Skip to content

fix(agent): fix inconsistent fc call record format in chained non-native function calling #515

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
zerob13 merged 2 commits into from
Jun 15, 2025
Merged

fix(agent): fix inconsistent fc call record format in chained non-native function calling #515

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
0c1ce9a
fix(agent): fix inconsistent fc call record format in chained non-nat…
neoragex2002 Jun 15, 2025
d6c168c
fix: fix coderabbit warnings
neoragex2002 Jun 15, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
231 changes: 141 additions & 90 deletions src/main/presenter/llmProviderPresenter/index.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -634,6 +634,7 @@ export class LLMProviderPresenter implements ILlmProviderPresenter {
const supportsFunctionCall = modelConfig?.functionCall || false

if (supportsFunctionCall) {
// Native Function Calling:
// Add original tool call message from assistant
const lastAssistantMsg = conversationMessages.findLast(
(m) => m.role === 'assistant'
Expand Down Expand Up @@ -674,88 +675,89 @@ export class LLMProviderPresenter implements ILlmProviderPresenter {
: JSON.stringify(toolResponse.content),
tool_call_id: toolCall.id
})

// Yield the 'end' event for ThreadPresenter
// ThreadPresenter needs this event to update the structured message state (DB/UI).
// Yield tool end event with response
yield {
type: 'response',
data: {
eventId,
tool_call: 'end',
tool_call_id: toolCall.id,
tool_call_response:
typeof toolResponse.content === 'string'
? toolResponse.content
: JSON.stringify(toolResponse.content), // Simplified content for UI
tool_call_name: toolCall.name,
tool_call_params: toolCall.arguments, // Original params
tool_call_server_name: toolDef.server.name,
tool_call_server_icons: toolDef.server.icons,
tool_call_server_description: toolDef.server.description,
tool_call_response_raw: toolResponse.rawData // Full raw data
}
}
} else {
// Non-native function calling: Append call and response differently
// Non-native FC: Add tool result to conversation history for next LLM turn.

// 1. Format tool execution result into prompt-defined text.
const formattedToolResultText = `<function_call>${JSON.stringify({ function_call_result: { name: toolCall.name, arguments: toolCall.arguments, response: toolResponse.content } })}</function_call>`

// 1. Append tool call info to the last assistant message
const lastAssistantMessage = conversationMessages.findLast(
(message) => message.role === 'assistant'
// 2. Add a role: 'assistant' message to conversationMessages (containing the result text).
// Find or create the last assistant message to append the result text
let lastAssistantMessage = conversationMessages.findLast(
(m) => m.role === 'assistant'
)
if (lastAssistantMessage) {
const toolCallInfo = `\n<function_call>
{
"function_call": ${JSON.stringify(
{
id: toolCall.id,
name: toolCall.name,
arguments: toolCall.arguments // Keep original args here
},
null,
2
)}
}
</function_call>\n`

if (lastAssistantMessage) {
// Append formatted result text to the existing assistant message's content
if (typeof lastAssistantMessage.content === 'string') {
lastAssistantMessage.content += toolCallInfo
lastAssistantMessage.content += formattedToolResultText + '\n'
} else if (Array.isArray(lastAssistantMessage.content)) {
// Find the last text part or add a new one
const lastTextPart = lastAssistantMessage.content.findLast(
(part) => part.type === 'text'
)
if (lastTextPart) {
lastTextPart.text += toolCallInfo
} else {
lastAssistantMessage.content.push({ type: 'text', text: toolCallInfo })
}
}
}

// 2. Create a user message containing the tool response
const toolResponseContent =
'以下是刚刚执行的工具调用响应，请根据响应内容更新你的回答：\n' +
JSON.stringify({
role: 'tool', // Indicate it's a tool response
content:
typeof toolResponse.content === 'string'
? toolResponse.content
: JSON.stringify(toolResponse.content), // Stringify complex content
tool_call_id: toolCall.id
})

// Append to last user message or create new one
const lastMessage = conversationMessages[conversationMessages.length - 1]
if (lastMessage && lastMessage.role === 'user') {
if (typeof lastMessage.content === 'string') {
lastMessage.content += '\n' + toolResponseContent
} else if (Array.isArray(lastMessage.content)) {
lastMessage.content.push({
lastAssistantMessage.content.push({
type: 'text',
text: toolResponseContent
text: formattedToolResultText + '\n'
})
} else {
// If content is undefined or null, set it as an array with the new text part
lastAssistantMessage.content = [
{ type: 'text', text: formattedToolResultText + '\n' }
]
}
} else {
// Create a new assistant message just for the tool result feedback
conversationMessages.push({
role: 'user',
content: toolResponseContent
role: 'assistant',
content: [{ type: 'text', text: formattedToolResultText + '\n' }] // Content should be an array for multi-part messages
})
lastAssistantMessage = conversationMessages[conversationMessages.length - 1] // Update lastAssistantMessage reference
}
}

// Yield tool end event with response
yield {
type: 'response',
data: {
eventId,
tool_call: 'end',
tool_call_id: toolCall.id,
tool_call_response: toolResponse.content, // Simplified content for UI
tool_call_name: toolCall.name,
tool_call_params: toolCall.arguments, // Original params
tool_call_server_name: toolDef.server.name,
tool_call_server_icons: toolDef.server.icons,
tool_call_server_description: toolDef.server.description,
tool_call_response_raw: toolResponse.rawData // Full raw data

// 3. Add a role: 'user' message to conversationMessages (containing prompt text).
const userPromptText = '以上是你刚执行的工具调用及其响应信息，已帮你插入，请仔细阅读工具响应，并继续你的回答。';
conversationMessages.push({
role: 'user',
content: [{ type: 'text', text: userPromptText }] // Content should be an array
});

// Yield tool end event for ThreadPresenter to save the result
// This event is separate from the messages added to conversationMessages.
// ThreadPresenter uses this to save the raw result into the structured Assistant message block in DB.
yield {
type: 'response', // Still a response event, but indicates tool execution ended
data: {
eventId,
tool_call: 'end', // Indicate tool execution ended
tool_call_id: toolCall.id,
tool_call_response: toolResponse.content, // Simplified content for UI/ThreadPresenter
tool_call_name: toolCall.name,
tool_call_params: toolCall.arguments, // Original params
tool_call_server_name: toolDef.server.name,
tool_call_server_icons: toolDef.server.icons,
tool_call_server_description: toolDef.server.description,
tool_call_response_raw: toolResponse.rawData // Full raw data for ThreadPresenter to store
}
}
}
} catch (toolError) {
Expand All @@ -768,30 +770,79 @@ export class LLMProviderPresenter implements ILlmProviderPresenter {
const errorMessage =
toolError instanceof Error ? toolError.message : String(toolError)

// Yield tool error event
yield {
type: 'response', // Still a response event, but indicates tool error
data: {
eventId,
tool_call: 'error',
tool_call_id: toolCall.id,
tool_call_name: toolCall.name,
tool_call_params: toolCall.arguments,
tool_call_response: errorMessage, // Error message as response
tool_call_server_name: toolDef.server.name,
tool_call_server_icons: toolDef.server.icons,
tool_call_server_description: toolDef.server.description
const supportsFunctionCallInAgent = modelConfig?.functionCall || false;
if (supportsFunctionCallInAgent) {
// Native FC Error Handling: Add role: 'tool' message with error
conversationMessages.push({
role: 'tool',
content: `The tool call with ID ${toolCall.id} and name ${toolCall.name} failed to execute: ${errorMessage}`,
tool_call_id: toolCall.id
});

// Yield the 'error' event for ThreadPresenter
yield {
type: 'response', // Still a response event, but indicates tool error
data: {
eventId,
tool_call: 'error', // Indicate tool execution error
tool_call_id: toolCall.id,
tool_call_name: toolCall.name,
tool_call_params: toolCall.arguments,
tool_call_response: errorMessage, // Error message as response
tool_call_server_name: toolDef.server.name,
tool_call_server_icons: toolDef.server.icons,
tool_call_server_description: toolDef.server.description
}
};
} else {
// Non-native FC Error Handling: Add error to Assistant content and add User prompt.

// 1. Construct error text
const formattedErrorText = `编号为 ${toolCall.id} 的工具 ${toolCall.name} 调用执行失败: ${errorMessage}`;

// 2. Add formattedErrorText to Assistant content
let lastAssistantMessage = conversationMessages.findLast(m => m.role === 'assistant');
if (lastAssistantMessage) {
if (typeof lastAssistantMessage.content === 'string') {
lastAssistantMessage.content += '\n' + formattedErrorText + '\n';
} else if (Array.isArray(lastAssistantMessage.content)) {
lastAssistantMessage.content.push({ type: 'text', text: '\n' + formattedErrorText + '\n' });
} else {
lastAssistantMessage.content = [{ type: 'text', text: '\n' + formattedErrorText + '\n' }];
}
} else {
conversationMessages.push({
role: 'assistant',
content: [{ type: 'text', text: formattedErrorText + '\n' }]
});
}
}

// Add error message to conversation history for the LLM
conversationMessages.push({
role: 'user', // Or 'tool' with error? Use user for now.
content: `Error executing tool ${toolCall.name}: ${errorMessage}`
})
// Decide if the loop should continue after a tool error.
// For now, let's assume it should try to continue if possible.
// needContinueConversation might need adjustment based on error type.
// 3. Add a role: 'user' message (prompt text)
const userPromptText = '以上是你刚调用的工具及其执行的错误信息，已帮你插入，请根据情况继续回答或重新尝试。';
conversationMessages.push({
role: 'user',
content: [{ type: 'text', text: userPromptText }]
});

// Yield the 'error' event for ThreadPresenter
yield {
type: 'response', // Still a response event, but indicates tool error
data: {
eventId,
tool_call: 'error', // Indicate tool execution error
tool_call_id: toolCall.id,
tool_call_name: toolCall.name,
tool_call_params: toolCall.arguments,
tool_call_response: errorMessage, // Error message as response
tool_call_server_name: toolDef.server.name,
tool_call_server_icons: toolDef.server.icons,
tool_call_server_description: toolDef.server.description
}
}
// Decide if the loop should continue after a tool error.
// For now, let's assume it should try to continue if possible.
// needContinueConversation might need adjustment based on error type.
}
}
} // End of tool execution loop

Expand Down
7 changes: 6 additions & 1 deletion src/main/presenter/llmProviderPresenter/providers/openAICompatibleProvider.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1114,7 +1114,12 @@ export class OpenAICompatibleProvider extends BaseLLMProvider {
}

// Generate a unique ID if not provided in the parsed content
const id = parsedCall.id || functionName || `${fallbackIdPrefix}-${index}-${Date.now()}`
const id =
parsedCall.id ??
(functionName
? `${functionName}-${index}-${Date.now()}`
: `${fallbackIdPrefix}-${index}-${Date.now()}`)

// console.log(
// `[parseFunctionCalls] Finalizing tool call for match ${index}: ID='${id}', Name='${functionName}', Args='${functionArgs}'`
// ) // Log final object details
Expand Down