Retries and external outcomes
An integration failure must be classified before it is retried. A request refused before effects, an accepted asynchronous action and a lost response after submission are different conditions. The recovery objective is to discover or complete the existing intended action while preserving its company, resource, source version and operation identity.
Choose the next action from evidence
Section titled “Choose the next action from evidence”| Observation | Interpretation | Next action |
|---|---|---|
| Capability/resource or identity precondition refused | The requested path is not admitted | Correct the authorized context or stale input; do not bypass the guard. |
| Version conflict | Another accepted change may have occurred | Reload and compare the current record/source before deciding again. |
| Pending outbox/connector command identity exists | Work is accepted but not yet a confirmed outcome | Inspect that command/lease/receipt; do not create an independent duplicate. |
POST timeout or submitted_unknown |
The provider may have accepted the action | Reconcile using the retained stable identity; do not assume failure. |
| Read projection is partial or stale | Current completeness is not established | Continue or repair the source read under its checkpoint contract. |
| Hash, company or contract mismatch | Returned evidence cannot safely bind to this action | Stop the affected path and diagnose the mismatch before progressing. |
An HTTP error code by itself cannot establish how far a multi-step workflow progressed. Identify the owner from the provider inventory and transaction reference before using a retry button or operational command.
Reconstruct one operation
Section titled “Reconstruct one operation”Start with the selected company and current authorized actor. Identify the source record/version/hash, the approved snapshot, request key and any external command or document reference. Check the local transition/event and last confirmed provider observation. Distinguish an attempt timestamp from a successful receipt. Use identifiers and enumerated errors in investigation artifacts, not raw provider payloads, mailbox contents or credentials.
For connector bookkeeping, AccountingConnectorClient.submit classifies transport
failure as unknown and does not automatically repeat the POST. The worker’s local
command and export states tell recovery whether to await, reconcile, fail or retry
an admitted step. A new idempotency key describes a new command; generating one is
not a safe way to recover an uncertain existing write.
For archive projection, incomplete coverage keeps the durable source sequence eligible for retry. Replaying that source page is different from replaying a financial write. For document routing, a known committed request can return via its checked RPC without another source-copy attempt. The attachment copy before that transaction still has its own failure/cleanup boundary.
Synthetic diagnostic examples
Section titled “Synthetic diagnostic examples”Unknown accounting result. Export A has a stable local key and a connector request timed out. Read the existing command/provider record under its mapping. Only the owning workflow can decide whether the evidence permits completion or another attempt. Do not treat a missing browser success message as proof that bookkeeping did not occur.
Partial source read. Mailbox B indexed metadata but reports original gaps. Preserve that gap, validate its attested binding and continue source synchronization under the existing cursor rule. Downloading a similarly named file from another message does not repair the original’s evidence.
Stale review. Candidate C changed after a human opened it. Retain the user’s intent, reload the current version and source hash, and ask for the operation’s review again where necessary. Do not change the expected version to the newest value without examining the changed content.
Escalation and proof
Section titled “Escalation and proof”Escalate when identity/value differs, current provider truth cannot be established, or the action requires a product decision. Recovery is complete only when the specific workflow’s outcome is verified. A healthy scheduler, cleared error or successful retry response is insufficient by itself. Preserve the concrete remaining limitation when full evidence is unavailable.
Use synthetic local clients for deterministic refusal, timeout and replay cases. Real provider actions require the current authorized target and workflow; Vi Målar is not a Fortnox write-test target. The delivery PR records executed tests and reused evidence separately. This troubleshooting reference does not authorize manual database rewrites, checkpoint deletion or repeating a real send.