run-ai-employee

Run an AI Employee

This starts an existing AI employee working. The work runs in the cloud and takes minutes, so this skill is fire-and-handle: start it, hand the human the task ID, and let them check back. You do NOT sit in a polling loop.

Prerequisite: load operate-ai-employees first. It sets up the dozaldevs CLI and the key. You also need the employee's slug (its stable identifier, distinct from role_name, which is only the display name). If you don't know the slug, find the employee first (manage-ai-employees). A key needs Member level or above to run one.

Exact bodies and answers: dozaldevs run-dozal --help and ../operate-ai-employees/references/customer-api.md. Answers are short; --detailed gives every field. --detailed needs dozaldevs 0.5.0 or newer (an older CLI reads the flag as needing a value).

Trigger

An employee with no inputs needs no data; an empty body is sent for you:

dozaldevs run-dozal "$SLUG"

When it takes inputs, pass them by key, with shell variables in place of the values:

dozaldevs run-dozal "$SLUG" --data inputs.topic="$TOPIC"

• 202 gives a task_id and a status_url. Keep the task_id.
• A 200 with deduped: true means the same trigger was already received, so nothing new started.
• Missing required inputs come back as 422 MISSING_REQUIRED_INPUTS with the list in missing. Ask the human for those values and retry.
• 400 MISSING_CONNECTION means an app the employee needs isn't connected. 402 PAYMENT_REQUIRED means the account can't run work right now. Tell the human; a person fixes both.

Check first with a dry run

--dry_run true checks the trigger without starting anything. Use it to learn which inputs are required (the 422 lists them), or how many runs a list would start:

dozaldevs run-dozal "$SLUG" --dry_run true

One run per line

Some employees take a list as one input and start a run for each line. The 202 for those has a runs list (one task_id per line) and a failures list for any line that didn't start. Report each run and each failure to the human. A list over the limit is refused up front with 422 TOO_MANY_ITEMS, and a dry run returns a run_count so you can check first. How an employee gets set up this way is in the "One run per line" section of ../operate-ai-employees/references/customer-api.md.

dozaldevs run-dozal "$SLUG" --data inputs.addresses="$ADDRESSES"

What to tell the human (fire-and-handle)

Respond right away, do NOT wait for completion:

"Started! Your Weekly Recap Writer is on it (task a1b2c3d4). It usually takes a few minutes. Ask me 'is the recap done?' whenever and I'll check."

Then stop. The human controls when to check.

Checking later

When the human asks if it's done, load monitor-and-review-ai-employee-work and make one status call with the task_id. If they want to wait for it, that skill covers how.

Inputs

The keys are defined by the employee. If the human's request implies one ("run the recap for last week"), pass it: --data inputs.report_date="$REPORT_DATE". Unsure which inputs exist? Do a dry run first.