Interactive chat mode with slash commands for controlling execution, saving work, and managing sessions.
Interactive mode starts a chat session where you compose prompts, switch runtime modes (plan, brainstorm, debug, etc.), and manage session state via slash commands.
amplifier run --mode chat
# Or simply with no prompt:
amplifier runWithin an interactive session, you compose prompts and issue slash commands. Runtime mode switching is provided by the amplifier-bundle-modes bundle (/mode <name>, /modes, /mode off); other commands are app-CLI built-ins documented below.
| Command | Purpose | Notes |
|---|---|---|
/save [file] |
Save transcript | To session directory |
/clear |
Clear history | Keeps session active |
/status |
Show session info | Active mode, message count, tools count |
/tools |
List tools | Shows loaded capabilities |
/config |
Show configuration | Full mount plan |
/help |
List commands | Quick reference |
/exit |
Exit this session | No arguments |
/quit |
Exit this session | Alias for /exit; no arguments |
| Ctrl-C | Interrupt execution | During an active turn |
/modes |
List available modes (from amplifier-bundle-modes) |
|
/mode <name> |
Activate a runtime mode | Mode name or shortcut |
/mode off |
Deactivate the current mode |
The app-CLI built-ins shown above are available in every interactive session.
The /mode* commands come from amplifier-bundle-modes when that bundle is
composed into your active configuration (which is the case for every standard
amplifier bundle).
Typing a leading / opens command and known argument-choice menus
automatically. Tab has the same prefix behavior whether a menu is already
open or not: it accepts one match, or extends every matching value's literal
shared prefix and leaves the menu unselected. It does not cycle an unselected
menu. Use Up/Down or Shift-Tab to preview a choice; Tab, Enter, or
Space then accepts that explicit choice without submitting the prompt.
For command names, unselected Enter accepts an exact match (case
insensitively) or one unique match. An ambiguous prefix remains open with a
short “Type more or choose” hint. For arguments, suggestions are advisory:
unselected Enter always submits the exact current line, including partial,
custom, quoted, multiword, or trailing-space arguments. This is also true
after opening an argument menu manually with automatic popups disabled.
For example, if /product-council, /product-council-here, and /provider
are available, /p then Tab becomes /pro. A second Tab leaves /pro
unselected because the names are still ambiguous. A real /p alias is an
exact shorter command and prevents that extension; Enter accepts it with a
space. The available command catalog is conditional on the built-ins and
currently mounted modes, skills, and providers.
To disable only the automatic popup, add this to ~/.amplifier/settings.yaml:
ui: {slash_popup: {enabled: false}}The default is enabled; this disables every automatic command and argument
popup (manual Tab remains available). enabled must be the YAML boolean false, not the
string "false". The setting takes effect for fresh and resumed interactive
sessions started after the change; restart the session to apply it.
This experiment opens menus on text insertion, not deletion: Backspace
closes an open menu until you type again or press Tab.
Runtime modes are runtime behavior overlays that modify how the assistant operates — restricting tools, contributing context, agents, or skills, and gating destructive actions. Plan mode is a common built-in:
> /mode plan
✓ Mode 'plan' activated
> Analyze auth system and suggest improvements
[AI analyzes without changes — writes blocked]
> /mode off
✓ Mode cleared
> Implement the improvements
[AI now makes changes]See amplifier-bundle-modes for the full list of built-in modes (plan, careful, explore, etc.) and how to author your own.
/save — Persist Transcript:
> /save auth_refactor.json
✓ Saved to ~/.amplifier/projects/<project>/sessions/<session-id>/auth_refactor.jsonSaves: All messages, session config, timestamp. Location: Session directory ~/.amplifier/projects/<project>/sessions/<session-id>/.
/clear — Reset Context: Clears conversation history, session stays active. Use when switching topics or context grows too large.
/status — Session Information:
> /status
Mode: plan | Messages: 42 | Providers: anthropic | Tools: 8/tools — Capability Discovery:
> /tools
filesystem - File operations
bash - Shell commands
web - Web search/fetch
task - Agent delegation> /mode plan
✓ Mode 'plan' activated
> Analyze this codebase for security vulnerabilities
[AI provides analysis]
> Show me the top 3 most critical issues
[AI explains issues]
> /mode off
✓ Mode cleared
> Fix the SQL injection vulnerability in auth.py
[AI makes the fix]> /mode plan
> Review the payment processing module for improvement opportunities
[AI provides recommendations]
> /save payment_analysis.json
✓ Transcript saved
> /mode off
> Implement recommendation #1: Extract payment validation
[AI makes changes]
> /status
Mode: off | Messages: 15 | ...
> /mode plan
> Review the changes we just made
[AI analyzes recent changes]Session 1: Planning
> /mode plan
> Create a plan for migrating to the new API
[AI creates detailed plan]
> /save api_migration_plan.json
✓ Transcript saved
> /exitSession 2: Resume and Implement
# Resume the most recent session
$ amplifier continue
Resuming session: a1b2c3d4
Messages: 5
> Implement step 1 of the migration plan
[AI implements with full context]
> /save api_migration_progress.jsonAlternative: Resume specific session
$ amplifier session list
Recent Sessions:
a1b2c3d4 5 messages
e5f6g7h8 12 messages
$ amplifier session resume a1b2c3d4
# Or use: amplifier continue> /tools
Available Tools:
filesystem, bash, web, task
> /config
[Shows the full mount plan including providers, tools, agents]
> Use the web tool to fetch documentation from anthropic.com
[AI uses web tool]Good candidates:
- Large codebase reviews
- Architecture analysis
- Security audits
- Refactoring planning
- Exploring unfamiliar code
Not needed for:
- Small, focused changes
- Well-understood modifications
- Following existing patterns
Context grows with messages:
- Every message adds to context
- Large context = slower responses + higher cost
- Use
/clearwhen switching topics
Save before clearing:
> /save before_clear.json
> /clear
# Now start freshNaming convention suggestions:
/save feature_name.json # Feature work
/save bug_fix_issue_123.json # Bug fixes
/save review_module_name.json # Code reviews
/save planning_migration.json # Planning sessionsSaved to session directory at ~/.amplifier/projects/<project>/sessions/<session-id>/.
Use interactive mode when:
- Iterative development
- Need to adjust mid-task
- Want to review before proceeding
- Working on complex, multi-step tasks
Use single mode when:
- One-off commands
- Scripting / automation
- Simple, well-defined tasks
# Single shot:
amplifier run "Reply with just OK"
# Interactive:
amplifier run --mode chatYou can start interactive sessions with specific bundles:
# Use a specific bundle
amplifier run --bundle dev --mode chatSee:
- Bundle Guide — bundle composition and structure
- Modes Bundle — the runtime modes system
Slash commands only work in interactive chat mode:
# ✗ Won't work
amplifier run "/mode plan analyze this code"
# ✓ Works
amplifier run --mode chat
> /mode plan
> analyze this codeCheck permissions on the session directory:
ls -ld ~/.amplifier/projects/<project>/sessions/<session-id>/Check /tools output to verify filesystem/bash tools are loaded. Confirm amplifier-bundle-modes is composed into your active bundle (it is, in every standard amplifier bundle). Verify with /config.
Previous documentation referenced /think and /do as plan-mode toggles. These have been replaced by the unified runtime modes system (/mode plan and /mode off) provided by amplifier-bundle-modes. The legacy commands are no longer wired into the CLI.