Skill-memory agent

Self-editing agent on Gemini 3.5 Flash. Three-tier markdown memory (system / active / archive), per-user isolation, web UI + CLI.

pip install -r requirements.txt
$env:GEMINI_API_KEY = "..."

Web server (multi-user, mock login):

uvicorn server:app --host 127.0.0.1 --port 8000

Open http://127.0.0.1:8000 → pick a mock user → chat.

CLI (single user per process):

$env:AGENT_USER_ID = "alice"
python agent.py

1. Read frontmatter of every skill under data/users/<user_id>/skills/.
2. Gemini picks relevant skills from non-system tiers.
3. System tier always loads. Active picks load whole-body. Archive picks load chosen ## sections.
4. Gemini answers (google_search grounding available).
5. Reflect call returns skill edits — applied to disk.
6. Turn appended to sessions/<session_id>.jsonl.
7. On /api/session/end (or Ctrl-C in CLI) → session summarized into a new active skill.

data/users/<user_id>/
  skills/
    system/*.md         # always loaded (identity, response style, hard rules)
    *.md                # active tier, frontmatter-matched
    archive/*.md        # archive tier, section-retrieved by ## heading
  sessions/
    <session_id>.jsonl  # one turn per line

---
name: skill-name
description: One-line description used for matching against prompts
---

Body. For archive-tier, structure with ## sections for granular retrieval.

mock_users.json lists allowed user ids. Login POSTs to /api/login which sets a signed cookie. current_user dependency in server.py re-validates the cookie on every request.

Single seam for Google OAuth: replace /api/login with the OAuth callback. Cookie payload shape ({"user_id": "..."}) and current_user dependency stay identical. Nothing else in the codebase touches identity.

Mock login has no password — bind to 127.0.0.1 only.

• GEMINI_API_KEY or GOOGLE_API_KEY — required.
• AGENT_USER_ID — CLI only. Must exist in mock_users.json.
• SECRET_KEY — optional. Auto-generated at .secret_key if absent.
• AGENT_DEBUG=1 — prints raw LLM JSON from reflect/session calls.