| 1 | # Tool interruption and recovery |
| 2 | |
| 3 | Reasonix records execution facts so every provider tool call has an honest result. These facts describe uncertainty; they do not grant or revoke access to later tools. |
| 4 | |
| 5 | ## Outcome model |
| 6 | |
| 7 | | Outcome | Meaning | Later behavior | |
| 8 | | --- | --- | --- | |
| 9 | | `not_started` | The durable start boundary was never crossed. | The model may issue a new call. | |
| 10 | | completed or failed | The host received a reliable terminal result. | The result and execution metadata are retained. | |
| 11 | | interrupted after start | Cancellation happened after the tool body began. Side effects may exist. | The model is advised to inspect state. Tools remain available. | |
| 12 | | `unknown` | Recovery found a durable start with no reliable terminal result. | A paired unknown result is added once. Tools remain available. | |
| 13 | |
| 14 | A call-start record is persisted before entering the tool body. Cancellation stops process trees and waits for started calls to settle; calls that never start receive an explicit result. Stopping a local process does not prove that a remote side effect did not occur. |
| 15 | |
| 16 | Recovery never automatically replays a call, synthesizes success from current state, blocks the same arguments, or installs a global read-only barrier. The reconstructed context lists which calls did not start, completed, or remain unknown. It advises retrying read-only or idempotent work as needed and checking external state before repeating a side effect. Each unknown record is added once to durable context rather than dynamically changing the system prompt every round. |
| 17 | |
| 18 | ## Compatibility API |
| 19 | |
| 20 | Desktop and Serve retain `GET /tool-recovery`, `POST /tool-recovery`, `GetToolRecoveryForTab`, and `ResolveToolRecoveryForTab` for wire compatibility. Queries return immutable historical facts with `retired=true` and `retryEnabled=false`. Action endpoints return `tool_recovery_retired`; they cannot inspect hidden arguments, confirm, reject, replay, or resume an operation. Old recovery cards render as read-only history without action buttons. |
| 21 | |
| 22 | Current runs write `unknown` only as an execution fact. Legacy `user_confirmed`, decision fields, Auto Guard sidecars, and recovery-action state are decoded without rewriting their facts; current runs do not create them, and none can recreate a gate. The `recovery_model` and Auto Guard settings are accepted when reading old configuration but omitted when saving current configuration. |
| 23 | |
| 24 | There is no host-level exactly-once guarantee for an external service that offers no authoritative receipt or idempotency mechanism. The model and user must inspect that service before repeating an uncertain side effect. |
| 25 |