Common Problems

The application remains on the startup screen

Quit the application completely, launch it once more, and preserve the first error. If it still fails, do not delete the application data directory or edit the database manually. Record the application version, operating-system version, and startup-stage message for further diagnosis.

Model requests fail

Check the following in order:

whether the API URL is a complete http or https address;
whether the Token and URL come from the same Provider;
whether the model identifier matches the provider's actual name;
whether the model is enabled;
whether the selected compatibility configuration matches the provider's protocol;
network connectivity, account balance, rate limits, and service status.

A complete configuration does not mean that the remote service has been verified. Do not put the complete Token in screenshots or any troubleshooting material you plan to share.

The Agent keeps waiting for approval

Open the relevant conversation and inspect the specific pending action. After approval or rejection, the Agent will attempt to continue the same Run. Notifications only provide reminders; they cannot substitute for approval. If the state still does not update, refresh the conversation and check again rather than creating a duplicate task.

An MCP Server cannot be enabled

Only local stdio is currently supported; HTTP URLs are not accepted.
Confirm that the executable and working directory still exist.
After saving the configuration, you must enable it separately and authorize launch in the confirmation dialog.
Changes to the program, arguments, working directory, or code entry point invalidate the previous authorization.
If the Catalog is incomplete, inspect the Server's Tool names and schemas rather than enabling automatic execution.

If the invocation result is unknown, do not retry automatically; the call may already have reached the Server.

A Skill does not appear or trigger

A Workspace Skill must be located at .agents/skills/<name>/SKILL.md with the exact capitalization.
Check name and description in the YAML front matter.
Confirm that the Skill is enabled and that the current task opened the correct project.
After making changes, test in a new task; a Run that has already started continues to use its frozen version.
An overly broad description can trigger incorrectly, while an overly narrow one may not trigger.

Browser automation is unavailable

Confirm that the built-in browser-automation capability is allowed in MCP settings and approve capability activation for the current task. Web permissions, sensitive sites, file uploads, downloads, and network targets may still require separate approval or be denied by security policy.

Sites that require a camera, microphone, location, or web notifications may not work because the managed browser denies these web permissions by default.

A Scheduled Automation did not run

Confirm that the application was running at the scheduled time.
Confirm that the task is Active and its health state is not Blocked.
Check that the model, project, target conversation, and permission mode remain available.
Check whether the Run is queued, running, or waiting for approval.
After fixing a Blocked configuration, save the task again; restoring an external resource does not clear Blocked automatically.

Scheduling is not a second-precision timer. If a Run is already active, the same task does not start another Run in parallel.

A native notification did not appear

Check whether the operating system allows Captain Who to send notifications, then review the actual Run and attention state on the Scheduled page. A notification is not proof that a task completed successfully. The current interface has no Toast fallback for system notifications. In rare cases, if the application exits after a notification is shown but before acknowledgement is saved, the notification may appear once more after restart.

An Office, PDF, or image task fails

Confirm that the file format is supported and, for a scanned PDF, whether it contains a text layer.
Confirm that the output directory is writable and that no other program has locked the input file.
Image generation requires a complete, enabled Provider configuration; local “ready” status does not verify the remote model and key.
A remote image request that times out may already have incurred a charge or produced an image; do not resubmit it immediately.

If the problem remains unresolved, use Diagnostics and Logs to prepare a minimal, redacted reproduction report.