Most Agent Bridge work is an ordinary Telegram conversation. You ask a coding agent to inspect a repository, make a change or show evidence, and then reply when a decision is needed. Slash commands are for changing the workstream controls, not for telling the agent how to write code.
These commands refer to the unified Telegram interactive bot. Provider-locked Telegram bots and the Discord and Slack surfaces differ. If you haven't connected a bot, begin with the Farstax setup guide.
Keep each project in the right conversation
A direct message is one workstream. In a Telegram supergroup with forum topics enabled, separate topics such as “OSS” and “Platform” become separate workstreams: each has its own conversation, queue, provider preference and native provider sessions. Replies go back to the originating topic. Add the bot with permissions to receive the messages you intend it to handle; the topic workstreams guide covers that setup.
In the topic for a project, you can write:
Investigate the intermittent release CI failure in owner/repo. Collect evidence first and do not change the release configuration yet.
“What did you find?” is a normal follow-up in that same topic. You do not need to start a new session for each message.
/cli — switch coding agents in this workstream
Send /cli to see the active provider, available authenticated CLIs and the inline provider picker. Choose Claude Code, Codex or another provider actually offered in that menu. It changes the selection for this conversation or topic, not all chats.
For example, ask Claude to propose a test, use /cli to choose Codex, then ask Codex to challenge the test. Agent Bridge retains the conversation and useful handoff context, but each provider maintains its own native session. A running Claude process and its private context are not transplanted into Codex.
Automatic fallback is separate from manual selection. It depends on configured alternatives, capacity and safe recovery; externally executed work must not be blindly repeated. Provider-locked bots cannot change provider. Use the working /cli picker rather than depending on a suggested /switch shortcut.
/stop and /reset — very different controls
/stop cancels and fences the current active Run. Use it if the agent is taking an unsafe direction, then ask what may already have changed. It cannot undo edits, commits, deployments or messages already sent.
/reset deliberately clears the scoped workstream's conversation history, provider session, working scratchpad and pending continuation. It is not a refresh button and does not roll back external work. If you only need to redirect the current agent, send a normal message or use /stop.
Both commands operate in the conversation or forum topic where you send them.
/btw and /context — ask without disrupting the main task
/btw <question> asks a fresh, one-off, read-only side question without disturbing the current provider session. For example:
/btw What does exit code 78 mean in a systemd service?
Use an ordinary follow-up when the answer should form part of the continuing work. A side question does not give authority for changes.
/context reports the workstream's retained turn count, pending queue and latest turn. It does not expose the complete private context window of Claude Code or Codex. This is useful when deciding whether to continue, hand off or deliberately reset a long conversation.
/queue_mode — what happens when you send a message mid-task
Send /queue_mode to choose how new messages are handled while this workstream is busy.
augment: Agent Bridge tries to incorporate new instructions into the active work when supported; otherwise it can cancel and resume through its managed queue. Do not assume a message always reaches the provider immediately.
interrupt: The runtime cancels the current execution and processes the newly queued instruction.
queue: The new message waits until the active work can finish, without asking it to interrupt.
The menu also offers Use configured default. This is one lane of work, not three concurrent agents. Use /stop if you mean to terminate the active run rather than steer it with another request.
/models, /effort and /skills — inspect available capabilities
/models presents model preferences supported by the selected provider; /effort presents available reasoning-effort preferences. Choices vary with the actual CLI configuration and are not a promise that your external subscription allows each option.
/skills prints the bundled Agent Bridge Skills catalogue and local install/repair hints. It is not a verified list of the Skills currently installed, enabled and discoverable by the active provider. For engineering work, ask the provider to check its effective Skills and any repository AGENTS.md instructions, then see the requirements-to-PR guide.
/sensors — inspect observations
/sensors lists the built-in agent-bridge and server checks plus any external Sensors the operator has configured. Choose a check or Run all sensors. Reports are observations, not automatic repair requests. Website-specific monitoring is not built in; the Sensors and Routines guide explains the additional configuration and separately authorised scheduling.
Less common commands
/start gives a basic readiness/help response. /usage is specific to supported Codex routes rather than a universal provider-quota command. /narration on|off|status is an Antigravity/Agy control, not a Claude or Codex setting. /autonomy appears only when first-class Telegram autonomy is configured; its explicit approval/status/stop controls concern a bounded Episode, not permission to make arbitrary production changes.
You can agree a scheduled Routine in normal conversation. The owner and agent must confirm its instruction, schedule and time zone before creating it through the scoped helper. There is no general /schedule Telegram command to assume.
A useful daily pattern
In the right topic, describe the result you want. Use /context to inspect the retained workstream, /btw for an unrelated question, /cli when another provider would help, and /stop when the active task should end. Reserve /reset for starting a genuinely new conversation.
A successful task is still proved by repository changes, test output and any resulting review—not by the number of commands used.
Prefer not to maintain the workspace infrastructure? Apply for the Farstax beta.
Prefer to run it yourself? Agent Bridge is open source.