Skip to content

features.md

Latest commit

 

History

History
225 lines (167 loc) · 7.35 KB

features.md

File metadata and controls

225 lines (167 loc) · 7.35 KB

Features

Thinking Modes

Mode Description
Disabled Standard responses without extended thinking
Concise Brief, efficient analysis
Verbose Detailed, thorough reasoning
Comprehensive Maximum depth analysis

Set via /settings > claude > set-thinking.

Operation Modes

Mode Permissions Description
Normal acceptEdits Can edit files, shows Allow/Deny buttons for other tools (e.g. Bash)
Trusted bypassPermissions Full auto-approval of all operations
Safe plan Analyze only, no modifications
Don't Ask dontAsk Auto-denies any tool that isn't pre-approved
Delegate delegate Restricts to Teammate + Task tools only (agent teams)

Set via /settings > permissions > set-mode.

Effort Levels

Level Description
min Quick responses, minimal processing
low Simple analysis and code completion
medium Standard effort with good accuracy
high Thorough analysis and robust output
max Maximum depth, best quality

Set via /settings > claude > set-effort.

Advanced SDK Features

Feature Description
Fast Mode Toggle Opus 4.6 speed-optimized API (2.5x faster, same quality) via /fast
1M Token Context Beta: increases context window to 1M tokens
Sandbox Mode Granular sandbox with network rules, filesystem ACLs, excluded commands
File Checkpointing Enables file change tracking for rewind support
Agent Teams Experimental multi-agent collaboration (delegate mode)
Hooks System Passive SDK callbacks for tool use, notification, and task observability
Additional Directories Multi-repo access — let Claude read/write across multiple directories
Fork Session Branch a conversation into a new independent session
AskUserQuestion Claude can ask clarifying questions mid-session via Discord buttons
Interactive Permission Requests Allow/Deny buttons when Claude wants to use unapproved tools (replaces CLI TUI prompt)
Channel Monitoring Auto-detect bot/webhook messages and trigger Claude investigation in a thread
Thread-per-Session Dedicated Discord thread for each /claude-thread conversation with custom names

Toggles available via /settings > claude.

Agent System

7 built-in agents optimized for specific tasks:

Agent Specialization
code-review Deep code analysis, security, and style checks
debug Error diagnosis and root cause analysis
architect System design and architecture planning
refactor Code restructuring and improvement
test-writer Test generation and coverage analysis
docs Documentation generation
security Security vulnerability scanning

Usage

/agent start code-review    # Start an agent
/agent list                  # View all available agents
/agent status                # Check running agent status
/agent stop                  # Stop current agent

Agents run through the SDK using native AgentDefinition support with tailored system prompts.

MCP Server Management

Manage Model Context Protocol servers for Claude:

/mcp list        # Show configured MCP servers
/mcp add         # Add a new MCP server
/mcp remove      # Remove an MCP server
/mcp status      # Check connection status
/mcp test        # Test a server connection
/mcp toggle      # Enable/disable a server mid-session
/mcp reconnect   # Reconnect a failed server mid-session

MCP configuration is stored in .claude/mcp.json:

{
  "mcpServers": {
    "my-server": {
      "command": "npx",
      "args": ["-y", "my-mcp-server"],
      "env": {
        "API_KEY": "your-key"
      }
    }
  }
}

Mid-Session Controls

Change model or permissions without restarting your session:

/claude-control set-model claude-sonnet-4-20250514
/claude-control set-permissions bypassPermissions
/claude-control stop-task <task_id>
/claude-control status
/claude-control interrupt

These operations apply immediately to the active query.

File Rewind

Undo file changes made during a Claude session:

/rewind turn:3              # Rewind to turn 3
/rewind turn:5 dry_run:true # Preview what would change

Requires file checkpointing to be enabled via /settings.

Structured Output

Force Claude to respond in JSON matching a specific schema:

  1. Enable via /settings > output > toggle-structured-output
  2. Set schema via /settings > output > set-output-schema

Example schema:

{
  "type": "object",
  "properties": {
    "summary": { "type": "string" },
    "confidence": { "type": "number" }
  }
}

When enabled, responses follow json_schema output format through the SDK.

Thread-per-Session

Keep conversations organized with dedicated Discord threads:

  • /claude-thread — Start a new Claude session in its own thread
    • name — Custom thread title (required)
    • prompt — Your initial prompt (required)
  • Per-channel session tracking — Regular /claude commands in the same channel automatically reuse the last session (no need to pass session_id)
  • Thread auto-naming — Threads are named with your custom title or a truncated version of the prompt
  • Seamless continuation — Subsequent /claude commands inside a session thread continue that conversation

How It Works

  1. Run /claude-thread name:"Fix auth bug" prompt:"Review the auth module"
  2. A new thread titled Fix auth bug is created
  3. Claude's response streams into the thread
  4. Any further /claude commands in that thread continue the same session
  5. Use /claude-thread again in the main channel to start a fresh conversation in a new thread

Channel Monitoring

Automatically investigate alerts from other bots or webhooks:

  1. A monitored bot posts in the configured channel
  2. Messages are batched over a 30-second debounce window
  3. A thread is created on the alert message
  4. Claude analyzes the alert and streams its investigation into the thread

Setup

MONITOR_CHANNEL_ID=123456789012345678    # Channel to watch
MONITOR_BOT_IDS=987654321,111111111      # Bot/webhook user IDs to respond to

Requires the Message Content privileged intent enabled in the Discord Developer Portal. The bot needs Read Messages, Create Public Threads, and Send Messages in Threads permissions in the monitored channel.

Role-Based Access Control (RBAC)

Restrict dangerous commands to authorized users and roles:

Restricted Command Categories

Category Commands
Shell /shell, /exec, /run, /terminal
Git /git, /commit, /push, /pull, /branch
System /shutdown, /restart, /config
Admin /admin

Configuration

Set via environment variables in .env:

# Comma-separated Discord Role IDs
ADMIN_ROLE_IDS=123456789,987654321

# Comma-separated Discord User IDs (bypass role checks)
ADMIN_USER_IDS=111111111,222222222

Behavior

  • Neither variable set: RBAC disabled — all commands open (default)
  • One or both set: Only users with a listed role or user ID can run restricted commands
  • Denied users: Receive an ephemeral "Permission Denied" message (only they can see it)
  • Unrestricted commands: Always available to everyone (e.g., /claude, /settings, /help)