| name | debug-diagnosis |
| description | MIR pipeline debugging and failure diagnosis for SunoScribe. Use when analyzing bad or empty scores, warnings, failed tasks, missing artifacts, vocal separation issues, F0 or octave errors, beat/downbeat errors, quantization problems, MusicXML/MIDI export failures, RVC quality issues, or when designing admin/debug artifact views. |
Debug Diagnosis
Use this skill to diagnose failures by walking the artifact lineage. Do not jump straight to model swapping or prompt changes.
Diagnosis Order
Follow the pipeline order:
MediaAsset -> CanonicalAudio -> StemSet -> F0Track -> NoteCandidateSet -> RhythmGrid -> ScoreRevision -> Exports -> RVC
At each stage, ask:
- Did the expected artifact exist?
- Was the artifact generated by the correct upstream input?
- Does its metadata match the downstream assumption?
- Is the failure visible in a debug plot or warning?
Failure Categories
Classify issues into one primary category:
media_ingest: unreadable file, wrong duration, failed video extraction, invalid sample rate.separation: missing vocals, strong accompaniment leakage, phase or artifact damage.f0: octave errors, unvoiced errors, missing fundamentals, unstable vibrato/slide handling.note_segmentation: fragmented notes, merged notes, ornaments mistaken for melody.rhythm: wrong tempo, tempo double/half, weak downbeat, unstable beat grid.quantization: wrong note values, misplaced notes, bad pickup measure handling.score_ir: invalid semantic score, missing track, inconsistent note IDs.export: MIDI or MusicXML generated from the wrong revision or invalid score data.frontend: OSMD rendering, revision selection, artifact listing, or playback issue.rvc: bad corrected F0, wrong transpose, model mismatch, converted vocal artifacts, mix imbalance.
Required Evidence
Prefer concrete evidence over broad claims:
- artifact path or ID
- warning text
- confidence value
- small timeline range
- note ID or measure number
- debug image name
- reproduction input
Debug Artifacts
First-version admin/debug views should include:
- waveform
- spectrogram
- vocals/accompaniment previews
- F0 curve with voiced/unvoiced and note candidates
- beat/downbeat markers
- quantized note overlay
- export metadata
- RVC corrected F0 overlay
Output Format
When reporting a diagnosis, use:
Symptom:
Likely stage:
Evidence:
Root cause:
Recommended fix:
Validation:
If evidence is insufficient, state exactly which artifact is missing.
Non-Goals
- Do not recommend RAG for production debugging.
- Do not accept a successful final export as proof of correct analysis.
- Do not tune HMM/DTW/agent prompts before checking the upstream representation.
- Do not hide failure by generating empty or placeholder scores.
