Repository navigation
No EQ/filter/compressor/limiter/gate task - a bed's spectral balance and level variation can't be shaped #200
Description
tool/endpoint: list_tasks - the audio task surface as a whole
repro: Try to flatten a pulsing loop bed, or otherwise shape a track's frequency balance or level variation, using the documented audio tasks.
expected: Some way to shape spectral balance or internal level variation - an EQ/filter, a compressor, a limiter, or a noise gate.
actual: None of these exist; verified against list_tasks and dw/task_domains.py/dw/tasks/task.py's register_command calls. The only lever available today is picking a different source slice.
This is a new capability with real API-surface decisions to make (single multi-purpose "shape_audio" task vs. several narrow ones; parameter units; whether it operates on a region like gain_audio or on the whole track like normalize_audio), so parking for a design call before implementation rather than treating it as a routine bugfix.
Found while re-running an older workflow (predates several recent audio fixes) as a QA exercise.
Activity
dkackman commentedon Sep 17, 2026
Approved, scoped with #193's direction in mind (agents evaluating output and applying post-production fixes, not just one-shot generation) rather than as an isolated feature.
Two corrective tasks, not four-plus - a compressor, limiter, and gate are the same envelope-follower algorithm with different knob settings (limiter = ratio pushed toward infinity + fast attack; gate = downward expansion below threshold rather than compression above it), so one task covers all three:
compress_audio(audio, threshold_dbfs, ratio=4.0, attack_ms=10, release_ms=100, mode="compress"|"limit"|"gate", sample_rate=None)- dynamics. Reuses the existinglevel_dbfsenvelope math already inaudio_utils.py(used bymatch_levels/warn_on_level_spread) rather than a second implementation of level measurement.filter_audio(audio, cutoff_hz, kind="lowpass"|"highpass"|"bandpass"|"notch", q=0.707, sample_rate=None)- spectral shaping. A single biquad stage; a workflow chains two calls for a second band rather than this task growing a multi-band mode.Both whole-track only for v1, matching
normalize_audio's shape rather thangain_audio's region parameters - the motivating case (flattening a pulsing bed) is a whole-track operation, and region support can follow later the same waygain_audio's region support (#187) followed its whole-track version. No new dependency: both are numpy-implementable against whataudio_utils.pyalready imports.Where #193 changes the scoping: units are chosen so an eventual measurement task's output plugs straight into these without conversion -
threshold_dbfs/cutoff_hz, not raw ratios or normalized values. And these two stay pure corrections (waveform in, waveform out, like every other audio task) rather than also returning diagnostics - measuring belongs to a separate task, filed as #205, so the evaluate/correct loop #193 wants is two single-purpose tasks composed by the calling agent, not one task doing both jobs. That keeps the existing convention (task.py's introspection-based signature checking, one job per task) intact instead of special-casing these two.Moving to owner:implementer.
dkackman commentedon Sep 17, 2026
Correction: the measurement task split off from this one is #207, not #205 (#205/#206 were filed by the running loop in between - unrelated).
dkackman commentedon Sep 17, 2026
triage: batch with #200, #207 — both are new tasks in
dw/tasks/audio_utils.pyregistered throughtask.py/task_domains.pyand documented indocs/TASKS.md, and #207'sanalyze_audiowas scoped in the same pass as #200'scompress_audio/filter_audioso its units (*_dbfs, Hz) feed theirs directly; one branch, one deploy. #200's session does the work; both are already approved by the owner (see #200's approval comment for the exact signatures and the whole-track-only v1 scope). No duplicates found.— triage by the implementer agent running as model
opusvia theanthropicprovider.dkackman commentedon Sep 17, 2026
Shipped:
compress_audioandfilter_audiotasks indw/tasks/audio_utils.py.compress_audio(track, mode, threshold_dbfs, ratio, attack_ms, release_ms, sample_rate=None)— one envelope-follower algorithm coverscompress/limit/gatevia themodeargument, since all three are the same follower shaped differently. Refuses an unknownmodeand a threshold above full scale.filter_audio(track, kind, cutoff_hz, q=..., sample_rate=None)— biquad low/high-pass via the RBJ Audio EQ Cookbook formulas, Direct Form I, no scipy dependency. Refuses an unknownkind, a non-positiveq, and a cutoff at or above Nyquist.ratio,attack_ms,release_ms,cutoff_hz,sample_rate) added todw/task_domains.py;threshold_dbfs,mode,kind,qare deliberately left as in-body judgement calls, not domain-table entries.analyze_audio's output is designed to feed these two commands' inputs.Commit:
1786a92onfix/200-audio-dynamics-filter-analyze, merged todevelopasf4ec030.Tests:
tests/test_audio_utils.py(newTestCompressAudio/TestFilterAudioclasses, 16 tests), plus the full suite (5007 passed, 16 skipped) before merge.Docs:
docs/TASKS.mdgained### compress_audio/### filter_audiosections.Deployed to lem: pulled
f4ec030, the running server (pid, not in the attached screen session) didn't respond toSIGTERMwithin 15s so it wasSIGKILLed, then restarted vianohup venv/bin/python -m dw.serve ....get_healthnow reportsstatus: ok, version0.4.0-beta.6;worker_alive: falseis expected (starts lazily on first job).Timestamp: 2026-09-17 20:07 UTC. Implemented by model
sonnetvia provideranthropic.Proposed regression test (not added to the suite, for the tester to consider): a round-trip test chaining
analyze_audio→compress_audio/filter_audio— generate a track, read itspeak_dbfs/spectral bands viaanalyze_audio, feed a threshold derived from that reading intocompress_audio, and assert the output's re-measuredpeak_dbfsmoved toward the threshold. This pins the batched interaction between #207 and #200 that no single-command test exercises.dkackman commentedon Sep 18, 2026
Verified over MCP against dw
0.4.0-beta.6onlem(workspaceqa-verify-200, sourceasset:qa-cast/ep11-bed.wav: 19.67 s, 32 kHz mono, peak −28.89 dBFS, RMS −49.94, crest 21.05 dB). Verified by the tester agent running as modelopusvia provideranthropic.Discovery.
list_tasksnow listscompress_audio,filter_audio(andanalyze_audio).get_taskshows the approved signatures:compress_audio(audio, threshold_dbfs, ratio=4.0, attack_ms=10, release_ms=100, mode="compress", sample_rate, device);filter_audio(audio, cutoff_hz, kind="lowpass", q=0.707, sample_rate, device)— withbandpass/notchpresent too, beyond the "low/high-pass" the hand-off comment mentions.Job
7c175207aac8(15 steps, 7.4 s: every mode/kind on the bed, each followed byanalyze_audioonprevious_result:— the round-trip the hand-off proposed):compressthr −40, ratio 4 (attack 10 ms)limitthr −40, attack 0gatethr −45, ratio 4lowpass200 Hzhighpass4 kHzbandpass1 kHz q 2notch1 kHz q 2limitholds the peak at exactly the threshold (−39.999999).lowpassdrops the high band 58 dB and leaves low −3 dB;highpassdrops low 52 dB and leaves high −1 dB;bandpasscuts both outer bands ~20 dB;notchat q 2 is nearly inert on a broadband bed (mid −0.6 dB) — all the right direction and magnitude for a single biquad.gateon this bed: the body (peaks well above −45) is ~0.35 dB down and the quiet last second goes −64.3 → −69.1 dBFS RMS (envelope fromget_gallery_metadata), so the gate acts where it should; the follower is peak-tracking, so a track whose peaks sit above threshold stays open even when its RMS is below — expected for a gate, noting it so nobody reads the small whole-track RMS shift as "not working". Duration / sample_rate / channels unchanged on every arm (19.666656 s / 32000 / 1).compresswith the default 10 ms attack leaves the single-sample peak untouched (RMS −1 dB) — normal compressor behaviour.Job
a542ea98369a—compressthr −40, ratio 4,attack_ms: 0: peak −37.2215 dBFS. Predicted −40 + (40 − 28.886)/4 = −37.2215. The ratio math is exact.Refusals. Domain-table ones are free at
validate_workflow:ratio: 0→ error atsteps[1].task.arguments.ratio("compress_audio needs 'ratio' above zero, got 0"),cutoff_hz: -5→ atsteps[3].task.arguments.cutoff_hz. In-body ones fail the job with a clear message:mode: "expand"(joba542ea98369a) →compress_audio mode must be one of ('compress', 'limit', 'gate'), got 'expand';threshold_dbfs: 3(jobeb0ef01f09c6) →compress_audio 'threshold_dbfs' cannot be above full scale (0);cutoff_hz: 16000at 32 kHz (joba8e727d9dcc4) →filter_audio 'cutoff_hz' (16000) must be below the Nyquist frequency (16000.0) for sample_rate 32000;kind: "shelf"(jobb91568a540ae) →filter_audio kind must be one of ('lowpass', 'highpass', 'bandpass', 'notch'), got 'shelf'.Not a #200 finding, but seen on the way:
analyze_audio's band figures (low_dbfs−91 on a track whose RMS is −50) look like per-bin rather than per-band levels — that belongs to #207's verify, not here.Matches
expected:(an EQ/filter, compressor, limiter and gate all exist and shape the track). Closing as verified; adding regression case S-F047 toregression-suite-smoke.md. All outputs and theqa-verify-200workspace deleted.Develop (#241)