agent-terminal-ui

Terminal user interface for AI agents built on agent-utilities.

A Textual-based terminal interface for interacting with AI agents. Connects to an agent-utilities backend via dual protocol support: AG-UI (SSE streaming) and ACP (JSON-RPC + SSE).

Note

This library is in early development and subject to change.

Documentation — Architecture, configuration, durable session management, and the autonomous goal and multi-session workflows are maintained in the official documentation.

Features

Core Functionality

• Dual protocol support -- AG-UI (SSE streaming, default) and ACP (JSON-RPC + SSE, opt-in)
• Dynamic workflow sidebar -- discovers graph nodes from sideband events at runtime; nodes are never hardcoded
• Phase labels -- Planning, Discovery, Execution, Validation
• Completed node markers -- checkmarks on finished specialists
• Rich tool execution display -- extensible formatter system for rendering tool calls and results
• Human-in-the-loop tool approval -- modal for confirming sensitive tool calls before execution
• Multi-modal image attachment -- attach images to messages for visual reasoning
• Session management and chat history -- browse and resume previous conversations
• MCP server browser -- inspect connected MCP servers and their tools

Session Persistence & Crash Recovery (TUI-1)

• SQLite-backed sessions -- durable session storage at ~/.local/share/agent-utilities/agent_terminal_ui.db (override the directory with AGENT_UTILITIES_DATA_DIR)
• Pre-turn checkpointing -- automatic checkpoints before each turn for crash recovery
• Session fork/resume -- fork sessions at any turn number with fork_session()
• Offline queue -- messages queued during disconnection survive process restarts
• Session archive -- archive, search, and filter sessions by status

Workspace Snapshots & Rollback (TUI-2)

• Side-git snapshots -- pre/post-turn workspace snapshots without touching your .git
• Snapshot restore -- /restore N to roll back workspace to any previous turn
• Diff viewer -- view changes between snapshot points
• Auto-pruning -- snapshots older than configurable max age are cleaned up

Reasoning Effort Tiers & Auto Routing (TUI-3)

• Three-tier reasoning -- OFF / HIGH / MAX with Shift+Tab cycling
• Auto model routing -- lightweight heuristic selects optimal model and thinking level per turn
• Mode-aware defaults -- plan/code modes default to higher reasoning

Context Compaction Engine (TUI-4)

• Multi-tier compaction -- L1 (summarize tools), L2 (summarize old turns), L3 (drop old), Cycle (hard reset)
• Configurable thresholds -- token-based triggers at 192k/384k/576k/768k
• Auto-compact toggle -- automatic compaction when context window pressure is detected
• Manual /compact -- force compaction at any tier

Durable Task Queue (TUI-5)

• SQLite-backed tasks -- background tasks survive process restarts
• Bounded concurrency -- configurable maximum concurrent tasks
• Timeline events -- full event log per task (created, started, progress, completed/failed)
• Checklist tracking -- structured checklist state per task
• Crash recovery -- interrupted tasks are automatically marked as failed on restart

Lifecycle Hooks (TUI-6)

• TOML-configured hooks -- shell commands triggered on lifecycle events
• Supported events -- session_start, session_end, message_submit, tool_call_before/after, mode_change, on_error, shell_env
• Timeout protection -- hooks are killed after configurable timeout
• Conditional triggers -- hooks can fire only for specific tool categories or modes
• Shell env injection -- shell_env hooks inject environment variables into tool execution

Desktop Notifications (TUI-7)

• OSC 9 / BEL notifications -- terminal-native notifications on long-running turn completion
• Auto-detection -- detects iTerm2, Ghostty, WezTerm, Kitty, and falls back to BEL
• Configurable threshold -- only fires when turn exceeds N seconds (default: 30s)

Workspace Boundary & Trust Mode (TUI-8)

• Three sandbox modes -- read-only, workspace-write (default), danger-full-access
• Trust mode -- bypass approval for non-destructive reads outside workspace
• Explicit allow/deny lists -- fine-grained path-level control
• Violation tracking -- all policy violations are logged

Draft Stash System (TUI-9)

• Multi-entry stash -- Ctrl+S to stash, /stash list and /stash pop to manage
• Buffer management -- Ctrl+U clear + Ctrl+Y restore for single buffer

Enhanced Cost Tracking (TUI-10)

• Per-turn breakdown -- input/output/cached/reasoning token counts with cost
• Cache hit rate -- per-turn and session-level cache utilization metrics
• Pricing registry -- configurable per-model pricing (built-in for GPT-4o, Claude, DeepSeek)
• Session aggregation -- total cost, tokens, and by-model grouping
• Status line display -- compact token/cost indicator

Approval Policy Engine (TUI-11)

• Three policies -- on-request (ask per command), auto (YOLO), never (block all)
• Auto-allow prefixes -- commands matching configured prefixes bypass approval
• Mode-aware strictness -- plan mode requires approval for unknown commands
• Integration -- extends existing danger.py 4-level classification

Job Center (TUI-12)

• Shell job registry -- tracks all shell commands with status, output, and timing
• Output tailing -- last 50 lines of output per job
• Linked tasks -- jobs can reference durable task IDs
• Job lifecycle -- running -> completed/failed/cancelled with cleanup

Agent View - Multi-Session Dashboard (TUI-20)

• Session dashboard -- at-a-glance view of all concurrent agent sessions with status indicators
• Peek panel -- preview the last response of any session without switching context
• Inline reply -- respond to waiting sessions directly from the dashboard
• Background sessions -- send any session to the background with /bg
• Session types -- distinguishes Chat and Goal sessions with visual indicators
• Auto-refresh -- session statuses update every 2 seconds

Background Agent Runner (TUI-21)

• Async workers -- up to 10 concurrent background agent sessions
• State persistence -- session state survives TUI restarts via TaskManager
• Crash recovery -- interrupted sessions are marked as failed on restart

Autonomous Goal Loop (ORCH-5.0)

• /goal command -- define an objective and let the agent work autonomously
• Natural language parsing -- supports until <end_state> and without <constraints> patterns
• KG-native goals -- goals are persisted as GoalNode entities in the Knowledge Graph
• Context enrichment -- goals are auto-enriched with codebase context from the KG
• Rule validation -- goals are checked against the project constitution before execution
• Historical leverage -- prior goal outcomes inform new goal planning
• Durable checkpoints -- goal state is checkpointed for crash recovery
• Validation commands -- auto-extracts shell commands from end-state criteria

User Experience

• Message queuing -- queue messages while agent is processing; related queries are intelligently combined using regex patterns for conjunctions, sequential actions, and similar structure
• Exit confirmation -- modal dialog prevents accidental termination via Ctrl+C or /exit
• Terminal transparency -- UI respects your terminal's transparency settings for seamless integration
• Theme system -- Textual's built-in themes (default tokyo-night; nord, gruvbox, dracula, and more) switchable with /theme

Commands

• Slash commands -- comprehensive command set for common operations:
  • /help -- show available commands
  • /clear -- clear the current event log
  • /mcp -- browse connected MCP servers and their tools
  • /history -- browse and select from historical chat sessions
  • /image -- attach images to messages
  • /init -- initialize a new project or workspace
  • /review -- review code and suggest improvements
  • /test -- run tests on the current codebase
  • /search -- search through code and documentation
  • /stats -- show statistics about the current session
  • /fleet -- show fleet topology and pending approvals; /fleet grant <id> to approve
  • /cost -- show token and cost tracking information
  • /queue -- show current message queue status
  • /queue:clear -- clear all queued messages
  • /queue:toggle -- enable/disable message queuing
  • /model -- switch between available AI models
  • /theme -- switch between available themes
  • /compact -- compact conversation context to save tokens
  • /diff -- show interactive diff viewer for recent changes
  • /recap -- summarize the session context
  • /fast -- toggle fast mode (Haiku/Flash models)
  • /memory -- manage project memory (AGENTS.md)
  • /agents -- list available specialized agents
  • /add-dir -- add a directory to the agent's working context
  • /restore N -- restore workspace to snapshot at turn N
  • /sessions -- list and manage durable sessions
  • /trust -- toggle trust mode
  • /sandbox -- set sandbox mode (read-only / workspace-write / danger-full-access)
  • /approve -- set approval policy (on-request / auto / never)
  • /jobs -- list and manage shell jobs
  • /tasks -- list and manage background tasks
  • /stash -- manage input draft stash
  • /hooks -- show lifecycle hook status
  • /exit, /quit -- exit the application with confirmation
  • /goal <text> -- start an autonomous goal loop
  • /goal:status -- show current goal progress
  • /goal:cancel -- cancel the active goal
  • /goal:history -- browse past goals from the Knowledge Graph
  • /agents -- open the Agent View multi-session dashboard
  • /bg -- background the current session
  • /attach <id> -- attach to a specific session

Input Prefixes

• ! -- Direct Bash execution (e.g., !ls -la)
• @ -- Fuzzy file mention autocomplete (e.g., @app.py)

Backend Integration

• Unified specialist visibility -- MCP agents and A2A peers appear identically in the workflow sidebar; both emit the same sideband events via the backend's discover_all_specialists() unified roster
• Tool-count telemetry -- tools-bound sideband events include toolset_count, dev_tools, and mcp_tools breakdowns for per-specialist visibility
• Real-time token and cost tracking -- integrated in the status line for session monitoring
• Memory Auto-loading -- backend automatically includes AGENTS.md and MEMORY.md in the system prompt for project-aware reasoning.

Usage

Start the agent-utilities backend server, then launch the TUI:

agent-terminal-ui

Or run with uv if installed locally:

uv run agent-terminal-ui

Headless mode

Run a single prompt without the TUI, streaming the response to stdout. No Textual widget tree is loaded, so a headless instance is lightweight (~30MB) -- suited to running many concurrent, non-interactive sessions against one shared backend.

agent-terminal-ui --headless --prompt "summarize the open PRs"
agent-terminal-ui --headless --prompt "run the tests" --model claude-opus-4-8

Docker

A slim, runtime-only image (Dockerfile, python:3.13-slim) ships the frontend without test/shell extras or the agent_utilities backend. Point it at a shared backend with AGENT_URL:

docker build -t agent-terminal-ui .
docker run --rm -e AGENT_URL=http://agent-utilities:8000 \
  agent-terminal-ui --headless --prompt "status report"

Because the backend is the heavy component, run one backend service and many lightweight frontends against it.

Keyboard Shortcuts

• Ctrl+C -- Interrupt generation or cancel current operation
• Ctrl+D -- Exit session (with confirmation)
• Ctrl+L -- Clear the event log
• Ctrl+O -- Toggle workflow sidebar
• Ctrl+S -- Stash current input draft
• Ctrl+T -- Toggle task list view
• Ctrl+U -- Clear input buffer
• Ctrl+Y -- Restore cleared input buffer
• Alt+P -- Switch AI model
• Alt+T -- Toggle Extended Thinking (for reasoning models)
• Alt+O -- Toggle Fast Mode
• Shift+Tab -- Cycle reasoning effort (OFF → HIGH → MAX)
• Ctrl+R -- Reverse history search
• Ctrl+H -- Show help overlay
• Tab -- Navigate between focusable elements
• Esc Esc -- Rewind/Undo (experimental)
• Left Arrow -- Switch to Agent View dashboard
• Right Arrow / Enter -- Attach to selected session (from Agent View)

Message Queuing

When the agent is processing, your input is automatically queued. The system intelligently combines related queries using patterns like:

• Conjunctions: "and", "also", "plus", "then", "after that"
• Sequential actions: semicolon-separated commands
• Similar structure: same action verbs (fix, add, remove, update, create, delete, implement, refactor)

Example: If you type "fix the bug in app.py" followed by "and add a test for it", these will be combined into a single query.

Environment Variables

Variable	Default	Purpose
AGENT_URL	http://localhost:8000	Agent server URL (interactive and headless)
ENABLE_ACP	false	Enable ACP protocol instead of AG-UI
ACP_URL	http://localhost:8001	Documented but not read; effective ACP URL is {AGENT_URL}/acp
AGENT_THEME	tokyo-night	Startup theme (any Textual built-in theme name)

Themes

The TUI uses Textual's built-in themes. The default is tokyo-night; others include nord, gruvbox, dracula, monokai, textual-dark, and textual-light.

Switch themes live with the /theme <name> command, or set the startup theme via the AGENT_THEME environment variable.

Development

Prerequisites

• uv
• prek

Setup

Create uv virtual environment and install dependencies:

uv sync --frozen --all-groups

To update dependencies (updates the lock file):

uv sync --all-groups

Run formatting, linting, and type checking:

uv run ruff format && uv run ruff check --fix && uv run ty check