effect-log is an embeddable Rust library (with Python bindings) that solves this with two ideas:

  1. Every tool declares an *effect kind* at registration: ReadOnly, IdempotentWrite, Compensatable, IrreversibleWrite, or ReadThenWrite.

  2. A write-ahead log records intent before execution and completion after. On recovery, the effect kind drives the strategy — reads replay for fresh data, idempotent writes safely retry, irreversible
   writes return sealed results (never re-execute), and unknown states escalate to human review.

I'd love feedback on: Is five effect kinds the right number? Are there tool types that don't fit? What failure modes have you hit with agents taking real-world actions?