Content
# Tool List
> Claude Code · Codex · Gemini multi-agent collaboration system. Based on **Verification-First** principle.
## Philosophy: Verification-First
**Claude Code is the main executor**. It directly performs code editing, file analysis, planning, and WebSearch.
Codex / Gemini delegation is only used for **cross-validation purposes**:
- Reduce hallucinations with independent model perspectives
- Adversarial review of design and assumptions
- Fact-checking based on Google index (Gemini)
- Analysis of long-context documents (> 2M) (Gemini)
**No automatic delegation**. The user decides the timing of the call. Claude only suggests a cross-validation option with `/codex` in high-risk contexts (migration/auth/crypto/security).
## Entry Points
All external agent calls go through **user-callable slash skills**:
| Skill | Purpose | Mode |
|---|---|---|
| `/codex` | Code review, design challenges, research and implementation delegation | `review` · `adversarial-review` · `explore` · `task` · `resume` |
| `/gemini` | Fact-checking, large document analysis, independent research | `research` · `review` · `factcheck` · `explore` · `task` |
| `/openclaw` | Browser, GUI, Telegram (M4 remote) | — |
The skill collects the current context (git state, recent conversations, diff) and presents **3-5 recommended menu options**. The user selects a mode to execute in **background + allow failure**.
- `rescue` / `task` (side-effect possible): 1-line echo-confirm before execution
- `review` / `adversarial-review` / `explore` (read-only): immediate execution
## Execution Hierarchy
```
User
│
▼
Claude Code ────────────────── direct execution (code editing, analysis, judgment)
│
├─ /codex → scripts/codex-dispatch.sh → codex-companion → Codex CLI (gpt-5.4 / 5.3)
├─ /gemini → scripts/gemini-dispatch.sh → gemini-companion → Gemini CLI (2.5-pro / 2.5-flash)
└─ /openclaw → SSH m4 → OpenClaw (browser, canvas, JS rendering, session)
```
**Dispatch wrapper** absorbs internal path changes:
- Automatic injection of `CLAUDE.md` in git root (`--no-context` to disable)
- Automatic addition of `--background` to all calls
- `codex-dispatch.sh resume` — resume previous task thread (Codex only)
- `codex-dispatch.sh wait <job-id>` — emit `<task-notification>` on completion
## Responsibility Separation
| Role | Responsible |
|---|---|
| **Claude** | Main executor. Editing, analysis, planning, direct search. Orchestration judgment |
| **Codex** | Independent perspective review, adversarial, implementation and research delegated by user |
| **Gemini** | Google index fact-checking, research, large document summarization |
| **OpenClaw** | Browser automation, login session, chart rendering, JS SPA crawling |
Detailed decision flow: [`ROUTING_TABLE.md`](./ROUTING_TABLE.md)
## Quick Start
```bash
# Check synchronization status
bash scripts/sync.sh --check
# Deploy adapters/ changes to ~/CLAUDE.md, ~/.codex/, ~/.gemini/
bash scripts/sync.sh
# Dispatch health check
bash scripts/codex-dispatch.sh health
bash scripts/gemini-dispatch.sh health
# Create AGENTS.md stub for other projects (automatically referencing CLAUDE.md)
bash scripts/init-project-agents.sh
```
## Structure
```
adapters/ Agent settings SoT (distributed by sync.sh to each agent home)
claude_global.md → ~/CLAUDE.md
codex_home.md → ~/AGENTS.md (Codex home scope, global load)
gemini.md → ~/.gemini/GEMINI.md
USER_CONTEXT.md User environment SoT (appended to ~/CLAUDE.md + ~/AGENTS.md)
scripts/ Orchestration scripts
sync.sh Distribute adapters to agent homes
codex-dispatch.sh Codex call stabilization wrapper
gemini-dispatch.sh Gemini call stabilization wrapper
guard.sh Destructive command blocking hook
sync-codex-skills.sh codex:true skills → ~/.codex/skills/
init-project-agents.sh Create AGENTS.md stub for other projects
hooks/ Claude Code hook scripts
device/ Manual installation per device (codex wrapper, plist, etc.)
config/ Environment and tool settings (settings_common.json, statusline.sh, mcp_setup.sh, …)
docs/ Operation documents (orchestration roadmap, remote-wake-pc, …)
context/ Local detailed context (notion-rules, statusline)
examples/ Reusable templates and prompts
agent_config.yaml Model tier, complexity classification, fallback chain
ROUTING_TABLE.md Task → agent routing decision table
ROADMAP.md Milestones and design decision history
```
Refer to [`docs/REPO_LAYOUT.md`](./docs/REPO_LAYOUT.md) for folder organization rules and lifecycle.
## Model Tier
Select models by complexity in `agent_config.yaml`.
| Complexity | Codex | Gemini | Claude |
|---|---|---|---|
| low | `gpt-5.3-codex-spark` | `gemini-2.5-flash-lite` | haiku |
| medium | `gpt-5.3-codex` (high) | `gemini-2.5-flash` | sonnet |
| high | `gpt-5.3-codex` (high) | `gemini-2.5-flash` | sonnet (high) |
| ultra | `gpt-5.4` (xhigh) | `gemini-2.5-pro` | opus |
## Multi-Device
Synchronize using git. Pull and run `bash scripts/sync.sh` on other devices.
Refer to `scripts/device/` for manual installation per device (wrapper, local path, etc.).
## Known Tradeoffs
- **User must decide call timing**. Intentional omission based on Verification-First principle.
- **Gemini struggles with complex multi-part prompts**. Narrow down to single clear questions.
- **Mesh collaboration (triangulate/debate) disabled**. Removed to prevent "collaboration theater". To be restored with defined orchestration contracts.
