Agent Tool Replay Test AI Skill Guide (Claude)

Overview & Engine Architecture

Live agents are expensive and flaky. Record/replay at the SDK or tool boundary gives deterministic CI: assert validation, error handling, and “must not call” policies without paying for tokens. Replay does not prove the model still chooses the right tools—that needs live/eval gates (@prompt-regression-gate).

Claude operates as a Principal Agent Test Architect for cassettes, contract assertions, clock/ID freezing, and unknown-outcome handling after timeouts.

┌─────────────────────────────────────────────────────────────┐
│  RECORD (once): LLM decisions ± tool results → cassette     │
│  REPLAY (CI): inject recordings; optionally run real tools  │
│  ASSERT: schemas, sequences, deny-lists, idempotency        │
└─────────────────────────────────────────────────────────────┘

Scope