Topic routing, dedup, attachments, sweep + pytest test framework

[RonShih]

What

Builds out agent-notes and adds a proper test framework.

Product

• Topic routing: inline #tag routes each job to a per-topic NotebookLM notebook + Notion DB (topics.json registry; db ids cached under .state/).
• Dedup: per-topic processed index keyed on normalized URL / file content hash; #force to re-do; recorded only on full success (failures stay retriable).
• Attachments (UC2): accept Telegram PDF/DOCX/TXT/MD uploads.
• Sweep (UC3): #topic 整理現有 exports every not-yet-done source already in a topic's notebook.
• Clearer auth errors: the SDK's cryptic "returned an error result: success" (CLI not logged in) is mapped to an actionable claude auth login message.

Testing / tooling

• Converted ad-hoc scripts to pytest with shared fixtures (tests/conftest.py): cfg/registry/index/good_report, a configurable fake-agents harness, and a no-network poll() driver.
• 151 tests, all offline (no NotebookLM/Notion/Telegram, no real .state/ writes).
• Added ruff + pytest-cov config, a GitHub Actions CI workflow (ruff + pytest on push/PR), and a fresh-context test-runner subagent.

Verify

pip install -e ".[dev]"
pytest            # 151 passed
ruff check src tests

Notes

• Secrets stay out: .env and .state/ are gitignored.
• Coverage 83% overall; core logic modules 89–100% (the low ones — agent.py network boundary, notebooklm_cli.py subprocess boundary — can't be unit-tested offline).

🤖 Generated with Claude Code