cc-recovery

How does Claude Code recover from its own errors? Track retry, investigate, fix, and pivot patterns across all your sessions.

npx cc-recovery
Drop your ~/.claude/projects/ folder here
or click to select
99.0%
of errors are self-recovered — no human help needed
6,512 errors across 1,993 sessions
1,993
Sessions
6,512
Errors analyzed
99.0%
Self-recover
61.9%
Retries thrash
Recovery strategy
What Claude does immediately after an error
retry
55.2%
3,597
investigate
27.5%
1,792
fix
14.1%
918
pivot
2.1%
140
ask
1.0%
64
rollback
0.0%
1
61.9%
of retries thrash 3+ times on the same tool
Retrying without changing approach — the most common anti-pattern
Recovery by tool
How Claude recovers from each tool's errors
Bash
retry 78% | investigate 12% | fix 6%
Read
investigate 41% | retry 35% | fix 22%
Edit
investigate 71% | fix 20% | retry 8%
WebFetch
retry 58% | investigate 34% | fix 6%
Write
fix 55% | investigate 35% | retry 9%
Glob
retry 40% | fix 39% | investigate 20%
What this means
  • 99% self-recovery — Claude almost never asks for human help after an error
  • 55% retry first — the default instinct is to try again, especially for Bash (78%) and WebFetch (58%)
  • Edit errors → investigate (71%) — when Edit fails (old_string not found), Claude reads the file first before retrying
  • 62% thrashing rate — most retries loop 3+ times before switching strategy. The biggest efficiency gap
  • Rollback is almost never used (0.0%) — git revert/reset is the rarest recovery pattern
  • Only 1% ask for help — Claude prefers to fix errors autonomously, even when struggling