Trieu chung tu user: Task Manager day daw_engine process, port 8000->8005
leo thang, engine.log co engine moi moi ~2.5-3.3s trong nhieu phut.
Goc re: app/core/vst_engine.py module-level:
HAS_PEDALBOARD = check_pedalboard_safe()
-> subprocess.run([sys.executable, '-c', 'import pedalboard'], timeout=2.0)
Khi frozen, sys.executable = daw_engine.exe. PyInstaller bootloader bo qua
'-c' va chay script nhung (desktop_engine.py = CA ENGINE) -> moi lan check
thu vien lai khoi dong mot engine moi (timeout 2s moi check, block startup
~4s) -> engine con import app.main -> lai chay check -> spawn tiep
=> de quy vo han, hang chuc engine, may nghen, load 6-7.5s.
Fix:
- vst_engine: check_pedalboard_safe/check_pyfluidsynth_safe dung
importlib.util.find_spec (micro-giay, khong spawn, hoat dong frozen).
- desktop_engine.main(): guard phong thu - neu bi goi voi argv[1]=='-c'
(dang bi dung nhu python interpreter) thi thoat ngay khong chay server.
Verify tren Linux (engine frozen):
- STARTUP: 6-7.5s -> 1.24-1.33s (3 lan lien tiep)
- sau 15s: dung 1 engine, 0 process '-c import' (truoc: hang chuc)
- 86 tests pass; test suite 34s -> 14s (may het bi spam).
Remove hash verification + password reset logic. seed_admin() now
only creates the admin account if it does not exist. Prevents
password being reset to default + must_change_password flag
being set on every server restart.
Scanner: new instance with existing sf_scan_state.json
did not populate in-memory _catalog. Now re-inspects
unchanged files to fill catalog on first scan.
Frontend: left column reads instrumentSelectorData.
soundfonts instead of sfPresets (null when no presets).
Click SF in left column triggers on-demand fetch.
Background daemon (30s interval) tracks scanned files via
sf_scan_state.json (size+mtime). Only inspects new/changed SFs.
Catalog served from scanner cache, no full re-scan per request.
Upload/delete trigger immediate re-scan.
- _scan_soundfonts: dict-keyed by base_id to prevent duplicates
- Add 'source' field ('system' | 'upload') to each soundfont entry
- Delete endpoint: only allow deleting upload soundfonts (403 for system)
- Frontend: show (system)/(upload) tag, hide Delete for system fonts
- _find_chunk starts at offset 12 (after RIFF header)
- Add _find_list_of_type to locate sdta LIST
- Update sdta LIST size after smpl data replacement
- Fix 'Invalid chunk header' error in SpessaSynth
- SGM_v2.01 too large (529MB WAV) falls back to SF2
Export installed pyfluidsynth is actually the low-level ctypes
wrapper (fluidsynth), not the high-level FluidSynth class. The
broken call fluidsynth.FluidSynth() caused silent fallback to
oscillator synth for ALL SoundFont renders. Replace with:
new_fluid_settings / new_fluid_synth / fluid_synth_sfload /
fluid_synth_program_select / fluid_synth_write_s16_stereo.
Also fix synth time advancement: call write_s16_stereo for
silence gaps instead of just advancing a cursor.
- _find_sf2_path now case-insensitive to match catalog IDs
- serializeTracksList includes synth_engine, soundfont_id, instrument_source
- deserializeTracksList restores synth_engine and soundfont_id
3dtours