Chat

The Astonish agent operates through an LLM-driven tool-use loop. You describe a goal, and the agent plans, executes tools, observes results, and iterates until the task is complete.

How It Works

1. You send a message (Studio or CLI)
2. The LLM analyzes the request and selects appropriate tools
3. Tools execute, and results feed back to the LLM
4. The LLM decides: respond to you or invoke more tools
5. Loop continues until the task is resolved

There is no fixed pipeline. The agent dynamically adapts based on tool outputs, errors, and intermediate findings.

Studio

The Studio chat interface at http://localhost:9393 provides the full visual experience:

• Real-time streaming responses
• Inline tool execution visualization with expandable cards
• File diffs for write operations
• Artifact previews (reports, generated files)
• Model selector in the toolbar (per-session pin; see Studio Chat)
• Session history sidebar with search
• Token usage tracking

CLI

The terminal chat interface provides the same agent capabilities:

astonish chat                                      # New session
astonish chat -p anthropic -m claude-sonnet-4-20250514  # Pin provider/model on the new session
astonish chat --resume <session-id>                # Resume a session
astonish chat -r <session-id>                      # Short form
astonish chat --auto-approve                       # Skip tool confirmations
astonish chat --debug                              # Enable debug output
astonish chat -w /path/to/project                  # Set working directory

-p / -m on a new session persist as a session pin (same concept as the Studio Model control). See Chat Commands for --no-pin, --clear-model, and astonish chat model.

Authentication Required

The CLI requires authentication before use. Run astonish login <server-url> first. See Remote CLI.

Confirmation System

Destructive or sensitive tool calls require confirmation before execution:

• auto-approve — Tools like read_file, grep_search, memory_search run without asking
• always-confirm — Tools like shell_command, write_file prompt you first

In Studio, confirmations appear as inline approval cards. In CLI, you see a [y/N] prompt.

Use --auto-approve on the CLI to skip all confirmations (use with caution).

Slash Commands

Slash commands are available in both Studio and CLI:

Command	Description
/help	Show available commands
/status	Show this session's provider, model (including pin), tools, and memory status
/new	Start a fresh conversation (new session)
/compact	Show context window usage and compaction status
/distill	Distill the current session into a reusable flow
/fleet	Start a fleet session
/drill	Create a drill suite with guided wizard

Multi-Turn Context

The agent maintains full context within a session. Reference previous messages, build on prior tool results, and incrementally refine solutions. For long-running sessions, the agent automatically compacts context when approaching token limits.

See Sessions for persistence details and Tools Overview for the complete tool catalog.