notebooklm-py

notebooklm-py 0.8.5 imports typing_extensions at runtime but fails to declare it as a dependency, so clean installs via pipx/uv/mise crash on startup. Workaround: install typing_extensions into the same isolated environment, e.g. pipx inject notebooklm-py typing_extensions or uv pip install --python <notebooklm-py venv>/bin/python typing_extensions. Permanent fix: upgrade to a release containing PR #2500, which declares typing_extensions as an unconditional dependency and adds CI checks for clean installs.

Create a LoopBoundPrimitive base class with a set_bound_loop method that stores the loop and fires a hook _on_loop_rebind(old_loop, new_loop) only when the loop actually changes. Subclasses can override the hook for clear-on-rebind behaviour. This consolidates duplicated bookkeeping across multiple components. Example: src/notebooklm/_loop_bound.py.

In notebooklm-py 0.8.5 on Windows, notebooklm login --master-token --browser chrome captures the oauth_token successfully, but teardown in src/notebooklm/_browser/oauth_token.py raises TargetClosedError at context.close() inside the finally block (because Chrome already closed itself). The exception escapes finally, so the captured token is discarded and master_token.json is never written. Workaround: patch oauth_token.py to wrap page.close(), context.close(), and browser_obj.close() each in try/except Exception: pass, then rerun the login command — it should report Master-token login OK. A permanent fix is in PR #2522; upgrade notebooklm-py once it is merged/released instead of keeping the local patch.

This happens because Google now returns an HTTP 200 signed-out landing page for GET / on notebook.google.com, while the capture logic only checks that the browser is on the app host and has a URL-scoped SID cookie. An expired SID therefore passes validation and overwrites storage_state.json.

Immediate workaround: disable headless re-auth so the expired browser profile cannot be captured:

unset NOTEBOOKLM_HEADLESS_REAUTH
# or
export NOTEBOOKLM_HEADLESS_REAUTH=0

Also ensure allow_headless=False, then restore a valid storage_state.json or authenticate again.

Code fix: in src/notebooklm/_browser/browser_capture.py, update both the headless path around _settle_capture_candidate/_captured_sid_is_usable and the CDP path to verify session liveness before writing storage. Either require the SNlM0e token on the captured app page, or perform the /login redirect check: a live session redirects back to the app, while a signed-out session goes to accounts.google.com. On failure, call the same refusal path used for SID-less captures, such as _refuse_incomplete_capture, and retain the existing authentication state. Add a regression test using a token-less app landing with an expired SID in the browser context.

This is a known bug (affects main @ 1d8920f, MCP _resolve_by_title and _normalize_for_match): str.casefold() converts 'İ' (U+0130) into two characters, 'i' + U+0307 (combining dot above), so a lowercase query like istanbul never matches the folded title İstanbul notları, raising NotebookNotFoundError. Workaround: type the title with its original exact characters, including the capital 'İ' (e.g., query İstanbul instead of istanbul) — exact-character queries match because both sides fold identically. Permanent fix: strip combining marks during normalization, e.g. in _normalize_for_match use unicodedata.normalize('NFKD', s) and remove characters where unicodedata.combining(c) is true before/after casefold(), so 'İ' folds to plain 'i'.