Files
qwen-code/packages/cli/src/utils/nonInteractiveHelpers.ts

625 lines
19 KiB
TypeScript

/**
* @license
* Copyright 2025 Qwen Team
* SPDX-License-Identifier: Apache-2.0
*/
import type {
Config,
ToolResultDisplay,
TaskResultDisplay,
OutputUpdateHandler,
ToolCallRequestInfo,
ToolCallResponseInfo,
SessionMetrics,
} from '@qwen-code/qwen-code-core';
import {
OutputFormat,
ToolErrorType,
getMCPServerStatus,
} from '@qwen-code/qwen-code-core';
import type { Part, PartListUnion } from '@google/genai';
import type {
CLIUserMessage,
Usage,
PermissionMode,
CLISystemMessage,
} from '../nonInteractive/types.js';
import { CommandService } from '../services/CommandService.js';
import { BuiltinCommandLoader } from '../services/BuiltinCommandLoader.js';
import type { JsonOutputAdapterInterface } from '../nonInteractive/io/BaseJsonOutputAdapter.js';
import { computeSessionStats } from '../ui/utils/computeStats.js';
/**
* Normalizes various part list formats into a consistent Part[] array.
*
* @param parts - Input parts in various formats (string, Part, Part[], or null)
* @returns Normalized array of Part objects
*/
export function normalizePartList(parts: PartListUnion | null): Part[] {
if (!parts) {
return [];
}
if (typeof parts === 'string') {
return [{ text: parts }];
}
if (Array.isArray(parts)) {
return parts.map((part) =>
typeof part === 'string' ? { text: part } : (part as Part),
);
}
return [parts as Part];
}
/**
* Extracts user message parts from a CLI protocol message.
*
* @param message - User message sourced from the CLI protocol layer
* @returns Extracted parts or null if the message lacks textual content
*/
export function extractPartsFromUserMessage(
message: CLIUserMessage | undefined,
): PartListUnion | null {
if (!message) {
return null;
}
const content = message.message?.content;
if (typeof content === 'string') {
return content;
}
if (Array.isArray(content)) {
const parts: Part[] = [];
for (const block of content) {
if (!block || typeof block !== 'object' || !('type' in block)) {
continue;
}
if (block.type === 'text' && 'text' in block && block.text) {
parts.push({ text: block.text });
} else {
parts.push({ text: JSON.stringify(block) });
}
}
return parts.length > 0 ? parts : null;
}
return null;
}
/**
* Extracts usage metadata from the Gemini client's debug responses.
*
* @param geminiClient - The Gemini client instance
* @returns Usage information or undefined if not available
*/
export function extractUsageFromGeminiClient(
geminiClient: unknown,
): Usage | undefined {
if (
!geminiClient ||
typeof geminiClient !== 'object' ||
typeof (geminiClient as { getChat?: unknown }).getChat !== 'function'
) {
return undefined;
}
try {
const chat = (geminiClient as { getChat: () => unknown }).getChat();
if (
!chat ||
typeof chat !== 'object' ||
typeof (chat as { getDebugResponses?: unknown }).getDebugResponses !==
'function'
) {
return undefined;
}
const responses = (
chat as {
getDebugResponses: () => Array<Record<string, unknown>>;
}
).getDebugResponses();
for (let i = responses.length - 1; i >= 0; i--) {
const metadata = responses[i]?.['usageMetadata'] as
| Record<string, unknown>
| undefined;
if (metadata) {
const promptTokens = metadata['promptTokenCount'];
const completionTokens = metadata['candidatesTokenCount'];
const totalTokens = metadata['totalTokenCount'];
const cachedTokens = metadata['cachedContentTokenCount'];
return {
input_tokens: typeof promptTokens === 'number' ? promptTokens : 0,
output_tokens:
typeof completionTokens === 'number' ? completionTokens : 0,
total_tokens:
typeof totalTokens === 'number' ? totalTokens : undefined,
cache_read_input_tokens:
typeof cachedTokens === 'number' ? cachedTokens : undefined,
};
}
}
} catch (error) {
console.debug('Failed to extract usage metadata:', error);
}
return undefined;
}
/**
* Computes Usage information from SessionMetrics using computeSessionStats.
* Aggregates token usage across all models in the session.
*
* @param metrics - Session metrics from uiTelemetryService
* @returns Usage object with token counts
*/
export function computeUsageFromMetrics(metrics: SessionMetrics): Usage {
const stats = computeSessionStats(metrics);
const { models } = metrics;
// Sum up output tokens (candidates) and total tokens across all models
const totalOutputTokens = Object.values(models).reduce(
(acc, model) => acc + model.tokens.candidates,
0,
);
const totalTokens = Object.values(models).reduce(
(acc, model) => acc + model.tokens.total,
0,
);
const usage: Usage = {
input_tokens: stats.totalPromptTokens,
output_tokens: totalOutputTokens,
cache_read_input_tokens: stats.totalCachedTokens,
};
// Only include total_tokens if it's greater than 0
if (totalTokens > 0) {
usage.total_tokens = totalTokens;
}
return usage;
}
/**
* Load slash command names using CommandService
*
* @param config - Config instance
* @returns Promise resolving to array of slash command names
*/
async function loadSlashCommandNames(config: Config): Promise<string[]> {
const controller = new AbortController();
try {
const service = await CommandService.create(
[new BuiltinCommandLoader(config)],
controller.signal,
);
const names = new Set<string>();
const commands = service.getCommands();
for (const command of commands) {
names.add(command.name);
}
return Array.from(names).sort();
} catch (error) {
if (config.getDebugMode()) {
console.error(
'[buildSystemMessage] Failed to load slash commands:',
error,
);
}
return [];
} finally {
controller.abort();
}
}
/**
* Build system message for SDK
*
* Constructs a system initialization message including tools, MCP servers,
* and model configuration. System messages are independent of the control
* system and are sent before every turn regardless of whether control
* system is available.
*
* Note: Control capabilities are NOT included in system messages. They
* are only included in the initialize control response, which is handled
* separately by SystemController.
*
* @param config - Config instance
* @param sessionId - Session identifier
* @param permissionMode - Current permission/approval mode
* @returns Promise resolving to CLISystemMessage
*/
export async function buildSystemMessage(
config: Config,
sessionId: string,
permissionMode: PermissionMode,
): Promise<CLISystemMessage> {
const toolRegistry = config.getToolRegistry();
const tools = toolRegistry ? toolRegistry.getAllToolNames() : [];
const mcpServers = config.getMcpServers();
const mcpServerList = mcpServers
? Object.keys(mcpServers).map((name) => ({
name,
status: getMCPServerStatus(name),
}))
: [];
// Load slash commands
const slashCommands = await loadSlashCommandNames(config);
// Load subagent names from config
let agentNames: string[] = [];
try {
const subagentManager = config.getSubagentManager();
const subagents = await subagentManager.listSubagents();
agentNames = subagents.map((subagent) => subagent.name);
} catch (error) {
if (config.getDebugMode()) {
console.error('[buildSystemMessage] Failed to load subagents:', error);
}
}
const systemMessage: CLISystemMessage = {
type: 'system',
subtype: 'init',
uuid: sessionId,
session_id: sessionId,
cwd: config.getTargetDir(),
tools,
mcp_servers: mcpServerList,
model: config.getModel(),
permissionMode,
slash_commands: slashCommands,
qwen_code_version: config.getCliVersion() || 'unknown',
agents: agentNames,
};
return systemMessage;
}
/**
* Creates an output update handler specifically for Task tool subagent execution.
* This handler monitors TaskResultDisplay updates and converts them to protocol messages
* using the unified adapter's subagent APIs. All emitted messages will have parent_tool_use_id set to
* the task tool's callId.
*
* @param config - Config instance for getting output format
* @param taskToolCallId - The task tool's callId to use as parent_tool_use_id for all subagent messages
* @param adapter - The unified adapter instance (JsonOutputAdapter or StreamJsonOutputAdapter)
* @returns An object containing the output update handler
*/
export function createTaskToolProgressHandler(
config: Config,
taskToolCallId: string,
adapter: JsonOutputAdapterInterface | undefined,
): {
handler: OutputUpdateHandler;
} {
// Track previous TaskResultDisplay states per tool call to detect changes
const previousTaskStates = new Map<string, TaskResultDisplay>();
// Track which tool call IDs have already emitted tool_use to prevent duplicates
const emittedToolUseIds = new Set<string>();
// Track which tool call IDs have already emitted tool_result to prevent duplicates
const emittedToolResultIds = new Set<string>();
/**
* Builds a ToolCallRequestInfo object from a tool call.
*
* @param toolCall - The tool call information
* @returns ToolCallRequestInfo object
*/
const buildRequest = (
toolCall: NonNullable<TaskResultDisplay['toolCalls']>[number],
): ToolCallRequestInfo => ({
callId: toolCall.callId,
name: toolCall.name,
args: toolCall.args || {},
isClientInitiated: true,
prompt_id: '',
response_id: undefined,
});
/**
* Builds a ToolCallResponseInfo object from a tool call.
*
* @param toolCall - The tool call information
* @returns ToolCallResponseInfo object
*/
const buildResponse = (
toolCall: NonNullable<TaskResultDisplay['toolCalls']>[number],
): ToolCallResponseInfo => ({
callId: toolCall.callId,
error:
toolCall.status === 'failed'
? new Error(toolCall.error || 'Tool execution failed')
: undefined,
errorType:
toolCall.status === 'failed' ? ToolErrorType.EXECUTION_FAILED : undefined,
resultDisplay: toolCall.resultDisplay,
responseParts: toolCall.responseParts || [],
});
/**
* Checks if a tool call has result content that should be emitted.
*
* @param toolCall - The tool call information
* @returns True if the tool call has result content to emit
*/
const hasResultContent = (
toolCall: NonNullable<TaskResultDisplay['toolCalls']>[number],
): boolean => {
// Check resultDisplay string
if (
typeof toolCall.resultDisplay === 'string' &&
toolCall.resultDisplay.trim().length > 0
) {
return true;
}
// Check responseParts - only check existence, don't parse for performance
if (toolCall.responseParts && toolCall.responseParts.length > 0) {
return true;
}
// Failed status should always emit result
return toolCall.status === 'failed';
};
/**
* Emits tool_use for a tool call if it hasn't been emitted yet.
*
* @param toolCall - The tool call information
* @param fallbackStatus - Optional fallback status if toolCall.status should be overridden
*/
const emitToolUseIfNeeded = (
toolCall: NonNullable<TaskResultDisplay['toolCalls']>[number],
fallbackStatus?: 'executing' | 'awaiting_approval',
): void => {
if (emittedToolUseIds.has(toolCall.callId)) {
return;
}
const toolCallToEmit: NonNullable<TaskResultDisplay['toolCalls']>[number] =
fallbackStatus
? {
...toolCall,
status: fallbackStatus,
}
: toolCall;
if (
toolCallToEmit.status === 'executing' ||
toolCallToEmit.status === 'awaiting_approval'
) {
if (adapter?.processSubagentToolCall) {
adapter.processSubagentToolCall(toolCallToEmit, taskToolCallId);
emittedToolUseIds.add(toolCall.callId);
}
}
};
/**
* Emits tool_result for a tool call if it hasn't been emitted yet and has content.
*
* @param toolCall - The tool call information
*/
const emitToolResultIfNeeded = (
toolCall: NonNullable<TaskResultDisplay['toolCalls']>[number],
): void => {
if (emittedToolResultIds.has(toolCall.callId)) {
return;
}
if (!hasResultContent(toolCall)) {
return;
}
// Mark as emitted even if we skip, to prevent duplicate emits
emittedToolResultIds.add(toolCall.callId);
if (adapter) {
const request = buildRequest(toolCall);
const response = buildResponse(toolCall);
// For subagent tool results, we need to pass parentToolUseId
// The adapter implementations accept an optional parentToolUseId parameter
if (
'emitToolResult' in adapter &&
typeof adapter.emitToolResult === 'function'
) {
adapter.emitToolResult(request, response, taskToolCallId);
} else {
adapter.emitToolResult(request, response);
}
}
};
/**
* Processes a tool call, ensuring tool_use and tool_result are emitted exactly once.
*
* @param toolCall - The tool call information
* @param previousCall - The previous state of the tool call (if any)
*/
const processToolCall = (
toolCall: NonNullable<TaskResultDisplay['toolCalls']>[number],
previousCall?: NonNullable<TaskResultDisplay['toolCalls']>[number],
): void => {
const isCompleted =
toolCall.status === 'success' || toolCall.status === 'failed';
const isExecuting =
toolCall.status === 'executing' ||
toolCall.status === 'awaiting_approval';
const wasExecuting =
previousCall &&
(previousCall.status === 'executing' ||
previousCall.status === 'awaiting_approval');
// Emit tool_use if needed
if (isExecuting) {
// Normal case: tool call is executing or awaiting approval
emitToolUseIfNeeded(toolCall);
} else if (isCompleted && !emittedToolUseIds.has(toolCall.callId)) {
// Edge case: tool call appeared with result already (shouldn't happen normally,
// but handle it gracefully by emitting tool_use with 'executing' status first)
emitToolUseIfNeeded(toolCall, 'executing');
} else if (wasExecuting && isCompleted) {
// Status changed from executing to completed - ensure tool_use was emitted
emitToolUseIfNeeded(toolCall, 'executing');
}
// Emit tool_result if tool call is completed
if (isCompleted) {
emitToolResultIfNeeded(toolCall);
}
};
const outputUpdateHandler = (
callId: string,
outputChunk: ToolResultDisplay,
) => {
// Only process TaskResultDisplay (Task tool updates)
if (
typeof outputChunk === 'object' &&
outputChunk !== null &&
'type' in outputChunk &&
outputChunk.type === 'task_execution'
) {
const taskDisplay = outputChunk as TaskResultDisplay;
const previous = previousTaskStates.get(callId);
// If no adapter, just track state (for non-JSON modes)
if (!adapter) {
previousTaskStates.set(callId, taskDisplay);
return;
}
// Only process if adapter supports subagent APIs
if (
!adapter.processSubagentToolCall ||
!adapter.emitSubagentErrorResult
) {
previousTaskStates.set(callId, taskDisplay);
return;
}
if (taskDisplay.toolCalls) {
if (!previous || !previous.toolCalls) {
// First time seeing tool calls - process all initial ones
for (const toolCall of taskDisplay.toolCalls) {
processToolCall(toolCall);
}
} else {
// Compare with previous state to find new/changed tool calls
for (const toolCall of taskDisplay.toolCalls) {
const previousCall = previous.toolCalls.find(
(tc) => tc.callId === toolCall.callId,
);
processToolCall(toolCall, previousCall);
}
}
}
// Handle task-level errors (status: 'failed', 'cancelled')
if (
taskDisplay.status === 'failed' ||
taskDisplay.status === 'cancelled'
) {
const previousStatus = previous?.status;
// Only emit error result if status changed to failed/cancelled
if (
previousStatus !== 'failed' &&
previousStatus !== 'cancelled' &&
previousStatus !== undefined
) {
const errorMessage =
taskDisplay.terminateReason ||
(taskDisplay.status === 'cancelled'
? 'Task was cancelled'
: 'Task execution failed');
// Use subagent adapter's emitSubagentErrorResult method
adapter.emitSubagentErrorResult(errorMessage, 0, taskToolCallId);
}
}
// Handle subagent initial message (prompt) in non-interactive mode with json/stream-json output
// Emit when this is the first update (previous is undefined) and task starts
if (
!previous &&
taskDisplay.taskPrompt &&
!config.isInteractive() &&
(config.getOutputFormat() === OutputFormat.JSON ||
config.getOutputFormat() === OutputFormat.STREAM_JSON)
) {
// Emit the user message with the correct parent_tool_use_id
adapter.emitUserMessage(
[{ text: taskDisplay.taskPrompt }],
taskToolCallId,
);
}
// Update previous state
previousTaskStates.set(callId, taskDisplay);
}
};
// No longer need to attach adapter to handler - task.ts uses TaskResultDisplay.message instead
return {
handler: outputUpdateHandler,
};
}
/**
* Converts function response parts to a string representation.
* Handles functionResponse parts specially by extracting their output content.
*
* @param parts - Array of Part objects to convert
* @returns String representation of the parts
*/
export function functionResponsePartsToString(parts: Part[]): string {
return parts
.map((part) => {
if ('functionResponse' in part) {
const content = part.functionResponse?.response?.['output'] ?? '';
return content;
}
return JSON.stringify(part);
})
.join('');
}
/**
* Extracts content from a tool call response for inclusion in tool_result blocks.
* Uses functionResponsePartsToString to properly handle functionResponse parts,
* which correctly extracts output content from functionResponse objects rather
* than simply concatenating text or JSON.stringify.
*
* @param response - Tool call response information
* @returns String content for the tool_result block, or undefined if no content available
*/
export function toolResultContent(
response: ToolCallResponseInfo,
): string | undefined {
if (
typeof response.resultDisplay === 'string' &&
response.resultDisplay.trim().length > 0
) {
return response.resultDisplay;
}
if (response.responseParts && response.responseParts.length > 0) {
// Always use functionResponsePartsToString to properly handle
// functionResponse parts that contain output content
return functionResponsePartsToString(response.responseParts);
}
if (response.error) {
return response.error.message;
}
return undefined;
}