# Advanced Features
Comprehensive guide to Claude Code's advanced capabilities including planning mode, extended thinking, auto mode, background tasks, permission modes, print mode (non-interactive), session management, interactive features, channels, voice dictation, remote control, web sessions, desktop app, task list, prompt suggestions, git worktrees, sandboxing, managed settings, and configuration.
## Table of Contents
1. [Overview](#overview)
2. [Planning Mode](#planning-mode)
3. [Extended Thinking](#extended-thinking)
4. [Auto Mode](#auto-mode)
5. [Background Tasks](#background-tasks)
6. [Monitor Tool (Event-Driven Streams)](#monitor-tool-event-driven-streams)
7. [Dynamic Workflows](#dynamic-workflows)
8. [Scheduled Tasks](#scheduled-tasks)
9. [Permission Modes](#permission-modes)
10. [Headless Mode](#headless-mode)
11. [Session Management](#session-management)
12. [Cross-Session Messaging](#cross-session-messaging)
13. [Interactive Features](#interactive-features)
14. [Output Styles](#output-styles)
15. [Status Line](#status-line)
16. [TUI Mode (Fullscreen)](#tui-mode-fullscreen)
17. [Voice Dictation](#voice-dictation)
18. [Channels](#channels)
19. [Chrome Integration](#chrome-integration)
20. [Remote Control](#remote-control)
21. [Web Sessions](#web-sessions)
22. [Desktop App](#desktop-app)
23. [Task List](#task-list)
24. [Prompt Suggestions](#prompt-suggestions)
25. [Git Worktrees](#git-worktrees)
26. [Sandboxing](#sandboxing)
27. [Managed Settings (Enterprise)](#managed-settings-enterprise)
28. [Configuration and Settings](#configuration-and-settings)
29. [Trust and Permission Scoping](#trust-and-permission-scoping)
30. [Agent Teams](#agent-teams)
31. [Best Practices](#best-practices)
32. [Additional Resources](#additional-resources)
---
## Overview
Advanced features in Claude Code extend the core capabilities with planning, reasoning, automation, and control mechanisms. These features enable sophisticated workflows for complex development tasks, code review, automation, and multi-session management.
**Key advanced features include:**
- **Planning Mode**: Create detailed implementation plans before coding
- **Extended Thinking**: Deep reasoning for complex problems
- **Auto Mode**: Background safety classifier reviews each action before execution
- **Background Tasks**: Run long operations without blocking the conversation
- **Permission Modes**: Control what Claude can do (`manual` — formerly `default`, `acceptEdits`, `plan`, `auto`, `dontAsk`, `bypassPermissions`)
- **Print Mode**: Run Claude Code non-interactively for automation and CI/CD (`claude -p`)
- **Session Management**: Manage multiple work sessions
- **Interactive Features**: Keyboard shortcuts, multi-line input, and command history
- **Voice Dictation**: Push-to-talk voice input with 20-language STT support
- **Channels**: MCP servers push messages into running sessions (Research Preview)
- **Remote Control**: Control Claude Code from Claude.ai or the Claude app
- **Web Sessions**: Run Claude Code in the browser at claude.ai/code
- **Desktop App**: Standalone app for visual diff review and multiple sessions
- **Task List**: Persistent task tracking across context compactions
- **Prompt Suggestions**: Smart command suggestions based on context
- **Git Worktrees**: Isolated worktree branches for parallel work
- **Sandboxing**: OS-level filesystem and network isolation
- **Managed Settings**: Enterprise deployment via plist, Registry, or managed files
- **Configuration**: Customize behavior with JSON configuration files
---
## Planning Mode
Planning mode allows Claude to think through complex tasks before implementing them, creating a detailed plan that you can review and approve.
### What is Planning Mode?
Planning mode is a two-phase approach:
1. **Planning Phase**: Claude analyzes the task and creates a detailed implementation plan
2. **Implementation Phase**: After approval, Claude executes the plan
### When to Use Planning Mode
✅ Use planning mode for:
- Complex multi-file refactoring
- New feature implementations
- Architectural changes
- Database migrations
- Major API redesigns
❌ Don't use planning mode for:
- Simple bug fixes
- Formatting changes
- Single-file edits
- Quick queries
### Activating Planning Mode
**Slash command**:
```bash
/plan Implement user authentication system
```
**CLI flag**:
```bash
claude --permission-mode plan
```
**Set as default**:
```json
{
"permissions": {
"defaultMode": "plan"
}
}
```
**Keyboard shortcut**:
- `Shift + Tab` - Toggle permission modes (including plan)
- `Alt + M` - Toggle permission modes (Windows/Linux)
### Benefits of Planning Mode
- **Structured implementation approach**: Detailed breakdown of implementation steps
- **Review and approval**: Approve or adjust the plan before execution
- **Risk identification**: Identify potential issues before implementation
- **Clear phases**: Organized implementation phases and milestones
### Example: Feature Implementation
```
User: /plan Add a real-time notification system
Claude: I'll create a plan for implementing real-time notifications.
## Implementation Plan
### Phase 1: Infrastructure Setup
1. Install WebSocket libraries (socket.io)
2. Set up WebSocket server
3. Configure connection management
### Phase 2: Backend Implementation
4. Create notification service
5. Implement notification queue
6. Add database schema for notifications
7. Create API endpoints
### Phase 3: Frontend Integration
8. Set up WebSocket client
9. Create notification UI components
10. Implement notification state management
11. Add notification permissions
### Phase 4: Testing & Deployment
12. Write unit tests
13. Write integration tests
14. Update documentation
15. Deploy to staging
Estimated time: 3-4 hours
Files to modify: 12
New files to create: 8
Ready to proceed? (yes/no/modify plan)
User: yes
Claude: [Begins implementation following the plan]
```
### Plan Modifications
```
User: Modify the plan - skip the queue for now, we'll add it later
Claude: Updated plan:
[Shows modified plan with queue removed]
User: Looks good, proceed
Claude: [Implements modified plan]
```
### Planning Mode Configuration
Planning mode is activated via the CLI flag or slash command:
```bash
# Activate plan mode via CLI
claude --permission-mode plan
# Or use the /plan slash command inside the REPL
/plan Implement user authentication system
```
**Model alias for planning**: Use `opusplan` as a model alias to use Opus for planning and Sonnet for execution:
```bash
claude --model opusplan "design and implement the new API"
```
**Edit plan externally**: Press `Ctrl+G` to open the current plan in your external editor for detailed modifications.
> **v2.1.112 update**: Plan files are now named after the prompt that produced them (instead of random words), making them easier to browse and reuse.
> **v2.1.136 update — plan-mode write blocks are unconditional**: Plan mode now blocks all file writes, including when a matching `Edit(...)` rule exists in `permissions.allow`. Previously a permissive `Edit(...)` rule could let writes through in plan mode; that bypass is closed. If a workflow depended on the older behavior, exit plan mode (`Shift+Tab`) before editing.
---
## Extended Thinking
Extended thinking allows Claude to spend more time reasoning about complex problems before providing a solution.
### What is Extended Thinking?
Extended thinking is a deliberate, step-by-step reasoning process where Claude:
- Breaks down complex problems
- Considers multiple approaches
- Evaluates trade-offs
- Reasons through edge cases
### Activating Extended Thinking
**Keyboard shortcut**:
- `Option + T` (macOS) / `Alt + T` (Windows/Linux) - Toggle extended thinking
**Automatic activation**:
- Enabled by default for all models (Opus 5, Opus 4.8, Opus 4.7, Sonnet 4.6, Haiku 4.5)
- Opus 5 / Opus 4.8: Adaptive reasoning with effort levels: `low` (○), `medium` (◐), `high` (●), `xhigh`, `max`. The default is `high` on Opus 5 (v2.1.219), Opus 4.8 (v2.1.154), Opus 4.6, and Sonnet 4.6, and `xhigh` on Opus 4.7. `xhigh` is available on Opus 5, Opus 4.8, and Opus 4.7 (it falls back to `high` on Opus 4.6 / Sonnet 4.6). `max` works on Opus 5 and Opus 4.8/4.7/4.6 and Sonnet 4.6 (session-only). Haiku 4.5 has no effort levels. Opus 5, Opus 4.8, and Opus 4.7 have a 1M-token native context window (1M context fix landed in v2.1.117 — before that, `/context` miscounted Opus 4.7 against a 200K window and triggered premature autocompact). Since v2.1.129, `/context` shows its visualization in-UI only; the ASCII viz no longer leaks into the conversation context (~1.6k tokens saved per call), so `/context` is safe to invoke freely.
- Pro/Max subscribers on Opus 4.6 / Sonnet 4.6: default effort was raised from `medium` to `high` in v2.1.117.
- Other models: Fixed budget up to 31,999 tokens
**Configuration methods**:
- Toggle: `Alt+T` / `Option+T`, or via `/config`
- View reasoning: `Ctrl+O` (verbose mode)
- Set effort: `/effort` command or `--effort` flag
**Custom budget**:
```bash
export MAX_THINKING_TOKENS=1024
```
**Effort level** (supported on Opus 5, Opus 4.8, Opus 4.7, Opus 4.6, and Sonnet 4.6 — not Haiku 4.5):
```bash
export CLAUDE_CODE_EFFORT_LEVEL=high # low (○), medium (◐), high (●), xhigh (Opus 5/4.8/4.7), or max — default is high on Opus 5 and Opus 4.8
```
**CLI flag**:
```bash
claude --effort high "complex architectural review"
```
**Slash command**:
```
/effort high
```
> **Note:** The keyword "ultrathink" in prompts activates deep reasoning mode. Effort levels `low`, `medium`, `high`, and `max` are supported on Opus 5, Opus 4.8, Opus 4.7, Opus 4.6, and Sonnet 4.6 (Haiku 4.5 has none). `xhigh` is available on Opus 5, Opus 4.8, and Opus 4.7. The default effort is `high` on Opus 5, Opus 4.8 (and Opus 4.6 / Sonnet 4.6) and `xhigh` on Opus 4.7. Unlike Opus 4.8 and Opus 4.7, which pin their default effort on first run, Opus 5 has no such hold — a level you previously set carries over. The `/effort` menu also offers `ultracode`, which is **not** a model effort level — it sends `xhigh` and has Claude orchestrate dynamic workflows (session-only).
### Safety-Classifier Fallback on Opus 5
Separate from the [`fallbackModel` setting](#fallback-models-fallbackmodel) (which handles overload and unavailability), Claude Code applies a **category-based fallback** when a safety classifier flags a request. It requires Claude Code v2.1.219 or later.
- **Opus 5**: cybersecurity-flagged requests re-run on Opus 4.8. Biology-flagged requests end with a refusal instead, because Opus 5 runs its own biology classifiers with no fallback model. You get those refusals from the first flagged request.
- Workloads in offensive security or biology — penetration testing, Capture the Flag (CTF) exercises, and biology-adjacent codebases — trigger fallback frequently, often on the first request.
Why this matters here: if you run [security-review subagents](../04-subagents/README.md) or CTF/pentest work on Opus 5, expect silent re-runs on Opus 4.8 (different effort ceiling and context window) for cybersecurity prompts, and outright refusals for biology-adjacent ones. Pin a different model explicitly if that behavior gets in the way.
**Controlling the switch** — the `switchModelsOnFlag` setting (v2.1.170+, default `true`) decides whether the fallback happens silently. Set it to `false` and Claude Code pauses instead, letting you choose between switching models and editing the prompt. It appears in `/config` as **Switch models when a message is flagged**.
```json
{
"switchModelsOnFlag": false
}
```
### Benefits of Extended Thinking
- **Deep reasoning**: Spend more time thinking about complex problems
- **Better architectural decisions**: Make informed choices with comprehensive evaluation
- **Consideration of edge cases**: Think through potential issues and scenarios
- **Thorough analysis**: Structured approach to complex problem-solving
### Example: Architectural Decision
```
User: ultrathink Should we use microservices or a monolith for our e-commerce platform?
Claude: Let me think through this architectural decision carefully...
## Context Analysis
- Team size: 5 developers
- Expected traffic: 10k daily users initially
- Need to scale to 100k+ users
- Limited DevOps resources
- 6-month timeline to MVP
## Approach 1: Microservices
Pros:
+ Independent scaling
+ Technology flexibility
+ Team independence
+ Fault isolation
Cons:
- Operational complexity
- Network overhead
- Distributed tracing needed
- Requires DevOps expertise
- Slower initial development
## Approach 2: Monolith
Pros:
+ Simpler deployment
+ Faster initial development
+ Easier debugging
+ Lower operational overhead
+ Single codebase
Cons:
- Scaling challenges later
- All-or-nothing deployments
- Technology lock-in
- Potential for tight coupling
## Decision Factors
1. Team size (5 devs) → Favors monolith
2. DevOps resources (limited) → Favors monolith
3. Timeline (6 months) → Favors monolith
4. Future scale (100k users) → Achievable with monolith + caching
5. Complexity budget → Limited
## Recommendation
Start with a **modular monolith**:
- Single deployment unit
- Clear module boundaries
- Clean interfaces between modules
- Prepare for future extraction to microservices if needed
This gives you:
✓ Fast initial development
✓ Low operational complexity
✓ Option to extract services later
✓ Clear architecture
Migration path: When you reach 50k+ users or specific modules need independent scaling, extract those modules into microservices.
**Recommendation**: Implement a modular monolith with clean module boundaries.
This approach balances your current constraints (team size, timeline, DevOps resources) with future scalability needs. You can migrate specific modules to microservices later when there's a clear business need.
```
### Extended Thinking Configuration
Extended thinking is controlled via environment variables, keyboard shortcuts, and CLI flags:
```bash
# Set thinking token budget
export MAX_THINKING_TOKENS=16000
# Set effort level (Opus 5, Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 4.6): low (○), medium (◐), high (●), xhigh (Opus 5/4.8/4.7), or max — default is high on Opus 5 and Opus 4.8
export CLAUDE_CODE_EFFORT_LEVEL=high
```
Toggle during a session with `Alt+T` / `Option+T`, set effort with `/effort`, or configure via `/config`.
> **Lean system prompt (v2.1.154):** The lean system prompt is now the **default** for all models except Haiku, Sonnet, and Opus 4.7-and-earlier, reducing baseline token overhead on Opus 5 and Opus 4.8.
---
## Auto Mode
Auto Mode is a permission mode that uses a background safety classifier to review each action before execution. It allows Claude to work autonomously while blocking dangerous operations. It's available on all plans, but requires an eligible model (Claude Opus 5, Opus 4.6+, Sonnet 4.6+, or Fable 5 on the Anthropic API and Claude Platform on AWS; Opus 5, Sonnet 5, Opus 4.7, Opus 4.8, or Fable 5 on Bedrock, Vertex, Foundry, and signed-in Claude apps gateway sessions). On Team and Enterprise it is on by default — administrators can turn it off for the organization in managed settings.
### Requirements
Auto mode is available only when your account meets all of these requirements:
- **Plan**: all plans.
- **Organization**: on Team and Enterprise, auto mode is available by default. Administrators can turn it off for the organization by setting `permissions.disableAutoMode` to `"disable"` in managed settings.
- **Model**: on the Anthropic API and Claude Platform on AWS — Claude Opus 4.6 or later (Opus 5 included), Sonnet 4.6 or later, or Fable 5. On Amazon Bedrock, Google Cloud's Agent Platform (Vertex AI), Microsoft Foundry, and signed-in Claude apps gateway sessions — only Claude Sonnet 5, Opus 4.7 or later (Opus 5 included), and Fable 5. Older models — Sonnet 4.5, Opus 4.5, Haiku, and claude-3 models — are not supported on any provider.
- **Provider**: available by default on the Anthropic API, Claude Platform on AWS, Amazon Bedrock, Google Cloud's Agent Platform (Vertex AI), Microsoft Foundry, and signed-in Claude apps gateway sessions. In v2.1.158 through v2.1.206, auto mode was off on all of these except the Anthropic API and Claude Platform on AWS until you set `CLAUDE_CODE_ENABLE_AUTO_MODE=1`; v2.1.207 removed the requirement. The variable is still accepted for compatibility and has no effect from v2.1.207 onward.
- **Classifier**: adds extra token cost, except on Enterprise plans and Claude API accounts, where v2.1.278+ runs the check server-side at no charge
### Enabling Auto Mode
```bash
# Unlock auto mode with CLI flag (no longer required for Max subscribers on Opus 4.7 — access it directly)
claude --enable-auto-mode
# Then cycle to it with Shift+Tab in the REPL
```
> **v2.1.112 update**: Auto mode no longer requires the `--enable-auto-mode` flag. Max subscribers access it directly on Opus 4.7.
> **v2.1.158 update**: Auto mode became available on Bedrock, Vertex, and Foundry for Opus 4.7/4.8, gated behind `CLAUDE_CODE_ENABLE_AUTO_MODE=1`.
>
> **v2.1.207 update**: That opt-in requirement was removed. Auto mode is now available by default on Bedrock, Vertex AI, Microsoft Foundry, and signed-in Claude apps gateway sessions for Claude Sonnet 5, Opus 4.7, Opus 4.8, and Fable 5 (and Opus 5 from v2.1.219) — no flag or env var needed. Administrators can disable it with `disableAutoMode` in managed settings. `CLAUDE_CODE_ENABLE_AUTO_MODE` is still accepted for compatibility but has no effect from v2.1.207 onward.
Or set it as the default permission mode:
```bash
claude --permission-mode auto
```
Setting via config:
```json
{
"permissions": {
"defaultMode": "auto"
}
}
```
### How the Classifier Works
The background classifier evaluates each action using the following decision order:
1. **Allow/deny rules** -- Explicit permission rules are checked first
2. **Read-only/edits auto-approved** -- File reads and edits pass automatically
3. **Classifier** -- The background classifier reviews the action
4. **Fallback** -- Falls back to prompting after 3 consecutive or 20 total blocks
### Default Blocked Actions
Auto mode blocks the following by default:
| Blocked Action | Example |
|----------------|---------|
| Pipe-to-shell installs | `curl \| bash` |
| Sending sensitive data externally | API keys, credentials over network |
| Production deploys | Deploy commands targeting production |
| Mass deletion | `rm -rf` on large directories |
| IAM changes | Permission and role modifications |
| Force push to main | `git push --force origin main` |
> **More decisions moved to the classifier (v2.1.218)**: The classifier also decides removals targeting the filesystem root or home directory, such as `rm -rf /` and `rm -rf ~`, including when the removal sits inside command or process substitution. Before v2.1.218, the plain forms prompted for approval instead, and the substitution forms prompted in v2.1.208 through v2.1.217. The background-`&` and suspicious-Windows-path checks likewise no longer open permission dialogs — the classifier judges them.
### Default Allowed Actions
| Allowed Action | Example |
|----------------|---------|
| Local file operations | Read, write, edit project files |
| Declared dependency installs | `npm install`, `pip install` from manifest |
| Read-only HTTP | `curl` for fetching documentation |
| Pushing to current branch | `git push origin feature-branch` |
### Configuring Auto Mode
**Print default rules as JSON**:
```bash
claude auto-mode defaults
```
**Restore default auto-mode configuration** (v2.1.212), with a confirmation prompt (`--yes` to skip):
```bash
claude auto-mode reset
claude auto-mode reset --yes
```
**Configure trusted infrastructure** via the `autoMode.environment` managed setting for enterprise deployments. This allows administrators to define trusted CI/CD environments, deployment targets, and infrastructure patterns.
#### Extending defaults with `"$defaults"` (v2.1.118)
Since v2.1.118, `autoMode.allow`, `autoMode.soft_deny`, and `autoMode.environment` accept a `"$defaults"` token that **appends** your rules to the built-in list instead of replacing it. Before v2.1.118, any user-defined array silently clobbered the built-ins.
#### Unconditional blocks with `autoMode.hard_deny` (v2.1.136)
`autoMode.hard_deny` (v2.1.136+) is an array of classifier rules that block a class of actions **regardless of inferred user intent**. Use this for actions that must never run in auto mode — for example, `rm -rf` on root paths or `git push --force` on protected branches. Unlike `soft_deny`, hard-deny rules are not negotiable by the classifier.
```json
{
"autoMode": {
"hard_deny": ["Bash(rm -rf /:*)", "Bash(git push --force*)"]
}
}
```
**Before (replaces built-ins — pre-v2.1.118 behavior):**
```json
{
"autoMode": {
"allow": ["Bash(gh pr list:*)"]
}
}
```
**After (extends built-ins — v2.1.118+):**
```json
{
"autoMode": {
"allow": ["$defaults", "Bash(gh pr list:*)"],
"soft_deny": ["$defaults", "Bash(kubectl delete:*)"],
"environment": ["$defaults", "trusted-ci.internal"]
}
}
```
Use `"$defaults"` to keep the shipped baseline rules while layering organization- or project-specific additions on top.
#### Classifying every shell command with `autoMode.classifyAllShell` (v2.1.193)
`autoMode.classifyAllShell` (boolean, v2.1.193+) routes **all** Bash/PowerShell commands through the auto-mode classifier. Enable it when you want the classifier to inspect every shell command in the session.
```json
{
"autoMode": {
"classifyAllShell": true
}
}
```
The same release surfaces a **denial reason** when auto mode blocks an action — visible in the transcript, the denial toast, and the recently-denied list under `/permissions` (v2.1.193+).
#### Built-in intent-based protection (v2.1.183)
Separate from user-configured `hard_deny`, auto mode blocks the following destructive commands by default unless you explicitly asked for them this session:
- `git reset --hard`, `git checkout -- .`, `git clean -fd`, `git stash drop`
- `git commit --amend` (when the commit wasn't made by the agent this session)
- `terraform destroy`, `pulumi destroy`, `cdk destroy` (unless you asked for the specific stack)
This is built-in default protection driven by inferred intent — you don't need to add these to `hard_deny` yourself.
### Fallback Behavior
When the classifier is uncertain, auto mode falls back to prompting the user:
- After **3 consecutive** classifier blocks
- After **20 total** classifier blocks in a session
This ensures the user always retains control when the classifier cannot confidently approve an action.
### Seeding Auto-Mode-Equivalent Permissions (No Team Plan Required)
If you don't have a Team plan or want a simpler approach without the background classifier, you can seed your `~/.claude/settings.json` with a conservative baseline of safe permission rules. The script starts with read-only and local-inspection rules, then lets you opt into edits, tests, local git writes, package installs, and GitHub write actions only when you want them.
**File:** `09-advanced-features/setup-auto-mode-permissions.py`
```bash
# Preview what would be added (no changes written)
python3 09-advanced-features/setup-auto-mode-permissions.py --dry-run
# Apply the conservative baseline
python3 09-advanced-features/setup-auto-mode-permissions.py
# Add more capability only when you need it
python3 09-advanced-features/setup-auto-mode-permissions.py --include-edits --include-tests
python3 09-advanced-features/setup-auto-mode-permissions.py --include-git-write --include-packages
```
The script adds rules across these categories:
| Category | Examples |
|----------|---------|
| Core read-only tools | `Read(*)`, `Grep(*)`, `Agent(*)`, `WebSearch(*)`, `WebFetch(*)` |
| Local inspection | `Bash(git status:*)`, `Bash(git log:*)`, `Bash(git diff:*)`, `Bash(cat:*)` |
| Optional edits | `Edit(*)` |
| Optional test/build | `Bash(pytest:*)`, `Bash(python3 -m pytest:*)`, `Bash(cargo test:*)` |
| Optional git writes | `Bash(git add:*)`, `Bash(git commit:*)`, `Bash(git checkout:*)`, `Bash(git switch:*)`, `Bash(git stash:*)`, `Bash(git tag:*)` |
| Package managers | `Bash(npm ci:*)`, `Bash(npm install:*)`, `Bash(pip install:*)`, `Bash(pip3 install:*)` |
| Build & test | `Bash(make:*)`, `Bash(pytest:*)`, `Bash(go test:*)` |
| Common shell | `Bash(ls:*)`, `Bash(cat:*)`, `Bash(find:*)` |
| GitHub CLI | `Bash(gh pr view:*)`, `Bash(gh pr create:*)`, `Bash(gh issue list:*)` |
Dangerous operations (`rm -rf`, `sudo`, force push, `DROP TABLE`, `terraform destroy`, etc.) are intentionally excluded. The script is idempotent — running it twice won't duplicate rules.
---
## Background Tasks
Background tasks allow long-running operations to execute without blocking your conversation.
### What Are Background Tasks?
Background tasks run asynchronously while you continue working:
- Long test suites
- Build processes
- Database migrations
- Deployment scripts
- Analysis tools
**Basic Usage:**
```bash
User: Run tests in background
Claude: Started task bg-1234
/task list # Show all tasks
/task status bg-1234 # Check progress
/task show bg-1234 # View output
/task cancel bg-1234 # Cancel task
```
### Starting Background Tasks
```
User: Run the full test suite in the background
Claude: Starting tests in background (task-id: bg-1234)
You can continue working while tests run.
[You can continue the conversation]
User: Meanwhile, let's refactor the auth module
Claude: [Works on auth module while tests run in background]
[Later, when tests complete]
Claude: 📢 Background task bg-1234 completed:
✅ 245 tests passed
❌ 3 tests failed
View results: /task show bg-1234
```
### Managing Background Tasks
**List active tasks**:
```
User: /task list
Active background tasks:
1. [bg-1234] Running tests (50% complete, 2min remaining)
2. [bg-1235] Building Docker image (25% complete, 8min remaining)
3. [bg-1236] Deploying to staging (90% complete, 30sec remaining)
```
**Check task status**:
```
User: /task status bg-1234
Task bg-1234: Running tests
Status: In progress
Progress: 120/245 tests (49%)
Started: 2025-11-08 10:30:15
Estimated completion: 2025-11-08 10:34:22
```
**View task output**:
```
User: /task show bg-1234
[Shows live output from the test run]
```
**Cancel task**:
```
User: /task cancel bg-1234
Cancelled background task bg-1234
```
### Example: Parallel Development
```
User: Run the build in the background
Claude: Starting build... (task-id: bg-5001)
User: Also run the linter in background
Claude: Starting linter... (task-id: bg-5002)
User: While those run, let's implement the new API endpoint
Claude: [Implements API endpoint while build and linter run]
[10 minutes later]
Claude: 📢 Build completed successfully (bg-5001)
📢 Linter found 12 issues (bg-5002)
User: Show me the linter issues
Claude: [Shows linter output from bg-5002]
```
### Configuration
There is no `settings.json` block for background tasks. The feature is controlled by an environment variable:
```bash
# Turn background tasks off entirely
export CLAUDE_CODE_DISABLE_BACKGROUND_TASKS=true
```
Concurrency is not a background-task setting either — how many agents run at once is governed by `CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS` (default `20`).
---
## Monitor Tool (Event-Driven Streams)
> **New in v2.1.98**: The Monitor tool lets Claude watch a background command's stdout and react the moment a matching event appears — replacing polling loops and `sleep` for waiting on long-running processes.
Monitor attaches to any shell command that writes to stdout. Each stdout line from the command becomes a notification that wakes the session. Claude specifies the command; the harness streams output and delivers events as they fire. See the related [Background Tasks](#background-tasks) section for launching the underlying processes.
> **Deadlines (v2.1.271)**: every Monitor watch now carries a deadline — at most 30 minutes, or 10 minutes in single-prompt `-p` runs. When the deadline passes, Claude is notified so it can re-arm the watch. There is no longer a way to arm a watch that never times out.
### Why It Matters
Polling with `/loop` or `sleep` burns a full API round-trip every cycle, whether or not anything changed. Monitor stays silent until an event fires, consuming **zero tokens** while the command is quiet. When an event does occur, Claude reacts immediately — no delayed discovery waiting for the next poll tick. For anything that runs longer than a few minutes, this is both cheaper and faster than poll loops.
### Two Common Patterns
**Stream filters** watch continuous output from a long-running source. The command runs forever; every matching line is an event.
```bash
tail -f /var/log/app.log | grep --line-buffered "ERROR"
```
**Poll-and-emit filters** check a source periodically and only emit when something changes. Use this for APIs, databases, or anything without a native stream.
```bash
last=$(date -u +%Y-%m-%dT%H:%M:%SZ)
while true; do
gh api "repos/owner/repo/issues/123/comments?since=$last" || true
last=$(date -u +%Y-%m-%dT%H:%M:%SZ)
sleep 30
done
```
### Concrete Example
"Start my dev server and monitor it for errors." Claude launches the server as a background task, attaches a Monitor filter (`tail -F server.log | grep --line-buffered -E "ERROR|FATAL"`), and the session goes quiet. The moment an error line appears in the log, Claude wakes up, reads the error, and can react — restart the server, fix the bug, or surface it to you — without you having to check in.
> **Warning**: When piping into `grep`, **always** use `grep --line-buffered`. Without it, grep buffers stdout in 4KB chunks, which can delay events by minutes on low-traffic streams. This is the #1 way Monitor breaks in practice — if your filter seems silent when it shouldn't be, check for the `--line-buffered` flag first.
---
## Dynamic Workflows
> **New in v2.1.154**
Dynamic workflows let Claude orchestrate tens to hundreds of background [subagents](../04-subagents/README.md) **deterministically** — fan-out, pipelines, and parallel stages encoded in a script rather than left to the model's improvisation. Where a single agent holds one context window, a workflow decomposes a task across many agents and recombines their results.
As of v2.1.219, dynamic workflows default to a **medium size guideline (aim for fewer than 10 agents)** — or to **small** when you are signed in on a Pro plan, since v2.1.271. Pick another size — or unrestricted — via **Dynamic workflow size** in `/config`, or set the `workflowSizeGuideline` key in your settings file. The running-workflow status line shows the active size and points to `/config` for changing it.
### When to Use Them
- **Comprehensive coverage** — audit or review across many files/dimensions in parallel.
- **Confidence** — generate independent perspectives, then adversarially verify findings before committing.
- **Scale beyond one context** — large migrations, broad sweeps, or research that no single context can hold.
For a one-off task you already understand, a single agent (or a direct edit) is still the right tool — workflows pay off when the work fans out.
### Launching and Viewing
- **Launch**: ask Claude to create a workflow for the task (e.g. "run a workflow to review every file in `src/`"). Claude authors the orchestration script and runs it in the background.
- **View**: the `/workflows` command shows running and completed workflow runs with live progress.
- **`ultracode`**: selecting `ultracode` in the `/effort` menu turns this on for the session — it sends `xhigh` to the model *and* has Claude orchestrate dynamic workflows by default. It is session-only and not accepted in the settings file. (As of v2.1.160 the trigger keyword is `ultracode`; the bare word "workflow" no longer triggers a run.)
Workflows build on the subagent model — see [Subagents](../04-subagents/README.md) for how individual agents are defined and scoped.
---
## Scheduled Tasks
Scheduled Tasks let you run prompts automatically on a recurring schedule or as one-time reminders. Tasks are session-scoped — they run while Claude Code is active and are cleared when the session ends. Available since v2.1.72+.
> **Marketed as "Routines" on claude.com (2026-05-14)**: Anthropic's product blog introduces this surface as **Routines**. The CLI command stays `/schedule`; this guide uses the original "Scheduled Tasks" naming for continuity. If you see "Routines" in claude.com docs or the desktop app, it refers to the same feature.
### The `/loop` command
```bash
# Explicit interval
/loop 5m check if the deployment finished
# Natural language
/loop check build status every 30 minutes
```
Standard 5-field cron expressions are also supported for precise scheduling.
### One-time reminders
Set reminders that fire once at a specific time:
```
remind me at 3pm to push the release branch
in 45 minutes, run the integration tests
```
### Managing scheduled tasks
| Tool | Description |
|------|-------------|
| `CronCreate` | Create a new scheduled task |
| `CronList` | List all active scheduled tasks. Since v2.1.136, output also includes the qualifier(s) and the scheduled prompt body, so you can audit what each cron will run without opening it. |
| `CronDelete` | Remove a scheduled task |
**Limits and behavior**:
- Up to **50 scheduled tasks** per session
- Session-scoped — cleared when the session ends
- Recurring tasks auto-expire after **3 days**
- Tasks only fire while Claude Code is running — no catch-up for missed fires
### Behavior details
| Aspect | Detail |
|--------|--------|
| **Recurring jitter** | Up to 10% of the interval (max 15 minutes) |
| **One-shot jitter** | Up to 90 seconds on :00/:30 boundaries |
| **Missed fires** | No catch-up — skipped if Claude Code was not running |
| **Persistence** | Not persisted across restarts |
### Cloud Scheduled Tasks
Use `/schedule` to create Cloud scheduled tasks that run on Anthropic infrastructure:
```
/schedule daily at 9am run the test suite and report failures
```
Cloud scheduled tasks persist across restarts and do not require Claude Code to be running locally.
### Disabling scheduled tasks
```bash
export CLAUDE_CODE_DISABLE_CRON=1
```
> **`/schedule` auto-disabled by API-key tiers (v2.1.139)**: Cloud `/schedule` is silently unavailable when any of `ANTHROPIC_API_KEY`, `ANTHROPIC_AUTH_TOKEN`, or `apiKeyHelper` is set — even if you are also logged in with claude.ai. The same condition disables [Remote Control](#disabling-remote-control-disableremotecontrol-v21128), claude.ai MCP connectors, and notification preferences. Unset the API key (or run on a Pro/Max OAuth tier) to use `/schedule`. Local `CronCreate` is unaffected.
### Example: monitoring a deployment
```
/loop 5m check the deployment status of the staging environment.
If the deploy succeeded, notify me and stop looping.
If it failed, show the error logs.
```
> **Tip**: Scheduled tasks are session-scoped. For persistent automation that survives restarts, use CI/CD pipelines, GitHub Actions, or Desktop App scheduled tasks instead.
---
## Permission Modes
Permission modes control what actions Claude can take without explicit approval.
### Available Permission Modes
| Mode | Behavior |
|---|---|
| `manual` | Read files only; prompts for all other actions. Renamed from `default` in v2.1.200 — `default` is still accepted as an alias |
| `acceptEdits` | Read and edit files; prompts for commands |
| `plan` | Read files only (research mode, no edits) |
| `auto` | All actions with background safety classifier checks. Requires an eligible model (Opus 5, Sonnet 5, Opus 4.7/4.8, or Fable 5 on most providers) and provider — available on all plans, see [Auto Mode](#auto-mode) |
| `bypassPermissions` | All actions, no permission checks (dangerous) |
| `dontAsk` | Only pre-approved tools execute; all others denied |
> **Note**: The interactive default mode was renamed from `default` to **Manual** in v2.1.200 (across the CLI, `--help`, VS Code, and JetBrains), and a grey ⏸ badge appears in the footer while it is active (v2.1.203). Both `--permission-mode manual` and `--permission-mode default` work, as do `"defaultMode": "manual"` and `"defaultMode": "default"` in settings. Note that the settings key is `permissions.defaultMode` — there is no `permissions.mode` key, so the examples below use the canonical spelling.
Cycle through modes with `Shift+Tab` in the CLI. Set a default with the `--permission-mode` flag or the `permissions.defaultMode` setting.
> **Plan mode defers shell commands to the classifier (v2.1.218)**: When [auto mode](#auto-mode) is available and the `useAutoModeDuringPlan` setting is on — which it is by default — the classifier reviews shell commands during planning instead of prompting you. Approved commands run, and rejected ones are blocked. Plan mode still blocks file writes unconditionally.
As of v2.1.160, even `acceptEdits` prompts before writing shell-startup files (`.zshenv`, `.zlogin`, `.bash_login`, `~/.config/git/`) and code-executing build configs (`.npmrc`, `.yarnrc*`, `bunfig.toml`, `.bazelrc`, `.pre-commit-config.yaml`, `.devcontainer/`, …), which could otherwise lead to unintended command execution.
> **`--dangerously-skip-permissions` extended path coverage (v2.1.121, v2.1.126)**: The `--dangerously-skip-permissions` CLI flag (and equivalent `bypassPermissions` mode) now bypasses prompts for writes to a much broader allowlist — `.claude/skills/`, `.claude/agents/`, `.claude/commands/`, `.claude/`, `.git/`, `.vscode/`, and shell config files. Catastrophic removal commands (`rm -rf /`, etc.) still prompt in this mode (in auto mode they are judged by the classifier instead — see [Auto Mode](#auto-mode)). Treat the flag as a sharper tool than before; use it only in throwaway sandboxes.
> **Windows shell detection (v2.1.120, v2.1.126)**: Git for Windows / Git Bash is no longer required. When Git Bash is absent, Claude Code uses PowerShell as the shell tool. From v2.1.126 PowerShell is the *primary* shell when the PowerShell tool is enabled, and detection covers PowerShell 7 installed via the Microsoft Store, MSI without PATH, or as a `.NET global tool`.
> **PowerShell tool enabled by default on Windows for Bedrock/Vertex/Foundry (v2.1.143)**: As of v2.1.143, the PowerShell tool is **enabled by default on Windows** for Bedrock, Vertex, and Foundry users. Claude Code invokes PowerShell with `-ExecutionPolicy Bypass` so scripts run even if the system policy is `Restricted`. To make Claude Code honor the system execution policy, set `CLAUDE_CODE_POWERSHELL_RESPECT_EXECUTION_POLICY=1`. To disable the PowerShell tool entirely, set `CLAUDE_CODE_USE_POWERSHELL_TOOL=0`.
### Activation Methods
**Keyboard shortcut**:
```bash
Shift + Tab # Cycle through all 6 modes
```
**Slash command**:
```bash
/plan # Enter plan mode
```
**CLI flag**:
```bash
claude --permission-mode plan
claude --permission-mode auto
```
**Setting**:
```json
{
"permissions": {
"defaultMode": "auto"
}
}
```
### Permission Mode Examples
#### Default Mode
Claude asks for confirmation on significant actions:
```
User: Fix the bug in auth.ts
Claude: I need to modify src/auth.ts to fix the bug.
The change will update the password validation logic.
Approve this change? (yes/no/show)
```
#### Plan Mode
Review implementation plan before execution:
```
User: /plan Implement user authentication system
Claude: I'll create a plan for implementing authentication.
## Implementation Plan
[Detailed plan with phases and steps]
Ready to proceed? (yes/no/modify)
```
#### Accept Edits Mode
Automatically accept file modifications:
```
User: acceptEdits
User: Fix the bug in auth.ts
Claude: [Makes changes without asking]
```
### Use Cases
**Code Review**:
```
User: claude --permission-mode plan
User: Review this PR and suggest improvements
Claude: [Reads code, provides feedback, but cannot modify]
```
**Pair Programming**:
```
User: claude --permission-mode default
User: Let's implement the feature together
Claude: [Asks for approval before each change]
```
**Automated Tasks**:
```
User: claude --permission-mode acceptEdits
User: Fix all linting issues in the codebase
Claude: [Auto-accepts file edits without asking]
```
---
## Headless Mode
Print mode (`claude -p`) allows Claude Code to run without interactive input, perfect for automation and CI/CD. This is the non-interactive mode, replacing the older `--headless` flag.
### What is Print Mode?
Print mode enables:
- Automated script execution
- CI/CD integration
- Batch processing
- Scheduled tasks
### Running in Print Mode (Non-Interactive)
```bash
# Run specific task
claude -p "Run all tests"
# Process piped content
cat error.log | claude -p "Analyze these errors"
# CI/CD integration (GitHub Actions)
- name: AI Code Review
run: claude -p "Review PR"
```
### Additional Print Mode Usage Examples
```bash
# Run a specific task with output capture
claude -p "Run all tests and generate coverage report"
# With structured output
claude -p --output-format json "Analyze code quality"
# With input from stdin
echo "Analyze code quality" | claude -p "explain this"
```
### Example: CI/CD Integration
**GitHub Actions**:
```yaml
# .github/workflows/code-review.yml
name: AI Code Review
on: [pull_request]
jobs:
review:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- name: Install Claude Code
run: npm install -g @anthropic-ai/claude-code
- name: Run Claude Code Review
env:
ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
run: |
claude -p --output-format json \
--max-turns 3 \
"Review this PR for:
- Code quality issues
- Security vulnerabilities
- Performance concerns
- Test coverage
Output results as JSON" > review.json
- name: Post Review Comment
uses: actions/github-script@v7
with:
script: |
const fs = require('fs');
const review = JSON.parse(fs.readFileSync('review.json', 'utf8'));
github.rest.issues.createComment({
issue_number: context.issue.number,
owner: context.repo.owner,
repo: context.repo.repo,
body: JSON.stringify(review, null, 2)
});
```
### Print Mode Configuration
Print mode (`claude -p`) supports several flags for automation:
```bash
# Limit autonomous turns
claude -p --max-turns 5 "refactor this module"
# Structured JSON output
claude -p --output-format json "analyze this codebase"
# With schema validation
claude -p --json-schema '{"type":"object","properties":{"issues":{"type":"array"}}}' \
"find bugs in this code"
# Disable session persistence
claude -p --no-session-persistence "one-off analysis"
```
### Safe Mode (troubleshooting)
`--safe-mode` (and the `CLAUDE_CODE_SAFE_MODE` environment variable, e.g. `CLAUDE_CODE_SAFE_MODE=1`) starts Claude Code with **all customizations disabled** — CLAUDE.md, plugins, skills, hooks, and MCP servers are all turned off.
```bash
# Launch with every customization disabled
claude --safe-mode
# Equivalent via environment variable
CLAUDE_CODE_SAFE_MODE=1 claude
```
It is a troubleshooting tool: when a custom config is causing problems, launch in safe mode to isolate whether the issue is in your setup or in Claude Code itself.
---
## Session Management
Manage multiple Claude Code sessions effectively.
### Session Management Commands
| Command | Description |
|---------|-------------|
| `/resume` | Resume a conversation by ID or name |
| `/rename` | Name the current session |
| `/fork [prompt]` | Copy the conversation into a new independent background session and keep working here (v2.1.212+) |
| `/subtask ` | Spawn a forked subagent that inherits the full conversation and reports its result back here (v2.1.212+) |
| `/branch [name]` | Switch into a copy of the conversation at this point, preserving the original |
| `claude -c` | Continue most recent conversation |
| `claude -r "session"` | Resume session by name or ID |
### Resuming Sessions
**Continue last conversation**:
```bash
claude -c
```
**Resume a named session**:
```bash
claude -r "auth-refactor" "finish this PR"
```
**Rename the current session** (inside the REPL):
```
/rename auth-refactor
```
> **v2.1.212 update**: Typing `/resume` (with no arguments) in the agent view now opens a picker of past sessions — including sessions removed from the visible list — and resumes the chosen one as a background session.
### Forking and Branching Sessions
Three commands make copies of a conversation, and they differ in *where the copy runs*:
`/subtask ` spawns a forked subagent that inherits the full conversation and works on the task while you keep working — its own row in `claude agents`, with the result returned to your conversation when it finishes:
```
/subtask Investigate why the auth tests are flaky
```
`/fork [prompt]` copies the conversation into a new **background session** instead. The copy starts with everything up to now and runs independently — nothing comes back to this conversation:
```
/fork Try the OAuth approach end to end
```
To switch into a copy yourself rather than delegating at all, use `/branch [name]`, which preserves the original and lets you return to it with `/resume`:
```
/branch try-oauth-instead
```
> **Note**: `/fork` and `/subtask` swapped roles in **v2.1.212**. Before v2.1.161 `/fork` was an alias for `/branch`; from v2.1.161 to v2.1.211 it started a forked subagent — the behavior now carried by `/subtask`. When agent view is turned off, `/subtask` is unavailable and `/fork` retains the forked-subagent behavior.
Or fork from the CLI:
```bash
claude --resume auth-refactor --fork-session "try OAuth instead"
```
### Session Persistence
Sessions are automatically saved and can be resumed:
```bash
# Continue last conversation
claude -c
# Resume specific session by name or ID
claude -r "auth-refactor"
# Resume and fork for experimentation
claude --resume auth-refactor --fork-session "alternative approach"
```
### Usage-Limit Auto-Continue (v2.1.234)
As of **v2.1.234**, a session blocked on a claude.ai usage limit auto-continues once that limit resets — no manual re-prompt needed. Toggle this from `/config` under "Continue automatically at usage limit."
### Session Recap (v2.1.108)
When you return to a session after being away, Claude can show a brief recap of what was accomplished. This is enabled by default for users with telemetry disabled (Bedrock, Vertex, Foundry users).
> **OTEL telemetry — re-enable feedback survey (v2.1.136+)**: Organizations capturing OpenTelemetry data can re-enable Anthropic's session-quality survey by setting `CLAUDE_CODE_ENABLE_FEEDBACK_SURVEY_FOR_OTEL=1`. The survey is off by default in OTEL deployments because it was previously redirected away from telemetry pipelines.
> **OTEL telemetry — `assistant_response` log event (v2.1.193+)**: Claude Code emits a `claude_code.assistant_response` OpenTelemetry log event carrying the model's response text, letting OTEL pipelines capture what Claude said alongside the existing tool/event telemetry.
**Control recap behavior:**
```bash
/recap # manually trigger a recap
/config # toggle auto-recap on/off
```
Or via environment variable:
```bash
CLAUDE_CODE_ENABLE_AWAY_SUMMARY=0 claude # disable recap
CLAUDE_CODE_ENABLE_AWAY_SUMMARY=1 claude # force enable recap
```
---
## Cross-Session Messaging
> **Added in v2.1.224**, extended through v2.1.239. Available on **macOS, Linux, and (since v2.1.239) Windows**.
Sessions used to be islands. Cross-session messaging lets one Claude Code session talk to
another — including sessions on your other machines and your cloud sessions — so you can
hand a question to the session that already has the right context loaded instead of
re-explaining it.
### Discovering Sessions
`ListAgents` lists everything you can address: subagents you spawned, other local sessions
on this machine, your cloud sessions, and (when Remote Control is connected) sessions on
your other machines. Each row is labeled by kind, and since v2.1.229 rows also carry
`offline` and `cloud` labels so you can tell a reachable session from a dormant one.
The **name in each row is the address** — that is what you send to.
### Sending a Message
`SendMessage` takes the target and the message:
```text
SendMessage({ to: "", message: "What did you conclude about the retry logic?" })
```
Since v2.1.232, a bare name is enough — you no longer need to append a disambiguating ref
unless two rows genuinely share the same name.
### Waiting for a Session to Go Idle (`notify_when_idle`, v2.1.236)
When the session you are messaging is mid-task, you usually want to know when it finishes
rather than poll it. `SendMessage` takes a `notify_when_idle` input for exactly that:
```text
SendMessage({
to: "auth-refactor",
message: "ping me when the migration finishes",
notify_when_idle: true
})
```
It is **opt-in and one-shot** — the target session sends a single notice the next time it
goes idle, and then the subscription is done. There is no polling loop and no repeated
notification if the session goes busy and idle again.
Two related v2.1.239 changes: `ListAgents` now also reports **the session's own name** (so
a session can tell others how to address it) alongside its live teammates, and cross-session
messaging became available on **Windows**.
### `@`-Mention Shorthand (v2.1.232)
Instead of calling the tool explicitly, you can `@`-mention a session directly in your
prompt to reach it:
```text
@auth-refactor did the migration tests pass?
```
### Controlling What Arrives: `crossSessionInbound`
Inbound messages are governed by the `crossSessionInbound` setting (v2.1.224+):
| Value | Behavior |
|---|---|
| `"accept"` | Inbound messages are delivered to Claude in this session |
| `"hold"` | You see a notice that a message arrived; it is not delivered |
| `"refuse"` | Inbound messages are dropped |
The values form a ladder — `accept < hold < refuse` — and **project and local settings
apply only when they are stricter** than the user-scope value. A project can tighten
inbound delivery, never loosen it. Since v2.1.232 the setting also has a `/config` row,
"Messages from your other sessions."
### Reach and Limits
- Local sessions on the same machine, plus your cloud sessions.
- Remote Control sessions on your other machines, addressable by name (v2.1.225).
- A cloud session **receives** your message but cannot message a local session back yet —
read its answer in its own transcript.
- macOS and Linux from v2.1.224; Windows since v2.1.239.
---
## Interactive Features
### Keyboard Shortcuts
Claude Code supports keyboard shortcuts for efficiency. Here's the complete reference from official docs:
| Shortcut | Description |
|----------|-------------|
| `Ctrl+C` | Cancel current input/generation |
| `Ctrl+D` | Exit Claude Code |
| `Ctrl+G` | Edit plan in external editor |
| `Ctrl+L` | Redraw the screen (repaint only — the double-press `/clear` shortcut was removed in v2.1.238) |
| `Ctrl+O` | Toggle verbose output (view reasoning) |
| `Ctrl+R` | Reverse search history. Defaults to **all prompts across all projects** (v2.1.129+); press `Ctrl+S` inside the picker to narrow to the current project. Earlier versions defaulted to project-only. |
| `Ctrl+T` | Toggle task list view |
| `Ctrl+B` | Background running tasks |
| `Esc+Esc` | Rewind code/conversation |
| `Shift+Tab` / `Alt+M` | Toggle permission modes |
| `Option+P` / `Alt+P` | Switch model |
| `Option+T` / `Alt+T` | Toggle extended thinking |
| `Option+O` / `Alt+O` | Toggle fast mode (`/fast`) |
| `Ctrl+X` `Ctrl+K` | Stop all background subagents |
| `Ctrl+Enter` (or `Ctrl+X` `Ctrl+S`) | Send now — interrupt the current turn and send every queued message at once. Sent and queued messages stay gray until the model receives them (v2.1.275) |
| `Ctrl+S` | Stash the current prompt; press again to restore it |
| `Ctrl+_` | Undo the last edit to the prompt input |
| `:` | Type `:` at the start of a word to open emoji shortcode completion, e.g. `:heart:` (v2.1.217+) |
**Line Editing (standard readline shortcuts):**
| Shortcut | Action |
|----------|--------|
| `Ctrl + A` | Move to line start |
| `Ctrl + E` | Move to line end |
| `Ctrl + K` | Cut to end of line |
| `Ctrl + U` | Cut to start of line |
| `Ctrl + W` | Delete word backward |
| `Ctrl + Y` | Paste (yank) |
| `Tab` | Autocomplete |
| `↑ / ↓` | Command history |
### Accessibility
Screen reader mode (v2.1.208+) switches the CLI to a plain-text rendering mode designed for screen readers. Enable it with the CLI flag, an environment variable, or a settings key:
```bash
claude --ax-screen-reader
```
```bash
export CLAUDE_AX_SCREEN_READER=1
```
```json
{
"axScreenReader": true
}
```
### Customizing keybindings
Create custom keyboard shortcuts by running `/keybindings`, which opens `~/.claude/keybindings.json` for editing (v2.1.18+).
**Configuration format**:
```json
{
"$schema": "https://www.schemastore.org/claude-code-keybindings.json",
"bindings": [
{
"context": "Chat",
"bindings": {
"ctrl+e": "chat:externalEditor",
"ctrl+u": null,
"ctrl+k ctrl+s": "chat:stash"
}
},
{
"context": "Confirmation",
"bindings": {
"ctrl+a": "confirmation:yes"
}
}
]
}
```
Set a binding to `null` to unbind a default shortcut.
### Available contexts
Keybindings are scoped to specific UI contexts:
| Context | Key Actions |
|---------|-------------|
| **Chat** | `submit`, `cancel`, `cycleMode`, `modelPicker`, `thinkingToggle`, `undo`, `externalEditor`, `stash`, `imagePaste` |
| **Confirmation** | `yes`, `no`, `previous`, `next`, `nextField`, `cycleMode`, `toggleExplanation` |
| **Global** | `interrupt`, `exit`, `toggleTodos`, `toggleTranscript` |
| **Autocomplete** | `accept`, `dismiss`, `next`, `previous` |
| **HistorySearch** | `search`, `previous`, `next` |
| **Settings** | Context-specific settings navigation |
| **Tabs** | Tab switching and management |
| **Help** | Help panel navigation |
There are 18 contexts total including `Transcript`, `Task`, `ThemePicker`, `Attachments`, `Footer`, `MessageSelector`, `DiffDialog`, `ModelPicker`, and `Select`.
### Chord support
Keybindings support chord sequences (multi-key combinations):
```
"ctrl+k ctrl+s" → Two-key sequence: press ctrl+k, then ctrl+s
"ctrl+shift+p" → Simultaneous modifier keys
```
**Keystroke syntax**:
- **Modifiers**: `ctrl`, `alt` (or `opt`), `shift`, `meta` (or `cmd`)
- **Uppercase implies Shift**: `K` is equivalent to `shift+k`
- **Special keys**: `escape`, `enter`, `return`, `tab`, `space`, `backspace`, `delete`, arrow keys
### Reserved and conflicting keys
| Key | Status | Notes |
|-----|--------|-------|
| `Ctrl+C` | Reserved | Cannot be rebound (interrupt) |
| `Ctrl+D` | Reserved | Cannot be rebound (exit) |
| `Ctrl+B` | Terminal conflict | tmux prefix key |
| `Ctrl+A` | Terminal conflict | GNU Screen prefix key |
| `Ctrl+Z` | Terminal conflict | Process suspend |
> **Tip**: If a shortcut does not work, check for conflicts with your terminal emulator or multiplexer.
### Tab Completion
Claude Code provides intelligent tab completion:
```
User: /rew
→ /rewind
User: /plu
→ /plugin
User: /plugin
→ /plugin install
→ /plugin enable
→ /plugin disable
```
### Command History
Access previous commands:
```
User: <↑> # Previous command
User: <↓> # Next command
User: Ctrl+R # Search history
(reverse-i-search)`test': run all tests
```
### Multi-line Input
For complex queries, use multi-line mode:
```bash
User: \
> Long complex prompt
> spanning multiple lines
> \end
```
**Example:**
```
User: \
> Implement a user authentication system
> with the following requirements:
> - JWT tokens
> - Email verification
> - Password reset
> - 2FA support
> \end
Claude: [Processes the multi-line request]
```
### Inline Editing
Edit commands before sending:
```
User: Deploy to prodcutionuction
[Edit in-place before sending]
```
### Vim Mode
Enable Vi/Vim keybindings for text editing:
**Activation**:
- Enable via `/config` (toggle "Editor / Vim mode") or in `~/.claude/settings.json` under `editorMode: "vim"`. The standalone `/vim` slash command was removed (see [issue #43370](https://github.com/anthropics/claude-code/issues/43370)); vim mode is now configuration-driven.
- Mode switching with `Esc` for NORMAL, `i/a/o` for INSERT, `v` for VISUAL, `V` for VISUAL-LINE (v2.1.118+)
**Navigation keys**:
- `h` / `l` - Move left/right
- `j` / `k` - Move down/up
- `w` / `b` / `e` - Move by word
- `0` / `$` - Move to line start/end
- `gg` / `G` - Jump to start/end of text
**Text objects**:
- `iw` / `aw` - Inner/around word
- `i"` / `a"` - Inner/around quoted string
- `i(` / `a(` - Inner/around parentheses
**Visual modes (v2.1.118+)**:
| Key | Mode | Behavior |
|-----|------|----------|
| `v` | Visual | Character-wise selection with visual feedback; extend with motion keys |
| `V` | Visual-line | Line-wise selection; always selects whole lines |
| `y` | Yank | Copy the current visual selection |
| `d` / `x` | Delete | Delete the current visual selection |
| `c` | Change | Delete selection and enter INSERT mode |
| `Esc` | Exit | Return to NORMAL mode |
Visual selections are highlighted in the input field so you can see exactly what will be yanked, deleted, or changed before you commit the operator.
### Bash Mode
Execute shell commands directly with `!` prefix:
```bash
! npm test
! git status
! cat src/index.js
```
Use this for quick command execution without switching contexts.
**Since v2.1.193:** bash mode (`!`) has live file-path autocomplete, so paths complete as you type your shell command without leaving the prompt.
**Since v2.1.186:** the output of a `!` command is now automatically sent to Claude, which responds to it. To keep the previous behavior where the output is only added to context without a response, set `"respondToBashCommands": false` in `settings.json`.
---
## Output Styles
Output styles change **how** Claude responds, not what it knows. They modify the system prompt to set role, tone, and default response format. Reach for one when you keep re-prompting for the same voice every turn, or when you want Claude acting as something other than a software engineer.
For instructions about your project or codebase, use [CLAUDE.md](../02-memory/) instead — that is a different mechanism with different tradeoffs.
### Built-in styles
| Style | Behavior |
|-------|----------|
| **Default** | The standard system prompt, tuned for completing software engineering tasks efficiently |
| **Proactive** | Claude executes immediately and makes reasonable assumptions instead of pausing for routine decisions. Stronger autonomous-execution guidance than auto mode, but it does **not** change your permission mode — you still see permission prompts |
| **Explanatory** | Adds educational "Insights" between steps, explaining implementation choices and codebase patterns |
| **Learning** | Collaborative learn-by-doing. Claude shares insights *and* leaves `TODO(human)` markers for you to implement small, strategic pieces yourself |
| **Concise** (v2.1.237) | Claude leads with the result and skips preamble and narration. Thoroughness is unchanged — only the framing around the answer is dropped. Select it in `/config` → Output style, or set `"outputStyle": "Concise"` |
### Selecting a style
Run `/config` and choose **Output style**, or run `/output-style
