fix(openai): preserve reasoning for local OpenAI-compatible models via R1 toggle by ch405canova-sudo · Pull Request #1119 · Zoo-Code-Org/Zoo-Code

ch405canova-sudo · 2026-08-04T09:09:02Z

Summary

Local OpenAI-compatible reasoning models (llama.cpp llama-server, LM Studio, Ollama's OpenAI endpoint, vLLM, etc.) stream a reasoning_content field, but Zoo Code strips it from the follow-up context for these providers — so the model can never see its own reasoning chain on the next turn.

Root cause

openAiR1FormatEnabled (exposed as a UI checkbox via R1FormatSetting and declared in provider-settings.ts) only forces the R1 request format in openai.ts. It never propagates to info.preserveReasoning, which is what gates reasoning retention in Task.ts:

// src/core/task/Task.ts
const shouldPreserveForApi = this.api.getModel().info.preserveReasoning === true

OpenAiHandler.getModel() builds its ModelInfo from openAiCustomModelInfo ?? openAiModelInfoSaneDefaults — neither sets preserveReasoning. For the built-in OpenAI-compatible provider the value is therefore always undefined, and reasoning is stripped from messages sent back to the API.

Fix

When the user enables the R1 format toggle, getModel() now sets preserveReasoning: true on the returned model info:

return {
    id,
    info: this.options.openAiR1FormatEnabled ? { ...info, preserveReasoning: true } : info,
    ...params,
}

This preserves reasoning_content in the assistant history sent to the model on follow-up turns. Default behaviour (toggle off) is unchanged.

Tests

Added two getModel cases to src/api/providers/__tests__/openai.spec.ts:

preserveReasoning is true when openAiR1FormatEnabled is on
preserveReasoning stays undefined by default

…a R1 toggle The openAiR1FormatEnabled toggle (UI + provider-settings) only forced the R1 request format, but getModel() never set info.preserveReasoning. As a result Task.ts (shouldPreserveForApi = info.preserveReasoning === true) stripped reasoning_content from follow-up context for every local OpenAI-compatible reasoning model (llama.cpp, LM Studio, Ollama) — there was no way to feed the chain back. Enable the toggle and getModel() now sets preserveReasoning: true so the reasoning chain is preserved in the next-turn context. Default behaviour unchanged.

coderabbitai · 2026-08-04T09:09:50Z

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Advanced

Run ID: 47647a28-5868-4e9a-a101-5d375d0e186a

📥 Commits

Reviewing files that changed from the base of the PR and between f4a69cc and 60b1b45.

📒 Files selected for processing (2)

src/api/providers/__tests__/openai.spec.ts
src/api/providers/openai.ts

Included review availability: Your plan provides up to 4 included reviews per hour; 3 remain after this review.

📜 Recent review details

⏰ Context from checks skipped due to timeout. (4)

GitHub Check: e2e-mock
GitHub Check: compile
GitHub Check: platform-unit-test (ubuntu-latest)
GitHub Check: platform-unit-test (windows-latest)

🧰 Additional context used

📓 Path-based instructions (5)

Treat model, provider, MCP, path, command, and tool data as untrusted.