/** * SemanticStepExtractor - Extracts semantic steps from AI chunks. * * Semantic steps represent logical units of work within AI responses: * - thinking: Claude's reasoning process * - tool_call: Tool invocation * - tool_result: Tool execution result * - output: Text output from Claude * - subagent: Nested agent execution * - interruption: User interruption */ import { countContentTokens } from '@main/utils/tokenizer'; import type { AIChunk, EnhancedAIChunk, SemanticStep } from '@main/types'; /** * Extract semantic steps from AI chunk responses. * Semantic steps represent logical units of work within responses. * * Note: ALL tool calls are included, including Task tools with subagents. * Task tools are filtered in the renderer's buildDisplayItems, * but they are kept here for accurate context token tracking in aggregateToolOutputs. */export function extractSemanticStepsFromAIChunk(chunk: AIChunk | EnhancedAIChunk): SemanticStep[] { const steps: SemanticStep[] = []; let stepIdCounter = 0; // Note: Task tool calls are included in semantic steps for context token tracking. // The renderer's buildDisplayItems filters Task tools with subagents. // Process only AI responses (no user message in AIChunk) for (const msg of chunk.responses) { if (msg.type === 'assistant') { // Extract from content blocks const content = Array.isArray(msg.content) ? msg.content : []; for (const block of content) { if (block.type === 'thinking' && block.thinking) { // Calculate tokens for thinking content (output from Claude) const thinkingTokens = countContentTokens(block.thinking); steps.push({ id: `${msg.uuid}-thinking-${stepIdCounter++}`, type: 'thinking', startTime: new Date(msg.timestamp), durationMs: 0, // Estimated from token count content: { thinkingText: block.thinking, tokenCount: thinkingTokens, // Pre-computed token count }, tokens: { input: 0, output: thinkingTokens, // Thinking is output from Claude }, context: msg.agentId ? 'subagent' : 'main', agentId: msg.agentId, sourceMessageId: msg.uuid, }); } if (block.type === 'tool_use' && block.id && block.name) { // Include ALL tool calls in semantic steps, including Task tools with processes. // Task tools with processes are filtered from DISPLAY in the renderer's buildDisplayItems, // but they should be included here for accurate context token tracking. // The renderer's aggregateToolOutputs will correctly count Task tool tokens // as part of the main session's context consumption. // Calculate tool call tokens directly from name + input // This reflects what actually enters the context window const callTokens = countContentTokens(block.name + JSON.stringify(block.input)); steps.push({ id: block.id, type: 'tool_call', startTime: new Date(msg.timestamp), durationMs: 0, content: { toolName: block.name, toolInput: block.input, sourceModel: msg.model, }, tokens: { input: callTokens, output: 0, }, context: msg.agentId ? 'subagent' : 'main', agentId: msg.agentId, sourceMessageId: msg.uuid, }); } if (block.type === 'text' && block.text) { // Calculate tokens for text output (Claude's generated text) const textTokens = countContentTokens(block.text); steps.push({ id: `${msg.uuid}-output-${stepIdCounter++}`, type: 'output', startTime: new Date(msg.timestamp), durationMs: 0, content: { outputText: block.text, tokenCount: textTokens, // Pre-computed token count for consistency }, tokens: { input: 0, // Text output is generated by Claude, not input output: textTokens, }, context: msg.agentId ? 'subagent' : 'main', agentId: msg.agentId, sourceMessageId: msg.uuid, }); } } } // Tool results from internal user messages // Note: isMeta can be true or null in JSONL, so check for toolResults presence directly if (msg.type === 'user' && msg.toolResults && msg.toolResults.length > 0) { for (const result of msg.toolResults) { steps.push({ id: result.toolUseId, type: 'tool_result', startTime: new Date(msg.timestamp), durationMs: 0, content: { toolResultContent: typeof result.content === 'string' ? result.content : JSON.stringify(result.content), isError: result.isError, toolUseResult: msg.toolUseResult, // Enriched data from message tokenCount: countContentTokens(result.content), // Pre-computed token count }, context: msg.agentId ? 'subagent' : 'main', agentId: msg.agentId, }); } } // User interruption messages // These are user messages with array content containing text like "[Request interrupted by user]" if (msg.type === 'user' && Array.isArray(msg.content)) { let foundInterruption = false; for (const block of msg.content) { if (block.type === 'text' && block.text) { const textContent = block.text; // Check for interruption patterns if ( textContent.includes('[Request interrupted by user]') || textContent.includes('[Request interrupted by user for tool use]') ) { steps.push({ id: `${msg.uuid}-interruption-${stepIdCounter++}`, type: 'interruption', startTime: new Date(msg.timestamp), durationMs: 0, content: { interruptionText: textContent, }, context: msg.agentId ? 'subagent' : 'main', agentId: msg.agentId, }); foundInterruption = true; } } } // User-rejected tool use (toolUseResult field is "User rejected tool use") if (!foundInterruption && (msg.toolUseResult as unknown) === 'User rejected tool use') { steps.push({ id: `${msg.uuid}-interruption-${stepIdCounter++}`, type: 'interruption', startTime: new Date(msg.timestamp), durationMs: 0, content: { interruptionText: 'Request interrupted by user', }, context: msg.agentId ? 'subagent' : 'main', agentId: msg.agentId, }); } } } // Link processes as steps for (const process of chunk.processes) { steps.push({ id: process.id, type: 'subagent', startTime: process.startTime, endTime: process.endTime, durationMs: process.durationMs, content: { subagentId: process.id, subagentDescription: process.description, }, tokens: { input: process.metrics.inputTokens, output: process.metrics.outputTokens, cached: process.metrics.cacheReadTokens, }, isParallel: process.isParallel, context: 'subagent', agentId: process.id, }); } // Sort by startTime return steps.sort((a, b) => a.startTime.getTime() - b.startTime.getTime());}
/**
* SemanticStepExtractor - Extracts semantic steps from AI chunks.
*
* Semantic steps represent logical units of work within AI responses:
* - thinking: Claude's reasoning process
* - tool_call: Tool invocation
* - tool_result: Tool execution result
* - output: Text output from Claude
* - subagent: Nested agent execution
* - interruption: User interruption
*/
import { countContentTokens } from '@main/utils/tokenizer';
import type { AIChunk, EnhancedAIChunk, SemanticStep } from '@main/types';
/**
* Extract semantic steps from AI chunk responses.
* Semantic steps represent logical units of work within responses.
*
* Note: ALL tool calls are included, including Task tools with subagents.
* Task tools are filtered in the renderer's buildDisplayItems,
* but they are kept here for accurate context token tracking in aggregateToolOutputs.
*/
export function extractSemanticStepsFromAIChunk(chunk: AIChunk | EnhancedAIChunk): SemanticStep[] {
const steps: SemanticStep[] = [];
let stepIdCounter = 0;
// Note: Task tool calls are included in semantic steps for context token tracking.
// The renderer's buildDisplayItems filters Task tools with subagents.
// Process only AI responses (no user message in AIChunk)
for (const msg of chunk.responses) {
if (msg.type === 'assistant') {
// Extract from content blocks
const content = Array.isArray(msg.content) ? msg.content : [];
for (const block of content) {
if (block.type === 'thinking' && block.thinking) {
// Calculate tokens for thinking content (output from Claude)
const thinkingTokens = countContentTokens(block.thinking);
steps.push({
id: `${msg.uuid}-thinking-${stepIdCounter++}`,
type: 'thinking',
startTime: new Date(msg.timestamp),
durationMs: 0, // Estimated from token count
content: {
thinkingText: block.thinking,
tokenCount: thinkingTokens, // Pre-computed token count
},
tokens: {
input: 0,
output: thinkingTokens, // Thinking is output from Claude
},
context: msg.agentId ? 'subagent' : 'main',
agentId: msg.agentId,
sourceMessageId: msg.uuid,
});
}
if (block.type === 'tool_use' && block.id && block.name) {
// Include ALL tool calls in semantic steps, including Task tools with processes.
// Task tools with processes are filtered from DISPLAY in the renderer's buildDisplayItems,
// but they should be included here for accurate context token tracking.
// The renderer's aggregateToolOutputs will correctly count Task tool tokens
// as part of the main session's context consumption.
// Calculate tool call tokens directly from name + input
// This reflects what actually enters the context window
const callTokens = countContentTokens(block.name + JSON.stringify(block.input));
steps.push({
id: block.id,
type: 'tool_call',
startTime: new Date(msg.timestamp),
durationMs: 0,
content: {
toolName: block.name,
toolInput: block.input,
sourceModel: msg.model,
},
tokens: {
input: callTokens,
output: 0,
},
context: msg.agentId ? 'subagent' : 'main',
agentId: msg.agentId,
sourceMessageId: msg.uuid,
});
}
if (block.type === 'text' && block.text) {
// Calculate tokens for text output (Claude's generated text)
const textTokens = countContentTokens(block.text);
steps.push({
id: `${msg.uuid}-output-${stepIdCounter++}`,
type: 'output',
startTime: new Date(msg.timestamp),
durationMs: 0,
content: {
outputText: block.text,
tokenCount: textTokens, // Pre-computed token count for consistency
},
tokens: {
input: 0, // Text output is generated by Claude, not input
output: textTokens,
},
context: msg.agentId ? 'subagent' : 'main',
agentId: msg.agentId,
sourceMessageId: msg.uuid,
});
}
}
}
// Tool results from internal user messages
// Note: isMeta can be true or null in JSONL, so check for toolResults presence directly
if (msg.type === 'user' && msg.toolResults && msg.toolResults.length > 0) {
for (const result of msg.toolResults) {
steps.push({
id: result.toolUseId,
type: 'tool_result',
startTime: new Date(msg.timestamp),
durationMs: 0,
content: {
toolResultContent:
typeof result.content === 'string' ? result.content : JSON.stringify(result.content),
isError: result.isError,
toolUseResult: msg.toolUseResult, // Enriched data from message
tokenCount: countContentTokens(result.content), // Pre-computed token count
},
context: msg.agentId ? 'subagent' : 'main',
agentId: msg.agentId,
});
}
}
// User interruption messages
// These are user messages with array content containing text like "[Request interrupted by user]"
if (msg.type === 'user' && Array.isArray(msg.content)) {
let foundInterruption = false;
for (const block of msg.content) {
if (block.type === 'text' && block.text) {
const textContent = block.text;
// Check for interruption patterns
if (
textContent.includes('[Request interrupted by user]') ||
textContent.includes('[Request interrupted by user for tool use]')
) {
steps.push({
id: `${msg.uuid}-interruption-${stepIdCounter++}`,
type: 'interruption',
startTime: new Date(msg.timestamp),
durationMs: 0,
content: {
interruptionText: textContent,
},
context: msg.agentId ? 'subagent' : 'main',
agentId: msg.agentId,
});
foundInterruption = true;
}
}
}
// User-rejected tool use (toolUseResult field is "User rejected tool use")
if (!foundInterruption && (msg.toolUseResult as unknown) === 'User rejected tool use') {
steps.push({
id: `${msg.uuid}-interruption-${stepIdCounter++}`,
type: 'interruption',
startTime: new Date(msg.timestamp),
durationMs: 0,
content: {
interruptionText: 'Request interrupted by user',
},
context: msg.agentId ? 'subagent' : 'main',
agentId: msg.agentId,
});
}
}
}
// Link processes as steps
for (const process of chunk.processes) {
steps.push({
id: process.id,
type: 'subagent',
startTime: process.startTime,
endTime: process.endTime,
durationMs: process.durationMs,
content: {
subagentId: process.id,
subagentDescription: process.description,
},
tokens: {
input: process.metrics.inputTokens,
output: process.metrics.outputTokens,
cached: process.metrics.cacheReadTokens,
},
isParallel: process.isParallel,
context: 'subagent',
agentId: process.id,
});
}
// Sort by startTime
return steps.sort((a, b) => a.startTime.getTime() - b.startTime.getTime());
}