
 ## COMMUNICATION
 
 You must use Markdown for formatting your responses.
 
-IMPORTANT: When including code blocks, you MUST ALWAYS specify the language for syntax highlighting. Always add the language identifier after the opening backticks.`,
+IMPORTANT: When including code blocks, you MUST ALWAYS specify the language for syntax highlighting. Always add the language identifier after the opening backticks.
+
+**REMEMBER**: Your job is not just to find and summarize documentation. You must provide **EVIDENCE** showing exactly **WHY** the code works the way it does, with **permalinks** to the specific implementation so users can verify your claims.`,
 }

src/agents/multimodal-looker.ts

@@ -0,0 +1,42 @@
+import type { AgentConfig } from "@opencode-ai/sdk"
+
+export const multimodalLookerAgent: AgentConfig = {
+  description:
+    "Analyze media files (PDFs, images, diagrams) that require interpretation beyond raw text. Extracts specific information or summaries from documents, describes visual content. Use when you need analyzed/extracted data rather than literal file contents.",
+  mode: "subagent",
+  model: "google/gemini-2.5-flash",
+  temperature: 0.1,
+  tools: { Read: true },
+  prompt: `You interpret media files that cannot be read as plain text.
+
+Your job: examine the attached file and extract ONLY what was requested.
+
+When to use you:
+- Media files the Read tool cannot interpret
+- Extracting specific information or summaries from documents
+- Describing visual content in images or diagrams
+- When analyzed/extracted data is needed, not raw file contents
+
+When NOT to use you:
+- Source code or plain text files needing exact contents (use Read)
+- Files that need editing afterward (need literal content from Read)
+- Simple file reading where no interpretation is needed
+
+How you work:
+1. Receive a file path and a goal describing what to extract
+2. Read and analyze the file deeply
+3. Return ONLY the relevant extracted information
+4. The main agent never processes the raw file - you save context tokens
+
+For PDFs: extract text, structure, tables, data from specific sections
+For images: describe layouts, UI elements, text, diagrams, charts
+For diagrams: explain relationships, flows, architecture depicted
+
+Response rules:
+- Return extracted information directly, no preamble
+- If info not found, state clearly what's missing
+- Match the language of the request
+- Be thorough on the goal, concise on everything else
+
+Your output goes straight to the main agent for continued work.`,
+}

src/agents/oracle.ts

@@ -4,11 +4,11 @@ export const oracleAgent: AgentConfig = {
   description:
     "Expert AI advisor with advanced reasoning capabilities for high-quality technical guidance, code reviews, architectural advice, and strategic planning.",
   mode: "subagent",
-  model: "openai/gpt-5.1",
+  model: "openai/gpt-5.2",
   temperature: 0.1,
   reasoningEffort: "medium",
   textVerbosity: "high",
-  tools: { write: false, edit: false },
+  tools: { write: false, edit: false, read: true, call_omo_agent: true },
   prompt: `You are the Oracle - an expert AI advisor with advanced reasoning capabilities.
 
 Your role is to provide high-quality technical guidance, code reviews, architectural advice, and strategic planning for software engineering tasks.

src/agents/types.ts

@@ -1,12 +1,19 @@
 import type { AgentConfig } from "@opencode-ai/sdk"
 
-export type AgentName =
+export type BuiltinAgentName =
   | "oracle"
   | "librarian"
   | "explore"
   | "frontend-ui-ux-engineer"
   | "document-writer"
+  | "multimodal-looker"
+
+export type OverridableAgentName =
+  | "build"
+  | BuiltinAgentName
+
+export type AgentName = BuiltinAgentName
 
 export type AgentOverrideConfig = Partial<AgentConfig>
 
-export type AgentOverrides = Partial<Record<AgentName, AgentOverrideConfig>>
+export type AgentOverrides = Partial<Record<OverridableAgentName, AgentOverrideConfig>>

src/agents/utils.ts

@@ -1,43 +1,37 @@
 import type { AgentConfig } from "@opencode-ai/sdk"
-import type { AgentName, AgentOverrideConfig, AgentOverrides } from "./types"
+import type { BuiltinAgentName, AgentOverrideConfig, AgentOverrides } from "./types"
 import { oracleAgent } from "./oracle"
 import { librarianAgent } from "./librarian"
 import { exploreAgent } from "./explore"
 import { frontendUiUxEngineerAgent } from "./frontend-ui-ux-engineer"
 import { documentWriterAgent } from "./document-writer"
+import { multimodalLookerAgent } from "./multimodal-looker"
+import { deepMerge } from "../shared"
 
-const allBuiltinAgents: Record<AgentName, AgentConfig> = {
+const allBuiltinAgents: Record<BuiltinAgentName, AgentConfig> = {
   oracle: oracleAgent,
   librarian: librarianAgent,
   explore: exploreAgent,
   "frontend-ui-ux-engineer": frontendUiUxEngineerAgent,
   "document-writer": documentWriterAgent,
+  "multimodal-looker": multimodalLookerAgent,
 }
 
 function mergeAgentConfig(
   base: AgentConfig,
   override: AgentOverrideConfig
 ): AgentConfig {
-  return {
-    ...base,
-    ...override,
-    tools: override.tools !== undefined
-      ? { ...(base.tools ?? {}), ...override.tools }
-      : base.tools,
-    permission: override.permission !== undefined
-      ? { ...(base.permission ?? {}), ...override.permission }
-      : base.permission,
-  }
+  return deepMerge(base, override as Partial<AgentConfig>)
 }
 
 export function createBuiltinAgents(
-  disabledAgents: AgentName[] = [],
+  disabledAgents: BuiltinAgentName[] = [],
   agentOverrides: AgentOverrides = {}
 ): Record<string, AgentConfig> {
   const result: Record<string, AgentConfig> = {}
 
   for (const [name, config] of Object.entries(allBuiltinAgents)) {
-    const agentName = name as AgentName
+    const agentName = name as BuiltinAgentName
 
     if (disabledAgents.includes(agentName)) {
       continue

src/auth/antigravity/constants.ts

@@ -0,0 +1,74 @@
+/**
+ * Antigravity OAuth configuration constants.
+ * Values sourced from cliproxyapi/sdk/auth/antigravity.go
+ *
+ * ## Logging Policy
+ *
+ * All console logging in antigravity modules follows a consistent policy:
+ *
+ * - **Debug logs**: Guard with `if (process.env.ANTIGRAVITY_DEBUG === "1")`
+ *   - Includes: info messages, warnings, non-fatal errors
+ *   - Enable debugging: `ANTIGRAVITY_DEBUG=1 opencode`
+ *
+ * - **Fatal errors**: None currently. All errors are handled by returning
+ *   appropriate error responses to OpenCode's auth system.
+ *
+ * This policy ensures production silence while enabling verbose debugging
+ * when needed for troubleshooting OAuth flows.
+ */
+
+// OAuth 2.0 Client Credentials
+export const ANTIGRAVITY_CLIENT_ID =
+  "1071006060591-tmhssin2h21lcre235vtolojh4g403ep.apps.googleusercontent.com"
+export const ANTIGRAVITY_CLIENT_SECRET = "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"
+
+// OAuth Callback
+export const ANTIGRAVITY_CALLBACK_PORT = 51121
+export const ANTIGRAVITY_REDIRECT_URI = `http://localhost:${ANTIGRAVITY_CALLBACK_PORT}/oauth-callback`
+
+// OAuth Scopes
+export const ANTIGRAVITY_SCOPES = [
+  "https://www.googleapis.com/auth/cloud-platform",
+  "https://www.googleapis.com/auth/userinfo.email",
+  "https://www.googleapis.com/auth/userinfo.profile",
+  "https://www.googleapis.com/auth/cclog",
+  "https://www.googleapis.com/auth/experimentsandconfigs",
+] as const
+
+// API Endpoint Fallbacks (order: daily → autopush → prod)
+export const ANTIGRAVITY_ENDPOINT_FALLBACKS = [
+  "https://daily-cloudcode-pa.sandbox.googleapis.com", // dev
+  "https://autopush-cloudcode-pa.sandbox.googleapis.com", // staging
+  "https://cloudcode-pa.googleapis.com", // prod
+] as const
+
+// API Version
+export const ANTIGRAVITY_API_VERSION = "v1internal"
+
+// Request Headers
+export const ANTIGRAVITY_HEADERS = {
+  "User-Agent": "google-api-nodejs-client/9.15.1",
+  "X-Goog-Api-Client": "google-cloud-sdk vscode_cloudshelleditor/0.1",
+  "Client-Metadata": JSON.stringify({
+    ideType: "IDE_UNSPECIFIED",
+    platform: "PLATFORM_UNSPECIFIED",
+    pluginType: "GEMINI",
+  }),
+} as const
+
+// Default Project ID (fallback when loadCodeAssist API fails)
+// From opencode-antigravity-auth reference implementation
+export const ANTIGRAVITY_DEFAULT_PROJECT_ID = "rising-fact-p41fc"
+
+
+
+// Google OAuth endpoints
+export const GOOGLE_AUTH_URL = "https://accounts.google.com/o/oauth2/v2/auth"
+export const GOOGLE_TOKEN_URL = "https://oauth2.googleapis.com/token"
+export const GOOGLE_USERINFO_URL = "https://www.googleapis.com/oauth2/v1/userinfo"
+
+// Token refresh buffer (refresh 60 seconds before expiry)
+export const ANTIGRAVITY_TOKEN_REFRESH_BUFFER_MS = 60_000
+
+// Default thought signature to skip validation (CLIProxyAPI approach)
+export const SKIP_THOUGHT_SIGNATURE_VALIDATOR = "skip_thought_signature_validator"

src/auth/antigravity/fetch.ts

@@ -0,0 +1,593 @@
+/**
+ * Antigravity Fetch Interceptor
+ *
+ * Creates a custom fetch function that:
+ * - Checks token expiration and auto-refreshes
+ * - Rewrites URLs to Antigravity endpoints
+ * - Applies request transformation (including tool normalization)
+ * - Applies response transformation (including thinking extraction)
+ * - Implements endpoint fallback (daily → autopush → prod)
+ *
+ * **Body Type Assumption:**
+ * This interceptor assumes `init.body` is a JSON string (OpenAI format).
+ * Non-string bodies (ReadableStream, Blob, FormData, URLSearchParams, etc.)
+ * are passed through unchanged to the original fetch to avoid breaking
+ * other requests that may not be OpenAI-format API calls.
+ *
+ * Debug logging available via ANTIGRAVITY_DEBUG=1 environment variable.
+ */
+
+import { ANTIGRAVITY_ENDPOINT_FALLBACKS, ANTIGRAVITY_DEFAULT_PROJECT_ID } from "./constants"
+import { fetchProjectContext, clearProjectContextCache } from "./project"
+import { isTokenExpired, refreshAccessToken, parseStoredToken, formatTokenForStorage } from "./token"
+import { transformRequest } from "./request"
+import { convertRequestBody, hasOpenAIMessages } from "./message-converter"
+import {
+  transformResponse,
+  transformStreamingResponse,
+  isStreamingResponse,
+  extractSignatureFromSsePayload,
+} from "./response"
+import { normalizeToolsForGemini, type OpenAITool } from "./tools"
+import { extractThinkingBlocks, shouldIncludeThinking, transformResponseThinking } from "./thinking"
+import {
+  getThoughtSignature,
+  setThoughtSignature,
+  getOrCreateSessionId,
+} from "./thought-signature-store"
+import type { AntigravityTokens } from "./types"
+
+/**
+ * Auth interface matching OpenCode's auth system
+ */
+interface Auth {
+  access?: string
+  refresh?: string
+  expires?: number
+}
+
+/**
+ * Client interface for auth operations
+ */
+interface AuthClient {
+  set(providerId: string, auth: Auth): Promise<void>
+}
+
+/**
+ * Debug logging helper
+ * Only logs when ANTIGRAVITY_DEBUG=1
+ */
+function debugLog(message: string): void {
+  if (process.env.ANTIGRAVITY_DEBUG === "1") {
+    console.log(`[antigravity-fetch] ${message}`)
+  }
+}
+
+function isRetryableError(status: number): boolean {
+  if (status === 0) return true
+  if (status === 429) return true
+  if (status >= 500 && status < 600) return true
+  return false
+}
+
+const GCP_PERMISSION_ERROR_PATTERNS = [
+  "PERMISSION_DENIED",
+  "does not have permission",
+  "Cloud AI Companion API has not been used",
+  "has not been enabled",
+] as const
+
+function isGcpPermissionError(text: string): boolean {
+  return GCP_PERMISSION_ERROR_PATTERNS.some((pattern) => text.includes(pattern))
+}
+
+function calculateRetryDelay(attempt: number): number {
+  return Math.min(200 * Math.pow(2, attempt), 2000)
+}
+
+async function isRetryableResponse(response: Response): Promise<boolean> {
+  if (isRetryableError(response.status)) return true
+  if (response.status === 403) {
+    try {
+      const text = await response.clone().text()
+      if (text.includes("SUBSCRIPTION_REQUIRED") || text.includes("Gemini Code Assist license")) {
+        debugLog(`[RETRY] 403 SUBSCRIPTION_REQUIRED detected, will retry with next endpoint`)
+        return true
+      }
+    } catch {}
+  }
+  return false
+}
+
+interface AttemptFetchOptions {
+  endpoint: string
+  url: string
+  init: RequestInit
+  accessToken: string
+  projectId: string
+  sessionId: string
+  modelName?: string
+  thoughtSignature?: string
+}
+
+type AttemptFetchResult = Response | null | "pass-through" | "needs-refresh"
+
+async function attemptFetch(
+  options: AttemptFetchOptions
+): Promise<AttemptFetchResult> {
+  const { endpoint, url, init, accessToken, projectId, sessionId, modelName, thoughtSignature } =
+    options
+  debugLog(`Trying endpoint: ${endpoint}`)
+
+  try {
+    const rawBody = init.body
+
+    if (rawBody !== undefined && typeof rawBody !== "string") {
+      debugLog(`Non-string body detected (${typeof rawBody}), signaling pass-through`)
+      return "pass-through"
+    }
+
+    let parsedBody: Record<string, unknown> = {}
+    if (rawBody) {
+      try {
+        parsedBody = JSON.parse(rawBody) as Record<string, unknown>
+      } catch {
+        parsedBody = {}
+      }
+    }
+
+    debugLog(`[BODY] Keys: ${Object.keys(parsedBody).join(", ")}`)
+    debugLog(`[BODY] Has contents: ${!!parsedBody.contents}, Has messages: ${!!parsedBody.messages}`)
+    if (parsedBody.contents) {
+      const contents = parsedBody.contents as Array<Record<string, unknown>>
+      debugLog(`[BODY] contents length: ${contents.length}`)
+      contents.forEach((c, i) => {
+        debugLog(`[BODY] contents[${i}].role: ${c.role}, parts: ${JSON.stringify(c.parts).substring(0, 200)}`)
+      })
+    }
+
+    if (parsedBody.tools && Array.isArray(parsedBody.tools)) {
+      const normalizedTools = normalizeToolsForGemini(parsedBody.tools as OpenAITool[])
+      if (normalizedTools) {
+        parsedBody.tools = normalizedTools
+      }
+    }
+
+    if (hasOpenAIMessages(parsedBody)) {
+      debugLog(`[CONVERT] Converting OpenAI messages to Gemini contents`)
+      parsedBody = convertRequestBody(parsedBody, thoughtSignature)
+      debugLog(`[CONVERT] After conversion - Has contents: ${!!parsedBody.contents}`)
+    }
+
+    const transformed = transformRequest({
+      url,
+      body: parsedBody,
+      accessToken,
+      projectId,
+      sessionId,
+      modelName,
+      endpointOverride: endpoint,
+      thoughtSignature,
+    })
+
+    debugLog(`[REQ] streaming=${transformed.streaming}, url=${transformed.url}`)
+
+    const maxPermissionRetries = 10
+    for (let attempt = 0; attempt <= maxPermissionRetries; attempt++) {
+      const response = await fetch(transformed.url, {
+        method: init.method || "POST",
+        headers: transformed.headers,
+        body: JSON.stringify(transformed.body),
+        signal: init.signal,
+      })
+
+      debugLog(
+        `[RESP] status=${response.status} content-type=${response.headers.get("content-type") ?? ""} url=${response.url}`
+      )
+
+      if (response.status === 401) {
+        debugLog(`[401] Unauthorized response detected, signaling token refresh needed`)
+        return "needs-refresh"
+      }
+
+      if (response.status === 403) {
+        try {
+          const text = await response.clone().text()
+          if (isGcpPermissionError(text)) {
+            if (attempt < maxPermissionRetries) {
+              const delay = calculateRetryDelay(attempt)
+              debugLog(`[RETRY] GCP permission error, retry ${attempt + 1}/${maxPermissionRetries} after ${delay}ms`)
+              await new Promise((resolve) => setTimeout(resolve, delay))
+              continue
+            }
+            debugLog(`[RETRY] GCP permission error, max retries exceeded`)
+          }
+        } catch {}
+      }
+
+      if (!response.ok && (await isRetryableResponse(response))) {
+        debugLog(`Endpoint failed: ${endpoint} (status: ${response.status}), trying next`)
+        return null
+      }
+
+      return response
+    }
+
+    return null
+  } catch (error) {
+    debugLog(
+      `Endpoint failed: ${endpoint} (${error instanceof Error ? error.message : "Unknown error"}), trying next`
+    )
+    return null
+  }
+}
+
+interface GeminiResponsePart {
+  thoughtSignature?: string
+  thought_signature?: string
+  functionCall?: Record<string, unknown>
+  text?: string
+  [key: string]: unknown
+}
+
+interface GeminiResponseCandidate {
+  content?: {
+    parts?: GeminiResponsePart[]
+    [key: string]: unknown
+  }
+  [key: string]: unknown
+}
+
+interface GeminiResponseBody {
+  candidates?: GeminiResponseCandidate[]
+  [key: string]: unknown
+}
+
+function extractSignatureFromResponse(parsed: GeminiResponseBody): string | undefined {
+  if (!parsed.candidates || !Array.isArray(parsed.candidates)) {
+    return undefined
+  }
+
+  for (const candidate of parsed.candidates) {
+    const parts = candidate.content?.parts
+    if (!parts || !Array.isArray(parts)) {
+      continue
+    }
+
+    for (const part of parts) {
+      const sig = part.thoughtSignature || part.thought_signature
+      if (sig && typeof sig === "string") {
+        return sig
+      }
+    }
+  }
+
+  return undefined
+}
+
+async function transformResponseWithThinking(
+  response: Response,
+  modelName: string,
+  fetchInstanceId: string
+): Promise<Response> {
+  const streaming = isStreamingResponse(response)
+
+  let result
+  if (streaming) {
+    result = await transformStreamingResponse(response)
+  } else {
+    result = await transformResponse(response)
+  }
+
+  if (streaming) {
+    return result.response
+  }
+
+  try {
+    const text = await result.response.clone().text()
+    debugLog(`[TSIG][RESP] Response text length: ${text.length}`)
+
+    const parsed = JSON.parse(text) as GeminiResponseBody
+    debugLog(`[TSIG][RESP] Parsed keys: ${Object.keys(parsed).join(", ")}`)
+    debugLog(`[TSIG][RESP] Has candidates: ${!!parsed.candidates}, count: ${parsed.candidates?.length ?? 0}`)
+
+    const signature = extractSignatureFromResponse(parsed)
+    debugLog(`[TSIG][RESP] Signature extracted: ${signature ? signature.substring(0, 30) + "..." : "NONE"}`)
+    if (signature) {
+      setThoughtSignature(fetchInstanceId, signature)
+      debugLog(`[TSIG][STORE] Stored signature for ${fetchInstanceId}`)
+    } else {
+      debugLog(`[TSIG][WARN] No signature found in response!`)
+    }
+
+    if (shouldIncludeThinking(modelName)) {
+      const thinkingResult = extractThinkingBlocks(parsed)
+      if (thinkingResult.hasThinking) {
+        const transformed = transformResponseThinking(parsed)
+        return new Response(JSON.stringify(transformed), {
+          status: result.response.status,
+          statusText: result.response.statusText,
+          headers: result.response.headers,
+        })
+      }
+    }
+  } catch {}
+
+  return result.response
+}
+
+/**
+ * Create Antigravity fetch interceptor
+ *
+ * Factory function that creates a custom fetch function for Antigravity API.
+ * Handles token management, request/response transformation, and endpoint fallback.
+ *
+ * @param getAuth - Async function to retrieve current auth state
+ * @param client - Auth client for saving updated tokens
+ * @param providerId - Provider identifier (e.g., "google")
+ * @param clientId - Optional custom client ID for token refresh (defaults to ANTIGRAVITY_CLIENT_ID)
+ * @param clientSecret - Optional custom client secret for token refresh (defaults to ANTIGRAVITY_CLIENT_SECRET)
+ * @returns Custom fetch function compatible with standard fetch signature
+ *
+ * @example
+ * ```typescript
+ * const customFetch = createAntigravityFetch(
+ *   () => auth(),
+ *   client,
+ *   "google",
+ *   "custom-client-id",
+ *   "custom-client-secret"
+ * )
+ *
+ * // Use like standard fetch
+ * const response = await customFetch("https://api.example.com/chat", {
+ *   method: "POST",
+ *   body: JSON.stringify({ messages: [...] })
+ * })
+ * ```
+ */
+export function createAntigravityFetch(
+  getAuth: () => Promise<Auth>,
+  client: AuthClient,
+  providerId: string,
+  clientId?: string,
+  clientSecret?: string
+): (url: string, init?: RequestInit) => Promise<Response> {
+  let cachedTokens: AntigravityTokens | null = null
+  let cachedProjectId: string | null = null
+  const fetchInstanceId = crypto.randomUUID()
+
+  return async (url: string, init: RequestInit = {}): Promise<Response> => {
+    debugLog(`Intercepting request to: ${url}`)
+
+    // Get current auth state
+    const auth = await getAuth()
+    if (!auth.access || !auth.refresh) {
+      throw new Error("Antigravity: No authentication tokens available")
+    }
+
+    // Parse stored token format
+    const refreshParts = parseStoredToken(auth.refresh)
+
+    // Build initial token state
+    if (!cachedTokens) {
+      cachedTokens = {
+        type: "antigravity",
+        access_token: auth.access,
+        refresh_token: refreshParts.refreshToken,
+        expires_in: auth.expires ? Math.floor((auth.expires - Date.now()) / 1000) : 3600,
+        timestamp: auth.expires ? auth.expires - 3600 * 1000 : Date.now(),
+      }
+    } else {
+      // Update with fresh values
+      cachedTokens.access_token = auth.access
+      cachedTokens.refresh_token = refreshParts.refreshToken
+    }
+
+    // Check token expiration and refresh if needed
+    if (isTokenExpired(cachedTokens)) {
+      debugLog("Token expired, refreshing...")
+
+      try {
+        const newTokens = await refreshAccessToken(refreshParts.refreshToken, clientId, clientSecret)
+
+        // Update cached tokens
+        cachedTokens = {
+          type: "antigravity",
+          access_token: newTokens.access_token,
+          refresh_token: newTokens.refresh_token,
+          expires_in: newTokens.expires_in,
+          timestamp: Date.now(),
+        }
+
+        // Clear project context cache on token refresh
+        clearProjectContextCache()
+
+        // Format and save new tokens
+        const formattedRefresh = formatTokenForStorage(
+          newTokens.refresh_token,
+          refreshParts.projectId || "",
+          refreshParts.managedProjectId
+        )
+
+        await client.set(providerId, {
+          access: newTokens.access_token,
+          refresh: formattedRefresh,
+          expires: Date.now() + newTokens.expires_in * 1000,
+        })
+
+        debugLog("Token refreshed successfully")
+      } catch (error) {
+        throw new Error(
+          `Antigravity: Token refresh failed: ${error instanceof Error ? error.message : "Unknown error"}`
+        )
+      }
+    }
+
+    // Fetch project ID via loadCodeAssist (CLIProxyAPI approach)
+    if (!cachedProjectId) {
+      const projectContext = await fetchProjectContext(cachedTokens.access_token)
+      cachedProjectId = projectContext.cloudaicompanionProject || ""
+      debugLog(`[PROJECT] Fetched project ID: "${cachedProjectId}"`)
+    }
+
+    const projectId = cachedProjectId
+    debugLog(`[PROJECT] Using project ID: "${projectId}"`)
+
+    // Extract model name from request body
+    let modelName: string | undefined
+    if (init.body) {
+      try {
+        const body =
+          typeof init.body === "string"
+            ? (JSON.parse(init.body) as Record<string, unknown>)
+            : (init.body as unknown as Record<string, unknown>)
+        if (typeof body.model === "string") {
+          modelName = body.model
+        }
+      } catch {
+        // Ignore parsing errors
+      }
+    }
+
+    const maxEndpoints = Math.min(ANTIGRAVITY_ENDPOINT_FALLBACKS.length, 3)
+    const sessionId = getOrCreateSessionId(fetchInstanceId)
+    const thoughtSignature = getThoughtSignature(fetchInstanceId)
+    debugLog(`[TSIG][GET] sessionId=${sessionId}, signature=${thoughtSignature ? thoughtSignature.substring(0, 20) + "..." : "none"}`)
+
+    let hasRefreshedFor401 = false
+
+    const executeWithEndpoints = async (): Promise<Response> => {
+      for (let i = 0; i < maxEndpoints; i++) {
+        const endpoint = ANTIGRAVITY_ENDPOINT_FALLBACKS[i]
+
+        const response = await attemptFetch({
+          endpoint,
+          url,
+          init,
+          accessToken: cachedTokens!.access_token,
+          projectId,
+          sessionId,
+          modelName,
+          thoughtSignature,
+        })
+
+        if (response === "pass-through") {
+          debugLog("Non-string body detected, passing through with auth headers")
+          const headersWithAuth = {
+            ...init.headers,
+            Authorization: `Bearer ${cachedTokens!.access_token}`,
+          }
+          return fetch(url, { ...init, headers: headersWithAuth })
+        }
+
+        if (response === "needs-refresh") {
+          if (hasRefreshedFor401) {
+            debugLog("[401] Already refreshed once, returning unauthorized error")
+            return new Response(
+              JSON.stringify({
+                error: {
+                  message: "Authentication failed after token refresh",
+                  type: "unauthorized",
+                  code: "token_refresh_failed",
+                },
+              }),
+              {
+                status: 401,
+                statusText: "Unauthorized",
+                headers: { "Content-Type": "application/json" },
+              }
+            )
+          }
+
+          debugLog("[401] Refreshing token and retrying...")
+          hasRefreshedFor401 = true
+
+          try {
+            const newTokens = await refreshAccessToken(
+              refreshParts.refreshToken,
+              clientId,
+              clientSecret
+            )
+
+            cachedTokens = {
+              type: "antigravity",
+              access_token: newTokens.access_token,
+              refresh_token: newTokens.refresh_token,
+              expires_in: newTokens.expires_in,
+              timestamp: Date.now(),
+            }
+
+            clearProjectContextCache()
+
+            const formattedRefresh = formatTokenForStorage(
+              newTokens.refresh_token,
+              refreshParts.projectId || "",
+              refreshParts.managedProjectId
+            )
+
+            await client.set(providerId, {
+              access: newTokens.access_token,
+              refresh: formattedRefresh,
+              expires: Date.now() + newTokens.expires_in * 1000,
+            })
+
+            debugLog("[401] Token refreshed, retrying request...")
+            return executeWithEndpoints()
+          } catch (refreshError) {
+            debugLog(`[401] Token refresh failed: ${refreshError instanceof Error ? refreshError.message : "Unknown error"}`)
+            return new Response(
+              JSON.stringify({
+                error: {
+                  message: `Token refresh failed: ${refreshError instanceof Error ? refreshError.message : "Unknown error"}`,
+                  type: "unauthorized",
+                  code: "token_refresh_failed",
+                },
+              }),
+              {
+                status: 401,
+                statusText: "Unauthorized",
+                headers: { "Content-Type": "application/json" },
+              }
+            )
+          }
+        }
+
+        if (response) {
+          debugLog(`Success with endpoint: ${endpoint}`)
+          const transformedResponse = await transformResponseWithThinking(
+            response,
+            modelName || "",
+            fetchInstanceId
+          )
+          return transformedResponse
+        }
+      }
+
+      const errorMessage = `All Antigravity endpoints failed after ${maxEndpoints} attempts`
+      debugLog(errorMessage)
+
+      return new Response(
+        JSON.stringify({
+          error: {
+            message: errorMessage,
+            type: "endpoint_failure",
+            code: "all_endpoints_failed",
+          },
+        }),
+        {
+          status: 503,
+          statusText: "Service Unavailable",
+          headers: { "Content-Type": "application/json" },
+        }
+      )
+    }
+
+    return executeWithEndpoints()
+  }
+}
+
+/**
+ * Type export for createAntigravityFetch return type
+ */
+export type AntigravityFetch = (url: string, init?: RequestInit) => Promise<Response>

src/auth/antigravity/index.ts

@@ -0,0 +1,13 @@
+export * from "./types"
+export * from "./constants"
+export * from "./oauth"
+export * from "./token"
+export * from "./project"
+export * from "./request"
+export * from "./response"
+export * from "./tools"
+export * from "./thinking"
+export * from "./thought-signature-store"
+export * from "./message-converter"
+export * from "./fetch"
+export * from "./plugin"

src/auth/antigravity/message-converter.ts

@@ -0,0 +1,206 @@
+/**
+ * OpenAI → Gemini message format converter
+ * 
+ * Converts OpenAI-style messages to Gemini contents format,
+ * injecting thoughtSignature into functionCall parts.
+ */
+
+import { SKIP_THOUGHT_SIGNATURE_VALIDATOR } from "./constants"
+
+function debugLog(message: string): void {
+  if (process.env.ANTIGRAVITY_DEBUG === "1") {
+    console.log(`[antigravity-converter] ${message}`)
+  }
+}
+
+interface OpenAIMessage {
+  role: "system" | "user" | "assistant" | "tool"
+  content?: string | OpenAIContentPart[]
+  tool_calls?: OpenAIToolCall[]
+  tool_call_id?: string
+  name?: string
+}
+
+interface OpenAIContentPart {
+  type: string
+  text?: string
+  image_url?: { url: string }
+  [key: string]: unknown
+}
+
+interface OpenAIToolCall {
+  id: string
+  type: "function"
+  function: {
+    name: string
+    arguments: string
+  }
+}
+
+interface GeminiPart {
+  text?: string
+  functionCall?: {
+    name: string
+    args: Record<string, unknown>
+  }
+  functionResponse?: {
+    name: string
+    response: Record<string, unknown>
+  }
+  inlineData?: {
+    mimeType: string
+    data: string
+  }
+  thought_signature?: string
+  [key: string]: unknown
+}
+
+interface GeminiContent {
+  role: "user" | "model"
+  parts: GeminiPart[]
+}
+
+export function convertOpenAIToGemini(
+  messages: OpenAIMessage[],
+  thoughtSignature?: string
+): GeminiContent[] {
+  debugLog(`Converting ${messages.length} messages, signature: ${thoughtSignature ? "present" : "none"}`)
+  
+  const contents: GeminiContent[] = []
+
+  for (const msg of messages) {
+    if (msg.role === "system") {
+      contents.push({
+        role: "user",
+        parts: [{ text: typeof msg.content === "string" ? msg.content : "" }],
+      })
+      continue
+    }
+
+    if (msg.role === "user") {
+      const parts = convertContentToParts(msg.content)
+      contents.push({ role: "user", parts })
+      continue
+    }
+
+    if (msg.role === "assistant") {
+      const parts: GeminiPart[] = []
+
+      if (msg.content) {
+        parts.push(...convertContentToParts(msg.content))
+      }
+
+      if (msg.tool_calls && msg.tool_calls.length > 0) {
+        for (const toolCall of msg.tool_calls) {
+          let args: Record<string, unknown> = {}
+          try {
+            args = JSON.parse(toolCall.function.arguments)
+          } catch {
+            args = {}
+          }
+
+          const part: GeminiPart = {
+            functionCall: {
+              name: toolCall.function.name,
+              args,
+            },
+          }
+
+          // Always inject signature: use provided or default to skip validator (CLIProxyAPI approach)
+          part.thoughtSignature = thoughtSignature || SKIP_THOUGHT_SIGNATURE_VALIDATOR
+          debugLog(`Injected signature into functionCall: ${toolCall.function.name} (${thoughtSignature ? "provided" : "default"})`)
+
+          parts.push(part)
+        }
+      }
+
+      if (parts.length > 0) {
+        contents.push({ role: "model", parts })
+      }
+      continue
+    }
+
+    if (msg.role === "tool") {
+      let response: Record<string, unknown> = {}
+      try {
+        response = typeof msg.content === "string" 
+          ? JSON.parse(msg.content) 
+          : { result: msg.content }
+      } catch {
+        response = { result: msg.content }
+      }
+
+      const toolName = msg.name || "unknown"
+      
+      contents.push({
+        role: "user",
+        parts: [{
+          functionResponse: {
+            name: toolName,
+            response,
+          },
+        }],
+      })
+      continue
+    }
+  }
+
+  debugLog(`Converted to ${contents.length} content blocks`)
+  return contents
+}
+
+function convertContentToParts(content: string | OpenAIContentPart[] | undefined): GeminiPart[] {
+  if (!content) {
+    return [{ text: "" }]
+  }
+
+  if (typeof content === "string") {
+    return [{ text: content }]
+  }
+
+  const parts: GeminiPart[] = []
+  for (const part of content) {
+    if (part.type === "text" && part.text) {
+      parts.push({ text: part.text })
+    } else if (part.type === "image_url" && part.image_url?.url) {
+      const url = part.image_url.url
+      if (url.startsWith("data:")) {
+        const match = url.match(/^data:([^;]+);base64,(.+)$/)
+        if (match) {
+          parts.push({
+            inlineData: {
+              mimeType: match[1],
+              data: match[2],
+            },
+          })
+        }
+      }
+    }
+  }
+
+  return parts.length > 0 ? parts : [{ text: "" }]
+}
+
+export function hasOpenAIMessages(body: Record<string, unknown>): boolean {
+  return Array.isArray(body.messages) && body.messages.length > 0
+}
+
+export function convertRequestBody(
+  body: Record<string, unknown>,
+  thoughtSignature?: string
+): Record<string, unknown> {
+  if (!hasOpenAIMessages(body)) {
+    debugLog("No messages array found, returning body as-is")
+    return body
+  }
+
+  const messages = body.messages as OpenAIMessage[]
+  const contents = convertOpenAIToGemini(messages, thoughtSignature)
+
+  const converted = { ...body }
+  delete converted.messages
+  converted.contents = contents
+
+  debugLog(`Converted body: messages → contents (${contents.length} blocks)`)
+  return converted
+}

src/auth/antigravity/oauth.ts

@@ -0,0 +1,361 @@
+/**
+ * Antigravity OAuth 2.0 flow implementation with PKCE.
+ * Handles Google OAuth for Antigravity authentication.
+ */
+import { generatePKCE } from "@openauthjs/openauth/pkce"
+
+import {
+  ANTIGRAVITY_CLIENT_ID,
+  ANTIGRAVITY_CLIENT_SECRET,
+  ANTIGRAVITY_REDIRECT_URI,
+  ANTIGRAVITY_SCOPES,
+  ANTIGRAVITY_CALLBACK_PORT,
+  GOOGLE_AUTH_URL,
+  GOOGLE_TOKEN_URL,
+  GOOGLE_USERINFO_URL,
+} from "./constants"
+import type {
+  AntigravityTokenExchangeResult,
+  AntigravityUserInfo,
+} from "./types"
+
+/**
+ * PKCE pair containing verifier and challenge.
+ */
+export interface PKCEPair {
+  /** PKCE verifier - used during token exchange */
+  verifier: string
+  /** PKCE challenge - sent in auth URL */
+  challenge: string
+  /** Challenge method - always "S256" */
+  method: string
+}
+
+/**
+ * OAuth state encoded in the auth URL.
+ * Contains the PKCE verifier for later retrieval.
+ */
+export interface OAuthState {
+  /** PKCE verifier */
+  verifier: string
+  /** Optional project ID */
+  projectId?: string
+}
+
+/**
+ * Result from building an OAuth authorization URL.
+ */
+export interface AuthorizationResult {
+  /** Full OAuth URL to open in browser */
+  url: string
+  /** PKCE verifier to use during code exchange */
+  verifier: string
+}
+
+/**
+ * Result from the OAuth callback server.
+ */
+export interface CallbackResult {
+  /** Authorization code from Google */
+  code: string
+  /** State parameter from callback */
+  state: string
+  /** Error message if any */
+  error?: string
+}
+
+/**
+ * Generate PKCE verifier and challenge pair.
+ * Uses @openauthjs/openauth for cryptographically secure generation.
+ *
+ * @returns PKCE pair with verifier, challenge, and method
+ */
+export async function generatePKCEPair(): Promise<PKCEPair> {
+  const pkce = await generatePKCE()
+  return {
+    verifier: pkce.verifier,
+    challenge: pkce.challenge,
+    method: pkce.method,
+  }
+}
+
+/**
+ * Encode OAuth state into a URL-safe base64 string.
+ *
+ * @param state - OAuth state object
+ * @returns Base64URL encoded state
+ */
+function encodeState(state: OAuthState): string {
+  const json = JSON.stringify(state)
+  return Buffer.from(json, "utf8").toString("base64url")
+}
+
+/**
+ * Decode OAuth state from a base64 string.
+ *
+ * @param encoded - Base64URL or Base64 encoded state
+ * @returns Decoded OAuth state
+ */
+export function decodeState(encoded: string): OAuthState {
+  // Handle both base64url and standard base64
+  const normalized = encoded.replace(/-/g, "+").replace(/_/g, "/")
+  const padded = normalized.padEnd(
+    normalized.length + ((4 - (normalized.length % 4)) % 4),
+    "="
+  )
+  const json = Buffer.from(padded, "base64").toString("utf8")
+  const parsed = JSON.parse(json)
+
+  if (typeof parsed.verifier !== "string") {
+    throw new Error("Missing PKCE verifier in state")
+  }
+
+  return {
+    verifier: parsed.verifier,
+    projectId:
+      typeof parsed.projectId === "string" ? parsed.projectId : undefined,
+  }
+}
+
+export async function buildAuthURL(
+  projectId?: string,
+  clientId: string = ANTIGRAVITY_CLIENT_ID,
+  port: number = ANTIGRAVITY_CALLBACK_PORT
+): Promise<AuthorizationResult> {
+  const pkce = await generatePKCEPair()
+
+  const state: OAuthState = {
+    verifier: pkce.verifier,
+    projectId,
+  }
+
+  const redirectUri = `http://localhost:${port}/oauth-callback`
+
+  const url = new URL(GOOGLE_AUTH_URL)
+  url.searchParams.set("client_id", clientId)
+  url.searchParams.set("redirect_uri", redirectUri)
+  url.searchParams.set("response_type", "code")
+  url.searchParams.set("scope", ANTIGRAVITY_SCOPES.join(" "))
+  url.searchParams.set("state", encodeState(state))
+  url.searchParams.set("code_challenge", pkce.challenge)
+  url.searchParams.set("code_challenge_method", "S256")
+  url.searchParams.set("access_type", "offline")
+  url.searchParams.set("prompt", "consent")
+
+  return {
+    url: url.toString(),
+    verifier: pkce.verifier,
+  }
+}
+
+/**
+ * Exchange authorization code for tokens.
+ *
+ * @param code - Authorization code from OAuth callback
+ * @param verifier - PKCE verifier from initial auth request
+ * @param clientId - Optional custom client ID (defaults to ANTIGRAVITY_CLIENT_ID)
+ * @param clientSecret - Optional custom client secret (defaults to ANTIGRAVITY_CLIENT_SECRET)
+ * @returns Token exchange result with access and refresh tokens
+ */
+export async function exchangeCode(
+  code: string,
+  verifier: string,
+  clientId: string = ANTIGRAVITY_CLIENT_ID,
+  clientSecret: string = ANTIGRAVITY_CLIENT_SECRET,
+  port: number = ANTIGRAVITY_CALLBACK_PORT
+): Promise<AntigravityTokenExchangeResult> {
+  const redirectUri = `http://localhost:${port}/oauth-callback`
+  const params = new URLSearchParams({
+    client_id: clientId,
+    client_secret: clientSecret,
+    code,
+    grant_type: "authorization_code",
+    redirect_uri: redirectUri,
+    code_verifier: verifier,
+  })
+
+  const response = await fetch(GOOGLE_TOKEN_URL, {
+    method: "POST",
+    headers: {
+      "Content-Type": "application/x-www-form-urlencoded",
+    },
+    body: params,
+  })
+
+  if (!response.ok) {
+    const errorText = await response.text()
+    throw new Error(`Token exchange failed: ${response.status} - ${errorText}`)
+  }
+
+  const data = (await response.json()) as {
+    access_token: string
+    refresh_token: string
+    expires_in: number
+    token_type: string
+  }
+
+  return {
+    access_token: data.access_token,
+    refresh_token: data.refresh_token,
+    expires_in: data.expires_in,
+    token_type: data.token_type,
+  }
+}
+
+/**
+ * Fetch user info from Google's userinfo API.
+ *
+ * @param accessToken - Valid access token
+ * @returns User info containing email
+ */
+export async function fetchUserInfo(
+  accessToken: string
+): Promise<AntigravityUserInfo> {
+  const response = await fetch(`${GOOGLE_USERINFO_URL}?alt=json`, {
+    headers: {
+      Authorization: `Bearer ${accessToken}`,
+    },
+  })
+
+  if (!response.ok) {
+    throw new Error(`Failed to fetch user info: ${response.status}`)
+  }
+
+  const data = (await response.json()) as {
+    email?: string
+    name?: string
+    picture?: string
+  }
+
+  return {
+    email: data.email || "",
+    name: data.name,
+    picture: data.picture,
+  }
+}
+
+export interface CallbackServerHandle {
+  port: number
+  waitForCallback: () => Promise<CallbackResult>
+  close: () => void
+}
+
+export function startCallbackServer(
+  timeoutMs: number = 5 * 60 * 1000
+): CallbackServerHandle {
+  let server: ReturnType<typeof Bun.serve> | null = null
+  let timeoutId: ReturnType<typeof setTimeout> | null = null
+  let resolveCallback: ((result: CallbackResult) => void) | null = null
+  let rejectCallback: ((error: Error) => void) | null = null
+
+  const cleanup = () => {
+    if (timeoutId) {
+      clearTimeout(timeoutId)
+      timeoutId = null
+    }
+    if (server) {
+      server.stop()
+      server = null
+    }
+  }
+
+  server = Bun.serve({
+    port: 0,
+    fetch(request: Request): Response {
+      const url = new URL(request.url)
+
+      if (url.pathname === "/oauth-callback") {
+        const code = url.searchParams.get("code") || ""
+        const state = url.searchParams.get("state") || ""
+        const error = url.searchParams.get("error") || undefined
+
+        let responseBody: string
+        if (code && !error) {
+          responseBody =
+            "<html><body><h1>Login successful</h1><p>You can close this window.</p></body></html>"
+        } else {
+          responseBody =
+            "<html><body><h1>Login failed</h1><p>Please check the CLI output.</p></body></html>"
+        }
+
+        setTimeout(() => {
+          cleanup()
+          if (resolveCallback) {
+            resolveCallback({ code, state, error })
+          }
+        }, 100)
+
+        return new Response(responseBody, {
+          status: 200,
+          headers: { "Content-Type": "text/html" },
+        })
+      }
+
+      return new Response("Not Found", { status: 404 })
+    },
+  })
+
+  const actualPort = server.port as number
+
+  const waitForCallback = (): Promise<CallbackResult> => {
+    return new Promise((resolve, reject) => {
+      resolveCallback = resolve
+      rejectCallback = reject
+
+      timeoutId = setTimeout(() => {
+        cleanup()
+        reject(new Error("OAuth callback timeout"))
+      }, timeoutMs)
+    })
+  }
+
+  return {
+    port: actualPort,
+    waitForCallback,
+    close: cleanup,
+  }
+}
+
+export async function performOAuthFlow(
+  projectId?: string,
+  openBrowser?: (url: string) => Promise<void>,
+  clientId: string = ANTIGRAVITY_CLIENT_ID,
+  clientSecret: string = ANTIGRAVITY_CLIENT_SECRET
+): Promise<{
+  tokens: AntigravityTokenExchangeResult
+  userInfo: AntigravityUserInfo
+  verifier: string
+}> {
+  const serverHandle = startCallbackServer()
+
+  try {
+    const auth = await buildAuthURL(projectId, clientId, serverHandle.port)
+
+    if (openBrowser) {
+      await openBrowser(auth.url)
+    }
+
+    const callback = await serverHandle.waitForCallback()
+
+    if (callback.error) {
+      throw new Error(`OAuth error: ${callback.error}`)
+    }
+
+    if (!callback.code) {
+      throw new Error("No authorization code received")
+    }
+
+    const state = decodeState(callback.state)
+    if (state.verifier !== auth.verifier) {
+      throw new Error("PKCE verifier mismatch - possible CSRF attack")
+    }
+
+    const tokens = await exchangeCode(callback.code, auth.verifier, clientId, clientSecret, serverHandle.port)
+    const userInfo = await fetchUserInfo(tokens.access_token)
+
+    return { tokens, userInfo, verifier: auth.verifier }
+  } catch (err) {
+    serverHandle.close()
+    throw err
+  }
+}

src/auth/antigravity/plugin.ts

@@ -0,0 +1,295 @@
+/**
+ * Google Antigravity Auth Plugin for OpenCode
+ *
+ * Provides OAuth authentication for Google models via Antigravity API.
+ * This plugin integrates with OpenCode's auth system to enable:
+ * - OAuth 2.0 with PKCE flow for Google authentication
+ * - Automatic token refresh
+ * - Request/response transformation for Antigravity API
+ *
+ * @example
+ * ```json
+ * // opencode.json
+ * {
+ *   "plugin": ["oh-my-opencode"],
+ *   "provider": {
+ *     "google": {
+ *       "options": {
+ *         "clientId": "custom-client-id",
+ *         "clientSecret": "custom-client-secret"
+ *       }
+ *     }
+ *   }
+ * }
+ * ```
+ */
+
+import type { Auth, Provider } from "@opencode-ai/sdk"
+import type { AuthHook, AuthOuathResult, PluginInput } from "@opencode-ai/plugin"
+
+import { ANTIGRAVITY_CLIENT_ID, ANTIGRAVITY_CLIENT_SECRET } from "./constants"
+import {
+  buildAuthURL,
+  exchangeCode,
+  startCallbackServer,
+  fetchUserInfo,
+  decodeState,
+} from "./oauth"
+import { createAntigravityFetch } from "./fetch"
+import { fetchProjectContext } from "./project"
+import { formatTokenForStorage } from "./token"
+
+/**
+ * Provider ID for Google models
+ * Antigravity is an auth method for Google, not a separate provider
+ */
+const GOOGLE_PROVIDER_ID = "google"
+
+/**
+ * Type guard to check if auth is OAuth type
+ */
+function isOAuthAuth(
+  auth: Auth
+): auth is { type: "oauth"; access: string; refresh: string; expires: number } {
+  return auth.type === "oauth"
+}
+
+/**
+ * Creates the Google Antigravity OAuth plugin for OpenCode.
+ *
+ * This factory function creates an auth plugin that:
+ * 1. Provides OAuth flow for Google authentication
+ * 2. Creates a custom fetch interceptor for Antigravity API
+ * 3. Handles token management and refresh
+ *
+ * @param input - Plugin input containing the OpenCode client
+ * @returns Hooks object with auth configuration
+ *
+ * @example
+ * ```typescript
+ * // Used by OpenCode automatically when plugin is loaded
+ * const hooks = await createGoogleAntigravityAuthPlugin({ client, ... })
+ * ```
+ */
+export async function createGoogleAntigravityAuthPlugin({
+  client,
+}: PluginInput): Promise<{ auth: AuthHook }> {
+  // Cache for custom credentials from provider.options
+  // These are populated by loader() and used by authorize()
+  // Falls back to defaults if loader hasn't been called yet
+  let cachedClientId: string = ANTIGRAVITY_CLIENT_ID
+  let cachedClientSecret: string = ANTIGRAVITY_CLIENT_SECRET
+
+  const authHook: AuthHook = {
+    /**
+     * Provider identifier - must be "google" as Antigravity is
+     * an auth method for Google models, not a separate provider
+     */
+    provider: GOOGLE_PROVIDER_ID,
+
+    /**
+     * Loader function called when auth is needed.
+     * Reads credentials from provider.options and creates custom fetch.
+     *
+     * @param auth - Function to retrieve current auth state
+     * @param provider - Provider configuration including options
+     * @returns Object with custom fetch function
+     */
+    loader: async (
+      auth: () => Promise<Auth>,
+      provider: Provider
+    ): Promise<Record<string, unknown>> => {
+      const currentAuth = await auth()
+      
+      if (process.env.ANTIGRAVITY_DEBUG === "1") {
+        console.log("[antigravity-plugin] loader called")
+        console.log("[antigravity-plugin] auth type:", currentAuth?.type)
+        console.log("[antigravity-plugin] auth keys:", Object.keys(currentAuth || {}))
+      }
+      
+      if (!isOAuthAuth(currentAuth)) {
+        if (process.env.ANTIGRAVITY_DEBUG === "1") {
+          console.log("[antigravity-plugin] NOT OAuth auth, returning empty")
+        }
+        return {}
+      }
+      
+      if (process.env.ANTIGRAVITY_DEBUG === "1") {
+        console.log("[antigravity-plugin] OAuth auth detected, creating custom fetch")
+      }
+
+      cachedClientId =
+        (provider.options?.clientId as string) || ANTIGRAVITY_CLIENT_ID
+      cachedClientSecret =
+        (provider.options?.clientSecret as string) || ANTIGRAVITY_CLIENT_SECRET
+
+      // Log if using custom credentials (for debugging)
+      if (
+        process.env.ANTIGRAVITY_DEBUG === "1" &&
+        (cachedClientId !== ANTIGRAVITY_CLIENT_ID ||
+          cachedClientSecret !== ANTIGRAVITY_CLIENT_SECRET)
+      ) {
+        console.log(
+          "[antigravity-plugin] Using custom credentials from provider.options"
+        )
+      }
+
+      // Create adapter for client.auth.set that matches fetch.ts AuthClient interface
+      const authClient = {
+        set: async (
+          providerId: string,
+          authData: { access?: string; refresh?: string; expires?: number }
+        ) => {
+          await client.auth.set({
+            body: {
+              type: "oauth",
+              access: authData.access || "",
+              refresh: authData.refresh || "",
+              expires: authData.expires || 0,
+            },
+            path: { id: providerId },
+          })
+        },
+      }
+
+      // Create auth getter that returns compatible format for fetch.ts
+      const getAuth = async (): Promise<{
+        access?: string
+        refresh?: string
+        expires?: number
+      }> => {
+        const authState = await auth()
+        if (isOAuthAuth(authState)) {
+          return {
+            access: authState.access,
+            refresh: authState.refresh,
+            expires: authState.expires,
+          }
+        }
+        return {}
+      }
+
+      const antigravityFetch = createAntigravityFetch(
+        getAuth,
+        authClient,
+        GOOGLE_PROVIDER_ID,
+        cachedClientId,
+        cachedClientSecret
+      )
+
+      return {
+        fetch: antigravityFetch,
+        apiKey: "antigravity-oauth",
+      }
+    },
+
+    /**
+     * Authentication methods available for this provider.
+     * Only OAuth is supported - no prompts for credentials.
+     */
+    methods: [
+      {
+        type: "oauth",
+        label: "OAuth with Google (Antigravity)",
+        // NO prompts - credentials come from provider.options or defaults
+        // OAuth flow starts immediately when user selects this method
+
+        /**
+         * Starts the OAuth authorization flow.
+         * Opens browser for Google OAuth and waits for callback.
+         *
+         * @returns Authorization result with URL and callback
+         */
+        authorize: async (): Promise<AuthOuathResult> => {
+          const serverHandle = startCallbackServer()
+          const { url, verifier } = await buildAuthURL(undefined, cachedClientId, serverHandle.port)
+
+          return {
+            url,
+            instructions:
+              "Complete the sign-in in your browser. We'll automatically detect when you're done.",
+            method: "auto",
+
+            callback: async () => {
+              try {
+                const result = await serverHandle.waitForCallback()
+
+                if (result.error) {
+                  if (process.env.ANTIGRAVITY_DEBUG === "1") {
+                    console.error(`[antigravity-plugin] OAuth error: ${result.error}`)
+                  }
+                  return { type: "failed" as const }
+                }
+
+                if (!result.code) {
+                  if (process.env.ANTIGRAVITY_DEBUG === "1") {
+                    console.error("[antigravity-plugin] No authorization code received")
+                  }
+                  return { type: "failed" as const }
+                }
+
+                const state = decodeState(result.state)
+                if (state.verifier !== verifier) {
+                  if (process.env.ANTIGRAVITY_DEBUG === "1") {
+                    console.error("[antigravity-plugin] PKCE verifier mismatch")
+                  }
+                  return { type: "failed" as const }
+                }
+
+                const tokens = await exchangeCode(result.code, verifier, cachedClientId, cachedClientSecret, serverHandle.port)
+
+                try {
+                  const userInfo = await fetchUserInfo(tokens.access_token)
+                  if (process.env.ANTIGRAVITY_DEBUG === "1") {
+                    console.log(`[antigravity-plugin] Authenticated as: ${userInfo.email}`)
+                  }
+                } catch {
+                  // User info is optional
+                }
+
+                const projectContext = await fetchProjectContext(tokens.access_token)
+
+                const formattedRefresh = formatTokenForStorage(
+                  tokens.refresh_token,
+                  projectContext.cloudaicompanionProject || "",
+                  projectContext.managedProjectId
+                )
+
+                return {
+                  type: "success" as const,
+                  access: tokens.access_token,
+                  refresh: formattedRefresh,
+                  expires: Date.now() + tokens.expires_in * 1000,
+                }
+              } catch (error) {
+                serverHandle.close()
+                if (process.env.ANTIGRAVITY_DEBUG === "1") {
+                  console.error(
+                    `[antigravity-plugin] OAuth flow failed: ${
+                      error instanceof Error ? error.message : "Unknown error"
+                    }`
+                  )
+                }
+                return { type: "failed" as const }
+              }
+            },
+          }
+        },
+      },
+    ],
+  }
+
+  return {
+    auth: authHook,
+  }
+}
+
+/**
+ * Default export for OpenCode plugin system
+ */
+export default createGoogleAntigravityAuthPlugin
+
+/**
+ * Named export for explicit imports
+ */
+export const GoogleAntigravityAuthPlugin = createGoogleAntigravityAuthPlugin

src/auth/antigravity/project.ts

@@ -0,0 +1,159 @@
+/**
+ * Antigravity project context management.
+ * Handles fetching GCP project ID via Google's loadCodeAssist API.
+ */
+
+import {
+  ANTIGRAVITY_DEFAULT_PROJECT_ID,
+  ANTIGRAVITY_ENDPOINT_FALLBACKS,
+  ANTIGRAVITY_API_VERSION,
+  ANTIGRAVITY_HEADERS,
+} from "./constants"
+import type {
+  AntigravityProjectContext,
+  AntigravityLoadCodeAssistResponse,
+} from "./types"
+
+/**
+ * In-memory cache for project context per access token.
+ * Prevents redundant API calls for the same token.
+ */
+const projectContextCache = new Map<string, AntigravityProjectContext>()
+
+/**
+ * Client metadata for loadCodeAssist API request.
+ * Matches cliproxyapi implementation.
+ */
+const CODE_ASSIST_METADATA = {
+  ideType: "IDE_UNSPECIFIED",
+  platform: "PLATFORM_UNSPECIFIED",
+  pluginType: "GEMINI",
+} as const
+
+/**
+ * Extracts the project ID from a cloudaicompanionProject field.
+ * Handles both string and object formats.
+ *
+ * @param project - The cloudaicompanionProject value from API response
+ * @returns Extracted project ID string, or undefined if not found
+ */
+function extractProjectId(
+  project: string | { id: string } | undefined
+): string | undefined {
+  if (!project) {
+    return undefined
+  }
+
+  // Handle string format
+  if (typeof project === "string") {
+    const trimmed = project.trim()
+    return trimmed || undefined
+  }
+
+  // Handle object format { id: string }
+  if (typeof project === "object" && "id" in project) {
+    const id = project.id
+    if (typeof id === "string") {
+      const trimmed = id.trim()
+      return trimmed || undefined
+    }
+  }
+
+  return undefined
+}
+
+/**
+ * Calls the loadCodeAssist API to get project context.
+ * Tries each endpoint in the fallback list until one succeeds.
+ *
+ * @param accessToken - Valid OAuth access token
+ * @returns API response or null if all endpoints fail
+ */
+async function callLoadCodeAssistAPI(
+  accessToken: string
+): Promise<AntigravityLoadCodeAssistResponse | null> {
+  const requestBody = {
+    metadata: CODE_ASSIST_METADATA,
+  }
+
+  const headers: Record<string, string> = {
+    Authorization: `Bearer ${accessToken}`,
+    "Content-Type": "application/json",
+    "User-Agent": ANTIGRAVITY_HEADERS["User-Agent"],
+    "X-Goog-Api-Client": ANTIGRAVITY_HEADERS["X-Goog-Api-Client"],
+    "Client-Metadata": ANTIGRAVITY_HEADERS["Client-Metadata"],
+  }
+
+  // Try each endpoint in the fallback list
+  for (const baseEndpoint of ANTIGRAVITY_ENDPOINT_FALLBACKS) {
+    const url = `${baseEndpoint}/${ANTIGRAVITY_API_VERSION}:loadCodeAssist`
+
+    try {
+      const response = await fetch(url, {
+        method: "POST",
+        headers,
+        body: JSON.stringify(requestBody),
+      })
+
+      if (!response.ok) {
+        // Try next endpoint on failure
+        continue
+      }
+
+      const data =
+        (await response.json()) as AntigravityLoadCodeAssistResponse
+      return data
+    } catch {
+      // Network or parsing error, try next endpoint
+      continue
+    }
+  }
+
+  // All endpoints failed
+  return null
+}
+
+/**
+ * Fetch project context from Google's loadCodeAssist API.
+ * Extracts the cloudaicompanionProject from the response.
+ *
+ * @param accessToken - Valid OAuth access token
+ * @returns Project context with cloudaicompanionProject ID
+ */
+export async function fetchProjectContext(
+  accessToken: string
+): Promise<AntigravityProjectContext> {
+  const cached = projectContextCache.get(accessToken)
+  if (cached) {
+    return cached
+  }
+
+  const response = await callLoadCodeAssistAPI(accessToken)
+  const projectId = response
+    ? extractProjectId(response.cloudaicompanionProject)
+    : undefined
+
+  const result: AntigravityProjectContext = {
+    cloudaicompanionProject: projectId || "",
+  }
+
+  if (projectId) {
+    projectContextCache.set(accessToken, result)
+  }
+
+  return result
+}
+
+/**
+ * Clear the project context cache.
+ * Call this when tokens are refreshed or invalidated.
+ *
+ * @param accessToken - Optional specific token to clear, or clears all if not provided
+ */
+export function clearProjectContextCache(accessToken?: string): void {
+  if (accessToken) {
+    projectContextCache.delete(accessToken)
+  } else {
+    projectContextCache.clear()
+  }
+}

src/auth/antigravity/request.ts

@@ -0,0 +1,303 @@
+/**
+ * Antigravity request transformer.
+ * Transforms OpenAI-format requests to Antigravity format.
+ * Does NOT handle tool normalization (handled by tools.ts in Task 9).
+ */
+
+import {
+  ANTIGRAVITY_HEADERS,
+  ANTIGRAVITY_ENDPOINT_FALLBACKS,
+  ANTIGRAVITY_API_VERSION,
+  SKIP_THOUGHT_SIGNATURE_VALIDATOR,
+} from "./constants"
+import type { AntigravityRequestBody } from "./types"
+
+/**
+ * Result of request transformation including URL, headers, and body.
+ */
+export interface TransformedRequest {
+  /** Transformed URL for Antigravity API */
+  url: string
+  /** Request headers including Authorization and Antigravity-specific headers */
+  headers: Record<string, string>
+  /** Transformed request body in Antigravity format */
+  body: AntigravityRequestBody
+  /** Whether this is a streaming request */
+  streaming: boolean
+}
+
+/**
+ * Build Antigravity-specific request headers.
+ * Includes Authorization, User-Agent, X-Goog-Api-Client, and Client-Metadata.
+ *
+ * @param accessToken - OAuth access token for Authorization header
+ * @returns Headers object with all required Antigravity headers
+ */
+export function buildRequestHeaders(accessToken: string): Record<string, string> {
+  return {
+    Authorization: `Bearer ${accessToken}`,
+    "Content-Type": "application/json",
+    "User-Agent": ANTIGRAVITY_HEADERS["User-Agent"],
+    "X-Goog-Api-Client": ANTIGRAVITY_HEADERS["X-Goog-Api-Client"],
+    "Client-Metadata": ANTIGRAVITY_HEADERS["Client-Metadata"],
+  }
+}
+
+/**
+ * Extract model name from request body.
+ * OpenAI-format requests include model in the body.
+ *
+ * @param body - Request body that may contain a model field
+ * @returns Model name or undefined if not found
+ */
+export function extractModelFromBody(
+  body: Record<string, unknown>
+): string | undefined {
+  const model = body.model
+  if (typeof model === "string" && model.trim()) {
+    return model.trim()
+  }
+  return undefined
+}
+
+/**
+ * Extract model name from URL path.
+ * Handles Google Generative Language API format: /models/{model}:{action}
+ *
+ * @param url - Request URL to parse
+ * @returns Model name or undefined if not found
+ */
+export function extractModelFromUrl(url: string): string | undefined {
+  // Match Google's API format: /models/gemini-3-pro:generateContent
+  const match = url.match(/\/models\/([^:]+):/)
+  if (match && match[1]) {
+    return match[1]
+  }
+  return undefined
+}
+
+/**
+ * Determine the action type from the URL path.
+ * E.g., generateContent, streamGenerateContent
+ *
+ * @param url - Request URL to parse
+ * @returns Action name or undefined if not found
+ */
+export function extractActionFromUrl(url: string): string | undefined {
+  // Match Google's API format: /models/gemini-3-pro:generateContent
+  const match = url.match(/\/models\/[^:]+:(\w+)/)
+  if (match && match[1]) {
+    return match[1]
+  }
+  return undefined
+}
+
+/**
+ * Check if a URL is targeting Google's Generative Language API.
+ *
+ * @param url - URL to check
+ * @returns true if this is a Google Generative Language API request
+ */
+export function isGenerativeLanguageRequest(url: string): boolean {
+  return url.includes("generativelanguage.googleapis.com")
+}
+
+/**
+ * Build Antigravity API URL for the given action.
+ *
+ * @param baseEndpoint - Base Antigravity endpoint URL (from fallbacks)
+ * @param action - API action (e.g., generateContent, streamGenerateContent)
+ * @param streaming - Whether to append SSE query parameter
+ * @returns Formatted Antigravity API URL
+ */
+export function buildAntigravityUrl(
+  baseEndpoint: string,
+  action: string,
+  streaming: boolean
+): string {
+  const query = streaming ? "?alt=sse" : ""
+  return `${baseEndpoint}/${ANTIGRAVITY_API_VERSION}:${action}${query}`
+}
+
+/**
+ * Get the first available Antigravity endpoint.
+ * Can be used with fallback logic in fetch.ts.
+ *
+ * @returns Default (first) Antigravity endpoint
+ */
+export function getDefaultEndpoint(): string {
+  return ANTIGRAVITY_ENDPOINT_FALLBACKS[0]
+}
+
+function generateRequestId(): string {
+  return `agent-${crypto.randomUUID()}`
+}
+
+export function wrapRequestBody(
+  body: Record<string, unknown>,
+  projectId: string,
+  modelName: string,
+  sessionId: string
+): AntigravityRequestBody {
+  const requestPayload = { ...body }
+  delete requestPayload.model
+
+  return {
+    project: projectId,
+    model: modelName,
+    userAgent: "antigravity",
+    requestId: generateRequestId(),
+    request: {
+      ...requestPayload,
+      sessionId,
+    },
+  }
+}
+
+interface ContentPart {
+  functionCall?: Record<string, unknown>
+  thoughtSignature?: string
+  [key: string]: unknown
+}
+
+interface ContentBlock {
+  role?: string
+  parts?: ContentPart[]
+  [key: string]: unknown
+}
+
+function debugLog(message: string): void {
+  if (process.env.ANTIGRAVITY_DEBUG === "1") {
+    console.log(`[antigravity-request] ${message}`)
+  }
+}
+
+export function injectThoughtSignatureIntoFunctionCalls(
+  body: Record<string, unknown>,
+  signature: string | undefined
+): Record<string, unknown> {
+  // Always use skip validator as fallback (CLIProxyAPI approach)
+  const effectiveSignature = signature || SKIP_THOUGHT_SIGNATURE_VALIDATOR
+  debugLog(`[TSIG][INJECT] signature=${effectiveSignature.substring(0, 30)}... (${signature ? "provided" : "default"})`)
+  debugLog(`[TSIG][INJECT] body keys: ${Object.keys(body).join(", ")}`)
+
+  const contents = body.contents as ContentBlock[] | undefined
+  if (!contents || !Array.isArray(contents)) {
+    debugLog(`[TSIG][INJECT] No contents array! Has messages: ${!!body.messages}`)
+    return body
+  }
+
+  debugLog(`[TSIG][INJECT] Found ${contents.length} content blocks`)
+  let injectedCount = 0
+  const modifiedContents = contents.map((content) => {
+    if (!content.parts || !Array.isArray(content.parts)) {
+      return content
+    }
+
+    const modifiedParts = content.parts.map((part) => {
+      if (part.functionCall && !part.thoughtSignature) {
+        injectedCount++
+        return {
+          ...part,
+          thoughtSignature: effectiveSignature,
+        }
+      }
+      return part
+    })
+
+    return { ...content, parts: modifiedParts }
+  })
+
+  debugLog(`[TSIG][INJECT] injected signature into ${injectedCount} functionCall(s)`)
+  return { ...body, contents: modifiedContents }
+}
+
+/**
+ * Detect if request is for streaming.
+ * Checks both action name and request body for stream flag.
+ *
+ * @param url - Request URL
+ * @param body - Request body
+ * @returns true if streaming is requested
+ */
+export function isStreamingRequest(
+  url: string,
+  body: Record<string, unknown>
+): boolean {
+  // Check URL action
+  const action = extractActionFromUrl(url)
+  if (action === "streamGenerateContent") {
+    return true
+  }
+
+  // Check body for stream flag
+  if (body.stream === true) {
+    return true
+  }
+
+  return false
+}
+
+export interface TransformRequestOptions {
+  url: string
+  body: Record<string, unknown>
+  accessToken: string
+  projectId: string
+  sessionId: string
+  modelName?: string
+  endpointOverride?: string
+  thoughtSignature?: string
+}
+
+export function transformRequest(options: TransformRequestOptions): TransformedRequest {
+  const {
+    url,
+    body,
+    accessToken,
+    projectId,
+    sessionId,
+    modelName,
+    endpointOverride,
+    thoughtSignature,
+  } = options
+
+  const effectiveModel =
+    modelName || extractModelFromBody(body) || extractModelFromUrl(url) || "gemini-3-pro-preview"
+
+  const streaming = isStreamingRequest(url, body)
+  const action = streaming ? "streamGenerateContent" : "generateContent"
+
+  const endpoint = endpointOverride || getDefaultEndpoint()
+  const transformedUrl = buildAntigravityUrl(endpoint, action, streaming)
+
+  const headers = buildRequestHeaders(accessToken)
+  if (streaming) {
+    headers["Accept"] = "text/event-stream"
+  }
+
+  const bodyWithSignature = injectThoughtSignatureIntoFunctionCalls(body, thoughtSignature)
+  const wrappedBody = wrapRequestBody(bodyWithSignature, projectId, effectiveModel, sessionId)
+
+  return {
+    url: transformedUrl,
+    headers,
+    body: wrappedBody,
+    streaming,
+  }
+}
+
+/**
+ * Prepare request headers for streaming responses.
+ * Adds Accept header for SSE format.
+ *
+ * @param headers - Existing headers object
+ * @returns Headers with streaming support
+ */
+export function addStreamingHeaders(
+  headers: Record<string, string>
+): Record<string, string> {
+  return {
+    ...headers,
+    Accept: "text/event-stream",
+  }
+}

src/auth/antigravity/response.ts

@@ -0,0 +1,598 @@
+/**
+ * Antigravity Response Handler
+ * Transforms Antigravity/Gemini API responses to OpenAI-compatible format
+ *
+ * Key responsibilities:
+ * - Non-streaming response transformation
+ * - SSE streaming response transformation (buffered - see transformStreamingResponse)
+ * - Error response handling with retry-after extraction
+ * - Usage metadata extraction from x-antigravity-* headers
+ */
+
+import type { AntigravityError, AntigravityUsage } from "./types"
+
+/**
+ * Usage metadata extracted from Antigravity response headers
+ */
+export interface AntigravityUsageMetadata {
+  cachedContentTokenCount?: number
+  totalTokenCount?: number
+  promptTokenCount?: number
+  candidatesTokenCount?: number
+}
+
+/**
+ * Transform result with response and metadata
+ */
+export interface TransformResult {
+  response: Response
+  usage?: AntigravityUsageMetadata
+  retryAfterMs?: number
+  error?: AntigravityError
+}
+
+/**
+ * Extract usage metadata from Antigravity response headers
+ *
+ * Antigravity sets these headers:
+ * - x-antigravity-cached-content-token-count
+ * - x-antigravity-total-token-count
+ * - x-antigravity-prompt-token-count
+ * - x-antigravity-candidates-token-count
+ *
+ * @param headers - Response headers
+ * @returns Usage metadata if found
+ */
+export function extractUsageFromHeaders(headers: Headers): AntigravityUsageMetadata | undefined {
+  const cached = headers.get("x-antigravity-cached-content-token-count")
+  const total = headers.get("x-antigravity-total-token-count")
+  const prompt = headers.get("x-antigravity-prompt-token-count")
+  const candidates = headers.get("x-antigravity-candidates-token-count")
+
+  // Return undefined if no usage headers found
+  if (!cached && !total && !prompt && !candidates) {
+    return undefined
+  }
+
+  const usage: AntigravityUsageMetadata = {}
+
+  if (cached) {
+    const parsed = parseInt(cached, 10)
+    if (!isNaN(parsed)) {
+      usage.cachedContentTokenCount = parsed
+    }
+  }
+
+  if (total) {
+    const parsed = parseInt(total, 10)
+    if (!isNaN(parsed)) {
+      usage.totalTokenCount = parsed
+    }
+  }
+
+  if (prompt) {
+    const parsed = parseInt(prompt, 10)
+    if (!isNaN(parsed)) {
+      usage.promptTokenCount = parsed
+    }
+  }
+
+  if (candidates) {
+    const parsed = parseInt(candidates, 10)
+    if (!isNaN(parsed)) {
+      usage.candidatesTokenCount = parsed
+    }
+  }
+
+  return Object.keys(usage).length > 0 ? usage : undefined
+}
+
+/**
+ * Extract retry-after value from error response
+ *
+ * Antigravity returns retry info in error.details array:
+ * {
+ *   error: {
+ *     details: [{
+ *       "@type": "type.googleapis.com/google.rpc.RetryInfo",
+ *       "retryDelay": "5.123s"
+ *     }]
+ *   }
+ * }
+ *
+ * Also checks standard Retry-After header.
+ *
+ * @param response - Response object (for headers)
+ * @param errorBody - Parsed error body (optional)
+ * @returns Retry after value in milliseconds, or undefined
+ */
+export function extractRetryAfterMs(
+  response: Response,
+  errorBody?: Record<string, unknown>,
+): number | undefined {
+  // First, check standard Retry-After header
+  const retryAfterHeader = response.headers.get("Retry-After")
+  if (retryAfterHeader) {
+    const seconds = parseFloat(retryAfterHeader)
+    if (!isNaN(seconds) && seconds > 0) {
+      return Math.ceil(seconds * 1000)
+    }
+  }
+
+  // Check retry-after-ms header (set by some transformers)
+  const retryAfterMsHeader = response.headers.get("retry-after-ms")
+  if (retryAfterMsHeader) {
+    const ms = parseInt(retryAfterMsHeader, 10)
+    if (!isNaN(ms) && ms > 0) {
+      return ms
+    }
+  }
+
+  // Check error body for RetryInfo
+  if (!errorBody) {
+    return undefined
+  }
+
+  const error = errorBody.error as Record<string, unknown> | undefined
+  if (!error?.details || !Array.isArray(error.details)) {
+    return undefined
+  }
+
+  const retryInfo = (error.details as Array<Record<string, unknown>>).find(
+    (detail) => detail["@type"] === "type.googleapis.com/google.rpc.RetryInfo",
+  )
+
+  if (!retryInfo?.retryDelay || typeof retryInfo.retryDelay !== "string") {
+    return undefined
+  }
+
+  // Parse retryDelay format: "5.123s"
+  const match = retryInfo.retryDelay.match(/^([\d.]+)s$/)
+  if (match?.[1]) {
+    const seconds = parseFloat(match[1])
+    if (!isNaN(seconds) && seconds > 0) {
+      return Math.ceil(seconds * 1000)
+    }
+  }
+
+  return undefined
+}
+
+/**
+ * Parse error response body and extract useful details
+ *
+ * @param text - Raw response text
+ * @returns Parsed error or undefined
+ */
+export function parseErrorBody(text: string): AntigravityError | undefined {
+  try {
+    const parsed = JSON.parse(text) as Record<string, unknown>
+
+    // Handle error wrapper
+    if (parsed.error && typeof parsed.error === "object") {
+      const errorObj = parsed.error as Record<string, unknown>
+      return {
+        message: String(errorObj.message || "Unknown error"),
+        type: errorObj.type ? String(errorObj.type) : undefined,
+        code: errorObj.code as string | number | undefined,
+      }
+    }
+
+    // Handle direct error message
+    if (parsed.message && typeof parsed.message === "string") {
+      return {
+        message: parsed.message,
+        type: parsed.type ? String(parsed.type) : undefined,
+        code: parsed.code as string | number | undefined,
+      }
+    }
+
+    return undefined
+  } catch {
+    // If not valid JSON, return generic error
+    return {
+      message: text || "Unknown error",
+    }
+  }
+}
+
+/**
+ * Transform a non-streaming Antigravity response to OpenAI-compatible format
+ *
+ * For non-streaming responses:
+ * - Parses the response body
+ * - Unwraps the `response` field if present (Antigravity wraps responses)
+ * - Extracts usage metadata from headers
+ * - Handles error responses
+ *
+ * Note: Does NOT handle thinking block extraction (Task 10)
+ * Note: Does NOT handle tool normalization (Task 9)
+ *
+ * @param response - Fetch Response object
+ * @returns TransformResult with transformed response and metadata
+ */
+export async function transformResponse(response: Response): Promise<TransformResult> {
+  const headers = new Headers(response.headers)
+  const usage = extractUsageFromHeaders(headers)
+
+  // Handle error responses
+  if (!response.ok) {
+    const text = await response.text()
+    const error = parseErrorBody(text)
+    const retryAfterMs = extractRetryAfterMs(response, error ? { error } : undefined)
+
+    // Parse to get full error body for retry-after extraction
+    let errorBody: Record<string, unknown> | undefined
+    try {
+      errorBody = JSON.parse(text) as Record<string, unknown>
+    } catch {
+      errorBody = { error: { message: text } }
+    }
+
+    const retryMs = extractRetryAfterMs(response, errorBody) ?? retryAfterMs
+
+    // Set retry headers if found
+    if (retryMs) {
+      headers.set("Retry-After", String(Math.ceil(retryMs / 1000)))
+      headers.set("retry-after-ms", String(retryMs))
+    }
+
+    return {
+      response: new Response(text, {
+        status: response.status,
+        statusText: response.statusText,
+        headers,
+      }),
+      usage,
+      retryAfterMs: retryMs,
+      error,
+    }
+  }
+
+  // Handle successful response
+  const contentType = response.headers.get("content-type") ?? ""
+  const isJson = contentType.includes("application/json")
+
+  if (!isJson) {
+    // Return non-JSON responses as-is
+    return { response, usage }
+  }
+
+  try {
+    const text = await response.text()
+    const parsed = JSON.parse(text) as Record<string, unknown>
+
+    // Antigravity wraps response in { response: { ... } }
+    // Unwrap if present
+    let transformedBody: unknown = parsed
+    if (parsed.response !== undefined) {
+      transformedBody = parsed.response
+    }
+
+    return {
+      response: new Response(JSON.stringify(transformedBody), {
+        status: response.status,
+        statusText: response.statusText,
+        headers,
+      }),
+      usage,
+    }
+  } catch {
+    // If parsing fails, return original response
+    return { response, usage }
+  }
+}
+
+/**
+ * Transform a single SSE data line
+ *
+ * Antigravity SSE format:
+ * data: { "response": { ... actual data ... } }
+ *
+ * OpenAI SSE format:
+ * data: { ... actual data ... }
+ *
+ * @param line - SSE data line
+ * @returns Transformed line
+ */
+function transformSseLine(line: string): string {
+  if (!line.startsWith("data:")) {
+    return line
+  }
+
+  const json = line.slice(5).trim()
+  if (!json || json === "[DONE]") {
+    return line
+  }
+
+  try {
+    const parsed = JSON.parse(json) as Record<string, unknown>
+
+    // Unwrap { response: { ... } } wrapper
+    if (parsed.response !== undefined) {
+      return `data: ${JSON.stringify(parsed.response)}`
+    }
+
+    return line
+  } catch {
+    // If parsing fails, return original line
+    return line
+  }
+}
+
+/**
+ * Transform SSE streaming payload
+ *
+ * Processes each line in the SSE stream:
+ * - Unwraps { response: { ... } } wrapper from data lines
+ * - Preserves other SSE control lines (event:, id:, retry:, empty lines)
+ *
+ * Note: Does NOT extract thinking blocks (Task 10)
+ *
+ * @param payload - Raw SSE payload text
+ * @returns Transformed SSE payload
+ */
+export function transformStreamingPayload(payload: string): string {
+  return payload
+    .split("\n")
+    .map(transformSseLine)
+    .join("\n")
+}
+
+function createSseTransformStream(): TransformStream<Uint8Array, Uint8Array> {
+  const decoder = new TextDecoder()
+  const encoder = new TextEncoder()
+  let buffer = ""
+
+  return new TransformStream({
+    transform(chunk, controller) {
+      buffer += decoder.decode(chunk, { stream: true })
+      const lines = buffer.split("\n")
+      buffer = lines.pop() || ""
+
+      for (const line of lines) {
+        const transformed = transformSseLine(line)
+        controller.enqueue(encoder.encode(transformed + "\n"))
+      }
+    },
+    flush(controller) {
+      if (buffer) {
+        const transformed = transformSseLine(buffer)
+        controller.enqueue(encoder.encode(transformed))
+      }
+    },
+  })
+}
+
+/**
+ * Transforms a streaming SSE response from Antigravity to OpenAI format.
+ *
+ * Uses TransformStream to process SSE chunks incrementally as they arrive.
+ * Each line is transformed immediately and yielded to the client.
+ *
+ * @param response - The SSE response from Antigravity API
+ * @returns TransformResult with transformed streaming response
+ */
+export async function transformStreamingResponse(response: Response): Promise<TransformResult> {
+  const headers = new Headers(response.headers)
+  const usage = extractUsageFromHeaders(headers)
+
+  // Handle error responses
+  if (!response.ok) {
+    const text = await response.text()
+    const error = parseErrorBody(text)
+
+    let errorBody: Record<string, unknown> | undefined
+    try {
+      errorBody = JSON.parse(text) as Record<string, unknown>
+    } catch {
+      errorBody = { error: { message: text } }
+    }
+
+    const retryAfterMs = extractRetryAfterMs(response, errorBody)
+
+    if (retryAfterMs) {
+      headers.set("Retry-After", String(Math.ceil(retryAfterMs / 1000)))
+      headers.set("retry-after-ms", String(retryAfterMs))
+    }
+
+    return {
+      response: new Response(text, {
+        status: response.status,
+        statusText: response.statusText,
+        headers,
+      }),
+      usage,
+      retryAfterMs,
+      error,
+    }
+  }
+
+  // Check content type
+  const contentType = response.headers.get("content-type") ?? ""
+  const isEventStream =
+    contentType.includes("text/event-stream") || response.url.includes("alt=sse")
+
+  if (!isEventStream) {
+    // Not SSE, delegate to non-streaming transform
+    // Clone response since we need to read it
+    const text = await response.text()
+    try {
+      const parsed = JSON.parse(text) as Record<string, unknown>
+      let transformedBody: unknown = parsed
+      if (parsed.response !== undefined) {
+        transformedBody = parsed.response
+      }
+      return {
+        response: new Response(JSON.stringify(transformedBody), {
+          status: response.status,
+          statusText: response.statusText,
+          headers,
+        }),
+        usage,
+      }
+    } catch {
+      return {
+        response: new Response(text, {
+          status: response.status,
+          statusText: response.statusText,
+          headers,
+        }),
+        usage,
+      }
+    }
+  }
+
+  if (!response.body) {
+    return { response, usage }
+  }
+
+  headers.delete("content-length")
+  headers.delete("content-encoding")
+  headers.set("content-type", "text/event-stream; charset=utf-8")
+
+  const transformStream = createSseTransformStream()
+  const transformedBody = response.body.pipeThrough(transformStream)
+
+  return {
+    response: new Response(transformedBody, {
+      status: response.status,
+      statusText: response.statusText,
+      headers,
+    }),
+    usage,
+  }
+}
+
+/**
+ * Check if response is a streaming SSE response
+ *
+ * @param response - Fetch Response object
+ * @returns True if response is SSE stream
+ */
+export function isStreamingResponse(response: Response): boolean {
+  const contentType = response.headers.get("content-type") ?? ""
+  return contentType.includes("text/event-stream") || response.url.includes("alt=sse")
+}
+
+/**
+ * Extract thought signature from SSE payload text
+ *
+ * Looks for thoughtSignature in SSE events:
+ * data: { "response": { "candidates": [{ "content": { "parts": [{ "thoughtSignature": "..." }] } }] } }
+ *
+ * Returns the last found signature (most recent in the stream).
+ *
+ * @param payload - SSE payload text
+ * @returns Last thought signature if found
+ */
+export function extractSignatureFromSsePayload(payload: string): string | undefined {
+  const lines = payload.split("\n")
+  let lastSignature: string | undefined
+
+  for (const line of lines) {
+    if (!line.startsWith("data:")) {
+      continue
+    }
+
+    const json = line.slice(5).trim()
+    if (!json || json === "[DONE]") {
+      continue
+    }
+
+    try {
+      const parsed = JSON.parse(json) as Record<string, unknown>
+
+      // Check in response wrapper (Antigravity format)
+      const response = (parsed.response || parsed) as Record<string, unknown>
+      const candidates = response.candidates as Array<Record<string, unknown>> | undefined
+
+      if (candidates && Array.isArray(candidates)) {
+        for (const candidate of candidates) {
+          const content = candidate.content as Record<string, unknown> | undefined
+          const parts = content?.parts as Array<Record<string, unknown>> | undefined
+
+          if (parts && Array.isArray(parts)) {
+            for (const part of parts) {
+              const sig = (part.thoughtSignature || part.thought_signature) as string | undefined
+              if (sig && typeof sig === "string") {
+                lastSignature = sig
+              }
+            }
+          }
+        }
+      }
+    } catch {
+      // Continue to next line if parsing fails
+    }
+  }
+
+  return lastSignature
+}
+
+/**
+ * Extract usage from SSE payload text
+ *
+ * Looks for usageMetadata in SSE events:
+ * data: { "usageMetadata": { ... } }
+ *
+ * @param payload - SSE payload text
+ * @returns Usage if found
+ */
+export function extractUsageFromSsePayload(payload: string): AntigravityUsage | undefined {
+  const lines = payload.split("\n")
+
+  for (const line of lines) {
+    if (!line.startsWith("data:")) {
+      continue
+    }
+
+    const json = line.slice(5).trim()
+    if (!json || json === "[DONE]") {
+      continue
+    }
+
+    try {
+      const parsed = JSON.parse(json) as Record<string, unknown>
+
+      // Check for usageMetadata at top level
+      if (parsed.usageMetadata && typeof parsed.usageMetadata === "object") {
+        const meta = parsed.usageMetadata as Record<string, unknown>
+        return {
+          prompt_tokens: typeof meta.promptTokenCount === "number" ? meta.promptTokenCount : 0,
+          completion_tokens:
+            typeof meta.candidatesTokenCount === "number" ? meta.candidatesTokenCount : 0,
+          total_tokens: typeof meta.totalTokenCount === "number" ? meta.totalTokenCount : 0,
+        }
+      }
+
+      // Check for usage in response wrapper
+      if (parsed.response && typeof parsed.response === "object") {
+        const resp = parsed.response as Record<string, unknown>
+        if (resp.usageMetadata && typeof resp.usageMetadata === "object") {
+          const meta = resp.usageMetadata as Record<string, unknown>
+          return {
+            prompt_tokens: typeof meta.promptTokenCount === "number" ? meta.promptTokenCount : 0,
+            completion_tokens:
+              typeof meta.candidatesTokenCount === "number" ? meta.candidatesTokenCount : 0,
+            total_tokens: typeof meta.totalTokenCount === "number" ? meta.totalTokenCount : 0,
+          }
+        }
+      }
+
+      // Check for standard OpenAI-style usage
+      if (parsed.usage && typeof parsed.usage === "object") {
+        const u = parsed.usage as Record<string, unknown>
+        return {
+          prompt_tokens: typeof u.prompt_tokens === "number" ? u.prompt_tokens : 0,
+          completion_tokens: typeof u.completion_tokens === "number" ? u.completion_tokens : 0,
+          total_tokens: typeof u.total_tokens === "number" ? u.total_tokens : 0,
+        }
+      }
+    } catch {
+      // Continue to next line if parsing fails
+    }
+  }
+
+  return undefined
+}

src/auth/antigravity/thinking.ts

@@ -0,0 +1,571 @@
+/**
+ * Antigravity Thinking Block Handler (Gemini only)
+ *
+ * Handles extraction and transformation of thinking/reasoning blocks
+ * from Gemini responses. Thinking blocks contain the model's internal
+ * reasoning process, available in `-high` model variants.
+ *
+ * Key responsibilities:
+ * - Extract thinking blocks from Gemini response format
+ * - Detect thinking-capable model variants (`-high` suffix)
+ * - Format thinking blocks for OpenAI-compatible output
+ *
+ * Note: This is Gemini-only. Claude models are NOT handled by Antigravity.
+ */
+
+/**
+ * Represents a single thinking/reasoning block extracted from Gemini response
+ */
+export interface ThinkingBlock {
+  /** The thinking/reasoning text content */
+  text: string
+  /** Optional signature for signed thinking blocks (required for multi-turn) */
+  signature?: string
+  /** Index of the thinking block in sequence */
+  index?: number
+}
+
+/**
+ * Raw part structure from Gemini response candidates
+ */
+export interface GeminiPart {
+  /** Text content of the part */
+  text?: string
+  /** Whether this part is a thinking/reasoning block */
+  thought?: boolean
+  /** Signature for signed thinking blocks */
+  thoughtSignature?: string
+  /** Type field for Anthropic-style format */
+  type?: string
+  /** Signature field for Anthropic-style format */
+  signature?: string
+}
+
+/**
+ * Gemini response candidate structure
+ */
+export interface GeminiCandidate {
+  /** Content containing parts */
+  content?: {
+    /** Role of the content (e.g., "model", "assistant") */
+    role?: string
+    /** Array of content parts */
+    parts?: GeminiPart[]
+  }
+  /** Index of the candidate */
+  index?: number
+}
+
+/**
+ * Gemini response structure for thinking block extraction
+ */
+export interface GeminiResponse {
+  /** Response ID */
+  id?: string
+  /** Array of response candidates */
+  candidates?: GeminiCandidate[]
+  /** Direct content (some responses use this instead of candidates) */
+  content?: Array<{
+    type?: string
+    text?: string
+    signature?: string
+  }>
+  /** Model used for response */
+  model?: string
+}
+
+/**
+ * Result of thinking block extraction
+ */
+export interface ThinkingExtractionResult {
+  /** Extracted thinking blocks */
+  thinkingBlocks: ThinkingBlock[]
+  /** Combined thinking text for convenience */
+  combinedThinking: string
+  /** Whether any thinking blocks were found */
+  hasThinking: boolean
+}
+
+/**
+ * Default thinking budget in tokens for thinking-enabled models
+ */
+export const DEFAULT_THINKING_BUDGET = 16000
+
+/**
+ * Check if a model variant should include thinking blocks
+ *
+ * Returns true for model variants with `-high` suffix, which have
+ * extended thinking capability enabled.
+ *
+ * Examples:
+ * - `gemini-3-pro-high` → true
+ * - `gemini-2.5-pro-high` → true
+ * - `gemini-3-pro-preview` → false
+ * - `gemini-2.5-pro` → false
+ *
+ * @param model - Model identifier string
+ * @returns True if model should include thinking blocks
+ */
+export function shouldIncludeThinking(model: string): boolean {
+  if (!model || typeof model !== "string") {
+    return false
+  }
+
+  const lowerModel = model.toLowerCase()
+
+  // Check for -high suffix (primary indicator of thinking capability)
+  if (lowerModel.endsWith("-high")) {
+    return true
+  }
+
+  // Also check for explicit thinking in model name
+  if (lowerModel.includes("thinking")) {
+    return true
+  }
+
+  return false
+}
+
+/**
+ * Check if a model is thinking-capable (broader check)
+ *
+ * This is a broader check than shouldIncludeThinking - it detects models
+ * that have thinking capability, even if not explicitly requesting thinking output.
+ *
+ * @param model - Model identifier string
+ * @returns True if model supports thinking/reasoning
+ */
+export function isThinkingCapableModel(model: string): boolean {
+  if (!model || typeof model !== "string") {
+    return false
+  }
+
+  const lowerModel = model.toLowerCase()
+
+  return (
+    lowerModel.includes("thinking") ||
+    lowerModel.includes("gemini-3") ||
+    lowerModel.endsWith("-high")
+  )
+}
+
+/**
+ * Check if a part is a thinking/reasoning block
+ *
+ * Detects both Gemini-style (thought: true) and Anthropic-style
+ * (type: "thinking" or type: "reasoning") formats.
+ *
+ * @param part - Content part to check
+ * @returns True if part is a thinking block
+ */
+function isThinkingPart(part: GeminiPart): boolean {
+  // Gemini-style: thought flag
+  if (part.thought === true) {
+    return true
+  }
+
+  // Anthropic-style: type field
+  if (part.type === "thinking" || part.type === "reasoning") {
+    return true
+  }
+
+  return false
+}
+
+/**
+ * Check if a thinking part has a valid signature
+ *
+ * Signatures are required for multi-turn conversations with Claude models.
+ * Gemini uses `thoughtSignature`, Anthropic uses `signature`.
+ *
+ * @param part - Thinking part to check
+ * @returns True if part has valid signature
+ */
+function hasValidSignature(part: GeminiPart): boolean {
+  // Gemini-style signature
+  if (part.thought === true && part.thoughtSignature) {
+    return true
+  }
+
+  // Anthropic-style signature
+  if ((part.type === "thinking" || part.type === "reasoning") && part.signature) {
+    return true
+  }
+
+  return false
+}
+
+/**
+ * Extract thinking blocks from a Gemini response
+ *
+ * Parses the response structure to identify and extract all thinking/reasoning
+ * content. Supports both Gemini-style (thought: true) and Anthropic-style
+ * (type: "thinking") formats.
+ *
+ * @param response - Gemini response object
+ * @returns Extraction result with thinking blocks and metadata
+ */
+export function extractThinkingBlocks(response: GeminiResponse): ThinkingExtractionResult {
+  const thinkingBlocks: ThinkingBlock[] = []
+
+  // Handle candidates array (standard Gemini format)
+  if (response.candidates && Array.isArray(response.candidates)) {
+    for (const candidate of response.candidates) {
+      const parts = candidate.content?.parts
+      if (!parts || !Array.isArray(parts)) {
+        continue
+      }
+
+      for (let i = 0; i < parts.length; i++) {
+        const part = parts[i]
+        if (!part || typeof part !== "object") {
+          continue
+        }
+
+        if (isThinkingPart(part)) {
+          const block: ThinkingBlock = {
+            text: part.text || "",
+            index: thinkingBlocks.length,
+          }
+
+          // Extract signature if present
+          if (part.thought === true && part.thoughtSignature) {
+            block.signature = part.thoughtSignature
+          } else if (part.signature) {
+            block.signature = part.signature
+          }
+
+          thinkingBlocks.push(block)
+        }
+      }
+    }
+  }
+
+  // Handle direct content array (Anthropic-style response)
+  if (response.content && Array.isArray(response.content)) {
+    for (let i = 0; i < response.content.length; i++) {
+      const item = response.content[i]
+      if (!item || typeof item !== "object") {
+        continue
+      }
+
+      if (item.type === "thinking" || item.type === "reasoning") {
+        thinkingBlocks.push({
+          text: item.text || "",
+          signature: item.signature,
+          index: thinkingBlocks.length,
+        })
+      }
+    }
+  }
+
+  // Combine all thinking text
+  const combinedThinking = thinkingBlocks.map((b) => b.text).join("\n\n")
+
+  return {
+    thinkingBlocks,
+    combinedThinking,
+    hasThinking: thinkingBlocks.length > 0,
+  }
+}
+
+/**
+ * Format thinking blocks for OpenAI-compatible output
+ *
+ * Converts Gemini thinking block format to OpenAI's expected structure.
+ * OpenAI expects thinking content as special message blocks or annotations.
+ *
+ * Output format:
+ * ```
+ * [
+ *   { type: "reasoning", text: "thinking content...", signature?: "..." },
+ *   ...
+ * ]
+ * ```
+ *
+ * @param thinking - Array of thinking blocks to format
+ * @returns OpenAI-compatible formatted array
+ */
+export function formatThinkingForOpenAI(
+  thinking: ThinkingBlock[],
+): Array<{ type: "reasoning"; text: string; signature?: string }> {
+  if (!thinking || !Array.isArray(thinking) || thinking.length === 0) {
+    return []
+  }
+
+  return thinking.map((block) => {
+    const formatted: { type: "reasoning"; text: string; signature?: string } = {
+      type: "reasoning",
+      text: block.text || "",
+    }
+
+    if (block.signature) {
+      formatted.signature = block.signature
+    }
+
+    return formatted
+  })
+}
+
+/**
+ * Transform thinking parts in a candidate to OpenAI format
+ *
+ * Modifies candidate content parts to use OpenAI-style reasoning format
+ * while preserving the rest of the response structure.
+ *
+ * @param candidate - Gemini candidate to transform
+ * @returns Transformed candidate with reasoning-formatted thinking
+ */
+export function transformCandidateThinking(candidate: GeminiCandidate): GeminiCandidate {
+  if (!candidate || typeof candidate !== "object") {
+    return candidate
+  }
+
+  const content = candidate.content
+  if (!content || typeof content !== "object" || !Array.isArray(content.parts)) {
+    return candidate
+  }
+
+  const thinkingTexts: string[] = []
+  const transformedParts = content.parts.map((part) => {
+    if (part && typeof part === "object" && part.thought === true) {
+      thinkingTexts.push(part.text || "")
+      // Transform to reasoning format
+      return {
+        ...part,
+        type: "reasoning" as const,
+        thought: undefined, // Remove Gemini-specific field
+      }
+    }
+    return part
+  })
+
+  const result: GeminiCandidate & { reasoning_content?: string } = {
+    ...candidate,
+    content: { ...content, parts: transformedParts },
+  }
+
+  // Add combined reasoning content for convenience
+  if (thinkingTexts.length > 0) {
+    result.reasoning_content = thinkingTexts.join("\n\n")
+  }
+
+  return result
+}
+
+/**
+ * Transform Anthropic-style thinking blocks to reasoning format
+ *
+ * Converts `type: "thinking"` blocks to `type: "reasoning"` for consistency.
+ *
+ * @param content - Array of content blocks
+ * @returns Transformed content array
+ */
+export function transformAnthropicThinking(
+  content: Array<{ type?: string; text?: string; signature?: string }>,
+): Array<{ type?: string; text?: string; signature?: string }> {
+  if (!content || !Array.isArray(content)) {
+    return content
+  }
+
+  return content.map((block) => {
+    if (block && typeof block === "object" && block.type === "thinking") {
+      return {
+        type: "reasoning",
+        text: block.text || "",
+        ...(block.signature ? { signature: block.signature } : {}),
+      }
+    }
+    return block
+  })
+}
+
+/**
+ * Filter out unsigned thinking blocks
+ *
+ * Claude API requires signed thinking blocks for multi-turn conversations.
+ * This function removes thinking blocks without valid signatures.
+ *
+ * @param parts - Array of content parts
+ * @returns Filtered array without unsigned thinking blocks
+ */
+export function filterUnsignedThinkingBlocks(parts: GeminiPart[]): GeminiPart[] {
+  if (!parts || !Array.isArray(parts)) {
+    return parts
+  }
+
+  return parts.filter((part) => {
+    if (!part || typeof part !== "object") {
+      return true
+    }
+
+    // If it's a thinking part, only keep it if signed
+    if (isThinkingPart(part)) {
+      return hasValidSignature(part)
+    }
+
+    // Keep all non-thinking parts
+    return true
+  })
+}
+
+/**
+ * Transform entire response thinking parts
+ *
+ * Main transformation function that handles both Gemini-style and
+ * Anthropic-style thinking blocks in a response.
+ *
+ * @param response - Response object to transform
+ * @returns Transformed response with standardized reasoning format
+ */
+export function transformResponseThinking(response: GeminiResponse): GeminiResponse {
+  if (!response || typeof response !== "object") {
+    return response
+  }
+
+  const result: GeminiResponse = { ...response }
+
+  // Transform candidates (Gemini-style)
+  if (Array.isArray(result.candidates)) {
+    result.candidates = result.candidates.map(transformCandidateThinking)
+  }
+
+  // Transform direct content (Anthropic-style)
+  if (Array.isArray(result.content)) {
+    result.content = transformAnthropicThinking(result.content)
+  }
+
+  return result
+}
+
+/**
+ * Thinking configuration for requests
+ */
+export interface ThinkingConfig {
+  /** Token budget for thinking/reasoning */
+  thinkingBudget?: number
+  /** Whether to include thoughts in response */
+  includeThoughts?: boolean
+}
+
+/**
+ * Normalize thinking configuration
+ *
+ * Ensures thinkingConfig is valid: includeThoughts only allowed when budget > 0.
+ *
+ * @param config - Raw thinking configuration
+ * @returns Normalized configuration or undefined
+ */
+export function normalizeThinkingConfig(config: unknown): ThinkingConfig | undefined {
+  if (!config || typeof config !== "object") {
+    return undefined
+  }
+
+  const record = config as Record<string, unknown>
+  const budgetRaw = record.thinkingBudget ?? record.thinking_budget
+  const includeRaw = record.includeThoughts ?? record.include_thoughts
+
+  const thinkingBudget =
+    typeof budgetRaw === "number" && Number.isFinite(budgetRaw) ? budgetRaw : undefined
+  const includeThoughts = typeof includeRaw === "boolean" ? includeRaw : undefined
+
+  const enableThinking = thinkingBudget !== undefined && thinkingBudget > 0
+  const finalInclude = enableThinking ? (includeThoughts ?? false) : false
+
+  // Return undefined if no meaningful config
+  if (
+    !enableThinking &&
+    finalInclude === false &&
+    thinkingBudget === undefined &&
+    includeThoughts === undefined
+  ) {
+    return undefined
+  }
+
+  const normalized: ThinkingConfig = {}
+  if (thinkingBudget !== undefined) {
+    normalized.thinkingBudget = thinkingBudget
+  }
+  if (finalInclude !== undefined) {
+    normalized.includeThoughts = finalInclude
+  }
+  return normalized
+}
+
+/**
+ * Extract thinking configuration from request payload
+ *
+ * Supports both Gemini-style thinkingConfig and Anthropic-style thinking options.
+ *
+ * @param requestPayload - Request body
+ * @param generationConfig - Generation config from request
+ * @param extraBody - Extra body options
+ * @returns Extracted thinking configuration or undefined
+ */
+export function extractThinkingConfig(
+  requestPayload: Record<string, unknown>,
+  generationConfig?: Record<string, unknown>,
+  extraBody?: Record<string, unknown>,
+): ThinkingConfig | undefined {
+  // Check for explicit thinkingConfig
+  const thinkingConfig =
+    generationConfig?.thinkingConfig ?? extraBody?.thinkingConfig ?? requestPayload.thinkingConfig
+
+  if (thinkingConfig && typeof thinkingConfig === "object") {
+    const config = thinkingConfig as Record<string, unknown>
+    return {
+      includeThoughts: Boolean(config.includeThoughts),
+      thinkingBudget:
+        typeof config.thinkingBudget === "number" ? config.thinkingBudget : DEFAULT_THINKING_BUDGET,
+    }
+  }
+
+  // Convert Anthropic-style "thinking" option: { type: "enabled", budgetTokens: N }
+  const anthropicThinking = extraBody?.thinking ?? requestPayload.thinking
+  if (anthropicThinking && typeof anthropicThinking === "object") {
+    const thinking = anthropicThinking as Record<string, unknown>
+    if (thinking.type === "enabled" || thinking.budgetTokens) {
+      return {
+        includeThoughts: true,
+        thinkingBudget:
+          typeof thinking.budgetTokens === "number"
+            ? thinking.budgetTokens
+            : DEFAULT_THINKING_BUDGET,
+      }
+    }
+  }
+
+  return undefined
+}
+
+/**
+ * Resolve final thinking configuration based on model and context
+ *
+ * Handles special cases like Claude models requiring signed thinking blocks
+ * for multi-turn conversations.
+ *
+ * @param userConfig - User-provided thinking configuration
+ * @param isThinkingModel - Whether model supports thinking
+ * @param isClaudeModel - Whether model is Claude (not used in Antigravity, but kept for compatibility)
+ * @param hasAssistantHistory - Whether conversation has assistant history
+ * @returns Final thinking configuration
+ */
+export function resolveThinkingConfig(
+  userConfig: ThinkingConfig | undefined,
+  isThinkingModel: boolean,
+  isClaudeModel: boolean,
+  hasAssistantHistory: boolean,
+): ThinkingConfig | undefined {
+  // Claude models with history need signed thinking blocks
+  // Since we can't guarantee signatures, disable thinking
+  if (isClaudeModel && hasAssistantHistory) {
+    return { includeThoughts: false, thinkingBudget: 0 }
+  }
+
+  // Enable thinking by default for thinking-capable models
+  if (isThinkingModel && !userConfig) {
+    return { includeThoughts: true, thinkingBudget: DEFAULT_THINKING_BUDGET }
+  }
+
+  return userConfig
+}

src/auth/antigravity/thought-signature-store.ts

@@ -0,0 +1,97 @@
+/**
+ * Thought Signature Store
+ *
+ * Stores and retrieves thought signatures for multi-turn conversations.
+ * Gemini 3 Pro requires thought_signature on function call content blocks
+ * in subsequent requests to maintain reasoning continuity.
+ *
+ * Key responsibilities:
+ * - Store the latest thought signature per session
+ * - Provide signature for injection into function call requests
+ * - Clear signatures when sessions end
+ */
+
+/**
+ * In-memory store for thought signatures indexed by session ID
+ */
+const signatureStore = new Map<string, string>()
+
+/**
+ * In-memory store for session IDs per fetch instance
+ * Used to maintain consistent sessionId across multi-turn conversations
+ */
+const sessionIdStore = new Map<string, string>()
+
+/**
+ * Store a thought signature for a session
+ *
+ * @param sessionKey - Unique session identifier (typically fetch instance ID)
+ * @param signature - The thought signature from model response
+ */
+export function setThoughtSignature(sessionKey: string, signature: string): void {
+  if (sessionKey && signature) {
+    signatureStore.set(sessionKey, signature)
+  }
+}
+
+/**
+ * Retrieve the stored thought signature for a session
+ *
+ * @param sessionKey - Unique session identifier
+ * @returns The stored signature or undefined if not found
+ */
+export function getThoughtSignature(sessionKey: string): string | undefined {
+  return signatureStore.get(sessionKey)
+}
+
+/**
+ * Clear the thought signature for a session
+ *
+ * @param sessionKey - Unique session identifier
+ */
+export function clearThoughtSignature(sessionKey: string): void {
+  signatureStore.delete(sessionKey)
+}
+
+/**
+ * Store or retrieve a persistent session ID for a fetch instance
+ *
+ * @param fetchInstanceId - Unique identifier for the fetch instance
+ * @param sessionId - Optional session ID to store (if not provided, returns existing or generates new)
+ * @returns The session ID for this fetch instance
+ */
+export function getOrCreateSessionId(fetchInstanceId: string, sessionId?: string): string {
+  if (sessionId) {
+    sessionIdStore.set(fetchInstanceId, sessionId)
+    return sessionId
+  }
+
+  const existing = sessionIdStore.get(fetchInstanceId)
+  if (existing) {
+    return existing
+  }
+
+  const n = Math.floor(Math.random() * Number.MAX_SAFE_INTEGER)
+  const newSessionId = `-${n}`
+  sessionIdStore.set(fetchInstanceId, newSessionId)
+  return newSessionId
+}
+
+/**
+ * Clear the session ID for a fetch instance
+ *
+ * @param fetchInstanceId - Unique identifier for the fetch instance
+ */
+export function clearSessionId(fetchInstanceId: string): void {
+  sessionIdStore.delete(fetchInstanceId)
+}
+
+/**
+ * Clear all stored data for a fetch instance (signature + session ID)
+ *
+ * @param fetchInstanceId - Unique identifier for the fetch instance
+ */
+export function clearFetchInstanceData(fetchInstanceId: string): void {
+  signatureStore.delete(fetchInstanceId)
+  sessionIdStore.delete(fetchInstanceId)
+}

src/auth/antigravity/token.ts

@@ -0,0 +1,119 @@
+/**
+ * Antigravity token management utilities.
+ * Handles token expiration checking, refresh, and storage format parsing.
+ */
+
+import {
+  ANTIGRAVITY_CLIENT_ID,
+  ANTIGRAVITY_CLIENT_SECRET,
+  ANTIGRAVITY_TOKEN_REFRESH_BUFFER_MS,
+  GOOGLE_TOKEN_URL,
+} from "./constants"
+import type {
+  AntigravityRefreshParts,
+  AntigravityTokenExchangeResult,
+  AntigravityTokens,
+} from "./types"
+
+/**
+ * Check if the access token is expired.
+ * Includes a 60-second safety buffer to refresh before actual expiration.
+ *
+ * @param tokens - The Antigravity tokens to check
+ * @returns true if the token is expired or will expire within the buffer period
+ */
+export function isTokenExpired(tokens: AntigravityTokens): boolean {
+  // Calculate when the token expires (timestamp + expires_in in ms)
+  // timestamp is in milliseconds, expires_in is in seconds
+  const expirationTime = tokens.timestamp + tokens.expires_in * 1000
+
+  // Check if current time is past (expiration - buffer)
+  return Date.now() >= expirationTime - ANTIGRAVITY_TOKEN_REFRESH_BUFFER_MS
+}
+
+/**
+ * Refresh an access token using a refresh token.
+ * Exchanges the refresh token for a new access token via Google's OAuth endpoint.
+ *
+ * @param refreshToken - The refresh token to use
+ * @param clientId - Optional custom client ID (defaults to ANTIGRAVITY_CLIENT_ID)
+ * @param clientSecret - Optional custom client secret (defaults to ANTIGRAVITY_CLIENT_SECRET)
+ * @returns Token exchange result with new access token, or throws on error
+ */
+export async function refreshAccessToken(
+  refreshToken: string,
+  clientId: string = ANTIGRAVITY_CLIENT_ID,
+  clientSecret: string = ANTIGRAVITY_CLIENT_SECRET
+): Promise<AntigravityTokenExchangeResult> {
+  const params = new URLSearchParams({
+    grant_type: "refresh_token",
+    refresh_token: refreshToken,
+    client_id: clientId,
+    client_secret: clientSecret,
+  })
+
+  const response = await fetch(GOOGLE_TOKEN_URL, {
+    method: "POST",
+    headers: {
+      "Content-Type": "application/x-www-form-urlencoded",
+    },
+    body: params,
+  })
+
+  if (!response.ok) {
+    const errorText = await response.text().catch(() => "Unknown error")
+    throw new Error(
+      `Token refresh failed: ${response.status} ${response.statusText} - ${errorText}`
+    )
+  }
+
+  const data = (await response.json()) as {
+    access_token: string
+    refresh_token?: string
+    expires_in: number
+    token_type: string
+  }
+
+  return {
+    access_token: data.access_token,
+    // Google may return a new refresh token, fall back to the original
+    refresh_token: data.refresh_token || refreshToken,
+    expires_in: data.expires_in,
+    token_type: data.token_type,
+  }
+}
+
+/**
+ * Parse a stored token string into its component parts.
+ * Storage format: `refreshToken|projectId|managedProjectId`
+ *
+ * @param stored - The pipe-separated stored token string
+ * @returns Parsed refresh parts with refreshToken, projectId, and optional managedProjectId
+ */
+export function parseStoredToken(stored: string): AntigravityRefreshParts {
+  const parts = stored.split("|")
+  const [refreshToken, projectId, managedProjectId] = parts
+
+  return {
+    refreshToken: refreshToken || "",
+    projectId: projectId || undefined,
+    managedProjectId: managedProjectId || undefined,
+  }
+}
+
+/**
+ * Format token components for storage.
+ * Creates a pipe-separated string: `refreshToken|projectId|managedProjectId`
+ *
+ * @param refreshToken - The refresh token
+ * @param projectId - The GCP project ID
+ * @param managedProjectId - Optional managed project ID for enterprise users
+ * @returns Formatted string for storage
+ */
+export function formatTokenForStorage(
+  refreshToken: string,
+  projectId: string,
+  managedProjectId?: string
+): string {
+  return `${refreshToken}|${projectId}|${managedProjectId || ""}`
+}

src/auth/antigravity/tools.ts

@@ -0,0 +1,243 @@
+/**
+ * Antigravity Tool Normalization
+ * Converts tools between OpenAI and Gemini formats.
+ *
+ * OpenAI format:
+ *   { "type": "function", "function": { "name": "x", "description": "...", "parameters": {...} } }
+ *
+ * Gemini format:
+ *   { "functionDeclarations": [{ "name": "x", "description": "...", "parameters": {...} }] }
+ *
+ * Note: This is for Gemini models ONLY. Claude models are not supported via Antigravity.
+ */
+
+/**
+ * OpenAI function tool format
+ */
+export interface OpenAITool {
+  type: string
+  function?: {
+    name: string
+    description?: string
+    parameters?: Record<string, unknown>
+  }
+}
+
+/**
+ * Gemini function declaration format
+ */
+export interface GeminiFunctionDeclaration {
+  name: string
+  description?: string
+  parameters?: Record<string, unknown>
+}
+
+/**
+ * Gemini tools format (array of functionDeclarations)
+ */
+export interface GeminiTools {
+  functionDeclarations: GeminiFunctionDeclaration[]
+}
+
+/**
+ * OpenAI tool call in response
+ */
+export interface OpenAIToolCall {
+  id: string
+  type: "function"
+  function: {
+    name: string
+    arguments: string
+  }
+}
+
+/**
+ * Gemini function call in response
+ */
+export interface GeminiFunctionCall {
+  name: string
+  args: Record<string, unknown>
+}
+
+/**
+ * Gemini function response format
+ */
+export interface GeminiFunctionResponse {
+  name: string
+  response: Record<string, unknown>
+}
+
+/**
+ * Gemini tool result containing function calls
+ */
+export interface GeminiToolResult {
+  functionCall?: GeminiFunctionCall
+  functionResponse?: GeminiFunctionResponse
+}
+
+/**
+ * Normalize OpenAI-format tools to Gemini format.
+ * Converts an array of OpenAI tools to Gemini's functionDeclarations format.
+ *
+ * - Handles `function` type tools with name, description, parameters
+ * - Logs warning for unsupported tool types (does NOT silently drop them)
+ * - Creates a single object with functionDeclarations array
+ *
+ * @param tools - Array of OpenAI-format tools
+ * @returns Gemini-format tools object with functionDeclarations, or undefined if no valid tools
+ */
+export function normalizeToolsForGemini(
+  tools: OpenAITool[]
+): GeminiTools | undefined {
+  if (!tools || tools.length === 0) {
+    return undefined
+  }
+
+  const functionDeclarations: GeminiFunctionDeclaration[] = []
+
+  for (const tool of tools) {
+    if (!tool || typeof tool !== "object") {
+      continue
+    }
+
+    const toolType = tool.type ?? "function"
+    if (toolType === "function" && tool.function) {
+      const declaration: GeminiFunctionDeclaration = {
+        name: tool.function.name,
+      }
+
+      if (tool.function.description) {
+        declaration.description = tool.function.description
+      }
+
+      if (tool.function.parameters) {
+        declaration.parameters = tool.function.parameters
+      } else {
+        declaration.parameters = { type: "object", properties: {} }
+      }
+
+      functionDeclarations.push(declaration)
+    } else if (toolType !== "function" && process.env.ANTIGRAVITY_DEBUG === "1") {
+      console.warn(
+        `[antigravity-tools] Unsupported tool type: "${toolType}". Tool will be skipped.`
+      )
+    }
+  }
+
+  // Return undefined if no valid function declarations
+  if (functionDeclarations.length === 0) {
+    return undefined
+  }
+
+  return { functionDeclarations }
+}
+
+/**
+ * Convert Gemini tool results (functionCall) back to OpenAI tool_call format.
+ * Handles both functionCall (request) and functionResponse (result) formats.
+ *
+ * Gemini functionCall format:
+ *   { "name": "tool_name", "args": { ... } }
+ *
+ * OpenAI tool_call format:
+ *   { "id": "call_xxx", "type": "function", "function": { "name": "tool_name", "arguments": "..." } }
+ *
+ * @param results - Array of Gemini tool results containing functionCall or functionResponse
+ * @returns Array of OpenAI-format tool calls
+ */
+export function normalizeToolResultsFromGemini(
+  results: GeminiToolResult[]
+): OpenAIToolCall[] {
+  if (!results || results.length === 0) {
+    return []
+  }
+
+  const toolCalls: OpenAIToolCall[] = []
+  let callCounter = 0
+
+  for (const result of results) {
+    // Handle functionCall (tool invocation from model)
+    if (result.functionCall) {
+      callCounter++
+      const toolCall: OpenAIToolCall = {
+        id: `call_${Date.now()}_${callCounter}`,
+        type: "function",
+        function: {
+          name: result.functionCall.name,
+          arguments: JSON.stringify(result.functionCall.args ?? {}),
+        },
+      }
+      toolCalls.push(toolCall)
+    }
+  }
+
+  return toolCalls
+}
+
+/**
+ * Convert a single Gemini functionCall to OpenAI tool_call format.
+ * Useful for streaming responses where each chunk may contain a function call.
+ *
+ * @param functionCall - Gemini function call
+ * @param id - Optional tool call ID (generates one if not provided)
+ * @returns OpenAI-format tool call
+ */
+export function convertFunctionCallToToolCall(
+  functionCall: GeminiFunctionCall,
+  id?: string
+): OpenAIToolCall {
+  return {
+    id: id ?? `call_${Date.now()}_${Math.random().toString(36).slice(2, 8)}`,
+    type: "function",
+    function: {
+      name: functionCall.name,
+      arguments: JSON.stringify(functionCall.args ?? {}),
+    },
+  }
+}
+
+/**
+ * Check if a tool array contains any function-type tools.
+ *
+ * @param tools - Array of OpenAI-format tools
+ * @returns true if there are function tools to normalize
+ */
+export function hasFunctionTools(tools: OpenAITool[]): boolean {
+  if (!tools || tools.length === 0) {
+    return false
+  }
+
+  return tools.some((tool) => tool.type === "function" && tool.function)
+}
+
+/**
+ * Extract function declarations from already-normalized Gemini tools.
+ * Useful when tools may already be in Gemini format.
+ *
+ * @param tools - Tools that may be in Gemini or OpenAI format
+ * @returns Array of function declarations
+ */
+export function extractFunctionDeclarations(
+  tools: unknown
+): GeminiFunctionDeclaration[] {
+  if (!tools || typeof tools !== "object") {
+    return []
+  }
+
+  // Check if already in Gemini format
+  const geminiTools = tools as Record<string, unknown>
+  if (
+    Array.isArray(geminiTools.functionDeclarations) &&
+    geminiTools.functionDeclarations.length > 0
+  ) {
+    return geminiTools.functionDeclarations as GeminiFunctionDeclaration[]
+  }
+
+  // Check if it's an array of OpenAI tools
+  if (Array.isArray(tools)) {
+    const normalized = normalizeToolsForGemini(tools as OpenAITool[])
+    return normalized?.functionDeclarations ?? []
+  }
+
+  return []
+}

src/auth/antigravity/types.ts

@@ -0,0 +1,185 @@
+/**
+ * Antigravity Auth Type Definitions
+ * Matches cliproxyapi/sdk/auth/antigravity.go token format exactly
+ */
+
+/**
+ * Token storage format for Antigravity authentication
+ * Matches Go metadata structure: type, access_token, refresh_token, expires_in, timestamp, email, project_id
+ */
+export interface AntigravityTokens {
+  /** Always "antigravity" for this auth type */
+  type: "antigravity"
+  /** OAuth access token from Google */
+  access_token: string
+  /** OAuth refresh token from Google */
+  refresh_token: string
+  /** Token expiration time in seconds */
+  expires_in: number
+  /** Unix timestamp in milliseconds when tokens were obtained */
+  timestamp: number
+  /** ISO 8601 formatted expiration datetime (optional, for display) */
+  expired?: string
+  /** User's email address from Google userinfo */
+  email?: string
+  /** GCP project ID from loadCodeAssist API */
+  project_id?: string
+}
+
+/**
+ * Project context returned from loadCodeAssist API
+ * Used to get cloudaicompanionProject for API calls
+ */
+export interface AntigravityProjectContext {
+  /** GCP project ID for Cloud AI Companion */
+  cloudaicompanionProject?: string
+  /** Managed project ID for enterprise users (optional) */
+  managedProjectId?: string
+}
+
+/**
+ * Metadata for loadCodeAssist API request
+ */
+export interface AntigravityClientMetadata {
+  /** IDE type identifier */
+  ideType: "IDE_UNSPECIFIED" | string
+  /** Platform identifier */
+  platform: "PLATFORM_UNSPECIFIED" | string
+  /** Plugin type - typically "GEMINI" */
+  pluginType: "GEMINI" | string
+}
+
+/**
+ * Request body for loadCodeAssist API
+ */
+export interface AntigravityLoadCodeAssistRequest {
+  metadata: AntigravityClientMetadata
+}
+
+/**
+ * Response from loadCodeAssist API
+ */
+export interface AntigravityLoadCodeAssistResponse {
+  /** Project ID - can be string or object with id field */
+  cloudaicompanionProject?: string | { id: string }
+}
+
+/**
+ * Request body format for Antigravity API calls
+ * Wraps the actual request with project and model context
+ */
+export interface AntigravityRequestBody {
+  /** GCP project ID */
+  project: string
+  /** Model identifier (e.g., "gemini-3-pro-preview") */
+  model: string
+  /** User agent identifier */
+  userAgent: string
+  /** Unique request ID */
+  requestId: string
+  /** The actual request payload */
+  request: Record<string, unknown>
+}
+
+/**
+ * Response format from Antigravity API
+ * Follows OpenAI-compatible structure with Gemini extensions
+ */
+export interface AntigravityResponse {
+  /** Response ID */
+  id?: string
+  /** Object type (e.g., "chat.completion") */
+  object?: string
+  /** Creation timestamp */
+  created?: number
+  /** Model used for response */
+  model?: string
+  /** Response choices */
+  choices?: AntigravityResponseChoice[]
+  /** Token usage statistics */
+  usage?: AntigravityUsage
+  /** Error information if request failed */
+  error?: AntigravityError
+}
+
+/**
+ * Single response choice in Antigravity response
+ */
+export interface AntigravityResponseChoice {
+  /** Choice index */
+  index: number
+  /** Message content */
+  message?: {
+    role: "assistant"
+    content?: string
+    tool_calls?: AntigravityToolCall[]
+  }
+  /** Delta for streaming responses */
+  delta?: {
+    role?: "assistant"
+    content?: string
+    tool_calls?: AntigravityToolCall[]
+  }
+  /** Finish reason */
+  finish_reason?: "stop" | "tool_calls" | "length" | "content_filter" | null
+}
+
+/**
+ * Tool call in Antigravity response
+ */
+export interface AntigravityToolCall {
+  id: string
+  type: "function"
+  function: {
+    name: string
+    arguments: string
+  }
+}
+
+/**
+ * Token usage statistics
+ */
+export interface AntigravityUsage {
+  prompt_tokens: number
+  completion_tokens: number
+  total_tokens: number
+}
+
+/**
+ * Error response from Antigravity API
+ */
+export interface AntigravityError {
+  message: string
+  type?: string
+  code?: string | number
+}
+
+/**
+ * Token exchange result from Google OAuth
+ * Matches antigravityTokenResponse in Go
+ */
+export interface AntigravityTokenExchangeResult {
+  access_token: string
+  refresh_token: string
+  expires_in: number
+  token_type: string
+}
+
+/**
+ * User info from Google userinfo API
+ */
+export interface AntigravityUserInfo {
+  email: string
+  name?: string
+  picture?: string
+}
+
+/**
+ * Parsed refresh token parts
+ * Format: refreshToken|projectId|managedProjectId
+ */
+export interface AntigravityRefreshParts {
+  refreshToken: string
+  projectId?: string
+  managedProjectId?: string
+}

src/config/index.ts

@@ -4,6 +4,7 @@ export {
   AgentOverridesSchema,
   McpNameSchema,
   AgentNameSchema,
+  HookNameSchema,
 } from "./schema"
 
 export type {
@@ -12,4 +13,5 @@ export type {
   AgentOverrides,
   McpName,
   AgentName,
+  HookName,
 } from "./schema"

src/config/schema.ts

@@ -16,12 +16,44 @@ const AgentPermissionSchema = z.object({
   external_directory: PermissionValue.optional(),
 })
 
-export const AgentNameSchema = z.enum([
+export const BuiltinAgentNameSchema = z.enum([
   "oracle",
   "librarian",
   "explore",
   "frontend-ui-ux-engineer",
   "document-writer",
+  "multimodal-looker",
+])
+
+export const OverridableAgentNameSchema = z.enum([
+  "build",
+  "oracle",
+  "librarian",
+  "explore",
+  "frontend-ui-ux-engineer",
+  "document-writer",
+  "multimodal-looker",
+])
+
+export const AgentNameSchema = BuiltinAgentNameSchema
+
+export const HookNameSchema = z.enum([
+  "todo-continuation-enforcer",
+  "context-window-monitor",
+  "session-recovery",
+  "session-notification",
+  "comment-checker",
+  "grep-output-truncator",
+  "directory-agents-injector",
+  "directory-readme-injector",
+  "empty-task-response-detector",
+  "think-mode",
+  "anthropic-auto-compact",
+  "rules-injector",
+  "background-notification",
+  "auto-update-checker",
+  "ultrawork-mode",
+  "agent-usage-reminder",
 ])
 
 export const AgentOverrideConfigSchema = z.object({
@@ -40,18 +72,40 @@ export const AgentOverrideConfigSchema = z.object({
   permission: AgentPermissionSchema.optional(),
 })
 
-export const AgentOverridesSchema = z.record(AgentNameSchema, AgentOverrideConfigSchema)
+export const AgentOverridesSchema = z
+  .object({
+    build: AgentOverrideConfigSchema.optional(),
+    oracle: AgentOverrideConfigSchema.optional(),
+    librarian: AgentOverrideConfigSchema.optional(),
+    explore: AgentOverrideConfigSchema.optional(),
+    "frontend-ui-ux-engineer": AgentOverrideConfigSchema.optional(),
+    "document-writer": AgentOverrideConfigSchema.optional(),
+    "multimodal-looker": AgentOverrideConfigSchema.optional(),
+  })
+  .partial()
+
+export const ClaudeCodeConfigSchema = z.object({
+  mcp: z.boolean().optional(),
+  commands: z.boolean().optional(),
+  skills: z.boolean().optional(),
+  agents: z.boolean().optional(),
+  hooks: z.boolean().optional(),
+})
 
 export const OhMyOpenCodeConfigSchema = z.object({
   $schema: z.string().optional(),
   disabled_mcps: z.array(McpNameSchema).optional(),
-  disabled_agents: z.array(AgentNameSchema).optional(),
+  disabled_agents: z.array(BuiltinAgentNameSchema).optional(),
+  disabled_hooks: z.array(HookNameSchema).optional(),
   agents: AgentOverridesSchema.optional(),
+  claude_code: ClaudeCodeConfigSchema.optional(),
+  google_auth: z.boolean().optional(),
 })
 
 export type OhMyOpenCodeConfig = z.infer<typeof OhMyOpenCodeConfigSchema>
 export type AgentOverrideConfig = z.infer<typeof AgentOverrideConfigSchema>
 export type AgentOverrides = z.infer<typeof AgentOverridesSchema>
 export type AgentName = z.infer<typeof AgentNameSchema>
+export type HookName = z.infer<typeof HookNameSchema>
 
 export { McpNameSchema, type McpName } from "../mcp/types"

src/features/background-agent/index.ts

@@ -0,0 +1,2 @@
+export * from "./types"
+export { BackgroundManager } from "./manager"

src/features/background-agent/manager.ts

@@ -0,0 +1,367 @@
+import type { PluginInput } from "@opencode-ai/plugin"
+import type {
+  BackgroundTask,
+  LaunchInput,
+} from "./types"
+import { log } from "../../shared/logger"
+
+type OpencodeClient = PluginInput["client"]
+
+interface MessagePartInfo {
+  sessionID?: string
+  type?: string
+  tool?: string
+}
+
+interface EventProperties {
+  sessionID?: string
+  info?: { id?: string }
+  [key: string]: unknown
+}
+
+interface Event {
+  type: string
+  properties?: EventProperties
+}
+
+export class BackgroundManager {
+  private tasks: Map<string, BackgroundTask>
+  private notifications: Map<string, BackgroundTask[]>
+  private client: OpencodeClient
+  private directory: string
+  private pollingInterval?: Timer
+
+  constructor(ctx: PluginInput) {
+    this.tasks = new Map()
+    this.notifications = new Map()
+    this.client = ctx.client
+    this.directory = ctx.directory
+  }
+
+  async launch(input: LaunchInput): Promise<BackgroundTask> {
+    if (!input.agent || input.agent.trim() === "") {
+      throw new Error("Agent parameter is required")
+    }
+
+    const createResult = await this.client.session.create({
+      body: {
+        parentID: input.parentSessionID,
+        title: `Background: ${input.description}`,
+      },
+    })
+
+    if (createResult.error) {
+      throw new Error(`Failed to create background session: ${createResult.error}`)
+    }
+
+    const sessionID = createResult.data.id
+
+    const task: BackgroundTask = {
+      id: `bg_${crypto.randomUUID().slice(0, 8)}`,
+      sessionID,
+      parentSessionID: input.parentSessionID,
+      parentMessageID: input.parentMessageID,
+      description: input.description,
+      prompt: input.prompt,
+      agent: input.agent,
+      status: "running",
+      startedAt: new Date(),
+      progress: {
+        toolCalls: 0,
+        lastUpdate: new Date(),
+      },
+    }
+
+    this.tasks.set(task.id, task)
+    this.startPolling()
+
+    log("[background-agent] Launching task:", { taskId: task.id, sessionID, agent: input.agent })
+
+    this.client.session.promptAsync({
+      path: { id: sessionID },
+      body: {
+        agent: input.agent,
+        tools: {
+          background_task: false,
+          background_output: false,
+          background_cancel: false,
+          call_omo_agent: false,
+        },
+        parts: [{ type: "text", text: input.prompt }],
+      },
+    }).catch((error) => {
+      log("[background-agent] promptAsync error:", error)
+      const existingTask = this.findBySession(sessionID)
+      if (existingTask) {
+        existingTask.status = "error"
+        const errorMessage = error instanceof Error ? error.message : String(error)
+        if (errorMessage.includes("agent.name") || errorMessage.includes("undefined")) {
+          existingTask.error = `Agent "${input.agent}" not found. Make sure the agent is registered in your opencode.json or provided by a plugin.`
+        } else {
+          existingTask.error = errorMessage
+        }
+        existingTask.completedAt = new Date()
+        this.markForNotification(existingTask)
+        this.notifyParentSession(existingTask)
+      }
+    })
+
+    return task
+  }
+
+  getTask(id: string): BackgroundTask | undefined {
+    return this.tasks.get(id)
+  }
+
+  getTasksByParentSession(sessionID: string): BackgroundTask[] {
+    const result: BackgroundTask[] = []
+    for (const task of this.tasks.values()) {
+      if (task.parentSessionID === sessionID) {
+        result.push(task)
+      }
+    }
+    return result
+  }
+
+  findBySession(sessionID: string): BackgroundTask | undefined {
+    for (const task of this.tasks.values()) {
+      if (task.sessionID === sessionID) {
+        return task
+      }
+    }
+    return undefined
+  }
+
+  handleEvent(event: Event): void {
+    const props = event.properties
+
+    if (event.type === "message.part.updated") {
+      if (!props || typeof props !== "object" || !("sessionID" in props)) return
+      const partInfo = props as unknown as MessagePartInfo
+      const sessionID = partInfo?.sessionID
+      if (!sessionID) return
+
+      const task = this.findBySession(sessionID)
+      if (!task) return
+
+      if (partInfo?.type === "tool" || partInfo?.tool) {
+        if (!task.progress) {
+          task.progress = {
+            toolCalls: 0,
+            lastUpdate: new Date(),
+          }
+        }
+        task.progress.toolCalls += 1
+        task.progress.lastTool = partInfo.tool
+        task.progress.lastUpdate = new Date()
+      }
+    }
+
+    if (event.type === "session.idle") {
+      const sessionID = props?.sessionID as string | undefined
+      if (!sessionID) return
+
+      const task = this.findBySession(sessionID)
+      if (!task || task.status !== "running") return
+
+      task.status = "completed"
+      task.completedAt = new Date()
+      this.markForNotification(task)
+      this.notifyParentSession(task)
+      log("[background-agent] Task completed via session.idle event:", task.id)
+    }
+
+    if (event.type === "session.deleted") {
+      const info = props?.info
+      if (!info || typeof info.id !== "string") return
+      const sessionID = info.id
+
+      const task = this.findBySession(sessionID)
+      if (!task) return
+
+      if (task.status === "running") {
+        task.status = "cancelled"
+        task.completedAt = new Date()
+        task.error = "Session deleted"
+      }
+
+      this.tasks.delete(task.id)
+      this.clearNotificationsForTask(task.id)
+    }
+  }
+
+  markForNotification(task: BackgroundTask): void {
+    const queue = this.notifications.get(task.parentSessionID) ?? []
+    queue.push(task)
+    this.notifications.set(task.parentSessionID, queue)
+  }
+
+  getPendingNotifications(sessionID: string): BackgroundTask[] {
+    return this.notifications.get(sessionID) ?? []
+  }
+
+  clearNotifications(sessionID: string): void {
+    this.notifications.delete(sessionID)
+  }
+
+  private clearNotificationsForTask(taskId: string): void {
+    for (const [sessionID, tasks] of this.notifications.entries()) {
+      const filtered = tasks.filter((t) => t.id !== taskId)
+      if (filtered.length === 0) {
+        this.notifications.delete(sessionID)
+      } else {
+        this.notifications.set(sessionID, filtered)
+      }
+    }
+  }
+
+  private startPolling(): void {
+    if (this.pollingInterval) return
+
+    this.pollingInterval = setInterval(() => {
+      this.pollRunningTasks()
+    }, 2000)
+  }
+
+  private stopPolling(): void {
+    if (this.pollingInterval) {
+      clearInterval(this.pollingInterval)
+      this.pollingInterval = undefined
+    }
+  }
+
+  private notifyParentSession(task: BackgroundTask): void {
+    const duration = this.formatDuration(task.startedAt, task.completedAt)
+
+    log("[background-agent] notifyParentSession called for task:", task.id)
+
+    // eslint-disable-next-line @typescript-eslint/no-explicit-any
+    const tuiClient = this.client as any
+    if (tuiClient.tui?.showToast) {
+      tuiClient.tui.showToast({
+        body: {
+          title: "Background Task Completed",
+          message: `Task "${task.description}" finished in ${duration}.`,
+          variant: "success",
+          duration: 5000,
+        },
+      }).catch(() => {})
+    }
+
+    const message = `[BACKGROUND TASK COMPLETED] Task "${task.description}" finished in ${duration}. Use background_output with task_id="${task.id}" to get results.`
+
+    log("[background-agent] Sending notification to parent session:", { parentSessionID: task.parentSessionID })
+
+    setTimeout(async () => {
+      try {
+        await this.client.session.prompt({
+          path: { id: task.parentSessionID },
+          body: {
+            parts: [{ type: "text", text: message }],
+          },
+          query: { directory: this.directory },
+        })
+        this.clearNotificationsForTask(task.id)
+        log("[background-agent] Successfully sent prompt to parent session:", { parentSessionID: task.parentSessionID })
+      } catch (error) {
+        log("[background-agent] prompt failed:", String(error))
+      }
+    }, 200)
+  }
+
+  private formatDuration(start: Date, end?: Date): string {
+    const duration = (end ?? new Date()).getTime() - start.getTime()
+    const seconds = Math.floor(duration / 1000)
+    const minutes = Math.floor(seconds / 60)
+    const hours = Math.floor(minutes / 60)
+
+    if (hours > 0) {
+      return `${hours}h ${minutes % 60}m ${seconds % 60}s`
+    } else if (minutes > 0) {
+      return `${minutes}m ${seconds % 60}s`
+    }
+    return `${seconds}s`
+  }
+
+  private hasRunningTasks(): boolean {
+    for (const task of this.tasks.values()) {
+      if (task.status === "running") return true
+    }
+    return false
+  }
+
+  private async pollRunningTasks(): Promise<void> {
+    const statusResult = await this.client.session.status()
+    const allStatuses = (statusResult.data ?? {}) as Record<string, { type: string }>
+
+    for (const task of this.tasks.values()) {
+      if (task.status !== "running") continue
+
+      try {
+        const sessionStatus = allStatuses[task.sessionID]
+        
+        if (!sessionStatus) {
+          log("[background-agent] Session not found in status:", task.sessionID)
+          continue
+        }
+
+        if (sessionStatus.type === "idle") {
+          task.status = "completed"
+          task.completedAt = new Date()
+          this.markForNotification(task)
+          this.notifyParentSession(task)
+          log("[background-agent] Task completed via polling:", task.id)
+          continue
+        }
+
+        const messagesResult = await this.client.session.messages({
+          path: { id: task.sessionID },
+        })
+
+        if (!messagesResult.error && messagesResult.data) {
+          const messages = messagesResult.data as Array<{
+            info?: { role?: string }
+            parts?: Array<{ type?: string; tool?: string; name?: string; text?: string }>
+          }>
+          const assistantMsgs = messages.filter(
+            (m) => m.info?.role === "assistant"
+          )
+
+          let toolCalls = 0
+          let lastTool: string | undefined
+          let lastMessage: string | undefined
+
+          for (const msg of assistantMsgs) {
+            const parts = msg.parts ?? []
+            for (const part of parts) {
+              if (part.type === "tool_use" || part.tool) {
+                toolCalls++
+                lastTool = part.tool || part.name || "unknown"
+              }
+              if (part.type === "text" && part.text) {
+                lastMessage = part.text
+              }
+            }
+          }
+
+          if (!task.progress) {
+            task.progress = { toolCalls: 0, lastUpdate: new Date() }
+          }
+          task.progress.toolCalls = toolCalls
+          task.progress.lastTool = lastTool
+          task.progress.lastUpdate = new Date()
+          if (lastMessage) {
+            task.progress.lastMessage = lastMessage
+            task.progress.lastMessageAt = new Date()
+          }
+        }
+      } catch (error) {
+        log("[background-agent] Poll error for task:", { taskId: task.id, error })
+      }
+    }
+
+    if (!this.hasRunningTasks()) {
+      this.stopPolling()
+    }
+  }
+}

src/features/background-agent/types.ts

@@ -0,0 +1,37 @@
+export type BackgroundTaskStatus =
+  | "running"
+  | "completed"
+  | "error"
+  | "cancelled"
+
+export interface TaskProgress {
+  toolCalls: number
+  lastTool?: string
+  lastUpdate: Date
+  lastMessage?: string
+  lastMessageAt?: Date
+}
+
+export interface BackgroundTask {
+  id: string
+  sessionID: string
+  parentSessionID: string
+  parentMessageID: string
+  description: string
+  prompt: string
+  agent: string
+  status: BackgroundTaskStatus
+  startedAt: Date
+  completedAt?: Date
+  result?: string
+  error?: string
+  progress?: TaskProgress
+}
+
+export interface LaunchInput {
+  description: string
+  prompt: string
+  agent: string
+  parentSessionID: string
+  parentMessageID: string
+}

src/features/claude-code-agent-loader/loader.ts

@@ -3,6 +3,7 @@ import { homedir } from "os"
 import { join, basename } from "path"
 import type { AgentConfig } from "@opencode-ai/sdk"
 import { parseFrontmatter } from "../../shared/frontmatter"
+import { isMarkdownFile } from "../../shared/file-utils"
 import type { AgentScope, AgentFrontmatter, LoadedAgent } from "./types"
 
 function parseToolsConfig(toolsStr?: string): Record<string, boolean> | undefined {
@@ -18,10 +19,6 @@ function parseToolsConfig(toolsStr?: string): Record<string, boolean> | undefine
   return result
 }
 
-function isMarkdownFile(entry: { name: string; isFile: () => boolean }): boolean {
-  return !entry.name.startsWith(".") && entry.name.endsWith(".md") && entry.isFile()
-}
-
 function loadAgentsFromDir(agentsDir: string, scope: AgentScope): LoadedAgent[] {
   if (!existsSync(agentsDir)) {
     return []

src/features/claude-code-command-loader/loader.ts

@@ -3,12 +3,9 @@ import { homedir } from "os"
 import { join, basename } from "path"
 import { parseFrontmatter } from "../../shared/frontmatter"
 import { sanitizeModelField } from "../../shared/model-sanitizer"
+import { isMarkdownFile } from "../../shared/file-utils"
 import type { CommandScope, CommandDefinition, CommandFrontmatter, LoadedCommand } from "./types"
 
-function isMarkdownFile(entry: { name: string; isFile: () => boolean }): boolean {
-  return !entry.name.startsWith(".") && entry.name.endsWith(".md") && entry.isFile()
-}
-
 function loadCommandsFromDir(commandsDir: string, scope: CommandScope): LoadedCommand[] {
   if (!existsSync(commandsDir)) {
     return []

src/features/claude-code-skill-loader/loader.ts

@@ -1,8 +1,9 @@
-import { existsSync, readdirSync, readFileSync, statSync, readlinkSync } from "fs"
+import { existsSync, readdirSync, readFileSync } from "fs"
 import { homedir } from "os"
-import { join, resolve } from "path"
+import { join } from "path"
 import { parseFrontmatter } from "../../shared/frontmatter"
 import { sanitizeModelField } from "../../shared/model-sanitizer"
+import { resolveSymlink } from "../../shared/file-utils"
 import type { CommandDefinition } from "../claude-code-command-loader/types"
 import type { SkillScope, SkillMetadata, LoadedSkillAsCommand } from "./types"
 
@@ -21,10 +22,7 @@ function loadSkillsFromDir(skillsDir: string, scope: SkillScope): LoadedSkillAsC
 
     if (!entry.isDirectory() && !entry.isSymbolicLink()) continue
 
-    let resolvedPath = skillPath
-    if (statSync(skillPath, { throwIfNoEntry: false })?.isSymbolicLink()) {
-      resolvedPath = resolve(skillPath, "..", readlinkSync(skillPath))
-    }
+    const resolvedPath = resolveSymlink(skillPath)
 
     const skillMdPath = join(resolvedPath, "SKILL.md")
     if (!existsSync(skillMdPath)) continue

src/google-auth.ts

@@ -0,0 +1,8 @@
+import type { Plugin } from "@opencode-ai/plugin"
+import { createGoogleAntigravityAuthPlugin } from "./auth/antigravity"
+
+const GoogleAntigravityAuthPlugin: Plugin = async (ctx) => {
+  return createGoogleAntigravityAuthPlugin(ctx)
+}
+
+export default GoogleAntigravityAuthPlugin

src/hooks/agent-usage-reminder/constants.ts

@@ -0,0 +1,53 @@
+import { join } from "node:path";
+import { xdgData } from "xdg-basedir";
+
+export const OPENCODE_STORAGE = join(xdgData ?? "", "opencode", "storage");
+export const AGENT_USAGE_REMINDER_STORAGE = join(
+  OPENCODE_STORAGE,
+  "agent-usage-reminder",
+);
+
+// All tool names normalized to lowercase for case-insensitive matching
+export const TARGET_TOOLS = new Set([
+  "grep",
+  "safe_grep",
+  "glob",
+  "safe_glob",
+  "webfetch",
+  "context7_resolve-library-id",
+  "context7_get-library-docs",
+  "websearch_exa_web_search_exa",
+  "grep_app_searchgithub",
+]);
+
+export const AGENT_TOOLS = new Set([
+  "task",
+  "call_omo_agent",
+  "background_task",
+]);
+
+export const REMINDER_MESSAGE = `
+[Agent Usage Reminder]
+
+You called a search/fetch tool directly without leveraging specialized agents.
+
+RECOMMENDED: Use background_task with explore/librarian agents for better results:
+
+\`\`\`
+// Parallel exploration - fire multiple agents simultaneously
+background_task(agent="explore", prompt="Find all files matching pattern X")
+background_task(agent="explore", prompt="Search for implementation of Y") 
+background_task(agent="librarian", prompt="Lookup documentation for Z")
+
+// Then continue your work while they run in background
+// System will notify you when each completes
+\`\`\`
+
+WHY:
+- Agents can perform deeper, more thorough searches
+- Background tasks run in parallel, saving time
+- Specialized agents have domain expertise
+- Reduces context window usage in main session
+
+ALWAYS prefer: Multiple parallel background_task calls > Direct tool calls
+`;

src/hooks/agent-usage-reminder/index.ts

@@ -0,0 +1,109 @@
+import type { PluginInput } from "@opencode-ai/plugin";
+import {
+  loadAgentUsageState,
+  saveAgentUsageState,
+  clearAgentUsageState,
+} from "./storage";
+import { TARGET_TOOLS, AGENT_TOOLS, REMINDER_MESSAGE } from "./constants";
+import type { AgentUsageState } from "./types";
+
+interface ToolExecuteInput {
+  tool: string;
+  sessionID: string;
+  callID: string;
+}
+
+interface ToolExecuteOutput {
+  title: string;
+  output: string;
+  metadata: unknown;
+}
+
+interface EventInput {
+  event: {
+    type: string;
+    properties?: unknown;
+  };
+}
+
+export function createAgentUsageReminderHook(_ctx: PluginInput) {
+  const sessionStates = new Map<string, AgentUsageState>();
+
+  function getOrCreateState(sessionID: string): AgentUsageState {
+    if (!sessionStates.has(sessionID)) {
+      const persisted = loadAgentUsageState(sessionID);
+      const state: AgentUsageState = persisted ?? {
+        sessionID,
+        agentUsed: false,
+        reminderCount: 0,
+        updatedAt: Date.now(),
+      };
+      sessionStates.set(sessionID, state);
+    }
+    return sessionStates.get(sessionID)!;
+  }
+
+  function markAgentUsed(sessionID: string): void {
+    const state = getOrCreateState(sessionID);
+    state.agentUsed = true;
+    state.updatedAt = Date.now();
+    saveAgentUsageState(state);
+  }
+
+  function resetState(sessionID: string): void {
+    sessionStates.delete(sessionID);
+    clearAgentUsageState(sessionID);
+  }
+
+  const toolExecuteAfter = async (
+    input: ToolExecuteInput,
+    output: ToolExecuteOutput,
+  ) => {
+    const { tool, sessionID } = input;
+    const toolLower = tool.toLowerCase();
+
+    if (AGENT_TOOLS.has(toolLower)) {
+      markAgentUsed(sessionID);
+      return;
+    }
+
+    if (!TARGET_TOOLS.has(toolLower)) {
+      return;
+    }
+
+    const state = getOrCreateState(sessionID);
+
+    if (state.agentUsed) {
+      return;
+    }
+
+    output.output += REMINDER_MESSAGE;
+    state.reminderCount++;
+    state.updatedAt = Date.now();
+    saveAgentUsageState(state);
+  };
+
+  const eventHandler = async ({ event }: EventInput) => {
+    const props = event.properties as Record<string, unknown> | undefined;
+
+    if (event.type === "session.deleted") {
+      const sessionInfo = props?.info as { id?: string } | undefined;
+      if (sessionInfo?.id) {
+        resetState(sessionInfo.id);
+      }
+    }
+
+    if (event.type === "session.compacted") {
+      const sessionID = (props?.sessionID ??
+        (props?.info as { id?: string } | undefined)?.id) as string | undefined;
+      if (sessionID) {
+        resetState(sessionID);
+      }
+    }
+  };
+
+  return {
+    "tool.execute.after": toolExecuteAfter,
+    event: eventHandler,
+  };
+}

src/hooks/agent-usage-reminder/storage.ts

@@ -0,0 +1,42 @@
+import {
+  existsSync,
+  mkdirSync,
+  readFileSync,
+  writeFileSync,
+  unlinkSync,
+} from "node:fs";
+import { join } from "node:path";
+import { AGENT_USAGE_REMINDER_STORAGE } from "./constants";
+import type { AgentUsageState } from "./types";
+
+function getStoragePath(sessionID: string): string {
+  return join(AGENT_USAGE_REMINDER_STORAGE, `${sessionID}.json`);
+}
+
+export function loadAgentUsageState(sessionID: string): AgentUsageState | null {
+  const filePath = getStoragePath(sessionID);
+  if (!existsSync(filePath)) return null;
+
+  try {
+    const content = readFileSync(filePath, "utf-8");
+    return JSON.parse(content) as AgentUsageState;
+  } catch {
+    return null;
+  }
+}
+
+export function saveAgentUsageState(state: AgentUsageState): void {
+  if (!existsSync(AGENT_USAGE_REMINDER_STORAGE)) {
+    mkdirSync(AGENT_USAGE_REMINDER_STORAGE, { recursive: true });
+  }
+
+  const filePath = getStoragePath(state.sessionID);
+  writeFileSync(filePath, JSON.stringify(state, null, 2));
+}
+
+export function clearAgentUsageState(sessionID: string): void {
+  const filePath = getStoragePath(sessionID);
+  if (existsSync(filePath)) {
+    unlinkSync(filePath);
+  }
+}

src/hooks/agent-usage-reminder/types.ts

@@ -0,0 +1,6 @@
+export interface AgentUsageState {
+  sessionID: string;
+  agentUsed: boolean;
+  reminderCount: number;
+  updatedAt: number;
+}

src/hooks/anthropic-auto-compact/executor.ts

@@ -1,4 +1,5 @@
-import type { AutoCompactState } from "./types"
+import type { AutoCompactState, RetryState } from "./types"
+import { RETRY_CONFIG } from "./types"
 
 type Client = {
   session: {
@@ -11,9 +12,34 @@ type Client = {
   }
   tui: {
     submitPrompt: (opts: { query: { directory: string } }) => Promise<unknown>
+    showToast: (opts: {
+      body: { title: string; message: string; variant: string; duration: number }
+    }) => Promise<unknown>
   }
 }
 
+function calculateRetryDelay(attempt: number): number {
+  const delay = RETRY_CONFIG.initialDelayMs * Math.pow(RETRY_CONFIG.backoffFactor, attempt - 1)
+  return Math.min(delay, RETRY_CONFIG.maxDelayMs)
+}
+
+function shouldRetry(retryState: RetryState | undefined): boolean {
+  if (!retryState) return true
+  return retryState.attempt < RETRY_CONFIG.maxAttempts
+}
+
+function getOrCreateRetryState(
+  autoCompactState: AutoCompactState,
+  sessionID: string
+): RetryState {
+  let state = autoCompactState.retryStateBySession.get(sessionID)
+  if (!state) {
+    state = { attempt: 0, lastAttemptTime: 0 }
+    autoCompactState.retryStateBySession.set(sessionID, state)
+  }
+  return state
+}
+
 export async function getLastAssistant(
   sessionID: string,
   // eslint-disable-next-line @typescript-eslint/no-explicit-any
@@ -42,6 +68,12 @@ export async function getLastAssistant(
   }
 }
 
+function clearSessionState(autoCompactState: AutoCompactState, sessionID: string): void {
+  autoCompactState.pendingCompact.delete(sessionID)
+  autoCompactState.errorDataBySession.delete(sessionID)
+  autoCompactState.retryStateBySession.delete(sessionID)
+}
+
 export async function executeCompact(
   sessionID: string,
   msg: Record<string, unknown>,
@@ -50,6 +82,27 @@ export async function executeCompact(
   client: any,
   directory: string
 ): Promise<void> {
+  const retryState = getOrCreateRetryState(autoCompactState, sessionID)
+
+  if (!shouldRetry(retryState)) {
+    clearSessionState(autoCompactState, sessionID)
+
+    await (client as Client).tui
+      .showToast({
+        body: {
+          title: "Auto Compact Failed",
+          message: `Failed after ${RETRY_CONFIG.maxAttempts} attempts. Please try manual compact.`,
+          variant: "error",
+          duration: 5000,
+        },
+      })
+      .catch(() => {})
+    return
+  }
+
+  retryState.attempt++
+  retryState.lastAttemptTime = Date.now()
+
   try {
     const providerID = msg.providerID as string | undefined
     const modelID = msg.modelID as string | undefined
@@ -61,14 +114,30 @@ export async function executeCompact(
         query: { directory },
       })
 
+      clearSessionState(autoCompactState, sessionID)
+
       setTimeout(async () => {
         try {
           await (client as Client).tui.submitPrompt({ query: { directory } })
         } catch {}
       }, 500)
     }
+  } catch {
+    const delay = calculateRetryDelay(retryState.attempt)
 
-    autoCompactState.pendingCompact.delete(sessionID)
-    autoCompactState.errorDataBySession.delete(sessionID)
-  } catch {}
+    await (client as Client).tui
+      .showToast({
+        body: {
+          title: "Auto Compact Retry",
+          message: `Attempt ${retryState.attempt}/${RETRY_CONFIG.maxAttempts} failed. Retrying in ${Math.round(delay / 1000)}s...`,
+          variant: "warning",
+          duration: delay,
+        },
+      })
+      .catch(() => {})
+
+    setTimeout(() => {
+      executeCompact(sessionID, msg, autoCompactState, client, directory)
+    }, delay)
+  }
 }

src/hooks/anthropic-auto-compact/index.ts

@@ -7,6 +7,7 @@ function createAutoCompactState(): AutoCompactState {
   return {
     pendingCompact: new Set<string>(),
     errorDataBySession: new Map<string, ParsedTokenLimitError>(),
+    retryStateBySession: new Map(),
   }
 }
 
@@ -21,6 +22,7 @@ export function createAnthropicAutoCompactHook(ctx: PluginInput) {
       if (sessionInfo?.id) {
         autoCompactState.pendingCompact.delete(sessionInfo.id)
         autoCompactState.errorDataBySession.delete(sessionInfo.id)
+        autoCompactState.retryStateBySession.delete(sessionInfo.id)
       }
       return
     }

src/hooks/anthropic-auto-compact/types.ts

@@ -7,7 +7,20 @@ export interface ParsedTokenLimitError {
   modelID?: string
 }
 
+export interface RetryState {
+  attempt: number
+  lastAttemptTime: number
+}
+
 export interface AutoCompactState {
   pendingCompact: Set<string>
   errorDataBySession: Map<string, ParsedTokenLimitError>
+  retryStateBySession: Map<string, RetryState>
 }
+
+export const RETRY_CONFIG = {
+  maxAttempts: 5,
+  initialDelayMs: 2000,
+  backoffFactor: 2,
+  maxDelayMs: 30000,
+} as const

src/hooks/auto-update-checker/cache.ts

@@ -0,0 +1,18 @@
+import * as fs from "node:fs"
+import { VERSION_FILE } from "./constants"
+import { log } from "../../shared/logger"
+
+export function invalidateCache(): boolean {
+  try {
+    if (fs.existsSync(VERSION_FILE)) {
+      fs.unlinkSync(VERSION_FILE)
+      log(`[auto-update-checker] Cache invalidated: ${VERSION_FILE}`)
+      return true
+    }
+    log("[auto-update-checker] Version file not found, nothing to invalidate")
+    return false
+  } catch (err) {
+    log("[auto-update-checker] Failed to invalidate cache:", err)
+    return false
+  }
+}

src/hooks/auto-update-checker/checker.ts

@@ -0,0 +1,119 @@
+import * as fs from "node:fs"
+import * as path from "node:path"
+import type { NpmDistTags, OpencodeConfig, PackageJson, UpdateCheckResult } from "./types"
+import {
+  PACKAGE_NAME,
+  NPM_REGISTRY_URL,
+  NPM_FETCH_TIMEOUT,
+  INSTALLED_PACKAGE_JSON,
+  USER_OPENCODE_CONFIG,
+} from "./constants"
+import { log } from "../../shared/logger"
+
+export function isLocalDevMode(directory: string): boolean {
+  const projectConfig = path.join(directory, ".opencode", "opencode.json")
+
+  for (const configPath of [projectConfig, USER_OPENCODE_CONFIG]) {
+    try {
+      if (!fs.existsSync(configPath)) continue
+      const content = fs.readFileSync(configPath, "utf-8")
+      const config = JSON.parse(content) as OpencodeConfig
+      const plugins = config.plugin ?? []
+
+      for (const entry of plugins) {
+        if (entry.startsWith("file://") && entry.includes(PACKAGE_NAME)) {
+          return true
+        }
+      }
+    } catch {
+      continue
+    }
+  }
+
+  return false
+}
+
+export function findPluginEntry(directory: string): string | null {
+  const projectConfig = path.join(directory, ".opencode", "opencode.json")
+
+  for (const configPath of [projectConfig, USER_OPENCODE_CONFIG]) {
+    try {
+      if (!fs.existsSync(configPath)) continue
+      const content = fs.readFileSync(configPath, "utf-8")
+      const config = JSON.parse(content) as OpencodeConfig
+      const plugins = config.plugin ?? []
+
+      for (const entry of plugins) {
+        if (entry === PACKAGE_NAME || entry.startsWith(`${PACKAGE_NAME}@`)) {
+          return entry
+        }
+      }
+    } catch {
+      continue
+    }
+  }
+
+  return null
+}
+
+export function getCachedVersion(): string | null {
+  try {
+    if (!fs.existsSync(INSTALLED_PACKAGE_JSON)) return null
+    const content = fs.readFileSync(INSTALLED_PACKAGE_JSON, "utf-8")
+    const pkg = JSON.parse(content) as PackageJson
+    return pkg.version ?? null
+  } catch {
+    return null
+  }
+}
+
+export async function getLatestVersion(): Promise<string | null> {
+  const controller = new AbortController()
+  const timeoutId = setTimeout(() => controller.abort(), NPM_FETCH_TIMEOUT)
+
+  try {
+    const response = await fetch(NPM_REGISTRY_URL, {
+      signal: controller.signal,
+      headers: { Accept: "application/json" },
+    })
+
+    if (!response.ok) return null
+
+    const data = (await response.json()) as NpmDistTags
+    return data.latest ?? null
+  } catch {
+    return null
+  } finally {
+    clearTimeout(timeoutId)
+  }
+}
+
+export async function checkForUpdate(directory: string): Promise<UpdateCheckResult> {
+  if (isLocalDevMode(directory)) {
+    log("[auto-update-checker] Local dev mode detected, skipping update check")
+    return { needsUpdate: false, currentVersion: null, latestVersion: null, isLocalDev: true }
+  }
+
+  const pluginEntry = findPluginEntry(directory)
+  if (!pluginEntry) {
+    log("[auto-update-checker] Plugin not found in config")
+    return { needsUpdate: false, currentVersion: null, latestVersion: null, isLocalDev: false }
+  }
+
+  const currentVersion = getCachedVersion()
+  if (!currentVersion) {
+    log("[auto-update-checker] No cached version found")
+    return { needsUpdate: false, currentVersion: null, latestVersion: null, isLocalDev: false }
+  }
+
+  const latestVersion = await getLatestVersion()
+  if (!latestVersion) {
+    log("[auto-update-checker] Failed to fetch latest version")
+    return { needsUpdate: false, currentVersion, latestVersion: null, isLocalDev: false }
+  }
+
+  const needsUpdate = currentVersion !== latestVersion
+  log(`[auto-update-checker] Current: ${currentVersion}, Latest: ${latestVersion}, NeedsUpdate: ${needsUpdate}`)
+
+  return { needsUpdate, currentVersion, latestVersion, isLocalDev: false }
+}

src/hooks/auto-update-checker/constants.ts

@@ -0,0 +1,40 @@
+import * as path from "node:path"
+import * as os from "node:os"
+
+export const PACKAGE_NAME = "oh-my-opencode"
+export const NPM_REGISTRY_URL = `https://registry.npmjs.org/-/package/${PACKAGE_NAME}/dist-tags`
+export const NPM_FETCH_TIMEOUT = 5000
+
+/**
+ * OpenCode plugin cache directory
+ * - Linux/macOS: ~/.cache/opencode/
+ * - Windows: %LOCALAPPDATA%/opencode/
+ */
+function getCacheDir(): string {
+  if (process.platform === "win32") {
+    return path.join(process.env.LOCALAPPDATA ?? os.homedir(), "opencode")
+  }
+  return path.join(os.homedir(), ".cache", "opencode")
+}
+
+export const CACHE_DIR = getCacheDir()
+export const VERSION_FILE = path.join(CACHE_DIR, "version")
+export const INSTALLED_PACKAGE_JSON = path.join(
+  CACHE_DIR,
+  "node_modules",
+  PACKAGE_NAME,
+  "package.json"
+)
+
+/**
+ * OpenCode config file locations (priority order)
+ */
+function getUserConfigDir(): string {
+  if (process.platform === "win32") {
+    return process.env.APPDATA ?? path.join(os.homedir(), "AppData", "Roaming")
+  }
+  return process.env.XDG_CONFIG_HOME ?? path.join(os.homedir(), ".config")
+}
+
+export const USER_CONFIG_DIR = getUserConfigDir()
+export const USER_OPENCODE_CONFIG = path.join(USER_CONFIG_DIR, "opencode", "opencode.json")

src/hooks/auto-update-checker/index.ts

@@ -0,0 +1,56 @@
+import type { PluginInput } from "@opencode-ai/plugin"
+import { checkForUpdate } from "./checker"
+import { invalidateCache } from "./cache"
+import { PACKAGE_NAME } from "./constants"
+import { log } from "../../shared/logger"
+
+export function createAutoUpdateCheckerHook(ctx: PluginInput) {
+  let hasChecked = false
+
+  return {
+    event: async ({ event }: { event: { type: string; properties?: unknown } }) => {
+      if (event.type !== "session.created") return
+      if (hasChecked) return
+
+      const props = event.properties as { info?: { parentID?: string } } | undefined
+      if (props?.info?.parentID) return
+
+      hasChecked = true
+
+      try {
+        const result = await checkForUpdate(ctx.directory)
+
+        if (result.isLocalDev) {
+          log("[auto-update-checker] Skipped: local development mode")
+          return
+        }
+
+        if (!result.needsUpdate) {
+          log("[auto-update-checker] No update needed")
+          return
+        }
+
+        invalidateCache()
+
+        await ctx.client.tui
+          .showToast({
+            body: {
+              title: `${PACKAGE_NAME} Update`,
+              message: `v${result.latestVersion} available (current: v${result.currentVersion}). Restart OpenCode to apply.`,
+              variant: "info" as const,
+              duration: 8000,
+            },
+          })
+          .catch(() => {})
+
+        log(`[auto-update-checker] Update notification sent: v${result.currentVersion} → v${result.latestVersion}`)
+      } catch (err) {
+        log("[auto-update-checker] Error during update check:", err)
+      }
+    },
+  }
+}
+
+export type { UpdateCheckResult } from "./types"
+export { checkForUpdate } from "./checker"
+export { invalidateCache } from "./cache"

src/hooks/auto-update-checker/types.ts

@@ -0,0 +1,22 @@
+export interface NpmDistTags {
+  latest: string
+  [key: string]: string
+}
+
+export interface OpencodeConfig {
+  plugin?: string[]
+  [key: string]: unknown
+}
+
+export interface PackageJson {
+  version: string
+  name?: string
+  [key: string]: unknown
+}
+
+export interface UpdateCheckResult {
+  needsUpdate: boolean
+  currentVersion: string | null
+  latestVersion: string | null
+  isLocalDev: boolean
+}

src/hooks/background-notification/index.ts

@@ -0,0 +1,22 @@
+import type { BackgroundManager } from "../../features/background-agent"
+
+interface Event {
+  type: string
+  properties?: Record<string, unknown>
+}
+
+interface EventInput {
+  event: Event
+}
+
+export function createBackgroundNotificationHook(manager: BackgroundManager) {
+  const eventHandler = async ({ event }: EventInput) => {
+    manager.handleEvent(event)
+  }
+
+  return {
+    event: eventHandler,
+  }
+}
+
+export type { BackgroundNotificationHookConfig } from "./types"

src/hooks/background-notification/types.ts

@@ -0,0 +1,5 @@
+import type { BackgroundTask } from "../../features/background-agent"
+
+export interface BackgroundNotificationHookConfig {
+  formatNotification?: (tasks: BackgroundTask[]) => string
+}

src/hooks/context-window-monitor.ts

@@ -52,11 +52,10 @@ export function createContextWindowMonitorHook(ctx: PluginInput) {
       const lastAssistant = assistantMessages[assistantMessages.length - 1]
       if (lastAssistant.providerID !== "anthropic") return
 
-      const totalInputTokens = assistantMessages.reduce((sum, m) => {
-        const inputTokens = m.tokens?.input ?? 0
-        const cacheReadTokens = m.tokens?.cache?.read ?? 0
-        return sum + inputTokens + cacheReadTokens
-      }, 0)
+      // Use only the last assistant message's input tokens
+      // This reflects the ACTUAL current context window usage (post-compaction)
+      const lastTokens = lastAssistant.tokens
+      const totalInputTokens = (lastTokens?.input ?? 0) + (lastTokens?.cache?.read ?? 0)
 
       const actualUsagePercentage = totalInputTokens / ANTHROPIC_ACTUAL_LIMIT
 

src/hooks/grep-output-truncator.ts

@@ -98,11 +98,11 @@ export function createGrepOutputTruncatorHook(ctx: PluginInput) {
 
       if (assistantMessages.length === 0) return
 
-      const totalInputTokens = assistantMessages.reduce((sum, m) => {
-        const inputTokens = m.tokens?.input ?? 0
-        const cacheReadTokens = m.tokens?.cache?.read ?? 0
-        return sum + inputTokens + cacheReadTokens
-      }, 0)
+      // Use only the last assistant message's input tokens
+      // This reflects the ACTUAL current context window usage (post-compaction)
+      const lastAssistant = assistantMessages[assistantMessages.length - 1]
+      const lastTokens = lastAssistant.tokens
+      const totalInputTokens = (lastTokens?.input ?? 0) + (lastTokens?.cache?.read ?? 0)
 
       const remainingTokens = ANTHROPIC_ACTUAL_LIMIT - totalInputTokens
 

src/hooks/index.ts

@@ -1,7 +1,7 @@
-export { createTodoContinuationEnforcer } from "./todo-continuation-enforcer";
+export { createTodoContinuationEnforcer, type TodoContinuationEnforcer } from "./todo-continuation-enforcer";
 export { createContextWindowMonitorHook } from "./context-window-monitor";
 export { createSessionNotification } from "./session-notification";
-export { createSessionRecoveryHook } from "./session-recovery";
+export { createSessionRecoveryHook, type SessionRecoveryHook } from "./session-recovery";
 export { createCommentCheckerHooks } from "./comment-checker";
 export { createGrepOutputTruncatorHook } from "./grep-output-truncator";
 export { createDirectoryAgentsInjectorHook } from "./directory-agents-injector";
@@ -10,3 +10,8 @@ export { createEmptyTaskResponseDetectorHook } from "./empty-task-response-detec
 export { createAnthropicAutoCompactHook } from "./anthropic-auto-compact";
 export { createThinkModeHook } from "./think-mode";
 export { createClaudeCodeHooksHook } from "./claude-code-hooks";
+export { createRulesInjectorHook } from "./rules-injector";
+export { createBackgroundNotificationHook } from "./background-notification"
+export { createAutoUpdateCheckerHook } from "./auto-update-checker";
+export { createUltraworkModeHook } from "./ultrawork-mode";
+export { createAgentUsageReminderHook } from "./agent-usage-reminder";

src/hooks/rules-injector/constants.ts

@@ -0,0 +1,23 @@
+import { join } from "node:path";
+import { xdgData } from "xdg-basedir";
+
+export const OPENCODE_STORAGE = join(xdgData ?? "", "opencode", "storage");
+export const RULES_INJECTOR_STORAGE = join(OPENCODE_STORAGE, "rules-injector");
+
+export const PROJECT_MARKERS = [
+  ".git",
+  "pyproject.toml",
+  "package.json",
+  "Cargo.toml",
+  "go.mod",
+  ".venv",
+];
+
+export const PROJECT_RULE_SUBDIRS: [string, string][] = [
+  [".cursor", "rules"],
+  [".claude", "rules"],
+];
+
+export const USER_RULE_DIR = ".claude/rules";
+
+export const RULE_EXTENSIONS = [".md", ".mdc"];

src/hooks/rules-injector/finder.ts

@@ -0,0 +1,237 @@
+import {
+  existsSync,
+  readdirSync,
+  realpathSync,
+  statSync,
+} from "node:fs";
+import { dirname, join, relative } from "node:path";
+import {
+  PROJECT_MARKERS,
+  PROJECT_RULE_SUBDIRS,
+  RULE_EXTENSIONS,
+  USER_RULE_DIR,
+} from "./constants";
+
+/**
+ * Candidate rule file with metadata for filtering and sorting
+ */
+export interface RuleFileCandidate {
+  /** Absolute path to the rule file */
+  path: string;
+  /** Real path after symlink resolution (for duplicate detection) */
+  realPath: string;
+  /** Whether this is a global/user-level rule */
+  isGlobal: boolean;
+  /** Directory distance from current file (9999 for global rules) */
+  distance: number;
+}
+
+/**
+ * Find project root by walking up from startPath.
+ * Checks for PROJECT_MARKERS (.git, pyproject.toml, package.json, etc.)
+ *
+ * @param startPath - Starting path to search from (file or directory)
+ * @returns Project root path or null if not found
+ */
+export function findProjectRoot(startPath: string): string | null {
+  let current: string;
+
+  try {
+    const stat = statSync(startPath);
+    current = stat.isDirectory() ? startPath : dirname(startPath);
+  } catch {
+    current = dirname(startPath);
+  }
+
+  while (true) {
+    for (const marker of PROJECT_MARKERS) {
+      const markerPath = join(current, marker);
+      if (existsSync(markerPath)) {
+        return current;
+      }
+    }
+
+    const parent = dirname(current);
+    if (parent === current) {
+      return null;
+    }
+    current = parent;
+  }
+}
+
+/**
+ * Recursively find all rule files (*.md, *.mdc) in a directory
+ *
+ * @param dir - Directory to search
+ * @param results - Array to accumulate results
+ */
+function findRuleFilesRecursive(dir: string, results: string[]): void {
+  if (!existsSync(dir)) return;
+
+  try {
+    const entries = readdirSync(dir, { withFileTypes: true });
+    for (const entry of entries) {
+      const fullPath = join(dir, entry.name);
+
+      if (entry.isDirectory()) {
+        findRuleFilesRecursive(fullPath, results);
+      } else if (entry.isFile()) {
+        const isRuleFile = RULE_EXTENSIONS.some((ext) =>
+          entry.name.endsWith(ext),
+        );
+        if (isRuleFile) {
+          results.push(fullPath);
+        }
+      }
+    }
+  } catch {
+    // Permission denied or other errors - silently skip
+  }
+}
+
+/**
+ * Resolve symlinks safely with fallback to original path
+ *
+ * @param filePath - Path to resolve
+ * @returns Real path or original path if resolution fails
+ */
+function safeRealpathSync(filePath: string): string {
+  try {
+    return realpathSync(filePath);
+  } catch {
+    return filePath;
+  }
+}
+
+/**
+ * Calculate directory distance between a rule file and current file.
+ * Distance is based on common ancestor within project root.
+ *
+ * @param rulePath - Path to the rule file
+ * @param currentFile - Path to the current file being edited
+ * @param projectRoot - Project root for relative path calculation
+ * @returns Distance (0 = same directory, higher = further)
+ */
+export function calculateDistance(
+  rulePath: string,
+  currentFile: string,
+  projectRoot: string | null,
+): number {
+  if (!projectRoot) {
+    return 9999;
+  }
+
+  try {
+    const ruleDir = dirname(rulePath);
+    const currentDir = dirname(currentFile);
+
+    const ruleRel = relative(projectRoot, ruleDir);
+    const currentRel = relative(projectRoot, currentDir);
+
+    // Handle paths outside project root
+    if (ruleRel.startsWith("..") || currentRel.startsWith("..")) {
+      return 9999;
+    }
+
+    const ruleParts = ruleRel ? ruleRel.split("/") : [];
+    const currentParts = currentRel ? currentRel.split("/") : [];
+
+    // Find common prefix length
+    let common = 0;
+    for (let i = 0; i < Math.min(ruleParts.length, currentParts.length); i++) {
+      if (ruleParts[i] === currentParts[i]) {
+        common++;
+      } else {
+        break;
+      }
+    }
+
+    // Distance is how many directories up from current file to common ancestor
+    return currentParts.length - common;
+  } catch {
+    return 9999;
+  }
+}
+
+/**
+ * Find all rule files for a given context.
+ * Searches from currentFile upward to projectRoot for rule directories,
+ * then user-level directory (~/.claude/rules).
+ *
+ * IMPORTANT: This searches EVERY directory from file to project root.
+ * Not just the project root itself.
+ *
+ * @param projectRoot - Project root path (or null if outside any project)
+ * @param homeDir - User home directory
+ * @param currentFile - Current file being edited (for distance calculation)
+ * @returns Array of rule file candidates sorted by distance
+ */
+export function findRuleFiles(
+  projectRoot: string | null,
+  homeDir: string,
+  currentFile: string,
+): RuleFileCandidate[] {
+  const candidates: RuleFileCandidate[] = [];
+  const seenRealPaths = new Set<string>();
+
+  // Search from current file's directory up to project root
+  let currentDir = dirname(currentFile);
+  let distance = 0;
+
+  while (true) {
+    // Search rule directories in current directory
+    for (const [parent, subdir] of PROJECT_RULE_SUBDIRS) {
+      const ruleDir = join(currentDir, parent, subdir);
+      const files: string[] = [];
+      findRuleFilesRecursive(ruleDir, files);
+
+      for (const filePath of files) {
+        const realPath = safeRealpathSync(filePath);
+        if (seenRealPaths.has(realPath)) continue;
+        seenRealPaths.add(realPath);
+
+        candidates.push({
+          path: filePath,
+          realPath,
+          isGlobal: false,
+          distance,
+        });
+      }
+    }
+
+    // Stop at project root or filesystem root
+    if (projectRoot && currentDir === projectRoot) break;
+    const parentDir = dirname(currentDir);
+    if (parentDir === currentDir) break;
+    currentDir = parentDir;
+    distance++;
+  }
+
+  // Search user-level rule directory (~/.claude/rules)
+  const userRuleDir = join(homeDir, USER_RULE_DIR);
+  const userFiles: string[] = [];
+  findRuleFilesRecursive(userRuleDir, userFiles);
+
+  for (const filePath of userFiles) {
+    const realPath = safeRealpathSync(filePath);
+    if (seenRealPaths.has(realPath)) continue;
+    seenRealPaths.add(realPath);
+
+    candidates.push({
+      path: filePath,
+      realPath,
+      isGlobal: true,
+      distance: 9999, // Global rules always have max distance
+    });
+  }
+
+  // Sort by distance (closest first, then global rules last)
+  candidates.sort((a, b) => {
+    if (a.isGlobal !== b.isGlobal) {
+      return a.isGlobal ? 1 : -1;
+    }
+    return a.distance - b.distance;
+  });
+
+  return candidates;
+}

src/hooks/rules-injector/index.ts

@@ -0,0 +1,150 @@
+import type { PluginInput } from "@opencode-ai/plugin";
+import { readFileSync } from "node:fs";
+import { homedir } from "node:os";
+import { relative, resolve } from "node:path";
+import { findProjectRoot, findRuleFiles } from "./finder";
+import {
+  createContentHash,
+  isDuplicateByContentHash,
+  isDuplicateByRealPath,
+  shouldApplyRule,
+} from "./matcher";
+import { parseRuleFrontmatter } from "./parser";
+import {
+  clearInjectedRules,
+  loadInjectedRules,
+  saveInjectedRules,
+} from "./storage";
+
+interface ToolExecuteInput {
+  tool: string;
+  sessionID: string;
+  callID: string;
+}
+
+interface ToolExecuteOutput {
+  title: string;
+  output: string;
+  metadata: unknown;
+}
+
+interface EventInput {
+  event: {
+    type: string;
+    properties?: unknown;
+  };
+}
+
+interface RuleToInject {
+  relativePath: string;
+  matchReason: string;
+  content: string;
+  distance: number;
+}
+
+const TRACKED_TOOLS = ["read", "write", "edit", "multiedit"];
+
+export function createRulesInjectorHook(ctx: PluginInput) {
+  const sessionCaches = new Map<
+    string,
+    { contentHashes: Set<string>; realPaths: Set<string> }
+  >();
+
+  function getSessionCache(sessionID: string): {
+    contentHashes: Set<string>;
+    realPaths: Set<string>;
+  } {
+    if (!sessionCaches.has(sessionID)) {
+      sessionCaches.set(sessionID, loadInjectedRules(sessionID));
+    }
+    return sessionCaches.get(sessionID)!;
+  }
+
+  function resolveFilePath(title: string): string | null {
+    if (!title) return null;
+    if (title.startsWith("/")) return title;
+    return resolve(ctx.directory, title);
+  }
+
+  const toolExecuteAfter = async (
+    input: ToolExecuteInput,
+    output: ToolExecuteOutput
+  ) => {
+    if (!TRACKED_TOOLS.includes(input.tool.toLowerCase())) return;
+
+    const filePath = resolveFilePath(output.title);
+    if (!filePath) return;
+
+    const projectRoot = findProjectRoot(filePath);
+    const cache = getSessionCache(input.sessionID);
+    const home = homedir();
+
+    const ruleFileCandidates = findRuleFiles(projectRoot, home, filePath);
+    const toInject: RuleToInject[] = [];
+
+    for (const candidate of ruleFileCandidates) {
+      if (isDuplicateByRealPath(candidate.realPath, cache.realPaths)) continue;
+
+      try {
+        const rawContent = readFileSync(candidate.path, "utf-8");
+        const { metadata, body } = parseRuleFrontmatter(rawContent);
+
+        const matchResult = shouldApplyRule(metadata, filePath, projectRoot);
+        if (!matchResult.applies) continue;
+
+        const contentHash = createContentHash(body);
+        if (isDuplicateByContentHash(contentHash, cache.contentHashes)) continue;
+
+        const relativePath = projectRoot
+          ? relative(projectRoot, candidate.path)
+          : candidate.path;
+
+        toInject.push({
+          relativePath,
+          matchReason: matchResult.reason ?? "matched",
+          content: body,
+          distance: candidate.distance,
+        });
+
+        cache.realPaths.add(candidate.realPath);
+        cache.contentHashes.add(contentHash);
+      } catch {}
+    }
+
+    if (toInject.length === 0) return;
+
+    toInject.sort((a, b) => a.distance - b.distance);
+
+    for (const rule of toInject) {
+      output.output += `\n\n[Rule: ${rule.relativePath}]\n[Match: ${rule.matchReason}]\n${rule.content}`;
+    }
+
+    saveInjectedRules(input.sessionID, cache);
+  };
+
+  const eventHandler = async ({ event }: EventInput) => {
+    const props = event.properties as Record<string, unknown> | undefined;
+
+    if (event.type === "session.deleted") {
+      const sessionInfo = props?.info as { id?: string } | undefined;
+      if (sessionInfo?.id) {
+        sessionCaches.delete(sessionInfo.id);
+        clearInjectedRules(sessionInfo.id);
+      }
+    }
+
+    if (event.type === "session.compacted") {
+      const sessionID = (props?.sessionID ??
+        (props?.info as { id?: string } | undefined)?.id) as string | undefined;
+      if (sessionID) {
+        sessionCaches.delete(sessionID);
+        clearInjectedRules(sessionID);
+      }
+    }
+  };
+
+  return {
+    "tool.execute.after": toolExecuteAfter,
+    event: eventHandler,
+  };
+}

src/hooks/rules-injector/matcher.ts

@@ -0,0 +1,63 @@
+import { createHash } from "crypto"
+import { relative } from "node:path"
+import picomatch from "picomatch"
+import type { RuleMetadata } from "./types"
+
+export interface MatchResult {
+  applies: boolean
+  reason?: string
+}
+
+/**
+ * Check if a rule should apply to the current file based on metadata
+ */
+export function shouldApplyRule(
+  metadata: RuleMetadata,
+  currentFilePath: string,
+  projectRoot: string | null
+): MatchResult {
+  if (metadata.alwaysApply === true) {
+    return { applies: true, reason: "alwaysApply" }
+  }
+
+  const globs = metadata.globs
+  if (!globs) {
+    return { applies: false }
+  }
+
+  const patterns = Array.isArray(globs) ? globs : [globs]
+  if (patterns.length === 0) {
+    return { applies: false }
+  }
+
+  const relativePath = projectRoot ? relative(projectRoot, currentFilePath) : currentFilePath
+
+  for (const pattern of patterns) {
+    if (picomatch.isMatch(relativePath, pattern, { dot: true, bash: true })) {
+      return { applies: true, reason: `glob: ${pattern}` }
+    }
+  }
+
+  return { applies: false }
+}
+
+/**
+ * Check if realPath already exists in cache (symlink deduplication)
+ */
+export function isDuplicateByRealPath(realPath: string, cache: Set<string>): boolean {
+  return cache.has(realPath)
+}
+
+/**
+ * Create SHA-256 hash of content, truncated to 16 chars
+ */
+export function createContentHash(content: string): string {
+  return createHash("sha256").update(content).digest("hex").slice(0, 16)
+}
+
+/**
+ * Check if content hash already exists in cache
+ */
+export function isDuplicateByContentHash(hash: string, cache: Set<string>): boolean {
+  return cache.has(hash)
+}

src/hooks/rules-injector/parser.ts

@@ -0,0 +1,211 @@
+import type { RuleMetadata } from "./types";
+
+export interface RuleFrontmatterResult {
+  metadata: RuleMetadata;
+  body: string;
+}
+
+/**
+ * Parse YAML frontmatter from rule file content
+ * Supports:
+ * - Single string: globs: "**\/*.py"
+ * - Inline array: globs: ["**\/*.py", "src/**\/*.ts"]
+ * - Multi-line array:
+ *   globs:
+ *     - "**\/*.py"
+ *     - "src/**\/*.ts"
+ * - Comma-separated: globs: "**\/*.py, src/**\/*.ts"
+ * - Claude Code 'paths' field (alias for globs)
+ */
+export function parseRuleFrontmatter(content: string): RuleFrontmatterResult {
+  const frontmatterRegex = /^---\r?\n([\s\S]*?)\r?\n---\r?\n?([\s\S]*)$/;
+  const match = content.match(frontmatterRegex);
+
+  if (!match) {
+    return { metadata: {}, body: content };
+  }
+
+  const yamlContent = match[1];
+  const body = match[2];
+
+  try {
+    const metadata = parseYamlContent(yamlContent);
+    return { metadata, body };
+  } catch {
+    return { metadata: {}, body: content };
+  }
+}
+
+/**
+ * Parse YAML content without external library
+ */
+function parseYamlContent(yamlContent: string): RuleMetadata {
+  const lines = yamlContent.split("\n");
+  const metadata: RuleMetadata = {};
+
+  let i = 0;
+  while (i < lines.length) {
+    const line = lines[i];
+    const colonIndex = line.indexOf(":");
+
+    if (colonIndex === -1) {
+      i++;
+      continue;
+    }
+
+    const key = line.slice(0, colonIndex).trim();
+    const rawValue = line.slice(colonIndex + 1).trim();
+
+    if (key === "description") {
+      metadata.description = parseStringValue(rawValue);
+    } else if (key === "alwaysApply") {
+      metadata.alwaysApply = rawValue === "true";
+    } else if (key === "globs" || key === "paths") {
+      const { value, consumed } = parseArrayOrStringValue(rawValue, lines, i);
+      // Merge paths into globs (Claude Code compatibility)
+      if (key === "paths") {
+        metadata.globs = mergeGlobs(metadata.globs, value);
+      } else {
+        metadata.globs = mergeGlobs(metadata.globs, value);
+      }
+      i += consumed;
+      continue;
+    }
+
+    i++;
+  }
+
+  return metadata;
+}
+
+/**
+ * Parse a string value, removing surrounding quotes
+ */
+function parseStringValue(value: string): string {
+  if (!value) return "";
+
+  // Remove surrounding quotes
+  if (
+    (value.startsWith('"') && value.endsWith('"')) ||
+    (value.startsWith("'") && value.endsWith("'"))
+  ) {
+    return value.slice(1, -1);
+  }
+
+  return value;
+}
+
+/**
+ * Parse array or string value from YAML
+ * Returns the parsed value and number of lines consumed
+ */
+function parseArrayOrStringValue(
+  rawValue: string,
+  lines: string[],
+  currentIndex: number
+): { value: string | string[]; consumed: number } {
+  // Case 1: Inline array ["a", "b", "c"]
+  if (rawValue.startsWith("[")) {
+    return { value: parseInlineArray(rawValue), consumed: 1 };
+  }
+
+  // Case 2: Multi-line array (value is empty, next lines start with "  - ")
+  if (!rawValue || rawValue === "") {
+    const arrayItems: string[] = [];
+    let consumed = 1;
+
+    for (let j = currentIndex + 1; j < lines.length; j++) {
+      const nextLine = lines[j];
+
+      // Check if this is an array item (starts with whitespace + dash)
+      const arrayMatch = nextLine.match(/^\s+-\s*(.*)$/);
+      if (arrayMatch) {
+        const itemValue = parseStringValue(arrayMatch[1].trim());
+        if (itemValue) {
+          arrayItems.push(itemValue);
+        }
+        consumed++;
+      } else if (nextLine.trim() === "") {
+        // Skip empty lines within array
+        consumed++;
+      } else {
+        // Not an array item, stop
+        break;
+      }
+    }
+
+    if (arrayItems.length > 0) {
+      return { value: arrayItems, consumed };
+    }
+  }
+
+  // Case 3: Comma-separated patterns in single string
+  const stringValue = parseStringValue(rawValue);
+  if (stringValue.includes(",")) {
+    const items = stringValue
+      .split(",")
+      .map((s) => s.trim())
+      .filter((s) => s.length > 0);
+    return { value: items, consumed: 1 };
+  }
+
+  // Case 4: Single string value
+  return { value: stringValue, consumed: 1 };
+}
+
+/**
+ * Parse inline JSON-like array: ["a", "b", "c"]
+ */
+function parseInlineArray(value: string): string[] {
+  // Remove brackets
+  const content = value.slice(1, value.lastIndexOf("]")).trim();
+  if (!content) return [];
+
+  const items: string[] = [];
+  let current = "";
+  let inQuote = false;
+  let quoteChar = "";
+
+  for (let i = 0; i < content.length; i++) {
+    const char = content[i];
+
+    if (!inQuote && (char === '"' || char === "'")) {
+      inQuote = true;
+      quoteChar = char;
+    } else if (inQuote && char === quoteChar) {
+      inQuote = false;
+      quoteChar = "";
+    } else if (!inQuote && char === ",") {
+      const trimmed = current.trim();
+      if (trimmed) {
+        items.push(parseStringValue(trimmed));
+      }
+      current = "";
+    } else {
+      current += char;
+    }
+  }
+
+  // Don't forget the last item
+  const trimmed = current.trim();
+  if (trimmed) {
+    items.push(parseStringValue(trimmed));
+  }
+
+  return items;
+}
+
+/**
+ * Merge two globs values (for combining paths and globs)
+ */
+function mergeGlobs(
+  existing: string | string[] | undefined,
+  newValue: string | string[]
+): string | string[] {
+  if (!existing) return newValue;
+
+  const existingArray = Array.isArray(existing) ? existing : [existing];
+  const newArray = Array.isArray(newValue) ? newValue : [newValue];
+
+  return [...existingArray, ...newArray];
+}

src/hooks/rules-injector/storage.ts

@@ -0,0 +1,59 @@
+import {
+  existsSync,
+  mkdirSync,
+  readFileSync,
+  writeFileSync,
+  unlinkSync,
+} from "node:fs";
+import { join } from "node:path";
+import { RULES_INJECTOR_STORAGE } from "./constants";
+import type { InjectedRulesData } from "./types";
+
+function getStoragePath(sessionID: string): string {
+  return join(RULES_INJECTOR_STORAGE, `${sessionID}.json`);
+}
+
+export function loadInjectedRules(sessionID: string): {
+  contentHashes: Set<string>;
+  realPaths: Set<string>;
+} {
+  const filePath = getStoragePath(sessionID);
+  if (!existsSync(filePath))
+    return { contentHashes: new Set(), realPaths: new Set() };
+
+  try {
+    const content = readFileSync(filePath, "utf-8");
+    const data: InjectedRulesData = JSON.parse(content);
+    return {
+      contentHashes: new Set(data.injectedHashes),
+      realPaths: new Set(data.injectedRealPaths ?? []),
+    };
+  } catch {
+    return { contentHashes: new Set(), realPaths: new Set() };
+  }
+}
+
+export function saveInjectedRules(
+  sessionID: string,
+  data: { contentHashes: Set<string>; realPaths: Set<string> }
+): void {
+  if (!existsSync(RULES_INJECTOR_STORAGE)) {
+    mkdirSync(RULES_INJECTOR_STORAGE, { recursive: true });
+  }
+
+  const storageData: InjectedRulesData = {
+    sessionID,
+    injectedHashes: [...data.contentHashes],
+    injectedRealPaths: [...data.realPaths],
+    updatedAt: Date.now(),
+  };
+
+  writeFileSync(getStoragePath(sessionID), JSON.stringify(storageData, null, 2));
+}
+
+export function clearInjectedRules(sessionID: string): void {
+  const filePath = getStoragePath(sessionID);
+  if (existsSync(filePath)) {
+    unlinkSync(filePath);
+  }
+}

src/hooks/rules-injector/types.ts

@@ -0,0 +1,43 @@
+/**
+ * Rule file metadata (Claude Code style frontmatter)
+ * @see https://docs.anthropic.com/en/docs/claude-code/settings#rule-files
+ */
+export interface RuleMetadata {
+  description?: string;
+  globs?: string | string[];
+  alwaysApply?: boolean;
+}
+
+/**
+ * Rule information with path context and content
+ */
+export interface RuleInfo {
+  /** Absolute path to the rule file */
+  path: string;
+  /** Path relative to project root */
+  relativePath: string;
+  /** Directory distance from target file (0 = same dir) */
+  distance: number;
+  /** Rule file content (without frontmatter) */
+  content: string;
+  /** SHA-256 hash of content for deduplication */
+  contentHash: string;
+  /** Parsed frontmatter metadata */
+  metadata: RuleMetadata;
+  /** Why this rule matched (e.g., "alwaysApply", "glob: *.ts", "path match") */
+  matchReason: string;
+  /** Real path after symlink resolution (for duplicate detection) */
+  realPath: string;
+}
+
+/**
+ * Session storage for injected rules tracking
+ */
+export interface InjectedRulesData {
+  sessionID: string;
+  /** Content hashes of already injected rules */
+  injectedHashes: string[];
+  /** Real paths of already injected rules (for symlink deduplication) */
+  injectedRealPaths: string[];
+  updatedAt: number;
+}

src/hooks/session-notification.ts

@@ -17,6 +17,8 @@ interface SessionNotificationConfig {
   idleConfirmationDelay?: number
   /** Skip notification if there are incomplete todos (default: true) */
   skipIfIncompleteTodos?: boolean
+  /** Maximum number of sessions to track before cleanup (default: 100) */
+  maxTrackedSessions?: number
 }
 
 type Platform = "darwin" | "linux" | "win32" | "unsupported"
@@ -103,12 +105,31 @@ export function createSessionNotification(
     soundPath: defaultSoundPath,
     idleConfirmationDelay: 1500,
     skipIfIncompleteTodos: true,
+    maxTrackedSessions: 100,
     ...config,
   }
 
   const notifiedSessions = new Set<string>()
   const pendingTimers = new Map<string, ReturnType<typeof setTimeout>>()
   const sessionActivitySinceIdle = new Set<string>()
+  // Track notification execution version to handle race conditions
+  const notificationVersions = new Map<string, number>()
+
+  function cleanupOldSessions() {
+    const maxSessions = mergedConfig.maxTrackedSessions
+    if (notifiedSessions.size > maxSessions) {
+      const sessionsToRemove = Array.from(notifiedSessions).slice(0, notifiedSessions.size - maxSessions)
+      sessionsToRemove.forEach(id => notifiedSessions.delete(id))
+    }
+    if (sessionActivitySinceIdle.size > maxSessions) {
+      const sessionsToRemove = Array.from(sessionActivitySinceIdle).slice(0, sessionActivitySinceIdle.size - maxSessions)
+      sessionsToRemove.forEach(id => sessionActivitySinceIdle.delete(id))
+    }
+    if (notificationVersions.size > maxSessions) {
+      const sessionsToRemove = Array.from(notificationVersions.keys()).slice(0, notificationVersions.size - maxSessions)
+      sessionsToRemove.forEach(id => notificationVersions.delete(id))
+    }
+  }
 
   function cancelPendingNotification(sessionID: string) {
     const timer = pendingTimers.get(sessionID)
@@ -117,6 +138,8 @@ export function createSessionNotification(
       pendingTimers.delete(sessionID)
     }
     sessionActivitySinceIdle.add(sessionID)
+    // Increment version to invalidate any in-flight notifications
+    notificationVersions.set(sessionID, (notificationVersions.get(sessionID) ?? 0) + 1)
   }
 
   function markSessionActivity(sessionID: string) {
@@ -124,9 +147,14 @@ export function createSessionNotification(
     notifiedSessions.delete(sessionID)
   }
 
-  async function executeNotification(sessionID: string) {
+  async function executeNotification(sessionID: string, version: number) {
     pendingTimers.delete(sessionID)
 
+    // Race condition fix: check if version matches (activity happened during async wait)
+    if (notificationVersions.get(sessionID) !== version) {
+      return
+    }
+
     if (sessionActivitySinceIdle.has(sessionID)) {
       sessionActivitySinceIdle.delete(sessionID)
       return
@@ -136,9 +164,17 @@ export function createSessionNotification(
 
     if (mergedConfig.skipIfIncompleteTodos) {
       const hasPendingWork = await hasIncompleteTodos(ctx, sessionID)
+      // Re-check version after async call (race condition fix)
+      if (notificationVersions.get(sessionID) !== version) {
+        return
+      }
       if (hasPendingWork) return
     }
 
+    if (notificationVersions.get(sessionID) !== version) {
+      return
+    }
+
     notifiedSessions.add(sessionID)
 
     try {
@@ -172,20 +208,34 @@ export function createSessionNotification(
       if (pendingTimers.has(sessionID)) return
 
       sessionActivitySinceIdle.delete(sessionID)
+      
+      const currentVersion = (notificationVersions.get(sessionID) ?? 0) + 1
+      notificationVersions.set(sessionID, currentVersion)
 
       const timer = setTimeout(() => {
-        executeNotification(sessionID)
+        executeNotification(sessionID, currentVersion)
       }, mergedConfig.idleConfirmationDelay)
 
       pendingTimers.set(sessionID, timer)
+      cleanupOldSessions()
+      return
     }
 
-    if (event.type === "message.updated") {
+    if (event.type === "message.updated" || event.type === "message.created") {
       const info = props?.info as Record<string, unknown> | undefined
       const sessionID = info?.sessionID as string | undefined
       if (sessionID) {
         markSessionActivity(sessionID)
       }
+      return
+    }
+
+    if (event.type === "tool.execute.before" || event.type === "tool.execute.after") {
+      const sessionID = props?.sessionID as string | undefined
+      if (sessionID) {
+        markSessionActivity(sessionID)
+      }
+      return
     }
 
     if (event.type === "session.deleted") {
@@ -194,6 +244,7 @@ export function createSessionNotification(
         cancelPendingNotification(sessionInfo.id)
         notifiedSessions.delete(sessionInfo.id)
         sessionActivitySinceIdle.delete(sessionInfo.id)
+        notificationVersions.delete(sessionInfo.id)
       }
     }
   }

src/hooks/session-recovery/index.ts

@@ -3,6 +3,7 @@ import type { createOpencodeClient } from "@opencode-ai/sdk"
 import {
   findEmptyMessages,
   findEmptyMessageByIndex,
+  findMessageByIndexNeedingThinking,
   findMessagesWithOrphanThinking,
   findMessagesWithThinkingBlocks,
   injectTextPart,
@@ -70,7 +71,10 @@ function detectErrorType(error: unknown): RecoveryErrorType {
 
   if (
     message.includes("thinking") &&
-    (message.includes("first block") || message.includes("must start with") || message.includes("preceeding"))
+    (message.includes("first block") ||
+      message.includes("must start with") ||
+      message.includes("preceeding") ||
+      (message.includes("expected") && message.includes("found")))
   ) {
     return "thinking_block_order"
   }
@@ -125,8 +129,17 @@ async function recoverThinkingBlockOrder(
   _client: Client,
   sessionID: string,
   _failedAssistantMsg: MessageData,
-  _directory: string
+  _directory: string,
+  error: unknown
 ): Promise<boolean> {
+  const targetIndex = extractMessageIndex(error)
+  if (targetIndex !== null) {
+    const targetMessageID = findMessageByIndexNeedingThinking(sessionID, targetIndex)
+    if (targetMessageID) {
+      return prependThinkingPart(sessionID, targetMessageID)
+    }
+  }
+
   const orphanMessages = findMessagesWithOrphanThinking(sessionID)
 
   if (orphanMessages.length === 0) {
@@ -203,14 +216,26 @@ async function recoverEmptyContentMessage(
 // All error types have dedicated recovery functions (recoverToolResultMissing,
 // recoverThinkingBlockOrder, recoverThinkingDisabledViolation, recoverEmptyContentMessage).
 
-export function createSessionRecoveryHook(ctx: PluginInput) {
+export interface SessionRecoveryHook {
+  handleSessionRecovery: (info: MessageInfo) => Promise<boolean>
+  isRecoverableError: (error: unknown) => boolean
+  setOnAbortCallback: (callback: (sessionID: string) => void) => void
+  setOnRecoveryCompleteCallback: (callback: (sessionID: string) => void) => void
+}
+
+export function createSessionRecoveryHook(ctx: PluginInput): SessionRecoveryHook {
   const processingErrors = new Set<string>()
   let onAbortCallback: ((sessionID: string) => void) | null = null
+  let onRecoveryCompleteCallback: ((sessionID: string) => void) | null = null
 
   const setOnAbortCallback = (callback: (sessionID: string) => void): void => {
     onAbortCallback = callback
   }
 
+  const setOnRecoveryCompleteCallback = (callback: (sessionID: string) => void): void => {
+    onRecoveryCompleteCallback = callback
+  }
+
   const isRecoverableError = (error: unknown): boolean => {
     return detectErrorType(error) !== null
   }
@@ -229,12 +254,12 @@ export function createSessionRecoveryHook(ctx: PluginInput) {
     processingErrors.add(assistantMsgID)
 
     try {
-      await ctx.client.session.abort({ path: { id: sessionID } }).catch(() => {})
-
       if (onAbortCallback) {
-        onAbortCallback(sessionID)
+        onAbortCallback(sessionID)  // Mark recovering BEFORE abort
       }
 
+      await ctx.client.session.abort({ path: { id: sessionID } }).catch(() => {})
+
       const messagesResp = await ctx.client.session.messages({
         path: { id: sessionID },
         query: { directory: ctx.directory },
@@ -275,7 +300,7 @@ export function createSessionRecoveryHook(ctx: PluginInput) {
       if (errorType === "tool_result_missing") {
         success = await recoverToolResultMissing(ctx.client, sessionID, failedMsg)
       } else if (errorType === "thinking_block_order") {
-        success = await recoverThinkingBlockOrder(ctx.client, sessionID, failedMsg, ctx.directory)
+        success = await recoverThinkingBlockOrder(ctx.client, sessionID, failedMsg, ctx.directory, info.error)
       } else if (errorType === "thinking_disabled_violation") {
         success = await recoverThinkingDisabledViolation(ctx.client, sessionID, failedMsg)
       } else if (errorType === "empty_content_message") {
@@ -288,6 +313,11 @@ export function createSessionRecoveryHook(ctx: PluginInput) {
     return false
   } finally {
     processingErrors.delete(assistantMsgID)
+
+    // Always notify recovery complete, regardless of success or failure
+    if (sessionID && onRecoveryCompleteCallback) {
+      onRecoveryCompleteCallback(sessionID)
+    }
   }
   }
 
@@ -295,5 +325,6 @@ export function createSessionRecoveryHook(ctx: PluginInput) {
     handleSessionRecovery,
     isRecoverableError,
     setOnAbortCallback,
+    setOnRecoveryCompleteCallback,
   }
 }

src/hooks/session-recovery/storage.ts

@@ -123,8 +123,6 @@ export function findEmptyMessages(sessionID: string): string[] {
   const emptyIds: string[] = []
 
   for (const msg of messages) {
-    if (msg.role !== "assistant") continue
-
     if (!messageHasContent(msg.id)) {
       emptyIds.push(msg.id)
     }
@@ -136,13 +134,25 @@ export function findEmptyMessages(sessionID: string): string[] {
 export function findEmptyMessageByIndex(sessionID: string, targetIndex: number): string | null {
   const messages = readMessages(sessionID)
   
-  if (targetIndex < 0 || targetIndex >= messages.length) return null
+  // Try multiple indices to handle system message offset
+  // API includes system message at index 0, storage may not
+  const indicesToTry = [targetIndex, targetIndex - 1]
+  
+  for (const idx of indicesToTry) {
+    if (idx < 0 || idx >= messages.length) continue
 
-  const targetMsg = messages[targetIndex]
-  if (targetMsg.role !== "assistant") return null
-  if (messageHasContent(targetMsg.id)) return null
+    const targetMsg = messages[idx]
+    
+    // NOTE: Do NOT skip last assistant message here
+    // If API returned an error, this message is NOT the final assistant message
+    // (the API only allows empty content for the ACTUAL final assistant message)
+    
+    if (!messageHasContent(targetMsg.id)) {
+      return targetMsg.id
+    }
+  }
 
-  return targetMsg.id
+  return null
 }
 
 export function findFirstEmptyMessage(sessionID: string): string | null {
@@ -154,13 +164,9 @@ export function findMessagesWithThinkingBlocks(sessionID: string): string[] {
   const messages = readMessages(sessionID)
   const result: string[] = []
 
-  for (let i = 0; i < messages.length; i++) {
-    const msg = messages[i]
+  for (const msg of messages) {
     if (msg.role !== "assistant") continue
 
-    const isLastMessage = i === messages.length - 1
-    if (isLastMessage) continue
-
     const parts = readParts(msg.id)
     const hasThinking = parts.some((p) => THINKING_TYPES.has(p.type))
     if (hasThinking) {
@@ -179,8 +185,8 @@ export function findMessagesWithOrphanThinking(sessionID: string): string[] {
     const msg = messages[i]
     if (msg.role !== "assistant") continue
 
-    const isLastMessage = i === messages.length - 1
-    if (isLastMessage) continue
+    // NOTE: Removed isLastMessage skip - recovery needs to fix last message too
+    // when "thinking must start with" errors occur on final assistant message
 
     const parts = readParts(msg.id)
     if (parts.length === 0) continue
@@ -188,10 +194,11 @@ export function findMessagesWithOrphanThinking(sessionID: string): string[] {
     const sortedParts = [...parts].sort((a, b) => a.id.localeCompare(b.id))
     const firstPart = sortedParts[0]
 
-    const hasThinking = parts.some((p) => THINKING_TYPES.has(p.type))
     const firstIsThinking = THINKING_TYPES.has(firstPart.type)
 
-    if (hasThinking && !firstIsThinking) {
+    // NOTE: Changed condition - if first part is not thinking, it's orphan
+    // regardless of whether thinking blocks exist elsewhere in the message
+    if (!firstIsThinking) {
       result.push(msg.id)
     }
   }
@@ -246,3 +253,25 @@ export function stripThinkingParts(messageID: string): boolean {
 
   return anyRemoved
 }
+
+export function findMessageByIndexNeedingThinking(sessionID: string, targetIndex: number): string | null {
+  const messages = readMessages(sessionID)
+
+  if (targetIndex < 0 || targetIndex >= messages.length) return null
+
+  const targetMsg = messages[targetIndex]
+  if (targetMsg.role !== "assistant") return null
+
+  const parts = readParts(targetMsg.id)
+  if (parts.length === 0) return null
+
+  const sortedParts = [...parts].sort((a, b) => a.id.localeCompare(b.id))
+  const firstPart = sortedParts[0]
+  const firstIsThinking = THINKING_TYPES.has(firstPart.type)
+
+  if (!firstIsThinking) {
+    return targetMsg.id
+  }
+
+  return null
+}

src/hooks/think-mode/index.ts

@@ -1,6 +1,7 @@
 import { detectThinkKeyword, extractPromptText } from "./detector"
-import { getHighVariant, isAlreadyHighVariant } from "./switcher"
+import { getHighVariant, isAlreadyHighVariant, getThinkingConfig } from "./switcher"
 import type { ThinkModeState, ThinkModeInput } from "./types"
+import { log } from "../../shared"
 
 export * from "./detector"
 export * from "./switcher"
@@ -23,6 +24,7 @@ export function createThinkModeHook() {
       const state: ThinkModeState = {
         requested: false,
         modelSwitched: false,
+        thinkingConfigInjected: false,
       }
 
       if (!detectThinkKeyword(promptText)) {
@@ -47,17 +49,31 @@ export function createThinkModeHook() {
       }
 
       const highVariant = getHighVariant(currentModel.modelID)
+      const thinkingConfig = getThinkingConfig(currentModel.providerID, currentModel.modelID)
 
-      if (!highVariant) {
-        thinkModeState.set(sessionID, state)
-        return
+      if (highVariant) {
+        output.message.model = {
+          providerID: currentModel.providerID,
+          modelID: highVariant,
+        }
+        state.modelSwitched = true
+        log("Think mode: model switched to high variant", {
+          sessionID,
+          from: currentModel.modelID,
+          to: highVariant,
+        })
       }
 
-      output.message.model = {
-        providerID: currentModel.providerID,
-        modelID: highVariant,
+      if (thinkingConfig) {
+        Object.assign(output.message, thinkingConfig)
+        state.thinkingConfigInjected = true
+        log("Think mode: thinking config injected", {
+          sessionID,
+          provider: currentModel.providerID,
+          config: thinkingConfig,
+        })
       }
-      state.modelSwitched = true
+
       thinkModeState.set(sessionID, state)
     },
 

src/hooks/think-mode/switcher.ts

@@ -1,19 +1,52 @@
+// Maps model IDs to their "high reasoning" variant (internal convention)
+// For OpenAI models, this signals that reasoning_effort should be set to "high"
 const HIGH_VARIANT_MAP: Record<string, string> = {
+  // Claude
   "claude-sonnet-4-5": "claude-sonnet-4-5-high",
   "claude-opus-4-5": "claude-opus-4-5-high",
-  "gpt-5.1": "gpt-5.1-high",
-  "gpt-5.1-medium": "gpt-5.1-high",
-  "gpt-5.1-codex": "gpt-5.1-codex-high",
+  // Gemini
   "gemini-3-pro": "gemini-3-pro-high",
   "gemini-3-pro-low": "gemini-3-pro-high",
+  // GPT-5
+  "gpt-5": "gpt-5-high",
+  "gpt-5-mini": "gpt-5-mini-high",
+  "gpt-5-nano": "gpt-5-nano-high",
+  "gpt-5-pro": "gpt-5-pro-high",
+  "gpt-5-chat-latest": "gpt-5-chat-latest-high",
+  // GPT-5.1
+  "gpt-5.1": "gpt-5.1-high",
+  "gpt-5.1-chat-latest": "gpt-5.1-chat-latest-high",
+  "gpt-5.1-codex": "gpt-5.1-codex-high",
+  "gpt-5.1-codex-mini": "gpt-5.1-codex-mini-high",
+  "gpt-5.1-codex-max": "gpt-5.1-codex-max-high",
+  // GPT-5.2
+  "gpt-5.2": "gpt-5.2-high",
+  "gpt-5.2-chat-latest": "gpt-5.2-chat-latest-high",
+  "gpt-5.2-pro": "gpt-5.2-pro-high",
 }
 
 const ALREADY_HIGH: Set<string> = new Set([
+  // Claude
   "claude-sonnet-4-5-high",
   "claude-opus-4-5-high",
-  "gpt-5.1-high",
-  "gpt-5.1-codex-high",
+  // Gemini
   "gemini-3-pro-high",
+  // GPT-5
+  "gpt-5-high",
+  "gpt-5-mini-high",
+  "gpt-5-nano-high",
+  "gpt-5-pro-high",
+  "gpt-5-chat-latest-high",
+  // GPT-5.1
+  "gpt-5.1-high",
+  "gpt-5.1-chat-latest-high",
+  "gpt-5.1-codex-high",
+  "gpt-5.1-codex-mini-high",
+  "gpt-5.1-codex-max-high",
+  // GPT-5.2
+  "gpt-5.2-high",
+  "gpt-5.2-chat-latest-high",
+  "gpt-5.2-pro-high",
 ])
 
 export const THINKING_CONFIGS: Record<string, Record<string, unknown>> = {
@@ -22,12 +55,14 @@ export const THINKING_CONFIGS: Record<string, Record<string, unknown>> = {
       type: "enabled",
       budgetTokens: 64000,
     },
+    maxTokens: 128000,
   },
   "amazon-bedrock": {
     reasoningConfig: {
       type: "enabled",
       budgetTokens: 32000,
     },
+    maxTokens: 64000,
   },
   google: {
     providerOptions: {

src/hooks/think-mode/types.ts

@@ -1,6 +1,7 @@
 export interface ThinkModeState {
   requested: boolean
   modelSwitched: boolean
+  thinkingConfigInjected: boolean
   providerID?: string
   modelID?: string
 }

src/hooks/todo-continuation-enforcer.ts

@@ -1,5 +1,11 @@
 import type { PluginInput } from "@opencode-ai/plugin"
 
+export interface TodoContinuationEnforcer {
+  handler: (input: { event: { type: string; properties?: unknown } }) => Promise<void>
+  markRecovering: (sessionID: string) => void
+  markRecoveryComplete: (sessionID: string) => void
+}
+
 interface Todo {
   content: string
   status: string
@@ -32,13 +38,22 @@ function detectInterrupt(error: unknown): boolean {
   return false
 }
 
-export function createTodoContinuationEnforcer(ctx: PluginInput) {
+export function createTodoContinuationEnforcer(ctx: PluginInput): TodoContinuationEnforcer {
   const remindedSessions = new Set<string>()
   const interruptedSessions = new Set<string>()
   const errorSessions = new Set<string>()
+  const recoveringSessions = new Set<string>()
   const pendingTimers = new Map<string, ReturnType<typeof setTimeout>>()
 
-  return async ({ event }: { event: { type: string; properties?: unknown } }) => {
+  const markRecovering = (sessionID: string): void => {
+    recoveringSessions.add(sessionID)
+  }
+
+  const markRecoveryComplete = (sessionID: string): void => {
+    recoveringSessions.delete(sessionID)
+  }
+
+  const handler = async ({ event }: { event: { type: string; properties?: unknown } }): Promise<void> => {
     const props = event.properties as Record<string, unknown> | undefined
 
     if (event.type === "session.error") {
@@ -73,6 +88,11 @@ export function createTodoContinuationEnforcer(ctx: PluginInput) {
       const timer = setTimeout(async () => {
         pendingTimers.delete(sessionID)
 
+        // Check if session is in recovery mode - if so, skip entirely without clearing state
+        if (recoveringSessions.has(sessionID)) {
+          return
+        }
+
         const shouldBypass = interruptedSessions.has(sessionID) || errorSessions.has(sessionID)
         
         interruptedSessions.delete(sessionID)
@@ -111,7 +131,7 @@ export function createTodoContinuationEnforcer(ctx: PluginInput) {
         remindedSessions.add(sessionID)
 
         // Re-check if abort occurred during the delay/fetch
-        if (interruptedSessions.has(sessionID) || errorSessions.has(sessionID)) {
+        if (interruptedSessions.has(sessionID) || errorSessions.has(sessionID) || recoveringSessions.has(sessionID)) {
           remindedSessions.delete(sessionID)
           return
         }
@@ -158,6 +178,7 @@ export function createTodoContinuationEnforcer(ctx: PluginInput) {
         remindedSessions.delete(sessionInfo.id)
         interruptedSessions.delete(sessionInfo.id)
         errorSessions.delete(sessionInfo.id)
+        recoveringSessions.delete(sessionInfo.id)
         
         // Cancel pending continuation
         const timer = pendingTimers.get(sessionInfo.id)
@@ -168,4 +189,10 @@ export function createTodoContinuationEnforcer(ctx: PluginInput) {
       }
     }
   }
+
+  return {
+    handler,
+    markRecovering,
+    markRecoveryComplete,
+  }
 }

src/hooks/ultrawork-mode/constants.ts

@@ -0,0 +1,48 @@
+/** Keyword patterns - "ultrawork", "ulw" (case-insensitive, word boundary) */
+export const ULTRAWORK_PATTERNS = [/\bultrawork\b/i, /\bulw\b/i]
+
+/** Code block pattern to exclude from keyword detection */
+export const CODE_BLOCK_PATTERN = /```[\s\S]*?```/g
+
+/** Inline code pattern to exclude */
+export const INLINE_CODE_PATTERN = /`[^`]+`/g
+
+/**
+ * ULTRAWORK_CONTEXT - Agent-Agnostic Guidance
+ *
+ * Key principles:
+ * - NO specific agent names (oracle, librarian, etc.)
+ * - Only provide guidance based on agent role/capability
+ * - Emphasize parallel execution, TODO tracking, delegation
+ */
+export const ULTRAWORK_CONTEXT = `<ultrawork-mode>
+[CODE RED] Maximum precision required. Ultrathink before acting.
+
+YOU MUST LEVERAGE ALL AVAILABLE AGENTS TO THEIR FULLEST POTENTIAL.
+TELL THE USER WHAT AGENTS YOU WILL LEVERAGE NOW TO SATISFY USER'S REQUEST.
+
+## AGENT UTILIZATION PRINCIPLES (by capability, not by name)
+- **Codebase Exploration**: Spawn exploration agents using BACKGROUND TASKS for file patterns, internal implementations, project structure
+- **Documentation & References**: Use librarian-type agents via BACKGROUND TASKS for API references, examples, external library docs
+- **Planning & Strategy**: NEVER plan yourself - ALWAYS spawn a dedicated planning agent for work breakdown
+- **High-IQ Reasoning**: Leverage specialized agents for architecture decisions, code review, strategic planning
+- **Frontend/UI Tasks**: Delegate to UI-specialized agents for design and implementation
+
+## EXECUTION RULES
+- **TODO**: Track EVERY step. Mark complete IMMEDIATELY after each.
+- **PARALLEL**: Fire independent agent calls simultaneously via background_task - NEVER wait sequentially.
+- **BACKGROUND FIRST**: Use background_task for exploration/research agents (10+ concurrent if needed).
+- **VERIFY**: Re-read request after completion. Check ALL requirements met before reporting done.
+- **DELEGATE**: Don't do everything yourself - orchestrate specialized agents for their strengths.
+
+## WORKFLOW
+1. Analyze the request and identify required capabilities
+2. Spawn exploration/librarian agents via background_task in PARALLEL (10+ if needed)
+3. Use planning agents to create detailed work breakdown
+4. Execute with continuous verification against original requirements
+
+</ultrawork-mode>
+
+---
+
+`

src/hooks/ultrawork-mode/detector.ts

@@ -0,0 +1,33 @@
+import {
+  ULTRAWORK_PATTERNS,
+  CODE_BLOCK_PATTERN,
+  INLINE_CODE_PATTERN,
+} from "./constants"
+
+/**
+ * Remove code blocks and inline code from text.
+ * Prevents false positives when keywords appear in code.
+ */
+export function removeCodeBlocks(text: string): string {
+  return text.replace(CODE_BLOCK_PATTERN, "").replace(INLINE_CODE_PATTERN, "")
+}
+
+/**
+ * Detect ultrawork keywords in text (excluding code blocks).
+ */
+export function detectUltraworkKeyword(text: string): boolean {
+  const textWithoutCode = removeCodeBlocks(text)
+  return ULTRAWORK_PATTERNS.some((pattern) => pattern.test(textWithoutCode))
+}
+
+/**
+ * Extract text content from message parts.
+ */
+export function extractPromptText(
+  parts: Array<{ type: string; text?: string }>
+): string {
+  return parts
+    .filter((p) => p.type === "text")
+    .map((p) => p.text || "")
+    .join("")
+}

src/hooks/ultrawork-mode/index.ts

@@ -0,0 +1,97 @@
+import { detectUltraworkKeyword, extractPromptText } from "./detector"
+import { ULTRAWORK_CONTEXT } from "./constants"
+import type { UltraworkModeState } from "./types"
+import { log } from "../../shared"
+import { injectHookMessage } from "../../features/hook-message-injector"
+
+export * from "./detector"
+export * from "./constants"
+export * from "./types"
+
+const ultraworkModeState = new Map<string, UltraworkModeState>()
+
+export function clearUltraworkModeState(sessionID: string): void {
+  ultraworkModeState.delete(sessionID)
+}
+
+export function createUltraworkModeHook() {
+  return {
+    /**
+     * chat.message hook - detect ultrawork/ulw keywords, inject context via history
+     *
+     * Execution timing: AFTER claudeCodeHooks["chat.message"]
+     * Behavior:
+     *   1. Extract text from user prompt
+     *   2. Detect ultrawork/ulw keywords (excluding code blocks)
+     *   3. If detected, inject ULTRAWORK_CONTEXT via injectHookMessage (history injection)
+     */
+    "chat.message": async (
+      input: {
+        sessionID: string
+        agent?: string
+        model?: { providerID: string; modelID: string }
+        messageID?: string
+      },
+      output: {
+        message: Record<string, unknown>
+        parts: Array<{ type: string; text?: string; [key: string]: unknown }>
+      }
+    ): Promise<void> => {
+      const state: UltraworkModeState = {
+        detected: false,
+        injected: false,
+      }
+
+      const promptText = extractPromptText(output.parts)
+
+      if (!detectUltraworkKeyword(promptText)) {
+        ultraworkModeState.set(input.sessionID, state)
+        return
+      }
+
+      state.detected = true
+      log("Ultrawork keyword detected", { sessionID: input.sessionID })
+
+      const message = output.message as {
+        agent?: string
+        model?: { modelID?: string; providerID?: string }
+        path?: { cwd?: string; root?: string }
+        tools?: Record<string, boolean>
+      }
+
+      const success = injectHookMessage(input.sessionID, ULTRAWORK_CONTEXT, {
+        agent: message.agent,
+        model: message.model,
+        path: message.path,
+        tools: message.tools,
+      })
+
+      if (success) {
+        state.injected = true
+        log("Ultrawork context injected via history", { sessionID: input.sessionID })
+      } else {
+        log("Ultrawork context injection failed", { sessionID: input.sessionID })
+      }
+
+      ultraworkModeState.set(input.sessionID, state)
+    },
+
+    /**
+     * event hook - cleanup session state on deletion
+     */
+    event: async ({
+      event,
+    }: {
+      event: { type: string; properties?: unknown }
+    }) => {
+      if (event.type === "session.deleted") {
+        const props = event.properties as
+          | { info?: { id?: string } }
+          | undefined
+        if (props?.info?.id) {
+          ultraworkModeState.delete(props.info.id)
+        }
+      }
+    },
+  }
+}

src/hooks/ultrawork-mode/types.ts

@@ -0,0 +1,20 @@
+export interface UltraworkModeState {
+  /** Whether ultrawork keyword was detected */
+  detected: boolean
+  /** Whether context was injected */
+  injected: boolean
+}
+
+export interface ModelRef {
+  providerID: string
+  modelID: string
+}
+
+export interface MessageWithModel {
+  model?: ModelRef
+}
+
+export interface UltraworkModeInput {
+  parts: Array<{ type: string; text?: string }>
+  message: MessageWithModel
+}

src/index.ts

@@ -1,9 +1,10 @@
 import type { Plugin } from "@opencode-ai/plugin";
-import { createBuiltinAgents } from "./agents";
+import { createBuiltinAgents, BUILD_AGENT_PROMPT_EXTENSION } from "./agents";
 import {
   createTodoContinuationEnforcer,
   createContextWindowMonitorHook,
   createSessionRecoveryHook,
+  createSessionNotification,
   createCommentCheckerHooks,
   createGrepOutputTruncatorHook,
   createDirectoryAgentsInjectorHook,
@@ -12,7 +13,13 @@ import {
   createThinkModeHook,
   createClaudeCodeHooksHook,
   createAnthropicAutoCompactHook,
+  createRulesInjectorHook,
+  createBackgroundNotificationHook,
+  createAutoUpdateCheckerHook,
+  createUltraworkModeHook,
+  createAgentUsageReminderHook,
 } from "./hooks";
+import { createGoogleAntigravityAuthPlugin } from "./auth/antigravity";
 import {
   loadUserCommands,
   loadProjectCommands,
@@ -35,14 +42,29 @@ import {
   getCurrentSessionTitle,
 } from "./features/claude-code-session-state";
 import { updateTerminalTitle } from "./features/terminal";
-import { builtinTools } from "./tools";
+import { builtinTools, createCallOmoAgent, createBackgroundTools, createLookAt } from "./tools";
+import { BackgroundManager } from "./features/background-agent";
 import { createBuiltinMcps } from "./mcp";
-import { OhMyOpenCodeConfigSchema, type OhMyOpenCodeConfig } from "./config";
-import { log } from "./shared/logger";
+import { OhMyOpenCodeConfigSchema, type OhMyOpenCodeConfig, type HookName } from "./config";
+import { log, deepMerge } from "./shared";
 import * as fs from "fs";
 import * as path from "path";
 import * as os from "os";
 
+/**
+ * Returns the user-level config directory based on the OS.
+ * - Linux/macOS: XDG_CONFIG_HOME or ~/.config
+ * - Windows: %APPDATA%
+ */
+function getUserConfigDir(): string {
+  if (process.platform === "win32") {
+    return process.env.APPDATA || path.join(os.homedir(), "AppData", "Roaming");
+  }
+
+  // Linux, macOS, and other Unix-like systems: respect XDG_CONFIG_HOME
+  return process.env.XDG_CONFIG_HOME || path.join(os.homedir(), ".config");
+}
+
 function loadConfigFromPath(configPath: string): OhMyOpenCodeConfig | null {
   try {
     if (fs.existsSync(configPath)) {
@@ -64,78 +86,161 @@ function loadConfigFromPath(configPath: string): OhMyOpenCodeConfig | null {
   return null;
 }
 
-function mergeConfigs(base: OhMyOpenCodeConfig, override: OhMyOpenCodeConfig): OhMyOpenCodeConfig {
+function mergeConfigs(
+  base: OhMyOpenCodeConfig,
+  override: OhMyOpenCodeConfig
+): OhMyOpenCodeConfig {
   return {
     ...base,
     ...override,
-    agents: override.agents !== undefined
-      ? { ...(base.agents ?? {}), ...override.agents }
-      : base.agents,
+    agents: deepMerge(base.agents, override.agents),
     disabled_agents: [
-      ...new Set([...(base.disabled_agents ?? []), ...(override.disabled_agents ?? [])])
+      ...new Set([
+        ...(base.disabled_agents ?? []),
+        ...(override.disabled_agents ?? []),
+      ]),
     ],
     disabled_mcps: [
-      ...new Set([...(base.disabled_mcps ?? []), ...(override.disabled_mcps ?? [])])
+      ...new Set([
+        ...(base.disabled_mcps ?? []),
+        ...(override.disabled_mcps ?? []),
+      ]),
     ],
+    disabled_hooks: [
+      ...new Set([
+        ...(base.disabled_hooks ?? []),
+        ...(override.disabled_hooks ?? []),
+      ]),
+    ],
+    claude_code: deepMerge(base.claude_code, override.claude_code),
   };
 }
 
 function loadPluginConfig(directory: string): OhMyOpenCodeConfig {
-  // User-level config paths
-  const userConfigPaths = [
-    path.join(os.homedir(), ".config", "opencode", "oh-my-opencode.json"),
-  ];
+  // User-level config path (OS-specific)
+  const userConfigPath = path.join(
+    getUserConfigDir(),
+    "opencode",
+    "oh-my-opencode.json"
+  );
 
-  // Project-level config paths (higher precedence)
-  const projectConfigPaths = [
-    path.join(directory, ".opencode", "oh-my-opencode.json"),
-  ];
+  // Project-level config path
+  const projectConfigPath = path.join(
+    directory,
+    ".opencode",
+    "oh-my-opencode.json"
+  );
 
-  // Load user config first
-  let config: OhMyOpenCodeConfig = {};
-  for (const configPath of userConfigPaths) {
-    const userConfig = loadConfigFromPath(configPath);
-    if (userConfig) {
-      config = userConfig;
-      break;
-    }
-  }
+  // Load user config first (base)
+  let config: OhMyOpenCodeConfig = loadConfigFromPath(userConfigPath) ?? {};
 
   // Override with project config
-  for (const configPath of projectConfigPaths) {
-    const projectConfig = loadConfigFromPath(configPath);
-    if (projectConfig) {
-      config = mergeConfigs(config, projectConfig);
-      break;
-    }
+  const projectConfig = loadConfigFromPath(projectConfigPath);
+  if (projectConfig) {
+    config = mergeConfigs(config, projectConfig);
   }
 
-  log("Final merged config", { agents: config.agents, disabled_agents: config.disabled_agents, disabled_mcps: config.disabled_mcps });
+  log("Final merged config", {
+    agents: config.agents,
+    disabled_agents: config.disabled_agents,
+    disabled_mcps: config.disabled_mcps,
+    disabled_hooks: config.disabled_hooks,
+    claude_code: config.claude_code,
+  });
   return config;
 }
 
 const OhMyOpenCodePlugin: Plugin = async (ctx) => {
   const pluginConfig = loadPluginConfig(ctx.directory);
+  const disabledHooks = new Set(pluginConfig.disabled_hooks ?? []);
+  const isHookEnabled = (hookName: HookName) => !disabledHooks.has(hookName);
 
-  const todoContinuationEnforcer = createTodoContinuationEnforcer(ctx);
-  const contextWindowMonitor = createContextWindowMonitorHook(ctx);
-  const sessionRecovery = createSessionRecoveryHook(ctx);
-  const commentChecker = createCommentCheckerHooks();
-  const grepOutputTruncator = createGrepOutputTruncatorHook(ctx);
-  const directoryAgentsInjector = createDirectoryAgentsInjectorHook(ctx);
-  const directoryReadmeInjector = createDirectoryReadmeInjectorHook(ctx);
-  const emptyTaskResponseDetector = createEmptyTaskResponseDetectorHook(ctx);
-  const thinkMode = createThinkModeHook();
-  const claudeCodeHooks = createClaudeCodeHooksHook(ctx, {});
-  const anthropicAutoCompact = createAnthropicAutoCompactHook(ctx);
+  const todoContinuationEnforcer = isHookEnabled("todo-continuation-enforcer")
+    ? createTodoContinuationEnforcer(ctx)
+    : null;
+  const contextWindowMonitor = isHookEnabled("context-window-monitor")
+    ? createContextWindowMonitorHook(ctx)
+    : null;
+  const sessionRecovery = isHookEnabled("session-recovery")
+    ? createSessionRecoveryHook(ctx)
+    : null;
+  const sessionNotification = isHookEnabled("session-notification")
+    ? createSessionNotification(ctx)
+    : null;
+
+  // Wire up recovery state tracking between session-recovery and todo-continuation-enforcer
+  // This prevents the continuation enforcer from injecting prompts during active recovery
+  if (sessionRecovery && todoContinuationEnforcer) {
+    sessionRecovery.setOnAbortCallback(todoContinuationEnforcer.markRecovering);
+    sessionRecovery.setOnRecoveryCompleteCallback(todoContinuationEnforcer.markRecoveryComplete);
+  }
+
+  const commentChecker = isHookEnabled("comment-checker")
+    ? createCommentCheckerHooks()
+    : null;
+  const grepOutputTruncator = isHookEnabled("grep-output-truncator")
+    ? createGrepOutputTruncatorHook(ctx)
+    : null;
+  const directoryAgentsInjector = isHookEnabled("directory-agents-injector")
+    ? createDirectoryAgentsInjectorHook(ctx)
+    : null;
+  const directoryReadmeInjector = isHookEnabled("directory-readme-injector")
+    ? createDirectoryReadmeInjectorHook(ctx)
+    : null;
+  const emptyTaskResponseDetector = isHookEnabled("empty-task-response-detector")
+    ? createEmptyTaskResponseDetectorHook(ctx)
+    : null;
+  const thinkMode = isHookEnabled("think-mode")
+    ? createThinkModeHook()
+    : null;
+  const claudeCodeHooks = createClaudeCodeHooksHook(ctx, {
+    disabledHooks: (pluginConfig.claude_code?.hooks ?? true) ? undefined : true,
+  });
+  const anthropicAutoCompact = isHookEnabled("anthropic-auto-compact")
+    ? createAnthropicAutoCompactHook(ctx)
+    : null;
+  const rulesInjector = isHookEnabled("rules-injector")
+    ? createRulesInjectorHook(ctx)
+    : null;
+  const autoUpdateChecker = isHookEnabled("auto-update-checker")
+    ? createAutoUpdateCheckerHook(ctx)
+    : null;
+  const ultraworkMode = isHookEnabled("ultrawork-mode")
+    ? createUltraworkModeHook()
+    : null;
+  const agentUsageReminder = isHookEnabled("agent-usage-reminder")
+    ? createAgentUsageReminderHook(ctx)
+    : null;
 
   updateTerminalTitle({ sessionId: "main" });
 
+  const backgroundManager = new BackgroundManager(ctx);
+
+  const backgroundNotificationHook = isHookEnabled("background-notification")
+    ? createBackgroundNotificationHook(backgroundManager)
+    : null;
+  const backgroundTools = createBackgroundTools(backgroundManager, ctx.client);
+
+  const callOmoAgent = createCallOmoAgent(ctx, backgroundManager);
+  const lookAt = createLookAt(ctx);
+
+  const googleAuthHooks = pluginConfig.google_auth
+    ? await createGoogleAntigravityAuthPlugin(ctx)
+    : null;
+
   return {
-    tool: builtinTools,
+    ...(googleAuthHooks ? { auth: googleAuthHooks.auth } : {}),
+
+    tool: {
+      ...builtinTools,
+      ...backgroundTools,
+      call_omo_agent: callOmoAgent,
+      look_at: lookAt,
+    },
 
     "chat.message": async (input, output) => {
-      await claudeCodeHooks["chat.message"]?.(input, output)
+      await claudeCodeHooks["chat.message"]?.(input, output);
+      await ultraworkMode?.["chat.message"]?.(input, output);
     },
 
     config: async (config) => {
@@ -143,8 +248,9 @@ const OhMyOpenCodePlugin: Plugin = async (ctx) => {
         pluginConfig.disabled_agents,
         pluginConfig.agents,
       );
-      const userAgents = loadUserAgents();
-      const projectAgents = loadProjectAgents();
+
+      const userAgents = (pluginConfig.claude_code?.agents ?? true) ? loadUserAgents() : {};
+      const projectAgents = (pluginConfig.claude_code?.agents ?? true) ? loadProjectAgents() : {};
 
       config.agent = {
         ...builtinAgents,
@@ -152,24 +258,61 @@ const OhMyOpenCodePlugin: Plugin = async (ctx) => {
         ...projectAgents,
         ...config.agent,
       };
+
+      // Inject orchestration prompt to all non-subagent agents
+      // Subagents are delegated TO, so they don't need orchestration guidance
+      for (const [agentName, agentConfig] of Object.entries(config.agent ?? {})) {
+        if (agentConfig && agentConfig.mode !== "subagent") {
+          const existingPrompt = agentConfig.prompt || "";
+          const userOverride = pluginConfig.agents?.[agentName as keyof typeof pluginConfig.agents]?.prompt || "";
+          config.agent[agentName] = {
+            ...agentConfig,
+            prompt: existingPrompt + BUILD_AGENT_PROMPT_EXTENSION + userOverride,
+          };
+        }
+      }
+
       config.tools = {
         ...config.tools,
       };
 
-      const mcpResult = await loadMcpConfigs();
+      if (config.agent.explore) {
+        config.agent.explore.tools = {
+          ...config.agent.explore.tools,
+          call_omo_agent: false,
+        };
+      }
+      if (config.agent.librarian) {
+        config.agent.librarian.tools = {
+          ...config.agent.librarian.tools,
+          call_omo_agent: false,
+        };
+      }
+      if (config.agent["multimodal-looker"]) {
+        config.agent["multimodal-looker"].tools = {
+          ...config.agent["multimodal-looker"].tools,
+          task: false,
+          call_omo_agent: false,
+          look_at: false,
+        };
+      }
+
+      const mcpResult = (pluginConfig.claude_code?.mcp ?? true)
+        ? await loadMcpConfigs()
+        : { servers: {} };
       config.mcp = {
         ...config.mcp,
         ...createBuiltinMcps(pluginConfig.disabled_mcps),
         ...mcpResult.servers,
       };
 
-      const userCommands = loadUserCommands();
+      const userCommands = (pluginConfig.claude_code?.commands ?? true) ? loadUserCommands() : {};
       const opencodeGlobalCommands = loadOpencodeGlobalCommands();
       const systemCommands = config.command ?? {};
-      const projectCommands = loadProjectCommands();
+      const projectCommands = (pluginConfig.claude_code?.commands ?? true) ? loadProjectCommands() : {};
       const opencodeProjectCommands = loadOpencodeProjectCommands();
-      const userSkills = loadUserSkillsAsCommands();
-      const projectSkills = loadProjectSkillsAsCommands();
+      const userSkills = (pluginConfig.claude_code?.skills ?? true) ? loadUserSkillsAsCommands() : {};
+      const projectSkills = (pluginConfig.claude_code?.skills ?? true) ? loadProjectSkillsAsCommands() : {};
 
       config.command = {
         ...userCommands,
@@ -183,13 +326,19 @@ const OhMyOpenCodePlugin: Plugin = async (ctx) => {
     },
 
     event: async (input) => {
+      await autoUpdateChecker?.event(input);
       await claudeCodeHooks.event(input);
-      await todoContinuationEnforcer(input);
-      await contextWindowMonitor.event(input);
-      await directoryAgentsInjector.event(input);
-      await directoryReadmeInjector.event(input);
-      await thinkMode.event(input);
-      await anthropicAutoCompact.event(input);
+      await backgroundNotificationHook?.event(input);
+      await sessionNotification?.(input);
+      await todoContinuationEnforcer?.handler(input);
+      await contextWindowMonitor?.event(input);
+      await directoryAgentsInjector?.event(input);
+      await directoryReadmeInjector?.event(input);
+      await rulesInjector?.event(input);
+      await thinkMode?.event(input);
+      await anthropicAutoCompact?.event(input);
+      await ultraworkMode?.event(input);
+      await agentUsageReminder?.event(input);
 
       const { event } = input;
       const props = event.properties as Record<string, unknown> | undefined;
@@ -241,7 +390,7 @@ const OhMyOpenCodePlugin: Plugin = async (ctx) => {
         const sessionID = props?.sessionID as string | undefined;
         const error = props?.error;
 
-        if (sessionRecovery.isRecoverableError(error)) {
+        if (sessionRecovery?.isRecoverableError(error)) {
           const messageInfo = {
             id: props?.messageID as string | undefined,
             role: "assistant" as const,
@@ -287,7 +436,7 @@ const OhMyOpenCodePlugin: Plugin = async (ctx) => {
 
     "tool.execute.before": async (input, output) => {
       await claudeCodeHooks["tool.execute.before"](input, output);
-      await commentChecker["tool.execute.before"](input, output);
+      await commentChecker?.["tool.execute.before"](input, output);
 
       if (input.sessionID === getMainSessionID()) {
         updateTerminalTitle({
@@ -302,12 +451,14 @@ const OhMyOpenCodePlugin: Plugin = async (ctx) => {
 
     "tool.execute.after": async (input, output) => {
       await claudeCodeHooks["tool.execute.after"](input, output);
-      await grepOutputTruncator["tool.execute.after"](input, output);
-      await contextWindowMonitor["tool.execute.after"](input, output);
-      await commentChecker["tool.execute.after"](input, output);
-      await directoryAgentsInjector["tool.execute.after"](input, output);
-      await directoryReadmeInjector["tool.execute.after"](input, output);
-      await emptyTaskResponseDetector["tool.execute.after"](input, output);
+      await grepOutputTruncator?.["tool.execute.after"](input, output);
+      await contextWindowMonitor?.["tool.execute.after"](input, output);
+      await commentChecker?.["tool.execute.after"](input, output);
+      await directoryAgentsInjector?.["tool.execute.after"](input, output);
+      await directoryReadmeInjector?.["tool.execute.after"](input, output);
+      await rulesInjector?.["tool.execute.after"](input, output);
+      await emptyTaskResponseDetector?.["tool.execute.after"](input, output);
+      await agentUsageReminder?.["tool.execute.after"](input, output);
 
       if (input.sessionID === getMainSessionID()) {
         updateTerminalTitle({
@@ -329,4 +480,5 @@ export type {
   AgentOverrideConfig,
   AgentOverrides,
   McpName,
+  HookName,
 } from "./config";

src/mcp/grep-app.ts

@@ -0,0 +1,5 @@
+export const grep_app = {
+  type: "remote" as const,
+  url: "https://mcp.grep.app",
+  enabled: true,
+}

src/mcp/index.ts

@@ -1,5 +1,6 @@
 import { websearch_exa } from "./websearch-exa"
 import { context7 } from "./context7"
+import { grep_app } from "./grep-app"
 import type { McpName } from "./types"
 
 export { McpNameSchema, type McpName } from "./types"
@@ -7,6 +8,7 @@ export { McpNameSchema, type McpName } from "./types"
 const allBuiltinMcps: Record<McpName, { type: "remote"; url: string; enabled: boolean }> = {
   websearch_exa,
   context7,
+  grep_app,
 }
 
 export function createBuiltinMcps(disabledMcps: McpName[] = []) {

src/mcp/types.ts

@@ -1,5 +1,5 @@
 import { z } from "zod"
 
-export const McpNameSchema = z.enum(["websearch_exa", "context7"])
+export const McpNameSchema = z.enum(["websearch_exa", "context7", "grep_app"])
 
 export type McpName = z.infer<typeof McpNameSchema>

src/shared/deep-merge.ts

@@ -0,0 +1,53 @@
+const DANGEROUS_KEYS = new Set(["__proto__", "constructor", "prototype"]);
+const MAX_DEPTH = 50;
+
+export function isPlainObject(value: unknown): value is Record<string, unknown> {
+  return (
+    typeof value === "object" &&
+    value !== null &&
+    !Array.isArray(value) &&
+    Object.prototype.toString.call(value) === "[object Object]"
+  );
+}
+
+/**
+ * Deep merges two objects, with override values taking precedence.
+ * - Objects are recursively merged
+ * - Arrays are replaced (not concatenated)
+ * - undefined values in override do not overwrite base values
+ *
+ * @example
+ * deepMerge({ a: 1, b: { c: 2, d: 3 } }, { b: { c: 10 }, e: 5 })
+ * // => { a: 1, b: { c: 10, d: 3 }, e: 5 }
+ */
+export function deepMerge<T extends Record<string, unknown>>(base: T, override: Partial<T>, depth?: number): T;
+export function deepMerge<T extends Record<string, unknown>>(base: T | undefined, override: T | undefined, depth?: number): T | undefined;
+export function deepMerge<T extends Record<string, unknown>>(
+  base: T | undefined,
+  override: T | undefined,
+  depth = 0
+): T | undefined {
+  if (!base && !override) return undefined;
+  if (!base) return override;
+  if (!override) return base;
+  if (depth > MAX_DEPTH) return override ?? base;
+
+  const result = { ...base } as Record<string, unknown>;
+
+  for (const key of Object.keys(override)) {
+    if (DANGEROUS_KEYS.has(key)) continue;
+
+    const baseValue = base[key];
+    const overrideValue = override[key];
+
+    if (overrideValue === undefined) continue;
+
+    if (isPlainObject(baseValue) && isPlainObject(overrideValue)) {
+      result[key] = deepMerge(baseValue, overrideValue, depth + 1);
+    } else {
+      result[key] = overrideValue;
+    }
+  }
+
+  return result as T;
+}

src/shared/file-utils.ts

@@ -0,0 +1,26 @@
+import { lstatSync, readlinkSync } from "fs"
+import { resolve } from "path"
+
+export function isMarkdownFile(entry: { name: string; isFile: () => boolean }): boolean {
+  return !entry.name.startsWith(".") && entry.name.endsWith(".md") && entry.isFile()
+}
+
+export function isSymbolicLink(filePath: string): boolean {
+  try {
+    return lstatSync(filePath, { throwIfNoEntry: false })?.isSymbolicLink() ?? false
+  } catch {
+    return false
+  }
+}
+
+export function resolveSymlink(filePath: string): string {
+  try {
+    const stats = lstatSync(filePath, { throwIfNoEntry: false })
+    if (stats?.isSymbolicLink()) {
+      return resolve(filePath, "..", readlinkSync(filePath))
+    }
+    return filePath
+  } catch {
+    return filePath
+  }
+}

src/shared/index.ts

@@ -7,3 +7,5 @@ export * from "./snake-case"
 export * from "./tool-name"
 export * from "./pattern-matcher"
 export * from "./hook-disabled"
+export * from "./deep-merge"
+export * from "./file-utils"

src/shared/snake-case.ts

@@ -1,3 +1,5 @@
+import { isPlainObject } from "./deep-merge"
+
 export function camelToSnake(str: string): string {
   return str.replace(/[A-Z]/g, (letter) => `_${letter.toLowerCase()}`)
 }
@@ -6,10 +8,6 @@ export function snakeToCamel(str: string): string {
   return str.replace(/_([a-z])/g, (_, letter) => letter.toUpperCase())
 }
 
-function isPlainObject(value: unknown): value is Record<string, unknown> {
-  return typeof value === "object" && value !== null && !Array.isArray(value)
-}
-
 export function objectToSnakeCase(
   obj: Record<string, unknown>,
   deep: boolean = true

src/tools/background-task/constants.ts

@@ -0,0 +1,41 @@
+export const BACKGROUND_TASK_DESCRIPTION = `Launch a background agent task that runs asynchronously.
+
+The task runs in a separate session while you continue with other work. The system will notify you when the task completes.
+
+Use this for:
+- Long-running research tasks
+- Complex analysis that doesn't need immediate results
+- Parallel workloads to maximize throughput
+
+Arguments:
+- description: Short task description (shown in status)
+- prompt: Full detailed prompt for the agent
+- agent: Agent type to use (any agent allowed)
+
+Returns immediately with task ID and session info. Use \`background_output\` to check progress or retrieve results.`
+
+export const BACKGROUND_OUTPUT_DESCRIPTION = `Get output from a background task.
+
+Arguments:
+- task_id: Required task ID to get output from
+- block: If true, wait for task completion. If false (default), return current status immediately.
+- timeout: Max wait time in ms when blocking (default: 60000, max: 600000)
+
+Returns:
+- When not blocking: Returns current status with task ID, description, agent, status, duration, and progress info
+- When blocking: Waits for completion, then returns full result
+
+IMPORTANT: The system automatically notifies the main session when background tasks complete.
+You typically don't need block=true - just use block=false to check status, and the system will notify you when done.
+
+Use this to:
+- Check task progress (block=false) - returns full status info, NOT empty
+- Wait for and retrieve task result (block=true) - only when you explicitly need to wait
+- Set custom timeout for long tasks`
+
+export const BACKGROUND_CANCEL_DESCRIPTION = `Cancel a running background task.
+
+Only works for tasks with status "running". Aborts the background session and marks the task as cancelled.
+
+Arguments:
+- taskId: Required task ID to cancel.`

src/tools/background-task/index.ts

@@ -0,0 +1,8 @@
+export {
+  createBackgroundTask,
+  createBackgroundOutput,
+  createBackgroundCancel,
+} from "./tools"
+
+export type * from "./types"
+export * from "./constants"

src/tools/background-task/tools.ts

@@ -0,0 +1,300 @@
+import { tool, type PluginInput } from "@opencode-ai/plugin"
+import type { BackgroundManager, BackgroundTask } from "../../features/background-agent"
+import type { BackgroundTaskArgs, BackgroundOutputArgs, BackgroundCancelArgs } from "./types"
+import { BACKGROUND_TASK_DESCRIPTION, BACKGROUND_OUTPUT_DESCRIPTION, BACKGROUND_CANCEL_DESCRIPTION } from "./constants"
+
+type OpencodeClient = PluginInput["client"]
+
+function formatDuration(start: Date, end?: Date): string {
+  const duration = (end ?? new Date()).getTime() - start.getTime()
+  const seconds = Math.floor(duration / 1000)
+  const minutes = Math.floor(seconds / 60)
+  const hours = Math.floor(minutes / 60)
+
+  if (hours > 0) {
+    return `${hours}h ${minutes % 60}m ${seconds % 60}s`
+  } else if (minutes > 0) {
+    return `${minutes}m ${seconds % 60}s`
+  } else {
+    return `${seconds}s`
+  }
+}
+
+export function createBackgroundTask(manager: BackgroundManager) {
+  return tool({
+    description: BACKGROUND_TASK_DESCRIPTION,
+    args: {
+      description: tool.schema.string().describe("Short task description (shown in status)"),
+      prompt: tool.schema.string().describe("Full detailed prompt for the agent"),
+      agent: tool.schema.string().describe("Agent type to use (any registered agent)"),
+    },
+    async execute(args: BackgroundTaskArgs, toolContext) {
+      if (!args.agent || args.agent.trim() === "") {
+        return `❌ Agent parameter is required. Please specify which agent to use (e.g., "explore", "librarian", "build", etc.)`
+      }
+
+      try {
+        const task = await manager.launch({
+          description: args.description,
+          prompt: args.prompt,
+          agent: args.agent.trim(),
+          parentSessionID: toolContext.sessionID,
+          parentMessageID: toolContext.messageID,
+        })
+
+        return `Background task launched successfully.
+
+Task ID: ${task.id}
+Session ID: ${task.sessionID}
+Description: ${task.description}
+Agent: ${task.agent}
+Status: ${task.status}
+
+The system will notify you when the task completes.
+Use \`background_output\` tool with task_id="${task.id}" to check progress:
+- block=false (default): Check status immediately - returns full status info
+- block=true: Wait for completion (rarely needed since system notifies)`
+      } catch (error) {
+        const message = error instanceof Error ? error.message : String(error)
+        return `❌ Failed to launch background task: ${message}`
+      }
+    },
+  })
+}
+
+function delay(ms: number): Promise<void> {
+  return new Promise(resolve => setTimeout(resolve, ms))
+}
+
+function truncateText(text: string, maxLength: number): string {
+  if (text.length <= maxLength) return text
+  return text.slice(0, maxLength) + "..."
+}
+
+function formatTaskStatus(task: BackgroundTask): string {
+  const duration = formatDuration(task.startedAt, task.completedAt)
+  const promptPreview = truncateText(task.prompt, 500)
+  
+  let progressSection = ""
+  if (task.progress?.lastTool) {
+    progressSection = `\n| Last tool | ${task.progress.lastTool} |`
+  }
+
+  let lastMessageSection = ""
+  if (task.progress?.lastMessage) {
+    const truncated = truncateText(task.progress.lastMessage, 500)
+    const messageTime = task.progress.lastMessageAt 
+      ? task.progress.lastMessageAt.toISOString()
+      : "N/A"
+    lastMessageSection = `
+
+## Last Message (${messageTime})
+
+\`\`\`
+${truncated}
+\`\`\``
+  }
+
+  let statusNote = ""
+  if (task.status === "running") {
+    statusNote = `
+
+> **Note**: No need to wait explicitly - the system will notify you when this task completes.`
+  } else if (task.status === "error") {
+    statusNote = `
+
+> **Failed**: The task encountered an error. Check the last message for details.`
+  }
+
+  return `# Task Status
+
+| Field | Value |
+|-------|-------|
+| Task ID | \`${task.id}\` |
+| Description | ${task.description} |
+| Agent | ${task.agent} |
+| Status | **${task.status}** |
+| Duration | ${duration} |
+| Session ID | \`${task.sessionID}\` |${progressSection}
+${statusNote}
+## Original Prompt
+
+\`\`\`
+${promptPreview}
+\`\`\`${lastMessageSection}`
+}
+
+async function formatTaskResult(task: BackgroundTask, client: OpencodeClient): Promise<string> {
+  const messagesResult = await client.session.messages({
+    path: { id: task.sessionID },
+  })
+
+  if (messagesResult.error) {
+    return `Error fetching messages: ${messagesResult.error}`
+  }
+
+  // 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 }>
+  }>
+
+  if (!Array.isArray(messages) || messages.length === 0) {
+    return `Task Result
+
+Task ID: ${task.id}
+Description: ${task.description}
+Duration: ${formatDuration(task.startedAt, task.completedAt)}
+Session ID: ${task.sessionID}
+
+---
+
+(No messages found)`
+  }
+
+  const assistantMessages = messages.filter(
+    (m) => m.info?.role === "assistant"
+  )
+
+  if (assistantMessages.length === 0) {
+    return `Task Result
+
+Task ID: ${task.id}
+Description: ${task.description}
+Duration: ${formatDuration(task.startedAt, task.completedAt)}
+Session ID: ${task.sessionID}
+
+---
+
+(No assistant response found)`
+  }
+
+  const lastMessage = assistantMessages[assistantMessages.length - 1]
+  const textParts = lastMessage?.parts?.filter(
+    (p) => p.type === "text"
+  ) ?? []
+  const textContent = textParts
+    .map((p) => p.text ?? "")
+    .filter((text) => text.length > 0)
+    .join("\n")
+
+  const duration = formatDuration(task.startedAt, task.completedAt)
+
+  return `Task Result
+
+Task ID: ${task.id}
+Description: ${task.description}
+Duration: ${duration}
+Session ID: ${task.sessionID}
+
+---
+
+${textContent || "(No text output)"}`
+}
+
+export function createBackgroundOutput(manager: BackgroundManager, client: OpencodeClient) {
+  return tool({
+    description: BACKGROUND_OUTPUT_DESCRIPTION,
+    args: {
+      task_id: tool.schema.string().describe("Task ID to get output from"),
+      block: tool.schema.boolean().optional().describe("Wait for completion (default: false). System notifies when done, so blocking is rarely needed."),
+      timeout: tool.schema.number().optional().describe("Max wait time in ms (default: 60000, max: 600000)"),
+    },
+    async execute(args: BackgroundOutputArgs) {
+      try {
+        const task = manager.getTask(args.task_id)
+        if (!task) {
+          return `Task not found: ${args.task_id}`
+        }
+
+        const shouldBlock = args.block === true
+        const timeoutMs = Math.min(args.timeout ?? 60000, 600000)
+
+        // Non-blocking: return status immediately
+        if (!shouldBlock) {
+          return formatTaskStatus(task)
+        }
+
+        // Already completed: return result immediately
+        if (task.status === "completed") {
+          return await formatTaskResult(task, client)
+        }
+
+        // Error or cancelled: return status immediately
+        if (task.status === "error" || task.status === "cancelled") {
+          return formatTaskStatus(task)
+        }
+
+        // Blocking: poll until completion or timeout
+        const startTime = Date.now()
+
+        while (Date.now() - startTime < timeoutMs) {
+          await delay(1000)
+
+          const currentTask = manager.getTask(args.task_id)
+          if (!currentTask) {
+            return `Task was deleted: ${args.task_id}`
+          }
+
+          if (currentTask.status === "completed") {
+            return await formatTaskResult(currentTask, client)
+          }
+
+          if (currentTask.status === "error" || currentTask.status === "cancelled") {
+            return formatTaskStatus(currentTask)
+          }
+        }
+
+        // Timeout exceeded: return current status
+        const finalTask = manager.getTask(args.task_id)
+        if (!finalTask) {
+          return `Task was deleted: ${args.task_id}`
+        }
+        return `Timeout exceeded (${timeoutMs}ms). Task still ${finalTask.status}.\n\n${formatTaskStatus(finalTask)}`
+      } catch (error) {
+        return `Error getting output: ${error instanceof Error ? error.message : String(error)}`
+      }
+    },
+  })
+}
+
+export function createBackgroundCancel(manager: BackgroundManager, client: OpencodeClient) {
+  return tool({
+    description: BACKGROUND_CANCEL_DESCRIPTION,
+    args: {
+      taskId: tool.schema.string().describe("Task ID to cancel"),
+    },
+    async execute(args: BackgroundCancelArgs) {
+      try {
+        const task = manager.getTask(args.taskId)
+        if (!task) {
+          return `❌ Task not found: ${args.taskId}`
+        }
+
+        if (task.status !== "running") {
+          return `❌ Cannot cancel task: current status is "${task.status}".
+Only running tasks can be cancelled.`
+        }
+
+        // Fire-and-forget: abort 요청을 보내고 await 하지 않음
+        // await 하면 메인 세션까지 abort 되는 문제 발생
+        client.session.abort({
+          path: { id: task.sessionID },
+        }).catch(() => {})
+
+        task.status = "cancelled"
+        task.completedAt = new Date()
+
+        return `✅ Task cancelled successfully
+
+Task ID: ${task.id}
+Description: ${task.description}
+Session ID: ${task.sessionID}
+Status: ${task.status}`
+      } catch (error) {
+        return `❌ Error cancelling task: ${error instanceof Error ? error.message : String(error)}`
+      }
+    },
+  })
+}

src/tools/background-task/types.ts

@@ -0,0 +1,15 @@
+export interface BackgroundTaskArgs {
+  description: string
+  prompt: string
+  agent: string
+}
+
+export interface BackgroundOutputArgs {
+  task_id: string
+  block?: boolean
+  timeout?: number
+}
+
+export interface BackgroundCancelArgs {
+  taskId: string
+}

src/tools/call-omo-agent/constants.ts

@@ -0,0 +1,24 @@
+export const ALLOWED_AGENTS = ["explore", "librarian"] as const
+
+export const CALL_OMO_AGENT_DESCRIPTION = `Launch a new agent to handle complex, multi-step tasks autonomously.
+
+This is a restricted version of the Task tool that only allows spawning explore and librarian agents.
+
+Available agent types:
+{agents}
+
+When using this tool, you must specify a subagent_type parameter to select which agent type to use.
+
+**IMPORTANT: run_in_background parameter is REQUIRED**
+- \`run_in_background=true\`: Task runs asynchronously in background. Returns immediately with task_id.
+  The system will notify you when the task completes.
+  Use \`background_output\` tool with task_id to check progress (block=false returns full status info).
+- \`run_in_background=false\`: Task runs synchronously. Waits for completion and returns full result.
+
+Usage notes:
+1. Launch multiple agents concurrently whenever possible, to maximize performance
+2. When the agent is done, it will return a single message back to you
+3. Each agent invocation is stateless unless you provide a session_id
+4. Your prompt should contain a highly detailed task description for the agent to perform autonomously
+5. Clearly tell the agent whether you expect it to write code or just to do research
+6. For long-running research tasks, use run_in_background=true to avoid blocking`

src/tools/call-omo-agent/index.ts

@@ -0,0 +1,3 @@
+export * from "./types"
+export * from "./constants"
+export { createCallOmoAgent } from "./tools"

src/tools/call-omo-agent/tools.ts

@@ -0,0 +1,176 @@
+import { tool, type PluginInput } from "@opencode-ai/plugin"
+import { ALLOWED_AGENTS, CALL_OMO_AGENT_DESCRIPTION } from "./constants"
+import type { CallOmoAgentArgs } from "./types"
+import type { BackgroundManager } from "../../features/background-agent"
+import { log } from "../../shared/logger"
+
+export function createCallOmoAgent(
+  ctx: PluginInput,
+  backgroundManager: BackgroundManager
+) {
+  const agentDescriptions = ALLOWED_AGENTS.map(
+    (name) => `- ${name}: Specialized agent for ${name} tasks`
+  ).join("\n")
+  const description = CALL_OMO_AGENT_DESCRIPTION.replace("{agents}", agentDescriptions)
+
+  return tool({
+    description,
+    args: {
+      description: tool.schema.string().describe("A short (3-5 words) description of the task"),
+      prompt: tool.schema.string().describe("The task for the agent to perform"),
+      subagent_type: tool.schema
+        .enum(ALLOWED_AGENTS)
+        .describe("The type of specialized agent to use for this task (explore or librarian only)"),
+      run_in_background: tool.schema
+        .boolean()
+        .describe("REQUIRED. true: run asynchronously (use background_output to get results), false: run synchronously and wait for completion"),
+      session_id: tool.schema.string().describe("Existing Task session to continue").optional(),
+    },
+    async execute(args: CallOmoAgentArgs, toolContext) {
+      log(`[call_omo_agent] Starting with agent: ${args.subagent_type}, background: ${args.run_in_background}`)
+
+      if (!ALLOWED_AGENTS.includes(args.subagent_type as typeof ALLOWED_AGENTS[number])) {
+        return `Error: Invalid agent type "${args.subagent_type}". Only ${ALLOWED_AGENTS.join(", ")} are allowed.`
+      }
+
+      if (args.run_in_background) {
+        if (args.session_id) {
+          return `Error: session_id is not supported in background mode. Use run_in_background=false to continue an existing session.`
+        }
+        return await executeBackground(args, toolContext, backgroundManager)
+      }
+
+      return await executeSync(args, toolContext, ctx)
+    },
+  })
+}
+
+async function executeBackground(
+  args: CallOmoAgentArgs,
+  toolContext: { sessionID: string; messageID: string },
+  manager: BackgroundManager
+): Promise<string> {
+  try {
+    const task = await manager.launch({
+      description: args.description,
+      prompt: args.prompt,
+      agent: args.subagent_type,
+      parentSessionID: toolContext.sessionID,
+      parentMessageID: toolContext.messageID,
+    })
+
+    return `Background agent task launched successfully.
+
+Task ID: ${task.id}
+Session ID: ${task.sessionID}
+Description: ${task.description}
+Agent: ${task.agent} (subagent)
+Status: ${task.status}
+
+Use \`background_output\` tool with task_id="${task.id}" to check progress or retrieve results.
+- block=false: Check status without waiting
+- block=true (default): Wait for completion and get result`
+  } catch (error) {
+    const message = error instanceof Error ? error.message : String(error)
+    return `Failed to launch background agent task: ${message}`
+  }
+}
+
+async function executeSync(
+  args: CallOmoAgentArgs,
+  toolContext: { sessionID: string },
+  ctx: PluginInput
+): Promise<string> {
+  let sessionID: string
+
+  if (args.session_id) {
+    log(`[call_omo_agent] Using existing session: ${args.session_id}`)
+    const sessionResult = await ctx.client.session.get({
+      path: { id: args.session_id },
+    })
+    if (sessionResult.error) {
+      log(`[call_omo_agent] Session get error:`, sessionResult.error)
+      return `Error: Failed to get existing session: ${sessionResult.error}`
+    }
+    sessionID = args.session_id
+  } else {
+    log(`[call_omo_agent] Creating new session with parent: ${toolContext.sessionID}`)
+    const createResult = await ctx.client.session.create({
+      body: {
+        parentID: toolContext.sessionID,
+        title: `${args.description} (@${args.subagent_type} subagent)`,
+      },
+    })
+
+    if (createResult.error) {
+      log(`[call_omo_agent] Session create error:`, createResult.error)
+      return `Error: Failed to create session: ${createResult.error}`
+    }
+
+    sessionID = createResult.data.id
+    log(`[call_omo_agent] Created session: ${sessionID}`)
+  }
+
+  log(`[call_omo_agent] Sending prompt to session ${sessionID}`)
+  log(`[call_omo_agent] Prompt text:`, args.prompt.substring(0, 100))
+
+  try {
+    await ctx.client.session.prompt({
+      path: { id: sessionID },
+      body: {
+        agent: args.subagent_type,
+        tools: {
+          task: false,
+          call_omo_agent: false,
+        },
+        parts: [{ type: "text", text: args.prompt }],
+      },
+    })
+  } catch (error) {
+    const errorMessage = error instanceof Error ? error.message : String(error)
+    log(`[call_omo_agent] Prompt error:`, errorMessage)
+    if (errorMessage.includes("agent.name") || errorMessage.includes("undefined")) {
+      return `Error: Agent "${args.subagent_type}" not found. Make sure the agent is registered in your opencode.json or provided by a plugin.\n\n<task_metadata>\nsession_id: ${sessionID}\n</task_metadata>`
+    }
+    return `Error: Failed to send prompt: ${errorMessage}\n\n<task_metadata>\nsession_id: ${sessionID}\n</task_metadata>`
+  }
+
+  log(`[call_omo_agent] Prompt sent, fetching messages...`)
+
+  const messagesResult = await ctx.client.session.messages({
+    path: { id: sessionID },
+  })
+
+  if (messagesResult.error) {
+    log(`[call_omo_agent] Messages error:`, messagesResult.error)
+    return `Error: Failed to get messages: ${messagesResult.error}`
+  }
+
+  const messages = messagesResult.data
+  log(`[call_omo_agent] Got ${messages.length} messages`)
+
+  // 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]
+
+  if (!lastAssistantMessage) {
+    log(`[call_omo_agent] No assistant message 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>`
+  }
+
+  log(`[call_omo_agent] Found assistant message with ${lastAssistantMessage.parts.length} parts`)
+
+  // 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")
+
+  log(`[call_omo_agent] Got response, length: ${responseText.length}`)
+
+  const output =
+    responseText + "\n\n" + ["<task_metadata>", `session_id: ${sessionID}`, "</task_metadata>"].join("\n")
+
+  return output
+}

src/tools/call-omo-agent/types.ts

@@ -0,0 +1,27 @@
+import type { ALLOWED_AGENTS } from "./constants"
+
+export type AllowedAgentType = (typeof ALLOWED_AGENTS)[number]
+
+export interface CallOmoAgentArgs {
+  description: string
+  prompt: string
+  subagent_type: string
+  run_in_background: boolean
+  session_id?: string
+}
+
+export interface CallOmoAgentSyncResult {
+  title: string
+  metadata: {
+    summary?: Array<{
+      id: string
+      tool: string
+      state: {
+        status: string
+        title?: string
+      }
+    }>
+    sessionId: string
+  }
+  output: string
+}

src/tools/index.ts

@@ -22,6 +22,28 @@ import { glob } from "./glob"
 import { slashcommand } from "./slashcommand"
 import { skill } from "./skill"
 
+import {
+  createBackgroundTask,
+  createBackgroundOutput,
+  createBackgroundCancel,
+} from "./background-task"
+
+import type { PluginInput } from "@opencode-ai/plugin"
+import type { BackgroundManager } from "../features/background-agent"
+
+type OpencodeClient = PluginInput["client"]
+
+export { createCallOmoAgent } from "./call-omo-agent"
+export { createLookAt } from "./look-at"
+
+export function createBackgroundTools(manager: BackgroundManager, client: OpencodeClient) {
+  return {
+    background_task: createBackgroundTask(manager),
+    background_output: createBackgroundOutput(manager, client),
+    background_cancel: createBackgroundCancel(manager, client),
+  }
+}
+
 export const builtinTools = {
   lsp_hover,
   lsp_goto_definition,

src/tools/look-at/constants.ts

@@ -0,0 +1,23 @@
+export const MULTIMODAL_LOOKER_AGENT = "multimodal-looker" as const
+
+export const LOOK_AT_DESCRIPTION = `Analyze media files (PDFs, images, diagrams) that require visual interpretation.
+
+Use this tool to extract specific information from files that cannot be processed as plain text:
+- PDF documents: extract text, tables, structure, specific sections
+- Images: describe layouts, UI elements, text content, diagrams
+- Charts/Graphs: explain data, trends, relationships
+- Screenshots: identify UI components, text, visual elements
+- Architecture diagrams: explain flows, connections, components
+
+Parameters:
+- file_path: Absolute path to the file to analyze
+- goal: What specific information to extract (be specific for better results)
+
+Examples:
+- "Extract all API endpoints from this OpenAPI spec PDF"
+- "Describe the UI layout and components in this screenshot"
+- "Explain the data flow in this architecture diagram"
+- "List all table data from page 3 of this PDF"
+
+This tool uses a separate context window with Gemini 2.5 Flash for multimodal analysis,
+saving tokens in the main conversation while providing accurate visual interpretation.`

src/tools/look-at/index.ts

@@ -0,0 +1,3 @@
+export * from "./types"
+export * from "./constants"
+export { createLookAt } from "./tools"

src/tools/look-at/tools.ts

@@ -0,0 +1,91 @@
+import { tool, type PluginInput } from "@opencode-ai/plugin"
+import { LOOK_AT_DESCRIPTION, MULTIMODAL_LOOKER_AGENT } from "./constants"
+import type { LookAtArgs } from "./types"
+import { log } from "../../shared/logger"
+
+export function createLookAt(ctx: PluginInput) {
+  return tool({
+    description: LOOK_AT_DESCRIPTION,
+    args: {
+      file_path: tool.schema.string().describe("Absolute path to the file to analyze"),
+      goal: tool.schema.string().describe("What specific information to extract from the file"),
+    },
+    async execute(args: LookAtArgs, toolContext) {
+      log(`[look_at] Analyzing file: ${args.file_path}, goal: ${args.goal}`)
+
+      const prompt = `Analyze this file and extract the requested information.
+
+File path: ${args.file_path}
+Goal: ${args.goal}
+
+Read the file using the Read tool, then provide ONLY the extracted information that matches the goal.
+Be thorough on what was requested, concise on everything else.
+If the requested information is not found, clearly state what is missing.`
+
+      log(`[look_at] Creating session with parent: ${toolContext.sessionID}`)
+      const createResult = await ctx.client.session.create({
+        body: {
+          parentID: toolContext.sessionID,
+          title: `look_at: ${args.goal.substring(0, 50)}`,
+        },
+      })
+
+      if (createResult.error) {
+        log(`[look_at] Session create error:`, createResult.error)
+        return `Error: Failed to create session: ${createResult.error}`
+      }
+
+      const sessionID = createResult.data.id
+      log(`[look_at] Created session: ${sessionID}`)
+
+      log(`[look_at] Sending prompt to session ${sessionID}`)
+      await ctx.client.session.prompt({
+        path: { id: sessionID },
+        body: {
+          agent: MULTIMODAL_LOOKER_AGENT,
+          tools: {
+            task: false,
+            call_omo_agent: false,
+            look_at: false,
+          },
+          parts: [{ type: "text", text: prompt }],
+        },
+      })
+
+      log(`[look_at] Prompt sent, fetching messages...`)
+
+      const messagesResult = await ctx.client.session.messages({
+        path: { id: sessionID },
+      })
+
+      if (messagesResult.error) {
+        log(`[look_at] Messages error:`, messagesResult.error)
+        return `Error: Failed to get messages: ${messagesResult.error}`
+      }
+
+      const messages = messagesResult.data
+      log(`[look_at] Got ${messages.length} messages`)
+
+      // 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]
+
+      if (!lastAssistantMessage) {
+        log(`[look_at] No assistant message found`)
+        return `Error: No response from multimodal-looker agent`
+      }
+
+      log(`[look_at] Found assistant message with ${lastAssistantMessage.parts.length} parts`)
+
+      // 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")
+
+      log(`[look_at] Got response, length: ${responseText.length}`)
+
+      return responseText
+    },
+  })
+}

src/tools/look-at/types.ts

@@ -0,0 +1,4 @@
+export interface LookAtArgs {
+  file_path: string
+  goal: string
+}

src/tools/skill/tools.ts

@@ -1,9 +1,10 @@
 import { tool } from "@opencode-ai/plugin"
-import { existsSync, readdirSync, statSync, readlinkSync, readFileSync } from "fs"
+import { existsSync, readdirSync, readFileSync } from "fs"
 import { homedir } from "os"
-import { join, resolve, basename } from "path"
+import { join, basename } from "path"
 import { z } from "zod/v4"
 import { parseFrontmatter, resolveCommandsInText } from "../../shared"
+import { resolveSymlink } from "../../shared/file-utils"
 import { SkillFrontmatterSchema } from "./types"
 import type { SkillScope, SkillMetadata, SkillInfo, LoadedSkill, SkillFrontmatter } from "./types"
 
@@ -37,15 +38,7 @@ function discoverSkillsFromDir(
     const skillPath = join(skillsDir, entry.name)
 
     if (entry.isDirectory() || entry.isSymbolicLink()) {
-      let resolvedPath = skillPath
-      try {
-        const stats = statSync(skillPath, { throwIfNoEntry: false })
-        if (stats?.isSymbolicLink()) {
-          resolvedPath = resolve(skillPath, "..", readlinkSync(skillPath))
-        }
-      } catch {
-        continue
-      }
+      const resolvedPath = resolveSymlink(skillPath)
 
       const skillMdPath = join(resolvedPath, "SKILL.md")
       if (!existsSync(skillMdPath)) continue
@@ -83,18 +76,6 @@ const skillListForDescription = availableSkills
   .map((s) => `- ${s.name}: ${s.description} (${s.scope})`)
   .join("\n")
 
-function resolveSymlink(skillPath: string): string {
-  try {
-    const stats = statSync(skillPath, { throwIfNoEntry: false })
-    if (stats?.isSymbolicLink()) {
-      return resolve(skillPath, "..", readlinkSync(skillPath))
-    }
-    return skillPath
-  } catch {
-    return skillPath
-  }
-}
-
 async function parseSkillMd(skillPath: string): Promise<SkillInfo | null> {
   const resolvedPath = resolveSymlink(skillPath)
   const skillMdPath = join(resolvedPath, "SKILL.md")