[Feature]: 建议去掉injecttool标志对ai模型返回处理流程的影响分支

[davidjohn2025]

Summary

在会话过程中如果大量使用缩写，比如下列对话
usr：执行这个程序看看
ai：我们不需要工具，直接回答即可。用户说“执行这个程序看看”，但没有指定程序。可能是测试？直接回答无法执行。但根据指示，不需要工具时直接回答。所以回复说明没有程序。没有提供具体的程序代码，无法执行。请提供需要运行的程序内容或指令。
user：xxx.py
此时ai会根据上下文返回tool_json块，但是因为usr会话并未包含tool工具的关键字，导致ai即使返回正确的tool_json，但是agent并不会执行

Problem to solve

建议把
auth-profiles-CGF9hdRi.js里的的
//if (!injectTools) return streamFn(model, modifiedContext, options);这句话注释掉
const originalStreamOrPromise = streamFn(model, modifiedContext, options);
const wrappedStream = createAssistantMessageEventStream();
const processEvents = async () => {
try {
这样就能在任何tool_json块出现的时候正常调用工具，不会中断会话

Proposed solution

建议把
auth-profiles-CGF9hdRi.js里的的
//if (!injectTools) return streamFn(model, modifiedContext, options);这句话注释掉
const originalStreamOrPromise = streamFn(model, modifiedContext, options);
const wrappedStream = createAssistantMessageEventStream();
const processEvents = async () => {
try {
这样就能在任何tool_json块出现的时候正常调用工具，不会中断会话

Alternatives considered

No response

Impact

影响会话连贯

Evidence/examples

No response

Additional information

No response

[davidjohn2025]

建议把wrapWithToolCalling代码改为
function wrapWithToolCalling(streamFn, api) {
return (model, context, options) => {
let processEvents;
let wrappedStream;
const messages = context.messages || [];
const lastMsg = messages[messages.length - 1];
if (lastMsg?.role === "toolResult") {
const tr = lastMsg;
let resultText = "";
if (Array.isArray(tr.content)) {
for (const part of tr.content) if (part.type === "text" && part.text) resultText += part.text;
}
const feedbackPrompt = Tool ${tr.toolName || "unknown"} returned: ${resultText}\nPlease answer the original question based on this tool result.;
const feedbackContext = Object.assign({}, context, {
messages: [{
role: "user",
content: feedbackPrompt
}],
tools: [],
systemPrompt: ""
});
console.log([WebStreamMiddleware] tool result feedback, len=${feedbackPrompt.length});
//return streamFn(model, feedbackContext, options);
//对于toolresult提交给AI的返回，这里仿造用户交互的处理，检查返回中是否有tool_json块而不是当作普通文本显示
const originalStreamOrPromise = streamFn(model, feedbackContext, options);
wrappedStream = createAssistantMessageEventStream();
processEvents = async () => {
try {
const originalStream = await Promise.resolve(originalStreamOrPromise);
let accumulatedText = "";
let toolCallEmitted = false;
for await (const event of originalStream) if (event.type === "done") {
const finalMsg = event.message;
if (finalMsg && Array.isArray(finalMsg.content)) {
for (const part of finalMsg.content) if (part.type === "text" && part.text) accumulatedText = part.text;
}
console.log([wrapWithToolCalling] 25 ${accumulatedText});
const toolCall = extractToolCall(accumulatedText);
if (toolCall) {
toolCallEmitted = true;
const toolCallPart = {
type: "toolCall",
id: web_tool_${Date.now()},
name: toolCall.tool,
arguments: toolCall.parameters
};
const toolMsg = {
role: "assistant",
content: [toolCallPart],
stopReason: "toolUse",
api: model.api,
provider: model.provider,
model: model.id,
usage: finalMsg?.usage ?? {
input: 0,
output: 0,
cacheRead: 0,
cacheWrite: 0,
totalTokens: 0,
cost: {
input: 0,
output: 0,
cacheRead: 0,
cacheWrite: 0,
total: 0
}
},
timestamp: Date.now()
};
wrappedStream.push({
type: "toolcall_start",
contentIndex: 0,
partial: toolMsg
});
wrappedStream.push({
type: "toolcall_end",
contentIndex: 0,
toolCall: toolCallPart,
partial: toolMsg
});
wrappedStream.push({
type: "done",
reason: "toolUse",
message: toolMsg
});
} else wrappedStream.push(event);
} else if (!toolCallEmitted) wrappedStream.push(event);
} catch (err) {
wrappedStream.push({
type: "error",
reason: "error",
error: {
role: "assistant",
content: [],
stopReason: "error",
errorMessage: err instanceof Error ? err.message : String(err),
api: model.api,
provider: model.provider,
model: model.id,
usage: {
input: 0,
output: 0,
cacheRead: 0,
cacheWrite: 0,
totalTokens: 0,
cost: {
input: 0,
output: 0,
cacheRead: 0,
cacheWrite: 0,
total: 0
}
},
timestamp: Date.now()
}
});
} finally {
wrappedStream.end();
}
};
}else{
let userMessage = "";
const lastUserMsg = [...messages].toReversed().find((m) => m.role === "user");
if (lastUserMsg) {
if (typeof lastUserMsg.content === "string") userMessage = lastUserMsg.content;
else if (Array.isArray(lastUserMsg.content)) userMessage = lastUserMsg.content.filter((p) => p.type === "text").map((p) => p.text).join("");
}
userMessage = stripInboundMeta(userMessage);
if (!userMessage) return streamFn(model, context, options);
const hasAgentTools = (context.tools?.length ?? 0) > 0;
const injectTools = shouldInjectToolPrompt(api) && hasAgentTools && needsToolInjection(userMessage);
console.log([wrapWithToolCalling] 25 hasAgentTools=${hasAgentTools} shouldInjectToolPrompt=${shouldInjectToolPrompt(api)} needsToolInjection=${needsToolInjection(userMessage)});
const prompt = injectTools ? getToolPrompt(api) + userMessage : userMessage;
console.log([WebStreamMiddleware] api=${api} injectTools=${injectTools} promptLen=${prompt.length} userMsgLen=${userMessage.length});
const modifiedContext = Object.assign({}, context, {
messages: [{
role: "user",
content: prompt
}],
tools: [],
systemPrompt: ""
});
//if (!injectTools) return streamFn(model, modifiedContext, options);
const originalStreamOrPromise = streamFn(model, modifiedContext, options);
wrappedStream = createAssistantMessageEventStream();
processEvents = async () => {
try {
const originalStream = await Promise.resolve(originalStreamOrPromise);
let accumulatedText = "";
let toolCallEmitted = false;
for await (const event of originalStream) if (event.type === "done") {
const finalMsg = event.message;
if (finalMsg && Array.isArray(finalMsg.content)) {
for (const part of finalMsg.content) if (part.type === "text" && part.text) accumulatedText = part.text;
}
console.log([wrapWithToolCalling] 25 ${accumulatedText});
const toolCall = extractToolCall(accumulatedText);
if (toolCall) {
toolCallEmitted = true;
const toolCallPart = {
type: "toolCall",
id: web_tool_${Date.now()},
name: toolCall.tool,
arguments: toolCall.parameters
};
const toolMsg = {
role: "assistant",
content: [toolCallPart],
stopReason: "toolUse",
api: model.api,
provider: model.provider,
model: model.id,
usage: finalMsg?.usage ?? {
input: 0,
output: 0,
cacheRead: 0,
cacheWrite: 0,
totalTokens: 0,
cost: {
input: 0,
output: 0,
cacheRead: 0,
cacheWrite: 0,
total: 0
}
},
timestamp: Date.now()
};
wrappedStream.push({
type: "toolcall_start",
contentIndex: 0,
partial: toolMsg
});
wrappedStream.push({
type: "toolcall_end",
contentIndex: 0,
toolCall: toolCallPart,
partial: toolMsg
});
wrappedStream.push({
type: "done",
reason: "toolUse",
message: toolMsg
});
} else wrappedStream.push(event);
} else if (!toolCallEmitted) wrappedStream.push(event);
} catch (err) {
wrappedStream.push({
type: "error",
reason: "error",
error: {
role: "assistant",
content: [],
stopReason: "error",
errorMessage: err instanceof Error ? err.message : String(err),
api: model.api,
provider: model.provider,
model: model.id,
usage: {
input: 0,
output: 0,
cacheRead: 0,
cacheWrite: 0,
totalTokens: 0,
cost: {
input: 0,
output: 0,
cacheRead: 0,
cacheWrite: 0,
total: 0
}
},
timestamp: Date.now()
}
});
} finally {
wrappedStream.end();
}
};
}
queueMicrotask(() => void processEvents());
return wrappedStream;
};
}

这样当使用“读取/tmp/feishu_report.txt发给飞书”这类的级联动作，或者usr提问后，tool把toolresult回给ai后，ai认为需要对结果进行二次加工而返回tool_json后，agent无法解析tool_json进而造成工作中断