fix(llm): add disableResponseFormat option for proxies that reject response_format

[laurentftech]

Problem

Some OpenAI-compatible proxy endpoints return {"detail":"There was an error parsing the body"} when the request includes response_format — they support the chat completions API but not structured output extensions. This causes all LLM requests to fail silently (retrying with empty error messages).

Root cause: spec-gen sends response_format: {type: "json_schema", ...} or response_format: {type: "json_object"} for JSON completions. Endpoints like vLLM or custom gateways reject this field entirely.

Diagnosis: confirmed by testing with/without response_format via PowerShell — the request without it succeeds, the one with it fails.

Fix

Add disableResponseFormat: boolean option:

• LLMServiceOptions + Required<LLMServiceOptions>
• OpenAICompatibleProvider constructor (4th arg, default false)
• GenerationConfig in types/index.ts
• Wired through generate.ts from config

When true, response_format is omitted. The LLM still produces JSON via system prompt instructions; completeJSON() handles free-form JSON as it already does for models without schema support.

Also improves error messages: all provider errors now include the HTTP status code (HTTP 422: ...) instead of just the raw body — making retry log lines actionable.

Config

generation:
  provider: openai-compat
  openaiCompatBaseUrl: https://your-proxy/v1
  disableResponseFormat: true
  model: your-model

Tests

Two new unit tests on OpenAICompatibleProvider:

• disableResponseFormat=true → response_format absent from request body
• disableResponseFormat=false → response_format present (default behavior unchanged)

🤖 Generated with Claude Code