A developer has released openclaw-code-agent, an OpenClaw plugin that transforms one-shot coding agent interactions into fully orchestrated background sessions. Built to solve what the creator calls "vanilla OpenClaw" limitations—including silly mistakes and lack of plan approval—the plugin wraps Claude Code and Codex sessions in an async control layer that works over Telegram, Discord, or any OpenClaw-supported messaging platform.

Core Orchestration Features

The plugin provides six core tools for session management: agent_launch starts new sessions with configurable models and permission modes, agent_respond sends follow-up messages including plan approvals, agent_kill terminates sessions, agent_output reads buffered output, agent_sessions lists active and persisted sessions, and agent_stats tracks usage metrics. Chat commands (/agent, /agent_respond, /agent_kill, etc.) expose all tools directly in conversations.

Three Operational Modes

The plugin operates in three modes: 'ask' mode pauses for user approval before execution, 'delegate' mode auto-approves low-risk plans and only escalates for significant decisions, and permission modes (default, plan, acceptEdits, bypassPermissions) control what the agent can do. Claude Code exposes plan mode natively; Codex uses soft first-turn planning while staying in implement mode externally—users approve plans without seeing plan state in session status.

Session Lifecycle and Notifications

Sessions emit deterministic wake signals at turn completion with output previews. The smart waiting detector reduces false-positive escalations. Completed sessions persist for resume (configurable TTL, default 24h) before garbage collection. Notifications route to the originating Telegram thread with emojis: 🚕 for launch, 🔔 for agent questions, 📋 for plan ready, ✅ for completion, ❌ for failures.

Auto-Respond and Anti-Cascade Protection

The orchestrator automatically handles permission requests ('Yes, proceed.') and confirmations ('Yes, continue.') while forwarding architecture decisions, destructive operations, scope changes, and ambiguous questions to users. Critically, the plugin never launches new sessions from wake events—anti-cascade protection prevents notification loops when agents wait for input.

Supported Agents and Configuration

Claude Code gets full support via @anthropic-ai/claude-agent-sdk; Codex uses @openai/codex-sdk thread streaming. The pluggable AgentHarness interface allows adding other agents later. Configuration lives in ~/.openclaw/openclaw.json with options for maxSessions (default 5), permissionMode, planApproval strategy, idleTimeoutMinutes, sessionGcAgeMinutes, and workspace-based channel routing.

Key Takeaways

  • Six core tools enable full session lifecycle management from chat platforms
  • Three operational modes (ask, delegate) plus permission modes provide granular control over agent behavior
  • Anti-cascade protection prevents notification loops when agents wait for user input
  • Native support for Claude Code and Codex with pluggable architecture for future agents

The Bottom Line

This plugin fills a real gap for developers running AI coding agents at scale—vanilla ACP is fine for one-off tasks but useless for persistent, reviewable workflows. The auto-respond rules alone save hours of babysitting. If you're orchestrating Claude Code or Codex from chat, this is the layer OpenClaw should have shipped with.