Features¶

Slash commands¶

Commands are registered in commands.py (CommandProcessor.commands) and surfaced through the input overlay: type / to open a fuzzy-filtered menu (ESC to close, TAB to cycle, Enter to run). Aliases are collapsed in the menu, so /exit and its alias /quit appear once.

The registry groups commands by area, including aliases and :-suffixed subcommands:

Area	Commands
Core	`/help`, `/clear`, `/exit` (alias `/quit`), `/config`, `/keybindings`
Modes	`/plan`, `/chat`, `/build`
Conversation	`/compact`, `/context`, `/recap`, `/export`, `/diff`, `/focus`, `/fast`, `/image`
Models & cost	`/model`, `/stats` (alias `/cost`)
Theme	`/theme`
Queue	`/queue`, `/queue:clear`, `/queue:toggle`
Goals (autonomous)	`/goal`, `/goal:status`, `/goal:cancel`, `/goal:history`
Sessions & agents	`/agents`, `/attach`, `/bg`, `/history`
Knowledge graph	`/graph`, `/kb`, `/impact`, `/codemap`, `/search`
Engine surface	`/ask`, `/nl`, `/obs`, `/broker`, `/kvcache`
Capabilities & runs	`/capabilities`, `/capability`, `/run` (alias `/runs`)
Project & SDD	`/init`, `/sdd`, `/memory`, `/add-dir`, `/prompts`, `/review`, `/test`, `/simplify`
MCP & tools	`/mcp`, `/mcp:reload`, `/tools`, `/skills`, `/resources`
Ops	`/logs`, `/cron`, `/pipeline`, `/maintenance`

Engine-surface commands (`/graph/*` gateway routes)¶

These wrap the epistemic-graph engine surfaces exposed over the AU gateway (action-routed /graph/* twins of the graph_* MCP tools). Each degrades cleanly to an inline error when the engine build lacks that surface.

/ask <question> — DB-GPT-style data-analyst loop (POST /graph/ask-data, KG-2.308): schema-link → plan → execute → self-correct → synthesized answer.
/nl <question> — NL→query planned by the AU fleet LLM (POST /graph/nl-query, KG-2.305); shows the generated query plus rows. Prefix preview to dry-run the query without executing it.
/obs <promql> — PromQL instant query (POST /graph/promql, KG-2.310); /obs range <promql> runs a range query over the last 15m and renders a Unicode sparkline; /obs traces [service] searches distributed traces (POST /graph/traces).
/broker [stats|queues|exchanges] — engine message-broker status (POST /graph/broker, KG-2.310).
/kvcache — shared content-addressed KV-cache occupancy and dedup counters (POST /graph/kvcache, KG-2.306).

Some commands depend on optional backend endpoints and degrade or report an error when the backend does not expose them — see Known Issues.

Live capabilities and run replay¶

Alt+C, /capabilities [query], the global Textual command palette, and the Capabilities sidebar tab all open the same live gateway catalog. Selecting a capability fetches its detail descriptor and generates top-level inputs from the selected action's JSON Schema. The UI shows runtime availability, missing preconditions, legacy route metadata, renderer hint, and policy metadata instead of assuming that every catalog entry can execute. available with readiness=cold remains callable and is labeled as a first-call warm-up, not as degraded.

The palette labels each action's legacy REST metadata as non-frontend-executable and never constructs that route. Every execution goes through POST /api/capabilities/{capability_id}/invoke, where the gateway owns action routing, identity, policy, audit, and canonical run creation. The normal HTTP 202 response is a running acknowledgement, not a completed result; eventual output is read from the run's tool_result event.

Sensitive tool results stay redacted in replay. The inspector displays only the gateway event metadata and never calls a reveal/claim route, caches a revealed value, or copies one automatically. A future reveal-once control must be an explicit, owner-scoped action with no-store handling; until that contract is implemented end to end, the Terminal UI reports the redacted event as-is.

Preflight is non-executing. The palette reruns it immediately before invocation, blocks invalid, denied, queued, unavailable, and degraded actions, and requires a separate operator confirmation for allowed mutating actions. Unknown side effects fail closed through gateway policy. The gateway remains authoritative and rechecks identity and policy at execution.

If the acknowledgement is HTTP 202 approval_required, the palette keeps the exact request and server-bound approval, run, and session IDs. An explicit approve action grants the approval and resumes that same request with those IDs; denial does not execute it.

/run [run_id] opens live Mission Control for the canonical event stream. With no explicit or recently observed run ID, it opens a newest-first browser backed by GET /api/runs, scoped to the current session when available. Mission Control polls cursor replay from GET /api/runs/{run_id}/events, drains retained pages, deduplicates by sequence, and follows until an exact terminal event arrives. Only run_completed, run_failed, run_interrupted, run_cancelled, and error are terminal. graph_complete and final_output remain progress because final output and the authoritative run completion can follow graph completion. The view reports unknown, expired, empty, degraded, and truncated process-local history explicitly, including the missing sequence range when the bounded replay window forces a reset.

Session IDs remain stable across turns while each execution receives a distinct run ID; the palette and replay browser preserve both identities independently.

Machine-readable native/generated support is recorded in capability-coverage.json.

`/model`¶

Backed by the agent-utilities multi-model registry (GET /models):

/model or /model list — render the configured models in a table (id, name, provider, tier, tags, default marker, per-1M cost).
/model show — show full metadata for the active model (falls back to the registry default when no override is set).
/model set <id> — pick a model id for subsequent turns. The id is stored on the app and propagated by AgentClient as an x-agent-model-id header so the backend can override the registry default for the session.

Zero-cost / local models render as $0.00 / $0.00 rather than -.

`/goal`¶

Starts an autonomous goal loop. The objective string is parsed locally by the dependency-free goal.py parser — /goal <objective> until <end_state> without <constraints> — and submitted to the backend. See Goal Command.

Keyboard shortcuts¶

Application bindings (app.py::BINDINGS):

Binding	Action
`Ctrl+C`	Interrupt the current operation
`Ctrl+D`	Exit session (with confirmation)
`Ctrl+A`	Select all
`Ctrl+Q`	Quit
`Ctrl+H`	Show help
`Ctrl+U` / `Ctrl+Y`	Clear / restore input
`Ctrl+G`	Open input in `$EDITOR`
`Ctrl+R`	Reverse history search
`Alt+P`	Model picker
`Alt+T`	Toggle extended thinking
`Alt+O`	Toggle fast mode
`Alt+D`	Service dashboard
`Alt+C`	Live capability palette
`Shift+Tab`	Cycle reasoning effort / mode
`Esc Esc`	Rewind (experimental)

Main-screen bindings (screens/main.py::BINDINGS):

Binding	Action
`Ctrl+B`	Toggle workflow sidebar
`Ctrl+L`	Clear conversation
`Ctrl+V`	Toggle server-log panel

Input prefixes¶

! — run the rest of the line as a shell command (e.g. !ls -la).
@ — fuzzy file-mention autocomplete (e.g. @app.py), via the file suggestions overlay.

Headless mode¶

Run a single prompt without the TUI, streaming output to stdout:

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

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. See Architecture.

Conversation UI¶

Structured message blocks: user messages, agent responses (streaming Markdown), and expandable tool-call blocks with status-colored borders.
New blocks fade in on mount (tui/animation.py); motion is disabled automatically under TEXTUAL_ANIMATIONS=none / reduced-motion environments.
The conversation prunes to max_conversation_widgets (default 50) to bound memory; full history is persisted in SQLite (see Session Management).
Themes: default tokyo-night; switch live with /theme <name> across Textual's built-in themes, or set the startup theme via AGENT_THEME.