fix(analyzer): cap "Claude thinking" gaps so mid-turn suspensions aren't counted as thinking

CHANGELOG.md

@@ -5,6 +5,19 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [Unreleased]
+
+### Fixed
+
+- **Time breakdown: cap "Claude thinking" gaps so mid-turn suspensions aren't counted as thinking.**
+  Previously only the assistant-end→user gap was capped by `IDLE_THRESHOLD_MS`; the
+  user→assistant and tool_result→assistant gaps (both attributed to `claudeThink`) were
+  uncapped. So any long pause that landed *mid-turn* — an overnight gap after a tool result,
+  a credit stall, a remote-control handoff — was reported as hours of "Claude thinking."
+  These gaps are now capped at `THINK_CAP_MS` (10 min); the remainder is booked as `humanAway`.
+  On a real 16 h session with overnight gaps this moved ~9 h out of "thinking"
+  (11 h 34 m → 2 h 35 m) into away time, where it belongs.
+
 ## [1.0.0] - 2026-02-18
 
 ### Added

src/analyzer.test.ts

@@ -444,3 +444,37 @@ describe('analyzer: warmup cost', () => {
     expect(result.warmupCost.turnCount).toBe(1);
   });
 });
+
+describe('analyzer: thinking gaps are capped (mid-turn suspension ≠ thinking)', () => {
+  const CAP = 10 * 60 * 1000; // THINK_CAP_MS
+
+  it('caps a long tool_result→assistant gap: 10min think + remainder humanAway', () => {
+    const result = analyzeSession('s1', [
+      userMsg('2026-01-01T00:00:00Z'),
+      assistantMsg('2026-01-01T00:00:02Z', { toolUses: [{ id: 'tu1', name: 'Bash' }] }),
+      toolResultMsg('2026-01-01T00:00:12Z', ['tu1']),
+      assistantMsg('2026-01-01T02:00:12Z'), // 2h gap after the tool result (session suspended)
+    ]);
+    // 2s legit first-response think (user→assistant) + 10min cap on the suspended gap
+    expect(result.enhancedStats.claudeThink).toBe(2000 + CAP);
+    expect(result.enhancedStats.humanAway).toBe(2 * 60 * 60 * 1000 - CAP); // 1h50m booked as away
+  });
+
+  it('caps a long user→assistant first-response gap the same way', () => {
+    const result = analyzeSession('s1', [
+      userMsg('2026-01-01T00:00:00Z'),
+      assistantMsg('2026-01-01T03:00:00Z'), // 3h before first response (suspended)
+    ]);
+    expect(result.enhancedStats.claudeThink).toBe(CAP);
+    expect(result.enhancedStats.humanAway).toBe(3 * 60 * 60 * 1000 - CAP);
+  });
+
+  it('leaves a normal short think gap untouched (no spurious humanAway)', () => {
+    const result = analyzeSession('s1', [
+      userMsg('2026-01-01T00:00:00Z'),
+      assistantMsg('2026-01-01T00:05:00Z'), // 5min < cap
+    ]);
+    expect(result.enhancedStats.claudeThink).toBe(5 * 60 * 1000);
+    expect(result.enhancedStats.humanAway).toBe(0);
+  });
+});

src/analyzer.ts

@@ -6,6 +6,14 @@ import type {
 import { estimateCost } from './pricing.js';
 
 const IDLE_THRESHOLD_MS = 2 * 60 * 1000; // 2 minutes
+// A single uninterrupted model response effectively never exceeds this. A longer
+// gap before an assistant message means the session was SUSPENDED mid-turn
+// (overnight pause, credit stall, remote-control handoff) — not the model
+// "thinking" for hours. We cap the thinking slice at this and attribute the
+// remainder to humanAway, so long-pause sessions aren't reported as huge
+// "Claude thinking" time. (Only the assistant-end→user gap was capped before;
+// the user→assistant and tool_result→assistant gaps were not.)
+const THINK_CAP_MS = 10 * 60 * 1000; // 10 minutes
 const SUBAGENT_TOOLS = new Set(['Task', 'Agent']);
 
 export function analyzeSession(sessionId: string, messages: SessionMessage[]): SessionAnalysis {
@@ -176,6 +184,18 @@ function detectEnhancedPhases(messages: SessionMessage[]): EnhancedTimeSegment[]
     segments.push({ phase, startTime: start, endTime: end, durationMs: end - start, toolName });
   }
 
+  // Emit a "Claude thinking" (or planning) slice, capped at THINK_CAP_MS. Any
+  // excess is a mid-turn suspension, not thinking, so it's booked as humanAway.
+  function emitThink(start: number, end: number) {
+    const phase: EnhancedPhaseType = planModeActive ? 'planning' : 'claudeThink';
+    if (end - start <= THINK_CAP_MS) {
+      emit(phase, start, end);
+    } else {
+      emit(phase, start, start + THINK_CAP_MS);
+      emit('humanAway', start + THINK_CAP_MS, end);
+    }
+  }
+
   for (const msg of messages) {
     const ts = Date.parse(msg.timestamp);
     if (isNaN(ts)) continue;
@@ -225,14 +245,12 @@ function detectEnhancedPhases(messages: SessionMessage[]): EnhancedTimeSegment[]
 
       // Gap from last external user message → this assistant = Claude thinking (first response)
       if (lastExternalUserTs !== null) {
-        const phase: EnhancedPhaseType = planModeActive ? 'planning' : 'claudeThink';
-        emit(phase, lastExternalUserTs, ts);
+        emitThink(lastExternalUserTs, ts);
         lastExternalUserTs = null;
       }
       // Gap from last tool_result → this assistant = Claude thinking (mid-turn, processing results)
       else if (lastToolResultTs !== null) {
-        const phase: EnhancedPhaseType = planModeActive ? 'planning' : 'claudeThink';
-        emit(phase, lastToolResultTs, ts);
+        emitThink(lastToolResultTs, ts);
       }
 
       lastAssistantEndTs = ts;