Releases: devdaviddr/itsaagent-cli
Releases · devdaviddr/itsaagent-cli
v0.1.0 — Initial Alpha Release
ItsAAgent v0.1.0 — Initial Alpha Release
⚠️ This is an early alpha. The CLI interface, config schema, and tool API may change before a stable release.
A local-first ReAct agent for the terminal. No cloud. No API keys. Runs entirely on your machine via Ollama.
What's included
Agent
- ReAct loop: Thought → Action → Observation, repeating until
<answer>or the step limit - Resilient 3-fallback response parser — handles
<tool_call>XML, legacyTOOL:format, and bare JSON (scoped to post-</thought>text to avoid false positives) - Loop detection: same tool + args called 3× aborts the run
- Context management: 24 576-token window, oldest-first eviction, system prompt and original task always pinned
- OS-aware system prompt: injects platform/version at runtime so the model uses correct commands
Tools
bash— shell commands viaexecFile(no injection), 30s timeoutssh— ControlMaster persistence, Wake-on-LAN auto-recovery, password from env onlyread_file,write_file,glob,grep
CLI
ai run— one-shot taskai chat— interactive multi-turn sessionai check— health check (provider + model)ai models— list Ollama modelsai config— view/edit persistent config
TUI
- Live Ink terminal UI: spinner, streaming tokens, per-step status
- Auto-detects TTY; falls back to plain text in pipes/scripts
Provider abstraction
- Ollama (NDJSON streaming) — default
- OpenAI-compatible endpoints (SSE streaming)
Session logging
- Markdown log per session written to
~/.config/ai-cli/logs/with-vor-l
Tests
- 40 unit tests: parser fallbacks, context trim, loop detection, tool execution, runtime events
Quick start
ollama pull qwen2.5-coder:7b
git clone https://github.com/devdaviddr/itsaagent-cli.git
cd itsaagent-cli
npm install && npm run build && npm install -g .
ai check
ai run "what files are on my desktop?" -vKnown limitations
- No streaming to non-TTY (plain mode waits for full response before printing)
- SSH password must be in
SSH_PASSenv var — no interactive prompt yet - Context trimming is FIFO; no tool-result summarisation
- Windows is untested
Recommended model
qwen2.5-coder:7b — the entire tool chain is tuned for its output format. mistral:7b also works well.