# Troubleshooting

Find the failing layer before changing the application.

Source: https://docs.minds.sh/docs/operations/troubleshooting



A request passes through several layers: account identity, organization ownership, instance selection, networking, authentication, service availability, and the operation itself. Diagnose them in that order when the symptom is ambiguous.

## Cannot create a Mind [#cannot-create-a-mind]

Confirm that the Organizations page can load the current account's organizations. An IAM lookup 404 is not an empty organization list. If creation returns 500, check for a partially provisioned instance before submitting again. Record the request time, region, name, and visible error for support.

## The workspace says an instance is required [#the-workspace-says-an-instance-is-required]

Select a ready Mind in the instance selector. Account settings can load without a Mind; memory and engine views cannot. If an instance was just created, wait for the real provisioning state rather than relying on the fact that the form was submitted.

## An API request is unauthorized [#an-api-request-is-unauthorized]

Check the target API and credential type. The capability-required daemon expects `x-akasha-capability`; the SDK's Bearer token does not replace it. An authenticated request can still lack the required namespace, keyspace, or action. Some host-level operations require operator authority unavailable to an ordinary customer token.

## A route returns 404 or unavailable [#a-route-returns-404-or-unavailable]

Confirm the registered method and path, the correct base URL, and service configuration. The current SDKs include declarations that do not match all registered daemon routes. Use the [compatibility guide](/docs/developers/compatibility) and direct endpoint reference to isolate the mismatch.

## A wizard passes but the assistant cannot remember [#a-wizard-passes-but-the-assistant-cannot-remember]

Test a write and recall through the actual instance endpoint. Then repeat through the assistant, restart the agent process, and recall again. The initial setup wizard simulated verification; its repaired configuration step now identifies unavailable key minting. The audited standalone MCP backend still uses local in-memory storage, so its successful tool call does not establish a durable hosted connection.

## A setting or operation says it succeeded but nothing changed [#a-setting-or-operation-says-it-succeeded-but-nothing-changed]

Reload and read the state from its authoritative service. Known examples include swallowed settings errors, UI-only snapshot/log handlers, and acknowledgment-only checkpoint handlers. Report the action and the fresh read result separately.

## Send a useful report [#send-a-useful-report]

Include the UI page or API method/path, exact error, time, instance and organization identifiers, expected result, and whether the issue survives a reload. Include a screenshot when it explains the problem. Redact passwords, capabilities, tokens, and private record contents.