[FEATURE]: External Tool Execution API #20697

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before. (I tried but there are many issues I can't really tell)

Describe the enhancement you want to request

Problem

When using OpenCode (via sdk) as an AI backend via API, there's no way to:

1. Disable OpenCode's built-in tool execution
2. Execute tools locally
3. Send tool results back to OpenCode so the model can see them

Currently, tools can be configured with action: "allow" (executes immediately) or action: "ask" (waits for permission response), but:

• With "allow": OpenCode executes the tool itself
• With "ask": The permission API responds successfully, but the tool never completes

Desired Flow

1. App creates session with permission: [{ permission: "bash", pattern: "*", action: "deny" }] (or similar)
2. OpenCode detects tool is needed but doesn't execute it
3. App executes the tool locally
4. App sends the tool result back to OpenCode via API
5. OpenCode continues with the tool result in its context

Suggested API

Option A: A new endpoint to submit tool results:

POST /session/{id}/tool-result
{
  callID: "call_xxx",
  tool: "bash", 
  output: "...",
  status: "completed"
}

Option B: Support injecting tool results via a new message part type that OpenCode recognizes as a tool result rather than user input

Option C: Add a session configuration option like tool_execution: "external" that completely disables internal execution

Workaround

Currently we're using action: "allow" which works but means OpenCode executes the tools, not my app This loses the benefit of local execution (using user's environment, preserving context, etc.).

Environment

• OpenCode version: 1.3.13
• SDK: @opencode-ai/sdk