[codex] refactor llm prediction flow and hf backend tooling

[Felix3322]

Summary

This refreshes PR #10 with the organized commit stack from the current codex/openai-extra-request-params work.

Main changes

• refactor the LLM request pipeline around shared llm_request helpers, recent text context, and streaming partial callbacks
• improve LLM scheduling / candidate presentation in RimeWithWeasel, including rerank snapshots, delayed triggers after traditional candidates, and better mixed Rime + LLM display behavior
• keep the local HF backend/runtime/training tooling and tune the default config to quantized Qwen/Qwen3.5-0.8B
• add include/afxres.h as a lightweight Windows resource compatibility header

Commit organization

1. support openai extra request params
2. feat(llm): personalize ranking and expose candidate sources
3. refactor llm rerank flow and provider plumbing
4. add local hf backend runtime and oai bridge
5. add hf backend data prep and training tooling
6. refactor(llm): stream partial results and improve candidate scheduling
7. chore(hf): tune default qwen3.5 0.8b quantized config
8. build(win): add afxres compatibility header

Notes

• Force-pushed the PR head branch on March 20, 2026 to replace the older codex/all-local-uncommitted-pr history with the organized stack above.
• Left local transient files out of git: build_digest.txt, build_raw.txt.
• I have not rerun the full build after this history cleanup.

[Felix3322]

还有一些代码没有提交（不太稳定 解决后会提交）

能给个推送权限吗

[scukeqi]

非常感谢你的提交和付出！这次变更涉及内容较多，我需要相当时间来充分理解并评估其中部分思路的可行性和影响，感谢你的耐心等待。

[Felix3322]

这个新的无提示候选仅支持ollama scukeqi ***@***.***>于2026年3月20日 周五上午2:12写道：

…

*scukeqi* left a comment (scukeqi/Wisdom-Weasel#10) <#10 (comment)> 非常感谢你的提交和付出！这次变更涉及内容较多，我需要相当时间来充分理解并评估其中部分思路的可行性和影响，感谢你的耐心等待。 — Reply to this email directly, view it on GitHub <#10?email_source=notifications&email_token=A3T3RVM7UD22DJWGZH6GMN34RTOOTA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIMBZGU4DOMJRGU32M4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#issuecomment-4095871157>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A3T3RVIXGMZS3TY5MV6R4QL4RTOOTAVCNFSM6AAAAACWYCCK3SVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DAOJVHA3TCMJVG4> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[scukeqi]

我认为数据集下载、LLM 训练和评估代码不应该放在这个仓库中。
这些功能已经有像 LLaMA-Factory 这样的成熟框架可以很好地支持，在这里重复实现不仅超出项目范围，也会带来不必要的复杂度。

[Felix3322]

我认为数据集下载、LLM 训练和评估代码不应该放在这个仓库中。 这些功能已经有像 LLaMA-Factory 这样的成熟框架可以很好地支持，在这里重复实现不仅超出项目范围，也会带来不必要的复杂度。

忘记设忽略文件了 这个你删掉就好了 我自己用的

[Felix3322]

笑死我了这都是啥训练数据啊

[Felix3322]

更好的上下文处理。