Subagents

OpenCode and Cursor each have their own idea of a subagent. open-cursor keeps those catalogs separate: delegation from a Cursor model goes back through OpenCode's task tool.

This matters because the names visible in OpenCode are not guaranteed to match Cursor's built-in Task tool.

Requirements

The bridge activates only when OpenCode offers a tool named task and CURSOR_ACP_BRIDGE_JSON is enabled. It copies OpenCode's Task description and schema into the request so the model sees the currently allowed subagent_type values.

The default bridge setting is enabled. To disable it:

export CURSOR_ACP_BRIDGE_JSON=false

Bridge envelope

For Task calls, the model is instructed to emit one JSON object:

{
  "name": "task",
  "arguments": {
    "description": "inspect stream parser",
    "prompt": "Trace how snapshot events become text deltas and report the edge cases.",
    "subagent_type": "explore"
  }
}

The valid subagent_type comes from the Task description in the active OpenCode session. Do not copy a name from Cursor's own agent picker.

The bridge requires non-empty description, prompt, and subagent_type strings. Optional task_id and command fields must also be strings. Invalid envelopes stay as assistant text instead of becoming a tool call.

Execution boundary

The sequence is:

Cursor model
  → strict Task JSON envelope
  → open-cursor validates and converts it
  → OpenCode executes task
  → tool result returns on the next model request

OpenCode therefore keeps control of the Task tool and its permissions. The bridge does not invoke Cursor's built-in Task tool.

Resume interaction

The Task schema participates in the session-resume fingerprint. If the available OpenCode agents change between turns, the old resume entry is rejected rather than sending stale subagent metadata.

If a Task call appears as raw JSON, confirm that OpenCode exposed task, the requested subagent name is in its description, and CURSOR_ACP_BRIDGE_JSON has not been disabled.