ai: split OpenAI Responses and compatible providers

[martinhoyer]

Summary

Split the existing OpenAI integration into two distinct providers:

• openai uses OpenAI’s native /v1/responses endpoint.
• openai-compatible retains the Chat Completions implementation.

Motivation

Using OpenAI API with gpt-5.6 models currently fails on:

{
  "error": {
    "message": "Unsupported value: 'temperature' does not support 0.0 with this model. Only the default (1) value is supported.",
    "type": "invalid_request_error",
    "param": "temperature",
    "code": "unsupported_value"
  }
}

Shoehorning special temperature handling into the existing compatible provider did not solve the underlying problem. Modern OpenAI reasoning and tool-calling workflows ultimately need the Responses API, while third-party OpenAI-compatible services generally continue to expose Chat Completions.
Splitting openai and openai-compatible therefore seemed like the most practical and maintainable approach.

Changes

• Add a dedicated OpenAI Responses API client.
• Add [ai.openai] configuration, separate from [ai.openai_compat].
• Clearly reject legacy-only provider = "openai" configurations with migration guidance.
• Preserve and replay complete Responses output items, including encrypted reasoning state and original function-call identifiers.
• Classify retryable OpenAI errors correctly.
• Handle JSON mode requirements and cached-input token accounting.
• Update settings, examples, provider documentation, and design documentation.

Not sure if there should be separate design doc as well.

[rgushchin]

Note

This review was generated with the assistance of an AI tool.

Overall looks great and well-architected! A few quick observations and potential regression risks:

1. Default max_tokens (4,096) may truncate reasoning:
   In src/ai/mod.rs, omitting [ai.openai].max_tokens defaults to 4,096. On the Responses API, max_output_tokens bounds the sum of internal reasoning tokens and visible output. Reasoning models like gpt-5.6 with medium effort can easily exhaust 4,096 tokens during the reasoning phase alone, causing premature incomplete (truncated) responses. Consider defaulting to 65536 for the openai provider.

2. Handle status: "failed" in translate_ai_response:
   translate_ai_response checks resp.status == "incomplete", but if OpenAI returns a 200 OK with status: "failed" and an error object, it currently returns Ok(AiResponse) with empty content and tool calls. Propagating the error message directly would prevent confusing downstream ValidationErrors.

3. Global cache invalidation:
   Bumping CACHE_KEY_VERSION in src/ai/cache.rs to "sashiko-ai-cache-v2" invalidates cached responses for all providers (Gemini, Claude, Bedrock, Ollama, etc.). Users running benchmarks or workloads with response_cache = true will experience a full cold-cache restart across all backends.

4. base_url doc & normalization:
   docs/configuration.md lists [ai.openai].base_url as "model-derived", but default_base_url() statically points to https://api.openai.com/v1/responses. Also, OpenAiClient does not normalize paths (e.g. appending /responses), so custom proxy URLs like http://host/v1 without /responses will fail.

5. temperature suppression:
   translate_ai_request unconditionally passes temperature: None. This fixes gpt-5.6 rejecting 0.0, but prevents tuning temperature for non-reasoning models (like gpt-4o) or future models that support it on /v1/responses.

[martinhoyer]

Consider defaulting to 65536 for the openai provider.

Thanks, already noticed 4096 is too low. Trying 16k and it was quite enough for sol with xhigh reasoning for a patch review. Wouldn't 65k be an overkill?

[martinhoyer]

@rgushchin adressing feedback in fixup commits, thanks.

• Raised the default output budget from 4K to 16K, validated with Sol/xhigh, and now log incomplete reasons and token usage.
• Surface failed Responses as provider errors instead of empty responses.
• Scoped cache invalidation to OpenAI Responses only.
• Corrected base_url documentation and added /responses normalization.
• Forward temperature for supported models while omitting it for known reasoning models.