Short answer: After a consequential write times out, the safest state may be unknown. The provider could have completed the action while the response disappeared in transit.
Why timeout handling matters
Transport failure tells you the caller did not receive a usable response. It does not tell you whether the provider committed the external effect.
Safer recovery
Keep the original logical-operation identity, inspect durable operation state and reconcile provider truth where supported. If safety cannot be established, preserve uncertainty instead of guessing.
Claim boundary: Once does not claim universal exactly-once execution. Safety depends on stable operation identity, durable Once state, the provider integration and sufficiently authoritative provider truth. When the outcome cannot be established safely, Once can preserve uncertainty rather than assume another external execution is safe.
Use Once for this failure mode
Once is the execution-safety layer between autonomous software and supported consequential real-world side effects.
npx -y @once-agent/mcpOfficial MCP Registry identity: io.github.stringsofthemind-oss/once