teng-lin/notebooklm-py v0.8.0b3

on GitHub

This is a beta pre-release in the 0.8.0 line. Changes since v0.8.0b2:

Added

• MCP file upload now records and can verify a SHA-256 of the uploaded bytes.
  The /files/ul route computes the digest of exactly the received bytes and
  stores it in the completion record ({source_id, name, size, mime, sha256}),
  which await_upload surfaces. An optional ?sha256=<hex> lets a client assert
  transit integrity — a mismatch returns a clean 400 before the source is added
  (the single-use token rolls back and stays retryable)
  (#1889).
• MCP await_upload emits progress keepalives while it polls, so an MCP host
  no longer treats a long upload wait as a stalled call (ctx.report_progress at
  start and on each ~2s tick; best-effort, and a no-op without a client progress
  token) (#1889).
• MCP remote studio_download returns the report / data-table body inline
  (bounded, alongside the signed link) so a remote host can read a generated
  report without a second fetch
  (#1911).
• studio_status reports media_ready, studio_download reports
  size_bytes, and studio_retry raises an actionable error when an artifact
  can't be retried, instead of a bare failure
  (#1924).
• Deploy quick start for end users: the shipped Compose defaults to the
  :latest image, and each release now attaches a version-baked
  docker-compose.yml + env.example
  (#1904).

Changed

• MCP: the standalone studio_get_prompt tool was folded into studio_list
  (net −1 tool, ADR-0025 tool-surface consolidation). Each artifact's
  generation_prompt (the free-text prompt it was generated from, null when it
  records none) now rides the default studio_list summary listing, and
  studio_list(item=<artifact>) returns it for a single artifact resolved by id or
  exact title (the unified cross-type Studio resolver, as studio_delete /
  studio_rename use — not the old artifact-scoped title-prefix resolver). The
  prompt is decoded from the same LIST_ARTIFACTS row the listing already fetches,
  so this adds no extra request. The CLI notebooklm artifact get-prompt command
  is unchanged.
  (#1896)
• The in-app upload widget's token pool gets a longer (1h) TTL. The widget
  mints one single-use token per file up front but uploads them sequentially, so
  a later file in a large batch could hit an expired token mid-upload; the pool
  now uses WIDGET_UPLOAD_TTL instead of the 15-minute default. The single-use /
  signed-token model is unchanged
  (#1894).

Fixed

• MCP notebook_describe(include_metadata=true) no longer reports a source
  count that disagrees with its own sources list. The exposed
  metadata.notebook.sources_count is now re-projected to the enumerated
  (id-bearing, deduped) source count instead of the raw unfiltered
  GET_NOTEBOOK row count, which could inflate it (e.g. 168 vs 50) and mislead
  autonomous agents on quota / completeness. Source enumeration
  (source_list / metadata.sources) also dedups duplicate id-bearing rows
  (research re-imports, ghost/probe echoes), keeping the first occurrence
  (#1919).
• MCP research_import is safe for unattended use. The tool now imports via
  the timeout-tolerant import_sources_with_verification (as the CLI does), so a
  deep-research import that times out is reconciled against what the server
  actually committed instead of failing as if nothing imported (hiding a large
  partial import). It also gained optional cited_only (import only the sources
  the report cites) and max_sources (cap the count imported, applied after
  cited_only) so one call can't silently blow a notebook to its source cap.
  (#1920)
• source_add batch mode isolates a per-URL failure instead of aborting the
  whole batch: a bad URL now yields a per-item SourceAddError while the other
  sources in the same call still add, matching the tool's documented behavior
  (#1905).
• MCP studio_generate for interactive mind maps now normalizes its payload
  to the bare node tree and returns a terminal poll shape, fixing malformed
  generate requests and non-terminal status reports
  (#1908,
  #1914).
• MCP research_cancel distinguishes an already-cancelled run and surfaces
  the raw status code instead of a generic failure
  (#1922).
• Content-sanity flags bot-challenge / WAF interstitial pages, so a source
  that fetched a challenge page instead of real content is reported rather than
  silently ingested as thin content
  (#1929).
• suggest_prompts strips a leading list marker from returned prompts, and
  research_start drops redundant ids from its result
  (#1909).
• The MCP server warns at startup when the upload widget is enabled but
  stateless HTTP is disabled (FASTMCP_STATELESS_HTTP explicitly falsey),
  which would silently stop the widget from rendering
  (#1915).

Security

• The loopback DNS-rebinding Host guard is no longer bypassed on a tokenless
  local bind. The guard is skipped only when an external bind is opted into
  and authentication (bearer or OAuth) is configured — the
  ALLOW_EXTERNAL_BIND flag alone (e.g. set while --host stays at loopback) no
  longer disables it, closing a rebinding gap on an unauthenticated local server
  (#1935).
• Client-facing RPC errors no longer leak the obfuscated internal method id or
  the raw upstream status code, trimming information disclosure in error
  messages (#1931).

Full 0.8.0 cycle changelog: CHANGELOG.md. Install with pip install --pre notebooklm-py==0.8.0b3.