fix(openrouter): preserve reasoning details

[dannote]

Description

This PR preserves OpenAI-compatible reasoning_details across ReqLLM's provider boundary.

OpenRouter can return reasoning_details as OpenAI-compatible wire maps with provider-specific metadata such as type, signatures, encrypted-signature markers, indexes, and other opaque fields. ReqLLM already has a normalized %ReqLLM.Message.ReasoningDetails{} struct with provider_data; this change makes that contract explicit and lossless across decode, context storage, and provider-boundary re-encoding.

This change:

• Decodes OpenAI-compatible reasoning_details into %ReqLLM.Message.ReasoningDetails{}.
• Preserves unknown/provider-specific fields in provider_data for lossless round-trips.
• Decodes both non-streaming response messages and streaming delta metadata.
• Keeps normalized reasoning_details as the internal representation by default.
• Adds an explicit encode_context_to_openai_format/3 option, encode_reasoning_details?: true, for callers that intentionally need OpenAI-compatible wire maps.
• Makes OpenRouter and MiniMax own their provider-specific reasoning-details re-encoding at the provider boundary.
• Filters provider-specific reasoning details so OpenRouter does not accidentally emit MiniMax/Anthropic/Google details, and MiniMax preserves its own wire shape.

#791 has landed. This branch has been rebased onto current agentjido/main as a single reasoning-details commit.

Type of Change

• Bug fix (non-breaking change fixing an issue)
• New feature (non-breaking change adding functionality)
• Breaking change (fix or feature causing existing functionality to change)
• Documentation update

Breaking Changes

None expected.

The internal representation becomes stricter and more consistent: OpenAI-compatible reasoning_details decoded by ReqLLM.Provider.Defaults are normalized to %ReqLLM.Message.ReasoningDetails{}. Existing raw maps are still tolerated in provider-specific encoders where they were already accepted.

Testing

• Tests pass (mix test)
• Quality checks pass (mix quality)

Ran focused tests:

mix test test/providers/openrouter_test.exs test/req_llm/provider/defaults_test.exs test/providers/minimax_test.exs test/provider/openai/responses_api_unit_test.exs test/provider/openai_structured_output_test.exs
Result: 306 passed

Ran full tests:

mix test
Result: 3523 passed, 11 skipped, 144 excluded

Ran project quality gate:

mix quality
Result: passed

GitHub CI is green on commit a18e86c0e5d8359309ee9e632936bc50cf4f157e.

Model Compatibility

No live model compatibility fixtures were recorded for this PR.

The change is local decode/encode behavior for OpenAI-compatible reasoning metadata and provider-boundary request rendering. It is covered by unit/provider tests for:

• default OpenAI-compatible response decoding
• streaming reasoning-details metadata decoding
• explicit OpenAI-compatible reasoning-details encoding
• OpenRouter response decode and request re-encode round-trip
• OpenRouter provider-data/signature preservation
• MiniMax provider-specific reasoning-details wire encoding
• provider filtering so unrelated providers' reasoning details are not emitted

Checklist

• My code follows the project's style guidelines
• I have updated the documentation accordingly
• I have added tests that prove my fix/feature works
• All new and existing tests pass
• My commits follow conventional commit format
• I have NOT edited CHANGELOG.md (it is auto-generated by git_ops)

Docs note: no docs were updated because this is an internal provider-boundary correctness fix for existing reasoning_details data, not a new public user-facing API.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a18e86c0e5

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve raw reasoning detail maps during encoding

When a caller reuses a legacy or persisted context whose reasoning_details are already raw OpenAI-compatible maps, the normal request path now enables encode_reasoning_details?: true and this catch-all filters every non-%ReasoningDetails{} entry out. That means those conversations silently stop sending the prior reasoning metadata on the next OpenAI-compatible request; before this change the raw maps were passed through unchanged.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Don't pre-normalize MiniMax stream details

For MiniMax streaming responses that include delta.reasoning_details, this now converts the raw MiniMax wire maps into generic %ReasoningDetails{} before ReqLLM.Providers.Minimax.decode_stream_event/2 pipes the chunks through its MiniMax-specific normalizer. That normalizer expects raw maps with string keys like "id" and "text"; given a struct, it treats it as a map and loses the text/signature fields, corrupting streamed MiniMax reasoning continuity.

Useful? React with 👍 / 👎.