Vertex AI: BadRequestError 'instructions: Extra inputs are not permitted' when running agents #4282
Description
Description
When running agents (using the Task tool) with Vertex AI (Google Cloud), OpenCode crashes with the following error:
litellm.BadRequestError: Vertex_aiException BadRequestError - b'{"type":"error","error":{"type":"invalid_request_error","message":"instructions: Extra inputs are not permitted"},"request_id":"req_vrtx_011CV5YEZHNBbDtN3P12S1cH"}'. Received Model Group=claude-sonnet-4-5 Available Model Group Fallbacks=None
Configuration
{
"provider": "vertex",
"vertexProjectId": "<my-project-id>",
"vertexRegion": "europe-west1",
"model": "claude-sonnet-4-5@20250514"
}Root Cause
Vertex AI's Anthropic API implementation doesn't support the instructions parameter that LiteLLM is trying to send when OpenCode spawns sub-agents via the Task tool. This is similar to issue #2038 which occurred with Bedrock.
The main OpenCode agent works fine, but whenever a sub-agent is created (e.g., when using the Task tool), the request includes an instructions field that Vertex AI rejects.
Expected Behavior
Agents should work seamlessly with Vertex AI just like the main OpenCode instance does.
Actual Behavior
Sub-agents fail immediately with the "Extra inputs are not permitted" error, causing OpenCode to crash.
Environment
- Provider: Google Vertex AI
- Model: claude-sonnet-4-5@20250514
- Region: europe-west1
- OpenCode version: Latest (as of Nov 2025) =>
1.0.62
Suggested Fix
Similar to how the main agent request is formatted for Vertex AI compatibility, sub-agent requests should also avoid sending unsupported parameters like instructions when using Vertex AI as the provider.