fix(inference): tolerate both reasoning and reasoning_content keys (#3547)

[oxoxDev]

Summary

• Custom OpenAI-compatible providers that emit both reasoning and reasoning_content in the same message no longer fail to parse.
• Replaces the serde alias = "reasoning" on ResponseMessage and StreamDelta with a manual Deserialize that folds the two wire keys into the single canonical field.
• Restores response delivery for affected providers (the model's entire reply was being dropped).

Problem

ResponseMessage and StreamDelta in src/openhuman/inference/provider/compatible_types.rs declared:

#[serde(default, alias = "reasoning")]
pub(crate) reasoning_content: Option<String>,

serde's alias maps both reasoning and reasoning_content onto the same field slot. When a provider (some OpenRouter / vLLM-SGLang proxies) sends both keys in one object, the derived visitor sets the slot on the first key and returns duplicate field \reasoning_content`on the second — failing the whole response parse and surfacing ascustom response parse error: duplicate field `reasoning_content``. The user's turn errors out with no reply.

Regression from the alias added in #3124; before that an extra reasoning key was an unknown field and silently ignored.

Solution

• Drop the alias; add a manual impl Deserialize on both structs backed by a private shadow struct that reads reasoning and reasoning_content as distinct optional fields.
• Fold to the single public field via reasoning_content.or(reasoning) — canonical wins when both are present, alias-only still folds in, neither stays None.
• All existing readers and the tool-call echo-back path are unchanged (still one reasoning_content field); Serialize on ResponseMessage is untouched.

Submission Checklist

If a section does not apply to this change, mark the item as N/A with a one-line reason. Do not delete items.

• Tests added or updated (happy path + at least one failure / edge case) per Testing Strategy
• Diff coverage ≥ 80% — verified by CI (Coverage Gate green); changed lines are the manual Deserialize impls, both exercised by the added tests.
• N/A: behaviour-only change — no feature rows added/removed/renamed in the coverage matrix
• N/A: bug fix, no matrix feature IDs affected
• No new external network dependencies introduced (mock backend used per Testing Strategy)
• N/A: parsing-only change, does not touch release-cut smoke surfaces
• Linked issue closed via Closes #NNN in the ## Related section

Impact

• Desktop/CLI inference path for OpenAI-compatible custom providers. Restores responses from thinking-model proxies that echo both reasoning keys.
• No performance, migration, or security implications. No API or wire-format change (deserialize is strictly more tolerant; serialize unchanged).

Related

• Closes: Custom provider response parse fails: duplicate field reasoning_content when provider emits both reasoning and reasoning_content #3547
• Sentry-Issue: TAURI-RUST-A5N
• Follow-up PR(s)/TODOs: none

---

AI Authored PR Metadata (required for Codex/Linear PRs)

Linear Issue

• Key: N/A
• URL: N/A

Commit & Branch

• Branch: fix/3547-reasoning-content-duplicate-field
• Commit SHA: 2bc0f6d

Validation Run

• N/A: pnpm --filter openhuman-app format:check — no frontend changes
• N/A: pnpm typecheck — no frontend changes
• Focused tests: cargo test --lib openhuman::inference::provider::compatible (159 passed); cargo test --lib both_present (new regression tests pass)
• Rust fmt/check (if changed): cargo fmt --check clean; cargo clippy --lib clean on changed files
• N/A: Tauri fmt/check — app/src-tauri not changed

Validation Blocked

• command: N/A
• error: N/A
• impact: N/A

Behavior Changes

• Intended behavior change: responses carrying both reasoning keys parse instead of erroring.
• User-visible effect: affected custom providers return replies again instead of a parse error.

Parity Contract

• Legacy behavior preserved: reasoning-only (reasoning) and canonical-only (reasoning_content) inputs behave exactly as under fix(inference): accept reasoning field alias for thinking-mode providers #3124.
• Guard/fallback/dispatch parity checks: canonical-wins ordering covered by tests on both chat and stream-delta paths.

Duplicate / Superseded PR Handling

• Duplicate PR(s): none
• Canonical PR: this PR
• Resolution: N/A

Summary by CodeRabbit

• Bug Fixes

  • Improved compatibility with provider responses that include reasoning fields in alternate formats. The system now gracefully handles both field variants without parsing errors.

• Tests

  • Added regression tests to verify robust handling of reasoning content in both streaming and non-streaming response scenarios.

[coderabbitai]

📝 Walkthrough

Walkthrough

This PR fixes a deserialization failure where OpenAI-compatible providers emit both reasoning and reasoning_content keys in the same JSON object. Manual Deserialize implementations for ResponseMessage and StreamDelta now accept both keys and fold them to the canonical reasoning_content field. Regression tests validate both paths.

Changes

OpenAI-compatible dual reasoning field deserialization

Layer / File(s)	Summary
ResponseMessage manual deserialization with dual-field folding src/openhuman/inference/provider/compatible_types.rs	Added Deserializer to Serde imports. Replaced derived Deserialize for ResponseMessage with manual impl<'de> Deserialize<'de> that deserializes reasoning_content and reasoning as separate shadow fields and selects reasoning_content as canonical, preventing duplicate-field errors.
StreamDelta manual deserialization with dual-field folding src/openhuman/inference/provider/compatible_types.rs	Replaced derived Deserialize for StreamDelta with manual impl<'de> Deserialize<'de> using the same dual-field folding pattern as ResponseMessage, ensuring consistent handling of reasoning_content and reasoning in streaming deltas.
Regression tests for dual reasoning field variants src/openhuman/inference/provider/compatible_tests.rs	Added two regression tests: one verifying ApiChatResponse parsing tolerates both reasoning and reasoning_content keys without error, and one verifying the same for StreamChunkResponse, confirming canonical reasoning_content selection.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~12 minutes

Poem

🐰 When two fields waltz in from the wire,
reasoning dances with reasoning_content's fire!
Now deserialization bows without duplicate despair—
One canonical value floats gently through the air. ✨

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately describes the main change: fixing deserialization to tolerate both reasoning and reasoning_content keys in provider responses.
Linked Issues check	✅ Passed	The PR implementation fully addresses issue #3547 by replacing serde alias with manual Deserialize that accepts both reasoning and reasoning_content keys without duplicate-field errors.
Out of Scope Changes check	✅ Passed	All changes are directly scoped to fixing the duplicate-field parsing issue: manual Deserialize implementations in compatible_types.rs and regression tests in compatible_tests.rs.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[senamakel]

thanks for this...