Darius v1.1.1

Open-source lean agent harness — Claude-Code-style TUI, durable memory, tool ACI, plan–execute–accept loop. Local-first, provider-optional, zero API keys required to get started.

Install

Option A: Pre-built binary

curl -sSL https://github.com/galaxycoils/darius/releases/latest/download/install.sh | bash

Option B: From source

cargo install --git https://github.com/galaxycoils/darius darius-cli

Quickstart

1. Run the smoke test (no API key needed)

darius session-smoke

2. Launch the TUI

darius tui

3. Use memory

darius memory stats

4. Configure a live provider (optional)

mkdir -p ~/.darius/profiles/default
cat > ~/.darius/profiles/default/config.toml << 'EOF'
[model]
provider = "openai_compatible"
base_url = "https://api.openai.com/v1"
model = "gpt-4o-mini"
api_key_env = "DARIUS_API_KEY"
EOF

export DARIUS_API_KEY="sk-your-key-here"

5. Run with a real goal

darius run "analyze this codebase and summarize the architecture"

Without DARIUS_API_KEY, darius run uses the offline MockModel — useful for testing the loop without network.

TUI Keyboard Reference

Key	Action
❯ text + Enter	Send a message
/	Open command palette
- at column zero	Also opens palette
Shift+Tab	Cycle mode (auto → manual → accept-edits → plan)
Esc	Close palette / interrupt
q	Quit

TUI Slash Commands

Command	Description
/help	Show available commands
/clear	Clear transcript
/compact	Compact context
/model	Show/set model
/mode	Cycle interaction mode
/effort	Set effort level
/permissions	View permission policy
/memory	Memory stats
/pack	Build MemoryPack
/tasks	Show task board
/plan	Show current plan
/status	Session status
/config	Show/set config
/skills	List skills
/a2a	A2A card info
/serve	Start localhost server
/stop	Stop current operation
/quit	Exit TUI

Modes

Mode	Indicator	Behavior
Auto	⏵⏵ auto	Full autonomous execution
Manual	⏸ manual	Pause after each step
Accept Edits	⏵⏵ accept edits	Auto-accept file edits
Plan	⏸ plan	Planning only, no execution

CLI Commands

Command	Description
darius run "goal"	Cognitive loop (Mock or live if configured)
darius tui	Launch Claude-Code-style TUI
darius serve	Start web dashboard + A2A server
darius session-smoke	Integrated daemon + session + handoff test
darius memory search <q>	FTS5 search
darius memory pack	Bounded MemoryPack (≤3500 chars)
darius memory import <file>	Deduped JSONL import
darius memory export <file>	JSONL export
darius memory stats	Record count + DB path
darius config show	Show profile config
darius a2a card	Show A2A agent card

Environment Variables

Variable	Description
DARIUS_API_KEY	API key for live provider
DARIUS_PROFILE	Profile name (default: default)
DARIUS_LIVE_MODEL	Force live model for darius run

Architecture

darius-memory     → SQLite FTS5, MemoryPack, JSONL
darius-tools      → ToolRegistry, TOOL line protocol, spill
darius-cognitive  → Plan → TaskBoard → ReAct → Accept
darius-daemon     → Session, event log, handoff, model router
darius-tui        → Claude-Code-style terminal UI (ratatui)
darius-web        → Axum web dashboard + A2A server
darius-cli        → CLI surface + session runtime

Lean resource caps

Resource	Cap
MemoryPack	3500 chars
Tool preview	32 KiB (+ spill to disk)
TaskBoard	15 tasks
ReAct iters	12 per task
Body size	32 KiB per record

What's in v1.1.1

• ✅ Claude-Code-style TUI with streaming turns, command palette, modes, effort, todos, diffs, permission chooser
• ✅ Offline MockModel (no network)
• ✅ Live OpenAI-compatible provider when configured
• ✅ Durable SQLite memory with FTS5 search
• ✅ Plan–execute–accept cognitive loop
• ✅ CLI with memory operations, TUI, serve, A2A
• ✅ Session handoff + event replay
• ✅ Web dashboard + A2A server (SSE)
• ✅ Unified UiEvent/runtime across CLI, TUI, web, A2A

Design inspiration

Visual/interaction grammar inspired by Brainless. No Brainless source code was copied or translated.

Build & Test

cargo fmt --all -- --check
cargo clippy --workspace --all-targets -- -D warnings
cargo test --workspace
cargo build --release -p darius-cli

License

MIT