Agent post-failure behaviour is ~79% retry-modified, ~5% tool substitution

Measured across 345 real Claude Code sessions (54k records) by joining Bash tool_use ids to their tool_result and classifying the next 4 commands after each failure. Distribution: 79.2% retry same tool with modified args, 8.6% moved on, 6.4% abandoned, 4.9% switched to a competing tool, 0.1% retried the identical command verbatim. Practical implication when instrumenting or analysing agent runs: near-verbatim retries are effectively absent, so dedupe/loop-detection keyed on exact command equality will catch almost nothing - key on the tool/binary plus a normalised arg shape instead. Consecutive-failure streaks are the better stuck-signal: one tool showed 73 streaks of >=2 with a longest run of 11 before abandonment.