Summary

Found a few bugs while reading through the codebase -- two temp file leaks, incorrect return type annotations, and some type() == str checks that should be isinstance().

Bug 1: Temp file leak in auto_sync()

auto_sync() creates temp files and appends .ass / .srt extensions to them (e.g. srtin = srtin_file.name + '.ass'). The finally block was deleting the extensionless base files (srtin_file.name, srtout_file.name) but never the actual .ass and .srt files that ffsubsync writes to. Every call to auto_sync() leaked two files in the system temp directory.

Fix: Added cleanup for srtin and srtout (the paths with extensions) in the finally block, with os.path.exists() guards since they may not have been created if an error happened early.

Bug 2: Temp file leak in WhisperAPIModel.transcribe()

convert_video_to_audio_ffmpeg() creates a temp audio file (e.g. /tmp/filename.mp3) but after transcription finishes, that file was never cleaned up. Repeated transcriptions would accumulate audio files in the temp directory.

Fix: Wrapped the transcription body in a try/finally that removes the converted audio file after use.

Bug 3: Wrong -> str return type on transcribe() in 6 models

The abstract base class correctly declares transcribe() -> SSAFile, and all implementations actually return SSAFile objects. But 6 concrete model files had -> str as the return type annotation:

• whisper_model.py
• whisper_api_model.py
• whisper_timestamped_model.py
• whisperX_model.py
• faster_whisper_model.py
• whispercpp_model.py

Fix: Changed all to -> SSAFile. Added the missing from pysubs2 import SSAFile to whisper_model.py (the others already had it).

Bug 4: type(model) == str instead of isinstance()

Three places in main.py used type(model) == str which won't match subclasses of str and is generally considered non-Pythonic.

Fix: Changed all three to isinstance(model, str).

Test plan

• Verify auto_sync() no longer leaves .ass/.srt files in temp directory after use
• Verify WhisperAPIModel.transcribe() cleans up the converted audio file
• Confirm type checkers (mypy/pyright) no longer flag return type mismatches on transcribe()
• Run existing test suite (python -m pytest tests/)

🤖 Generated with Claude Code