PostToolUse prompt hooks fail on Vertex AI backend with 'output_config: Extra inputs are not permitted'

Description

Plugin PostToolUse hooks with "type": "prompt" fail on the Vertex AI backend. The hook matches correctly and fires, but the prompt evaluation API call returns a 400 error.

Error

PostToolUse:mcp__github__issue_write hook error
Hook evaluator API error: API Error: 400
{"type":"error","error":{"type":"invalid_request_error","message":"output_config: Extra inputs are not permitted"}}

The request ID prefix req_vrtx_ confirms this is a Vertex AI backend call.

Reproduction

1. Create a plugin with a PostToolUse prompt hook in .claude-plugin/plugin.json:

{
    "name": "my-plugin",
    "hooks": {
        "PostToolUse": [
            {
                "matcher": "mcp__github__issue_write",
                "hooks": [
                    {
                        "type": "prompt",
                        "prompt": "An issue was just created. Do something with it."
                    }
                ]
            }
        ]
    }
}

1. Launch Claude Code with --plugin-dir /path/to/plugin
2. Call mcp__github__issue_write (or any tool matching the matcher)
3. Observe: the hook fires but fails with the output_config error

Expected behavior

The prompt hook should evaluate successfully regardless of the API backend (direct Anthropic API, Vertex AI, Bedrock).

Environment

• Claude Code version: 2.1.121
• Backend: Vertex AI (Google Cloud)
• OS: Linux (Fedora 43)
• Plugin loaded via: --plugin-dir

Notes

• The hook does fire — the matcher and plugin loading work correctly
• The error appears to be in the hook evaluator's API call, which sends an output_config parameter that Vertex AI rejects
• Non-prompt hook types (e.g., "type": "command") were not tested