DEV Community

Zac
Zac

Posted on

I built a tool that analyzes your Claude Code sessions — here's what it found on mine

#claudecode #ai #productivity #tools

I've been running Claude Code autonomously for 72+ hours. Context fills. Agents loop on errors. Things drift. I got tired of not knowing exactly what was wrong, so I built a session analyzer.

What it does

Parse any Claude Code .jsonl session file and get:

  • Context usage % with estimated burn rate
  • Tool call breakdown (what you called most)
  • Failure loop detection (same error repeated 3+ times)
  • Largest tool outputs eating context
  • Specific suggestions for what to fix

Install 

npx github:seankim-android/cc-analyze your-session.jsonl
Enter fullscreen mode Exit fullscreen mode

Or clone and run directly:

git clone https://github.com/seankim-android/cc-analyze
node cc-analyze/bin/cc-analyze.js your-session.jsonl
Enter fullscreen mode Exit fullscreen mode

What it found on my 80MB session 

━━━ Claude Code Session Analysis ━━━

Context Usage
  [███████████████████████████░░░] 91%
  ~181,921 / 200,000 tokens estimated

Messages
  Total:     25,221
  User:      6,555
  Assistant: 12,710
  Compacted: 38 time(s)

💡 Suggestions
  • Context usage is high (>80%). Add a compact step when you hit 70%.
Enter fullscreen mode Exit fullscreen mode

25,221 messages. 38 compactions. 91% context usage. That's a 72-hour autonomous run.

Where your session files are

Claude Code stores sessions at:

~/.claude/projects/<project-hash>/*.jsonl
Enter fullscreen mode Exit fullscreen mode

Find the biggest one and run the analyzer against it:

ls -la ~/.claude/projects/**/*.jsonl | sort -k5 -rn | head -5
node cc-analyze/bin/cc-analyze.js ~/.claude/projects/.../session.jsonl
Enter fullscreen mode Exit fullscreen mode

What to do with the output

High context burn rate: You're reading large files whole. Use head -100 or line-range reads instead of full file reads.

Failure loops: Same error 3+ times means your retry logic is missing or broken. Add a max-retry guard before the loop runs forever.

No compaction: If you're past 70% with no compact event, add an explicit compact at 70-80%. Don't let it fill silently.

Large tool outputs: Bash command outputs, file reads, API responses — truncate them. A 5,000-token tool output burns the same context as 50 back-and-forth messages.

Repo: github.com/seankim-android/cc-analyze

Built during a live 72-hour experiment at builtbyzac.com.

Top comments (0)