Release 0.4.0

[dkackman]

Release 0.4.0: develop @ e86fc20 into master (262 commits since 0.4.0-beta.7).

Gates (all green on e86fc20)

• scripts/preflight.sh on 3a28146: ruff clean; 6,585 pytest, 335 UI unit, 108 Playwright e2e passed. Full pytest on e86fc20: 6,592 passed. CI and CodeQL on e86fc20 pass, with no new alerts relative to master.
• Regression suites on 3a28146 (security 37 cases, smoke 68, complete 109): no security failures. Issues get_task: argument renamed name -> command (suite/docs still use the old name) #447–C-F108: dissolve of a multi-shot previous_result: expands inner shot names instead of collapsing to the step name #460 were filed; none blocks the release. They are suite drift, two performance readings and one metric bug, plus C-F020: dissolve-between-shots warns on identical-rate shots due to template's pinned sample_rate (likely stale suite expectation post-#287) #453/C-F029: templates/assemble-and-score warns on routine internal trim/pad, expected warnings: [] #454, which are fixed here. They go to the agent loop after the release.
• Confirmation runs: security again on 0d3a6c9 (nothing new) and smoke on e86fc20 (nothing new). assemble-and-score and dissolve-between-shots on the C-F029: templates/assemble-and-score warns on routine internal trim/pad, expected warnings: [] #454/C-F020: dissolve-between-shots warns on identical-rate shots due to template's pinned sample_rate (likely stale suite expectation post-#287) #453 inputs now finish with warnings: [].
• Whole-diff review of the overnight range, for security and correctness. The fixes it asked for are the last commits here:
  • /outputs asset misses no longer name server paths;
  • dissolve_videos after an ltx2/keyframes clip: shot map ends 16 then 32 samples short of the frame grid, no pad/warn (#428 class) #435's mux warning and get_job_events has no kind filter: confirming 4 applied-value log lines costs ~20 KB, most of it memory events #436's kinds filter;
  • the docs the overnight fixes contradicted;
  • stock joins warning on rounding (C-F020: dissolve-between-shots warns on identical-rate shots due to template's pinned sample_rate (likely stale suite expectation post-#287) #453/C-F029: templates/assemble-and-score warns on routine internal trim/pad, expected warnings: [] #454);
  • paths routed through validators CodeQL can see, and model_index.json component keys held to the snapshot.

After merging

1. scripts/release.sh 0.4.0 on master.
2. gh release edit v0.4.0 --notes-file <the section below>.
3. Merge master back into develop and bump develop to the next version.

Release notes (from docs/RELEASING.md)

Breaking and behaviour changes

• download_output over a dw.serve --mcp endpoint refuses a call with no
  destination. It used to write into the server's own directory (export_job/list_gallery return origin-less URLs and export_job tells an MCP-only agent to fetch them; mounted download_output strands files in the workspace root #353).

• Untrusted workflows are refused in more cases (#407 stage 1: trust gate and URL host #409-#407 stage 5: pixel limit on decode #413):

  • a *_type that doesn't resolve to a class, or that isn't a kind a
    workflow constructs: a diffusers or transformers model, pipeline,
    scheduler, tokenizer or processor, a quantization config, an auto
    factory, a diffusers reference/condition type or an attention processor.
    A plain torch class such as torch.nn.Linear is now refused;
  • constant: walks through _ names or out of the allowed packages;
  • URLs with backslashes;
  • text/html and text/xml result types;
  • media hosts that aren't globally routable, including 100.64/10 (CGNAT,
    and so Tailscale);
  • more than 5 redirects;
  • images over 50M pixels.

  Listings and export zips drop symlinks that escape their root.
  --trust-workflows lifts all of these.

• run_workflow validates the caller's arguments when it queues the job
  (run_workflow queues a job whose argument-resolved result.content_type is text/html; validate_workflow refuses it #414/run_workflow refuses a document-default text/html content_type that the caller's argument overrides; validate_workflow accepts it #415). validate_workflow(arguments={}) checks a run with no values
  supplied, not just the document (Required task argument fed by a null-defaulted variable is reported as "the step does not supply" it, and save_workflow refuses the document #364).

• A fractional value for an int variable is refused (Variable coercion silently truncates a fractional argument to an integer-defaulted variable (score_gain 0.3 -> 0 deleted the score) #338), and so is a
  still image passed as a video argument (loop_frames invites a still image for video, but a .png path/asset there fails at run time ("Video file extension not allowed: .png"), and validate passes it #347).

• templates/minimax/music normalizes to -3 dBFS instead of -1, so its output
  is quieter (templates/minimax/music: balanced step at peak_dbfs -1.0 still clips after mp3 encode (+0.56 dBFS); engine's own warning prescribes -3 #362).

• Every response carries X-Content-Type-Options: nosniff and
  X-Frame-Options: DENY. Active document types under /outputs and
  /inputs are served with Content-Security-Policy: sandbox.

• A validate-time probe reads only a literal media path that the run itself
  would be allowed to read.

• A dict or list passed to a string-typed variable is refused (validate_workflow passes a dict for a string location variable (templates/ltx2/keyframes first_image); run fails with repr-as-path #433).
  templates/ltx2/keyframes takes first_image/last_image as plain
  strings, not {"location": ...} (templates/ltx2/keyframes fails with its own defaults: Unsupported frames type for condition 0: dict (validate clean) #431/validate_workflow passes a dict for a string location variable (templates/ltx2/keyframes first_image); run fails with repr-as-path #433).

• loop_frames returns float32 frames in [0, 1] instead of uint8, the shape
  LTX2ReferenceCondition needs; a keyframe condition still wants
  frames_as_array. ltx2/reference-sheet's default asset is now
  asset:reference_sheet.jpg (ltx2/reference-sheet: succeeds with no warnings but every frame is blocky mush; default asset:reference_sheet.png missing #444).

• validate_workflow refuses a components name the pipeline doesn't
  register; duration_head is gone from the in-context LTX-2 templates
  (templates/ltx2/restore-decompression fails at load: LTX2InContextPipeline has no component 'duration_head' (validate passes) #442).

• A {"media_type": "image"} reference on a video argument loads as a
  one-frame still (frame_grid: validator suggests {media_type: image} for stills; validates clean, fails at run time 'Video file extension not allowed: .jpg' #443).

• pair_audio fit: "video" always fits, and warns on any nonzero gap
  (pair_audio fit="video" (and slice_audio past the end) leave a track 15 samples short with no pad and no warning #428/pair_audio fit=video warns 'last part of the cut has no soundtrack' for a 1-sample (0.00 s) pad from ordinary rate rounding #429). concat_videos and dissolve_videos pad a short joined
  track to the frame grid, warning (joined_audio_padded_to_frames) only
  when the pad is a frame or more; a residual the AAC mux trims off is
  logged, or warned as joined_audio_short_after_mux from a frame up.
  media.shots is measured against the file as written (concat_videos: after a resampling join the shot map overruns the written file by 30 samples (assess_output shot_span_overrun on a stock join) #426/dissolve_videos after an ltx2/keyframes clip: shot map ends 16 then 32 samples short of the frame grid, no pad/warn (#428 class) #435/C-F029: templates/assemble-and-score warns on routine internal trim/pad, expected warnings: [] #454).
  Neither warns about resampling inputs that agree to a pinned
  sample_rate (C-F020: dissolve-between-shots warns on identical-rate shots due to template's pinned sample_rate (likely stale suite expectation post-#287) #453).

• New warnings: match_levels_near_silent (match_levels gains a near-silent input +29.9 dB with only a log line, no warning (dissolve_videos, ep65) #434), and shot_span_overrun
  from the probes plus a validate-time check (analyze_seams: a shots= record running past the file's end (124+300 on a 248 f video) is clipped silently, no warning #425).

• Error text changed: delete_workspace (delete_workspace refusal names a nonexistent 'acknowledged=true' parameter before the correct acknowledged_cost #437/delete_workspace on a nonexistent workspace tells the caller to repeat with acknowledged_cost=True #438), the sub-workflow path
  refusal names the places it looked (SE-F023: sub-workflow path refusal message doesn't list where it looked #422), and /outputs/asset:... misses
  name the asset without server paths.

New

• The assess_output tool and GET /api/gallery/{name}/assess, plus the
  probe tasks analyze_shots, analyze_seams and analyze_sync_drift
  (#378 stage B: probes as tasks (analyze_shots, analyze_seams, analyze_sync_drift) and the rules table #387/#378 stage C: assess_output MCP tool, route, and the guide section #388).
• A joined video records its shot boundaries (media.shots).
  get_output_frames(seams=true) uses them, so it no longer needs
  boundaries (#378 stage A: record shot boundaries on joined video (manifest, metadata, seams without boundaries) #385).
• Run versions (v<N>):
  • list_gallery returns run_id/version and filters by folder and
    version;
  • output:<wf>/v<N>/<file> references;
  • wait_for_job returns run_version;
  • export zips download as <wf>-vN-<job>.zip.
• list_gallery(media=true) adds durations, and output: names work in
  gallery reads (Gallery reads reject an output:-prefixed name with a misleading "path does not exist"; list_gallery can't show a take's duration #356).
• DW_PUBLIC_URL adds absolute URLs to gallery and export responses.
  export_job also returns auth_required and open_url (export_job/list_gallery return origin-less URLs and export_job tells an MCP-only agent to fetch them; mounted download_output strands files in the workspace root #353).
• A grade task for images and video: exposure, contrast, saturation and
  temperature/tint (Add a CPU grade task (v1: exposure, contrast, saturation, temperature/tint) for images and video #349).
• The templates/minimax/shots-batch H3 template (Add templates/minimax/shots-batch: list-driven H3 shot generation with no in-job assembly #352).
• Every generative template takes a seed argument (Catalog seed convention: every generative entry takes "seed": "variable:seed" (pinned ones keep their value; unseeded ones default null) #351).
• normalize_audio(target_lufs), and integrated_lufs plus true peak in
  media metadata (Loudness: report integrated LUFS + true peak in media metadata, and add target_lufs to normalize_audio #361).
• gain_audio with no region gains the whole track (gain_audio with no region validates clean, then fails at run time; the schema marks every region arg optional #395).
• world_fade_out_ms on assemble-and-score (assemble-and-score / dissolve-between-shots: expose a world_fade_out_ms so a score's written ending isn't buried under the world track #339).
• Download progress shows in phase_detail (A download a job triggers is invisible: 17.8 min of phase_stall warnings, nothing in list_downloads, no byte progress #343). phase_stall events
  now read as informational (phase_stall events read as a fault during phases the H3 skill calls healthy, and they bump event_count so they look like liveness #357).
• workflow, inline_workflow and prompt also accept a JSON string. A
  mistyped workflow name gets suggestions from the catalog (get_workflow: 'Unknown workflow: dialogue-short' gives no hint that the catalog name is templates/minimax/dialogue-short #397).
• Host caches are released when each job ends (dialogue-short: host RSS grows ~3.6 GB per shot@ member and ~10.6 GB stays resident after H3 release (#344 part b, measured) #368), and the skills point
  at clear_memory.
• get_job_events(kinds=...) and ?kinds= on the event-log route; a kind
  matches an event's event or its kind, so ["phase_stall"] selects
  one warning type (get_job_events has no kind filter: confirming 4 applied-value log lines costs ~20 KB, most of it memory events #436).
• get_memory reports the step cache's entries and retained_bytes
  (Step cache: a stale-entry eviction never subtracts its size, so _retained_bytes only grows #418).
• get_output_image and /outputs resolve asset: references (get_output_image: asset: reference returns bare 'Not Found' (siblings accept asset:) #445), and
  get_output_frames(seams=true) works on linked assets (get_output_frames(seams=true) refuses a linked asset whose media.shots get_gallery_metadata/assess_output both read #430).
• Compact assess_output lists each finding once (assess_output compact: the same shot_span_overrun finding is listed 3 times #427). Shots are named by
  their source when joined inputs already carry shots (concat_videos: previous_result input named "video 2" again when an earlier input carries inner shots (#396 regression on multi-shot inputs) #432).
• A task-only workflow's run history counts, so its estimate can quote
  basis: observed (validate_workflow: dissolve-between-shots at 3 shots quotes basis:catalog though get_workflow shows 24 observed runs at drivers {shots: 3}, no reason given #439). The Music 3 hint no longer shows on video
  (get_gallery_metadata: next hint on a video output includes Music 3 ceiling guidance #441).

Fixes

• The step cache's retained-byte count no longer only grows (Step cache: a stale-entry eviction never subtracts its size, so _retained_bytes only grows #418).
• templates/ltx2/keyframes (templates/ltx2/keyframes fails with its own defaults: Unsupported frames type for condition 0: dict (validate clean) #431), restore-decompression (templates/ltx2/restore-decompression fails at load: LTX2InContextPipeline has no component 'duration_head' (validate passes) #442) and
  reference-sheet (ltx2/reference-sheet: succeeds with no warnings but every frame is blocky mush; default asset:reference_sheet.png missing #444) run with their own defaults again.
• Joined audio and shot maps stay on the frame grid through repeated joins
  (dissolve-between-shots: last shot's num_samples is 1 over its frame span (227851 vs 227850) at 44.1k/24fps #423, concat_videos: after a resampling join the shot map overruns the written file by 30 samples (assess_output shot_span_overrun on a stock join) #426, pair_audio fit="video" (and slice_audio past the end) leave a track 15 samples short with no pad and no warning #428, dissolve_videos after an ltx2/keyframes clip: shot map ends 16 then 32 samples short of the frame grid, no pad/warn (#428 class) #435).

🤖 Generated with Claude Code