Skip to content

Talking to a desk

A message to a desk becomes a turn: the desk reads it with its instructions, thinks, calls tools, and answers. Every step is shown as it happens. In Chat a tool call is a folded card with its arguments and result, and the answer appears as it is written. Messages can carry pictures and documents, both ways; Pictures and documents explains how. A desk can also ask you something and carry on without waiting for the answer, as Questions, plans and permissions explains.

Slash commands

A message that starts with a slash is a command. They work in Chat, on Telegram and with bin/trading-agent send. Most are answered by code without asking a model; /btw and /review ask one, and say so. Several of them were borrowed from Claude Code, OpenClaw and Hermes Agent, which use the same names.

The last column says when a command is answered. At once means it is answered straight away, even while the desk is in the middle of a turn. Beside means it runs next to the turn, without waiting for it or interrupting it, and only one such command runs on a desk at a time. Queued means it waits until the running turn has finished, because it changes the desk.

Command What it does When
/status What the desk is doing and which models are ready At once
/usage Today's model calls, tokens, costs and Claude's limits At once
/context How full the context window is, and what fills it At once
/tools, /tools full The tools the desk can call, with the first sentence or the whole of each description At once
/skills Your skills and the ones the desk wrote, plus skills that cannot be used here and why At once
/subagents The desk's subagents, and today's runs with their outcome At once
/help The list At once
/stop Stops the running turn At once
/btw <question> A side question, such as /btw why did you sell INFY?, answered from the desk's conversation; neither the question nor the answer is kept in it Beside
/review [notes] The trade reviewer subagent judges today's orders and fills against your rules, with any notes you add Beside
/model Shows the model order; /model deepseek tries that model first from now on; /model auto goes back to the configured order. The model menu under the message box does the same Queued
/think Shows how hard each model thinks; /think off, low, medium, high, xhigh or max sets it from the next turn; /think default goes back to the configuration Queued
/new Forgets this conversation and starts a fresh one Queued
/compact Saves what matters, then compresses the conversation into a summary Queued
/rollover Summarises the day and starts a fresh session now Queued
/pause, /resume Stops the desk asking any model, and starts it again; commands still work while paused, except /btw and /review, which ask a model Queued
/<skill-name> ... Asks the desk to follow one of its skills Queued

/model and /think last until the gateway restarts; lasting settings belong in trading_agent.toml. For Claude, a new /think level restarts Claude Code at the next turn and resumes the same session, so the conversation is kept. Qwen and Gemma think or not, so every level except off turns thinking on. DeepSeek and OpenAI decide for themselves.

/btw asks the model that answered last. For Claude it runs in a throwaway copy of the session, in which every tool call is refused, and the copy is deleted afterwards.

On Telegram, a few commands are answered by the channel itself and cover every desk: /status, /positions, /pnl, /desk, /pause all, /resume all, /approve, /deny and /kill CONFIRM <reason>. The Telegram guide lists them. Typing / on the phone shows a menu of all the commands.

Steering a turn

A long turn can take minutes. A message you send while the desk is working does not wait for the turn to end: it goes to the desk's steering inbox, and the running turn picks it up at its next tool call. If the turn ends without another tool call, the message becomes the next turn, so nothing you write is lost. Chat shows the message inside the running turn, with whether it was handed over.

Steering and stopping a turn A message sent while a desk works waits in the steering inbox and is handed to the turn at its next tool round; /stop skips the queue. /stop, at once next tool round turn ended first: it becomes the next turn commands and other wake-ups wait here You, in Chat "Change of plan: end with BANANA" or the Stop button The desk a message from you while a model turn runs is steering /stop cancels the turn at once; tools already run are not undone Steering inbox one per desk; subagents have none Queue one turn at a time The running turn 1. model call 2. tool call, such as get_time 3. its result comes back 4. your message is added here 5. next model call sees it Qwen and Gemma: the agent loop adds it Claude: a PostToolUse hook hands it over Text appears in Chat as it is written your message, or /stop handed into the running turn
The orange dots are your message going into the steering inbox, and /stop skipping the queue; the blue dot is the message being handed to the running turn at its next tool round.

For Qwen and Gemma, the agent loop adds the message to the conversation before its next model call. For Claude, a PostToolUse hook hands it over after each tool call. Slash commands are never steering. The ones marked "At once" or "Beside" above are answered without waiting for the turn, and the rest wait in the queue.

Stopping a turn

/stop, or the Stop button in Chat, cancels the running turn at once. Tools that already ran are not undone: an order already placed stays placed. The turn answers "Stopped by the user", and the desk is ready for the next message.

On 17 October 2026, Qwen was asked to call get_time four times, and while it worked it was told to end its answer with the word BANANA; it did. A 700-word essay was then stopped mid-sentence, and the next message was answered normally.