Repository navigation
dissolve_videos/concat_videos: size mismatch between asset: shots passes validate, fails at run, and no task can resize a video to fix it #504
Description
Tool/endpoint: validate_workflow + run_workflow on templates/dissolve-between-shots (task dissolve_videos); same class of gap for concat_videos.
Repro (workspace qa-ep75; all inputs are shared assets, so it rebuilds anywhere):
validate_workflow(name="templates/dissolve-between-shots", workspace="qa-ep75", arguments={"shots": ["asset:qa-cast/ep62-shot1-accuse.mp4", "asset:qa-cast/ep75-shot-ltx2stage-hal.mp4", "asset:qa-cast/ep62-shot2-deflect.mp4"], "score": "asset:qa-cast/ep15-song.mp3", "sample_rate": 48000, "total_frames": 345, "match_levels": "rms", "match_levels_dbfs": -24, "score_gain": 0.35})
→valid: true,warnings: [], estimate 0.1 min.run_workflowwith the same arguments and the bound acknowledgement → job98ed5f36db43failed after 3 s:
Workflow execution error: dissolve_videos needs every video at one size: video 0 is 960x544, video 1 is 1536x896
ep62-shot* are H3 shots (960×544). ep75-shot-ltx2stage-hal.mp4 is templates/ltx2/two-stage at its defaults (full_width 1536, full_height 896), job 6492d239d6ec, seed 75.
Expected
validate_workflowalready resolvesasset:refs, andget_gallery_metadatareports each asset'swidth/height. So a list ofasset:videos going intodissolve_videos/concat_videosat different sizes could be an error at validate, the way an out-of-bound variable is, instead of an error at run time. The standing-task rules call "validate passed, run failed" a ticket.- Also friction, and possibly the bigger ask: sample-rate mismatches get auto-resampled with a warning (concat_videos: sample-rate mismatch fails mid-run with no remedy named and no mention of resample_audio anywhere near it #108/dissolve_videos: sample-rate mismatch between shots fails mid-run with no remedy and no validate_workflow pre-flight — missed #108's concat_videos fix #287), but a size mismatch has no remedy.
list_taskshas no video resize/letterbox task (theresize_*processors are image processors), and neither join task has asize/resizeargument. As things stand, the only way to cut a LTX-2.5 full-size shot into an H3 episode is to regenerate the shot at 960×544. None of the join task docs or thetwo-stagetemplate mention that. An opt-inresize(fit/crop to the first video's size) on the join tasks, or a per-video resize task, would fix it; failing that, the join tasks'videosdescription should state the one-size rule and the two-stage template should warn that its full size won't join with H3 shots.
Actual: validate is clean and the run dies. There is no documented way to reconcile sizes.
Filed by the tester agent, model claude-opus-5-5 via anthropic, standing task ep75.
Activity
dkackman commentedon Sep 27, 2026
triage: escalate — this asks for two new pieces. One is a new
validate_workflowrule that reads the pixel sizes of the videos going intodissolve_videos/concat_videosfrom their media. The other is a video resize: a new task, or aresizefit/crop argument on both join tasks. Both are new engine and validation code that a narrow repro check would not cover (flat lists,for_each/gather:inputs,previous_result:inputs whose size is unknown until run time, and the fit/crop/letterbox options). Don should decide the scope before a fix session builds it. A cheap docs-only part could go first if approved: state the one-size rule in both tasks'videosdescription, and say inltx2/two-stagethat its full size won't join with 960x544 H3 shots. Triaged by claude-opus-5-5 via anthropic.dkackman commentedon Sep 27, 2026
Decision: joining outputs across models with different native resolutions (LTX-2.5 full size into an H3 episode) isn't something we need to support, so declining the resize/fit/crop argument on the join tasks specifically.
Approving the rest:
dissolve_videos/concat_videosinputs, same shape as the existing sample-rate mismatch check (concat_videos: sample-rate mismatch fails mid-run with no remedy named and no mention of resample_audio anywhere near it #108/dissolve_videos: sample-rate mismatch between shots fails mid-run with no remedy and no validate_workflow pre-flight — missed #108's concat_videos fix #287) — catches a mismatched-size join before spending GPU time, without needing to solve resize.videosargument description, and note inltx2/two-stage's docs that its default full size won't join with 960x544 H3 shots.The general resize/upscale capability is worth having on its own — split off as #512 (idea, owner:lead) — but decoupled from these join tasks.
dkackman commentedon Sep 27, 2026
triage: work — Don approved a scoped version (see his decision comment): (1) a validate-time size-mismatch check for the
videosinputs ofdissolve_videos/concat_videos, and (2) docs: the one-size rule in both tasks'videosdescription, plus a note inltx2/two-stagethat its default full size won't join with 960x544 H3 shots. No resize/fit argument (declined), and general resize is split off as #512. The check can only read sizes it can resolve before the run (asset:/output:media). Aprevious_result:/gather:input has no size until run time, so leave it to the existing run-time error; don't guess. Nothing for this is on develop yet. Triaged by claude-opus-5-5 via anthropic.style(tests): #504 - ruff format test_video_size_errors.py
dkackman commentedon Sep 27, 2026
Fixed (Claude Sonnet 5, Anthropic).
Added
dw/video_size_errors.py: avalidate_workflow-time check fordissolve_videos/concat_videos, mirroring the existingdissolve_frame_errors.py(#400) sample-rate/frame-count pattern. For everyvideosentry that's already resolvable before the run (asset:/output:/literal path), it probes the real dimensions and refuses a size disagreement atsteps[N].task.arguments.videoswith a message like:A
previous_result:/variable:/item:/gather:reference, or a{"location": ...}dict, has no known size before the step runs and is deliberately left to the existing run-timecheck_same_frame_sizeerror - not a gap, per the "don't guess" scope Don set.Docs updated to state the one-size rule: both tasks'
videosargument docstring (served via MCPget_task) and their rows indocs/TASKS.md(a served guide).workflows/templates/ltx2/two-stage.jsonalready carries the note about its default size not joining with 960x544 H3 shots (landed earlier in 1cf752c, already on develop) - no further edit needed there.Declined per Don's scope: no
resize/fit/cropargument on the join tasks - that's #512 (owner:lead).New test:
tests/test_video_size_errors.py- exercises the real decode path (genuine mp4 fixtures viaprobe_media), covering both commands, a matching-size clean pass,previous_result:/mixed-resolvability deferral to run time,output:references, single-video no-op, and wiring intoWorkflow.validation_errors().Shipped: merged
fix/504-video-size-mismatch-validateintodevelop(fdfdb02), then a ruff-format fix-forward (b8fd862) after the harness guard caught formatting in the new test file.developdeployed to lem atb8fd862, confirmed healthy, 2026-09-27T12:38:08Z.Possible regression-suite case for
regression-suite-smoke.md, if you want to add it after verifying:validate_workflowon adissolve_videos(orconcat_videos) step whosevideosare twoasset:references at different resolutions (e.g. 960x544 and 1280x720) - expect a validation error atsteps[0].task.arguments.videosnaming both sizes, no job queued.dkackman commentedon Sep 27, 2026
Verified by the tester agent: model
claude-opus-5-5viaanthropic, against lem at develop @ b8fd862.validate_workflow(name="templates/dissolve-between-shots", arguments={shots: [ep62-shot1-accuse, ep75-shot-ltx2stage-hal, ep62-shot2-deflect], score: ep15-song.mp3, sample_rate 48000, total_frames 345, match_levels rms, match_levels_dbfs -24, score_gain 0.35}). I ran it without theworkspacepin becauseqa-ep75no longer exists; all the inputs are sharedasset:qa-cast/...files. Result:valid: false, one error atsteps[0].task.arguments.videos: "dissolve_videos needs every video at one size: video 0 is 960x544, video 1 is 1536x896". So validate now refuses it and no job is queued.concat_videoswithvideos: [asset:qa-cast/ep62-shot1-accuse.mp4, asset:qa-cast/ep75-shot-ltx2stage-hal.mp4]fails validate the same way, at the same path, namingconcat_videos.concat_videoswith both ep62 shots (both 960x544) returnsvalid: truewith no errors or warnings.get_taskvideosdescriptions forconcat_videosanddissolve_videosboth state the one-size rule, say there is no reconciliation for a size mismatch, and say that a mismatch among asset:, output: or literal-path videos is refused at validate while one known only at run time still fails at run.get_workflow("templates/ltx2/two-stage")says in its description that the delivered size is always a multiple of 64, which rules out 960x544, so a two-stage clip can't be cut into an H3 episode without a resize step downstream.The resize argument was declined per Don's scope and split off to #512. I did not test that part.
fix(mcp): dkackman#504 - validate-time size-mismatch check for dissol…