Minimize a reproducible failure without confusing a smaller trace with its root cause

An action replay executes recorded actions against a reset fixture. An observation replay feeds recorded tool responses to a policy, potentially without reproducing their real effects. A fresh policy rerun can make different choices even from a similar snapshot. Name which mode is used and preserve initial state, versions, seeds, time assumptions, and external-response handling. Stubbed responses are useful for diagnosis but may remove the very commitment or delay behavior that caused the real failure. Do not treat a replayed success under a weaker simulator as proof that the original environment was correct.

Keep replay inside controlled fixtures. Reissuing a state-changing production request can create another payment, message, or file mutation instead of merely observing history. Preserve action identifiers and the intended idempotency semantics when those are part of the bug. Verify a failure predicate on authoritative evidence, such as duplicate committed effects or a specific violated invariant, rather than any exception string. Classify invalid or incomplete replays as unresolved when they cannot assess the intended failure; converting them into ordinary passes can mislead minimization and causal interpretation.

Delta debugging systematically removes parts while retaining a chosen failure. Zeller and Hildebrandt's original work distinguishes minimality properties under that test contract; a one-deletion-minimal input need not be the globally smallest one. For agent traces, deletions can violate authentication, object creation, or causal ordering preconditions. Consider grouping dependent events or changing the fixture explicitly, then label that as a changed minimization space. Rewriting a tool response or amount can reveal another failure but is not the same operation as deleting an irrelevant read from an otherwise fixed action trace.

A minimized action sequence can demonstrate that two calls trigger duplicate effects without proving why the agent chose the retry or whether another independent defect also exists. Compare controlled counterfactuals and preserve the original trace alongside the reduction. The first visibly wrong output may be downstream of an earlier observation or state error, and a plausible blame explanation needs corroborating evidence. Classify environment, policy, grader, and instrumentation defects separately where the data permits it rather than assigning every failure to the last tool call.

Replay the original and reduced cases under the same declared boundary, verify decisive state effects, and test the allowed single deletions before claiming the chosen minimality. For intermittent failures, preserve a deterministic trigger where possible or use a repeated statistical protocol with uncertainty and stopping rules. Do not accept an occasional pass as proof that a removed event was necessary. Record transformation history, unresolved candidates, and which environmental inputs were fixed. A good regression retains the failure mechanism while remaining small enough to run often and inspect meaningfully.

The probability lab illustrates intermittent detection under an assumed model. Reset, explicitly set BOTH p=0.2 and n=10, predict at least one failure with probability about 0.8926, then Play or Step. Reset and re-set BOTH p=0.2 and n=20; predict about 0.9885. Edits retain t; playback highlights exact iid counts. There are no traces, deletions, stateful tools, or minimization algorithms. The calculation cannot validate a replay fixture or establish independence; it only makes the stipulated sampling risk of missing an intermittent failure explicit.