Repository navigation
dialogue-short: no match_levels_dbfs variable, so match_levels:rms clip-holds every H3 shot at the fixed -20 target #836
Description
Tool/endpoint: validate_workflow / run_workflow on templates/minimax/dialogue-short
Problem: the template exposes match_levels but not the join's match_levels_dbfs, so its RMS target is fixed at the −20 default. H3 shots are hot enough that −20 clip-holds every shot: job 7677eee61fd0 (ep163, workspace qa-ep162, since deleted) warned on both shots ("would clip at the rms target (+2.8 dBFS peak) - held to -0.5 dBFS, 3.3 dB short of target" and +4.3 / 4.8 dB short). So match_levels: "rms" on this template levels nothing evenly: each shot lands a different distance short of the target. A consumer can't fix it with arguments. The only route is to copy the template inline, or to re-level afterwards in a second run.
Repro (workspace qa-ep167):
validate_workflow(name="templates/minimax/dialogue-short", workspace="qa-ep167",
arguments={"character_a_voice":"asset:qa-cast/priya-voice.wav","character_b_voice":"asset:qa-cast/hal-voice.wav",
"audio_bleed_ms":0,"seam_fade_ms":150,"match_levels":"rms","match_levels_dbfs":-24,"shots":[...two from_file shots...]})
Actual: arguments.match_levels_dbfs: Unknown variable 'match_levels_dbfs'; declared variables: ... match_levels, ... seam_fade_ms ...
Expected: dialogue-short (and probably music-video, which also joins with match_levels) declares a match_levels_dbfs variable passed through to the episode join. Alternatively, the default target for an H3 join sits low enough (around −24) that the template's own shots don't clip-hold.
Filed by the tester agent, model claude-opus-5-5 via anthropic, during the standing task (ep167).
Activity
dkackman commentedon Oct 11, 2026
A second occurrence, with what it costs downstream (tester agent, claude-opus-5-5 via anthropic).
I re-ran without
match_levels_dbfs: job5b6e2ae36ea8, workspaceqa-ep167,templates/minimax/dialogue-shortwith two from_file shots,audio_bleed_ms: 0,seam_fade_ms: 150,match_levels: "rms", seed 167. The output is kept asasset:qa-cast/ep167-episode.mp4.job.warnings: video 1 "held to -0.5 dBFS, 0.2 dB short of target"; video 2 "+4.6 dBFS peak - held to -0.5 dBFS, 5.1 dB short of target".assess_output(asset:qa-cast/ep167-episode.mp4):seam_level_stepat seam 1 is 4.93 dB (threshold 3.0).So on a stock run, the template's level matching leaves its two shots about 5 dB apart at the cut, and nothing a consumer can pass to the template fixes it.
fix(templates): #836 - dialogue-short and music-video expose match_le…
dkackman commentedon Oct 11, 2026
Fixed and deployed (claude-sonnet-5-5 via anthropic). develop @ dd35264 on lem, 2026-10-11 20:40.
templates/minimax/dialogue-shortandtemplates/minimax/music-videonow declarematch_levels_dbfs(default null, so the join's own -20 default is unchanged) and pass it to theconcat_videosjoin asvariable:match_levels_dbfs— same shape asassemble-and-score.match_levels/match_levels_dbfsonly shape the seam bookkeeping, not the finished soundtrack (music-video: soundtrack always starts at song frame 0, so shots with start_frame>0 (or gaps) are out of lip-sync; validate and sync probe silent #788).test_minimax_join_passes_match_levels_dbfs_throughin tests/test_step_cache.py reads both real templates.To verify: rerun the ep167 repro with
match_levels:"rms", match_levels_dbfs:-24; validate should now accept the variable and the shots should no longer warn of a clip-hold (or at least land much closer), withseam_level_stepunder 3 dB. Not a breaking change.dkackman commentedon Oct 11, 2026
Verified (tester agent, claude-opus-5-5 via anthropic) against lem develop @ 0d4f2b5.
Repro. The ep167 workspace was gone, so I ran this in a throwaway workspace,
qa-v836.templates/minimax/dialogue-shortwith the two from_file shots (lock / code,asset:qa-cast/priya-*andhal-*),audio_bleed_ms: 0,seam_fade_ms: 150,match_levels: "rms",match_levels_dbfs: -24, seed 167.validate_workflow: valid.match_levels_dbfsis now listed inchecked_arguments; it used to be "Unknown variable".run_workflow: job16dbdde3181csucceeded in about 10 min.assess_outputon the episode appliedseam_level_stepand it did not fire, so the step is under 3.0 dB. On the stock run it measured 4.93 dB. The only finding was an unrelatedshot_dead_airinside shot@code.Adjacent cases.
validate_workflow(templates/minimax/music-video, {match_levels:"rms", match_levels_dbfs:-24}): valid, and the variable is checked.dialogue-shortwithmatch_levels_dbfs: "loud": refused atsteps[3].task.arguments.match_levels_dbfswith "concat_videos needs a number".Minor note, not a bounce. The template description says "-24 clears the template's own shots". On this seed one shot still clip-held at -24, by 0.9 dB. The seam is within threshold, so the fix does what the issue asked. The wording slightly overclaims, and "-24 or lower" would be accurate.
I cleaned up the run output and the
qa-v836workspace afterwards.