opencode-agent

Opencode Agent

Delegate tasks to opencode, an alternative AI coding agent. This skill runs opencode run as a subprocess, letting you hand off work to models like zai-coding-plan/glm-5-turbo without leaving your Claude Code session.

When to use this

• The user explicitly asks to run something with opencode
• The user wants to use a non-Claude model (GLM, etc.) for a specific task
• The user says "launch an opencode agent" or similar phrasing

How it works

Run the user's task via opencode run, passing the task as the message argument. The default model is zai-coding-plan/glm-5-turbo. If the user specifies a different model, pass it with the -m flag.

Basic execution

opencode run -m zai-coding-plan/glm-5-turbo "<task description>"

With a specific model

If the user requests a different model, override:

opencode run -m <provider/model> "<task description>"

With a custom prompt/agent

opencode run --prompt "<system prompt>" "<task description>"
opencode run --agent <agent-name> "<task description>"

Step-by-step

1. Parse the request. Extract the task the user wants opencode to perform. If the user specified a model, note it. Otherwise default to zai-coding-plan/glm-5-turbo.

2. Run opencode. Execute the task using Bash. Choose the right execution strategy based on the task's expected duration:

   Short tasks (file creation, simple edits, quick scripts) — run inline with a 2-minute timeout:

   opencode run -m zai-coding-plan/glm-5-turbo "<task>"
   

   Medium tasks (multi-file changes, test generation, refactoring) — use a longer timeout up to 10 minutes:

   opencode run -m zai-coding-plan/glm-5-turbo "<task>"   # with timeout set to 600000
   

   Long-running tasks (large codebases, complex generation, full project scaffolding) — run in the background so the user isn't blocked:

   opencode run -m zai-coding-plan/glm-5-turbo "<task>"   # with run_in_background: true
   

   When running in the background, you'll be notified when the task completes. Let the user know the task is running and you'll report back when it finishes.

   Always pass -m zai-coding-plan/glm-5-turbo unless the user specified a different model.

3. Verify the result. After opencode finishes, check that the expected output was produced. If opencode created or modified files, read them to confirm. Report the outcome to the user.

4. Handle errors. If opencode fails or isn't installed, tell the user clearly. Common issues:

   • opencode: command not found — opencode isn't installed. Suggest npm install -g opencode or checking their PATH.
   • Timeout — increase the timeout or run in the background. If the task is genuinely too large, break it into smaller subtasks.
   • Model not available — the requested model may not be configured. Suggest running opencode models to see available options.

Tips

• You can continue a previous opencode session with opencode run -c "<follow-up message>".
• To see what models are available: opencode models.
• The user's opencode configuration (API keys, providers) is managed separately from Claude Code — if a model isn't working, point them to opencode providers to set up credentials.