# Tool calls cut off by a crash are rerun when it is safe

> After a worker crash, calls that never started, read-only and idempotent calls run again, and MCP servers get a stable idempotency key.

Source: <https://everruns.com/docs/changelog/2026-10-13-tool-calls-survive-crashes/>

Date: 2026-10-13
Products: Platform, Cloud, Framework
Category: Improvement
Available: In 0.49.0
Pull request: everruns/everruns#4694
Pull request: everruns/everruns#4698
Pull request: everruns/everruns#4700

When a worker stops in the middle of a turn, the turn resumes on another worker, and fewer tool calls now end as “interrupted”:

* A call that had not started yet runs, even one that must run at most once, such as sending an email: such a call starts only after its start is recorded.
* A call to a tool marked read-only or idempotent runs again. This includes MCP tools that declare `readOnlyHint` or `idempotentHint`.
* Every MCP tool call carries an idempotency key that stays the same when the call runs again, so a server that honors it can return the first result instead of acting twice:

```json
{
  "method": "tools/call",
  "params": {
    "name": "send_email",
    "arguments": {},
    "_meta": {
      "io.everruns/idempotency-key": "9b6c0f3e-5d1a-5e4b-8f2a-0c7d4e6a1b23"
    }
  }
}
```

A short outage of the control plane also no longer stops a running turn, as long as it ends within the lease time.

See [Durable Execution](https://everruns.com/docs/explanation/durable-execution/).