Two surfaces: briglia subcommands in your shell, and /commands inside a conversation. The chat commands are one shared set — everything that works on Telegram works in the terminal, and vice versa.
Shell: briglia …
Command
What it does
briglia · briglia chat
Interactive chat with Briglia (the default command).
briglia menu
Setup and settings in your browser (English or Italian): pick ChatGPT subscription, OpenCode Go, OpenRouter or your own named servers (local or online); connect Telegram (the chat is detected when you message your bot); add keys, email, permissions and tools. Afterwards it is the hub for your AI lanes (a “Briglia is using” selector to switch; adding a lane doesn’t switch), model and effort, and has a Start/Stop switch. Works while Briglia is running.
briglia setup
Setup wizard: providers, keys, permissions, channels. Re-run any single section later.
briglia daemon
Headless mode: no local chat, messaging channels (Telegram) only. One conversation-owning instance at a time — briglia and briglia daemon share state, the second refuses to start.
briglia doctor [--online]
Check configuration, permissions and toolchain; --online also probes the configured services. Exit code 1 if problems were found.
briglia upgrade [-y] · briglia update
Update to the latest signed release; -y skips the confirmation.
briglia service install [--no-keepawake] | status | uninstall
Linux: run Briglia as an always-on systemd user service (auto-start at boot via linger; keep-awake support on Ubuntu Touch). Bare briglia service = status.
Post an event to an external-trigger watcher created by Briglia. Works even when Briglia isn't running.
briglia media-selftest
Verify the PDF/image pipeline (poppler/ImageMagick on Linux).
briglia --version · briglia --help
Version and help.
Hidden but present: briglia setup-api (JSON setup surface used by the Ubuntu Touch app) and briglia bundle-check (used by the updater). Internal __… commands are self-tests.
Chat: /commands
menu marks the eight commands shown in Telegram's “/” menu (since CLI 0.2.12, in this order: /status, /provider, /model, /effort, /prune, /upgrade, /commands, /stop); /commands prints the full standard set below. Typed arguments keep working everywhere; the buttons are a Telegram convenience.
Control
/stopmenu
Stop the current work immediately, including everything Briglia moved to the background (commands, subagents, results still waiting); stopped work does not restart later. Since CLI 0.2.50 the reply is honest: if something is still winding down after a short wait (for example a helper process, or memory archiving, which is never interrupted), it says what, with how long it has been running, and sends one “✅ The stopped request has ended” when it really has. A repeated /stop answers “Already stopping”. The terminal and the phone app's Stop button now show the reply too
/statusmenu
Show what Briglia is doing right now, including background jobs (since CLI 0.2.38), a stopped request that is still finishing, and a memory archive running in the background (since CLI 0.2.50). Since CLI 0.2.52 it no longer lists unknown charges — those stay in /spend and briglia doctor; only when a spending limit is set and paid work is paused does it show one line pointing to /spend
/prune [nosnapshot]menu
Free up Briglia's working memory. Since 0.2.13 a full-text snapshot of the conversation is saved first (see Storage & privacy); nosnapshot forces one prune without it, for a full disk. The report counts the new summary (0.2.14)
/continue
Show the rest of a long reply
Models
/model [id]menu
Show or switch the main model — on Telegram, tap a button for the OpenCode Go catalog or the ChatGPT models (GPT-6.1 Sol; GPT-6 Sol, Luna, Astra; GPT-5.6 Luna, Terra, Sol); other providers take a typed id
/provider [name]menu
Show or switch the LLM provider — on Telegram, one button per configured provider, and one per named server (since CLI 0.2.37; /provider <name> is case-insensitive). Switching to OpenCode Go or a server with no OpenAI key saved still switches, with a warning that web page reading will fail until you add the key
/effort [level]menu
Show or set the reasoning effort — on Telegram, buttons for the levels the current model accepts
/websearch [name]
Show or switch the web page-reading backend (openai, opencode, openrouter): page extraction and web_fetch compression. The research itself runs on your main model (since CLI 0.2.37). While /provider is chatgpt, page reading follows the subscription; while it is openrouter, page reading uses GPT-6 Luna through OpenRouter (OpenAI or Azure host). This choice applies on the other providers
/subagentmodels
Show or set the cheap subagent model lanes
/orprovider [slug|off]
OpenRouter only (CLI 0.2.30): pin the main agent's requests to one upstream host. No argument lists the hosts OpenRouter serves your current model from (name, price per million tokens, cache-read price, uptime) with tap buttons on Telegram; a base slug such as deepinfra covers every endpoint of that host, a full tag such as deepinfra/turbo targets one variant. A pin is hard: requests fail with OpenRouter's error instead of hopping when the host is unavailable. Validated against the live host list; /model onto a model the pinned host doesn't serve releases the pin and says so. Scope: the main model — the main agent and every subagent; while a pin is set the cheap subagent lanes (/subagentmodels) are bypassed so subagents run the main model on the pinned host, and they return when the pin is released. Web research keeps its own routing. /orprovider off returns to OpenRouter's automatic (sticky-by-session) routing.
/subagents [on|off]
Turn the Agent delegation tools on or off
/websubagent [on|off]
Turn the Web researcher subagent on or off (default on since CLI 0.2.29, off restores the legacy web_search/web_research_sweep tools byte for byte; needs subagents on and web search configured; while on, ordinary subagents delegate their web research to it instead of using the legacy web tools)
/subscription [login|status|cancel|logout]
Connect, inspect or disconnect your ChatGPT subscription (device-code sign-in), then /provider chatgpt
/cachestats
Recent token/cache statistics for the ChatGPT and OpenAI Responses providers (statistics, not credits)
System
/upgrademenu
Update Briglia to the latest release
/restart
Restart Briglia (reloads mcp.json and skills)
/archiveinline [on|off]
Since CLI 0.2.50 memory archiving runs in the background by default: Briglia keeps replying while the oldest messages are summarised, and the new summary replaces them at the start of the next turn. on restores the old behaviour — wait for the archive before replying — for a local model server that handles one request at a time; off returns to the background default
/commandsmenu
List standard Briglia commands
Account
/setname [name]
Set or change your name (asks for confirmation)
/deleteuserdata
Erase all memory and user data (asks for confirmation)
/exportmind [lite]
Save a memory backup (.mind) to this computer (add 'lite' to skip files)
/importmind [path]
Restore a memory backup (asks for confirmation)
/resumewatcher [id]
Review and re-arm watchers quarantined by a memory import
/switchbot [token]
Move Briglia to a different Telegram bot (guided, asks for confirmation)
/rotateaffinity
Present a new session identity to OpenCode/OpenRouter from the next request (no data change)
Terminal only
/help
Show the command list
/attach <path> [message]
Send a file, with optional text
/quit · /exit
Leave the chat (also Ctrl-C)
Power commands
Working but kept out of the table: /spend (shows today's and this month's spend with each limit; /spend turn|daily|monthly <usd|off> sets or removes a per-turn tool cap or a daily/monthly limit — all off by default; /more1, /more5, /more10 raise a reached daily/monthly limit for today/this month only; since CLI 0.2.38, when a limit is set and some cost can't be known, paid work pauses until you repair the spend file or send /spend accept-unknown, which counts only those unknown amounts as $0 and keeps every known charge — /more can't lift that pause; since CLI 0.2.42 an OpenRouter page-reading call that was cut at its time limit or whose connection dropped is also an unknown charge, and since CLI 0.2.44 the limit is 2 minutes on every backend and the same applies to page-reading calls on your OpenAI API key), /hide and /show (privacy mode for local surfaces), /transcribe_openai. Aliases: /pulisci = /prune, /riavvia = /restart, /comandi = /commands, /continua = /continue.
Two-step commands
/setname <name> then /setname confirm; /importmind <path> then /importmind confirm; /deleteuserdata and /switchbot ask for a confirmation too. /upgrade and /restart refuse while a turn is running — /stop first.
