Persistent OpenCode servers

Enable persistentProcess on an OpenCode agent to keep its server running between requests:

{
  "tier": "opencode",
  "model": "provider/your-model",
  "persistentProcess": true
}

Use a model available through your configured OpenCode provider. This setting preserves that model; it does not select a cheaper one. Restart AgentX after changing the setting, once active tasks have finished.

Requirements

Install OpenCode v2 and configure a provider using the OpenCode setup guide. Check your installation with opencode --version. This implementation was tested with v2.0.12 and its serve --stdio and run --server interfaces. See the official command reference.

Missing or older installations retain the existing CLI path and its installation error reporting. A failed server startup falls back before submitting the request. For v2, that fallback uses --standalone so the private server receives AgentX's environment. A dispatched request is never automatically replayed.

What stays warm

Each agent/channel/chat combination gets a dedicated loopback server with a random password. Provider and MCP services remain in that server. A small opencode run client still starts per turn to stream JSON output using the existing runtime. OpenCode v2 normally uses its own shared background service; AgentX's dedicated servers preserve each conversation's environment and MCP identity.

The pool has eight slots and closes servers after five idle minutes. Busy servers are never evicted. Model, workspace, permission or environment changes replace an idle server. Fresh-session requests also replace it and omit the resume ID. Cancellation, timeout, or failed execution invalidates the server. Daemon shutdown closes the pool after draining active work.

These limits are fixed for now. processPool settings and agentx process commands apply to the Claude pool, not this OpenCode pool.

Measuring startup

Execution events expose opencode.ready with startupMs and reused, plus opencode.first_output with elapsedMs. Ready time measures server acquisition; it excludes the small per-turn CLI client and model latency. Compare repeated requests before drawing conclusions about end-to-end response speed.

Check it worked

1. Send the agent two short messages in the same chat.
2. Browser: open each run's Task page from Live or Activity and read its live output.
3. The first run shows . opencode started ready=…ms. The second shows . opencode reused ready=…ms: the process was kept between requests.

If something is wrong

• Nothing changed after editing the setting: the daemon only reads persistentProcess when it starts. Restart it once running tasks have finished (see Restart without losing work).
• Every request shows . opencode using CLI fallback: …: the reason follows the colon. Check opencode --version reports v2 or newer and that a provider is configured, then restart the daemon.
• Replies use the wrong model: check the agent's model is available through your OpenCode provider. This setting never changes the model.