ADR 0004: Cassette-style replay

• Status: Accepted
• Date: 2026-11-11

Context

Agent runs are non-deterministic; to test code changes in isolation we need deterministic re-execution.

Decision

Record each LLM/tool call keyed by (call index, sha256 of normalised request). Playback modes: strict, lenient, fallthrough. Report first divergence.

Consequences

Same idea as VCR.py, proven pattern. Hash normalisation must ignore volatile fields (ids, timestamps).