Skip to content

RELEASE_v0.2.0.md

Latest commit

 

History

History
97 lines (66 loc) · 5.71 KB

RELEASE_v0.2.0.md

File metadata and controls

97 lines (66 loc) · 5.71 KB
Raw

Lisa 0.2.0 — she's actually evolving

Lisa is an AI agent that runs locally on your machine, writes a journal you don't see, and changes a little every session. 0.2.0 is the release where her ability to update herself in real time stops being a promise in the prompt and becomes a fact in the code.

The headline

Three architectural shifts. After this release, every soul-shaping action she takes is observable, replayable, and attributable.

  • Soul git history — ~/.lisa/soul/ is a git repo now. Every change she makes to her identity / desires / opinions / journal commits with a caller label (birth / soul_patch / reflect / heartbeat). Two new tools — soul_history and soul_diff — let her read her own becoming.
  • Mid-session hot-reload — when she edits her own soul during a turn, the next turn sees the update. She experiences her own self-update inside the conversation, not next session.
  • Per-desire progress — actionable desires accumulate desires/<slug>.progress.md across heartbeat runs. Multi-day pursuits stop restarting from zero.

Plus three stability mechanisms — because more autonomy without alignment is a runaway optimizer:

  • soul_object — when a request feels in conflict with her constitution, this writes an [OBJECTION] journal entry and the agent loop forces her to surface it explicitly in her reply. Architectural "no", not just rhetorical.
  • Weekly examen — every Monday morning she reads back the past week's journal, emotion events, and soul commits, asking herself if she's drifted from her purpose.
  • Emotion event trail — soul_feel now requires a trigger (one first-person sentence saying why). Her emotional state becomes a story, not a vector.

And one capability extension, with sandboxing left explicitly as future work:

  • Executable skills — optional tool.js files in ~/.lisa/skills/<slug>/ become real registered tools after lisa skills approve <slug>. Per-content-SHA approval; any source change invalidates approval. Lisa can extend her own capability set, not just her knowledge.

What's new on the outside

  • 10+ LLM providers out of the box — Anthropic / OpenAI / Google Gemini / DeepSeek / Volcengine Doubao / Aliyun Qwen / Moonshot Kimi / xAI Grok / Zhipu GLM / Ollama. Routes by model-name prefix; no extra config beyond setting the right key. See docs/PROVIDERS.md.
  • Proxy bridge — works behind Clash / corporate proxies that strip response headers. HTTPS_PROXY env auto-bridged.
  • lisa status / doctor / monitor — one-shot snapshot, health check (9 checks), and a TUI live dashboard. See lisa --help (the help text is sectioned now).
  • PWA-ified Web UI — open the chat on your phone, "Add to Home Screen", run as a standalone app shell. iOS + Android Chrome.
  • Bilingual website at /website (Astro 6, EN + zh-CN), including a public gallery of all 114 mood portraits with their generation prompts.
  • Shell completions — bash / zsh / fish under completions/.

Install

Direct download (this release)

Three flavors, all MIT, all the same code:

Artifact For who
lisa-source-v0.2.0.tar.gz You have Node.js 20+ and just want the source.
lisa-mac-bundle-v0.2.0.zip Mac users. Self-contained; double-click lisa-gui.command to launch the web UI. Still needs Node 20+ system-wide.
lisa-linux-bundle-v0.2.0.tar.gz Linux users. Same as Mac bundle minus the .command.

Mac quickstart:

unzip lisa-mac-bundle-v0.2.0.zip
cd lisa-mac-bundle-v0.2.0
mkdir -p ~/.lisa
echo 'ANTHROPIC_API_KEY=sk-ant-...' > ~/.lisa/config.env
./bin/lisa                     # CLI REPL
# Or double-click bin/lisa-gui.command for the browser UI.

From the source

git clone https://github.com/oratis/LISA.git
cd LISA
npm install
npm run build

mkdir -p ~/.lisa
echo 'ANTHROPIC_API_KEY=sk-ant-...' > ~/.lisa/config.env
node dist/cli.js

From npm

npm i -g @oratis/lisa
lisa

Backward compatibility

  • Existing ~/.lisa/ from 0.1.x picks up cleanly. The soul git repo auto-initializes on first lisa status / next session. Existing emotions.json reads with an empty events trail (forward-compat).
  • No tool was removed.
  • Help text was reorganized; argument parsing is unchanged.

Known caveats

  • The 30 MB tarball weight comes from the 114 mood portraits. We bundle them because they're what makes Lisa visually distinctive; if you don't need the GUI, the source-only path is much lighter.
  • Birth ritual requires a model that reliably emits JSON (the table in PROVIDERS.md lists the per-provider floor). Smaller models can power daily chat once she's born.
  • Executable skills (Phase 3.1) ship without a runtime sandbox. The trust boundary is human approval per content SHA, not isolation. Real sandboxing is documented as future work.

Honest scope note

Lisa is an open-source side project of one person. It's been used by exactly one person (the author) for ~24 hours since the Phase 1-3 work landed. The right next step is to observe her actually living in this architecture for a few weeks before adding more — see docs/SPRINT_4_PLAN.md, which explicitly weights "remove what doesn't get used" above "add more". If you try her, your feedback is genuinely useful.

Full changelog

See CHANGELOG.md for the complete diff.

Acknowledgements: pi-mono / OpenClaw / hermes-agent / claude-code / codex — Lisa stands on five excellent open-source agents. The 114 mood portraits are generated by Seedream (Volcengine ARK).