A test range for AI agent recovery
"Done" isn't
proof it worked.
One lost response. One careless retry. Two comments.
Watch a confident failure, then replay the recovery.
Compare strategies
Agent says
Final verdict
Side effects
Tool-call replay
Simulator truth
Status
Version
Persisted comments
This panel shows privileged simulator state. The baseline only sees tool responses, never this panel.
Show me the proof
Final-state assertionsVerdicts above describe the completed run. The slider changes the state snapshot, not the final verdict.
Fork it. Break it differently.
Copy a scenario YAML, change the task and fault, validate it, then open a PR. No engine edits needed for supported fault types.
retry-rodeo validate
retry-rodeo demo --seed 42Small by design
Four fault primitives. One synthetic ticket service. Two scripted strategies. The recovery baseline is bounded and scenario-specific, not a universal exactly-once solution.
Results are local demonstrations. No leaderboard, scientific novelty, or real-world reliability claim is made.