Summary

Three small, additive improvements to the OpenAI backend (src/llm/backends/openai.py + src/embedding_client.py) that broaden compatibility with OpenAI-compatible providers (DeepSeek, Alibaba Bailian, vLLM-hosted models, etc.) without changing behaviour on OpenAI itself.

1. provider_params → OpenAI SDK extra_body passthrough

ConfiguredModelSettings.overrides.provider_params is documented as a free-form passthrough for backend-specific knobs, but the OpenAI backend's _build_params was cherry-picking a small whitelist (top_p, frequency_penalty, presence_penalty, seed, verbosity) and silently dropping the rest. Any unrecognised key now flows through to the OpenAI SDK's extra_body parameter, which is the canonical way to pass provider-specific fields.

This unblocks e.g.:

[dialectic.levels.minimal.model_config.overrides]
provider_params = { thinking = { type = "disabled" } }

…on DeepSeek's v4 family, which defaults thinking.type=enabled (returns reasoning_content + rejects tool_choice=required).

2. Forward dimensions on embedding calls

embedding_client.py already validates that returned embedding length matches vector_dimensions, but it doesn't pass dimensions= on the wire. With Alibaba Bailian's text-embedding-v4, default output is 1024-d but the deployment's pgvector schema typically uses 1536. Calling with dimensions=self.vector_dimensions lets the operator-configured EMBEDDING_VECTOR_DIMENSIONS actually take effect — OpenAI's own text-embedding-3-* accepts the same parameter, so this is portable.

Applied at all three OpenAI embedding call sites (single embed, batch embed, batch embed with text ids).

3. Default structured-output to json_object + schema-in-prompt

When a Pydantic class is passed as response_format, the OpenAI backend was first trying chat.completions.parse() (Structured Outputs, json_schema mode), then falling back to a json_schema payload via _create_structured_response. Both paths use json_schema, which DeepSeek-v4 family and several Bailian / vLLM models reject (This response_format type is unavailable now / Failed to deserialize the JSON body).

Default now goes straight to {"type": "json_object"} (universally supported) with the target Pydantic schema injected as a system-message instruction. The existing repair_response_model_json machinery already handles minor JSON shape drift, so the looser enforcement is acceptable in exchange for portability. Streaming path (stream) gets the same treatment.

Trade-off: on OpenAI itself, this gives up strict-mode json_schema in favour of json_object + prompted schema. In practice the parsed output is virtually identical because gpt-* models comply with schema descriptions reliably; the repair logic catches any drift.

Why these three together

They surfaced together while wiring honcho into a stack that uses DeepSeek as the dialectic / deriver LLM and Alibaba Bailian (text-embedding-v4) for embeddings. Each change is independent and small but they share the same motivation: make the OpenAI backend gracefully cover the long tail of OpenAI-compatible providers.

End-to-end verification

Tested against:

• LLM: DeepSeek v4-flash via https://api.deepseek.com/v1 with thinking={type:disabled} injected through provider_params
• Embeddings: Alibaba Bailian text-embedding-v4 via https://dashscope.aliyuncs.com/compatible-mode/v1, dimensions=1536

Full dialectic chat across all 5 reasoning levels (minimal / low / medium / high / max) returns synthesized answers; deriver creates observations end-to-end; dreamer specialists (deduction + induction) complete a full update_peer_card cycle.

Notes

• No new dependencies.
• No behaviour change on OpenAI's own endpoints other than (a) embeddings now explicitly request dimensions (matches OpenAI's documented param) and (b) structured outputs route through json_object instead of json_schema (still produces valid Pydantic-parseable JSON via repair logic).
• The submodule is wired into a downstream consumer (ZVin-Chen/emotional_agent#17) that exercises the stack end-to-end.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Bug Fixes

  • Fixed embedding request dimension consistency across all operation types to ensure proper vector generation.
  • Improved structured JSON output handling with better validation, error detection, and automatic correction mechanisms.

• Improvements

  • Enhanced provider parameter compatibility by expanding support for additional configuration options.