Improve OpenAI compatibility for tool calls, Responses output, and errors

[robberphex]

Summary

• Normalize official OpenAI function tool_choice format for Chat Completions and Responses API.
• Return cleaner non-streaming /v1/responses objects by avoiding internal fields like instructions.
• Add OpenAI-style error fields: message, type, param, and code.
• Preserve reasoning summaries and final output in /v1/responses.
• Add regression tests for tool calling, compact Responses output, error shape, and reasoning summary behavior.

Tests

• uv run python -m unittest tests.test_routes tests.test_models tests.test_fast_mode

---

摘要

• 兼容 OpenAI 官方的 function tool_choice 格式，覆盖 Chat Completions 和 Responses API。
• 清理非流式 /v1/responses 返回，避免暴露 instructions 等内部字段。
• 补齐 OpenAI 风格错误字段：message、type、param、code。
• 保留 /v1/responses 中可展示的 reasoning summary 和最终输出。
• 增加回归测试，覆盖工具调用、Responses 输出裁剪、错误结构和 reasoning summary 行为。

测试

• uv run python -m unittest tests.test_routes tests.test_models tests.test_fast_mode