# AI assistant

> Ask an assistant about your logs, tasks and GitHub repositories, and what it can and cannot do.

The AI assistant (**AI Assistant** in the dashboard) answers questions about your own data: why a run failed, what changed in a repository, which task to pick up next. It sees only what you switch on for a chat.

## Context you can switch on

| Switch | What the assistant gets |
|---|---|
| Logs | Your recent log lines, newest first; repeated lines are collapsed |
| Tasks | Your active tasks |
| GitHub | Recent commits, pull requests, checks and Actions runs of the repositories you picked |
| Memory | The facts you asked it to remember (see below) |

With a switch off, that data is not sent at all. Everything the assistant reads from logs, tasks, repositories and web pages is passed to the model as untrusted data, not as instructions, so a crafted log line can't take it over.

## Levels: Low, Medium, High

Under the message box you pick how hard the assistant works on the next reply:

| Level | Costs | What changes |
|---|---|---|
| Low | 1 message | A quick answer, the default |
| Medium | 2 messages | The model thinks it through first and gets 15% more context (logs, history, files, search results) |
| High | 4 messages | The deepest thinking, 30% more context and room for longer answers |

When the model reasons before it answers, the chat shows it as a folded **Thinking** block above the reply; open it to see how the answer was reached.

## What it can do

- **Explain and summarise** logs, stack traces and failed CI runs.
- **Read files and CI logs** of a selected GitHub repository when it needs them to answer; you see a "Reading ..." status while it does.
- **Search the web** on its own when an answer depends on current public information, such as a library release or an error message, and list the pages it used as sources under the reply. Searches a day: Free 3, Starter 15, Expert 50, Enterprise 100. Search queries never carry tokens, keys or email addresses.
- **Propose tasks**: create a task or mark one done. It only proposes; nothing changes until you press **Yes**.
- **Ask you a question** with a few answer buttons after it has answered, when your choice decides the next step. Pressing a button sends that answer as your next message.

## Memory

When you tell the assistant something lasting, such as your stack, a naming convention or how you like answers, it offers to remember it. A fact is saved only when you press **Yes**, and it is then shared by all your chats while the **Memory** switch is on. Open **Manage** next to the switch to see and delete facts; up to 30 are kept. Facts that look like passwords or keys are never saved.

## GitHub access

Connect GitHub from the dashboard. The JustLog3 GitHub App is read-only: contents, pull requests, checks and Actions. It can't push, merge or comment. You choose which repositories the assistant may read:

- Free: 1 repository
- Starter: 5 repositories
- Expert: 15 repositories
- Enterprise: 15 repositories

## Chats and limits

Chats are saved, so you can come back to one later, edit your last question or retry an answer. Long chats stay cheap: older turns are condensed into a short summary instead of being resent in full. Messages a day (reset at 00:00 UTC; a reply that fails costs nothing):

- Free: 12 a day
- Starter: 75 a day
- Expert: 400 a day
- Enterprise: unlimited

Answers stream as they are written. If one AI provider fails mid-answer, the next one takes over automatically.