litellm/litellm-rust/crates/circuit-breaker/AGENTS.md

• fsm is a generic state machine runtime and knows nothing about breakers. breaker is a generic circuit breaker and knows nothing about Redis. Deciding which errors count as failures belongs to the consumer
• Machine::step stays pure. &self is config and time comes in on the event, never read inside the machine. The only clock read is Admission's drop, at the call-site boundary
• Transition rules are one match (state, event) per machine, with no macro DSL and no _ arm that would hide a new state from the compiler
• Shared counts a transition, bumps the generation and calls the observer only when the state's enum variant changes, matching Python bumping generation only when _state changes. Data inside a variant (failure counters, since) changes without a transition
• The observer runs under the lock so transitions are seen in generation order. Keep it to logging or a metric
• A disabled breaker is Option<Breaker>::None at the consumer, not a flag inside the machine. Config comes in already resolved, reading REDIS_CIRCUIT_BREAKER_* env is the consumer's job

Parity with Python RedisCircuitBreaker (litellm/caching/redis_cache.py)

• The rstest table matches_the_python_redis_circuit_breaker was checked case by case against the Python breaker. A behavior change here needs the same check, by driving the Python class with the same steps
• Divergences, each on purpose
  • monotonic Instant instead of time.time(), so a wall clock jump can't open or close the breaker
  • a probe Admission dropped without an outcome records a hard failure and reopens. Python's probe that dies on CancelledError (a BaseException) records nothing and leaves the breaker HALF_OPEN for good, and a dropped Rust future is common
  • _failure_count is dropped, it only fed a log line
  • no enabled flag, see above