Why This Exists

OpenClaw's built-in ACP is useful for simple one-shot tasks, but it falls short for real development work — no plan approval, no session persistence, no concurrent sessions. The CLI requires babysitting a terminal. So I built this plugin to fill the gap.

Three Execution Modes

• Ask — the agent proposes a plan and waits for your approval before touching a single file. Approve or redirect from chat.
• Delegate — auto-approves low-risk plans, escalates when a decision actually matters.
• Autonomous — fully hands-off. Notified when done or when it needs input.

Example Flow (Ask Mode)

1. You send a task: "Fix the bug in auth.ts"
2. The agent explores the codebase and drafts a plan
3. You get notified in the same Telegram thread with the plan
4. Reply "go" to approve — the agent implements with full permissions
5. You get a completion summary: files changed, cost, duration

Delegate Mode in Action

In delegate mode, the orchestrator reads the plan automatically. Routine changes get approved without interrupting you. Risky or ambiguous decisions come to you with the agent's exact question.

Key Features

• Plan → Execute workflow — review the agent's plan before it runs. Approve with a single reply.
• Session resume & fork — sessions can be resumed after a gateway restart or forked to try a different approach
• Thread-routed notifications — updates arrive in the exact Telegram topic or Discord thread where you launched the session
• Concurrent sessions — run multiple sessions simultaneously, each with a unique ID and human-readable name
• Multi-harness — supports Claude Code and Codex through the same interface and same commands
• Auto-respond rules — routine permission requests are handled automatically; real decisions are forwarded to you