fix(cerebras): preserve thinking inline for models that reject reasoning_content

[aaalaniz]

fix(cerebras): preserve thinking inline for models that reject reasoning_content

Fixes #10406. Supersedes the Cerebras fix in #10417.

Problem

Cerebras rejects multi-turn requests that replay messages[].reasoning_content with
400 wrong_api_format. Once a session has one assistant turn with thinking, every
subsequent turn fails.

Before

cerebras-before.mov

After

after2.mp4

Fix

Two halves, both required.

Replay side — a declarative model can declare how its thinking is replayed:

thinking_preservation_format	Replayed as
content_prepend	inline in content, plain text
content_xml	inline in content, wrapped in <think> tags
reasoning_content	the separate reasoning_content field (OpenAI-compatible default)

Request side — ModelInfo.request_params carries per-model static request-body
params. Cerebras models send {"reasoning_format": "parsed"} so reasoning comes back in
a structured field.

request_params mirrors the existing ModelConfig.request_params: same type, same
semantics, just sourced from the model declaration instead of user config. It is
unconditional and untyped — no enabled flag, no coupling to effort levels, and not gated
on whether reasoning is on, since reasoning_format describes the shape of the response
rather than whether the model should reason. It is declared per-model, so it only reaches
models that support it. Reserved keys (model, messages, stream, stream_options)
are skipped so a declaration cannot clobber the streaming setup.

Changes

• goose-provider-types/src/base.rs — ThinkingPreservationFormat enum, and two
  optional ModelInfo fields (thinking_preservation_format, request_params).
  DeclarativeProviderConfig.models is already Vec<ModelInfo>, so declarative JSON feeds
  both without extra plumbing.
• goose-provider-types/src/formats/openai.rs —
  OpenAiFormatOptions.thinking_preservation_format and inline_reasoning_content. It
  runs after merge_split_tool_call_messages, which uses reasoning_content to identify
  messages split from the same assistant turn. is_reserved_request_param_key is now pub
  so the provider reuses it rather than duplicating the list.
• goose-providers/src/openai.rs — custom_models retains full ModelInfo instead of
  just names (fetch_supported_models maps back to names). stream() looks the model up
  once and uses both fields: it passes the format into OpenAiFormatOptions, and merges
  request_params into the payload after create_request_with_options returns.
  apply_ollama_options is the precedent for provider-level payload shaping.
• goose-providers/src/declarative/definitions/cerebras.json — per-model
  thinking_preservation_format and request_params; adds gemma-4-31b.
• Mechanical (~8 lines) — ModelInfo literals gain the two new fields and two
  OpenAiFormatOptions literals gain ..Default::default(), across databricks.rs,
  declarative_providers.rs, provider_registry.rs, and tests/agent.rs.
• goose-self-test.yaml — adds Phase 3C, a manual multi-turn thinking-preservation
  procedure (AGENTS.md:71). It self-skips unless CEREBRAS_API_KEY is set and the
  session is on a Cerebras model, so the automated coverage is the unit tests below.

cerebras.json deliberately does not set "preserves_thinking": true:
should_preserve_thinking_by_default already defaults it to true for
ProviderEngine::OpenAI, and cerebras is "engine": "openai". There is a test asserting
that.

Limitations

Cerebras models do not receive reasoning_effort, so the thinking-effort setting has no
effect on them. This is not a regression — is_openai_responses_model matches only o\d+
and gpt-5*, so gpt-oss-120b, zai-glm-4.7, and gemma-4-31b get no reasoning param on
main today either. Mapping thinking effort onto these models is follow-up work.

Testing

Verified end to end against a live Cerebras key on all three models: multi-turn sessions
with thinking enabled now work, which is the scenario in #10406.

Unit tests added:

• ModelInfo deserializes both new fields, and defaults them to None when absent.
• Each of the three formats produces the expected content and drops reasoning_content.
• Inlining runs after the split-tool-call merge: two split assistant/tool pairs still
  reunite into one assistant message with both tool calls, with the thinking inlined once.
• cerebras.json declares the expected format and reasoning_format: parsed per model,
  and still defaults preserves_thinking to true.
• request_params cannot clobber stream, stream_options, model, or messages.

cargo fmt --all -- --check, cargo build --workspace --all-targets,
cargo clippy --all-targets -- -D warnings, and
cargo test -p goose-provider-types -p goose-providers all pass.

[aaalaniz]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Bravo.

Reviewed commit: 07511d53eb

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".

[filipkujawa]

Thanks for the fast turnaround on the split!

One required item before merge: the Azure AI Foundry provider (#10622) landed on main after your base, and it constructs ModelInfo (azure_foundry.rs, model_info_for_deployment) and OpenAiFormatOptions (openai_compatible.rs) without the new fields, so the branch merges cleanly textually but won't compile. Please rebase onto main and add the two new None fields / a ..Default::default() at those two sites- should be quick.

LGTM after the rebase.

[filipkujawa]

#10774 (comment)

[aaalaniz]

Thanks for the fast turnaround on the split!

One required item before merge: the Azure AI Foundry provider (#10622) landed on main after your base, and it constructs ModelInfo (azure_foundry.rs, model_info_for_deployment) and OpenAiFormatOptions (openai_compatible.rs) without the new fields, so the branch merges cleanly textually but won't compile. Please rebase onto main and add the two new None fields / a ..Default::default() at those two sites- should be quick.

LGTM after the rebase.

addressed, thank you!

[jamesn-cerebras]

Confirmed on our side, tested with a Cerebras reasoning model (gpt-oss-120b), thanks for the implementation!