Slash Commands

Overview

Agents support slash commands for quick actions. These work in both the web chat and Telegram channel. Commands are recognized by messages starting with /.

Built-in Commands

/clear

Clears the current conversation history for the active session. In v1.0.0, this command automatically archives the session to the secure session_archive.db before deletion, preventing accidental data loss.

Behavior:

• Archives current session state to session_archive.db
• Deletes all messages from the current session
• Resets the session summary
• Truncates the agent’s llm.log and sessrecap.log files
• Resets agent state to fresh plan mode
• Emits a session_clear event
• The agent responds with a confirmation message

Example:

/clear
> History cleared.

---

/clear-memory

Introduced in v0.6.91.

Clears the agent’s long-term memories. This removes all stored facts, preferences, and remembered information that the agent has accumulated across sessions.

Behavior:

• Deletes all memories stored via remember() or the Message Wrapper Protocol
• Each memory is soft-deleted (marked as expired) so it won’t appear in future recall() results
• The agent responds with a confirmation of how many memories were cleared

Example:

/clear-memory
> Cleared 12 memories.

Note: This only affects the agent’s long-term memory, not the current conversation history. To clear the conversation history, use /clear instead.

---

/help

Displays a list of available slash commands with brief descriptions. Certain commands (e.g. /restart) are hidden from non-super agents.

Response format:

Available commands:
- /clear — Clear chat history for this session
- /help — Show available commands
- /summary — Force regenerate session summary

---

/summary

Triggers conversation summarization for the current session. This compresses the conversation history into a concise summary, freeing up context space for the LLM.

Behavior:

• Runs the summarization algorithm on the current session
• Replaces the full message history with a summary
• The agent responds with a confirmation

Example:

/summary
> Session summary has been regenerated.

---

/stop

Stops the agent’s current processing loop immediately. Use this when the agent is stuck in a long-running response or you want to cancel an ongoing operation.

Behavior:

• Sends a stop signal to the agent runtime
• The current processing loop is interrupted
• The agent responds with a confirmation

Example:

/stop
> Stop signal sent.

---

/cwd

Displays the current workspace directory the agent is operating in.

Behavior:

• Reads the agent’s configured workspace path from the database
• Returns the absolute path

Example:

/cwd
> Current workspace: /workspace

---

/cd

Changes the agent’s workspace directory to a new path. Directory traversal via .. is blocked for security.

Permission: All agents (super-agent restriction has been lifted).

Behavior:

• Updates the agent’s workspace path in the database
• Destroys the existing Docker container so the new workspace is mounted on the next tool use
• Returns the new workspace path

Example:

/cd /workspace/projects/evonic
> Workspace changed to: /workspace/projects/evonic

Error cases:

/cd /nonexistent/path
> Error: directory does not exist: /nonexistent/path

/cd ../etc
> Error: path contains '..' which is not allowed: /workspace/../etc

---

/restart

Restarts the entire Evonic service. This command replaces the running process in-place, preserving the original execution mode (foreground stays foreground, daemon stays daemon).

Permission: Super agent only.

Behavior:

• Validates the caller is the super agent
• Persists caller info so the new process can send a “ready” notification
• Stops all channels cleanly (releasing Telegram long-poll)
• Closes inherited file descriptors
• Resolves the correct restart target (release or dev mode)
• Replaces the current process via os.execv

Example:

/restart
> Restarting...

The new process will send a notification to the caller once boot is complete.

---

/plan

Switches the agent to plan mode. In plan mode, the agent enters a deliberate planning phase with write tools blocked.

Behavior:

• Creates a fresh AgentState in plan mode
• Persists the state to the database
• The agent responds with a confirmation

Example:

/plan
> Switched to plan mode.

---

/exec

Introduced in v0.5.0.

Switches the agent to execute mode. In execute mode, write tools are unlocked and the agent can take actions.

Behavior:

• Creates a fresh AgentState in execute mode
• Persists the state to the database
• The agent responds with a confirmation

Example:

/exec
> Switched to execute mode.

---

/dump

Introduced in v1.0.0.

Exports the current session state as a JSONL file. This is invaluable for auditing, debugging, or migrating a conversation to another environment.

Behavior:

• Collects all messages, tool outputs, and agent state for the active session
• Generates a formatted JSONL dump file
• Saves the file to the agent’s artifacts directory
• Returns a download link or artifact filename

Example:

/dump
> Session dumped to artifact: session_dump_20260717.jsonl

---

/unfocus

Clears the agent’s focus mode, allowing it to accept messages from all sessions again. When an agent is focused on a specific session, other sessions are ignored.

Behavior:

• Checks if focus mode is currently active
• If active, clears the focus and focus reason
• Returns information about what was cleared
• If focus mode was already off, returns a notification

Example:

/unfocus
> Focus mode cleared (was: working on task #183). Agent sekarang bisa menerima semua session.

---

/status

Displays detailed status information about the agent, including model, mode, workplace, workspace, toggles, tools, skills, and channels.

Behavior:

• Queries the agent’s full configuration and state from the database
• Returns structured information including:

Field	Description
Model	Active LLM model (agent-specific or override)
Mode	Current agent mode (plan or execute)
Focus	Whether focus mode is active (and reason)
Plan file	Active plan file path (if any)
Workplace	Workplace name, type, and connection status
Workspace	Workspace directory path
Toggles	Sandbox, Safety Checker, Vision, Agent Messaging
Tools	Number of tools configured
Skills	Number of skills configured
Channels	Connected channels and their status

Example:

/status
> **Status — Adit**
>
> Model: Claude Sonnet 4 (claude-sonnet-4-20250514)
>
> Mode: execute
>
> Focus: yes — working on task #183
>
> Plan file: plan/kanban-task-183.md
>
> Workplace: Main Office (local, connected)
>
> Workspace: /workspace/docs-site
>
> Toggles:
>   Sandbox: enabled
>   Safety Checker: enabled
>   Vision: enabled
>   Agent Messaging: enabled
>
> Tools: 12
>
> Skills: 3

---

/shutdown

Introduced in v0.7.0.

Cleanly shuts down the entire Evonic server from within a conversation. No terminal access needed.

Permission: Super agent only.

Behavior:

• Validates the caller is the super agent
• Initiates a graceful server shutdown
• All channels are stopped cleanly
• The server process exits

Example:

/shutdown
> Shutting down Evonic...

---

/detach

Introduced in v0.8.0.

Moves a long-running command (builds, downloads, compilations) to the background so you can keep chatting while it runs. Progress is tracked persistently across the session.

Behavior:

• The process starts in the background immediately
• You can continue chatting without waiting for the process to finish
• When the job completes, the agent automatically notifies you with the result
• The background job tracker survives agent restarts within the same session

Example:

/detach
> Process detached with job ID #42. You'll be notified when it completes.

---

/investigate

Introduced in v0.8.0.

Inspects another agent’s context from the chat, surfacing session state, tool configuration, and runtime diagnostics without leaving the conversation.

Usage:

/investigate <agent-id> <context>

Parameters:

Part	Description
agent-id	The target agent’s slug ID (e.g. bookstore_bot)
context	What to inspect (session state, tools, config, etc.)

Behavior:

• Queries the target agent’s runtime configuration
• Returns structured diagnostic information
• No modification is made to the target agent

Example:

/investigate bookstore_bot tools
> Agent: bookstore_bot
> Tools (10):
>   - bash
>   - read
>   - save_artifact
>   ...

---

/sub

Introduced in v0.8.0.

Spawns a sub-agent directly from chat for parallel work. Sub-agents execute without planning delays and deliver responses via inter-agent forwarding.

Usage:

/sub <name> --description "<purpose>"

Behavior:

• Creates a lightweight sub-agent with a descriptive name
• The sub-agent inherits the parent’s model, tools, and skills
• Sub-agents skip the planning phase and execute directly
• Replies are automatically forwarded back to your session
• Naming-pattern enforcement prevents invalid sub-agent IDs

Example:

/sub researcher --description "Research topic X and summarize findings"
> Sub-agent 'researcher' spawned. It will report back when done.

See Sub-Agents for the full system documentation.

Slash Command Controls

Introduced in v1.1.0.

Each agent can now be configured with hidden and disabled slash command lists directly from the agent Settings tab. This gives you precise control over which commands are visible and usable per agent.

Setting	Description
hidden_slash_commands	Commands hidden from the /help listing but still executable
disabled_slash_commands	Commands fully blocked — they won’t appear and won’t execute

Behavior:

• Hidden commands won’t show up in /help but can still be invoked by name
• Disabled commands return a clear error if someone tries to invoke them
• Configure these per-agent in the Settings tab or via API

Example — hiding advanced commands from a public-facing agent:

# In the agent's hidden_slash_commands:
/restart, /shutdown

# In the agent's disabled_slash_commands:
/investigate, /sub

---

Guided Slash Commands

Introduced in v1.1.0.

Slash commands now support parameter metadata that enables live frontend hints and autocomplete as you type. When you start typing a command like /clear or /model, the UI shows the expected parameters so you never have to guess the syntax.

How it works:

• Each command exposes parameters, accepts_args, and to_dict metadata
• The frontend reads this metadata and renders inline hints and autocomplete suggestions
• Supported commands include /clear, /model, /sub, /investigate, and more

Example — typing /sub shows inline hints:

/sub <name> --description "<purpose>"

The <name> and --description placeholders appear as grey hints in the input field, and autocomplete fills them as you tab through.

---

How Commands Are Processed

When a message starts with /, the agent runtime intercepts it before sending to the LLM:

1. Parse the command name (text after /, before any space)
2. Match against known commands in the registry
3. Execute the corresponding handler and return the response
4. If the command is unknown, pass the message to the LLM normally

Implementation

Slash commands are implemented in backend/slash_commands.py using a registry pattern (SlashCommandRegistry) and integrated into the agent runtime (backend/agent_runtime.py). The command parsing uses regex (^/(\w+)(?:\s+(.*))?$) to extract command name and optional arguments.

New commands can be added by registering a handler function with command_registry.register(name, handler, description). Each handler receives (session_id, agent_id, external_user_id, channel_id, args) and returns a response string.