[Bug] 模型降智问题，感觉比较严重的，无法调用工具

[pridelizihao]

版本 / Version

2.0.94

部署方式 / Deployment

自建服务器 / Self-hosted

问题描述 / What happened?

好像从windsurf里面调用出来的模型比较弱智，经常报上下文不足的啊，总是工具调用失败的啊，这个是我没部署好的问题吗？

复现步骤 / Steps to reproduce

<img width="971" height="158" alt="Image" src="https://github.com/user-attachments/assets/f317d64f-04dd-44f2-a986-f100809dc4b9" />

<img width="953" height="380" alt="Image" src="https://github.com/user-attachments/assets/2108bcf9-43d4-41c5-841c-f58e77e29073" />

日志 / Logs

模型 / Model

No response

环境 / Environment

No response

[pridelizihao]

[Javkonline]

好像从windsurf里面调用出来的模型比较弱智，经常报上下文不足的啊，总是工具调用失败的，感觉简单的任务也没办法做出来了

[dwgx]

感谢反馈。这个方向后来已经拆成几个可验证的问题，而不是简单归因为单个模型“变弱”。

当前处理原则：

• 先确认客户端、模型、route、工具列表和日志；
• 区分是请求里没有工具、工具被兼容层压缩/剥离、模型没有产出 tool call，还是 native bridge 路由边界问题；
• 不把 Read/Grep/Glob/WebSearch/WebFetch 这类实验性 native bridge 路径说成默认生产修复。

后续请按最新版本提供 Probe[...]、ToolRoute[...] 和 BridgeResult[...] 附近日志，这个 issue 会继续作为“降智/工具失败”的具体复现桶来跟进。

[dwgx]

v2.0.97 VPS smoke passed for Claude-family tool calls at the time, but this issue needs more specific evidence before it can be treated as fixed.

Please provide:

• exact model,
• client and route,
• tool names/schema count,
• nearby Probe[...] / ToolRoute[...] / BridgeResult[...] logs from a current release.

Some model families and large tool schemas can still behave differently through Cascade. A single successful Claude smoke does not prove every client/model/tool combination is fixed.

[dwgx]

这个问题和 issue 178 类似，工具调用失败 / 降智的根因分析：

已确认的 bug（刚修复）：

1. tool-emulation.js flush() buffer 清空 bug — 流结束时 _findClosingBrace() 扫描的是已清空的 this.buffer 而不是保存的 remaining，导致 inToolCode / inBareCall 状态下的最后一个 tool call 被当成纯文本丢弃。模型发出了工具调用但代理没解析出来，下游客户端看到的就是"降智"。
2. messages.js tool_result 顺序错误 — 工具结果排在用户消息后面而不是前面，破坏了 OpenAI 格式的 tool chain 要求。

upstream 限制：

• Windsurf 的 SectionOverrideConfig（issue 164 分析）在某些模式下会抑制工具定义
• 非 Claude 模型（GPT / GLM / Kimi）的工具调用稳定性显著低于 Claude family

建议：

1. 更新到最新版（commit f68486e）
2. 日常任务优先 claude-sonnet-4.6 / claude-haiku-4.5
3. 如果降智问题仍在，请提供 Dashboard 日志中的模型名称和 tools=emu 字段

[dwgx]

This class of "degraded model / context not connected / tool failure" is not one single bug.

One confirmed direction was prompt-emulated tool definitions becoming too large or unstable, so the project added better diagnostics and a gated native-bridge path. The important boundary after later releases is:

• Production-mature native bridge scope remains Bash / shell_command / run_command.
• Read/Grep/Glob/WebSearch/WebFetch stay in protocol-lab unless explicitly gated by API key, account, model, and tool allowlist.
• WINDSURFAPI_NATIVE_TOOL_BRIDGE=all_mapped is not a general production fix for every client.
• Native bridge executes in the remote Windsurf workspace; it is not local IDE/MCP/client tool execution.

If the problem still reproduces, please attach the exact model, client, tool list, and Probe[...] / ToolRoute[...] / BridgeResult[...] logs. Without those logs, this issue remains a broad reproduction bucket rather than a single actionable bug.

[dwgx]

v2.0.139 已发布。这里先同步当前边界，避免误导：\n\n- native bridge 不是本地 IDE 工具调用的通用修复；它会让 Cascade 在远端 Windsurf workspace 执行内置工具。\n- 生产默认仍只把 Bash / shell_command / run_command 这条命令路径视作成熟 canary。\n- Read / Grep / Glob / WebSearch / WebFetch 仍放在 protocol-lab，需要明确模型/账号/API key gate 后小流量抓 trace；不会在生产默认全量打开。\n\n这版主要修了 CI full shard、dashboard 分页、native bridge 文档边界和 response cache 字节上限。工具调用/降智问题后续继续按具体模型+客户端+工具 schema 复现推进，不会把 Read/WebFetch 逆向未完成的部分说成已修。

[dwgx]

Update from v2.0.141:

This remains open as a concrete reproduction bucket, but the generic root cause in #164 is now closed as upstream/documented.

What changed in v2.0.141:

• New ToolRoute[...] log line for tool-bearing requests: requested tools, effective tools after tool_choice, native mapped/unmapped tools, native decision reason, and preamble tier.
• /v1/responses now drops forced tool_choice when it points to an unbridged server-side tool that was stripped during translation.
• README now documents that native bridge is remote Windsurf workspace execution, not a general local IDE tool fix.

For any new reproduction here, please include:

• client name/version,
• exact model,
• request route (/v1/chat/completions, /v1/messages, or /v1/responses),
• tool names/schema count,
• the ToolRoute[...] and Probe[...] log lines around the failed request.