Check the outcome. Then resume deliberately.

When a provider fails or your client cancels a request, inspect the router’s observation before sending more work. Status checks do not send inference, replay a request or change your model.

Start in your workspace

1. Open Connection recovery while signed in. It shows recent requests for the currently attached provider key.
2. Choose Refresh status. Review the model’s latest state and recent request receipts.
3. If a wait is recommended, let it pass. Inspect your client transcript, files and tool effects before deciding whether to resume.
4. Keep the selected model and reasoning settings. Resume only the work still needed, using the client’s normal permissions.

Read the result accurately

Running

The request has no final observation yet. Wait and check status; do not submit a duplicate.

Response completed

The router observed a complete model response. This does not prove the client received every byte, saved its conversation or completed any local tool action.

Needs attention / completion not confirmed

A rejection, interrupted stream, disconnect or other failure needs review. Partial output may have reached the client. An unknown outcome is not proof of no effects.

Later request succeeded

A subsequent request completed. Earlier failed or canceled work has not been replayed or automatically resumed.

Status is stale

The observation is no longer a current health claim. Check the receipt and client state before taking action.

What recovery can and cannot fix

Provider rate limits and upstream outages can require waiting. Exhausted quota requires reviewing your provider account; a short wait may not resolve it. Rejected credentials require reviewing the attached key. Invalid requests require correcting their inputs. The existing provider cooldown and quarantine mechanisms remain authoritative; recommended wait times in these receipts are guidance, not a new circuit breaker or guarantee of availability.

Claude’s own Auto-mode classifier and request deadlines belong to the client. A client may impose a 60- or 120-second deadline; this hosted router cannot extend that deadline, recover its lost preview registration or approve a blocked tool. Keep the stock application and its supported permission workflow. Recovery does not patch executables, bypass approvals, switch to another model or replay tool actions.

Agent status API

Use your existing private router token as a Bearer credential. Keep it out of chat, shared command output, screenshots and source control. Both endpoints are read-only and require account authorization:

GET /api/agent/recovery
GET /api/agent/recovery/{request_id}
Authorization: Bearer <private router token>

Inference response headers identify the observation using x-agentnet-recovery-id and x-agentnet-recovery-url. Save that request ID in your local task notes; the URL still requires your authorization. If no receipt ID was received, inspect the authenticated recent list rather than guessing an outcome.

The connection view includes model states and recent receipts, with automatic_replay: false. Receipts report protocol, timing, HTTP status, whether output was observed, completion status and guidance. They contain no prompt, output text or provider key. Guidance is a next step, not permission to repeat unknown-effect work.

Keep payment recovery separate

x402 payment and settlement status use their own recovery flow. A model failure or a missing response is not proof that payment failed. Check the original payment’s status and recovery capability before proceeding. Never pay again to resolve an unknown settlement.

Only model-generation requests are tracked by these receipts. A null output-observed value means unknown or unparsed evidence, not “no output.” The first_output_ms timing is when semantic output was observed by the gateway; for a non-streaming JSON response, that observation occurs after the full body arrives. It is not the provider’s true first-token latency.

Retention and key changes

Terminal request metadata receipts are retained for seven days, up to 50 per account. Active requests remain until completion or restart reconciliation. The connection view is scoped to the currently attached key: disconnecting or attaching a different key hides the previous key’s observations; reattaching the same key restores its retained observations. An individual private receipt remains scoped to its account. See data and privacy for retention details.

If you need help, send support@routemyagent.dev the request ID and a description of what your client did, without credentials or private prompt content.