Lesson 5 · 18 min
With the bot running, your paper queue becomes messages. Each line is a skill with an input; the bot puts each request in the right queue, runs it within that skill’s limits, and lets you know in the chat when it’s done.
Start lesson
Goal: send 5 to 10 tasks in skill format and track the batch with /fila and /jobs.
The path each batch request takes.
Scroll down: each step below lights up its part of the diagram.
One line per task. The name before the colon is the skill; the input comes after it; optional fields go after a vertical bar.
The bot looks up the skill in the catalog. It lists which queue the skill runs in, the maximum deadline, and how many attempts it gets. A skill that isn’t in the catalog is rejected right away.
Each queue has its own limit. All ten video requests can be added, but the render queue runs one at a time because they share the graphics card.
The worker builds the skill prompt with your input and calls the agent. It works on its own, without asking for confirmation, and ends with RESULT or ERRO.
When it finishes, the bot sends the result or the reason for failure in the chat. In the morning, the conversation is your report.
You send the whole list at once. The bot distributes tasks across queues, respects the limits, and replies to each one when it’s done.
Click each block. Queue and limit numbers come from the inemaccbot documentation.
<skill>: <input> [| field]. The input is what you put in the "Input" column of your queue on paper.
/skills lists the installed skills, and /ajuda <skill> shows each skill’s fields. A field written without the bar is rejected; it won’t silently become part of the input.
explicativo: what is a durable queue | vertical
| modelo=haiku and | esforco=high change the model and reasoning for that request only. | motor=codex runs that request through codex instead of claude.
Use these to test one request. The effective model is recorded in the job, so in the morning you know what was used for each result.
/fila shows, for each queue, what’s running, how many are waiting, their age, errors from the last 24 hours, and retries. /jobs lists jobs that are running and those that are finished.
/status j13 gives you details about a job. A number alone always means a job; a letter and # (A#9) means a flow, covered in the next lesson.
/furar j13 moves a pending job to the front of the queue. /cancelar j13 stops a pending or running job.
Canceling applies to the queue, not the world: anything already created outside it (a delivered file, a published video) stays there.
If a job fails with code 143, it wasn’t the agent: the service restarted in the middle (143 = 128 + 15, the shutdown signal). The job goes back in the queue if it has attempts left; otherwise, it fails.
A render takes 10 to 15 min. Two restarts in a row can use up both attempts for the same job. Before restarting or updating, check that the queue is empty.
Replace the examples with lines from your queue on paper. One message per task.
transcrever: https://link-da-gravacao-de-segunda transcrever: https://link-da-aula-gravada explicativo: what is a durable queue | vertical dublar: https://link-do-video-em-ingles transcrever: https://link-do-podcast | modelo=haiku
/fila /jobs /status j13 /furar j15 /cancelar j14
sqlite3 inemaccbot.db \
"select id,fila,tarefa,status,flow_ref from jobs where status in ('queued','running');"
# vazio → pode reiniciar. Com render rodando → espere.text · running 2 · waiting 1 · errors 24h 0 render · running 1 · waiting 2 · errors 24h 0 io · running 0 · waiting 0 · errors 24h 0
Illustration of the kind of information; the bot’s exact text may vary.
transcrever · text queue · audio → text (local Whisper)dublar · text queue · translates and dubsexplicativo · render queue · explainer videocurso, demo, reel · render queueimagem · io queue · standalone imageSome depend on other projects installed on the machine. /skills shows what your bot has.
Use your queue on paper. Start with tasks without a gate.
Send the bot 5 to 10 tasks from your list (transcribe, summarize, explainer…). Check /fila once, then go do something else. Don’t restart the bot until the queue is empty.
Done when: `/fila` showed the distributed batch, and the next day, each task has a reply in the chat (result or reason for failure).
The morning check is lesson 7. Original exercise: 03-filas.md (E4).
The explanation appears after your answer.
Question 1 of 3
You sent four explainer video requests. Why is only one showing as running?
Question 2 of 3
A job failed with code 143. What does that indicate?skill
Question 3 of 3
You want to try a cheaper model for just one request. How?
Choose an option. The explanation appears after you try.