OpenAI tool-call compatibility issue in flare that kills agentic loops

未关闭
  • 哪里不对？
#32 1 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
缺陷
描述清晰度
基本清楚
活跃度
活跃
技术栈
python
领域
api, backend

调研方向

Start by locating the completion stream code that raises “model called unregistered tool” for newly generated calls; the issue names no source file or test. Compare that path with the accepted history case, then trace how tool calls and recoverable errors reach clients. Done means an unregistered generated call can be handled by the client and the turn can continue; add coverage for that recovery path.

由索引模型根据 Issue 内容生成。

描述

area:api bug triage:verify-for-closure

What happens: when the model emits a tool_call whose function name isn't in the request's tools schema (e.g. run_commands instead of run_command), flare aborts the completion with model called unregistered tool …. Clients are designed to receive that call, return a "no such tool" result, and let the model correct itself in the same turn — the Open WebUI maintainer confirmed its loop does exactly that. Flare terminates the stream first, so the recovery never runs and the whole turn/agent loop dies. Seen live in OWUI: the loop just stops on that error.

Evidence:

The error is flare's own: model called unregistered tool … sits in the engine binary at offsets ~1127283/1127296 (/usr/local/bin/flare).
Control: the same unregistered name placed in history (assistant tool_calls + tool-role error result) is accepted — HTTP 200, model recovers and calls run_command correctly. So the rejection fires only on newly generated calls.
Black-box forcing didn't work (the model refuses to emit invalid names when instructed; 30 samples at temp 1.5, all run_command), but the check is deterministic once a bad name is generated — a canned invalid generation feeds straight into it.
vLLM/SGLang pass emitted tool calls through verbatim; name routing is the client's contract. That's why this never appeared before.
Suggested fix: don't hard-fail on an emitted name that isn't in the schema — either pass the call through as-is (client dispatches or errors), or surface it as a recoverable error event/tool-call so the client's own recovery can run.

主要语言
Rust
星标
800
派生
106
平均合并
1 天 3 小时
30 天内合并 PR
100

环境准备

  • 提供 Dockerfile 或 Docker Compose 文件
  • 没有 Pull Request 模板
  • 阅读贡献指南

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。

NVIDIA/hyperflare-inference 的其他 Issue

查看 NVIDIA/hyperflare-inference 的全部 Issue

相似的 Issue

更多 Rust Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。