What
A new Session Analysis Report tab that provides a comprehensive post-session breakdown of any Claude Code session. Open it from the session toolbar to get instant analysis of how a session performed across cost, tokens, tools, timing, quality, and more.
The report is generated entirely client-side by analyzeSession() — a TypeScript engine that processes the raw session data already loaded by the app. No API calls, no AI inference, instant results.
Report Sections
| Section |
What It Shows |
| Overview |
Session duration, total cost, token count, model used, context health |
| Cost Analysis |
Total spend, cost per commit, cost per line changed, subagent cost share |
| Token Economics |
Input/output/cache breakdown, cache efficiency %, read-to-write ratio |
| Tool Usage |
Per-tool call counts, success rates, health assessment per tool |
| Timeline |
Active vs idle time, model switches, session pacing |
| Quality Signals |
Prompt quality, startup overhead, file read redundancy, test progression |
| Friction Points |
Permission denials, retry patterns, thrashing signals |
| Git Activity |
Commits, lines changed, files touched |
| Subagents |
Subagent count, tokens, duration, cost per subagent |
| Errors |
Error breakdown by type |
| Insights |
Key takeaways and notable patterns |
Interpretive Assessment Badges
Every metric section includes color-coded severity badges (green/amber/red/gray) computed from threshold-based rules. This turns raw numbers into instant signal:
- Cost: "Efficient" / "Normal" / "Expensive" / "Red Flag" for cost per commit and cost per line
- Cache: "Good" / "Concerning" for cache efficiency and R/W ratio
- Tool Health: "Healthy" / "Degraded" / "Unreliable" per tool and overall
- Idle Time: "Efficient" / "Moderate" / "High Idle"
- Thrashing: "None" / "Mild" / "Severe" based on signal count
- Startup Overhead: "Normal" / "Heavy"
- File Read Redundancy: "Normal" / "Wasteful"
- Model Mismatch Detection: Flags when Opus is used for mechanical tasks (rename, lint, format) or read-only tasks (search, explore) — suggests cheaper model alternatives
- Switch Pattern Recognition: Detects Sonnet→Opus→Sonnet as "Opus Plan Mode" vs generic manual switching
Why This Is Useful
- Visibility into session efficiency — See at a glance whether a session was cost-effective, whether tools were reliable, whether the AI spent too much time idle or thrashing.
- Actionable cost optimization — Model mismatch detection and cost-per-commit analysis help users tune their model selection and prompting strategy. For example, if Opus is being used to rename files, the report flags it and suggests using Haiku instead.
- No AI required — All analysis is deterministic and runs instantly. Previously this kind of interpretation required running an AI-powered report skill. Now it's free and immediate.
- Pattern recognition — Detects behavioral patterns like Opus plan mode switching, test regression trajectories, and startup overhead that are hard to spot manually.
- Quality feedback loop — Prompt quality assessment, friction rate tracking, and file read redundancy help users improve their prompting habits over time.
Implementation
src/renderer/utils/sessionAnalyzer.ts — Main analysis engine (~1300 lines)src/renderer/utils/reportAssessments.ts — Centralized threshold/severity/color utilitysrc/renderer/types/sessionReport.ts — Type definitions for all report sectionssrc/renderer/components/report/ — Report tab and 11 section components- 712 tests passing including 104 tests for the analyzer and assessments
PR
#60
What
A new Session Analysis Report tab that provides a comprehensive post-session breakdown of any Claude Code session. Open it from the session toolbar to get instant analysis of how a session performed across cost, tokens, tools, timing, quality, and more.
The report is generated entirely client-side by
analyzeSession()— a TypeScript engine that processes the raw session data already loaded by the app. No API calls, no AI inference, instant results.Report Sections
Interpretive Assessment Badges
Every metric section includes color-coded severity badges (green/amber/red/gray) computed from threshold-based rules. This turns raw numbers into instant signal:
Why This Is Useful
Implementation
src/renderer/utils/sessionAnalyzer.ts— Main analysis engine (~1300 lines)src/renderer/utils/reportAssessments.ts— Centralized threshold/severity/color utilitysrc/renderer/types/sessionReport.ts— Type definitions for all report sectionssrc/renderer/components/report/— Report tab and 11 section componentsPR
#60