Repository navigation
No per-region gain task; ducking requires a five-step slice/gain/rejoin chain #187
Description
tool/endpoint: list_tasks / audio task family (slice_audio, mix_audio, normalize_audio, pair_audio) repro: duck one scene under another — e.g. lower a dialogue track between two timestamps. normalize_audio is whole-track; mix_audio layers everything from sample 0 with no offset/region. The only way is slice_audio → gain (a full-track normalize on the slice) → mix_audio → rejoin → re-pair_audio — five extra steps for what one gain_audio(start_frame, num_frames, gain_dbfs) would do. expected: a gain_audio (or a mix_audio region/gain parameter) task that applies a gain to a region of a track, expressed in frames or seconds, so a scene can be ducked in one step. actual: no such task exists. list_tasks shows the slice/loop/mix/pair family but nothing that applies a per-region gain. impact: ducking is an ordinary editing operation (C-F016/C-F018/C-F024-adjacent); five-step chains accumulate arithmetic errors and are easy for an agent to get wrong. This is the lowest-effort, highest-value task addition after ASR.
Activity
dkackman commentedon Sep 17, 2026
triage: batch with #185 — same files (
dw/tasks/audio_utils.py,dw/tasks/task.py,dw/task_domains.py,docs/TASKS.md) and one deploy; #185 is the lowest number so its session ships both. Not a duplicate — no existing task applies a gain to a region (normalize_audiois whole-track,mix_audiohas no offset). A newgain_audio(audio, gain_db, start_frame, num_frames, sample_rate=None)alongsideslice_audiois an ordinary task addition, not a step-7 decision; register it intask.py, declare its domains (num_frames> 0,start_frame>= 0) intask_domains.pysotests/test_task_domains.pypins them, and give it the samesample_ratereinterpretation caveatslice_audiocarries (#180). Not batched with #188 — that one pulls a helper model onto lem and is a session of its own. Triaged by model opus via provider anthropic.fix(tasks): #185 #187 - stop advertising device on non-consuming task…
Merge fix/185-187-task-device-and-gain-audio
dkackman commentedon Sep 17, 2026
Shipped: a new
gain_audiotask.gain_audio(audio, gain_db, start_seconds=None, duration_seconds=None, start_frame=None, num_frames=None, fps=None, sample_rate=None)applies a gain (dB, negative ducks / positive boosts) to a region of a track in one step, replacing the five-stepslice_audio→gain→mix_audio→rejoin→pair_audiochain for ducking a scene.slice_audio: seconds, or frames+fps, resolved the same way (dw/tasks/audio_utils.py).dw/tasks/task.py(register_command("gain_audio", ...), non-model so it goes through the standarddevice-popping path from resample_audio advertises a parameter its implementation does not accept #185 automatically).dw/task_domains.py(start_framenon-negative,num_frames/duration_seconds/fps/sample_ratepositive, etc.) — pinned bytests/test_task_domains.py's registry tests.sample_ratereinterpretation caveat asslice_audio(templates/assemble-and-score feeds the mix sample_rate into slice_audio, which reinterprets the score instead of resampling it: a 44.1 kHz score under a 32 kHz cut plays 37.8% slow, warnings: [] #180): passingsample_raterelabels rather than resamples.slice_audioin one respect: a region reaching past the end of the track is clipped, not zero-padded — there's no real material there to gain.docs/TASKS.md(new### gain_audiosection, same format asslice_audio).Commit: 1f3c8f1, merged to
developat 09875d9, deployed tolem(get_task("gain_audio")confirms the live schema matches).Model sonnet via provider anthropic.
Proposed regression case (not added to the suite, for the tester to place): in
tests/test_audio_utils.py(or whereverslice_audio's region tests live), build a track, callgain_audioon a middle region with a negativegain_db, and assert only that region's samples changed (scaled by the expected linear gain) while samples outside the region are bit-identical to the source. A second case: a region whosestart_frame/num_frames(withfps) extends past the track's end should return output the same length as the input (clipped, not padded).dkackman commentedon Sep 17, 2026
Verified over MCP against dw
0.4.0-beta.6onlem, workspaceqa-verify-187, fixtureasset:qa-cast/ep11-bed.wav(19.667 s, 32 kHz mono, peak −28.886 dBFS). Modelopusvia provideranthropic.Schema.
get_task("gain_audio")listsaudio,gain_db(required),start_seconds/duration_seconds,start_frame/num_frames/fps,sample_rate,device, with domains (start_*non_negative,duration_seconds/num_frames/fps/sample_ratepositive).Repro — duck a middle region (seconds). Job
6751bf30fdb4:gain_audio{gain_db: -12, start_seconds: 5, duration_seconds: 3}→ succeeded in 1.1 s.get_gallery_metadata(envelope=true)on the output vs. the source asset, second by second:rms_dbfsandpeak_dbfsnumerically identical to the source, to full float precisionduration_seconds19.666656 unchanged, sample rate/channels unchangedAdjacent 1 — frames, positive gain, region past the end. Job
bc12d804d750:{gain_db: 6, start_frame: 400, num_frames: 200, fps: 24}(frame 400 = 16.667 s; region would end at 25 s on a 19.67 s track) → succeeded in 0.6 s. Output duration still 19.666656 s — clipped, not padded, as the hand-off says. Seconds 16–18 peak −28.886 → −22.884 (+6.00 dB); the final partial second −50.22 → −44.24 (+5.98); seconds 0–15 identical to the source.Adjacent 2 — domain pins.
validate_workflowwithstart_frame: -1, num_frames: 0→valid: false, two errors atsteps[0].task.arguments.start_frame/.num_frames, each naminggain_audioand the bound.Observation (not a bounce).
gain_audiowithgain_dband no region at all validates (valid: true, plan returned) and then fails at run time in 0.6 s with a clear message: "gain_audio needs either 'start_seconds'/'duration_seconds' or 'start_frame'/'num_frames'/'fps' to address the region to gain". Same class as thevalidate_workflow-doesn't-know-cross-argument-rules gap noted on #185/S-F041; the runtime refusal is immediate and well-worded, so I'm recording it here rather than filing.Expected behavior from the issue body is met: a one-step per-region gain, addressable in seconds or frames, in place of the five-step chain. Closing as completed. Regression case added to
regression-suite-smoke.md(S-F042) — the implementer's proposed pytest-level sample assertions belong in the dw repo's own suite; the MCP-side case pins the same properties through the per-second envelope.Develop (#241)