Helpers and specialists
When a question needs several reads that do not depend on each other, such as the logs, the recent deploys and the error rate of one service, Halon hands them to helpers that read at the same time instead of one after another. Each helper gets one short brief, only reads, and reports back to Halon in a few lines, so you get the answer sooner.
You do not ask for helpers. Halon decides when a question has parts it can check side by side, and keeps for itself anything that changes something or depends on another check’s answer.
Where you see them
Section titled “Where you see them”In a chat in the dashboard, Halon’s steps show one step such as 3 checks at once, or One check by a helper, followed by the name of each check. Under it, each helper has its own list of steps, open while it works and folded once it is done. What it reported sits right below its name, or why it has no report, so you can read every result without opening anything.
- Click a helper’s name to see every step it took, and click a step to see what came back.
- Open the 3 checks at once step to read the brief Halon gave each helper.
In a Slack thread, the same step appears with one line per helper right under it, such as Logs of checkout with “2 steps, reported”. Each line updates in place while the helper reads and once it reports, and a helper with no report is marked as failed.
On an investigation’s page, under How it got there, the step shows as 3 checks at once with each helper folded to its name, how it ended and the numbers of its steps, such as “Reported · Steps 2, 3”, with what it reported right below. Click a helper’s name to see its steps. They keep the investigation’s own numbers, so the answer cites them like any other step, and following a link to one of them opens the helper that holds it.
What a helper can reach
Section titled “What a helper can reach”A helper only reads. It never changes anything, whatever it reads, and a change is always something Halon asks you about itself.
A helper has the same permissions as whoever Halon is working for, never more. In a chat that is the person who asked, and in an investigation it is Firefight’s own Halon agent, with the grants your workspace gave it.
Every read a helper makes is recorded under Gateway → Activity, with the chat or investigation it came from. See Activity.
Models and spending
Section titled “Models and spending”Helpers run on the quick model of your AI account. A check that has to reason across several results, such as reading code or comparing a failing run with one that passed, runs on the main model.
Helpers spend from the same budget as the question or investigation that started them. When too little is left, Halon reads the rest itself.
Limits
Section titled “Limits”| Limit | Value |
|---|---|
| Checks Halon hands off at once | 4 |
| Checks for one question, or for one investigation | 8 |
| Helpers running at once across your workspace | 12 |
When a limit is reached, Halon reads what is left itself.
When a helper has no report
Section titled “When a helper has no report”A helper that cannot finish says why where its report would be, and the others carry on.
| What you see | What happened |
|---|---|
| It ran out of the budget it was given before it could report. | Its share of the budget was spent first |
| It took too many steps without reporting. | It read for a long time without settling the check |
| It could not run. | Something failed on Firefight’s side |
| Stopped before it reported. | You stopped the answer or the investigation |
| It was interrupted before it reported. | Firefight restarted while it worked |
Pressing Stop in a chat, or stopping an investigation, stops its helpers too.
Specialists
Section titled “Specialists”When your workspace has connected a great many tools, and the AI model Halon runs on cannot search a long list of them, Halon hands every read of a connected system to a specialist instead of reading it itself. A specialist is a helper that knows one kind of system and holds only the tools for it, which keeps Halon’s answers quick. Halon picks the kind for each question.
| Specialist | What it reads |
|---|---|
| Code | Repositories, commits, pull requests, branches and files on your code hosts |
| Deploys and runs | Where apps run, their deployments, settings, builds, CI and pipeline runs, and what changed |
| Logs, metrics and traces | Logs, metrics, traces, errors and product analytics |
| Databases | Database health, connections, locks, slow queries, replication, schemas and read only queries |
| Cloud resources and cost | Cloud accounts, their resources, networking, DNS, quotas, limits and the bill |
| Tickets and docs | Issues and follow-up work, your knowledge pages, runbooks and the handbook |
| On-call and alerts | Who is on call, schedules, pages and the alerts behind them |
The step reads Asked a specialist, or Asked 3 specialists, with the kinds asked, and each specialist shows under it the way a helper does, in the dashboard and in a Slack thread. A kind with nothing connected, or with no tool that reads switched on, cannot be asked, and Halon says so.
A specialist only reads, as the person asking, exactly like a helper. It may point to a change, such as a rollback, but its report is a lead and never the reason for a change on its own. Before Halon asks you to approve anything, it makes the read that backs the change itself, and the confirmation shows that read under Why. See Halon reads before it proposes a change.
Specialists count toward the same limits as helpers and spend from the same budget. They work in chats only, never in investigations or watches.