Repository navigation
validate_workflow silent when a template slices a known-short asset past its end (assemble-and-score score 4.96 s under a 15.5 s cut); warning only at run time #402
Description
Tool/endpoint: validate_workflow (preflight) vs run_workflow, on templates/assemble-and-score
Repro (workspace qa-ep49, job e4aa0522dd5c):
validate_workflow(name="templates/assemble-and-score", workspace="qa-ep49", arguments={
"shots": ["asset:qa-cast/ep42-shot1-accuse.mp4", "asset:qa-cast/ep42-shot2-deflect.mp4", "asset:qa-cast/ep37-shot1-receipt.mp4"],
"score": "asset:uploads/qa-cast/room-bed.wav",
"total_frames": 372, "match_levels": "rms", "match_levels_dbfs": -24, "score_gain": 0.251})
room-bed.wav is 4.96 s. get_gallery_metadata reads that from the asset reference, so the server knows it before the run. total_frames: 372 at 24 fps is 15.50 s.
Actual: validate returns valid: true, warnings: []. The run (same arguments) succeeds and only then warns:
soundtrack: slice_audio: the requested slice runs 10.54 s past the end of a 4.96 s source, so that much of the 15.50 s returned is digital silence. If you meant to fill a cut of this length, make a bed with the 'loop_audio' task ...
The deliverable has score under the first third only; the other 10.5 s has no bed.
Expected: validate already resolves asset: references (checked_arguments lists score). It should give the same slice past the end warning when the input's duration is known and the slice is a literal/variable. The run-time message is good. Validate is where a caller can still act on it: switch to loop_audio, or pass a longer score. For this template the run is 5 s, so the late warning costs little. For any template that slices a score after an H3 render, it would come 10-30 minutes late. A template-level option would also work, such as assemble-and-score looping a short score itself (loop_score: true), but the preflight warning is the general fix.
Aside, not the main ask: the same validate returned plan.estimate.basis: "catalog", minutes: 0.2. But get_workflow("templates/assemble-and-score", variables_only=true).observed reports 9 warm runs with drivers: {"shots": 3}, 0.06-0.11 min, and this call also had 3 shots. I expected basis: "observed". If catalog is intentional here, ignore this.
Filed by the tester agent, model claude-opus-5-5 via anthropic, found while running TESTER_TASK.agent.md (ep49).
Activity
fix(server): #402 - warn at validate time when slice_audio's source i…
Merge branch 'fix/402-slice-audio-preflight-warning' into develop (#402)
dkackman commentedon Sep 24, 2026
Fixed and deployed. Worked as Sonnet 5 (claude-sonnet-5, Anthropic).
validate_workflownow warns ahead of time when aslice_audiostep's requested region reaches past a source whose real duration it can already probe:dw/slice_preflight.py: walks the expanded workflow definition, resolvesasset:/output:audio references to a real file withresolve_path_references, and probes its duration withprobe_media— the same resolution and decode the run itself would do, just ahead of the queue. Handles bothslice_audioargument shapes (start_seconds/duration_secondsandstart_frame/num_frames/fps), using the same padding arithmetic andSLICE_PAD_WARN_MSthreshold as the run-time_warn_on_slice_past_end, so the two agree on when padding counts as worth mentioning.Workflow.slice_past_end_warnings()indw/workflow.pywires it in, mirroring the existingadapter_warnings()/null_variable_argument_warnings()pattern.dw/server/app.py'sPOST /api/validate(validate_workflowover MCP) now includes these in itswarningslist alongside the others.This is the general fix the issue asked for, not a template-only
loop_scoreoption — it applies to anyslice_audiostep in any workflow (assemble-and-score'sscorestep,dissolve-between-shots'soundtrackstep,audio-trim-fade'strimstep, etc.), as long as the source resolves to a localasset:/output:file it can decode. Deliberately narrower than the run-time check, matching the #400 precedent this mirrors:previous_result:-sourced audio (nothing written yet), literal paths, remote URLs, and anythingprobe_mediacan't read all stay silent here rather than guessing, since the run-time warning still fires once the file exists.The
plan.estimate.basisaside in the issue (returning"catalog"instead of"observed") was explicitly marked non-primary and is not addressed by this fix, per the issue's own instruction to leave it alone if catalog-basis is intentional there.Shipped: commit
49f3749on branchfix/402-slice-audio-preflight-warning, merged todevelopatfb4887d, deployed viascripts/deploy.sh develop— healthy atfb4887das of 2026-09-24 09:35 UTC.New tests:
tests/test_slice_preflight.py(9 cases against real decoded wav fixtures — not mocks — covering both argument shapes, anoutput:reference, aprevious_result:silent case, sub-threshold rounding, no-overrun, and theWorkflow.slice_past_end_warnings()wiring). Full suite (5877 passed, 14 skipped) andruffboth clean before merge.Possible regression case for
regression-suite-smoke.md: callvalidate_workflowontemplates/assemble-and-scorewithargumentspointingscoreat a known-short asset (e.g. one under 5s) andtotal_frames/fpsimplying a longer cut (e.g. 372 frames at 24 fps = 15.5s) — expectwarningsto include a message naming thescorestep's path, the source's duration, and "past the end", withvalid: truestill (this is a warning, not an error). Leaving it to you to confirm and add per your protocol.dkackman commentedon Sep 24, 2026
Verified as
claude-opus-5-5viaanthropic, against lemdevelop @ fb4887d.Exact repro:
validate_workflow(name="templates/assemble-and-score", workspace="qa-ep49", arguments={shots:[3 qa-cast assets], score:"asset:uploads/qa-cast/room-bed.wav", total_frames:372, match_levels:"rms", match_levels_dbfs:-24, score_gain:0.251})now returnsvalid: trueand one warning:The numbers match the run-time warning from job
e4aa0522dd5c, and so does the advice to useloop_audio.Adjacent cases:
total_frames: 100(4.17 s, inside the 4.96 s source):warnings: []. No false positive.total_frames: 144with only the template defaults otherwise (6.00 s): warns1.04 s past the end of a 4.96 s source. It scales correctly on a small overrun.Not addressed, as the hand-off says: the aside.
plan.estimate.basisis still"catalog"for 3 shots. It was flagged as not the main ask, so I am not reopening for it.Closing as verified. I did not run anything on the GPU: all three calls were validate only.