Why is the AI not responding?#

Check pending approvals, then target availability and the model connection. Try a simple request without attachments. Record the error type without exposing keys. Stop an abnormal turn if needed; this does not undo completed work.

Why can I see a project but not open its files?#

Identify the computer that owns the file. A cloud or mirrored path is not a local path. Check execution target, file existence and permissions.

Why is an uploaded source not searchable knowledge?#

Upload, parsing, draft review and approved knowledge are different stages. Check for parsing errors, pending drafts or the wrong knowledge base. Try short keywords and material IDs.

Why does the same account show different content?#

Check computer, workspace and conversation identity. Same-named projects on different computers do not merge automatically, and sync scopes may differ.

Why did a schedule not run?#

Check enabled status, time zone, trigger rule, target runtime, model and run history. Do not assume execution or catch-up while the computer sleeps or the app is closed.

Why is an Android approval or notification missing?#

Approval cards require a real pending request and an online target. Notifications additionally depend on system permissions and push services. Inspect the conversation itself first.

How do I report a reproducible problem?#

Application version and operating system:
Mode: desktop / cloud web / remote / Android
Target and workspace: non-sensitive aliases
Steps:
Expected result:
Actual result and error:
Can it be reproduced consistently?

Hide account details, keys, private paths and client content in screenshots. Include the error time and time zone for comparison with run records.

Context or output budget errors after connecting a source#

Check that the desktop executing the task is on rc.50 or later. Large external tool sets use on-demand discovery instead of loading every schema at once; this does not enlarge the model context window. Source activation resumes the task with its permissions and plan preserved; Stop cancels pending automatic continuation. If the error persists, reduce input or oversized tool results, or select a model with a larger context window. This error does not by itself indicate a network failure.

Reasoning exhausted the allowance without an answer#

This means the server reported length with reasoning but no text or tool calls. It does not establish that history or tools filled the context. Check whether recovery is in progress. If the single recovery fails or capabilities are unknown, verify output capacity, request limits and supported reasoning controls in model capabilities and budgets. Partial text is preserved; incomplete tool calls are not executed. Stopping does not undo earlier successful tool operations.

ECS web startup feels slow#

Distinguish the login page, execution-target connection and conversation loading. A first visit or upgrade downloads new UI assets; slower networks can take longer than a cached revisit. rc.52 limits startup icon imports and loads diagram rendering when a diagram is displayed, reserving its space while loading. Updated hosts return permission states with the conversation list instead of requiring a request for each conversation. Older hosts retain compatible reconciliation. Avoid repeated reloads or duplicate conversations. Use Retry after a connection failure; report only versions, elapsed time, conversation count and whether this was a first visit, without sending conversation contents.

Desktop model reports account connectivity loss or offline grace#

Starting in rc.53, a transient account heartbeat failure does not block online-model requests with a valid sign-in credential. After connectivity returns, retry the conversation or choose Settings → Account → Sync model policy. Background verification also retries, with delays of up to five minutes during a continuing outage. Expired credentials must be renewed online; revoked authorization or a disabled account requires signing in again or contacting an administrator. Local projects, knowledge and separately configured local models remain usable. Account connectivity errors differ from model response timeouts. If only one model becomes slow, report its name, error time and whether the conversation was long, without including conversation content or keys.