[codex] support openai extra request params

[Felix3322]

Summary

This PR adds configurable passthrough request fields for the OpenAI-compatible backend so Wisdom-Weasel can talk to provider-specific OpenAI-style APIs that require extra parameters beyond the current fixed request body.

This is related to the follow-up discussion in #4 about supporting vendor-specific controls such as disabling thinking / reasoning output or limiting CoT / reasoning budgets on OpenAI-compatible services.

User impact

Before this change, the OpenAI-compatible request body was effectively hard-coded to:

• model
• messages
• max_tokens
• temperature

That worked for basic OpenAI-compatible services, but it blocked or limited compatibility with providers that expect extra JSON fields or custom HTTP headers. In practice that means users could not cleanly pass options like:

• reasoning_effort
• provider-specific thinking controls
• chat_template_kwargs.enable_thinking
• thinking_budget
• custom routing / vendor headers

As a result, some providers would either ignore the intended behavior, return overly expensive / verbose reasoning, or require users to patch the source code for every backend variation.

Root cause

The OpenAI-compatible provider and the memory-compression request path both constructed request payloads from a fixed set of fields in C++ and had no general mechanism to:

1. read arbitrary nested config from weasel.yaml
2. serialize that config back into JSON
3. merge it into the outgoing request body
4. append custom headers without hard-coding them in source

Fix

This PR introduces a reusable config-to-JSON helper and wires it into both OpenAI-style HTTP call paths.

Code changes

• add WeaselServer/ConfigJsonUtils.h
  • escapes JSON strings safely
  • recursively serializes Rime config maps / lists / scalars into JSON
  • loads string-like header maps from config
• extend OpenAICompatibleProvider with:
  • llm/openai/extra_body
  • llm/openai/extra_headers
• extend MemoryCompressor with:
  • llm/memory/extra_body
  • llm/memory/extra_headers
• preserve default Content-Type and Authorization behavior unless those headers are explicitly overridden in config
• update README.md with concrete YAML examples for passing reasoning / thinking-related vendor parameters

Example config enabled by this PR

llm:
  openai:
    api_url: "https://your-provider.example/v1/chat/completions"
    api_key: "your-api-key"
    model: "your-model"
    max_tokens: 20
    temperature: "0.6"
    extra_body:
      reasoning_effort: "low"
      thinking:
        type: "disabled"
      chat_template_kwargs:
        enable_thinking: false
        thinking_budget: 0

Validation

I did not rely on CLion's project model for validation.

Checks performed:

• verified GitHub CLI auth and repo access via gh auth status
• verified the real MSVC toolchain environment via msvc-latest.bat x64
• confirmed cl.exe and MSBuild.exe are available in that environment
• re-ran IDE-level inspections on the modified C++ files after the changes

Full repository build was not completed locally because this workspace does not currently have BOOST_ROOT / env.bat configured for the project build script.

Notes

This PR is intentionally scoped to the OpenAI-compatible / memory-compressor passthrough parameter work discussed in #4 comments. It does not attempt to implement the larger llama.cpp constrained decoding / beam search request from the issue title.

[Felix3322]

已按最新 review comment 处理：

• OpenAI 兼容后端现在会默认优先请求 JSON 输出（内置 response_format: {"type":"json_object"}，若用户未在 extra_body.response_format 中自定义）
• prompt 同步改成只返回 {"candidates":[...]} 结构
• 解析端补了更稳健的 JSON 字符串与数组解析，能处理转义字符，也会先尝试解析结构化 candidates
• 若某些兼容后端不支持该 JSON 约束，且没有用户自定义 response_format，会自动回退到原来的纯文本请求方式，避免兼容性倒退

本地已在独立干净 worktree 用真实 MSVC 环境重新编过 WeaselServer 相关目标，当前通过。

[KagaJiankui]

感谢极快的commit，可以考虑改变JSON和纯文本方式的顺序，因为有些提供商会主动拒绝不符合其规定格式的response_format参数：

• 在没有配置response_format类参数时，默认使用纯文本提示词与解析方法
• 只有在显式配置response_format参数之后才启用约束解码参数

如果您认为实现有困难或者就是觉得我事太多，可以忽略这个建议ovo

[Felix3322]

感谢极快的commit，可以考虑改变JSON和纯文本方式的顺序，因为有些提供商会主动拒绝不符合其规定格式的response_format参数：

• 在没有配置response_format类参数时，默认使用纯文本提示词与解析方法

• 只有在显式配置response_format参数之后才启用约束解码参数

如果您认为实现有困难或者就是觉得我事太多，可以忽略这个建议ovo

你快速删除文本会出现性能问题吗？

[Felix3322]

感谢极快的commit，可以考虑改变JSON和纯文本方式的顺序，因为有些提供商会主动拒绝不符合其规定格式的response_format参数：

• 在没有配置response_format类参数时，默认使用纯文本提示词与解析方法

• 只有在显式配置response_format参数之后才启用约束解码参数

如果您认为实现有困难或者就是觉得我事太多，可以忽略这个建议ovo

如果不是个例的话，backspace后就不计算候选了吧

[KagaJiankui]

你快速删除文本会出现性能问题吗？

我遇到的性能问题是在快速输入文本时卡顿，尤其是输入“`”符号时如果当前键入未上屏下一次键入一定无法输入，快速删除文本时没有遇到任何性能问题。

[Felix3322]

你快速删除文本会出现性能问题吗？

我遇到的性能问题是在快速输入文本时卡顿，尤其是输入“`”符号时如果当前键入未上屏下一次键入一定无法输入，快速删除文本时没有遇到任何性能问题。

我输入`甚至没反应

[Felix3322]

可能是我cpu缩缸了 反正我这装了这个就卡

[Felix3322]

我鼠标都动不了了...神经

[KagaJiankui]

可能是我cpu缩缸了 反正我这装了这个就卡

我笔记本液金涂歪了cpu上不了3.0G但是平常都还好，也就是有时候不太跟手，然后用本地模型经常报boost::archive相关的ipc异常

[Felix3322]

可能是我cpu缩缸了 反正我这装了这个就卡

我笔记本液金涂歪了cpu上不了3.0G但是平常都还好，也就是有时候不太跟手，然后用本地模型经常报boost::archive相关的ipc异常

我刚换完硅脂还是上不了4G（锁在65w了 国内还不管我这死地方的守候

[Felix3322]

已按这条最新建议调整：

• 默认不再主动注入 response_format
• 在没有显式配置 llm/openai/extra_body/response_format 时，走纯文本提示词 + 纯文本解析
• 只有当用户明确配置了 response_format 时，才走结构化 JSON 输出路径
• 若显式 JSON 输出解析失败，仍会回退到纯文本再试一次，避免完全无候选
• README 也补充了说明：response_format 现在是显式 opt-in 能力

本地已在干净 worktree 中重新用真实 MSVC 环境编过 WeaselServer 相关目标，当前通过。

对应提交：887d363 (default openai output to plain text)

[Felix3322]

Superseded by #10.

PR #10 contains the current combined local working state and should be used as the active review target going forward. Closing this older split PR to avoid fragmented review.

[Felix3322]

Closed as superseded by #10.