dw plugin: MiniMax H3 and LTX-2.5 composition skills (Part 4, package B)

[dkackman]

Part 4 of docs/proposals/agent-catalog-legibility.md, package B: model-specific knowledge packaged as a Claude Code plugin in this repo, one composition skill per model family. Spec: docs/superpowers/specs/2026-09-07-dw-plugin-skills-design.md. Everything is data (a marketplace manifest, a plugin manifest, two SKILL.md files, tests); the one engine change is the download_output message below.

The plugin

• .claude-plugin/marketplace.json at the root and plugins/dw/ beneath it, so the install is two commands the README shows: /plugin marketplace add dkackman/diffusers-workflow, /plugin install dw@diffusers-workflow. The plugin version is the engine's; scripts/release.sh bumps it and a test holds them equal.
• plugins/dw/skills/minimax-h3/SKILL.md: the shape decision over the family's templates (shot, sequence, audio), the rules that bite (17n+5 frames at 24 fps, the canvas limits, the LoRA/step coupling, Ref2VA input limits, no negative prompt), the cost quote with a fallback for templates that carry none, and the run-and-judge loop. Prompt format is not transcribed: the skill points at MiniMax's h3-prompt-writing skill and its two guides.
• plugins/dw/skills/ltx-2.5/SKILL.md: same outline for LTX-2.5 (the three-move distilled flow, 32-multiple sizes, 8k+1 frames, the latent handoff, the DFR fps note), with the trained caption spec quoted from LTX2_5_T2V_DEFAULT_SYSTEM_PROMPT and held equal to the diffusers constant by test.
• tests/test_plugin_skills.py pins every number a skill states to the diffusers module that enforces it, resolves every catalog name a skill mentions, and caps each skill at 12 KB.

The cold drill

Recorded in the proposal under "Plugin drill, 2026-09-08": a fresh session with the plugin, and one without, both given "a short multi-shot video with cuts between the shots" against lem. The skill fired on the bare prompt, chose templates/minimax/storyboard, wrote a Ref2VA Context-IR prompt, quoted cost and asked before an 11.6-minute run, and inspected the result. The control found the same template but ran unasked with no cost statement and then failed trying to download the file to the wrong machine. Both wrote the vendor's prompt format, because the storyboard template's default prompt is itself a Context-IR example.

Two things the drill caught, fixed here: the H3 skill said the reference-conditioned templates run 20 steps without a LoRA, but four of them keep the turbo LoRA at nine (now named and pinned); and storyboard.json had no cost (now the control run's 10.1 warm minutes). One engine fix from the same drill: download_output over dw.serve --mcp failed with an unwrapped PermissionError on a client-side path, surfaced as an anonymous tool error; an OSError now becomes a DwApiError that says the write happens on the server and names the client-side ways to see the file.

After the drill: the 30-second follow-up and Music 3

A 30-second request in the same session showed that the H3 skill never said what does not carry between generations. It now states that nothing does except a reference (no latent memory, no extension mode), gives the chain-or-cut rule, and says how a cuts piece gets one score and one voice; the README's cuts section says the same. Seven bundled prompts wrote a silent music field as None. where the guides' literal is N/A; fixed and held by test.

MiniMax Music 3 then went through the onboarding lifecycle: a vendor audit (docs/proposals/audits/2026-09-08-minimax-music3-audit.md) confirmed the ceiling rule and the 44.1 kHz output against three primary sources and found four small errors, all fixed (a stale trim-task path in music.json, music-video's 21-second ceiling over 20.7 seconds of slices, an invented "guillotine" mechanism in the README, and a text_encoder component in the quantization docs that Music 3 does not have). plugins/dw/skills/minimax-music3/SKILL.md covers a song, an instrumental, a score under a cut and the music video, with the caps, window, guider and output rate pinned to the diffusers modular pipeline, and defers captions to MiniMax's music-caption-rewriter skill, which the repo had never mentioned. Its cold drill is still to run.

Scoring that cut from the failed run's files turned up one more engine gap, fixed here: the audio tasks refused a video file's path (File extension not allowed: .mp4) though a finished cut's soundtrack is exactly what a scoring pass mixes under, and pair_audio could not take a path at all. load_audio now takes a video's muxed track, so every audio task accepts it, and pair_audio loads a path; a survey of the rest of the audio and video tasks found the frame-consuming ones already fine. A documentation review against the whole range (seventeen findings, all applied) rides along.

Full suite: 3225 passed.

🤖 Generated with Claude Code