Wiki · Start here
Your first chat
An agent is not a search box. Here is how to give it tasks you can verify — and what to do when something goes wrong.
What it looks like
Type hermes and a terminal conversation opens. You write plain English, the agent works:
In the conversation you see what the agent does: every tool it runs starts with ● and ends with
✓ (or ✗ when it failed). A blinking cursor means it is still thinking.
How to give tasks that can be verified
| Instead of | Write | Why |
|---|---|---|
| "Set up the server." | "Find the OS version and write it to version.txt." | You read the result and immediately know whether it was done right |
| "Tidy up the data." | "List the ten largest files in /var/log and delete nothing." | A small step you can check, with clear limits |
| "Fix it." | "Find the cause of the error in the log, tell me what it is and suggest two fixes — do not change anything yet." | Diagnosis before action keeps the agent from charging off in the wrong direction |
| "Check the website." | "Verify the page returns 200 and that the header declares utf-8; show me the exact command you used." | You want evidence, not an impression |
Golden rule: the better you can verify the result yourself, the better the agent works. If you cannot tell that a task is finished, neither can it.
Useful slash commands
Inside a chat you drive the agent with commands starting with a slash:
| Command | What it does |
|---|---|
/help | Lists everything you can do in the chat |
/new | Starts a fresh conversation (the old one stays saved) |
/model | Switches model mid-task (expensive one only for hard work) |
/context | Shows how full the conversation memory is |
/compress | Trims the conversation memory when it fills up |
/undo and /retry | Removes the last exchange / resends the last message |
/stop | Kills all running background tasks |
/queue (or /q) | Queues the next task for when the current one finishes |
/bg | Runs a task in the background while you carry on |
/status | Shows model, profile, working directory and session ID |
The full list lives in the official documentation (160+ commands) — link below.
Approvals: the agent asks before anything risky
Before running something that can change or delete things, an approval panel appears
(⚠ approval required) showing the exact command. You confirm, or refuse and say what to do instead.
A mode where the agent never asks (YOLO) exists, but we do not recommend it on a server that holds data —
nobody is watching what gets run.
Coming back later
Tips from experience
| Tip | Why |
|---|---|
| Ask it to inspect before acting | "Take a look…" instead of "Do…" saves you from cleaning up after a rushed change |
| Ask for evidence | "Show me the command you used." You verify both the result and that it was not invented |
When it goes wrong, use /stop | It halts running tasks. File changes can be reverted with /rollback |
| Do not hand over one giant task | Split it into steps you can check. Smaller steps mean fewer dead ends |
| Anything that must last belongs in a file | A conversation is forgotten; a file (a guide, a note, a skill) is not |
When it does not work first time
| What happens | What to do |
|---|---|
| The agent understood it differently | Type /undo and be more specific. Tell it what you expected |
| It loops or gets stuck | /stop, or start a new chat with /new and summarise what you know |
| The conversation memory is full | /context shows the state, /compress trims it |
| It broke something | /rollback restores files from a checkpoint; ask before system-wide changes |