Skills 2026-08-27 08:49 UTC
cdaflatest
Version [latest] committed by UditAkhourii. Update Category: Feature Addition & Enhancement. Commit details: Fix transcript guard false positives; consolidate skill trust guidance. This update introduces new capabilities and module extensions.
CommitRaw commit message
Fix transcript guard false positives; consolidate skill trust guidance The non-transcript guard from #1 substring-matched anywhere in the first 200 characters against a list including "sorry", "i cannot", "there is no" and "trained by". Six plausible on-camera lines -- "Sorry, I'm late", "I cannot believe how well this turned out", "There is no better way to start the morning" -- were all discarded, silently: the transcript vanishes and x-transcript-timing records none, so the sidecar looks legitimately speechless. Now two tiers. Assistant self-reference and explicit transcription refusals still match anywhere. Generic refusal phrasings must both open the reply and be paired with a reference to the recording, so ordinary narration survives. All 7 real lines pass, all 6 chatbot replies blocked. Adds test_local_body.py: drives describe_video_local end to end with ffmpeg and the model mocked, asserting the assembled body parses as a valid sidecar (including >1h timestamps and the brief/coarsen path). Its first version passed for the wrong reason -- the mocked ffmpeg never created the wav, so _transcribe bailed before the guard ran. Merging #1 also left two overlapping trust sections in SKILL.md; they are now one, keeping the PR's x-key table and phantom-brand-name warning. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>