opencode-keep-running

An OpenCode plugin that detects when the AI agent is stuck and automatically sends a "continue" prompt to keep the session running.

How It Works

1. Tracks activity — Monitors session.status, message.part.updated, message.updated, and tool.execute.after events
2. Detects stuck sessions — When a session is busy but no activity is detected within the threshold, it's considered stuck
3. Auto-continues — Aborts the stuck run and sends a configurable message to prompt the agent to keep going
4. Re-schedules monitoring — After a continue is sent, monitoring resumes so subsequent stucks are also detected (up to maxContinues)

Installation

opencode plugin bacluc-opencode-keep-running

Or add to your opencode.json:

{
  "plugin": ["bacluc-opencode-keep-running@<version>"]
}

Configuration

{
  "plugin": [
    [
      "opencode-keep-running",
      {
        "threshold": 120000,
        "message": "continue",
        "abortBeforeContinue": true,
        "abortCooldown": 3000,
        "maxContinues": 5,
        "continueCooldown": 30000,
        "debug": false
      }
    ]
  ]
}

Option	Type	Default	Description
threshold	number	120000 (2 min)	Inactivity threshold in ms before considering the agent stuck
message	string	"continue"	Message to send when the agent is stuck
abortBeforeContinue	boolean	true	Whether to abort the current run before sending continue
abortCooldown	number	3000 (3s)	Delay in ms after aborting before sending continue
maxContinues	number	5	Maximum continue attempts per session
continueCooldown	number	30000 (30s)	Cooldown in ms between continues
debug	boolean	false	Enable debug logging via opencode app.log

Behavior Details

• Continue count is tracked per session and resets only when a new session is created. Going idle does not reset the count — if the agent gets stuck 5 times in the same session, it stops trying.
• Activity events reset the inactivity timer, meaning any tool call, message part, or status change (including retry) keeps the session from being considered stuck.
• Re-scheduling: After a continue is sent, monitoring resumes automatically. If the agent gets stuck again, another continue is sent (subject to maxContinues and continueCooldown).

Development

npm install
npm test