Orion — AI Software Engineer for VS Code

Give Orion a task in your own language — "add a dark-mode toggle, test it, and update the README" — and it plans the work into a todo list, reads your project, edits code, runs tests and terminal commands with your approval, verifies its own work with real parsers/compilers, and reports honestly what changed and what did not. Orion replies in the language you write in.

Features

• Agentic loop with an explicit todo list — multi-item requests become a visible checklist; the loop keeps going until every item is done or asks you before stopping.
• Real verification, not vibes — verify_code checks 16 file kinds with actual parsers/compilers (py, js/ts, html, json, yaml, css, …), run_tests returns parsed pass/fail reports, and VS Code's own diagnostics, your active file and your selection are sent with every request.
• Safety by code, not by promise — dangerous terminal commands require an approval modal; destructive/networked git (init/push/pull/clone/reset) is blocked at code level; path-traversal guard; sensitive files (.env, keys) are never auto-overwritten; every overwrite keeps a .bak backup.
• Resilient by design — 3-tier LLM router (rotated Gemini keys → OpenRouter → Cloudflare), heartbeat status while the model thinks, and every run ends with exactly one terminal event so the UI never hangs or fakes success.
• Attachments — paste screenshots or drop image/text files into the chat.
• Performance dashboard — requests, tokens, estimated cost and latency.
• Project memory — per-project .orion/ folder with context docs, dependency map, todo list, backups and logs. Local only, gitignored, never packaged.

Requirements

• Python 3.11+ and Node.js 18+
• VS Code 1.125+
• At least one LLM provider key (Gemini recommended; OpenRouter/Cloudflare optional as fallback tiers)

Install

1. Extension

From a release build (recommended once published):

code --install-extension orion-ai-<version>.vsix --force

From source:

cd orion
npm install
npm run package                    # type-check + lint + production bundle
npx vsce package                   # produces orion-ai-<version>.vsix
code --install-extension orion-ai-<version>.vsix --force

2. Backend

cd backend
python -m venv .venv               # Windows: .venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env               # then add at least GEMINI_API_KEY_1

The extension auto-starts the backend when VS Code opens a workspace that contains (or sits near) this repository's backend/ folder. To run it manually instead:

cd backend
python -m uvicorn app.main:app --host 127.0.0.1 --port 8000

3. First run

1. Open your project's folder in VS Code.
2. Click the Orion icon in the activity bar → Click to open the Chat Panel (or Ctrl+Shift+P → Orion: Open Chat Panel).
3. Send a task. Watch the Activity view — it is Orion's live log (Thinking…, Reading folder structure…, Running test suite…).
4. Approve or reject terminal commands when the modal appears; multi-item tasks show a todo checklist box in the chat.

Where your data lives

Everything Orion remembers about a project lives in that project's .orion/ folder (session history, backups, logs, todos, dependency map). It is gitignored by design and never shipped; deleting it resets Orion's memory for that project without touching your code.

Troubleshooting

Security & privacy

The backend binds to loopback only, CORS is closed, risky terminals need your approval, and destructive git is blocked in code. The approval modal is a deliberate defense against prompt injection — do not disable it. Threat model and honest limitations: see SECURITY.md.

Development

Open orion/ in VS Code and press F5 for an Extension Development Host. Backend tests: python -m pytest backend/app/services/ -q (97 tests). Architecture and roadmap: docs/DEVELOPMENT_ROADMAP.md; refactor history and rules: docs/FASE_C_PLAN.md.

License

MIT — see LICENSE.