Support OpenAI Reasoning Models in the Model Adapter

未关闭
#171 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
48/100
Issue 类型
功能
描述清晰度
基本清楚
活跃度
活跃
技术栈
typescript

调研方向

Start with packages/models/src/providers/openai.ts around lines 138–160 and trace how the adapter builds request parameters and reads model sampling profiles. Check how existing tests construct standard model requests, then add coverage for reasoning and standard models as the work breakdown specifies. Done means the request fields match each model family’s requirements and the listed typecheck, test, and lint commands pass.

由索引模型根据 Issue 内容生成。

描述

bug good first issue

Summary

Codra's OpenAI adapter in packages/models/src/providers/openai.ts unconditionally sends temperature and max_tokens to every OpenAI model. Modern OpenAI reasoning models including the o-series (o1, o3, o3-mini, o4-mini) and newer GPT-5/GPT-6 family models in reasoning mode reject these parameters and return HTTP 400 Bad Request, making it impossible to route Codra reviews to any OpenAI reasoning model.

This is the parent issue. See sub-issues for the specific implementation tasks.

Work Breakdown

  • Sub-issue: Add reasoning model detection helper and conditional request body (covers max_completion_tokens and role: "developer")
  • Sub-issue: Add reasoning_effort support via model sampling profile
  • Sub-issue: Tests for reasoning and standard model request construction

Affected File

packages/models/src/providers/openai.ts lines 138-160:

const bodyPayload: Record<string, unknown> = {
  model,
  messages: [
    { role: "system", content: prompts.system },
    { role: "user", content: prompts.user },
  ],
  temperature,                        // rejected by o1, o3, o3-mini, o4-mini
  max_tokens: effectiveCeiling,       // must be max_completion_tokens for reasoning models
  response_format: { type: "json_object" },
};

OpenAI Reasoning Model API Requirements

Model family temperature max_tokens max_completion_tokens reasoning_effort
gpt-4o, gpt-4o-mini accepted accepted - -
gpt-5, gpt-5-mini accepted accepted - -
o1, o1-mini rejected rejected required optional
o3, o3-mini rejected rejected required supported
o4-mini rejected rejected required supported
gpt-6 (reasoning mode) rejected rejected required supported

Reasoning models generate internal chain-of-thought tokens before producing output. max_completion_tokens controls the ceiling over both reasoning and output tokens combined.

Done When

  • o1, o3, o3-mini, o4-mini models receive max_completion_tokens instead of max_tokens
  • temperature is omitted for reasoning models
  • role: "developer" replaces role: "system" for reasoning model system prompts
  • reasoning_effort is populated from the model sampling profile when available
  • Detection logic is forward-compatible (prefix-based, not a hardcoded list of model names)
  • Standard models (gpt-4o, gpt-4o-mini, gpt-5, custom base URLs) continue working unchanged
  • Tests cover reasoning model request construction and standard model construction
  • npm run typecheck, npm test, and npm run lint all pass
主要语言
TypeScript
星标
125
派生
23
平均合并
2 天 13 小时
30 天内合并 PR
12

环境准备

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

devarshishimpi/codra 的其他 Issue

查看 devarshishimpi/codra 的全部 Issue

相似的 Issue

更多 TypeScript Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。