Known gaps, observed from the code as it stands. This is a record of limitations, not a schedule — nothing here is committed work.
Settings are hardcoded. 6 values that a user might reasonably want to change — whisper model, transcription language, ollama model, summary chunk size — are literals inside the stage scripts. There is no config file, no environment variable, and no flag. Changing any of them means editing code and clearing a resume log.
Resume logs record completion, not provenance. A log entry says an item was processed; it does not say which model produced it or when. Re-running after changing a model therefore skips everything, and the only remedy is deleting the log and regenerating from scratch.
No tests. The pipeline has no automated verification. A stage that silently produces empty output would not be caught until someone read a transcript.
Upstream layout is an unversioned dependency. Feeds, publisher repositories, and page structures can change without notice, and the download stages have no contract to rely on. See External Services.
Single language. Transcription and grammar correction are both pinned to one locale, so multilingual episodes are transcribed as though they were not.
CONTENT_LICENSE.md.