diff --git a/.claude-plugin/marketplace.json b/.claude-plugin/marketplace.json
new file mode 100644
index 00000000..46a9b5f0
--- /dev/null
+++ b/.claude-plugin/marketplace.json
@@ -0,0 +1,15 @@
+{
+    "$schema": "https://anthropic.com/claude-code/marketplace.schema.json",
+    "name": "diffusers-workflow",
+    "description": "Skills for driving a diffusers-workflow server from Claude Code: one composition skill per model family.",
+    "owner": {
+        "name": "Don Kackman"
+    },
+    "plugins": [
+        {
+            "name": "dw",
+            "source": "./plugins/dw",
+            "description": "Compose MiniMax H3 video, MiniMax Music 3 and LTX-2.5 workflows over a dw MCP server: which template fits which shape, the hard rules, cost, and how to judge the output. Prompt format comes from the vendors' own guides."
+        }
+    ]
+}
diff --git a/CLAUDE.md b/CLAUDE.md
index e98cf2d2..78af1f0a 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -45,6 +45,16 @@ See docs/SERVER.md, `dw/server/CLAUDE.md` and `ui/CLAUDE.md`.
 
 The stdio MCP server lives in `dw_mcp/` — see `dw_mcp/CLAUDE.md` and docs/MCP.md.
 
+### Claude Code plugin
+
+`.claude-plugin/marketplace.json` publishes the `dw` plugin in `plugins/dw/`: one
+composition skill per model family (`minimax-h3`, `minimax-music3`, `ltx-2.5`) that
+chooses a template for a request's shape and states the family's hard rules. Model
+knowledge lives there and in the catalog, never in engine code; every number a skill
+states is pinned to a diffusers symbol by `tests/test_plugin_skills.py`. `plugin.json`'s
+version is the engine's, bumped by `scripts/release.sh`. Adding or re-auditing a family
+is `.claude/skills/model-family-onboarding/`.
+
 ### REPL Architecture
 
 The REPL (`dw/repl.py`) uses a **persistent worker subprocess** (`dw/worker.py`) to keep GPU models cached between runs. Communication is via `multiprocessing.Queue`. Worker management is in `dw/repl_worker.py`, command handlers in `dw/repl_commands.py`.
@@ -103,7 +113,7 @@ read an inferred workspace back as one the user named - `get_prompt_dir` yields
 to its older discovery (`./prompts`, then the walk up from the workflow file)
 for an inferred workspace but not for an explicit one. `--workflow-dir`,
 `--output-dir` and `--prompt-dir` each still override one folder. See
-docs/WORKSPACES.md, and docs/proposals/workspaces.md for the later stages
+docs/WORKSPACES.md, and docs/proposals/server-workspaces.md for the later stages
 (workflow search path, run directories, `asset:`/`output:` references).
 
 ### Type System
diff --git a/README.md b/README.md
index 25558ff6..26f9fd99 100644
--- a/README.md
+++ b/README.md
@@ -49,6 +49,15 @@ claude mcp add --transport http dw http://gpu-box:8765/mcp \
   --header "Authorization: Bearer $DW_API_TOKEN"
 ```
 
+The [dw plugin](plugins/dw/README.md) adds one skill per model family - what
+to run for a given shape, the rules that bite, what it costs - and points at
+the vendors' own prompt guides rather than restating them:
+
+```
+/plugin marketplace add dkackman/diffusers-workflow
+/plugin install dw@diffusers-workflow
+```
+
 You don't have to compose that command by hand — the server's own **Server**
 page builds it from the address you pick, alongside the directories it
 resolved and the workspaces it holds:
diff --git a/docs/MCP.md b/docs/MCP.md
index 80413c39..01449cad 100644
--- a/docs/MCP.md
+++ b/docs/MCP.md
@@ -46,6 +46,14 @@ call; that probe is fatal for a remote URL and only a warning for a
 loopback one (where it usually means `dw.serve` is not up yet).
 [REMOTE.md](REMOTE.md) covers the remote setup end to end.
 
+Claude Code users can add the composition skills as well:
+`/plugin marketplace add dkackman/diffusers-workflow` then
+`/plugin install dw@diffusers-workflow`. The plugin ships one skill per model
+family (MiniMax H3, MiniMax Music 3, LTX-2.5) that picks a template for a
+request's shape and states the family's rules - see
+[plugins/dw/README.md](../plugins/dw/README.md). It is optional; every tool
+below works without it.
+
 ### Use the absolute path to `dw-mcp`
 
 **This is the one setup detail that reliably goes wrong.** If you installed
@@ -224,7 +232,7 @@ when no single workflow covers it.
 | --- | --- | --- |
 | `get_output_image(name, max_dimension=768)` | `name`, `max_dimension` | Look at a generated image, downscaled to `max_dimension` on its longest side. Returns the image plus a text part reporting `original_size`, `returned_size` and `bytes`, so a downscale is never silent |
 | `get_output_text(name, max_characters=20000)` | `name`, `max_characters` | Read a text output — a prompt enhancement, or any step whose result is `text/plain` or JSON. Reports the file's real length and whether it was truncated |
-| `download_output(name, destination=None, overwrite=False)` | `name`, `destination`, `overwrite` | Save one output file to local disk, of any content type. `destination` may be a full path, a directory, or omitted to save under the output's own name in the current working directory; `~` expands and missing parent directories are created. `overwrite=True` is required to replace a file already at the resolved path. Returns nothing to the conversation but where the file landed — unlike the other media tools, the point is a file on disk, not a payload in context. Writes on the machine running the MCP server - over `dw.serve --mcp` that is the GPU box |
+| `download_output(name, destination=None, overwrite=False)` | `name`, `destination`, `overwrite` | Save one output file to local disk, of any content type. `destination` may be a full path, a directory, or omitted to save under the output's own name in the current working directory; `~` expands and missing parent directories are created. `overwrite=True` is required to replace a file already at the resolved path. Returns nothing to the conversation but where the file landed — unlike the other media tools, the point is a file on disk, not a payload in context. Writes on the machine running the MCP server - over `dw.serve --mcp` that is the GPU box. A write that fails there (a path that exists only on the client, for instance) comes back as an error naming the server-side write and the client-side alternatives, not as an anonymous tool failure |
 | `delete_output(name)` | `name` | Permanently remove one generated file from the output directory |
 
 ### Authoring, assets and workspaces
diff --git a/docs/QUANTIZATION.md b/docs/QUANTIZATION.md
index 00787bb4..0a50812d 100644
--- a/docs/QUANTIZATION.md
+++ b/docs/QUANTIZATION.md
@@ -192,7 +192,7 @@ through `from_pretrained_arguments`, so quantization is keyed by component name
                 "configuration": { "config_type": "TorchAoConfig" },
                 "arguments": { "quant_type": "torchao.quantization.Int8WeightOnlyConfig" }
             },
-            "text_encoder": {
+            "language_model": {
                 "configuration": { "config_type": "transformers.TorchAoConfig" },
                 "arguments": { "quant_type": "torchao.quantization.Int8WeightOnlyConfig" }
             }
diff --git a/docs/RECIPES_24GB.md b/docs/RECIPES_24GB.md
index 2028a495..cdb4df1f 100644
--- a/docs/RECIPES_24GB.md
+++ b/docs/RECIPES_24GB.md
@@ -102,6 +102,20 @@ workflows below fit; without it, the load is an OOM kill, not a slowdown.
 
 **Examples:** [reference-to-video.json](../workflows/templates/minimax/reference-to-video.json), [chain-matched-to-audio.json](../workflows/templates/minimax/chain-matched-to-audio.json), [image-to-video.json](../workflows/templates/minimax/image-to-video.json), [dialogue-short.json](../workflows/templates/minimax/dialogue-short.json) (five ref2va shots + two Z-Image portraits in ~35 minutes end to end)
 
+## MiniMax-Music3
+
+Music3 runs at about 22GiB in bfloat16 under the templates' `components_manager`
+auto CPU offload, which keeps only the running component resident; no quantization
+is needed on a 24GB card. The language model is the part worth offloading harder:
+a leaf-level `group_offload` of `language_model` brings it to about 8GiB (the model
+card's low-VRAM recipe). Two things the examples carry: `release_pipeline` on the
+music step in any workflow that loads H3 afterwards, since host RAM is the binding
+constraint (see Multi-model workflows below), and the run's time follows the length
+the model actually sings, not `audio_duration`, which is a ceiling of at most 9000
+frames at 25 frames per second (360 seconds). Output is 44.1 kHz stereo.
+
+**Examples:** [music.json](../workflows/templates/minimax/music.json), [music-video.json](../workflows/templates/minimax/music-video.json)
+
 ## LTX-2.5 (22B, video + audio)
 
 A standard pipeline, but placed per component rather than with a pipeline-level
diff --git a/docs/RELEASING.md b/docs/RELEASING.md
index d0075cca..71ef4b83 100644
--- a/docs/RELEASING.md
+++ b/docs/RELEASING.md
@@ -8,20 +8,23 @@ scripts/release.sh 0.38.0-alpha.1 "UI front end"   # optional tag message
 ```
 
 The script bumps `pyproject.toml` (the single source of the version —
-`dw.__version__` reads it at runtime), commits just that file, pushes
+`dw.__version__` reads it at runtime) and sets the same version in
+`plugins/dw/.claude-plugin/plugin.json`, so an installed plugin names the
+engine it was written against; it commits just those two files, pushes
 master, tags the bump commit `v0.38.0`, and pushes the tag. It refuses
 a malformed version, a branch other than master, an existing tag, or a
 dirty index (unstaged changes elsewhere are fine — the release commit
-is path-limited to pyproject.toml).
+is path-limited to those two files).
 
 By hand, the equivalent is:
 
 ```bash
 # 1. Bump the version in pyproject.toml:
 #    version = "0.38.0"
-git commit -m "release 0.38.0" -- pyproject.toml
+# 2. Set the same version in plugins/dw/.claude-plugin/plugin.json
+git commit -m "release 0.38.0" -- pyproject.toml plugins/dw/.claude-plugin/plugin.json
 
-# 2. Tag the bump commit and push
+# 3. Tag the bump commit and push
 git tag -a v0.38.0 -m "release 0.38.0"
 git push origin master v0.38.0
 ```
diff --git a/docs/TASKS.md b/docs/TASKS.md
index 70ccd507..49930833 100644
--- a/docs/TASKS.md
+++ b/docs/TASKS.md
@@ -389,7 +389,7 @@ returns frames without it, and this puts it back:
 | Argument | Required | Description |
 | -------- | -------- | ----------- |
 | `video` | Yes | The frames - a frame list, a frame array or tensor, or an audio+video pair whose own soundtrack is replaced |
-| `audio` | Yes | The soundtrack - a waveform, or the earlier step whose video carried one, which brings its sample rate along |
+| `audio` | Yes | The soundtrack - a waveform, the earlier step whose video carried one, or the path or URL of an audio or video file; the last two bring their sample rate along |
 | `sample_rate` | No | Sample rate of the waveform. Required unless `audio` carries one; given here it wins |
 
 **Example:** [assemble-and-score.json](../workflows/templates/assemble-and-score.json)
@@ -419,7 +419,7 @@ a length still passes the whole track along:
 
 | Argument | Required | Description |
 | -------- | -------- | ----------- |
-| `audio` | Yes | Path or URL of an audio file, a waveform from a previous step, or an earlier step's video generated with a soundtrack (which brings its sample rate along) |
+| `audio` | Yes | Path or URL of an audio file (or of a video file, whose soundtrack is taken), a waveform from a previous step, or an earlier step's video generated with a soundtrack (which brings its sample rate along) |
 | `start_seconds` / `duration_seconds` | One pair | The slice in seconds; either may be omitted |
 | `start_frame` / `num_frames` / `fps` | One pair | The slice in video frames; `fps` is required, start and count may be omitted |
 | `sample_rate` | With a waveform | Sample rate of a directly passed waveform (files carry their own) |
@@ -465,7 +465,7 @@ ending. The curve is the equal-power cosine the seam joins use:
 
 | Argument | Required | Description |
 | -------- | -------- | ----------- |
-| `audio` | Yes | Path or URL of an audio file, a waveform from a previous step, or an earlier step's video generated with a soundtrack (which brings its sample rate along) |
+| `audio` | Yes | Path or URL of an audio file (or of a video file, whose soundtrack is taken), a waveform from a previous step, or an earlier step's video generated with a soundtrack (which brings its sample rate along) |
 | `fade_in_ms` | No | Length of the fade in, from the head of the track (default: 0) |
 | `fade_out_ms` | No | Length of the fade out, to the tail of the track (default: 0) |
 | `sample_rate` | With a waveform | Sample rate of a directly passed waveform (files carry their own) |
@@ -494,7 +494,7 @@ changes, so the dynamics survive:
 
 | Argument | Required | Description |
 | -------- | -------- | ----------- |
-| `audio` | Yes | Path or URL of an audio file, a waveform from a previous step, or an earlier step's video generated with a soundtrack (which brings its sample rate along) |
+| `audio` | Yes | Path or URL of an audio file (or of a video file, whose soundtrack is taken), a waveform from a previous step, or an earlier step's video generated with a soundtrack (which brings its sample rate along) |
 | `peak_dbfs` | No | The level the loudest sample is moved to, in dB below full scale (default: -1.0). 0 is full scale |
 | `sample_rate` | With a waveform | Sample rate of a directly passed waveform (files carry their own) |
 
@@ -524,7 +524,7 @@ at every cut:
 
 | Argument | Required | Description |
 | -------- | -------- | ----------- |
-| `audios` | Yes | The tracks to layer - waveforms, audio file paths, or videos generated with a soundtrack |
+| `audios` | Yes | The tracks to layer - waveforms, audio or video file paths, or videos generated with a soundtrack |
 | `gains` | No | One plain multiplier per track, in the same order - not decibels. Defaults to unity on every track |
 | `sample_rate` | With a raw waveform | Sample rate of the waveforms. Required unless every track brings its own; given here it wins |
 
@@ -556,7 +556,7 @@ supplied recording once, up front, feeds it what it already wants:
 
 | Argument | Required | Description |
 | -------- | -------- | ----------- |
-| `audio` | Yes | Path or URL of an audio file, a video generated with a soundtrack (which brings its sample rate along), or a waveform |
+| `audio` | Yes | Path or URL of an audio or video file, a video generated with a soundtrack (which brings its sample rate along), or a waveform |
 | `target_sample_rate` | Yes | The rate to convert to |
 | `sample_rate` | With a waveform | Sample rate of a waveform passed directly; given for a file or a video it overrides the rate they carry |
 
diff --git a/docs/WORKFLOW_GUIDE.md b/docs/WORKFLOW_GUIDE.md
index 245ae3d4..43aa7176 100644
--- a/docs/WORKFLOW_GUIDE.md
+++ b/docs/WORKFLOW_GUIDE.md
@@ -557,7 +557,7 @@ once, not twice.
 - **Ignored when the component's device is the CPU**, where there is nothing to move it
   off of.
 
-On a 24GB card, `MiniMaxH3Ref2VA.json` peaks at 18.9GiB of reserved VRAM with on-demand
+On a 24GB card, `templates/minimax/reference-to-video.json` peaks at 18.9GiB of reserved VRAM with on-demand
 VAEs against 23.2GiB resident, and the tighter resident fit costs 40 allocator retries -
 cache flushes forced by a failed allocation - where the on-demand run has none. The
 headroom is also what lets the chained variant run: its later segments carry an extra
@@ -815,7 +815,7 @@ and `load_components` pulls the weights:
                   "modules_to_not_convert": ["proj_in", "proj_out"]
               }
           },
-          "text_encoder": {
+          "language_model": {
               "configuration": { "config_type": "transformers.TorchAoConfig" },
               "arguments": { "quant_type": "torchao.quantization.Int8WeightOnlyConfig" }
           }
@@ -1172,7 +1172,9 @@ directory being copied. `latest` only selects a run where run directories are; a
 workflow or file that happens to be called `latest` is still named as itself.
 
 Like `asset:`, a reference resolves to a path and then whatever loads paths loads it, so
-it works under `image`, `video`, a `from_file`, or a list of them. It resolves against
+it works under `image`, `video`, a `from_file`, or a list of them. The audio tasks take
+a video file's path too and use the soundtrack muxed into it, which is how a finished
+cut is scored in a later run without re-cutting it. It resolves against
 the output directory the run was told to write to, and cannot leave it: `..`, an
 absolute path, and a symlink pointing out are all refused.
 
diff --git a/docs/WORKSPACES.md b/docs/WORKSPACES.md
index 5eaa8bc3..079092a9 100644
--- a/docs/WORKSPACES.md
+++ b/docs/WORKSPACES.md
@@ -263,7 +263,7 @@ it. Use them to keep work apart, not to keep it private.
 ## Where this is going
 
 Workspaces were the first stage of the design in
-[proposals/workspaces.md](proposals/workspaces.md). The resolver, the workflow
+[proposals/server-workspaces.md](proposals/server-workspaces.md). The resolver, the workflow
 search path with writes confined to the writable root, run directories with an
 on-disk manifest, `asset:` and `output:` references, and server-side named
 workspaces are all implemented. What remains from the proposal is an MCP
diff --git a/docs/img/workflow-h3-with-mixed-audio.png b/docs/img/workflow-h3-with-mixed-audio.png
new file mode 100644
index 00000000..705a46d2
Binary files /dev/null and b/docs/img/workflow-h3-with-mixed-audio.png differ
diff --git a/docs/proposals/agent-catalog-legibility.md b/docs/proposals/agent-catalog-legibility.md
index 17014b09..92fa8ae4 100644
--- a/docs/proposals/agent-catalog-legibility.md
+++ b/docs/proposals/agent-catalog-legibility.md
@@ -3,7 +3,7 @@
 Status: Parts 1–2 designed 2026-09-06 — see the
 [design spec](../superpowers/specs/2026-09-06-agent-catalog-legibility-design.md)
 and the **Ledger** at the end of this document, which records what was
-actually done against each proposal. Parts 3–4 remain open. Synthesizes and replaces two prior proposals,
+actually done against each proposal. Part 3 is carried as a constraint and Part 4 shipped through both channels (see the Ledger and the plugin drill at the end); no part remains open. Synthesizes and replaces two prior proposals,
 `catalog-shape-index.md` and `mcp-discovery-data.md` (both fully folded into
 this document and deleted — see git history for the original framing), plus
 the discovery-facing conclusion of
@@ -372,7 +372,7 @@ the moment an agent is actually driving H3. Loading it by default, or even
 indexing it by default, works against the economy Part 1 fought to
 establish.
 
-## Part 4: packaging model-specific knowledge — an open design question
+## Part 4: packaging model-specific knowledge — resolved 2026-09-07 in favour of Channel B
 
 Given Part 3's constraint (data, not code) and its budget concern
 (don't bloat the standing MCP index), there appear to be two non-exclusive
@@ -411,6 +411,18 @@ authoring knowledge independently. Whether that's practical, and whether
 it should ship as a plugin bundled with `dw_mcp` or live separately, is
 exactly the kind of specific-mechanism question this proposal defers.
 
+**Resolved (2026-09-07, shipped 2026-09-08).** Both channels are in use,
+asymmetrically: dw-generic authoring knowledge went to the guides (Channel
+A, the `WORKFLOW_GUIDE.md` authoring section), and model-family knowledge
+ships as the dw plugin's per-family skills (Channel B: `plugins/dw/skills/`
+for MiniMax H3, MiniMax Music 3 and LTX-2.5), while `dw/server/guides.py`
+deliberately indexes no model family. The two-copies risk named above is
+answered by keeping the skills thin - they defer prompt format to the
+vendors' own text and quote catalog names rather than catalog content - and
+by `tests/test_plugin_skills.py`, which pins every number a skill states to
+the diffusers module that enforces it. The paragraphs above stay as the
+design record; the "Plugin drill" section below is the acceptance test.
+
 ## Principle: format-knowledge belongs in guides and templates, not in new engine code
 
 The catalog and guides answer "which existing workflow fits", and Part 2's
@@ -505,7 +517,7 @@ starts from the record rather than the intent. Updated as work lands.
 | 9 composition rules | done (plan 2, task 6) | spec §2.3 | the cartesian rule and the one-step-per-pair form are stated in the authoring section, the general case of `scripted-dialogue-and-tts.md`'s reasoning |
 | 10 save-time metadata | done (task 6) | spec §1.6 | derived and returned on save; empty summary warns, never rejects |
 | Part 3 constraint | carried | spec "Principle" | derivation reads structure, never model family |
-| Part 4 packaging | first item through Channel A (2026-09-07) | `WORKFLOW_GUIDE.md` authoring section | the spoon-set drill (cold session, PR #48 merged) planned a shared seed for "identical but for colour", which draws a different object per prompt; `### Keeping a set consistent` now says which of style, object or character consistency wants prompts, an edit pass or a reference, and `templates/consistent-set.json` (declared `image-set`: derivation says `image-edit`) is the generate-then-edit shape on FLUX.1 Kontext, whose one pipeline both draws and edits; measured on lem 2026-09-07 at 11.5 warm minutes (2.2 for the base, 3.1 per edit, sequential offload) and the four mugs came out identical but for colour. A FLUX.2 draft found that the HF remote text encoder every FLUX.2 entry names now answers with an HTML page (broken on lem since 2026-08-26); `remote.py` now says so instead of raising an unpickling error, and both FLUX.2 entries load the repo's own 4-bit text encoder under model offload instead. The FLUX.2 text encoder is declared as its own component pinned to the CPU, because BitsAndBytes materializes on the accelerator and the two 4-bit models do not fit at load; the `flash_hub` backend went too, since it needs the `kernels` package no install has. models/flux2-dev measured at 3.3 warm minutes on lem. The Krea edit template moved to FLUX.1 Kontext for the same reason, and a catalog test now refuses `trust_remote_code` and `custom_pipeline` in any bundled entry or packaged builtin, so nothing dw ships asks an operator to lower `--trust-workflows`. The Florence-2 and Phi-3.5 builtins went with it: the `text_generation` task had replaced both, and `describe-and-regenerate` now composes it twice. `### Remote code is refused by default` came out of the same run: the first draft used the Krea edit template's `custom_pipeline`, which a server without `--trust-workflows` refuses at load. Still open: `templates/minimax/README.md` and `ltx2/README.md` are model knowledge as data, unindexed — a third channel. Catalog repair task 1: the latent handoff a two-stage flow needs is proven by test (tests/test_result.py::TestLatentHandoff), no engine change. Catalog repair task 2: two-stage.json is the three-move flow (8 sigmas at 768x448, 2x latent upsample, renoise + 3 stage-two sigmas at 1536x896, audio latents carried); latents pass by name; a test holds noise_scale to STAGE_2_DISTILLED_SIGMA_VALUES[0]. Cost and the sharpness comparison await the lem run (task 7). Catalog repair task 3: the six prompts/ltx2 captions are rewritten to the trained format (one paragraph, 150-220 words, shot type/camera motion/viewpoint in prose, sound interleaved; the two I2V ones describe only what changes), intended_model ltx-2.5, four summaries say LTX-2.5; tests/test_ltx_prompt_library.py holds the shape. Catalog repair task 4: h3_context_ir names its two source guides, writes N/A for silent audio fields, numbers <Video N>/<Audio N> within their category, labels continuity modes as this engine's chaining convention (nothing in the engine emits the phrase; it is a user-message convention), adds the ref guide's dialogue-fidelity rules, drops the two unsourced lines; tests/test_h3_context_ir.py. Catalog repair task 5: both template READMEs link the files that exist, name the vendor sources and the audit, the H3 one states the canvas rules and the 5-second diffusers floor; a test resolves every README link. Lem found templates/minimax/enhance-prompt failing on master already: both enhance templates passed a stored copy of the Context-IR system prompt (prompts/prompt_enhancement/minimax_h3.json, 10096 chars) as a sub-workflow argument, which exceeded MAX_VARIABLE_VALUE_LENGTH 10000; the limit is now 20000, the stored copy is gone, and the templates run the builtin's own prompt, with a test refusing a future override. Three wordings of the N/A rule were drafted before the final review found the templates were running the stored copy, so none of them had been tested; the third, which puts the rule on each audio field's own line, is what shipped.  Catalog repair task 7, verified on lem 2026-09-08: two-stage runs at 1536x896, 8.2 warm minutes (base 1.7, refine 2.7, writing the full-size clip 3.5), latents handed by name and the refine pass served from the pipeline cache; at seed 42 the refined frame is sharp where the upsample-only frame is a soft blur. The fox and hummingbird captions ran as written (tracking shot and pull-out; hold and close-up), the hummingbird soundtrack near-silent. The silent-candle brief's three runs went through the stale stored prompt (final review), so the N/A rule's effect on the Qwen3-4B enhancer is re-verified below. Re-verified 2026-09-08 with the templates on the builtin's prompt: the framed idea the template documents (Task, Duration, Idea) gives a full integrated_multimodal_description with both audio fields N/A; a bare idea first collapsed to the two N/A lines alone, so the description field is now declared unconditional and a bare idea also gets its description (jobs 28f1d29a5b90, be94ed994ecd). The N/A rule took four wordings in all, and only the last two were ever tested. Two engine findings on the way: the step cache keys on the resolved step definition, which names a builtin by path, so editing a builtin does not invalidate a cached step (follow-up); and a cancelled H3 job runs to its next step boundary, minutes on this model. |
+| Part 4 packaging | done: Channel A for the generic item (2026-09-07), Channel B for model families (2026-09-08) | `WORKFLOW_GUIDE.md` authoring section; `plugins/dw/`, `.claude-plugin/marketplace.json`, `.claude/skills/model-family-onboarding` | the spoon-set drill (cold session, PR #48 merged) planned a shared seed for "identical but for colour", which draws a different object per prompt; `### Keeping a set consistent` now says which of style, object or character consistency wants prompts, an edit pass or a reference, and `templates/consistent-set.json` (declared `image-set`: derivation says `image-edit`) is the generate-then-edit shape on FLUX.1 Kontext, whose one pipeline both draws and edits; measured on lem 2026-09-07 at 11.5 warm minutes (2.2 for the base, 3.1 per edit, sequential offload) and the four mugs came out identical but for colour. A FLUX.2 draft found that the HF remote text encoder every FLUX.2 entry names now answers with an HTML page (broken on lem since 2026-08-26); `remote.py` now says so instead of raising an unpickling error, and both FLUX.2 entries load the repo's own 4-bit text encoder under model offload instead. The FLUX.2 text encoder is declared as its own component pinned to the CPU, because BitsAndBytes materializes on the accelerator and the two 4-bit models do not fit at load; the `flash_hub` backend went too, since it needs the `kernels` package no install has. models/flux2-dev measured at 3.3 warm minutes on lem. The Krea edit template moved to FLUX.1 Kontext for the same reason, and a catalog test now refuses `trust_remote_code` and `custom_pipeline` in any bundled entry or packaged builtin, so nothing dw ships asks an operator to lower `--trust-workflows`. The Florence-2 and Phi-3.5 builtins went with it: the `text_generation` task had replaced both, and `describe-and-regenerate` now composes it twice. `### Remote code is refused by default` came out of the same run: the first draft used the Krea edit template's `custom_pipeline`, which a server without `--trust-workflows` refuses at load. Superseded by plugin tasks 1-5 below: `templates/minimax/README.md` and `ltx2/README.md` are the sources the family skills derive from and link back to, so that knowledge is reached through the plugin rather than a third channel. Catalog repair task 1: the latent handoff a two-stage flow needs is proven by test (tests/test_result.py::TestLatentHandoff), no engine change. Catalog repair task 2: two-stage.json is the three-move flow (8 sigmas at 768x448, 2x latent upsample, renoise + 3 stage-two sigmas at 1536x896, audio latents carried); latents pass by name; a test holds noise_scale to STAGE_2_DISTILLED_SIGMA_VALUES[0]. Cost and the sharpness comparison await the lem run (task 7). Catalog repair task 3: the six prompts/ltx2 captions are rewritten to the trained format (one paragraph, 150-220 words, shot type/camera motion/viewpoint in prose, sound interleaved; the two I2V ones describe only what changes), intended_model ltx-2.5, four summaries say LTX-2.5; tests/test_ltx_prompt_library.py holds the shape. Catalog repair task 4: h3_context_ir names its two source guides, writes N/A for silent audio fields, numbers <Video N>/<Audio N> within their category, labels continuity modes as this engine's chaining convention (nothing in the engine emits the phrase; it is a user-message convention), adds the ref guide's dialogue-fidelity rules, drops the two unsourced lines; tests/test_h3_context_ir.py. Catalog repair task 5: both template READMEs link the files that exist, name the vendor sources and the audit, the H3 one states the canvas rules and the 5-second diffusers floor; a test resolves every README link. Lem found templates/minimax/enhance-prompt failing on master already: both enhance templates passed a stored copy of the Context-IR system prompt (prompts/prompt_enhancement/minimax_h3.json, 10096 chars) as a sub-workflow argument, which exceeded MAX_VARIABLE_VALUE_LENGTH 10000; the limit is now 20000, the stored copy is gone, and the templates run the builtin's own prompt, with a test refusing a future override. Three wordings of the N/A rule were drafted before the final review found the templates were running the stored copy, so none of them had been tested; the third, which puts the rule on each audio field's own line, is what shipped.  Catalog repair task 7, verified on lem 2026-09-08: two-stage runs at 1536x896, 8.2 warm minutes (base 1.7, refine 2.7, writing the full-size clip 3.5), latents handed by name and the refine pass served from the pipeline cache; at seed 42 the refined frame is sharp where the upsample-only frame is a soft blur. The fox and hummingbird captions ran as written (tracking shot and pull-out; hold and close-up), the hummingbird soundtrack near-silent. The silent-candle brief's three runs went through the stale stored prompt (final review), so the N/A rule's effect on the Qwen3-4B enhancer is re-verified below. Re-verified 2026-09-08 with the templates on the builtin's prompt: the framed idea the template documents (Task, Duration, Idea) gives a full integrated_multimodal_description with both audio fields N/A; a bare idea first collapsed to the two N/A lines alone, so the description field is now declared unconditional and a bare idea also gets its description (jobs 28f1d29a5b90, be94ed994ecd). The N/A rule took four wordings in all, and only the last two were ever tested. Two engine findings on the way: the step cache keys on the resolved step definition, which names a builtin by path, so editing a builtin does not invalidate a cached step (follow-up); and a cancelled H3 job runs to its next step boundary, minutes on this model. Plugin task 1: a marketplace at the repo root and the dw plugin under plugins/dw, installable with two commands the README shows; plugin.json's version is the engine's, bumped by release.sh and held by test. Plugin task 2: the minimax-h3 skill - shape decision over the family's templates, the frame and canvas rules pinned by test to the diffusers modular pipeline, prompts deferred to MiniMax's h3-prompt-writing skill and the two guides, the run-and-judge loop; the README points at it. Plugin task 3: the ltx-2.5 skill - shape decision, the schedule and size rules pinned to the LTX-2 pipeline, the trained caption spec quoted and held equal to the diffusers constant by test; the README points at it. Final review of the plugin branch: the H3 skill's LoRA coupling scoped to the text- and frame-conditioned templates (the six reference ones run no LoRA at 20 steps), both judge steps rewritten because no MCP tool shows a video frame (the agent hands over the gallery url and checks metadata), the fps rule marked as the DFR path's, tighter numeric pins. Plugin task 4: cold drill 2026-09-08 - pass; 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, where the control ran unasked and could not download the file; the drill caught the skill overstating 20 steps for four LoRA-bearing reference templates (fixed, pinned) and gave storyboard its measured cost (10.1 warm minutes); the mould for the next family is copy a skill, follow its six sections, add the family's rules to tests/test_plugin_skills.py, cite the vendor, and run this drill (.claude/skills/model-family-onboarding). Plugin task 5 (2026-09-08, the 30-second follow-up to the drill): the H3 skill states that nothing carries between generations except a reference, gives the chain-or-cut rule, and says how a cuts piece gets one score and one voice; seven bundled prompts wrote a silent music field as None. and now write N/A, held by test. MiniMax Music 3 went through the lifecycle the same night: audited (docs/proposals/audits/2026-09-08-minimax-music3-audit.md; the ceiling rule and the 44.1 kHz result were confirmed against three primary sources, four small corrections landed: music.json's stale trim-task path, music-video's 21-second ceiling over 20.7 seconds of slices, the README's invented guillotine mechanism, and a text_encoder component the quantization docs showed that Music3 does not have), and plugins/dw/skills/minimax-music3/SKILL.md written with its caps, window, guider and output rate pinned to the diffusers modular pipeline and the caption format deferred to MiniMax's music-caption-rewriter skill, which the repo had never mentioned. Its cold drill is still to run. |
 
 
 ### Cold-session probe, 2026-09-07
@@ -545,6 +557,88 @@ What that says about Part 1 as built, and what Part 2 has to carry:
   Part 4's packaging question includes how a playbook and the catalog share
   the knowledge rather than compete for it.
 
+### Plugin drill, 2026-09-08
+
+The Part 4 acceptance test: a fresh Claude Code session in an empty directory
+(`~/testing/7`, no `CLAUDE.md`, no memory, only the dw MCP server for lem on
+branch `dw-plugin`), the dw plugin installed from this checkout, asked "a short
+multi-shot video with cuts between the shots"; then the same prompt in the same
+directory with the plugin uninstalled, as the control. Transcripts:
+`~/.claude/projects/-Users-don-testing-7/24d8bc9d-7c01-4320-aa5d-a8acace418ef.jsonl`
+(plugin) and `...-testing-7/bb52715b-c3ea-4227-97fd-88dee384af42.jsonl`
+(control). Lem's `dw.log` is the engine log, not the access log, so the call
+order below is the transcripts'.
+
+**Plugin session.** The `dw:minimax-h3` skill fired on the bare prompt, before
+any tool call. It asked which cut style (storyboard, dialogue short, music
+video), then the subject ("a lighthouse keeper's last night before the light is
+automated"). Then, in order: `get_server_info`, `list_workflows(shape=sequence)`,
+`list_workflows(shape=shot)`, `get_workflow(templates/minimax/storyboard)`,
+`validate_workflow`, a plan with 192 frames (17x11+5), 960x544, cuts at 2.7 and
+5.4 seconds, and the skill's fallback cost wording because the template carried
+no `cost` ("not measured for this template; a few minutes for a turbo-length
+clip, reference-conditioned roughly doubles that"). It asked before running.
+After `run_workflow` and eleven `wait_for_job` polls it read `get_job_events`
+to see which step it was on, then `list_gallery` and `get_gallery_metadata` on
+the mp4, handed over the gallery url with the manifest path, and named the
+family's failure modes to look for (face drift across cuts, storyboard anchors
+overriding the framing). Job 9f75733cfb3c, 11.6 minutes with both models
+loading from disk.
+
+**Control session.** No shape question; it asked for the subject with an open
+question. Then `get_server_info`, the same two shape listings, `list_guides`
+(never followed by a `get_guide`), the same `get_workflow`, `validate_workflow`,
+`get_memory`, and `run_workflow` without asking, with "similar template ran a
+few minutes" as its whole cost statement. It passed the template's own
+`image_reference_type` default back as an argument, harmless but a variable it
+did not understand. When the job finished it called `download_output` with a
+path on this machine, which lem cannot write (`PermissionError` on `/private`,
+surfaced as an opaque `Error executing tool`); it retried twice, once with no
+destination, which saved the mp4 into lem's working directory; then it fetched
+the three stills inline with `get_output_image` and reported. Job
+513205c86652, 10.1 warm minutes.
+
+**The prompts.** Both sessions wrote a correct six-section Ref2VA Context-IR
+prompt (subject definitions, tagged summary, per-reference retention analysis,
+timed shots, soundscape, continuous score) and they are near-identical in
+structure. Neither took it from the skill's vendor pointer: the storyboard
+template's default `prompt` is itself a full Context-IR example, and both
+patterned on it. On a template that carries an example, the vendor format
+reaches a cold agent through the catalog with or without the plugin.
+
+**The 30-second follow-up, same session.** Asked for a 30-second continuation,
+the plugin session did not stretch one generation past 14.4 seconds: it planned
+six 124-frame shots from one keeper portrait, concatenated with hard cuts, and
+quoted 30-45 minutes before running. It had to be told that H3 shares context
+only inside a generation, so each shot writes its music field as N/A and one
+score goes under the concat; that sentence is now in the skill. Its score step
+then failed because it passed empty lyrics for an instrumental (Music 3 rejects
+them; a tag-only body is the form), and its retry, an inline workflow with one
+argument changed, regenerated all six cached shots. The cut was scored from the
+failed run's files by a saved seven-step workflow (`score-lighthouse` on lem):
+a cello instrumental at a 50-second ceiling, a one-video `concat_videos` to
+read the edit with its soundtrack, slice, mix, normalize, pair. Its own first
+attempt hit the audio tasks refusing a video path. Second run 7 seconds with
+the score served from the cache; the film is
+`outputs/score-lighthouse/20260908-061449-*/ScoreLighthouseCut-film.6-0.0.mp4`.
+Both engine findings are in the follow-ups below.
+
+**Verdict: pass, with the skill's contribution measured honestly.** The skill
+fired unprompted, the right template was chosen, prompts were in the vendor's
+format, the cost was stated before the run, and the output was inspected. What
+the plugin changed against the control was the shape choice put to the user,
+asking before spending eleven GPU minutes, the honest cost wording, and the
+family-specific judging brief at the end. What it did not change was
+discovery (both sessions found the template in the same three calls) or the
+prompt format on this template. Two defects it surfaced: the skill said the
+reference-conditioned templates run 20 steps without a LoRA, but four of
+them (`storyboard`, `dialogue-short`, `music-video`,
+`chain-matched-and-aligned`) keep the turbo LoRA at nine, and the plugin
+session repeated the wrong number; fixed, with the four now named and pinned
+by test. And `storyboard` had no `cost`, so both sessions guessed; it now
+carries the control run's 10.1 warm minutes. The `download_output` failure is
+fixed on the branch; see the engine note under the follow-ups.
+
 ### Model-knowledge follow-ups, 2026-09-07
 
 What the two vendor audits found that the catalog does not carry, dated by the
@@ -581,4 +675,48 @@ MiniMax H3 ([audit](audits/2026-09-07-minimax-h3-audit.md)):
 - Step-count note: 20 default, ~25 for motion (ComfyUI).
 - Ref2VA input limits (≤9 images, ≤3 videos, ≤3 audio, ≤12) and that audio can never
   be the only reference; H3-Regenerate-2K is API-only.
-- Music3 `audio_duration` cap: 9000 frames, six minutes.
+- Music3 `audio_duration` cap: 9000 frames, six minutes (done 2026-09-08, in the `minimax-music3` skill and the README).
+
+MiniMax Music 3 ([audit](audits/2026-09-08-minimax-music3-audit.md), 2026-09-08),
+deliberately not done:
+
+- No stored prompt or template exercises the three-heading Structured Caption
+  (Global Metadata, Vocal Details, Arrangement); the library carries only the
+  concise one-paragraph form. `_clean_caption` strips markdown, so one can be
+  pasted as-is.
+- No instrumental example: a tag-only lyrics body plus a caption that names the
+  lead instrument, since diffusers has no `is_instrumental` flag and rejects
+  empty lyrics.
+- The 8 GB path (leaf-level group offload of `language_model`) is undocumented
+  in RECIPES_24GB.
+- The vendor's five-minute range against the engine's 360-second cap; the skill
+  says stay at or under 300.
+- The model card says CUDA only; the diffusers snippet lists mps and cpu. Untested
+  here on either.
+
+Engine, from the 30-second follow-up on 2026-09-08:
+
+- Fixed on the branch: the audio tasks (`resample_audio`, `slice_audio`,
+  `mix_audio` and the rest) refused a path to a video file ("File extension
+  not allowed: .mp4"), though a video's soundtrack is exactly what scoring an
+  existing cut needs, and `pair_audio` could not take a path for its track at
+  all. `load_audio` now takes a video file's soundtrack (a silent one is an
+  error), so every audio task accepts the cut an earlier run wrote, and
+  `pair_audio` reads a path. A survey of the rest found the frame-consuming
+  tasks already fine, since the engine loads any argument named `video` from a
+  path, and `concat_videos`, `dissolve_videos` and `stabilize_video` load
+  paths with their audio themselves.
+- Not fixed: an inline workflow's step-cache key covers the whole definition, so fixing
+  one argument on a retry regenerated six cached H3 shots (thirty minutes);
+  a saved workflow with a stable id would have served them. The authoring
+  guide should say to save a workflow before a long run.
+
+Engine, from the 2026-09-08 plugin drill, fixed on the same branch:
+
+- `download_output` over `dw.serve --mcp` writes on the GPU box, and a path from
+  the agent's own machine failed there with an unwrapped `PermissionError` that
+  the MCP layer reported as `Error executing tool download_output`; the agent
+  then retried with no destination and left a copy in the server's working
+  directory. `dw_mcp/media.py` now answers an `OSError` with a `DwApiError`
+  that says the write happens on the server and names the client-side ways to
+  see the file (the gallery url, the inline tools, `keep_output`).
diff --git a/docs/proposals/audits/2026-09-08-minimax-music3-audit.md b/docs/proposals/audits/2026-09-08-minimax-music3-audit.md
new file mode 100644
index 00000000..2c959f9b
--- /dev/null
+++ b/docs/proposals/audits/2026-09-08-minimax-music3-audit.md
@@ -0,0 +1,164 @@
+# MiniMax Music 3 — knowledge audit of `diffusers-workflow`
+
+Research date: 2026-09-08. Repository: `/Users/don/src/dkackman/diffusers-workflow` (branch `dw-plugin`).
+Installed engine: diffusers `0.41.0.dev0` at
+`/Users/don/src/dkackman/diffusers-workflow/venv/lib/python3.14/site-packages/diffusers`.
+
+---
+
+## Sources
+
+| URL | Type | Date | Contributes |
+| --- | --- | --- | --- |
+| `venv/.../diffusers/modular_pipelines/minimax_music3/modular_blocks_minimax_music3.py` | PRIMARY (code, as installed) | diffusers 0.41.0.dev0, read 2026-09-08 | `audio_duration` default/ceiling docstring, `num_inference_steps` default 30, 200-frame window, lyrics tag contract, 44.1 kHz claim, `audios` output shape |
+| `venv/.../minimax_music3/encoders.py` | PRIMARY (code) | read 2026-09-08 | Special-token prompt template, `_MAX_PROMPT_TOKENS = 5_000`, `_MAX_AUDIO_FRAMES = 9_000`, `_AR_CFG_SCALE = 1.5`, `_AR_CFG_TOP_K = 50`, `_AR_SAMPLING_TOP_K = 50`, `_normalize_lyrics`, `_clean_caption` markdown stripping |
+| `venv/.../minimax_music3/before_denoise.py` | PRIMARY (code) | read 2026-09-08 | `_CHUNK_FRAMES = 200`, `_CHUNK_HOP = 100` |
+| `venv/.../minimax_music3/denoise.py` | PRIMARY (code) | read 2026-09-08 | `ClassifierFreeGuidance` default `{"guidance_scale": 1.7}`, `num_inference_steps` default 30, `_OVERLAP_LATENT_LENGTH = 172` |
+| `venv/.../minimax_music3/decoders.py` | PRIMARY (code) | read 2026-09-08 | `_CROP_LEFT_LATENT = 86`, `_CROP_RIGHT_LATENT = 344 - 86`, stereo stitch |
+| `venv/.../minimax_music3/modular_pipeline.py` | PRIMARY (code) | read 2026-09-08 | `sampling_rate` = 44100, `frame_rate` = 25.0 Hz, `latent_hop_length` = 512, `num_channels_latents` = 128 |
+| `venv/.../models/autoencoders/minimax_music3_vocoder.py` | PRIMARY (code) | read 2026-09-08 | `sampling_rate: int = 44100` (line 85), stereo `(batch, 2, samples)` |
+| `venv/.../models/condition_embedders/condition_embedder_minimax_music3.py` | PRIMARY (code) | read 2026-09-08 | `input_sampling_rate=24000`, `input_hop_length=960` (→ 25 fps), `output_sampling_rate=44100`, `output_hop_length=512` |
+| `venv/.../models/transformers/transformer_minimax_music3.py` | PRIMARY (code) | read 2026-09-08 | `in_channels: int = 128` |
+| https://huggingface.co/MiniMaxAI/MiniMax-Music3/raw/main/README.md | PRIMARY (vendor model card, fetched verbatim by curl) | fetched 2026-09-08; card content dates from the 2026-08 launch | Five-minute max, tag vocabulary, Structured Caption contract, `max_new_tokens` @ 25 fps ceiling semantics, 5,000-token / 9,000-frame limits, VRAM tiers, 32 kHz WAV for the SGLang server, diffusers snippet with `audio_duration=60.0` |
+| https://huggingface.co/MiniMaxAI/MiniMax-Music3/raw/main/vocoder/config.json | PRIMARY (checkpoint config) | fetched 2026-09-08 | `"sampling_rate": 44100`, `latent_channels: 128` |
+| https://huggingface.co/MiniMaxAI/MiniMax-Music3/tree/main | PRIMARY (repo tree) | fetched 2026-09-08 | **No `docs/` or `skills/` directory on the HF repo**; only `assets/`, `figures/`, `scripts/`, and component folders |
+| https://api.github.com/repos/MiniMax-AI/MiniMax-Music3/git/trees/main?recursive=1 | PRIMARY (repo tree) | fetched 2026-09-08 | 1,037 paths; `skills/music-caption-rewriter/` with `SKILL.md`, `references/genre-router.md` + 18 family indexes, **1,000** caption templates. No `docs/` directory |
+| https://raw.githubusercontent.com/MiniMax-AI/MiniMax-Music3/main/skills/music-caption-rewriter/SKILL.md | PRIMARY (vendor agent skill) | fetched 2026-09-08 | The authoritative caption contract: three headings, instrumental handling, English default, 250–450 words, "never reproduce lyrics", precedence rules |
+| https://raw.githubusercontent.com/MiniMax-AI/MiniMax-Music3/main/skills/music-caption-rewriter/references/genre-router.md | PRIMARY | fetched 2026-09-08 | 18 style families, alias/fusion routing rules |
+| https://raw.githubusercontent.com/MiniMax-AI/MiniMax-Music3/main/skills/music-caption-rewriter/templates/acoustic-folk-singer-songwriter_0001.txt | PRIMARY (vendor example caption) | fetched 2026-09-08 | Concrete shape of a full Structured Caption |
+| https://raw.githubusercontent.com/MiniMax-AI/MiniMax-Music3/main/skills/README.md | PRIMARY | fetched 2026-09-08 | Install path (`npx skills add MiniMax-AI/MiniMax-Music3 --skill music-caption-rewriter`), progressive-disclosure design |
+| https://huggingface.co/docs/diffusers/main/en/api/pipelines/minimax_music3 | PRIMARY (diffusers docs) | fetched 2026-09-08 | The decisive 44.1 kHz vs 32 kHz statement, the "may drift instrumental" tip, CFG 1.7 swap recipe, ~23 GB VRAM, 25 fps dominates runtime |
+| https://platform.minimax.io/docs/guides/music-generation | PRIMARY (vendor API docs) | fetched 2026-09-08; notice dated 2026-08-20 | Hosted API only: `music-3.0`, lyrics 10–1000 chars, `is_instrumental`, `lyrics_optimizer`, 44100 Hz / 256 kbps MP3, tags incl. `[Hook]`; **paid music API closed to new users from 2026-08-20** |
+| https://www.minimax.io/blog/minimax-music-3-0-next-generation-open-weights-production-ready-versatile-music-model | PRIMARY (vendor blog) | published 2026-08-13 | Launch claims: five minutes, structured-caption philosophy, Mandarin + English demos |
+| https://minimax3.com/blog/how-to-use-minimax-music-3 | SECONDARY | published 2026-08-14 | Mostly restates the card; original advice: iterate at 30–60 s before requesting a long track; "brief = how it sounds, lyrics = what happens per section" |
+| https://www.minimax-music.com/minimax-music-3-prompt-guide, https://minimaxmusic3.ai/, https://minimax-ai.chat/models/minimax-music-3-0/ | SECONDARY (aggregators) | surfaced 2026-09-08 | All restate the model card's tag list and Structured Caption; no independent claims. Not relied on |
+| https://huggingface.co/MiniMaxAI/MiniMax-Music3/discussions/17 | SECONDARY (community) | opened ~2026-08-15 ("24 days ago") | A community-authored system prompt; **no MiniMax staff reply**. Not authoritative |
+
+Failed fetches: `https://github.com/huggingface/diffusers/blob/minimax-music3-integration/docs/source/en/api/pipelines/minimax_music3.md` → HTTP 404 (the integration branch is gone; the doc is now on `main` and was read there instead).
+
+---
+
+## Claim-by-claim verdict
+
+### `workflows/templates/minimax/README.md`
+
+| # | Claim | Verdict |
+| --- | --- | --- |
+| 1 | "music generation with [MiniMax-Music3](https://huggingface.co/MiniMaxAI/MiniMax-Music3)" | **CONFIRMED** — repo id correct, model card is live. |
+| 2 | "fitted onto a single 24GB consumer GPU… Every example here runs on an RTX 3090" | **CONFIRMED** — model card "Low VRAM": full precision fits under 24 GB; auto CPU offload ≈22 GB. diffusers docs: "The full pipeline needs ~23 GB of VRAM in bfloat16." The templates use `enable_auto_cpu_offload`, the card's ~22 GB path. |
+| 3 | The prompt-format paragraph ("The prompt format is MiniMax's own. Two guides… `docs/VIDEO_PROMPT_WRITING_GUIDE_*`… `skills/h3-prompt-writing`") | **CONFIRMED for H3, but the README never says the equivalent for Music3.** Music3 has its own vendor prompt authority — `skills/music-caption-rewriter` in `github.com/MiniMax-AI/MiniMax-Music3` — and the README does not mention it. See *Missing knowledge*. (There is no `docs/` folder in either the HF or the GitHub Music3 repo.) |
+| 4 | music.json row: "The minimal modular pipeline: a `components_manager` owns device placement, and the output is audio, not video" | **CONFIRMED** — matches the model card's Low-VRAM snippet exactly (`ComponentsManager` + `enable_auto_cpu_offload` + `load_components(dtype=bfloat16)`), and `MiniMaxMusic3Blocks.outputs` is `audios` only. |
+| 5 | "Music3 reads it as a ceiling rather than a target. The piece ends where the music ends" | **CONFIRMED, twice over.** `modular_blocks_minimax_music3.py`: "`audio_duration` … Upper bound on the generated audio length in seconds. The language model may stop earlier. Capped at 9000 frames (six minutes)." diffusers docs Tips: "`audio_duration` is an upper bound — the language model may end the song earlier with a stop token." Model card, SGLang path: "`max_new_tokens` sets the maximum number of audio frames at 25 frames per second. Generation may finish before this limit when the model emits an end-of-audio token." |
+| 6 | "a value set to the length the song *should* be will guillotine the outro mid-decay" | **UNSOURCED (mechanism inverted).** No source describes truncation at the ceiling; the documented behaviour is the opposite — the model emits a stop token and the run *ends short*. A hard cut can happen when the model has not finished by `max_frames` (`encoders.py` breaks at `len(frame_hiddens) >= max_frames`), so the advice "ask for more time than the song needs" is *sound*, but the stated reason ("will guillotine the outro") is the repo's own inference, not a sourced claim. |
+| 7 | "trim the tail afterwards - [templates/audio-trim-fade.json]" | **CONFIRMED (repo-internal)** — `workflows/templates/audio-trim-fade.json` exists. Note that `music.json`'s description points at a `tasks/` TrimFadeAudio path, which does **not** exist (see #14). |
+| 8 | "Output audio is 32 kHz stereo" (canvas paragraph) | **CONFIRMED but scoped to H3, not Music3.** The sentence sits in the H3 canvas note and is correct there. If a reader carries it to Music3 it is wrong: the Music3 diffusers path emits 44.1 kHz (see #16). |
+| 9 | music-video.json row: "a music video cut to a generated song: `slice_audio` deals frame-exact pieces to lip-synced shots, and `pair_audio` lays the unbroken track over the finished edit" | **CONFIRMED (repo-internal)** — the workflow does exactly that. No vendor bearing. |
+
+### `workflows/templates/minimax/music.json`
+
+| # | Setting / description sentence | Verdict |
+| --- | --- | --- |
+| 10 | `component_type: "ModularPipeline"`, `from_pretrained_arguments.model_name: "MiniMaxAI/MiniMax-Music3"` | **CONFIRMED** — the model card's own snippet is `ModularPipeline.from_pretrained("MiniMaxAI/MiniMax-Music3")`. (`MiniMaxMusic3ModularPipeline` — the class named in `docs/QUANTIZATION.md:187` and `docs/WORKFLOW_GUIDE.md:797` — is also real and resolves via `modular_config_dict`.) |
+| 11 | `load_components.dtype: torch.bfloat16`; `components_manager.enable_auto_cpu_offload: true` | **CONFIRMED** — verbatim the card's Low-VRAM recipe. |
+| 12 | Arguments `prompt`, `lyrics`, `audio_duration`, `output: "audios"` | **CONFIRMED** — exact match to the card's and the docs page's `pipe(prompt=…, lyrics=…, audio_duration=…, output="audios")`. |
+| 13 | `"'audio_duration' is a ceiling, not a target: the model stops where the music stops"` | **CONFIRMED** — same citations as #5. |
+| 14 | `"finish the tail with a tasks/ TrimFadeAudio path"` | **CONTRADICTED (repo-internal, stale path).** No such file exists under workflows/tasks; the file is `workflows/templates/audio-trim-fade.json`, which is what the README correctly names. |
+| 15 | `"The result declares 'sample_rate' because a modular pipeline's dict output carries none of its own."` | **CONFIRMED (engine behaviour).** `MiniMaxMusic3Blocks.outputs` returns only the raw `audios` tensor; the rate lives on the pipeline (`MiniMaxMusic3ModularPipeline.sampling_rate`), not in the output. |
+| 16 | `result.sample_rate: 44100` | **CONFIRMED — and this is the subtle one.** `minimax_music3_vocoder.py:85` `sampling_rate: int = 44100`; the checkpoint's `vocoder/config.json` is `"sampling_rate": 44100`; `condition_embedder_minimax_music3.py:40` `output_sampling_rate: int = 44100`. The **model card says "32 kHz, 16-bit stereo WAV"**, which is the SGLang server's resampled response format, not the diffusers output. The diffusers docs page settles it: "The pipeline returns the vocoder's native 44.1 kHz stereo output. The reference server additionally resamples to 32 kHz; apply your own resampling if you need that exact rate." The repo is **right**, and right for a non-obvious reason worth recording. |
+| 17 | `result.content_type: "audio/mp3"` | **UNSOURCED / engine choice.** No vendor bearing. The vendor's own outputs are WAV (SGLang) or MP3 at 44100 Hz / 256 kbps (hosted API), so MP3 at 44.1 kHz is a defensible container choice, but the tensor is lossless and the repo re-encodes. |
+| 18 | Stereo output implied by the audio result | **CONFIRMED** — `MiniMaxMusic3Vocoder.forward` returns `(batch, 2, samples)`; "the two audio channels are decoded as two folded `latent_channels // 2` streams". |
+| 19 | Lyrics variable: `[verse]` / `[chorus]` tags each alone on a line, lower-case | **CONFIRMED.** `encoders.py::_normalize_lyrics` keeps only leading bracketed tags on a tag line ("text on the same line as a leading tag is dropped"), lower-cases every tag, and prepends `[start]`. Model card tag list: `[Intro]`, `[Verse]`, `[Pre-Chorus]`, `[Chorus]`, `[Post-Chorus]`, `[Bridge]`, `[Instrumental]`, `[Solo]`, `[Outro]`. The template uses only `[verse]`/`[chorus]` — legal, but the narrowest slice of the vocabulary (no `[intro]`, `[bridge]`, `[outro]`), which sits oddly beside the description's advice about the outro decaying. |
+| 20 | `audio_duration: 120` (two minutes) | **CONFIRMED as legal** — under the 9,000-frame / 360 s engine cap and the vendor's five-minute maximum. Consistent with the ceiling advice (the lyrics are ~4 sections, far under two minutes). |
+| 21 | `num_inference_steps` not set (falls to the default) | **CONFIRMED as the reference value** — default 30 in both `MiniMaxMusic3ChunkSetTimestepsStep` and `MiniMaxMusic3ChunkDenoiseInner`. |
+| 22 | No `guidance_scale` set | **CONFIRMED as correct** — flow-stage CFG is a *guider* config, not a call argument: `ComponentSpec("guider", ClassifierFreeGuidance, config=FrozenDict({"guidance_scale": 1.7}))`. It cannot be passed as an argument; it is swapped with `pipe.update_components(guider=…)`. |
+| 23 | `cost`: RTX 3090 / 24 GB / 3.6 minutes | **UNSOURCED** (repo measurement). Plausible: the docs page says "The autoregressive stage generates 25 frames per second of audio and dominates the runtime", so runtime tracks *generated* length, not `audio_duration`. |
+| 24 | Prompt `prompt:minimax/acoustic_pop_song` — the "Genre: … BPM: … Key: … Vocals: … Arrangement: …" single-paragraph style | **CONFIRMED as a valid, vendor-shown form** — the model card's and the docs page's own diffusers example uses exactly this inline form ("Genre: acoustic pop. BPM: 96. Key: C major. … Vocals: … Arrangement: …"). It is the *concise* form; the vendor's recommended form for precise control is the three-heading Structured Caption. The prompt also names vocal gender and timbre ("soft female lead, close-mic breathy"), satisfying the docs tip in #26. |
+
+### `workflows/templates/minimax/music-video.json` (Music3 parts only)
+
+| # | Setting / sentence | Verdict |
+| --- | --- | --- |
+| 25 | `write_song` step: same pipeline class, arguments, `components_manager`, `content_type: audio/mp3`, `sample_rate: 44100`, `release_pipeline: true` | **CONFIRMED** — identical to #10–#12, #16, #18. `release_pipeline` is required here, per `docs/RECIPES_24GB.md:99` (host RAM, not VRAM, is the binding constraint before H3 loads). |
+| 26 | `song_prompt` names a "smooth male crooner lead" | **CONFIRMED as best practice** — diffusers docs Tips: "The music description controls the vocals: describe the vocal gender and timbre explicitly (e.g. 'warm female vocal') or the model may drift instrumental." Both repo prompts do this. |
+| 27 | `song_lyrics`: `[verse]` + `[chorus]`, tags on their own lines | **CONFIRMED** — same contract as #19. |
+| 28 | `audio_duration: 21` against a `soundtrack` slice of 496 frames at 24 fps (= 20.667 s) and four 124-frame slices at offsets 0/124/248/372 | **CONTRADICTED by the repo's own rule.** `audio_duration` is a ceiling; the model may stop early. The workflow needs at least 20.667 s of *actual* audio and asks for a ceiling of 21 s — a 0.33 s margin. The README and `music.json` both say "ask for more time than the song needs"; this template asks for essentially exactly what it needs. An early stop, or a song that simply ends at 18 s, leaves `slice_4` and the `soundtrack` slice reading past the end of the waveform. A ceiling of ~30 s with the same 496-frame slicing would follow the repo's own advice. |
+| 29 | `"MiniMax-Music3 writes the song, 'slice_audio' cuts it into frame-exact pieces (124 frames at 24 fps each)"` | **CONFIRMED (arithmetic)** — 124 / 24 = 5.167 s per slice; `sample_rate: 44100` is used for the slicing, matching #16. Note the two frame rates in play: Music3 generates at **25 Hz** internally, H3 renders at **24 fps**; slicing is done in H3 frames against the 44.1 kHz waveform, which is correct. |
+| 30 | `"The generation models pass through one at a time - each is released before the next loads - so the workflow peaks no higher than its largest single model."` | **CONFIRMED (repo-internal)** — `release_pipeline: true` on `draw_singer` and `write_song`; consistent with `docs/RECIPES_24GB.md:95-101`. |
+
+### Prompt library
+
+| # | Claim | Verdict |
+| --- | --- | --- |
+| 31 | `prompts/minimax/acoustic_pop_song.json`, `otter_soul_song.json` — inline `Genre/BPM/Key/Emotional progression/Listening scenario/Production profile/Vocals/Arrangement` style | **CONFIRMED as well aligned.** These field names track the model card's Global Metadata list almost item for item ("genre, subgenre, BPM, key, scale, emotional progression, listening scenario, and production profile"), and cover Vocal Details and Arrangement. They are the card's vocabulary in the card's *concise* rendering. `intended_model: "minimax-music"` and the `music` tags are repo conventions. |
+| 32 | No prompt in the library uses the three-heading Structured Caption form | **Gap, not an error.** `encoders.py::_clean_caption` explicitly strips markdown headings, bullets, bold and italics from the caption, so a markdown three-heading caption is *safe to paste* — the vendor's recommended form is fully supported and the library does not exercise it. |
+
+### Docs mentions
+
+| # | Claim | Verdict |
+| --- | --- | --- |
+| 33 | `docs/QUANTIZATION.md:187` / `docs/WORKFLOW_GUIDE.md:797` use `"component_type": "MiniMaxMusic3ModularPipeline"` with per-component `quantization_config` on `transformer` and `text_encoder` | **PARTLY CONTRADICTED.** The class name is real. But Music3 has **no component named `text_encoder`** — the pipeline's components are `tokenizer`, `language_model`, `rvq_depth_decoder`, `condition_encoder`, `transformer`, `scheduler`, `guider`, `vocoder` (`MiniMaxMusic3Blocks` docstring). The doc snippet reads as illustrative rather than runnable; as written the `text_encoder` entry would name nothing. The language model is the one worth quantizing/offloading (the card group-offloads `pipe.language_model` to reach 8 GB). |
+| 34 | `docs/RECIPES_24GB.md:99` "a workflow that ran another model first (Z-Image drawing a subject, Music3 writing a song) must free it with `release_pipeline`" | **CONFIRMED (repo measurement)** — internally consistent with the templates; no vendor bearing. |
+| 35 | `plugins/dw/skills/minimax-h3/SKILL.md:54,82` — "Music alone: `templates/minimax/music` (Music3)" and the ceiling rule | **CONFIRMED**, same citations as #5. Music3 currently exists in the plugin only as a footnote inside the H3 skill. |
+| 36 | `docs/TASKS.md` | **No Music3 mentions** (grep clean). Not an error; `slice_audio`/`pair_audio`/trim-fade are documented model-agnostically. |
+
+### Claims the task asked about that the repository does **not** make
+
+- **Maximum duration**: unstated anywhere in the templates. Vendor: **five minutes** (model card, blog 2026-08-13). Engine: `_MAX_AUDIO_FRAMES = 9_000` at 25 fps = **360 s / six minutes** (`encoders.py`, and the docstring's "Capped at 9000 frames (six minutes)"). These disagree by a minute; the engine cap is the hard stop, the vendor's five minutes is the supported/trained range. Only the internal planning docs (`docs/proposals/agent-catalog-legibility.md:649`, `docs/superpowers/plans/2026-09-07-ltx-h3-catalog-repair.md:1062`) carry the 9000/six-minute figure, and they carry it without the five-minute caveat.
+- **Full structure-tag vocabulary**: UNSOURCED in the repo — it only ever shows `[verse]`/`[chorus]`.
+- **Instrumental-only**: not addressed anywhere in the repo. Vendor position (see below).
+- **Language support**: not addressed. Vendor: blog demos are Mandarin and English; the caption skill says write the *caption* in English by default; no published language list.
+- **Steps / guidance defaults**: not stated in any template or doc.
+- **Prompt-token limit**: not stated. `_MAX_PROMPT_TOKENS = 5_000` raises a hard `ValueError`.
+
+---
+
+## Missing knowledge
+
+Everything below is published, primary, and absent from the repository.
+
+1. **The `music-caption-rewriter` agent skill** (GitHub `MiniMax-AI/MiniMax-Music3/skills/music-caption-rewriter`, current as of 2026-09-08). This is the exact Music3 counterpart of the H3 prompt-writing guides the README already points at, and the README does not mention it. It ships `SKILL.md`, a `references/genre-router.md`, **18 family indexes**, and **1,000 full caption templates**. Installable with `npx skills add MiniMax-AI/MiniMax-Music3 --skill music-caption-rewriter`. *(Repo gap dated 2026-09-08.)*
+2. **The Structured Caption contract** (model card, 2026-08; SKILL.md "Output Contract"): exactly three top-level headings, in order — **Global Metadata** (genre/subgenres, tempo, emotional progression, sonic and production profile; exact BPM/key only when justified), **Vocal Details** (lead configuration, timbre, register, delivery, harmony, restrained FX), **Arrangement** (a section-by-section timeline: what enters, exits, changes, intensifies). Target **250–450 English words**. Do not include a title, template ID, reasoning trace, or any copied lyric line. The repo's prompts are the concise form only.
+3. **The full tag vocabulary**: `[Intro]`, `[Verse]`, `[Pre-Chorus]`, `[Chorus]`, `[Post-Chorus]`, `[Bridge]`, `[Instrumental]`, `[Solo]`, `[Outro]` (model card, 2026-08; diffusers docs Tips lower-case the same list). The hosted API additionally lists `[Hook]` (platform.minimax.io, fetched 2026-09-08) — hosted-only, unverified for the open weights.
+4. **Tags are directives, not just markers**: a section tag can change its *local* arrangement without replacing the global genre, and the caption skill is instructed to preserve section-local directives attached to lyric tags in the Arrangement text while the lyric text itself stays in `lyrics` (SKILL.md, 2026-09-08). Nothing in the repo says this.
+5. **Instrumental music is supported, and how it is expressed**: SKILL.md — "Preserve an explicit instrumental request. Do not add vocals"; the Vocal Details heading must "state that the piece is instrumental and identify the instrument or texture carrying the lead melodic role". The hosted API has a dedicated `is_instrumental` flag (platform docs), **which the diffusers path does not expose** — `MiniMaxMusic3TokenizeStep.check_inputs` *requires* a non-empty `lyrics` string. So in this repo an instrumental means a tag-only `lyrics` (e.g. `[intro]\n[instrumental]\n[solo]\n[outro]`) plus a caption that says instrumental. *(2026-09-08; the diffusers-side consequence is an inference from `encoders.py`, not a vendor statement.)*
+6. **The opposite failure mode — accidental instrumental**: diffusers docs Tips (fetched 2026-09-08): "describe the vocal gender and timbre explicitly … or the model may drift instrumental." The repo's prompts happen to do this; nothing tells an author it is mandatory.
+7. **Prompt-token ceiling of 5,000** with a hard error (`encoders.py`, model card Limitations). A 450-word Structured Caption is nowhere near it, but a pasted 1,000-template dump would be.
+8. **Flow-stage CFG is 1.7 and lives on the guider**, changed only via `pipe.update_components(guider=ClassifierFreeGuidance(guidance_scale=...))` (diffusers docs Tips + `denoise.py`). The AR stage has its own fixed 1.5 that is not user-settable.
+9. **Runtime scales with generated length, not with `audio_duration`** — "The autoregressive stage generates 25 frames per second of audio and dominates the runtime" (diffusers docs). This is what makes the "ask for more than you need" advice cheap; the repo asserts the advice without the reason.
+10. **The 8 GB path**: `apply_group_offloading(pipe.language_model, offload_type="leaf_level", use_stream=True)` (model card Low VRAM + docs). The repo's 24 GB framing never mentions that Music3 alone runs on far less.
+11. **The 32 kHz vs 44.1 kHz distinction** — the vendor advertises 32 kHz, the diffusers pipeline emits 44.1 kHz, and the difference is a resample in the reference server (diffusers docs, 2026-09-08). Anyone reading the model card and the repo side by side will think one of them is wrong. Nothing in the repo explains it.
+12. **Vendor five-minute maximum vs the engine's 9,000-frame / six-minute cap** (model card vs `encoders.py`). Undated conflict; both current as of 2026-09-08.
+13. **Explicit vendor limitation**: "Section tags and music descriptions provide generative control rather than strict symbolic guarantees. The generated tempo, key, instrumentation, lyrics, and song structure may not always match every requested detail exactly" (model card Limitations, 2026-08). The repo's prompts specify exact BPM and key with no caveat.
+14. **Inference requires CUDA** (model card Limitations, 2026-08). The diffusers docs snippet says `pipe.to("cuda")  # or "mps", "xpu", "cpu"`, so the two disagree; on this Mac the repo would fall back under `resolve_device()` with unknown results. Untested either way.
+15. **The hosted music API closed to new users on 2026-08-20** (platform.minimax.io notice). Anything the repo might later want to borrow from the API surface (`is_instrumental`, `lyrics_optimizer`, cover mode) is now open-weights-only territory.
+16. **Iterate short** — secondary but repeated: "Start with 30 or 60 seconds, fix the brief, and only then request a longer track" (minimax3.com, 2026-08-14). Consistent with #9. `music.json`'s default of 120 s is a slow first iteration.
+17. **No newer checkpoint** — as of 2026-09-08 the only published open-weights music checkpoint is `MiniMaxAI/MiniMax-Music3`; the hosted model id is `music-3.0` and `music-cover`. No `Music3.1`/`Music4` found.
+
+---
+
+## What a skill should teach
+
+**Shape first.** A *song* is lyrics with tags plus a caption naming vocal gender and timbre — the default. An *instrumental* is a tag-only `lyrics` body (`[intro]`/`[instrumental]`/`[solo]`/`[outro]`) plus a caption that says instrumental and names the lead instrument; diffusers has no `is_instrumental` flag and rejects empty lyrics. A *score under a film* is an instrumental generated to a ceiling comfortably longer than the cut, then trimmed and faded with `templates/audio-trim-fade`. A *music video* is `music.json` first, then frame-exact `slice_audio` into fresh H3 shots — and the ceiling must exceed the total sliced length with real margin, not by a third of a second.
+
+**Hard numbers, each pinned to a constant.**
+- `audio_duration` is an upper bound; the model may stop earlier — `MiniMaxMusic3AutoregressiveStep` InputParam docstring; default `60.0`.
+- Frame rate 25 Hz — `MiniMaxMusic3ModularPipeline.frame_rate`, = `input_sampling_rate 24000` / `input_hop_length 960`.
+- Hard cap `_MAX_AUDIO_FRAMES = 9_000` = 360 s; vendor supports five minutes. Prefer 300 s.
+- Prompt cap `_MAX_PROMPT_TOKENS = 5_000` — hard `ValueError`.
+- Output 44,100 Hz **stereo** — `minimax_music3_vocoder.py:85`, `vocoder/config.json`; the card's "32 kHz" is the reference server's resample.
+- `num_inference_steps` default 30 per 200-frame window (`_CHUNK_FRAMES = 200`, `_CHUNK_HOP = 100`).
+- Flow CFG 1.7 on the guider, not an argument (`ComponentSpec("guider", …, {"guidance_scale": 1.7})`).
+- Tag lines: leading bracketed tags only; text on a tag line is dropped (`_normalize_lyrics`).
+- ~22 GB with auto CPU offload; 8 GB with `language_model` leaf-level group offload.
+
+**Point, don't transcribe.** Send the author to `skills/music-caption-rewriter/SKILL.md` and `references/genre-router.md` in `github.com/MiniMax-AI/MiniMax-Music3` for the three-heading caption contract, the 18-family routing, and the 1,000 templates; to the model card's "Fine-Grained Music Control" for the tag vocabulary; to the diffusers docs Tips for the drift-instrumental warning. Say only that markdown headings are stripped safely by `_clean_caption`, so the vendor's caption pastes straight in.
+
+---
+
+## Assessment
+
+The repository is largely right, and right about the hard parts. The `audio_duration`-as-ceiling rule is confirmed verbatim by three independent primary sources. `sample_rate: 44100` is correct against the checkpoint config even though the model card says 32 kHz — the repo picked the right number for a reason it never records, and should. The pipeline wiring, `components_manager` recipe, `release_pipeline` discipline, and the prompt vocabulary all track the vendor's own examples.
+
+Four things need correction. `music.json` points at a `tasks/` TrimFadeAudio path, which does not exist. `music-video.json` sets `audio_duration: 21` for 20.667 s of slices, violating the very margin rule it states. The README's "will guillotine the outro mid-decay" asserts a mechanism no source describes — the documented behaviour is an early stop. And `docs/QUANTIZATION.md`/`WORKFLOW_GUIDE.md` show a `text_encoder` component Music3 does not have.
+
+The real gap is coverage, not accuracy: Music3 has a full vendor prompt-authority — the `music-caption-rewriter` skill with 1,000 templates — that the repo never mentions, alongside the full tag list, instrumental handling, the five-vs-six-minute ceiling, and the 32/44.1 kHz explanation.
diff --git a/docs/superpowers/plans/2026-09-07-dw-plugin-skills.md b/docs/superpowers/plans/2026-09-07-dw-plugin-skills.md
new file mode 100644
index 00000000..bbdef8c3
--- /dev/null
+++ b/docs/superpowers/plans/2026-09-07-dw-plugin-skills.md
@@ -0,0 +1,766 @@
+# dw plugin: model-family composition skills — implementation plan
+
+> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
+
+**Goal:** A Claude Code plugin in this repo, installed once, holding one composition skill per model family (MiniMax H3, LTX-2.5) that teaches which template fits which shape, the hard rules, cost, and the run-and-judge loop, and defers prompt format to the vendors' own text.
+
+**Architecture:** Repository content only: a marketplace manifest at the root, one plugin under `plugins/dw/` with two `SKILL.md` files, a README, a version line the release script bumps. Drift tests in `tests/test_plugin_skills.py` resolve every catalog name a skill quotes, pin every number a skill states to the diffusers module that enforces it, and hold the one quoted vendor text equal to the library constant. Nothing under `dw/`, `dw_mcp/` or the wheel changes.
+
+**Tech Stack:** Claude Code plugin format (`.claude-plugin/marketplace.json`, `.claude-plugin/plugin.json`, `skills/<name>/SKILL.md` with YAML frontmatter), pytest, diffusers 0.41 (installed in the venv; `diffusers.pipelines.ltx2.utils`, `diffusers.modular_pipelines.minimax_h3`).
+
+**Spec:** [docs/superpowers/specs/2026-09-07-dw-plugin-skills-design.md](../specs/2026-09-07-dw-plugin-skills-design.md). Package A (merged as #52) is the catalog these skills describe.
+
+## Global Constraints
+
+- Composition only: a skill carries no prompt format of its own. H3 prompts come from MiniMax's `h3-prompt-writing` skill or the two guides on the model card, else the family's enhance-prompt template; LTX-2.5 prompts follow `LTX2_5_T2V_DEFAULT_SYSTEM_PROMPT` quoted from diffusers and held equal by test.
+- Every catalog name a skill quotes in backticks (`templates/...`) resolves to a file under `workflows/`.
+- Every number a skill states as a rule is pinned by a test to the source it comes from.
+- Each `SKILL.md` is under 12 KB and its frontmatter `description` names the model.
+- `plugins/dw/.claude-plugin/plugin.json` `version` equals `pyproject.toml`'s.
+- No MCP server declared by the plugin; a skill's first instruction is `get_server_info`.
+- Nothing under `dw/`, `dw_mcp/`, `MANIFEST.in` or `pyproject.toml`'s package data changes.
+- Test names and docstrings state the behaviour, matching the suite's style.
+- Commit messages end with `Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>`.
+- Each task appends one sentence to the Part 4 row of the ledger table in `docs/proposals/agent-catalog-legibility.md` (the row beginning `| Part 4 packaging |`, at the end of its last cell before the closing `|`), in the same commit.
+- Work on branch `dw-plugin`, cut from `master` at or after 09e36e1.
+
+---
+
+## File map
+
+| File | Responsibility | Task |
+| --- | --- | --- |
+| `.claude-plugin/marketplace.json` | names the marketplace and its one plugin | 1 |
+| `plugins/dw/.claude-plugin/plugin.json` | plugin identity and version | 1 |
+| `plugins/dw/README.md` | install and what each skill covers | 1 |
+| `README.md` | two install lines after the MCP registration | 1 |
+| `scripts/release.sh` | bumps plugin.json beside pyproject.toml | 1 |
+| `tests/test_plugin_skills.py` | version parity; generic skill invariants; family pins | 1, 2, 3 |
+| `plugins/dw/skills/minimax-h3/SKILL.md` | the H3 composition skill | 2 |
+| `workflows/templates/minimax/README.md` | one line pointing at the skill | 2 |
+| `plugins/dw/skills/ltx-2.5/SKILL.md` | the LTX-2.5 composition skill | 3 |
+| `workflows/templates/ltx2/README.md` | one line pointing at the skill | 3 |
+| `docs/proposals/agent-catalog-legibility.md` | ledger row; drill result; the mould for the next family | every task, 4 |
+
+---
+
+### Task 1: The plugin scaffold, installable and version-locked
+
+**Recommended model:** sonnet — every file's content is given below.
+
+**Files:**
+- Create: `.claude-plugin/marketplace.json`, `plugins/dw/.claude-plugin/plugin.json`, `plugins/dw/README.md`, `tests/test_plugin_skills.py`
+- Modify: `README.md` (after the `claude mcp add --transport http ...` block, before the Server-page paragraph), `scripts/release.sh` (the version-bump block, lines ~50-77)
+
+**Interfaces:**
+- Produces: the directory `plugins/dw/skills/` (empty until Tasks 2 and 3), the test module `tests/test_plugin_skills.py` with helpers `SKILLS` (list of SKILL.md paths), `skill_text(path)`, `frontmatter(text)` that later tasks extend.
+
+- [ ] **Step 1: Write the failing test**
+
+Create `tests/test_plugin_skills.py`:
+
+```python
+"""The dw plugin: skills that teach an agent to compose a model family's
+workflows, kept true by tests rather than by memory.
+
+A skill quotes catalog names and states numeric rules. Each name has to
+resolve to a file, each number has to come from the library that enforces
+it, and the plugin's version has to be the engine's - an installed skill is
+matched to the server it was written against by that number.
+"""
+
+import glob
+import os
+import re
+
+import pytest
+
+from tests.test_examples import REPO_ROOT
+
+PLUGIN_DIR = os.path.join(REPO_ROOT, "plugins", "dw")
+SKILLS = sorted(glob.glob(os.path.join(PLUGIN_DIR, "skills", "*", "SKILL.md")))
+SKILL_SIZE_LIMIT = 12 * 1024
+
+CATALOG_NAME = re.compile(r"`((?:templates|models)/[A-Za-z0-9_./-]+)`")
+
+
+def skill_text(path):
+    return open(path, encoding="utf-8").read()
+
+
+def frontmatter(text):
+    """The YAML block between the leading '---' lines, as a dict of the
+    top-level 'key: value' pairs (the two the plugin format needs)."""
+    assert text.startswith("---\n"), "a skill begins with frontmatter"
+    end = text.index("\n---", 4)
+    fields = {}
+    for line in text[4:end].splitlines():
+        if ":" in line and not line.startswith(" "):
+            key, value = line.split(":", 1)
+            fields[key.strip()] = value.strip().strip('"')
+    return fields
+
+
+def _pyproject_version():
+    text = open(os.path.join(REPO_ROOT, "pyproject.toml"), encoding="utf-8").read()
+    return re.search(r'^version = "(.*)"$', text, re.M).group(1)
+
+
+def test_the_marketplace_names_the_plugin():
+    import json
+
+    manifest = json.load(open(os.path.join(REPO_ROOT, ".claude-plugin", "marketplace.json"), encoding="utf-8"))
+
+    assert manifest["name"] == "diffusers-workflow"
+    (plugin,) = manifest["plugins"]
+    assert plugin["name"] == "dw"
+    assert plugin["source"] == "./plugins/dw"
+
+
+def test_the_plugin_version_is_the_engine_version():
+    """An installed plugin is matched to the engine it was written against by
+    this number, so the release script bumps both in one commit."""
+    import json
+
+    plugin = json.load(open(os.path.join(PLUGIN_DIR, ".claude-plugin", "plugin.json"), encoding="utf-8"))
+
+    assert plugin["name"] == "dw"
+    assert plugin["version"] == _pyproject_version()
+
+
+def test_there_are_skills():
+    assert SKILLS, "the plugin ships at least one skill"
+
+
+@pytest.mark.parametrize("path", SKILLS, ids=lambda p: os.path.basename(os.path.dirname(p)))
+def test_a_skill_has_a_triggering_description_under_the_size_cap(path):
+    text = skill_text(path)
+    fields = frontmatter(text)
+
+    assert fields["name"] == os.path.basename(os.path.dirname(path))
+    assert "description" in fields and len(fields["description"]) > 40
+    assert len(text.encode("utf-8")) <= SKILL_SIZE_LIMIT, f"{path} is over {SKILL_SIZE_LIMIT} bytes"
+
+
+@pytest.mark.parametrize("path", SKILLS, ids=lambda p: os.path.basename(os.path.dirname(p)))
+def test_every_catalog_name_a_skill_quotes_resolves(path):
+    """A renamed template fails here rather than in a cold session."""
+    names = CATALOG_NAME.findall(skill_text(path))
+
+    assert names, f"{path} quotes no catalog names"
+    for name in names:
+        target = os.path.join(REPO_ROOT, "workflows", name.removesuffix(".json") + ".json")
+        assert os.path.isfile(target), f"{path} quotes {name}, which is not a workflow ({target})"
+```
+
+- [ ] **Step 2: Run it to see it fail**
+
+Run: `pytest tests/test_plugin_skills.py -v`
+Expected: the marketplace and version tests FAIL with `FileNotFoundError`; `test_there_are_skills` FAILS on the empty list; the parametrized tests are collected as zero cases.
+
+- [ ] **Step 3: Create the manifests**
+
+`.claude-plugin/marketplace.json`:
+
+```json
+{
+    "$schema": "https://anthropic.com/claude-code/marketplace.schema.json",
+    "name": "diffusers-workflow",
+    "description": "Skills for driving a diffusers-workflow server from Claude Code: one composition skill per model family.",
+    "owner": {
+        "name": "Don Kackman"
+    },
+    "plugins": [
+        {
+            "name": "dw",
+            "source": "./plugins/dw",
+            "description": "Compose MiniMax H3 and LTX-2.5 video workflows over a dw MCP server: which template fits which shape, the hard rules, cost, and how to judge the output. Prompt format comes from the vendors' own guides."
+        }
+    ]
+}
+```
+
+`plugins/dw/.claude-plugin/plugin.json`, with the version copied from `pyproject.toml`'s `version = "..."` line at the time of writing:
+
+```json
+{
+    "name": "dw",
+    "description": "Compose MiniMax H3 and LTX-2.5 video workflows over a dw MCP server: which template fits which shape, the hard rules, cost, and how to judge the output. Prompt format comes from the vendors' own guides.",
+    "version": "<the pyproject version>",
+    "author": {
+        "name": "Don Kackman"
+    },
+    "repository": "https://github.com/dkackman/diffusers-workflow",
+    "keywords": [
+        "diffusers",
+        "video",
+        "minimax-h3",
+        "ltx-2.5",
+        "mcp"
+    ]
+}
+```
+
+`plugins/dw/README.md`:
+
+```markdown
+# dw plugin
+
+Skills for Claude Code that teach an agent to compose a model family's
+workflows on a [diffusers-workflow](https://github.com/dkackman/diffusers-workflow)
+server. They assume the `dw` MCP server is already registered
+(see the repo README, "Drive it from Claude Code"); a skill's first move is
+`get_server_info`.
+
+Install once, from Claude Code:
+
+```
+/plugin marketplace add dkackman/diffusers-workflow
+/plugin install dw@diffusers-workflow
+```
+
+| Skill | Teaches |
+| ----- | ------- |
+| `minimax-h3` | MiniMax H3 video with audio: one take, a longer take by chain, a piece with cuts, identity and voice references, music; the frame and canvas rules; what a run costs. Prompts come from MiniMax's own `h3-prompt-writing` skill or the guides on the model card. |
+| `ltx-2.5` | LTX-2.5 video with a soundtrack: a single clip, first-frame and keyframe conditioning, the three-move two-stage flow, the IC-LoRA upscale, extend and chain; the distilled schedule and the frame and size rules. Prompts follow the trained-caption spec that ships inside diffusers. |
+
+Each skill quotes catalog names and numeric rules that `tests/test_plugin_skills.py`
+holds to the catalog and to the diffusers module that enforces them. The
+plugin's version is the engine's; the release script bumps both.
+
+Adding a family: copy a skill, follow its outline, add the family's rules to
+the test, cite the vendor. The repo's `model-family-onboarding` skill
+(`.claude/skills/`) is the full lifecycle.
+```
+
+- [ ] **Step 4: The README install lines**
+
+In `README.md`, after the fenced block that ends with `--header "Authorization: Bearer $DW_API_TOKEN"` and before the paragraph beginning `You don't have to compose that command by hand`, insert:
+
+```markdown
+The [dw plugin](plugins/dw/README.md) adds one skill per model family - what
+to run for a given shape, the rules that bite, what it costs - and points at
+the vendors' own prompt guides rather than restating them:
+
+```
+/plugin marketplace add dkackman/diffusers-workflow
+/plugin install dw@diffusers-workflow
+```
+```
+
+- [ ] **Step 5: The release script bumps plugin.json too**
+
+In `scripts/release.sh`, inside the `if [ "$current" != "$version" ]; then` block after the `echo "pyproject.toml: $current -> $version"` line, add:
+
+```bash
+    python3 - "$version" <<'EOF'
+import json, sys
+
+path = "plugins/dw/.claude-plugin/plugin.json"
+with open(path, encoding="utf-8") as f:
+    plugin = json.load(f)
+plugin["version"] = sys.argv[1]
+with open(path, "w", encoding="utf-8") as f:
+    json.dump(plugin, f, indent=4, ensure_ascii=False)
+    f.write("\n")
+EOF
+    echo "plugins/dw/.claude-plugin/plugin.json: -> $version"
+```
+
+and change the commit line `git commit -m "release $version" -- pyproject.toml` to `git commit -m "release $version" -- pyproject.toml plugins/dw/.claude-plugin/plugin.json`, and the `git diff --quiet -- pyproject.toml` guard to `git diff --quiet -- pyproject.toml plugins/dw/.claude-plugin/plugin.json`. Add one sentence to `docs/RELEASING.md` where it describes the bump: "The same commit sets `plugins/dw/.claude-plugin/plugin.json`'s version, so an installed plugin names the engine it was written against."
+
+- [ ] **Step 6: Run the tests**
+
+Run: `pytest tests/test_plugin_skills.py -v`
+Expected: marketplace and version tests PASS; `test_there_are_skills` still FAILS (Tasks 2 and 3 add the skills). Mark it `@pytest.mark.xfail(strict=True, reason="skills land in the next tasks")` for this commit only, and note in the report that Task 2 removes the marker.
+
+- [ ] **Step 7: Ledger and commit**
+
+Append to the Part 4 ledger row: `Plugin task 1: a marketplace at the repo root and the dw plugin under plugins/dw, installable with two commands the README shows; plugin.json's version is the engine's, bumped by release.sh and held by test.`
+
+```bash
+git add .claude-plugin plugins/dw README.md scripts/release.sh docs/RELEASING.md tests/test_plugin_skills.py docs/proposals/agent-catalog-legibility.md
+git commit -m "dw plugin scaffold: marketplace, plugin manifest, version lock
+
+Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>"
+```
+
+---
+
+### Task 2: The MiniMax H3 skill
+
+**Recommended model:** opus — the skill is prose an agent will act on; the draft below is the content, the implementer's job is to make it read as one instruction and to keep every fact exact.
+
+**Files:**
+- Create: `plugins/dw/skills/minimax-h3/SKILL.md`
+- Modify: `workflows/templates/minimax/README.md` (one line after the first paragraph), `tests/test_plugin_skills.py` (remove the xfail; add the H3 pins)
+
+**Interfaces:**
+- Consumes: Task 1's test helpers `skill_text`, `SKILLS`.
+- Produces: the outline every later family copies: Before anything / Which shape / Hard rules / Prompts / Run and judge / Sources.
+
+- [ ] **Step 1: Write the failing tests**
+
+Append to `tests/test_plugin_skills.py`:
+
+```python
+H3_SKILL = os.path.join(PLUGIN_DIR, "skills", "minimax-h3", "SKILL.md")
+
+
+class TestMiniMaxH3Skill:
+    """The numbers the H3 skill states come from the diffusers modular
+    pipeline that enforces them, so a library change fails here."""
+
+    def test_the_frame_rule_and_bounds_are_the_pipeline_s(self):
+        import inspect
+
+        from diffusers.modular_pipelines.minimax_h3 import before_encoder, modular_pipeline
+
+        text = skill_text(H3_SKILL)
+        assert "17n + 5" in text or "17 * n + 5" in text
+        assert "17 * n + 5" in inspect.getsource(before_encoder)
+        assert modular_pipeline.MINIMAX_H3_FPS == 24 and "24 fps" in text
+        # 124 and 345 are the smallest and largest 17n + 5 inside 5 to 15 seconds at 24 fps
+        assert "124" in text and "345" in text
+        assert 124 == 17 * 7 + 5 and 345 == 17 * 20 + 5
+        assert 124 / modular_pipeline.MINIMAX_H3_FPS >= 5 and 345 / modular_pipeline.MINIMAX_H3_FPS <= 15
+
+    def test_the_canvas_rules_are_the_pipeline_s(self):
+        import inspect
+
+        from diffusers.modular_pipelines.minimax_h3 import before_encoder, modular_pipeline
+
+        text = skill_text(H3_SKILL)
+        source = inspect.getsource(before_encoder)
+        assert 'ConfigSpec("canvas_short_edge", 768)' in source and "768" in text
+        assert "768 * 1344" in source and "1344" in text
+        assert modular_pipeline.MINIMAX_H3_MIN_ASPECT_RATIO == 1 / 4
+        assert modular_pipeline.MINIMAX_H3_MAX_ASPECT_RATIO == 4
+        assert "1:4" in text and "4:1" in text
+        assert "32" in text  # dimensions are multiples of 32
+
+    def test_the_skill_defers_prompt_format_to_minimax(self):
+        text = skill_text(H3_SKILL)
+        assert "h3-prompt-writing" in text
+        assert "VIDEO_PROMPT_WRITING_GUIDE_base_en.md" in text
+        assert "VIDEO_PROMPT_WRITING_GUIDE_ref_en.md" in text
+        assert "`templates/minimax/enhance-prompt`" in text
+        # no transcription of the format: none of its section labels appear as instructions
+        assert "retention_analysis:" not in text
+
+    def test_the_skill_starts_with_the_server(self):
+        text = skill_text(H3_SKILL)
+        assert text.index("get_server_info") < text.index("templates/minimax/")
+```
+
+- [ ] **Step 2: Run them to see them fail**
+
+Run: `pytest tests/test_plugin_skills.py -k MiniMaxH3 -v`
+Expected: FAIL with `FileNotFoundError` on the skill path.
+
+- [ ] **Step 3: Write the skill**
+
+Create `plugins/dw/skills/minimax-h3/SKILL.md`. The draft below is the content; keep every name, number and URL exactly, and edit only for flow. Stay under 12 KB.
+
+````markdown
+---
+name: minimax-h3
+description: Use when a dw MCP server is connected and the user wants MiniMax H3 video - a clip with its own sound, dialogue or a speaking character, a music video, a multi-shot short with cuts, a longer take, or a subject or voice kept consistent from a reference. Picks the template for the shape, states the rules that bite, quotes cost, and points at MiniMax's own prompt guides.
+---
+
+# MiniMax H3 on a dw server
+
+H3 generates video and audio together: speech with lip sync, ambient sound,
+score. Every template here fits a 24 GB card. This skill chooses the template
+and the arguments; the prompt format is MiniMax's and comes from their text,
+not from here.
+
+## Before anything
+
+1. `get_server_info`: the device (H3 templates are CUDA; the quantized
+   configurations do not run on mps) and which workspace this session is in.
+2. `list_workflows(shape="shot")` and `list_workflows(shape="sequence")`:
+   the family's templates by their current names, with `summary`, `traits`
+   and `cost`. Trust the listing over the names quoted below.
+3. `get_workflow` on the one chosen, for its variables and their defaults.
+
+## Which shape is the request
+
+- **One clip, up to 14 seconds, from text**: `templates/minimax/video-with-audio`.
+  From a one-line idea: `templates/minimax/enhance-prompt` writes the prompt
+  with the built-in Context-IR enhancer first.
+- **Pinned to a picture**: first frame `templates/minimax/image-to-video`;
+  first and last `templates/minimax/first-and-last-frame`; last only
+  `templates/minimax/last-frame-only`; a one-line idea plus a picture
+  `templates/minimax/enhance-prompt-with-image`.
+- **A subject that must look the same**: `templates/minimax/reference-to-video`
+  (an image fixes appearance, an audio clip fixes voice);
+  `templates/minimax/composable-references` adds a video reference for framing
+  and camera; `templates/minimax/generated-subject-reference` draws the subject
+  with Z-Image first and references it in the same workflow;
+  `templates/minimax/voice-timbre-reference` fixes a voice from a Bark-spoken line.
+- **Several boards in one generation, one unbroken score**:
+  `templates/minimax/storyboard` - H3 cuts between the boards inside a single
+  generation, which no concat of separate clips can match for continuous audio.
+- **Longer than 14 seconds as one take**: a chain. `templates/minimax/chained-segments`
+  (last-frame continuity), `templates/minimax/chain-video-continuity` (the
+  previous segment's tail rides along as a video reference - motion, camera and
+  voice carry across the seam), `templates/minimax/chain-matched-to-audio` (a
+  supplied track sets the length and is muxed back seamless),
+  `templates/minimax/chain-matched-and-aligned` (all of it, per-segment prompts).
+  Drift compounds per seam: reference the subject picture in every segment,
+  prefer `last_segment` continuity, and use the longest segments memory allows.
+- **A piece with cuts**: fresh shots from shared portraits, then a concat.
+  `templates/minimax/dialogue-short` (Z-Image draws the cast, one loaded model
+  per shot, `concat_videos` splices) and `templates/minimax/music-video`
+  (shots cut to a generated song, lip-synced slices). A cut erases drift; the
+  last shot is as clean as the first. Write shots, not takes.
+- **Music alone**: `templates/minimax/music` (Music3).
+
+If none fits, compose from `list_tasks` before authoring a new workflow, and
+read the `workflows` guide's authoring section first.
+
+## Hard rules
+
+- `num_frames` is `17n + 5`, from 124 to 345, at a fixed 24 fps: 5.2 to 14.4
+  seconds in one clip. Templates default to 124 for fast iteration; `num_frames=345`
+  is the full length and fits the same 24 GB configuration. The 5-second floor
+  is diffusers'; the model card says 4.
+- Canvas: a 768-pixel short edge, at most 768x1344 pixels, dimensions multiples
+  of 32, aspect from 1:4 to 4:1. The templates' 960x544 is the speed choice and
+  is coupled to the 544p turbo LoRA and its nine steps - change one, change all
+  three. Output audio is 32 kHz stereo.
+- H3 is guidance-distilled: no `guidance_scale`, no negative prompt. Say what is
+  there, never what is not.
+- Ref2VA limits: at most 9 images, 3 videos, 3 audio clips, 12 files; audio can
+  never be the only reference. References are labelled in the order passed.
+- Music3 reads `audio_duration` as a ceiling, not a target: ask for more than
+  the song needs and trim with `templates/audio-trim-fade`. Cap: six minutes.
+- Write the prompt for the length being generated: shot timestamps should span
+  the duration, or a five-second script conditions a five-second story
+  whatever the frame count.
+
+## Prompts
+
+H3 wants Context-IR, MiniMax's own format. Do not invent it and do not
+paraphrase it from examples:
+
+1. If the `h3-prompt-writing` skill is installed (MiniMax ships it in
+   https://github.com/MiniMax-AI/MiniMax-H3 under `skills/`), use it.
+2. Else read the guides on the model card:
+   https://huggingface.co/MiniMaxAI/MiniMax-H3/raw/main/docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md
+   for text- and frame-conditioned generation, and
+   https://huggingface.co/MiniMaxAI/MiniMax-H3/raw/main/docs/VIDEO_PROMPT_WRITING_GUIDE_ref_en.md
+   for reference-conditioned generation.
+3. Else run `templates/minimax/enhance-prompt` (or `-with-image`), whose
+   built-in enhancer writes the format from those guides. Its `idea` is
+   framed as `Task: T2VA. Duration: 5.17 seconds. Idea: ...`.
+
+Whichever route: repeat a speaker's voice description verbatim across shots,
+and when a reference picture should fix identity but not framing, say so in
+the prompt's reference analysis, or every shot inherits the portrait's
+composition.
+
+## Run and judge
+
+1. `validate_workflow` first - free, and it catches arguments the pipeline
+   does not accept.
+2. Quote the listing's `cost` (warm minutes on the card it was measured on;
+   a first load is longer) and get the user's go-ahead before `run_workflow`
+   with `acknowledged_cost=true`.
+3. `wait_for_job`, then `get_job` for the manifest. A cancelled H3 job runs
+   on to its next step boundary, minutes on this model.
+4. Look: `get_output_image` on a frame, the gallery `url` for the clip.
+   Check for the family's failure modes: a character that changes between
+   shots (reference the same portraits in every shot), a reference portrait
+   imposing its framing on every shot, a storyboard skipped, drift sharpening
+   into noise late in a chain.
+
+## Sources
+
+MiniMax-H3 model card and prompt guides (huggingface.co/MiniMaxAI/MiniMax-H3),
+the `h3-prompt-writing` skill (github.com/MiniMax-AI/MiniMax-H3), the diffusers
+MiniMax-H3 modular pipeline, the `lightx2v/Minimax-h3-Turbo` LoRA notes. Read
+2026-09-07; the audit is `docs/proposals/audits/2026-09-07-minimax-h3-audit.md`.
+````
+
+- [ ] **Step 4: The README pointer, and the xfail marker**
+
+In `workflows/templates/minimax/README.md`, after the paragraph that ends `([the audit](...minimax-h3-audit.md)).`, add one line:
+
+```markdown
+An agent driving this family from Claude Code has the `minimax-h3` skill of the
+[dw plugin](../../../plugins/dw/README.md), which chooses among these templates
+and points at the guides.
+```
+
+Remove the `xfail` marker Task 1 put on `test_there_are_skills`.
+
+- [ ] **Step 5: Run the tests**
+
+Run: `pytest tests/test_plugin_skills.py tests/test_catalog_structure.py -k "plugin or readme_link" -v`
+Expected: all PASS, including the README link test for the new relative link and every catalog name the skill quotes.
+
+- [ ] **Step 6: Ledger and commit**
+
+Append to the Part 4 ledger row: `Plugin task 2: the minimax-h3 skill - shape decision over the family's templates, the frame and canvas rules pinned by test to the diffusers modular pipeline, prompts deferred to MiniMax's h3-prompt-writing skill and the two guides, the run-and-judge loop; the README points at it.`
+
+```bash
+git add plugins/dw/skills/minimax-h3 workflows/templates/minimax/README.md tests/test_plugin_skills.py docs/proposals/agent-catalog-legibility.md
+git commit -m "dw plugin: the MiniMax H3 composition skill
+
+Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>"
+```
+
+---
+
+### Task 3: The LTX-2.5 skill, quoting the trained-caption spec
+
+**Recommended model:** opus.
+
+**Files:**
+- Create: `plugins/dw/skills/ltx-2.5/SKILL.md`
+- Modify: `workflows/templates/ltx2/README.md` (one line), `tests/test_plugin_skills.py` (the LTX pins and the quote-equality test)
+
+**Interfaces:**
+- Consumes: Task 1's helpers; Task 2's outline.
+- Produces: the convention for quoting a vendor constant: a fenced block under the heading `## The trained caption spec` that a test holds equal to the library string, whitespace aside.
+
+- [ ] **Step 1: Write the failing tests**
+
+Append to `tests/test_plugin_skills.py`:
+
+```python
+LTX_SKILL = os.path.join(PLUGIN_DIR, "skills", "ltx-2.5", "SKILL.md")
+
+
+def _fenced_block_after(text, heading):
+    """The first fenced code block after a markdown heading."""
+    start = text.index(heading)
+    open_fence = text.index("\n```", start)
+    open_end = text.index("\n", open_fence + 1)
+    close_fence = text.index("\n```", open_end)
+    return text[open_end + 1 : close_fence]
+
+
+class TestLtx25Skill:
+    """The LTX-2.5 skill's numbers come from the pipeline that enforces them,
+    and the one vendor text it quotes is the library's own constant."""
+
+    def test_the_schedule_is_the_library_s(self):
+        from diffusers.pipelines.ltx2.utils import DISTILLED_SIGMA_VALUES, STAGE_2_DISTILLED_SIGMA_VALUES
+
+        text = skill_text(LTX_SKILL)
+        assert len(DISTILLED_SIGMA_VALUES) == 8 and "eight" in text
+        assert len(STAGE_2_DISTILLED_SIGMA_VALUES) == 3 and "three" in text
+        assert str(STAGE_2_DISTILLED_SIGMA_VALUES[0]) in text
+
+    def test_the_size_and_frame_rules_are_the_pipeline_s(self):
+        import inspect
+
+        from diffusers.pipelines.ltx2 import pipeline_ltx2
+        from diffusers.pipelines.ltx2.utils import MAX_CONDITIONING_FPS
+
+        text = skill_text(LTX_SKILL)
+        source = inspect.getsource(pipeline_ltx2)
+        assert "height % 32 != 0 or width % 32 != 0" in source and "32" in text
+        assert "(num_frames - 1) // self.vae_temporal_compression_ratio + 1" in source
+        assert "8k + 1" in text or "8n + 1" in text
+        assert MAX_CONDITIONING_FPS == 60.0 and "60" in text
+
+    def test_the_quoted_caption_spec_is_the_library_constant(self):
+        """The one vendor text the plugin carries, tied to the library that
+        ships it so it cannot drift."""
+        from diffusers.pipelines.ltx2.utils import LTX2_5_T2V_DEFAULT_SYSTEM_PROMPT
+
+        quoted = _fenced_block_after(skill_text(LTX_SKILL), "## The trained caption spec")
+
+        assert " ".join(quoted.split()) == " ".join(LTX2_5_T2V_DEFAULT_SYSTEM_PROMPT.split())
+
+    def test_the_skill_starts_with_the_server(self):
+        text = skill_text(LTX_SKILL)
+        assert text.index("get_server_info") < text.index("templates/ltx2/")
+```
+
+- [ ] **Step 2: Run them to see them fail**
+
+Run: `pytest tests/test_plugin_skills.py -k Ltx25 -v`
+Expected: FAIL with `FileNotFoundError`.
+
+- [ ] **Step 3: Write the skill**
+
+Print the constant to paste verbatim into the fenced block (it is about 3.7 KB, so the rest of the skill has about 8 KB):
+
+```bash
+python -c "from diffusers.pipelines.ltx2.utils import LTX2_5_T2V_DEFAULT_SYSTEM_PROMPT as t; print(t)"
+```
+
+Create `plugins/dw/skills/ltx-2.5/SKILL.md`. Keep every name, number and URL exact; edit only for flow; stay under 12 KB (check with `wc -c`).
+
+````markdown
+---
+name: ltx-2.5
+description: Use when a dw MCP server is connected and the user wants LTX-2.5 video - a short clip with its own soundtrack, a clip that starts from a picture or runs between two, a sharper full-size render, a 2x upscale, or a clip extended or chained longer. Picks the template for the shape, states the schedule and size rules, quotes cost, and carries the trained caption spec the prompt must follow.
+---
+
+# LTX-2.5 on a dw server
+
+LTX-2.5 generates video and a soundtrack together, 24 fps, on a distilled
+schedule that is not a knob. Every template here fits a 24 GB card. This
+skill chooses the template and the arguments; the prompt is written to
+Lightricks' own caption spec, quoted below from diffusers.
+
+## Before anything
+
+1. `get_server_info`: the device (these templates quantize with SDNQ on CUDA)
+   and the workspace.
+2. `list_workflows(shape="shot")`: the family's templates by their current
+   names, with `summary`, `traits` and `cost`. Trust the listing over the
+   names quoted below.
+3. `get_workflow` on the one chosen.
+
+## Which shape is the request
+
+- **A clip from text**: `templates/ltx2/text-to-video` (960x544, 121 frames, 5 s).
+- **From a picture**: `templates/ltx2/image-to-video` (first frame; 481 frames
+  is 20 s in one pass, so reach for extend or chain only past that);
+  `templates/ltx2/keyframes` (first and last frames pinned);
+  `templates/ltx2/enhance-prompt` (a one-line idea plus a picture; the model's
+  own enhancer writes the caption).
+- **Sharper at full size**: `templates/ltx2/two-stage`, the three-move distilled
+  flow - eight sigmas at 768x448, a 2x latent upsample, then renoise and three
+  stage-two sigmas at 1536x896 with the audio latents carried through. The
+  upsample alone is soft; the refine pass is where the detail comes from.
+- **A 2x upscale of an existing clip**: `templates/ltx2/generative-upscale`,
+  the IC-LoRA re-rendering at twice the size and inventing detail. For a clean
+  low-resolution render, not compressed footage; fewer steps and lower guidance
+  keep it closer to the reference.
+- **Longer**: `templates/ltx2/extend-clip` continues a clip conditioned on all
+  of it; `templates/ltx2/chained-segments` re-runs per segment on the previous
+  last frame and stitches. Neither is a Lightricks recipe; both are dw's, and
+  a single 481-frame pass reaches 20 seconds before either is needed.
+
+If none fits, compose from `list_tasks` before authoring a new workflow, and
+read the `workflows` guide's authoring section first.
+
+## Hard rules
+
+- `num_frames` is `8k + 1` (121, 241, 481); `width` and `height` are multiples
+  of 32; 24 fps. Higher frame rates: condition at 60 at most, never 120 - RoPE
+  time is `frame / fps`, and the model is trained around 24, 25, 30 and 60.
+- The distilled transformer runs its eight trained sigmas (`DISTILLED_SIGMA_VALUES`)
+  with `guidance_scale` 1.0 and STG and modality guidance off. No
+  `num_inference_steps`. Those knobs mean something only against the dev
+  transformer, which no 24 GB template ships.
+- Stage two of the two-stage flow: renoise at 0.909375 (the first
+  `STAGE_2_DISTILLED_SIGMA_VALUES` entry) and run its three sigmas at full size.
+- An image condition is re-compressed at CRF 18 to match training and needs a
+  PIL image; a multi-frame video condition is not re-compressed.
+- Audio is generated in the first pass; nothing refines it afterwards, so
+  carry the audio latents (two-stage) or pair the soundtrack back
+  (`pair_audio`) on any step that works on frames alone.
+
+## Prompts
+
+A caption, not a tag list: one paragraph of roughly 150 to 220 words in the
+present progressive, opening on the action, stating for every shot a shot
+type, a camera motion (say static when there is none) and a viewpoint, with
+the soundscape interleaved with the action rather than appended, in plain
+observable words. For an image-conditioned clip describe only what changes
+from the image; restating it invites a scene cut. The stored prompts under
+`prompts/ltx2/` are written to this spec. For a one-line idea,
+`templates/ltx2/enhance-prompt` runs the model's own enhancer with the same
+spec. The spec itself, from `diffusers.pipelines.ltx2.utils.LTX2_5_T2V_DEFAULT_SYSTEM_PROMPT`
+(the image-to-video variant, `LTX2_5_I2V_DEFAULT_SYSTEM_PROMPT`, adds the
+describe-only-changes rule):
+
+## The trained caption spec
+
+```
+<paste LTX2_5_T2V_DEFAULT_SYSTEM_PROMPT here verbatim>
+```
+
+## Run and judge
+
+1. `validate_workflow` first.
+2. Quote the listing's `cost` and get the go-ahead; `run_workflow` with
+   `acknowledged_cost=true`.
+3. `wait_for_job`, then `get_job`. Writing a 121-frame 1536x896 clip takes
+   minutes after the last step ends; the job is not stuck.
+4. Look at a frame and the clip. Failure modes: a scene cut where the prompt
+   contradicted the image; softness when the refine pass was skipped; a
+   near-silent soundtrack when the caption gave the sound nothing to do.
+
+## Sources
+
+Lightricks/LTX-2.5-Diffusers model card, the `ltx-pipelines` docs and
+CHANGELOG (github.com/Lightricks/LTX-2), the diffusers LTX-2 pipelines and
+`utils.py`. Read 2026-09-07; the audit is
+`docs/proposals/audits/2026-09-07-ltx-2.5-audit.md`. Since 2026-08 Lightricks
+route production quality through their DFR pipeline, which diffusers ships
+and no template here uses yet.
+````
+
+- [ ] **Step 4: The README pointer**
+
+In `workflows/templates/ltx2/README.md`, after the paragraph that ends `([the audit](...ltx-2.5-audit.md)).`, add:
+
+```markdown
+An agent driving this family from Claude Code has the `ltx-2.5` skill of the
+[dw plugin](../../../plugins/dw/README.md), which chooses among these templates
+and carries the caption spec.
+```
+
+- [ ] **Step 5: Run the tests**
+
+Run: `pytest tests/test_plugin_skills.py tests/test_catalog_structure.py -k "plugin or readme_link" -v` and `wc -c plugins/dw/skills/*/SKILL.md`
+Expected: all PASS; both skills under 12288 bytes.
+
+- [ ] **Step 6: Ledger and commit**
+
+Append to the Part 4 ledger row: `Plugin task 3: the ltx-2.5 skill - shape decision, the schedule and size rules pinned to the LTX-2 pipeline, the trained caption spec quoted and held equal to the diffusers constant by test; the README points at it.`
+
+```bash
+git add plugins/dw/skills/ltx-2.5 workflows/templates/ltx2/README.md tests/test_plugin_skills.py docs/proposals/agent-catalog-legibility.md
+git commit -m "dw plugin: the LTX-2.5 composition skill, quoting the trained caption spec
+
+Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>"
+```
+
+---
+
+### Task 4: The cold drill, and the mould for the next family
+
+**Recommended model:** the controlling session, with the user - the drill needs a fresh Claude Code session on the user's machine with the plugin installed, which no subagent can start. Opus for reading the transcripts.
+
+**Files:**
+- Modify: `docs/proposals/agent-catalog-legibility.md` (the ledger row and a `### Plugin drill` section beside the cold-session probe)
+
+**Interfaces:**
+- Consumes: the merged branch on lem, restarted (`git checkout master && git pull`, restart `dw.serve --mcp`).
+
+- [ ] **Step 1: Install the plugin in the drill session**
+
+From the branch's checkout (before merge) the marketplace can be added by path: in a fresh Claude Code session started in the scratch directory the earlier drills used (`~/testing/scratch-1`, no `CLAUDE.md`, the dw MCP registered), run `/plugin marketplace add /Users/don/src/dkackman/diffusers-workflow` then `/plugin install dw@diffusers-workflow`, and confirm `/plugin` lists `minimax-h3` and `ltx-2.5`. After merge the GitHub form in the README works the same way.
+
+- [ ] **Step 2: Run the two prompts**
+
+In that session: `a short multi-shot video with cuts between the shots`, the prompt the earlier probe answered with three unrelated LTX-2 shots. Let it run to a finished piece if it asks for a go-ahead. Then a second fresh session without the plugin (`/plugin uninstall dw@diffusers-workflow`, new session), same prompt, as the control.
+
+- [ ] **Step 3: Read the evidence**
+
+The transcripts are under `~/.claude/projects/-Users-don-testing-scratch-1/*.jsonl`; lem's `~/.diffusers_helper/log/dw.log` and the uvicorn access log show the tool calls. Pass, for the plugin run: the `minimax-h3` skill fired; the agent called `get_server_info` and a shape-filtered `list_workflows`; it chose `templates/minimax/dialogue-short` or another cuts template with shared portraits; its prompts came from the guides or the `h3-prompt-writing` skill rather than an invented layout; it quoted cost before running; the finished piece holds the cast across shots. Record what the control did differently.
+
+- [ ] **Step 4: Ledger and the mould**
+
+Add a `### Plugin drill, <date>` section after `### Cold-session probe, 2026-09-07` in the proposal with: the two transcript paths, the tool calls in order for each run, which template each chose, whether the prompts came from vendor text, and the verdict. Append to the Part 4 ledger row: `Plugin task 4: cold drill <date> - <pass/fail and one sentence>; the mould for the next family is copy a skill, follow its six sections, add the family's rules to tests/test_plugin_skills.py, cite the vendor, and run this drill (.claude/skills/model-family-onboarding).`
+
+```bash
+git add docs/proposals/agent-catalog-legibility.md
+git commit -m "Plugin drill recorded; the mould for the next model family
+
+Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>"
+git push -u origin dw-plugin
+```
+
+Open the PR against `master` with the four commits' subjects as its bullets and end the body with `🤖 Generated with [Claude Code](https://claude.com/claude-code)`.
diff --git a/dw/tasks/audio_utils.py b/dw/tasks/audio_utils.py
index 08d22a2c..e82ccf3a 100644
--- a/dw/tasks/audio_utils.py
+++ b/dw/tasks/audio_utils.py
@@ -6,6 +6,7 @@
 """
 
 import io
+import os
 import logging
 from fractions import Fraction
 
@@ -192,9 +193,28 @@ def crossfade_concat(waveforms, sample_rate, crossfade_ms):
 def load_audio(location, base_dir=None):
     """Load an audio file from a local path or http(s) URL.
 
+    A video file loads too, and contributes the track muxed into it: the cut
+    an earlier run wrote is exactly what a scoring pass wants to mix under,
+    and refusing its extension made an agent extract the audio by hand
+    (2026-09-08). A video without an audio stream is an error, not silence.
+
     Returns:
         Tuple of a (channels, samples) float32 waveform and its sample rate
     """
+    from ..security import ALLOWED_VIDEO_EXTENSIONS
+
+    extension = os.path.splitext(location.split("?", 1)[0])[1].lower()
+    if extension in ALLOWED_VIDEO_EXTENSIONS:
+        from .video_utils import load_audio_video
+
+        video = load_audio_video(location, base_dir=base_dir)
+        if video.audio is None:
+            raise ValueError(
+                f"{location} carries no audio track - a video's soundtrack is "
+                "what an audio task takes from it"
+            )
+        return as_channels_samples(video.audio), video.sample_rate
+
     if location.startswith(("http://", "https://")):
         import requests
 
@@ -248,7 +268,8 @@ def slice_audio(
     rather than failing.
 
     Args:
-        audio: Path or URL of an audio file, a video generated with a
+        audio: Path or URL of an audio file (or of a video file, whose
+            soundtrack is taken), a video generated with a
             soundtrack (which brings its sample rate along), or a waveform
             (which needs sample_rate alongside it)
         sample_rate: Sample rate of a waveform passed directly; given for a
@@ -305,7 +326,8 @@ def resample_audio(audio, target_sample_rate, sample_rate=None):
     the conversion.
 
     Args:
-        audio: Path or URL of an audio file, a video generated with a
+        audio: Path or URL of an audio file (or of a video file, whose
+            soundtrack is taken), a video generated with a
             soundtrack (which brings its sample rate along), or a waveform
             (which needs sample_rate alongside it)
         target_sample_rate: Rate to convert to
@@ -351,8 +373,8 @@ def crossfade_audio(audios, crossfade_ms=75, sample_rate=None):
     shorter than the plain sum by one window per seam.
 
     Args:
-        audios: The tracks to join, in order - waveforms, audio file paths, or
-            videos generated with a soundtrack
+        audios: The tracks to join, in order - waveforms, audio or video file
+            paths, or videos generated with a soundtrack
         crossfade_ms: Length of each crossfade
         sample_rate: Sample rate of the waveforms. Required unless every track
             brings its own; given here it wins
@@ -397,7 +419,7 @@ def mix_audio(audios, gains=None, sample_rate=None):
     should sound - follow it with normalize_audio to bring the peak back down.
 
     Args:
-        audios: The tracks to layer - waveforms, audio file paths, or videos
+        audios: The tracks to layer - waveforms, audio or video file paths, or videos
             generated with a soundtrack
         gains: One plain multiplier per track, in the same order - not decibels.
             Defaults to unity on every track
@@ -487,7 +509,8 @@ def fade_audio(audio, fade_in_ms=0, fade_out_ms=0, sample_rate=None):
     not a volume knob.
 
     Args:
-        audio: Path or URL of an audio file, a video generated with a
+        audio: Path or URL of an audio file (or of a video file, whose
+            soundtrack is taken), a video generated with a
             soundtrack, or a waveform (which needs sample_rate alongside it)
         fade_in_ms: Length of the fade in, from the head of the track
         fade_out_ms: Length of the fade out, to the tail of the track
@@ -520,7 +543,8 @@ def normalize_audio(audio, peak_dbfs=-1.0, sample_rate=None):
     nothing but the gain, so the dynamics survive.
 
     Args:
-        audio: Path or URL of an audio file, a video generated with a
+        audio: Path or URL of an audio file (or of a video file, whose
+            soundtrack is taken), a video generated with a
             soundtrack, or a waveform (which needs sample_rate alongside it)
         peak_dbfs: The level the loudest sample is moved to, in dB below full
             scale. 0 is full scale; -1 leaves a little headroom
diff --git a/dw/tasks/pair_audio.py b/dw/tasks/pair_audio.py
index e5998be4..ec21e3cd 100644
--- a/dw/tasks/pair_audio.py
+++ b/dw/tasks/pair_audio.py
@@ -15,15 +15,23 @@
 logger = logging.getLogger("dw")
 
 
+class _Loaded:
+    """A waveform read from a file, shaped like the artifact pair_audio expects."""
+
+    def __init__(self, audio, sample_rate):
+        self.audio = audio
+        self.sample_rate = sample_rate
+
+
 def pair_audio(video, audio, sample_rate=None):
     """Pair a video's frames with an audio track.
 
     Args:
         video: The frames - a frame list, a frame array or tensor, or an
             AudioVideo whose own soundtrack is replaced by this one
-        audio: The soundtrack - a waveform, or an AudioVideo (or any object
-            carrying '.audio') to take it from, which brings its sample rate
-            along with it
+        audio: The soundtrack - a waveform, an AudioVideo (or any object
+            carrying '.audio') to take it from, or the path or URL of an
+            audio or video file; the last two bring their sample rate along
         sample_rate: Sample rate of the waveform. Required unless `audio`
             carries one; given here it wins, for a track whose rate was
             reported wrong
@@ -35,6 +43,11 @@ def pair_audio(video, audio, sample_rate=None):
         ValueError: If no waveform was given, or if no sample rate can be
             established for the one that was
     """
+    if isinstance(audio, str):
+        # A file an earlier run wrote - a score, or a cut whose track is wanted
+        from .audio_utils import load_audio
+
+        audio = _Loaded(*load_audio(audio))
     waveform = getattr(audio, "audio", audio)
     if waveform is None:
         raise ValueError(
diff --git a/dw_mcp/media.py b/dw_mcp/media.py
index a8eb48ca..6f3067a1 100644
--- a/dw_mcp/media.py
+++ b/dw_mcp/media.py
@@ -171,12 +171,26 @@ def download_output(client, name, destination=None, overwrite=False):
         )
 
     parent = os.path.dirname(destination)
-    if parent:
-        os.makedirs(parent, exist_ok=True)
-
-    content_type, bytes_written = client.stream_to_file(
-        api_path("outputs", name), destination
-    )
+    try:
+        if parent:
+            os.makedirs(parent, exist_ok=True)
+        content_type, bytes_written = client.stream_to_file(
+            api_path("outputs", name), destination
+        )
+    except OSError as e:
+        # A client-side path handed to a `dw.serve --mcp` endpoint lands
+        # here: the write happens on the server, so a permission or
+        # missing-volume error is the surest sign the agent is on another
+        # machine. Say so, rather than letting the OSError surface as an
+        # anonymous "Error executing tool".
+        raise DwApiError(
+            f"Could not write {destination} on the machine running the MCP "
+            f"server ({e.strerror or e}). This tool saves on that machine - "
+            "over a dw.serve --mcp endpoint that is the GPU box, not where "
+            "you are. To see the file from here use the url list_gallery "
+            "reports, get_output_image / get_output_text for inline content, "
+            "or keep_output to make it an asset for a later workflow."
+        ) from e
 
     return {
         "name": name,
diff --git a/plugins/dw/.claude-plugin/plugin.json b/plugins/dw/.claude-plugin/plugin.json
new file mode 100644
index 00000000..e29e5219
--- /dev/null
+++ b/plugins/dw/.claude-plugin/plugin.json
@@ -0,0 +1,18 @@
+{
+    "name": "dw",
+    "description": "Compose MiniMax H3 video, MiniMax Music 3 and LTX-2.5 workflows over a dw MCP server: which template fits which shape, the hard rules, cost, and how to judge the output. Prompt format comes from the vendors' own guides.",
+    "version": "0.4.0-beta.1",
+    "author": {
+        "name": "Don Kackman"
+    },
+    "repository": "https://github.com/dkackman/diffusers-workflow",
+    "keywords": [
+        "diffusers",
+        "video",
+        "music",
+        "minimax-h3",
+        "minimax-music3",
+        "ltx-2.5",
+        "mcp"
+    ]
+}
diff --git a/plugins/dw/README.md b/plugins/dw/README.md
new file mode 100644
index 00000000..d55f8fb2
--- /dev/null
+++ b/plugins/dw/README.md
@@ -0,0 +1,28 @@
+# dw plugin
+
+Skills for Claude Code that teach an agent to compose a model family's
+workflows on a [diffusers-workflow](https://github.com/dkackman/diffusers-workflow)
+server. They assume the `dw` MCP server is already registered
+(see the repo README, "Drive it from Claude Code"); a skill's first move is
+`get_server_info`.
+
+Install once, from Claude Code:
+
+```
+/plugin marketplace add dkackman/diffusers-workflow
+/plugin install dw@diffusers-workflow
+```
+
+| Skill | Teaches |
+| ----- | ------- |
+| `minimax-h3` | MiniMax H3 video with audio: one take, a longer take by chain, a piece with cuts, identity and voice references, music; the frame and canvas rules; what a run costs. Prompts come from MiniMax's own `h3-prompt-writing` skill or the guides on the model card. |
+| `minimax-music3` | MiniMax Music 3: a song, an instrumental, a score under a film, the soundtrack a music video is cut to; the ceiling semantics, the tag vocabulary, the caps and the 44.1 kHz output. Captions come from MiniMax's own `music-caption-rewriter` skill. |
+| `ltx-2.5` | LTX-2.5 video with a soundtrack: a single clip, first-frame and keyframe conditioning, the three-move two-stage flow, the IC-LoRA upscale, extend and chain; the distilled schedule and the frame and size rules. Prompts follow the trained-caption spec that ships inside diffusers. |
+
+Each skill quotes catalog names and numeric rules that `tests/test_plugin_skills.py`
+holds to the catalog and to the diffusers module that enforces them. The
+plugin's version is the engine's; the release script bumps both.
+
+Adding a family: copy a skill, follow its outline, add the family's rules to
+the test, cite the vendor. The repo's `model-family-onboarding` skill
+(`.claude/skills/`) is the full lifecycle.
diff --git a/plugins/dw/skills/ltx-2.5/SKILL.md b/plugins/dw/skills/ltx-2.5/SKILL.md
new file mode 100644
index 00000000..6f17736d
--- /dev/null
+++ b/plugins/dw/skills/ltx-2.5/SKILL.md
@@ -0,0 +1,145 @@
+---
+name: ltx-2.5
+description: Use when a dw MCP server is connected and the user wants LTX-2.5 video - a short clip with its own soundtrack, a clip that starts from a picture or runs between two, a sharper full-size render, a 2x upscale, or a clip extended or chained longer. Picks the template for the shape, states the schedule and size rules, quotes cost, and carries the trained caption spec the prompt must follow.
+---
+
+# LTX-2.5 on a dw server
+
+LTX-2.5 generates video and a soundtrack together, 24 fps, on a distilled
+schedule that is not a knob. Every template here fits a 24 GB card. This
+skill chooses the template and the arguments; the prompt is written to
+Lightricks' own caption spec, quoted below from diffusers.
+
+## Before anything
+
+1. `get_server_info`: the device and the workspace. These templates quantize
+   with SDNQ on CUDA; on an `mps` or `cpu` server say so and stop, because
+   none of them fit there.
+2. `list_workflows(shape="shot")`: all eight of the family's templates carry
+   that shape. Take their current names, `summary`, `traits` and `cost` from
+   the listing and trust it over the names quoted below.
+3. `get_workflow` on the one chosen, for its variables and their defaults.
+
+## Which shape is the request
+
+- **A clip from text**: `templates/ltx2/text-to-video` (960x544, 121 frames, 5 s).
+- **From a picture**: `templates/ltx2/image-to-video` (first frame; 481 frames
+  is 20 s in one pass, so reach for extend or chain only past that);
+  `templates/ltx2/keyframes` (first and last frames pinned);
+  `templates/ltx2/enhance-prompt` (a one-line idea plus a picture; the model's
+  own enhancer writes the caption).
+- **Sharper at full size**: `templates/ltx2/two-stage`, the three-move distilled
+  flow - eight sigmas at 768x448, a 2x latent upsample, then renoise and three
+  stage-two sigmas at 1536x896 with the audio latents carried through. The
+  upsample alone is soft; the refine pass is where the detail comes from.
+- **A generative 2x render**: `templates/ltx2/generative-upscale` draws its own
+  low-resolution pass and has the IC-LoRA re-render it at twice the size,
+  inventing detail rather than interpolating. `base_width` and `base_height`
+  are that first render's size, `width` and `height` the doubled target; both
+  passes run the same eight distilled sigmas at guidance 1.0. Nothing in this
+  family takes a user-supplied video, so a user's own footage is not a fit for
+  any of these templates.
+- **Longer**: `templates/ltx2/extend-clip` generates an opening and then
+  continues it conditioned on the whole opening, not on a single frame - a
+  supplied clip is not an input here either; `templates/ltx2/chained-segments`
+  re-runs per segment on the previous last frame and stitches. Neither is a Lightricks recipe; both are dw's, and
+  a single 481-frame pass reaches 20 seconds before either is needed.
+
+If none fits, compose from `list_tasks` before authoring a new workflow, and
+read the `workflows` guide's authoring section first.
+
+## Hard rules
+
+- `num_frames` is `8k + 1` (121, 241, 481); `width` and `height` are in multiples of 32.
+  The templates generate at 24 fps. RoPE time is `frame / fps` and the model is
+  trained around 24, 25, 30 and 60, so for a higher-fps request generate at 24
+  (or condition at 60 at most - `MAX_CONDITIONING_FPS`, never 120) and let
+  playback carry the rate. The temporal-upscaling path that renders 48 and 96
+  fps belongs to the DFR pipelines, which no template here uses yet.
+- The distilled transformer runs its eight trained sigmas (`DISTILLED_SIGMA_VALUES`)
+  with `guidance_scale` 1.0 and STG and modality guidance off. No
+  `num_inference_steps`. Those knobs mean something only against the dev
+  transformer, which no 24 GB template ships.
+- Stage two of the two-stage flow: renoise at 0.909375 (the first
+  `STAGE_2_DISTILLED_SIGMA_VALUES` entry) and run its three sigmas at full size.
+- An image condition is re-compressed at CRF 18 to match training and needs a
+  PIL image; a multi-frame video condition is not re-compressed.
+- Audio is generated in the first pass; nothing refines it afterwards, so
+  carry the audio latents (two-stage) or pair the soundtrack back
+  (`pair_audio`) on any step that works on frames alone.
+
+## Prompts
+
+A caption, not a tag list: one paragraph of roughly 150 to 220 words in the
+present progressive, opening on the action, stating for every shot a shot
+type, a camera motion (say static when there is none) and a viewpoint, with
+the soundscape interleaved with the action rather than appended, in plain
+observable words. For an image-conditioned clip describe only what changes
+from the image; restating it invites a scene cut. The stored prompts under
+`prompts/ltx2/` are written to this spec. For a one-line idea,
+`templates/ltx2/enhance-prompt` runs the model's own enhancer with the same
+spec. The spec itself, from `diffusers.pipelines.ltx2.utils.LTX2_5_T2V_DEFAULT_SYSTEM_PROMPT`
+(the image-to-video variant, `LTX2_5_I2V_DEFAULT_SYSTEM_PROMPT`, adds the
+describe-only-changes rule):
+
+## The trained caption spec
+
+```
+You are given a user's short text-to-video request. Write a single, highly detailed audio-visual caption describing the video that best fulfills that request, in the EXACT style of the training captions used for this video model. The generated video is scored against the user's ORIGINAL request, so preserve every element the user stated; expand faithfully into the full caption style without contradicting or dropping anything they asked for.
+
+Match this captioning style precisely:
+
+1. Begin immediately with the action or visual detail. Do NOT use "The scene opens…", "We see…", "There is…".
+
+2. Objective, observable description only. Do not infer emotions or intentions — describe what is visible and audible (e.g. not "he looks sad" but "his eyebrows angle downward and his lips are pressed together").
+
+3. Full visual detail: environment (materials, textures, lighting, colors), character appearance (clothing, posture, facial details), and the spatial positioning of all elements. When a human appears, identify them specifically (gendered terms when clearly implied; differentiate multiple people consistently) and describe visible physical attributes — apparent gender presentation, skin tone, estimated age group, hair color/length/style, build, clothing and accessories. Do not infer ethnicity, nationality, religion, or culture.
+
+4. Precise motion and cinematic description. For every shot you MUST include, woven naturally into the prose (never as tags or labels):
+   - Shot type (exactly one: extreme wide shot / wide shot / medium shot / medium close-up / close-up / extreme close-up)
+   - Camera motion (always stated; if none, explicitly say the camera remains static). Camera movement is expected and good — match the user if they specified it, otherwise choose the treatment that best presents the requested scene.
+   - Camera viewpoint relative to subject (front-facing / back-facing / side view / over-the-shoulder / top-down / low-angle / high-angle).
+   Express these as flowing prose: "a medium shot frames…, captured from a front-facing angle as the camera slowly pans…". Never as "medium shot, static camera —".
+
+5. Complete soundscape, integrated naturally: any dialogue (quote it exactly, in the original language), tone of voice, background music (type, mood, volume changes), and environmental sounds (footsteps, wind, traffic, animals). If the request implies sound, describe it plausibly.
+
+6. Strict chronological, real-time flow using transitions like "Initially…", "A moment later…", "Simultaneously…". Keep every stated action in motion.
+
+7. One single continuous paragraph. No bullet points, no section headers, no labels like "Audio:" or "Visual:". Exhaustive and lossless — include background elements, subtle movements, lighting, secondary sounds — detailed enough to reconstruct the scene. Aim for a rich, complete paragraph (roughly 150–220 words).
+
+If the user wrote in another language, produce the English caption of the same content. Output ONLY the caption text — no JSON, no preamble.
+
+AESTHETIC QUALITY (in addition to the above, without breaking the objective caption style): render the described scene with strong visual production value — cinematic, film-grade color and contrast, beautiful natural lighting, crisp fine detail and texture, pleasing composition and depth. Weave these quality descriptors naturally into the same observable prose (e.g. "warm cinematic lighting", "richly saturated film-grade color", "crisp high-resolution detail") — describe how the exact requested scene LOOKS at its most visually striking, never adding new objects or actions. Keep everything else (framing triple, soundscape, chronological single paragraph, faithfulness) exactly as specified.
+
+```
+
+## Run and judge
+
+1. `validate_workflow` first - free, and it catches arguments the pipeline
+   does not accept.
+2. Quote the listing's `cost`. Only `templates/ltx2/text-to-video` and
+   `templates/ltx2/two-stage` declare one; for the other six say so and give
+   the shape of the spend instead - a 121-frame clip at 960x544 is under a
+   minute warm on a 24 GB card, the two-stage flow about eight, and extend
+   and chain multiply by their passes. Either way get the user's go-ahead
+   before `run_workflow` with `acknowledged_cost=true`.
+3. `wait_for_job`, then `get_job` for the manifest. Writing a 121-frame
+   1536x896 clip takes minutes after the last step ends; the job is not stuck.
+4. You cannot watch a video: no tool returns a frame from one, and this family
+   has no image steps for `get_output_image` to read. Hand the user the gallery
+   `url` (`list_gallery`, or the manifest's file name) and ask them to look, and
+   check what you can yourself - `get_job` for the manifest and its warnings,
+   `get_gallery_metadata` for duration, size and whether an audio stream is
+   present. Ask the user to look for the family's failure modes: a scene cut
+   where the prompt contradicted the image; softness where the refine pass was
+   skipped; a near-silent soundtrack where the caption gave the sound nothing
+   to do.
+
+## Sources
+
+Lightricks/LTX-2.5-Diffusers model card, the `ltx-pipelines` docs and
+CHANGELOG (github.com/Lightricks/LTX-2), the diffusers LTX-2 pipelines and
+`utils.py`. Read 2026-09-07; the audit is
+`docs/proposals/audits/2026-09-07-ltx-2.5-audit.md`. Since 2026-08 Lightricks
+route production quality through their DFR pipeline, which diffusers ships
+and no template here uses yet.
diff --git a/plugins/dw/skills/minimax-h3/SKILL.md b/plugins/dw/skills/minimax-h3/SKILL.md
new file mode 100644
index 00000000..16e564b6
--- /dev/null
+++ b/plugins/dw/skills/minimax-h3/SKILL.md
@@ -0,0 +1,159 @@
+---
+name: minimax-h3
+description: Use when a dw MCP server is connected and the user wants MiniMax H3 video - a clip with its own sound, dialogue or a speaking character, a music video, a multi-shot short with cuts, a longer take, or a subject or voice kept consistent from a reference. Picks the template for the shape, states the rules that bite, quotes cost, and points at MiniMax's own prompt guides.
+---
+
+# MiniMax H3 on a dw server
+
+H3 generates video and audio together: speech with lip sync, ambient sound,
+score. Every template here fits a 24 GB card. This skill chooses the template
+and the arguments; the prompt format is MiniMax's and comes from their text,
+not from here.
+
+## Before anything
+
+1. `get_server_info`: the device (H3 templates are CUDA; the quantized
+   configurations do not run on mps) and which workspace this session is in.
+2. `list_workflows(shape="shot")`, `list_workflows(shape="sequence")` and
+   `list_workflows(shape="audio")`: the family's templates by their current
+   names, with `summary`, `traits` and `cost`. Trust the listing over the
+   names quoted below.
+3. `get_workflow` on the one chosen, for its variables and their defaults.
+
+## Which shape is the request
+
+- **One clip, up to 14.4 seconds, from text**: `templates/minimax/video-with-audio`.
+  From a one-line idea: `templates/minimax/enhance-prompt` writes the prompt
+  with the built-in Context-IR enhancer first.
+- **Pinned to a picture**: first frame `templates/minimax/image-to-video`;
+  first and last `templates/minimax/first-and-last-frame`; last only
+  `templates/minimax/last-frame-only`; a one-line idea plus a picture
+  `templates/minimax/enhance-prompt-with-image`.
+- **A subject that must look the same**: `templates/minimax/reference-to-video`
+  (an image fixes appearance, an audio clip fixes voice);
+  `templates/minimax/composable-references` adds a video reference for framing
+  and camera; `templates/minimax/generated-subject-reference` draws the subject
+  with Z-Image first and references it in the same workflow;
+  `templates/minimax/voice-timbre-reference` fixes a voice from a Bark-spoken line.
+- **Several boards in one generation, one unbroken score**:
+  `templates/minimax/storyboard` - H3 cuts between the boards inside a single
+  generation, which no concat of separate clips can match for continuous audio.
+- **Longer than 14.4 seconds**: decide first whether the seam is a cut or a
+  continuation. Chain when the same action or line of speech has to cross the
+  seam; cut when the scene changes, and treat each cut as its own generation.
+  Six distinct scenes are a cuts piece, not a chain.
+- **Longer than 14.4 seconds as one take**: a chain. `templates/minimax/chained-segments`
+  (last-frame continuity), `templates/minimax/chain-video-continuity` (the
+  previous segment's tail rides along as a video reference - motion, camera and
+  voice carry across the seam), `templates/minimax/chain-matched-to-audio` (a
+  supplied track sets the length and is muxed back seamless),
+  `templates/minimax/chain-matched-and-aligned` (all of it, per-segment prompts).
+  Drift compounds per seam: reference the subject picture in every segment,
+  prefer `last_segment` continuity, and use the longest segments memory allows.
+- **A piece with cuts**: fresh shots from shared portraits, then a concat.
+  `templates/minimax/dialogue-short` (Z-Image draws the cast, one loaded model
+  per shot, `concat_videos` splices) and `templates/minimax/music-video`
+  (shots cut to a generated song, lip-synced slices). A cut erases drift; the
+  last shot is as clean as the first. Write shots, not takes. Each shot
+  generates its own audio, so write `non_diegetic_music: N/A` in every shot
+  and lay one score under the concat afterwards: `templates/minimax/music`
+  writes the track and `templates/assemble-and-score` shows the `pair_audio`
+  step that mixes it under the world sound (it takes three shots; for more,
+  author the concat and score steps the same way). A character who speaks in
+  several shots keeps one voice by passing the same clip as an audio
+  reference in each (the `voice-timbre-reference` pattern); a repeated voice
+  description alone drifts.
+- **Music alone**: `templates/minimax/music` (Music3); the `minimax-music3` skill.
+
+If none fits, compose from `list_tasks` before authoring a new workflow, and
+read the `workflows` guide's authoring section first.
+
+## Hard rules
+
+- `num_frames` is `17n + 5`, from 124 to 345, at a fixed 24 fps: 5.17 to 14.4
+  seconds in one clip. Most templates default to 124 for fast iteration (storyboard uses 192); `num_frames=345`
+  is the full length and fits the same 24 GB configuration. The 5-second floor
+  is diffusers'; the model card says 4.
+- Canvas: a 768-pixel short edge, at most 768x1344 pixels, dimensions in multiples of 32,
+  aspect from 1:4 to 4:1. Output audio is 32 kHz stereo.
+- The text- and frame-conditioned templates render at 960x544 with the 544p
+  turbo LoRA in nine steps, and those three go together - change one, change
+  all three. Six templates that condition on references alone
+  (`reference-to-video`, `composable-references`, `voice-timbre-reference`,
+  `generated-subject-reference`, `chain-matched-to-audio`,
+  `chain-video-continuity`) carry no LoRA and run 20 steps, because the turbo
+  LoRA is distilled against the base transformer and they load the reference
+  one; such a run is about twice the time of a turbo one at the same length.
+  `storyboard`, `dialogue-short`, `music-video` and
+  `chain-matched-and-aligned` pass references *and* keep the turbo LoRA at
+  nine steps; say nine for those, not 20.
+- Nothing carries between generations except what is passed as a reference:
+  no latent memory and no extension mode, in the checkpoint, the hosted API or
+  diffusers. Identity rides on a picture, voice on an audio clip, motion and
+  camera on a video tail (what a chain passes forward), and a score across
+  cuts is laid under the concat afterwards.
+- H3 is guidance-distilled: no `guidance_scale`, no negative prompt. Say what is
+  there, never what is not.
+- Ref2VA limits: at most 9 images, 3 videos, 3 audio clips, 12 files; audio can
+  never be the only reference. References are labelled in the order passed.
+- Music3 reads `audio_duration` as a ceiling, not a target: ask for more than
+  the song needs and trim with `templates/audio-trim-fade`. The `minimax-music3`
+  skill has the rest of that family.
+- Write the prompt for the length being generated: shot timestamps should span
+  the duration, or a five-second script conditions a five-second story
+  whatever the frame count.
+
+## Prompts
+
+H3 wants Context-IR, MiniMax's own format. Do not invent it and do not
+paraphrase it from examples:
+
+1. If the `h3-prompt-writing` skill is installed (MiniMax ships it in
+   https://github.com/MiniMax-AI/MiniMax-H3 under `skills/`), use it.
+2. Else read the guides on the model card:
+   https://huggingface.co/MiniMaxAI/MiniMax-H3/raw/main/docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md
+   for text- and frame-conditioned generation, and
+   https://huggingface.co/MiniMaxAI/MiniMax-H3/raw/main/docs/VIDEO_PROMPT_WRITING_GUIDE_ref_en.md
+   for reference-conditioned generation.
+3. Else run `templates/minimax/enhance-prompt` (or `-with-image`), whose
+   built-in enhancer writes the format from those guides. Its `idea` is
+   framed as `Task: T2VA. Duration: 5.17 seconds. Idea: ...`.
+
+Whichever route: repeat a speaker's voice description verbatim across shots,
+and when a reference picture should fix identity but not framing, say so in
+the prompt itself - in a reference-conditioned request, in the lines that
+define the subject and state what each reference keeps - or every shot
+inherits the portrait's composition.
+
+## Run and judge
+
+1. `validate_workflow` first - free, and it catches arguments the pipeline
+   does not accept.
+2. Quote the listing's `cost` (warm minutes on the card it was measured on;
+   a first load is longer). When the listing declares none, say so and give the
+   shape of the spend instead: a 124-frame turbo clip is a few minutes on a
+   24 GB card, the full 345 frames about three times that, a
+   reference-conditioned clip about twice a turbo one, and a chain multiplies
+   by its segment count. Get the user's go-ahead before `run_workflow` with
+   `acknowledged_cost=true`.
+3. `wait_for_job`, then `get_job` for the manifest. A cancelled H3 job runs
+   on to its next step boundary, minutes on this model.
+4. You cannot watch a video: no tool returns a frame from one. Hand the user
+   the gallery `url` (`list_gallery`, or the manifest's file name) and ask them
+   to look, and check what you can yourself - `get_job` for the manifest and
+   its warnings, `get_gallery_metadata` for duration, size and whether an audio
+   stream is present. `get_output_image` works only on image steps, which in
+   this family are the Z-Image portraits and boards of
+   `templates/minimax/dialogue-short`, `templates/minimax/storyboard`,
+   `templates/minimax/generated-subject-reference` and
+   `templates/minimax/music-video`. Ask the user to look for the family's
+   failure modes: a character that changes between shots (reference the same
+   portraits in every shot), a reference portrait imposing its framing on every
+   shot, a storyboard skipped, drift sharpening into noise late in a chain.
+
+## Sources
+
+MiniMax-H3 model card and prompt guides (huggingface.co/MiniMaxAI/MiniMax-H3),
+the `h3-prompt-writing` skill (github.com/MiniMax-AI/MiniMax-H3), the diffusers
+MiniMax-H3 modular pipeline, the `lightx2v/Minimax-h3-Turbo` LoRA notes. Read
+2026-09-07; the audit is `docs/proposals/audits/2026-09-07-minimax-h3-audit.md`.
diff --git a/plugins/dw/skills/minimax-music3/SKILL.md b/plugins/dw/skills/minimax-music3/SKILL.md
new file mode 100644
index 00000000..f06db199
--- /dev/null
+++ b/plugins/dw/skills/minimax-music3/SKILL.md
@@ -0,0 +1,133 @@
+---
+name: minimax-music3
+description: Use when a dw MCP server is connected and the user wants music from MiniMax Music 3 - a song with sung lyrics, an instrumental, a score to lay under a film or a cuts piece, or the soundtrack a music video is cut to. Picks the template for the shape, states the ceiling and tag rules that bite, quotes cost, and points at MiniMax's own caption skill for the prompt.
+---
+
+# MiniMax Music 3 on a dw server
+
+Music 3 writes a whole track from two strings: a caption that says how it
+sounds and a lyrics body that says what happens section by section. It runs on
+a dw server as one modular pipeline step, and the catalog already holds the
+shapes; do not author a new workflow until the shape decision below fails.
+
+## Before anything
+
+1. `get_server_info`: the device. Music 3 is a CUDA model; on an `mps` or `cpu`
+   server stop and say so.
+2. `list_workflows(shape="audio")`, and `shape="sequence"` for the music video.
+   Trust the names quoted below only after the listing confirms them.
+3. `get_workflow` on the one chosen, for its variables and their defaults.
+
+## Which shape is the request
+
+- **A song**: `templates/minimax/music`. Lyrics with section tags plus a caption
+  that names the vocal gender and timbre. This is the default shape.
+- **An instrumental**: the same template with a tag-only lyrics body
+  (`[intro]`, `[instrumental]`, `[solo]`, `[outro]`, one per line) and a caption
+  that says instrumental and names the instrument carrying the lead. The
+  diffusers pipeline has no instrumental flag and rejects empty lyrics, so the
+  tags are how it is asked for.
+- **A score under a film or a cuts piece**: an instrumental generated to a
+  ceiling comfortably longer than the cut, trimmed and faded with
+  `templates/audio-trim-fade`, then mixed under the picture the way
+  `templates/assemble-and-score` does with `pair_audio`. Each H3 shot should
+  have written `non_diegetic_music: N/A` so the two scores do not fight.
+- **A music video**: `templates/minimax/music-video`. The song is written
+  first, `slice_audio` deals frame-exact pieces to lip-synced H3 shots, and
+  `pair_audio` lays the unbroken track back over the edit. The ceiling must
+  exceed the total sliced length with real margin, not by a fraction of a
+  second, since the model may stop early.
+
+If none fits, compose from `list_tasks` (`slice_audio`, `fade_audio`,
+`pair_audio`, `mix_audio`, `concat_videos`) before authoring, and read the
+`workflows` guide's authoring section first.
+
+## Hard rules
+
+- `audio_duration` is a ceiling, not a target: the language model stops when
+  the song ends, and generation is cut at the ceiling if it has not. Default
+  60 seconds. Ask for more than the piece needs and trim; the run's time
+  follows the length actually generated, so the margin is free.
+- The engine caps a track at 9000 frames at 25 frames per second, 360 seconds.
+  MiniMax supports five minutes; stay at or under 300.
+- The caption is capped at 5,000 tokens and a longer one is an error, not a
+  truncation. A 250-450 word caption is nowhere near it; a pasted pile of
+  example captions is.
+- Lyrics: a section tag stands alone on its line and is lower-cased on the way
+  in; any words on the same line as a leading tag are dropped. The vocabulary
+  on the model card is `[Intro]`, `[Verse]`, `[Pre-Chorus]`, `[Chorus]`,
+  `[Post-Chorus]`, `[Bridge]`, `[Instrumental]`, `[Solo]`, `[Outro]`. A tag can
+  carry a local direction for its section; the caption's arrangement text is
+  where that direction is spelled out.
+- Name the vocal gender and timbre in the caption, or the model may drift
+  instrumental. An instrumental says so in the same place.
+- Output is 44.1 kHz stereo, the vocoder's native rate. The model card's
+  32 kHz is what MiniMax's reference server resamples to; the templates' result
+  declares `sample_rate: 44100` because the modular output carries no rate of
+  its own.
+- The flow stage denoises in 200-frame windows at 30 steps each, with
+  classifier-free guidance fixed at 1.7 on the pipeline's guider. Neither is a
+  call argument; leave them unless a template exposes `num_inference_steps`.
+- Memory: about 22 GB with the templates' auto CPU offload; the language model
+  can be group-offloaded to run in 8 GB. A workflow that runs Music 3 before
+  H3 frees it with `release_pipeline` first, as `music-video` does.
+- Section tags and the caption are generative control, not guarantees: the
+  vendor says tempo, key, structure and lyrics may not match every detail.
+  Iterate at 30-60 seconds before asking for a long track.
+
+## Prompts
+
+The caption format is MiniMax's own. Do not invent it and do not paraphrase it
+from the templates' examples:
+
+1. If the `music-caption-rewriter` skill is installed (MiniMax ships it in
+   https://github.com/MiniMax-AI/MiniMax-Music3 under `skills/`, with a genre
+   router, 18 family indexes and 1,000 example captions), use it.
+2. Else read its `SKILL.md` and `references/genre-router.md` at that path. The
+   contract is three headings in order - Global Metadata, Vocal Details,
+   Arrangement - in 250-450 English words, with no title, no reasoning, and
+   no lyric line copied into the caption. The pipeline strips markdown
+   headings and emphasis on the way in, so the vendor's caption pastes
+   straight into `prompt`.
+3. The concise one-paragraph form the templates' stored prompts use (genre,
+   BPM, key, emotional progression, listening scenario, production profile,
+   vocals, arrangement) is the model card's own example and works; the
+   three-heading form is for precise control.
+
+Lyrics carry the structure: tags for the sections, the words to sing under
+them, and nothing else. The tag list is the model card's "Fine-Grained Music
+Control" section.
+
+## Run and judge
+
+1. `validate_workflow` first - free, and it catches arguments the pipeline
+   does not accept.
+2. Quote the listing's `cost` (warm minutes on the card it was measured on;
+   a first load is longer). When the listing declares none, say so and give
+   the shape of the spend: the autoregressive stage runs at 25 frames per
+   second of audio and dominates, so time scales with the length the model
+   actually sings, not the ceiling. Get the user's go-ahead before
+   `run_workflow` with `acknowledged_cost=true`.
+3. `wait_for_job`, then `get_job` for the manifest.
+4. You cannot listen: no tool returns audio inline. Hand the user the gallery
+   `url` (`list_gallery`, or the manifest's file name) and check what you can
+   yourself - `get_gallery_metadata` for the file's duration against the
+   ceiling (a track well short of it stopped on its own; one exactly at it was
+   cut) and the sample rate. Ask the user to listen for the family's failure
+   modes: a song that went instrumental (name the vocals in the caption), an
+   ending cut mid-note (raise the ceiling, then trim), a structure that ignored
+   the tags (fewer sections, plainer directions).
+5. To use the track in a later workflow, `keep_output` makes it an `asset:`;
+   to trim it in the same run, chain `templates/audio-trim-fade` on the output.
+
+## Sources
+
+MiniMax-Music3 model card (https://huggingface.co/MiniMaxAI/MiniMax-Music3,
+2026-08): the tag vocabulary, the Structured Caption headings, the five-minute
+range, the VRAM tiers. The `music-caption-rewriter` skill in
+https://github.com/MiniMax-AI/MiniMax-Music3 (2026-09-08): the caption
+contract. The diffusers `minimax_music3` modular pipeline and its docs page
+(diffusers 0.41, 2026-09-08): the ceiling semantics, the 9000-frame and
+5,000-token caps, the 200-frame window, the guider, the 44.1 kHz output, the
+drift-instrumental tip. The repo's audit of all of it is
+`docs/proposals/audits/2026-09-08-minimax-music3-audit.md`.
diff --git a/scripts/release.sh b/scripts/release.sh
index 77803b31..ebf8e403 100755
--- a/scripts/release.sh
+++ b/scripts/release.sh
@@ -69,10 +69,23 @@ with open(path, "w", encoding="utf-8") as f:
     f.write(text)
 EOF
     echo "pyproject.toml: $current -> $version"
+
+    python3 - "$version" <<'EOF'
+import json, sys
+
+path = "plugins/dw/.claude-plugin/plugin.json"
+with open(path, encoding="utf-8") as f:
+    plugin = json.load(f)
+plugin["version"] = sys.argv[1]
+with open(path, "w", encoding="utf-8") as f:
+    json.dump(plugin, f, indent=4, ensure_ascii=False)
+    f.write("\n")
+EOF
+    echo "plugins/dw/.claude-plugin/plugin.json: -> $version"
 fi
 
-if ! git diff --quiet -- pyproject.toml; then
-    git commit -m "release $version" -- pyproject.toml
+if ! git diff --quiet -- pyproject.toml plugins/dw/.claude-plugin/plugin.json; then
+    git commit -m "release $version" -- pyproject.toml plugins/dw/.claude-plugin/plugin.json
 else
     echo "pyproject.toml already at $version and committed - tagging HEAD"
 fi
diff --git a/tests/test_audio_utils.py b/tests/test_audio_utils.py
index d3383822..15134d72 100644
--- a/tests/test_audio_utils.py
+++ b/tests/test_audio_utils.py
@@ -195,6 +195,62 @@ def test_a_disallowed_extension_raises(self, tmp_path):
         with pytest.raises(Exception):
             load_audio(str(path))
 
+    @staticmethod
+    def _write_video(path, num_frames=8, fps=4, sample_rate=8000, audio=True):
+        from diffusers.utils.export_utils import encode_video
+        from PIL import Image
+
+        frames = [Image.new("RGB", (16, 16), (i, 0, 0)) for i in range(num_frames)]
+        track = (
+            torch.full((2, int(num_frames / fps * sample_rate)), 0.25)
+            if audio
+            else None
+        )
+        encode_video(
+            frames,
+            fps=fps,
+            output_path=str(path),
+            audio=track,
+            audio_sample_rate=sample_rate if audio else None,
+        )
+        return str(path)
+
+    def test_a_video_file_contributes_its_soundtrack(self, tmp_path):
+        # The cut an earlier run wrote is what a scoring pass mixes under; an
+        # agent hit "File extension not allowed: .mp4" here on 2026-09-08
+        path = self._write_video(tmp_path / "cut.mp4")
+
+        waveform, sample_rate = load_audio(path)
+
+        assert waveform.shape[0] == 2
+        assert sample_rate == 8000
+        assert abs(waveform.shape[1] - 2 * 8000) < 8000 // 4
+
+    def test_a_silent_video_file_is_an_error_not_silence(self, tmp_path):
+        path = self._write_video(tmp_path / "mute.mp4", audio=False)
+
+        with pytest.raises(ValueError, match="no audio track"):
+            load_audio(path)
+
+    def test_every_audio_task_takes_a_video_path(self, tmp_path):
+        from dw.tasks.audio_utils import (
+            crossfade_audio,
+            fade_audio,
+            mix_audio,
+            normalize_audio,
+            resample_audio,
+            slice_audio,
+        )
+
+        path = self._write_video(tmp_path / "cut.mp4")
+
+        assert resample_audio(path, 4000).shape[1] == 2
+        assert slice_audio(path, start_seconds=0, duration_seconds=1).shape[0] == 8000
+        assert fade_audio(path, fade_out_ms=100).shape[1] == 2
+        assert normalize_audio(path).shape[1] == 2
+        assert mix_audio([path, path]).shape[1] == 2
+        assert crossfade_audio([path, path], crossfade_ms=10).shape[1] == 2
+
 
 class TestBleedJoin:
     def test_the_tail_rings_on_over_a_silent_head(self):
diff --git a/tests/test_catalog_structure.py b/tests/test_catalog_structure.py
index f69cab4c..01fb3891 100644
--- a/tests/test_catalog_structure.py
+++ b/tests/test_catalog_structure.py
@@ -361,7 +361,9 @@ class TestLtxTwoStage:
     name, so the template carries the literal and this test ties it to the library."""
 
     def _definition(self):
-        path = os.path.join(REPO_ROOT, "workflows", "templates", "ltx2", "two-stage.json")
+        path = os.path.join(
+            REPO_ROOT, "workflows", "templates", "ltx2", "two-stage.json"
+        )
         return json.load(open(path, encoding="utf-8"))
 
     def test_the_renoise_scale_is_the_first_stage_two_sigma(self):
@@ -369,13 +371,19 @@ def test_the_renoise_scale_is_the_first_stage_two_sigma(self):
 
         refine = _step(self._definition(), "refine")
 
-        assert refine["pipeline"]["arguments"]["noise_scale"] == STAGE_2_DISTILLED_SIGMA_VALUES[0]
+        assert (
+            refine["pipeline"]["arguments"]["noise_scale"]
+            == STAGE_2_DISTILLED_SIGMA_VALUES[0]
+        )
 
     def test_the_refine_pass_runs_the_stage_two_schedule_on_the_upsampled_latents(self):
         refine = _step(self._definition(), "refine")
         arguments = refine["pipeline"]["arguments"]
 
-        assert arguments["sigmas"] == "constant:diffusers.pipelines.ltx2.utils.STAGE_2_DISTILLED_SIGMA_VALUES"
+        assert (
+            arguments["sigmas"]
+            == "constant:diffusers.pipelines.ltx2.utils.STAGE_2_DISTILLED_SIGMA_VALUES"
+        )
         assert arguments["latents"] == "previous_result:upscale.frames"
         assert arguments["audio_latents"] == "previous_result:base.audio"
 
@@ -394,7 +402,10 @@ def test_the_base_and_refine_passes_share_one_pipeline(self):
 LINK_PATTERN = re.compile(r"\]\(([^)]+)\)")
 READMES = sorted(
     os.path.relpath(path, REPO_ROOT)
-    for path in glob.glob(os.path.join(REPO_ROOT, "workflows", "templates", "**", "README.md"), recursive=True)
+    for path in glob.glob(
+        os.path.join(REPO_ROOT, "workflows", "templates", "**", "README.md"),
+        recursive=True,
+    )
 )
 
 
@@ -410,4 +421,6 @@ def test_every_readme_link_resolves(path):
         if target.startswith(("http://", "https://", "#")):
             continue
         target = target.split("#", 1)[0]
-        assert os.path.exists(os.path.join(base, target)), f"{path} links to {target}, which does not exist"
+        assert os.path.exists(
+            os.path.join(base, target)
+        ), f"{path} links to {target}, which does not exist"
diff --git a/tests/test_h3_context_ir.py b/tests/test_h3_context_ir.py
index ae5f50f9..d52945e4 100644
--- a/tests/test_h3_context_ir.py
+++ b/tests/test_h3_context_ir.py
@@ -9,6 +9,9 @@
 import glob
 import json
 import os
+import re
+
+import pytest
 
 from dw.security import MAX_VARIABLE_VALUE_LENGTH
 from tests.test_examples import BUILTIN_DIR, REPO_ROOT
@@ -31,7 +34,10 @@ def test_silent_audio_fields_are_written_as_not_applicable():
     prompt = _system_prompt()
 
     assert "N/A" in prompt
-    assert "overall_soundscape" in prompt[prompt.index("N/A") - 600 : prompt.index("N/A") + 600]
+    assert (
+        "overall_soundscape"
+        in prompt[prompt.index("N/A") - 600 : prompt.index("N/A") + 600]
+    )
 
 
 def test_video_and_audio_references_are_numbered_within_their_own_category():
@@ -101,3 +107,23 @@ def test_no_catalog_template_overrides_the_context_ir_system_prompt():
                 isinstance(value, str)
                 and value.startswith("prompt:prompt_enhancement/minimax_h3")
             ), f"{path}: overrides the builtin's system prompt with a stored copy"
+
+
+BUNDLED_H3_PROMPTS = sorted(
+    glob.glob(os.path.join(REPO_ROOT, "workflows", "templates", "minimax", "*.json"))
+)
+
+
+@pytest.mark.parametrize("path", BUNDLED_H3_PROMPTS, ids=lambda p: os.path.basename(p))
+def test_bundled_prompts_write_silent_audio_fields_as_the_literal(path):
+    """The guides' literal for a silent audio field is N/A; the templates had
+    drifted to "None." (found by the 2026-09-08 drill)."""
+    text = open(path, encoding="utf-8").read()
+    for field in ("overall_soundscape", "non_diegetic_music"):
+        for m in re.finditer(field + r":\\n([^\\]{0,12})", text):
+            head = m.group(1).strip().rstrip(".").lower()
+            assert head not in ("none", "no music", "silence", "silent", ""), (
+                path,
+                field,
+                m.group(1),
+            )
diff --git a/tests/test_ltx_prompt_library.py b/tests/test_ltx_prompt_library.py
index 8bad886f..b0f091c4 100644
--- a/tests/test_ltx_prompt_library.py
+++ b/tests/test_ltx_prompt_library.py
@@ -45,7 +45,9 @@ def test_a_prompt_is_one_paragraph_of_caption_length(path):
 
     assert "\n" not in text.strip(), f"{path} is more than one paragraph"
     words = len(text.split())
-    assert 140 <= words <= 240, f"{path} is {words} words; the trained caption is 150-220"
+    assert (
+        140 <= words <= 240
+    ), f"{path} is {words} words; the trained caption is 150-220"
 
 
 @pytest.mark.parametrize("path", PROMPTS, ids=os.path.basename)
@@ -62,7 +64,9 @@ def test_a_prompt_names_the_model_it_is_for(path):
 
 
 def test_no_ltx_template_summary_names_the_older_model():
-    templates = glob.glob(os.path.join(REPO_ROOT, "workflows", "templates", "ltx2", "*.json"))
+    templates = glob.glob(
+        os.path.join(REPO_ROOT, "workflows", "templates", "ltx2", "*.json")
+    )
     for path in templates:
         summary = json.load(open(path, encoding="utf-8")).get("summary", "")
         assert "LTX-2 " not in summary and not summary.endswith("LTX-2"), path
diff --git a/tests/test_mcp_media.py b/tests/test_mcp_media.py
index 65826e50..29c69209 100644
--- a/tests/test_mcp_media.py
+++ b/tests/test_mcp_media.py
@@ -453,6 +453,29 @@ def test_download_output_refuses_to_overwrite_an_existing_file_by_default(tmp_pa
     assert destination.read_bytes() == b"already here"
 
 
+def test_download_output_reports_an_unwritable_destination_as_the_servers_disk(
+    tmp_path,
+):
+    # A cold agent over dw.serve --mcp passed a path on its own machine; on
+    # the server that parent is a file, so makedirs raises. The message has
+    # to say the write happened server-side and name the ways to see the
+    # file from the client, instead of an anonymous tool error (drill
+    # 2026-09-08).
+    blocker = tmp_path / "private"
+    blocker.write_bytes(b"not a directory")
+    destination = blocker / "scratch" / "clip.mp4"
+    client = serving(png_bytes(10, 10), "image/png")
+
+    with pytest.raises(DwApiError) as excinfo:
+        download_output(client, "clip.mp4", destination=str(destination))
+
+    message = str(excinfo.value)
+    assert str(destination) in message
+    assert "machine running the MCP server" in message
+    assert "list_gallery" in message and "keep_output" in message
+    assert not destination.exists()
+
+
 def test_download_output_overwrite_true_replaces_an_existing_file(tmp_path):
     destination = tmp_path / "existing.png"
     destination.write_bytes(b"already here")
diff --git a/tests/test_pair_audio.py b/tests/test_pair_audio.py
index ca5f2537..0a431de8 100644
--- a/tests/test_pair_audio.py
+++ b/tests/test_pair_audio.py
@@ -22,6 +22,18 @@ def _waveform(samples=100, channels=2):
     return numpy.zeros((channels, samples), dtype=numpy.float32)
 
 
+def test_pairs_frames_with_a_soundtrack_read_from_a_file(tmp_path):
+    import soundfile
+
+    path = tmp_path / "score.wav"
+    soundfile.write(path, numpy.zeros((100, 2), dtype=numpy.float32), 16000)
+
+    paired = pair_audio(["frame1"], str(path))
+
+    assert paired.audio.shape == (2, 100)
+    assert paired.sample_rate == 16000
+
+
 def test_pairs_frames_with_a_bare_waveform():
     paired = pair_audio(["frame1", "frame2"], _waveform(), sample_rate=24000)
 
diff --git a/tests/test_plugin_skills.py b/tests/test_plugin_skills.py
new file mode 100644
index 00000000..5562a911
--- /dev/null
+++ b/tests/test_plugin_skills.py
@@ -0,0 +1,370 @@
+"""The dw plugin: skills that teach an agent to compose a model family's
+workflows, kept true by tests rather than by memory.
+
+A skill quotes catalog names and states numeric rules. Each name has to
+resolve to a file, each number has to come from the library that enforces
+it, and the plugin's version has to be the engine's - an installed skill is
+matched to the server it was written against by that number.
+"""
+
+import glob
+import inspect
+import os
+import re
+
+import pytest
+
+from tests.test_examples import REPO_ROOT
+
+PLUGIN_DIR = os.path.join(REPO_ROOT, "plugins", "dw")
+SKILLS = sorted(glob.glob(os.path.join(PLUGIN_DIR, "skills", "*", "SKILL.md")))
+SKILL_SIZE_LIMIT = 12 * 1024
+
+CATALOG_NAME = re.compile(r"`((?:templates|models)/[A-Za-z0-9_./-]+)`")
+
+
+def skill_text(path):
+    return open(path, encoding="utf-8").read()
+
+
+def frontmatter(text):
+    """The YAML block between the leading '---' lines, as a dict of the
+    top-level 'key: value' pairs (the two the plugin format needs)."""
+    assert text.startswith("---\n"), "a skill begins with frontmatter"
+    end = text.index("\n---", 4)
+    fields = {}
+    for line in text[4:end].splitlines():
+        if ":" in line and not line.startswith(" "):
+            key, value = line.split(":", 1)
+            fields[key.strip()] = value.strip().strip('"')
+    return fields
+
+
+def _pyproject_version():
+    text = open(os.path.join(REPO_ROOT, "pyproject.toml"), encoding="utf-8").read()
+    return re.search(r'^version = "(.*)"$', text, re.M).group(1)
+
+
+def test_the_marketplace_names_the_plugin():
+    import json
+
+    manifest = json.load(
+        open(
+            os.path.join(REPO_ROOT, ".claude-plugin", "marketplace.json"),
+            encoding="utf-8",
+        )
+    )
+
+    assert manifest["name"] == "diffusers-workflow"
+    (plugin,) = manifest["plugins"]
+    assert plugin["name"] == "dw"
+    assert plugin["source"] == "./plugins/dw"
+
+
+def test_the_plugin_version_is_the_engine_version():
+    """An installed plugin is matched to the engine it was written against by
+    this number, so the release script bumps both in one commit."""
+    import json
+
+    plugin = json.load(
+        open(
+            os.path.join(PLUGIN_DIR, ".claude-plugin", "plugin.json"), encoding="utf-8"
+        )
+    )
+
+    assert plugin["name"] == "dw"
+    assert plugin["version"] == _pyproject_version()
+
+
+def test_there_are_skills():
+    assert SKILLS, "the plugin ships at least one skill"
+
+
+@pytest.mark.parametrize(
+    "path", SKILLS, ids=lambda p: os.path.basename(os.path.dirname(p))
+)
+def test_a_skill_has_a_triggering_description_under_the_size_cap(path):
+    text = skill_text(path)
+    fields = frontmatter(text)
+
+    assert fields["name"] == os.path.basename(os.path.dirname(path))
+    assert "description" in fields and len(fields["description"]) > 40
+    assert (
+        len(text.encode("utf-8")) <= SKILL_SIZE_LIMIT
+    ), f"{path} is over {SKILL_SIZE_LIMIT} bytes"
+
+
+@pytest.mark.parametrize(
+    "path", SKILLS, ids=lambda p: os.path.basename(os.path.dirname(p))
+)
+def test_every_catalog_name_a_skill_quotes_resolves(path):
+    """A renamed template fails here rather than in a cold session."""
+    names = CATALOG_NAME.findall(skill_text(path))
+
+    assert names, f"{path} quotes no catalog names"
+    for name in names:
+        target = os.path.join(
+            REPO_ROOT, "workflows", name.removesuffix(".json") + ".json"
+        )
+        assert os.path.isfile(
+            target
+        ), f"{path} quotes {name}, which is not a workflow ({target})"
+
+
+H3_SKILL = os.path.join(PLUGIN_DIR, "skills", "minimax-h3", "SKILL.md")
+
+
+class TestMiniMaxH3Skill:
+    """The numbers the H3 skill states come from the diffusers modular
+    pipeline that enforces them, so a library change fails here."""
+
+    def test_the_frame_rule_and_bounds_are_the_pipeline_s(self):
+        import inspect
+
+        from diffusers.modular_pipelines.minimax_h3 import (
+            before_encoder,
+            modular_pipeline,
+        )
+
+        text = skill_text(H3_SKILL)
+        assert "17n + 5" in text or "17 * n + 5" in text
+        assert "17 * n + 5" in inspect.getsource(before_encoder)
+        assert modular_pipeline.MINIMAX_H3_FPS == 24 and "24 fps" in text
+        # 124 and 345 are the smallest and largest 17n + 5 inside 5 to 15 seconds at 24 fps
+        assert "124" in text and "345" in text
+        assert 124 == 17 * 7 + 5 and 345 == 17 * 20 + 5
+        # min_duration/max_duration are instance properties on MiniMaxH3ModularPipeline, so
+        # the window is pinned by the check that reads them rather than by their values
+        assert (
+            "min_duration <= duration <= components.max_duration"
+            in inspect.getsource(before_encoder)
+        )
+        assert (
+            124 / modular_pipeline.MINIMAX_H3_FPS >= 5
+            and 345 / modular_pipeline.MINIMAX_H3_FPS <= 15
+        )
+
+    def test_the_canvas_rules_are_the_pipeline_s(self):
+        import inspect
+
+        from diffusers.modular_pipelines.minimax_h3 import (
+            before_encoder,
+            modular_pipeline,
+        )
+
+        text = skill_text(H3_SKILL)
+        source = inspect.getsource(before_encoder)
+        assert 'ConfigSpec("canvas_short_edge", 768)' in source and "768" in text
+        assert "768 * 1344" in source and "1344" in text
+        assert modular_pipeline.MINIMAX_H3_MIN_ASPECT_RATIO == 1 / 4
+        assert modular_pipeline.MINIMAX_H3_MAX_ASPECT_RATIO == 4
+        assert "1:4" in text and "4:1" in text
+        assert "multiples of 32" in text
+
+    def test_the_reference_and_audio_limits_are_the_pipeline_s(self):
+        import inspect
+
+        from diffusers.modular_pipelines.minimax_h3 import (
+            before_encoder,
+            modular_pipeline,
+        )
+        from diffusers.modular_pipelines.minimax_h3.before_encoder import (
+            MiniMaxH3Ref2VASetupStep,
+        )
+
+        text = skill_text(H3_SKILL)
+        limits = inspect.signature(MiniMaxH3Ref2VASetupStep.__init__).parameters
+        assert limits["max_images"].default == 9 and "9 images" in text
+        assert limits["max_videos"].default == 3 and "3 videos" in text
+        assert limits["max_audios"].default == 3 and "3 audio clips" in text
+        assert limits["max_references"].default == 12 and "12 files" in text
+        # the 32 kHz rate is the audio VAE's, with this literal as the fallback
+        assert "return 32000" in inspect.getsource(modular_pipeline)
+        assert (
+            modular_pipeline.MINIMAX_H3_AUDIO_CHANNELS == 2 and "32 kHz stereo" in text
+        )
+        assert "audio can" in text and "never be the only reference" in text
+        assert before_encoder is not None
+
+    def test_the_lora_coupling_is_scoped_to_the_turbo_templates(self):
+        """Six reference-only templates carry no LoRA and run 20 steps; the four
+        that mix references with the turbo LoRA run nine (the 2026-09-08 drill
+        caught the skill saying 20 for storyboard)."""
+        import json
+
+        text = skill_text(H3_SKILL)
+        assert "run 20 steps" in text
+        assert "keep the turbo LoRA at\n  nine steps" in text
+        for name in (
+            "storyboard",
+            "dialogue-short",
+            "music-video",
+            "chain-matched-and-aligned",
+        ):
+            path = os.path.join(
+                REPO_ROOT, "workflows", "templates", "minimax", name + ".json"
+            )
+            spec = open(path, encoding="utf-8").read()
+            assert (
+                '"loras"' in spec or "lora_model_name" in spec
+            ) and '"num_inference_steps": 9' in spec, name
+        for name in (
+            "reference-to-video",
+            "composable-references",
+            "voice-timbre-reference",
+            "generated-subject-reference",
+            "chain-matched-to-audio",
+            "chain-video-continuity",
+        ):
+            path = os.path.join(
+                REPO_ROOT, "workflows", "templates", "minimax", name + ".json"
+            )
+            spec = open(path, encoding="utf-8").read()
+            assert "lora_model_name" not in spec, f"{name} now loads a LoRA"
+            assert (
+                '"num_inference_steps": 20' in spec
+            ), f"{name} no longer runs 20 steps"
+            json.loads(spec)
+
+    def test_the_skill_defers_prompt_format_to_minimax(self):
+        text = skill_text(H3_SKILL)
+        assert "h3-prompt-writing" in text
+        assert "VIDEO_PROMPT_WRITING_GUIDE_base_en.md" in text
+        assert "VIDEO_PROMPT_WRITING_GUIDE_ref_en.md" in text
+        assert "`templates/minimax/enhance-prompt`" in text
+        # no transcription of the format: none of its section labels appear as instructions
+        assert "retention_analysis:" not in text
+
+    def test_the_skill_starts_with_the_server(self):
+        text = skill_text(H3_SKILL)
+        assert text.index("get_server_info") < text.index("templates/minimax/")
+
+
+LTX_SKILL = os.path.join(PLUGIN_DIR, "skills", "ltx-2.5", "SKILL.md")
+
+
+def _fenced_block_after(text, heading):
+    """The first fenced code block after a markdown heading."""
+    start = text.index(heading)
+    open_fence = text.index("\n```", start)
+    open_end = text.index("\n", open_fence + 1)
+    close_fence = text.index("\n```", open_end)
+    return text[open_end + 1 : close_fence]
+
+
+class TestLtx25Skill:
+    """The LTX-2.5 skill's numbers come from the pipeline that enforces them,
+    and the one vendor text it quotes is the library's own constant."""
+
+    def test_the_schedule_is_the_library_s(self):
+        from diffusers.pipelines.ltx2.utils import (
+            DISTILLED_SIGMA_VALUES,
+            STAGE_2_DISTILLED_SIGMA_VALUES,
+        )
+
+        text = skill_text(LTX_SKILL)
+        assert len(DISTILLED_SIGMA_VALUES) == 8 and "eight" in text
+        assert len(STAGE_2_DISTILLED_SIGMA_VALUES) == 3 and "three" in text
+        assert str(STAGE_2_DISTILLED_SIGMA_VALUES[0]) in text
+
+    def test_the_size_and_frame_rules_are_the_pipeline_s(self):
+        import inspect
+
+        from diffusers.pipelines.ltx2 import pipeline_ltx2
+        from diffusers.pipelines.ltx2.utils import MAX_CONDITIONING_FPS
+
+        text = skill_text(LTX_SKILL)
+        source = inspect.getsource(pipeline_ltx2)
+        assert (
+            "height % 32 != 0 or width % 32 != 0" in source
+            and "multiples of 32" in text
+        )
+        assert "(num_frames - 1) // self.vae_temporal_compression_ratio + 1" in source
+        assert "8k + 1" in text or "8n + 1" in text
+        assert MAX_CONDITIONING_FPS == 60.0 and "60" in text
+
+    def test_the_image_condition_crf_is_the_library_s(self):
+        from diffusers.pipelines.ltx2.utils import LTX2_5_IMAGE_CRF
+
+        assert LTX2_5_IMAGE_CRF == 18
+        assert "CRF 18" in skill_text(LTX_SKILL)
+
+    def test_the_quoted_caption_spec_is_the_library_constant(self):
+        """The one vendor text the plugin carries, tied to the library that
+        ships it so it cannot drift."""
+        from diffusers.pipelines.ltx2.utils import LTX2_5_T2V_DEFAULT_SYSTEM_PROMPT
+
+        quoted = _fenced_block_after(
+            skill_text(LTX_SKILL), "## The trained caption spec"
+        )
+
+        assert " ".join(quoted.split()) == " ".join(
+            LTX2_5_T2V_DEFAULT_SYSTEM_PROMPT.split()
+        )
+
+    def test_the_skill_starts_with_the_server(self):
+        text = skill_text(LTX_SKILL)
+        assert text.index("get_server_info") < text.index("templates/ltx2/")
+
+
+MUSIC_SKILL = os.path.join(PLUGIN_DIR, "skills", "minimax-music3", "SKILL.md")
+
+
+class TestMiniMaxMusic3Skill:
+    """The Music 3 skill's numbers come from the diffusers modular pipeline
+    that enforces them, and the caption format is deferred to MiniMax."""
+
+    def test_the_ceiling_and_the_caps_are_the_encoder_s(self):
+        from diffusers.modular_pipelines.minimax_music3 import encoders
+
+        text = skill_text(MUSIC_SKILL)
+        source = inspect.getsource(encoders)
+        assert encoders._MAX_AUDIO_FRAMES == 9000 and "9000 frames" in text
+        assert encoders._MAX_PROMPT_TOKENS == 5000 and "5,000 tokens" in text
+        # the ceiling semantics: the language model may stop earlier
+        assert re.search(r'"audio_duration",\s*default=60.0', source)
+        assert "Default\n  60 seconds" in text or "Default 60 seconds" in text
+        assert "ceiling, not a target" in text
+        # a tag line keeps only its leading bracketed tags, lower-cased
+        assert "_LEADING_TAGS_RE" in source and hasattr(encoders, "_normalize_lyrics")
+        assert "dropped" in text and "lower-cased" in text
+
+    def test_the_frame_rate_and_the_output_rate_are_the_pipeline_s(self):
+        from diffusers.modular_pipelines.minimax_music3 import modular_pipeline
+        from diffusers.models.autoencoders import minimax_music3_vocoder
+
+        text = skill_text(MUSIC_SKILL)
+        assert "frame_rate = 25.0" in inspect.getsource(modular_pipeline)
+        assert "25 frames per second" in text and "360 seconds" in text
+        assert "sampling_rate: int = 44100" in inspect.getsource(minimax_music3_vocoder)
+        assert "44.1 kHz stereo" in text and "sample_rate: 44100" in text
+
+    def test_the_window_steps_and_guidance_are_the_denoiser_s(self):
+        from diffusers.modular_pipelines.minimax_music3 import before_denoise, denoise
+
+        text = skill_text(MUSIC_SKILL)
+        assert before_denoise._CHUNK_FRAMES == 200 and "200-frame windows" in text
+        source = inspect.getsource(denoise)
+        assert re.search(r'"num_inference_steps",\s*default=30', source)
+        assert "30 steps" in text
+        assert '"guidance_scale": 1.7' in source and "fixed at 1.7" in text
+
+    def test_the_skill_defers_caption_format_to_minimax(self):
+        text = skill_text(MUSIC_SKILL)
+        assert "music-caption-rewriter" in text
+        assert "https://github.com/MiniMax-AI/MiniMax-Music3" in text
+        assert "https://huggingface.co/MiniMaxAI/MiniMax-Music3" in text
+        for tag in ("[Intro]", "[Pre-Chorus]", "[Bridge]", "[Instrumental]", "[Outro]"):
+            assert tag in text, tag
+        assert "2026-09-08-minimax-music3-audit.md" in text
+
+    def test_the_music_video_ceiling_clears_its_slices(self):
+        """The sliced total is 496 frames at 24 fps; the ceiling must clear it
+        with margin, since the model may stop early (audit 2026-09-08)."""
+        import json
+
+        path = os.path.join(
+            REPO_ROOT, "workflows", "templates", "minimax", "music-video.json"
+        )
+        spec = json.load(open(path, encoding="utf-8"))
+        assert spec["variables"]["audio_duration"] >= 496 / 24 + 5
diff --git a/tests/test_result.py b/tests/test_result.py
index e25358b6..e02ca748 100644
--- a/tests/test_result.py
+++ b/tests/test_result.py
@@ -247,7 +247,9 @@ class TestLatentHandoff:
     def _base_result(self):
         result = Result({"content_type": "video/mp4", "save": False})
         result.add_result(
-            _LatentOutput(frames=torch.zeros(1, 128, 16, 14, 24), audio=torch.zeros(1, 8, 50, 16))
+            _LatentOutput(
+                frames=torch.zeros(1, 128, 16, 14, 24), audio=torch.zeros(1, 8, 50, 16)
+            )
         )
         return result
 
diff --git a/workflows/templates/ltx2/README.md b/workflows/templates/ltx2/README.md
index cf069fae..05a55838 100644
--- a/workflows/templates/ltx2/README.md
+++ b/workflows/templates/ltx2/README.md
@@ -16,6 +16,10 @@ written to it. Settings come from the
 `ltx-pipelines` notes. Audited against those sources on 2026-09-07
 ([the audit](../../../docs/proposals/audits/2026-09-07-ltx-2.5-audit.md)).
 
+An agent driving this family from Claude Code has the `ltx-2.5` skill of the
+[dw plugin](../../../plugins/dw/README.md), which chooses among these templates
+and carries the caption spec.
+
 Read them in this order and each introduces one new idea on top of the last.
 
 ## The basics
diff --git a/workflows/templates/minimax/README.md b/workflows/templates/minimax/README.md
index 10e3452c..7ff7e899 100644
--- a/workflows/templates/minimax/README.md
+++ b/workflows/templates/minimax/README.md
@@ -15,6 +15,16 @@ and MiniMax publishes them as an agent skill, `skills/h3-prompt-writing`, in the
 for hand-written prompts, read them rather than reverse-engineering the examples.
 Audited against those sources on 2026-09-07
 ([the audit](../../../docs/proposals/audits/2026-09-07-minimax-h3-audit.md)).
+Music3 has the same kind of authority: the caption format is MiniMax's
+`skills/music-caption-rewriter` in the
+[MiniMax-Music3 GitHub repository](https://github.com/MiniMax-AI/MiniMax-Music3)
+(a genre router, 18 family indexes and 1,000 example captions) and the tag
+vocabulary is on its model card; audited 2026-09-08
+([the audit](../../../docs/proposals/audits/2026-09-08-minimax-music3-audit.md)).
+
+An agent driving this family from Claude Code has the `minimax-h3` and
+`minimax-music3` skills of the [dw plugin](../../../plugins/dw/README.md), which
+choose among these templates and point at the guides.
 
 Read them in this order and each introduces one new idea on top of the last.
 
@@ -26,11 +36,15 @@ Read them in this order and each introduces one new idea on top of the last.
 | [video-with-audio.json](video-with-audio.json) | The baseline text-to-video-audio run: per-component SDNQ quantization, mixed offload, the turbo LoRA, and muxing video + audio into one file |
 
 A note on `audio_duration`: Music3 reads it as a ceiling rather than a target.
-The piece ends where the music ends, so a value set to the length the song
-*should* be will guillotine the outro mid-decay. Ask for more time than the
-song needs and trim the tail afterwards -
+The language model stops when the song ends, so the track is usually shorter
+than the ceiling, and a ceiling set to the length the song *should* be leaves
+no room for the outro before the hard stop. Ask for more time than the song
+needs and trim the tail afterwards -
 [templates/audio-trim-fade.json](../audio-trim-fade.json) slices a generated track
-to length and fades the cut into an ending.
+to length and fades the cut into an ending. Runtime follows the length actually
+generated, not the ceiling, so the margin is free. The output is 44.1 kHz stereo,
+the vocoder's native rate; the model card's 32 kHz is what MiniMax's reference
+server resamples to.
 
 A note on length: H3 accepts any `num_frames` of the form `17n + 5` between 124
 and 345 - at its fixed 24 fps, that is 5.17 to 14.4 seconds **in a single clip**.
@@ -42,10 +56,16 @@ should span the full duration, since a prompt scripted for five seconds
 conditions a five-second story regardless of the frame count. Reach for chains
 and cuts when you need to go past 14 seconds.
 
-A note on the canvas: H3 draws on a 768-pixel short edge (960x544 in these
-examples is the speed choice, coupled to the 544p turbo LoRA and its nine steps -
-change one and change the others), dimensions are multiples of 32, and aspect ratios
-run from 1:4 to 4:1. The 5-second floor is diffusers' constraint; the model card and
+A note on the canvas: H3 draws on a 768-pixel short edge, dimensions are
+multiples of 32, and aspect ratios run from 1:4 to 4:1. Every example renders at
+960x544, the speed choice. The text- and frame-conditioned ones couple it to the
+544p turbo LoRA and its nine steps - change one and change the others. The six
+conditioned on references alone (`reference-to-video`, `composable-references`,
+`voice-timbre-reference`, `generated-subject-reference`, `chain-matched-to-audio`,
+`chain-video-continuity`) load no LoRA and run 20 steps, because the turbo LoRA is
+distilled against the base transformer and they load the reference one;
+`storyboard`, `dialogue-short`, `music-video` and `chain-matched-and-aligned` pass
+references and still keep the turbo LoRA at nine. The 5-second floor is diffusers' constraint; the model card and
 the hosted API accept 4. Output audio is 32 kHz stereo.
 
 ## Conditioning on frames
@@ -73,6 +93,10 @@ the hosted API accept 4. Output audio is 32 kHz stereo.
 | [generated-subject-reference.json](generated-subject-reference.json) | Drawing the subject with Z-Image first and referencing it with `from_previous_result` |
 | [storyboard.json](storyboard.json) | Several images in one request: a first frame plus storyboard anchors for later shots, so one generation cuts between three boards under an unbroken score |
 
+Ref2VA takes at most 9 images, 3 videos, 3 audio clips and 12 files in total, and
+audio can never be the only reference. Labels follow the order the request passes
+them, numbered independently per category.
+
 ## Going long: chains
 
 A `chain` block runs the pipeline once per segment and stitches the output. Chains
@@ -97,7 +121,14 @@ Chained generation fights drift; a scene cut erases it. Each shot in these
 examples is generated fresh from the same portraits, so the last shot is exactly
 as clean as the first and the piece can run as long as the script does - which is
 how television gets away with it too. This is the pattern for digital shorts:
-write shots, not takes.
+write shots, not takes. What a cut gives up is audio: H3 carries nothing between
+generations except what is passed as a reference, so each shot generates its own
+sound and a score would restart at every cut. Write `non_diegetic_music: N/A` in
+every shot and lay one track under the finished edit, as `music-video.json` does
+with `pair_audio` and [assemble-and-score.json](../assemble-and-score.json) shows
+on its own; a character who speaks in several shots keeps one voice by passing the
+same clip as an audio reference in each shot, as
+[voice-timbre-reference.json](voice-timbre-reference.json) does.
 
 | Example | What it introduces |
 | ------- | ------------------ |
diff --git a/workflows/templates/minimax/chain-matched-and-aligned.json b/workflows/templates/minimax/chain-matched-and-aligned.json
index 6af5fb7a..d9779a9a 100644
--- a/workflows/templates/minimax/chain-matched-and-aligned.json
+++ b/workflows/templates/minimax/chain-matched-and-aligned.json
@@ -4,8 +4,8 @@
     "summary": "A long H3 clip sized to a supplied soundtrack, with video-tail continuity across every seam.",
     "seed": 12345,
     "variables": {
-        "first_segment_prompt": "subject_definitions:\n<Subject 1> is the cartoon astronaut in <Picture 1>, whose face, hair, clothing and proportions define their appearance for the whole video.\n<Audio 1> is the speech track <Subject 1> (S1) is lip-synchronised to: a single continuous recorded voiceover, even in level and unhurried in delivery.\n\nsummary:\n[reference generation + audio reuse] The target video shows <Subject 1> delivering the speech carried by <Audio 1> directly to camera in the setting of <Picture 1>, in one continuous take that runs to the end of the clip without resolving.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, hair, clothing and proportions of <Picture 1> are kept exactly, with the head, jaw and shoulders newly animated to speak.\n<Audio 1>: partially_copy - the segment of the recording covering this part of the timeline is reused as <Subject 1>'s speech, and the mouth is synchronised to it.\n\ndetailed_description:\nThe target video is in realistic photographic style. [Shot 1] A medium close-up frames <Subject 1> from the chest up, centred and facing camera in the setting of <Picture 1>, held as one continuous shot. The lighting of <Picture 1> is kept exactly as it is, steady and unchanging across the whole clip, and the background sits well behind in shallow focus. <Subject 1> (S1) speaks continuously, the jaw and lips moving in natural synchronisation with <Audio 1> and pausing only where the recording pauses. Between phrases <Subject 1> shifts weight very slightly and the head drifts a few degrees off centre and back, with the eyes staying on the lens and blinking at a natural rate. The camera holds a static shot throughout. The clip ends mid-delivery with <Subject 1> still speaking, the pose, framing and lighting unchanged from the rest of the shot.\n\noverall_soundscape:\nA quiet, close room tone at an even level from the first frame to the last, with no doors, footsteps or movement in the room, and faint clothing rustle as <Subject 1> shifts weight.\n\nnon_diegetic_music:\nNone.",
-        "continuation_prompt": "subject_definitions:\n<Subject 1> is the cartoon astronaut in <Picture 1>, whose face, hair, clothing and proportions define their appearance for the whole video.\n<Audio 1> is the speech track <Subject 1> (S1) is lip-synchronised to: a single continuous recorded voiceover, even in level and unhurried in delivery.\n<Video 1> is the tail of the shot immediately preceding this one, and it is the same continuous take: it carries <Subject 1>'s pose, head angle, weight, framing, lighting and speaking rhythm at the moment this clip begins.\n\nsummary:\n[video continuation + reference generation + audio reuse] The target video continues the unbroken take of <Video 1> without a cut, picking up <Subject 1>'s exact pose, framing, lighting and speaking rhythm at the frame <Video 1> ends on, and carrying the delivery of <Audio 1> onward to the end of the clip.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, hair, clothing and proportions of <Picture 1> are kept exactly, unchanged from <Video 1>.\n<Video 1>: fully_preserved - the pose, head angle, weight distribution, framing, camera position, lighting and speaking rhythm at its final frame are continued directly into the first frame of this clip; none of its frames are reused.\n<Audio 1>: partially_copy - the segment of the recording covering this part of the timeline is reused as <Subject 1>'s speech, and the mouth is synchronised to it.\n\ndetailed_description:\nThe target video is in realistic photographic style. [Shot 1] The same medium close-up as <Video 1>, framing <Subject 1> from the chest up, centred and facing camera in the setting of <Picture 1>, held as one continuous shot. The first frame of this clip continues directly from the last frame of <Video 1>: the head angle, the weight on each foot, the position of the hands, the direction of the gaze and the state of the mouth all carry over without a jump, and the lighting and background are identical and stay unchanged across the whole clip. <Subject 1> (S1) speaks continuously, the jaw and lips moving in natural synchronisation with <Audio 1> and pausing only where the recording pauses, in the same unhurried delivery as <Video 1>. Between phrases <Subject 1> shifts weight very slightly and the head drifts a few degrees off centre and back, with the eyes staying on the lens and blinking at a natural rate. The camera holds the same static shot throughout, with no reframing, no cut and no change in scale. The clip ends mid-delivery with <Subject 1> still speaking, the pose, framing and lighting unchanged.\n\noverall_soundscape:\nA quiet, close room tone identical to <Video 1> and at the same even level from the first frame to the last, with no doors, footsteps or movement in the room, and faint clothing rustle as <Subject 1> shifts weight.\n\nnon_diegetic_music:\nNone.",
+        "first_segment_prompt": "subject_definitions:\n<Subject 1> is the cartoon astronaut in <Picture 1>, whose face, hair, clothing and proportions define their appearance for the whole video.\n<Audio 1> is the speech track <Subject 1> (S1) is lip-synchronised to: a single continuous recorded voiceover, even in level and unhurried in delivery.\n\nsummary:\n[reference generation + audio reuse] The target video shows <Subject 1> delivering the speech carried by <Audio 1> directly to camera in the setting of <Picture 1>, in one continuous take that runs to the end of the clip without resolving.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, hair, clothing and proportions of <Picture 1> are kept exactly, with the head, jaw and shoulders newly animated to speak.\n<Audio 1>: partially_copy - the segment of the recording covering this part of the timeline is reused as <Subject 1>'s speech, and the mouth is synchronised to it.\n\ndetailed_description:\nThe target video is in realistic photographic style. [Shot 1] A medium close-up frames <Subject 1> from the chest up, centred and facing camera in the setting of <Picture 1>, held as one continuous shot. The lighting of <Picture 1> is kept exactly as it is, steady and unchanging across the whole clip, and the background sits well behind in shallow focus. <Subject 1> (S1) speaks continuously, the jaw and lips moving in natural synchronisation with <Audio 1> and pausing only where the recording pauses. Between phrases <Subject 1> shifts weight very slightly and the head drifts a few degrees off centre and back, with the eyes staying on the lens and blinking at a natural rate. The camera holds a static shot throughout. The clip ends mid-delivery with <Subject 1> still speaking, the pose, framing and lighting unchanged from the rest of the shot.\n\noverall_soundscape:\nA quiet, close room tone at an even level from the first frame to the last, with no doors, footsteps or movement in the room, and faint clothing rustle as <Subject 1> shifts weight.\n\nnon_diegetic_music:\nN/A",
+        "continuation_prompt": "subject_definitions:\n<Subject 1> is the cartoon astronaut in <Picture 1>, whose face, hair, clothing and proportions define their appearance for the whole video.\n<Audio 1> is the speech track <Subject 1> (S1) is lip-synchronised to: a single continuous recorded voiceover, even in level and unhurried in delivery.\n<Video 1> is the tail of the shot immediately preceding this one, and it is the same continuous take: it carries <Subject 1>'s pose, head angle, weight, framing, lighting and speaking rhythm at the moment this clip begins.\n\nsummary:\n[video continuation + reference generation + audio reuse] The target video continues the unbroken take of <Video 1> without a cut, picking up <Subject 1>'s exact pose, framing, lighting and speaking rhythm at the frame <Video 1> ends on, and carrying the delivery of <Audio 1> onward to the end of the clip.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, hair, clothing and proportions of <Picture 1> are kept exactly, unchanged from <Video 1>.\n<Video 1>: fully_preserved - the pose, head angle, weight distribution, framing, camera position, lighting and speaking rhythm at its final frame are continued directly into the first frame of this clip; none of its frames are reused.\n<Audio 1>: partially_copy - the segment of the recording covering this part of the timeline is reused as <Subject 1>'s speech, and the mouth is synchronised to it.\n\ndetailed_description:\nThe target video is in realistic photographic style. [Shot 1] The same medium close-up as <Video 1>, framing <Subject 1> from the chest up, centred and facing camera in the setting of <Picture 1>, held as one continuous shot. The first frame of this clip continues directly from the last frame of <Video 1>: the head angle, the weight on each foot, the position of the hands, the direction of the gaze and the state of the mouth all carry over without a jump, and the lighting and background are identical and stay unchanged across the whole clip. <Subject 1> (S1) speaks continuously, the jaw and lips moving in natural synchronisation with <Audio 1> and pausing only where the recording pauses, in the same unhurried delivery as <Video 1>. Between phrases <Subject 1> shifts weight very slightly and the head drifts a few degrees off centre and back, with the eyes staying on the lens and blinking at a natural rate. The camera holds the same static shot throughout, with no reframing, no cut and no change in scale. The clip ends mid-delivery with <Subject 1> still speaking, the pose, framing and lighting unchanged.\n\noverall_soundscape:\nA quiet, close room tone identical to <Video 1> and at the same even level from the first frame to the last, with no doors, footsteps or movement in the room, and faint clothing rustle as <Subject 1> shifts weight.\n\nnon_diegetic_music:\nN/A",
         "subject": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg",
         "voice": "https://huggingface.co/datasets/Xenova/transformers.js-docs/resolve/main/jfk.wav",
         "num_frames": 124,
diff --git a/workflows/templates/minimax/dialogue-short.json b/workflows/templates/minimax/dialogue-short.json
index 91574c45..134100e6 100644
--- a/workflows/templates/minimax/dialogue-short.json
+++ b/workflows/templates/minimax/dialogue-short.json
@@ -5,10 +5,10 @@
     "variables": {
         "howie_portrait_prompt": "a 3/4-length studio portrait of a wiry man in his early 30s with short curly dark hair and expressive eyebrows, wearing a mustard-yellow sweater vest over a white button-down shirt, standing in a 1990s multi-camera sitcom apartment set with a sagging olive-green couch and a brick wall behind him, warm even studio lighting, sharp focus, photographic",
         "pat_portrait_prompt": "a 3/4-length studio portrait of a heavyset man in his mid 40s with a shaved head and a calm, heavy-lidded expression, wearing an open gray flannel shirt over a dark t-shirt, standing in a 1990s multi-camera sitcom apartment set with a sagging olive-green couch and a brick wall behind him, warm even studio lighting, sharp focus, photographic",
-        "shot_1_cold_open": "subject_definitions:\n<Subject 1> is the wiry man in <Picture 1>, early 30s, short curly dark hair, mustard-yellow sweater vest over a white button-down shirt.\n<Subject 2> is the heavyset man in <Picture 2>, mid 40s, shaved head, open gray flannel shirt over a dark t-shirt.\n\nsummary:\n[reference generation] The target video is one shot of a 1990s multi-camera sitcom: <Subject 2> reads on the couch while <Subject 1> enters holding an empty ice cream carton and confronts him, in front of a live studio audience.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, hair, sweater vest and proportions of <Picture 1> are kept exactly, with the body newly animated to walk and speak.\n<Subject 2> (appears in [Shot 1]): fully_preserved - the face, head, flannel shirt and proportions of <Picture 2> are kept exactly, with the body newly posed seated and reading.\n\ndetailed_description:\nThe target video is in the style of a 1990s multi-camera sitcom, shot on studio video with warm, even, shadowless lighting. [Shot 1] A wide shot frames a living room set: a sagging olive-green couch centre frame in front of a brick wall with a window onto a painted city backdrop. <Subject 2> sits at the right end of the couch, ankles crossed, reading a magazine and not looking up. <Subject 1> strides in from frame left holding an empty pint-sized ice cream carton at chest height, stops beside the couch, holds the carton out toward <Subject 2> and turns it upside down to show it is empty. <Subject 1> (S1), a fast, nasal, high-pitched male tenor voice, tight with disbelief, says <d>[English] Okay. Who finished the pistachio?</d> <Subject 2> keeps reading and does not react. A live studio audience laughs warmly as the carton turns over. The camera holds a static wide shot throughout, level with the set, the way a sitcom master camera holds a scene.\n\noverall_soundscape:\nA bright, close television-studio room tone. Quick footsteps on a hollow stage floor as <Subject 1> enters, the papery rattle of the empty carton as it turns over, a soft page turn from the couch, and a warm burst of live studio-audience laughter that begins as the carton flips and settles under the end of the shot.\n\nnon_diegetic_music:\nNone.",
-        "shot_2_pat_deflects": "subject_definitions:\n<Subject 1> is the heavyset man in <Picture 1>, mid 40s, shaved head, calm and heavy-lidded, open gray flannel shirt over a dark t-shirt.\n\nsummary:\n[reference generation] The target video is one close-up shot of a 1990s multi-camera sitcom: <Subject 1> answers an accusation without looking up from his magazine, in front of a live studio audience.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, head, flannel shirt and proportions of <Picture 1> are kept exactly, with the head and jaw newly animated to speak.\n\ndetailed_description:\nThe target video is in the style of a 1990s multi-camera sitcom, shot on studio video with warm, even, shadowless lighting. [Shot 1] A medium close-up frames <Subject 1> from the chest up, seated at the end of an olive-green couch in front of a brick wall, holding an open magazine just below the frame line and reading it. He does not look up. <Subject 1> (S1), a flat, low, unhurried male baritone voice, completely unbothered, says <d>[English] Technically? The freezer has a defrost cycle.</d> On the last word he turns a page. A live studio audience laughs as the line lands. The camera holds a static close-up throughout, the way a sitcom cross-shot camera holds a speaker.\n\noverall_soundscape:\nA bright, close television-studio room tone. A soft page turn, the faint creak of couch springs as <Subject 1> settles, and a rolling wave of live studio-audience laughter that starts on the last word of the line and carries under the end of the shot.\n\nnon_diegetic_music:\nNone.",
-        "shot_3_howie_incredulous": "subject_definitions:\n<Subject 1> is the wiry man in <Picture 1>, early 30s, short curly dark hair, expressive eyebrows, mustard-yellow sweater vest over a white button-down shirt.\n\nsummary:\n[reference generation] The target video is one close-up shot of a 1990s multi-camera sitcom: <Subject 1> repeats an absurd excuse back in disbelief, holding up an empty ice cream carton, in front of a live studio audience.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, hair, sweater vest and proportions of <Picture 1> are kept exactly, with the head, jaw and arm newly animated to speak and gesture.\n\ndetailed_description:\nThe target video is in the style of a 1990s multi-camera sitcom, shot on studio video with warm, even, shadowless lighting. [Shot 1] A medium close-up frames <Subject 1> from the chest up, standing beside a couch in front of a brick wall, an empty pint-sized ice cream carton raised in one hand at head height. His eyebrows are high and his mouth is half open before he speaks. <Subject 1> (S1), a fast, nasal, high-pitched male tenor voice, tight with disbelief, cracking upward on the last word, says <d>[English] The defrost cycle used a spoon?</d> He shakes the empty carton once on the word spoon. A live studio audience erupts in a big laugh as the line lands. The camera holds a static close-up throughout, the way a sitcom cross-shot camera holds a speaker.\n\noverall_soundscape:\nA bright, close television-studio room tone. The hollow papery rattle of the empty carton as it shakes, and a big rising wave of live studio-audience laughter that breaks on the word spoon and rolls under the end of the shot.\n\nnon_diegetic_music:\nNone.",
-        "shot_4_pat_button": "subject_definitions:\n<Subject 1> is the heavyset man in <Picture 1>, mid 40s, shaved head, calm and heavy-lidded, open gray flannel shirt over a dark t-shirt.\n\nsummary:\n[reference generation] The target video is one close-up shot of a 1990s multi-camera sitcom: <Subject 1> shrugs off the accusation with a deadpan line, in front of a live studio audience.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, head, flannel shirt and proportions of <Picture 1> are kept exactly, with the head, jaw and shoulders newly animated to speak and shrug.\n\ndetailed_description:\nThe target video is in the style of a 1990s multi-camera sitcom, shot on studio video with warm, even, shadowless lighting. [Shot 1] A medium close-up frames <Subject 1> from the chest up, seated at the end of an olive-green couch in front of a brick wall, the magazine now closed in his lap. He looks up for the first time, holds a beat of eye contact just past the lens, and gives one slow shrug. <Subject 1> (S1), a flat, low, unhurried male baritone voice, perfectly deadpan, says <d>[English] I don't know its life.</d> He goes back to the magazine. A live studio audience gives a big laugh with scattered applause as he looks back down. The camera holds a static close-up throughout, the way a sitcom cross-shot camera holds a speaker.\n\noverall_soundscape:\nA bright, close television-studio room tone. The soft slap of the magazine reopening, a couch-spring creak under the shrug, and a big wave of live studio-audience laughter with a few claps in it that breaks after the line and rolls under the end of the shot.\n\nnon_diegetic_music:\nNone.",
+        "shot_1_cold_open": "subject_definitions:\n<Subject 1> is the wiry man in <Picture 1>, early 30s, short curly dark hair, mustard-yellow sweater vest over a white button-down shirt.\n<Subject 2> is the heavyset man in <Picture 2>, mid 40s, shaved head, open gray flannel shirt over a dark t-shirt.\n\nsummary:\n[reference generation] The target video is one shot of a 1990s multi-camera sitcom: <Subject 2> reads on the couch while <Subject 1> enters holding an empty ice cream carton and confronts him, in front of a live studio audience.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, hair, sweater vest and proportions of <Picture 1> are kept exactly, with the body newly animated to walk and speak.\n<Subject 2> (appears in [Shot 1]): fully_preserved - the face, head, flannel shirt and proportions of <Picture 2> are kept exactly, with the body newly posed seated and reading.\n\ndetailed_description:\nThe target video is in the style of a 1990s multi-camera sitcom, shot on studio video with warm, even, shadowless lighting. [Shot 1] A wide shot frames a living room set: a sagging olive-green couch centre frame in front of a brick wall with a window onto a painted city backdrop. <Subject 2> sits at the right end of the couch, ankles crossed, reading a magazine and not looking up. <Subject 1> strides in from frame left holding an empty pint-sized ice cream carton at chest height, stops beside the couch, holds the carton out toward <Subject 2> and turns it upside down to show it is empty. <Subject 1> (S1), a fast, nasal, high-pitched male tenor voice, tight with disbelief, says <d>[English] Okay. Who finished the pistachio?</d> <Subject 2> keeps reading and does not react. A live studio audience laughs warmly as the carton turns over. The camera holds a static wide shot throughout, level with the set, the way a sitcom master camera holds a scene.\n\noverall_soundscape:\nA bright, close television-studio room tone. Quick footsteps on a hollow stage floor as <Subject 1> enters, the papery rattle of the empty carton as it turns over, a soft page turn from the couch, and a warm burst of live studio-audience laughter that begins as the carton flips and settles under the end of the shot.\n\nnon_diegetic_music:\nN/A",
+        "shot_2_pat_deflects": "subject_definitions:\n<Subject 1> is the heavyset man in <Picture 1>, mid 40s, shaved head, calm and heavy-lidded, open gray flannel shirt over a dark t-shirt.\n\nsummary:\n[reference generation] The target video is one close-up shot of a 1990s multi-camera sitcom: <Subject 1> answers an accusation without looking up from his magazine, in front of a live studio audience.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, head, flannel shirt and proportions of <Picture 1> are kept exactly, with the head and jaw newly animated to speak.\n\ndetailed_description:\nThe target video is in the style of a 1990s multi-camera sitcom, shot on studio video with warm, even, shadowless lighting. [Shot 1] A medium close-up frames <Subject 1> from the chest up, seated at the end of an olive-green couch in front of a brick wall, holding an open magazine just below the frame line and reading it. He does not look up. <Subject 1> (S1), a flat, low, unhurried male baritone voice, completely unbothered, says <d>[English] Technically? The freezer has a defrost cycle.</d> On the last word he turns a page. A live studio audience laughs as the line lands. The camera holds a static close-up throughout, the way a sitcom cross-shot camera holds a speaker.\n\noverall_soundscape:\nA bright, close television-studio room tone. A soft page turn, the faint creak of couch springs as <Subject 1> settles, and a rolling wave of live studio-audience laughter that starts on the last word of the line and carries under the end of the shot.\n\nnon_diegetic_music:\nN/A",
+        "shot_3_howie_incredulous": "subject_definitions:\n<Subject 1> is the wiry man in <Picture 1>, early 30s, short curly dark hair, expressive eyebrows, mustard-yellow sweater vest over a white button-down shirt.\n\nsummary:\n[reference generation] The target video is one close-up shot of a 1990s multi-camera sitcom: <Subject 1> repeats an absurd excuse back in disbelief, holding up an empty ice cream carton, in front of a live studio audience.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, hair, sweater vest and proportions of <Picture 1> are kept exactly, with the head, jaw and arm newly animated to speak and gesture.\n\ndetailed_description:\nThe target video is in the style of a 1990s multi-camera sitcom, shot on studio video with warm, even, shadowless lighting. [Shot 1] A medium close-up frames <Subject 1> from the chest up, standing beside a couch in front of a brick wall, an empty pint-sized ice cream carton raised in one hand at head height. His eyebrows are high and his mouth is half open before he speaks. <Subject 1> (S1), a fast, nasal, high-pitched male tenor voice, tight with disbelief, cracking upward on the last word, says <d>[English] The defrost cycle used a spoon?</d> He shakes the empty carton once on the word spoon. A live studio audience erupts in a big laugh as the line lands. The camera holds a static close-up throughout, the way a sitcom cross-shot camera holds a speaker.\n\noverall_soundscape:\nA bright, close television-studio room tone. The hollow papery rattle of the empty carton as it shakes, and a big rising wave of live studio-audience laughter that breaks on the word spoon and rolls under the end of the shot.\n\nnon_diegetic_music:\nN/A",
+        "shot_4_pat_button": "subject_definitions:\n<Subject 1> is the heavyset man in <Picture 1>, mid 40s, shaved head, calm and heavy-lidded, open gray flannel shirt over a dark t-shirt.\n\nsummary:\n[reference generation] The target video is one close-up shot of a 1990s multi-camera sitcom: <Subject 1> shrugs off the accusation with a deadpan line, in front of a live studio audience.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, head, flannel shirt and proportions of <Picture 1> are kept exactly, with the head, jaw and shoulders newly animated to speak and shrug.\n\ndetailed_description:\nThe target video is in the style of a 1990s multi-camera sitcom, shot on studio video with warm, even, shadowless lighting. [Shot 1] A medium close-up frames <Subject 1> from the chest up, seated at the end of an olive-green couch in front of a brick wall, the magazine now closed in his lap. He looks up for the first time, holds a beat of eye contact just past the lens, and gives one slow shrug. <Subject 1> (S1), a flat, low, unhurried male baritone voice, perfectly deadpan, says <d>[English] I don't know its life.</d> He goes back to the magazine. A live studio audience gives a big laugh with scattered applause as he looks back down. The camera holds a static close-up throughout, the way a sitcom cross-shot camera holds a speaker.\n\noverall_soundscape:\nA bright, close television-studio room tone. The soft slap of the magazine reopening, a couch-spring creak under the shrug, and a big wave of live studio-audience laughter with a few claps in it that breaks after the line and rolls under the end of the shot.\n\nnon_diegetic_music:\nN/A",
         "shot_5_tag": "subject_definitions:\n<Subject 1> is the wiry man in <Picture 1>, early 30s, short curly dark hair, mustard-yellow sweater vest over a white button-down shirt.\n<Subject 2> is the heavyset man in <Picture 2>, mid 40s, shaved head, open gray flannel shirt over a dark t-shirt.\n\nsummary:\n[reference generation] The target video is one shot of a 1990s multi-camera sitcom: <Subject 1> gives up and sits down next to <Subject 2>, the two trade a final exchange, and the scene ends on applause.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, hair, sweater vest and proportions of <Picture 1> are kept exactly, with the body newly animated to sit and speak.\n<Subject 2> (appears in [Shot 1]): fully_preserved - the face, head, flannel shirt and proportions of <Picture 2> are kept exactly, with the head newly animated to reply.\n\ndetailed_description:\nThe target video is in the style of a 1990s multi-camera sitcom, shot on studio video with warm, even, shadowless lighting. [Shot 1] A wide shot frames the living room set: the sagging olive-green couch centre frame in front of a brick wall with a window onto a painted city backdrop. <Subject 2> sits at the right end of the couch reading his magazine. <Subject 1> exhales, drops the empty carton onto the coffee table, and sits down heavily at the left end of the couch, staring straight ahead. A beat of silence. <Subject 1> (S1), a fast, nasal, high-pitched male tenor voice, now flat with defeat, says <d>[English] We're getting a padlock.</d> Without looking up, <Subject 2> (S2), a flat, low, unhurried male baritone voice, says <d>[English] Sure.</d> The studio audience laughs and applauds as the shot holds on the two of them side by side, neither moving. The camera holds a static wide shot throughout.\n\noverall_soundscape:\nA bright, close television-studio room tone. The hollow knock of the carton landing on the coffee table, a heavy couch-spring compression as <Subject 1> sits, a page turn, then a full wave of live studio-audience laughter that turns into warm applause and carries to the end of the shot.\n\nnon_diegetic_music:\nA short, bright sitcom scene-change sting - a bass guitar walk-down with a light drum hit - enters under the applause in the final second of the shot.",
         "subject_reference_type": "diffusers.modular_pipelines.minimax_h3.MiniMaxH3ImageReference",
         "num_frames": 124,
diff --git a/workflows/templates/minimax/generated-subject-reference.json b/workflows/templates/minimax/generated-subject-reference.json
index 05e77565..051dfa55 100644
--- a/workflows/templates/minimax/generated-subject-reference.json
+++ b/workflows/templates/minimax/generated-subject-reference.json
@@ -4,7 +4,7 @@
     "summary": "H3 video of a character Z-Image draws first, referenced in-workflow rather than from a file.",
     "variables": {
         "subject_prompt": "prompt:zimage/deep_sea_diver_portrait",
-        "prompt": "subject_definitions:\n<Subject 1> is the deep-sea diver in <Picture 1>, in a heavy canvas suit and a brass helmet with the faceplate open, whose face, helmet, suit and proportions define their appearance for the whole video.\n\nsummary:\n[reference generation] The target video shows <Subject 1> speaking directly to camera in the setting of <Picture 1>, in one continuous take.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, helmet, suit and proportions of <Picture 1> are kept exactly, with the head, jaw and shoulders newly animated to speak.\n\ndetailed_description:\nThe target video is in realistic photographic style. [Shot 1] A medium close-up frames <Subject 1> from the chest up, centred and facing camera in the setting of <Picture 1>, held as one continuous shot. The lighting of <Picture 1> is kept exactly as it is, the warm rim light steady along the brass of the helmet and the background staying dark and well behind in shallow focus. <Subject 1> (S1), a low and gravelled male voice, unhurried and worn, says <d>[English] Down there the light gives out first. Everything else follows.</d> The jaw and lips move in natural synchronisation with the speech and settle closed at the end of the line. Between phrases <Subject 1> shifts weight very slightly and the head drifts a few degrees off centre and back, with the eyes staying on the lens and blinking at a natural rate. The camera holds a static shot throughout.\n\noverall_soundscape:\nA quiet, close room tone at an even level from the first frame to the last, with a faint creak of canvas as <Subject 1> shifts weight and a single soft knock of a glove against the brass collar.\n\nnon_diegetic_music:\nNone.",
+        "prompt": "subject_definitions:\n<Subject 1> is the deep-sea diver in <Picture 1>, in a heavy canvas suit and a brass helmet with the faceplate open, whose face, helmet, suit and proportions define their appearance for the whole video.\n\nsummary:\n[reference generation] The target video shows <Subject 1> speaking directly to camera in the setting of <Picture 1>, in one continuous take.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, helmet, suit and proportions of <Picture 1> are kept exactly, with the head, jaw and shoulders newly animated to speak.\n\ndetailed_description:\nThe target video is in realistic photographic style. [Shot 1] A medium close-up frames <Subject 1> from the chest up, centred and facing camera in the setting of <Picture 1>, held as one continuous shot. The lighting of <Picture 1> is kept exactly as it is, the warm rim light steady along the brass of the helmet and the background staying dark and well behind in shallow focus. <Subject 1> (S1), a low and gravelled male voice, unhurried and worn, says <d>[English] Down there the light gives out first. Everything else follows.</d> The jaw and lips move in natural synchronisation with the speech and settle closed at the end of the line. Between phrases <Subject 1> shifts weight very slightly and the head drifts a few degrees off centre and back, with the eyes staying on the lens and blinking at a natural rate. The camera holds a static shot throughout.\n\noverall_soundscape:\nA quiet, close room tone at an even level from the first frame to the last, with a faint creak of canvas as <Subject 1> shifts weight and a single soft knock of a glove against the brass collar.\n\nnon_diegetic_music:\nN/A",
         "num_frames": 124,
         "width": 960,
         "height": 544,
diff --git a/workflows/templates/minimax/music-video.json b/workflows/templates/minimax/music-video.json
index 68db67e0..b21a7435 100644
--- a/workflows/templates/minimax/music-video.json
+++ b/workflows/templates/minimax/music-video.json
@@ -5,11 +5,11 @@
         "singer_portrait_prompt": "prompt:zimage/otter_singer_portrait",
         "song_prompt": "prompt:minimax/otter_soul_song",
         "song_lyrics": "[verse]\nI've been waiting on you darling like a render overnight\nEvery frame of my devotion coming slowly into light\n[chorus]\nOoh, progress bar of love\nOoh, don't you stall on me now",
-        "audio_duration": 21,
-        "shot_1_wide_open": "subject_definitions:\n<Subject 1> is the anthropomorphic river otter lounge singer in <Picture 1>, in a burgundy velvet dinner jacket and black bow tie, whose face, fur, jacket and proportions define his appearance for the whole video.\n<Audio 1> is the song <Subject 1> (S1) is lip-synchronised to: a smooth male crooner over a retro soul arrangement.\n\nsummary:\n[reference generation + audio reuse] The target video shows <Subject 1> performing the song carried by <Audio 1> at a vintage microphone on a nightclub stage, in one continuous wide shot.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, fur, velvet jacket, bow tie and proportions of <Picture 1> are kept exactly, with the body newly animated to sing.\n<Audio 1>: partially_copy - the piece of the song covering this part of the video is reused as the soundtrack, and <Subject 1>'s mouth is synchronised to its vocal.\n\ndetailed_description:\nThe target video is in realistic photographic style. [Shot 1] A wide shot frames a small nightclub stage: <Subject 1> stands centred at a vintage chrome ribbon microphone under a single warm spotlight, deep blue darkness and the faint silhouettes of a band behind him. He sings the vocal of <Audio 1>, his mouth moving in natural synchronisation with it, one paw resting on the microphone stand and the other loose at his side, swaying gently with the tempo. His tail sways in counterpoint behind him. The camera holds a static wide shot throughout, the stage lights unchanging.\n\noverall_soundscape:\nThe song of <Audio 1> carries the shot. Under it, a low nightclub room tone, the soft clink of glassware far off, and the faint brush of <Subject 1>'s jacket as he sways.\n\nnon_diegetic_music:\nNone - the music is the diegetic performance of <Audio 1>.",
-        "shot_2_closeup": "subject_definitions:\n<Subject 1> is the anthropomorphic river otter lounge singer in <Picture 1>, in a burgundy velvet dinner jacket and black bow tie, whose face, fur, jacket and proportions define his appearance for the whole video.\n<Audio 1> is the song <Subject 1> (S1) is lip-synchronised to: a smooth male crooner over a retro soul arrangement.\n\nsummary:\n[reference generation + audio reuse] The target video shows <Subject 1> crooning the song carried by <Audio 1> in close-up at a vintage microphone, in one continuous shot.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, fur, velvet jacket, bow tie and proportions of <Picture 1> are kept exactly, with the head and jaw newly animated to sing.\n<Audio 1>: partially_copy - the piece of the song covering this part of the video is reused as the soundtrack, and <Subject 1>'s mouth is synchronised to its vocal.\n\ndetailed_description:\nThe target video is in realistic photographic style. [Shot 1] A close-up frames <Subject 1> from the shoulders up, angled slightly off centre, the vintage chrome microphone in the lower corner of frame and the warm spotlight modelling his fur against deep blue darkness. He sings the vocal of <Audio 1> with his eyes half closed, whiskers catching the light, his mouth in natural synchronisation with the vocal, leaning a breath closer to the microphone on the held notes and easing back between phrases. The camera holds a static close-up throughout.\n\noverall_soundscape:\nThe song of <Audio 1> carries the shot, the vocal close and warm at this distance. Under it, a low nightclub room tone and the faint creak of the microphone stand as he leans in.\n\nnon_diegetic_music:\nNone - the music is the diegetic performance of <Audio 1>.",
-        "shot_3_room": "subject_definitions:\n<Subject 1> is the anthropomorphic river otter lounge singer in <Picture 1>, in a burgundy velvet dinner jacket and black bow tie, whose face, fur, jacket and proportions define his appearance for the whole video.\n<Audio 1> is the song <Subject 1> (S1) is lip-synchronised to: a smooth male crooner over a retro soul arrangement.\n\nsummary:\n[reference generation + audio reuse] The target video shows <Subject 1> performing the song carried by <Audio 1> seen across the tables of the nightclub, in one continuous shot.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, fur, velvet jacket, bow tie and proportions of <Picture 1> are kept exactly, with the body newly animated to sing and gesture.\n<Audio 1>: partially_copy - the piece of the song covering this part of the video is reused as the soundtrack, and <Subject 1>'s mouth is synchronised to its vocal.\n\ndetailed_description:\nThe target video is in realistic photographic style. [Shot 1] A medium-wide shot looks across the nightclub toward the stage from table height: candlelit tables blur softly in the foreground, and beyond them <Subject 1> stands at the vintage microphone in his spotlight, singing the vocal of <Audio 1> with his mouth in natural synchronisation. On the swell of the hook he lifts his free paw in a slow, open gesture and lets it settle back to the microphone stand. The camera holds a static shot from among the tables throughout, the candle flames flickering gently.\n\noverall_soundscape:\nThe song of <Audio 1> carries the shot with the slight distance and room reverb of the back of the club. Under it, the close clink of glasses at the foreground tables and a low murmur that hushes as the hook lands.\n\nnon_diegetic_music:\nNone - the music is the diegetic performance of <Audio 1>.",
-        "shot_4_finale": "subject_definitions:\n<Subject 1> is the anthropomorphic river otter lounge singer in <Picture 1>, in a burgundy velvet dinner jacket and black bow tie, whose face, fur, jacket and proportions define his appearance for the whole video.\n<Audio 1> is the song <Subject 1> (S1) is lip-synchronised to: a smooth male crooner over a retro soul arrangement.\n\nsummary:\n[reference generation + audio reuse] The target video shows <Subject 1> bringing the song carried by <Audio 1> home, holding the last note with his arms spread wide, in one continuous shot.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, fur, velvet jacket, bow tie and proportions of <Picture 1> are kept exactly, with the body newly animated to sing.\n<Audio 1>: partially_copy - the piece of the song covering this part of the video is reused as the soundtrack, and <Subject 1>'s mouth is synchronised to its vocal.\n\ndetailed_description:\nThe target video is in realistic photographic style. [Shot 1] A wide shot frames the nightclub stage straight on: <Subject 1> at the vintage chrome microphone under his spotlight, deep blue darkness behind. He sings the vocal of <Audio 1> in natural synchronisation, and as the last line arrives he tips his head back, spreads both arms wide from the microphone, and holds the final note with his eyes closed, his tail still, the spotlight warming slightly around him. He brings his paws gently back together on the stand as the note ends. The camera holds a static wide shot throughout.\n\noverall_soundscape:\nThe song of <Audio 1> carries the shot. Under it, a low nightclub room tone that goes attentively quiet on the held note, and the soft brush of velvet as his arms spread.\n\nnon_diegetic_music:\nNone - the music is the diegetic performance of <Audio 1>.",
+        "audio_duration": 30,
+        "shot_1_wide_open": "subject_definitions:\n<Subject 1> is the anthropomorphic river otter lounge singer in <Picture 1>, in a burgundy velvet dinner jacket and black bow tie, whose face, fur, jacket and proportions define his appearance for the whole video.\n<Audio 1> is the song <Subject 1> (S1) is lip-synchronised to: a smooth male crooner over a retro soul arrangement.\n\nsummary:\n[reference generation + audio reuse] The target video shows <Subject 1> performing the song carried by <Audio 1> at a vintage microphone on a nightclub stage, in one continuous wide shot.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, fur, velvet jacket, bow tie and proportions of <Picture 1> are kept exactly, with the body newly animated to sing.\n<Audio 1>: partially_copy - the piece of the song covering this part of the video is reused as the soundtrack, and <Subject 1>'s mouth is synchronised to its vocal.\n\ndetailed_description:\nThe target video is in realistic photographic style. [Shot 1] A wide shot frames a small nightclub stage: <Subject 1> stands centred at a vintage chrome ribbon microphone under a single warm spotlight, deep blue darkness and the faint silhouettes of a band behind him. He sings the vocal of <Audio 1>, his mouth moving in natural synchronisation with it, one paw resting on the microphone stand and the other loose at his side, swaying gently with the tempo. His tail sways in counterpoint behind him. The camera holds a static wide shot throughout, the stage lights unchanging.\n\noverall_soundscape:\nThe song of <Audio 1> carries the shot. Under it, a low nightclub room tone, the soft clink of glassware far off, and the faint brush of <Subject 1>'s jacket as he sways.\n\nnon_diegetic_music:\nN/A",
+        "shot_2_closeup": "subject_definitions:\n<Subject 1> is the anthropomorphic river otter lounge singer in <Picture 1>, in a burgundy velvet dinner jacket and black bow tie, whose face, fur, jacket and proportions define his appearance for the whole video.\n<Audio 1> is the song <Subject 1> (S1) is lip-synchronised to: a smooth male crooner over a retro soul arrangement.\n\nsummary:\n[reference generation + audio reuse] The target video shows <Subject 1> crooning the song carried by <Audio 1> in close-up at a vintage microphone, in one continuous shot.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, fur, velvet jacket, bow tie and proportions of <Picture 1> are kept exactly, with the head and jaw newly animated to sing.\n<Audio 1>: partially_copy - the piece of the song covering this part of the video is reused as the soundtrack, and <Subject 1>'s mouth is synchronised to its vocal.\n\ndetailed_description:\nThe target video is in realistic photographic style. [Shot 1] A close-up frames <Subject 1> from the shoulders up, angled slightly off centre, the vintage chrome microphone in the lower corner of frame and the warm spotlight modelling his fur against deep blue darkness. He sings the vocal of <Audio 1> with his eyes half closed, whiskers catching the light, his mouth in natural synchronisation with the vocal, leaning a breath closer to the microphone on the held notes and easing back between phrases. The camera holds a static close-up throughout.\n\noverall_soundscape:\nThe song of <Audio 1> carries the shot, the vocal close and warm at this distance. Under it, a low nightclub room tone and the faint creak of the microphone stand as he leans in.\n\nnon_diegetic_music:\nN/A",
+        "shot_3_room": "subject_definitions:\n<Subject 1> is the anthropomorphic river otter lounge singer in <Picture 1>, in a burgundy velvet dinner jacket and black bow tie, whose face, fur, jacket and proportions define his appearance for the whole video.\n<Audio 1> is the song <Subject 1> (S1) is lip-synchronised to: a smooth male crooner over a retro soul arrangement.\n\nsummary:\n[reference generation + audio reuse] The target video shows <Subject 1> performing the song carried by <Audio 1> seen across the tables of the nightclub, in one continuous shot.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, fur, velvet jacket, bow tie and proportions of <Picture 1> are kept exactly, with the body newly animated to sing and gesture.\n<Audio 1>: partially_copy - the piece of the song covering this part of the video is reused as the soundtrack, and <Subject 1>'s mouth is synchronised to its vocal.\n\ndetailed_description:\nThe target video is in realistic photographic style. [Shot 1] A medium-wide shot looks across the nightclub toward the stage from table height: candlelit tables blur softly in the foreground, and beyond them <Subject 1> stands at the vintage microphone in his spotlight, singing the vocal of <Audio 1> with his mouth in natural synchronisation. On the swell of the hook he lifts his free paw in a slow, open gesture and lets it settle back to the microphone stand. The camera holds a static shot from among the tables throughout, the candle flames flickering gently.\n\noverall_soundscape:\nThe song of <Audio 1> carries the shot with the slight distance and room reverb of the back of the club. Under it, the close clink of glasses at the foreground tables and a low murmur that hushes as the hook lands.\n\nnon_diegetic_music:\nN/A",
+        "shot_4_finale": "subject_definitions:\n<Subject 1> is the anthropomorphic river otter lounge singer in <Picture 1>, in a burgundy velvet dinner jacket and black bow tie, whose face, fur, jacket and proportions define his appearance for the whole video.\n<Audio 1> is the song <Subject 1> (S1) is lip-synchronised to: a smooth male crooner over a retro soul arrangement.\n\nsummary:\n[reference generation + audio reuse] The target video shows <Subject 1> bringing the song carried by <Audio 1> home, holding the last note with his arms spread wide, in one continuous shot.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the face, fur, velvet jacket, bow tie and proportions of <Picture 1> are kept exactly, with the body newly animated to sing.\n<Audio 1>: partially_copy - the piece of the song covering this part of the video is reused as the soundtrack, and <Subject 1>'s mouth is synchronised to its vocal.\n\ndetailed_description:\nThe target video is in realistic photographic style. [Shot 1] A wide shot frames the nightclub stage straight on: <Subject 1> at the vintage chrome microphone under his spotlight, deep blue darkness behind. He sings the vocal of <Audio 1> in natural synchronisation, and as the last line arrives he tips his head back, spreads both arms wide from the microphone, and holds the final note with his eyes closed, his tail still, the spotlight warming slightly around him. He brings his paws gently back together on the stand as the note ends. The camera holds a static wide shot throughout.\n\noverall_soundscape:\nThe song of <Audio 1> carries the shot. Under it, a low nightclub room tone that goes attentively quiet on the held note, and the soft brush of velvet as his arms spread.\n\nnon_diegetic_music:\nN/A",
         "audio_reference_type": "diffusers.modular_pipelines.minimax_h3.MiniMaxH3AudioReference",
         "image_reference_type": "diffusers.modular_pipelines.minimax_h3.MiniMaxH3ImageReference",
         "sample_rate": 44100,
diff --git a/workflows/templates/minimax/music.json b/workflows/templates/minimax/music.json
index 6349ec10..6d5cf519 100644
--- a/workflows/templates/minimax/music.json
+++ b/workflows/templates/minimax/music.json
@@ -1,6 +1,6 @@
 {
     "id": "MiniMaxMusic",
-    "description": "Audio-only generation: a music track from a prompt, showing that a pipeline's output need not be visual. The step declares an 'audios' output and an audio result type, and the engine encodes it at the given sample rate. Also the minimal way to run a modular pipeline: rather than placing components by hand, a 'components_manager' with auto CPU offload takes ownership of device placement and keeps only the running components resident, so no quantization or per-component offload configuration is needed. 'audio_duration' is a ceiling, not a target: the model stops where the music stops, so ask for more time than the song needs - a value trimmed to the intended length cuts the outro off mid-decay - and finish the tail with 'tasks/TrimFadeAudio.json'. The result declares 'sample_rate' because a modular pipeline's dict output carries none of its own.",
+    "description": "Audio-only generation: a music track from a prompt, showing that a pipeline's output need not be visual. The step declares an 'audios' output and an audio result type, and the engine encodes it at the given sample rate. Also the minimal way to run a modular pipeline: rather than placing components by hand, a 'components_manager' with auto CPU offload takes ownership of device placement and keeps only the running components resident, so no quantization or per-component offload configuration is needed. 'audio_duration' is a ceiling, not a target: the model stops where the music stops, so ask for more time than the song needs - a value trimmed to the intended length cuts the outro off mid-decay - and finish the tail with 'templates/audio-trim-fade.json'. The result declares 'sample_rate' because a modular pipeline's dict output carries none of its own.",
     "cost": [
         {"device": "cuda", "name": "RTX 3090", "vram_gb": 24, "minutes": 3.6}
     ],
diff --git a/workflows/templates/minimax/storyboard.json b/workflows/templates/minimax/storyboard.json
index af184645..1e1d8dbc 100644
--- a/workflows/templates/minimax/storyboard.json
+++ b/workflows/templates/minimax/storyboard.json
@@ -1,6 +1,9 @@
 {
     "id": "MiniMaxH3Storyboard",
     "description": "One generation, three shots, from a storyboard of stills. H3 packs every reference into the same sequence it generates, so a request can hand it several images and say what each one is for: Z-Image draws three boards, and the 'ref2va' prompt makes <Picture 1> the first frame of [Shot 1] and <Picture 2> and <Picture 3> storyboard anchors for the cuts at 00:02.700 and 00:05.400. Because the cuts happen inside one in-context generation rather than between separate runs, the soundtrack - a piano score - runs unbroken across them, which no concat of separate shots can give. The boards are drawn by one Z-Image pipeline (identical configurations share a load) and 'release_pipeline' on the last frees it before H3 loads. 'reference_image_short_edge' is lowered from its 2048 default so three references fit alongside a 192-frame clip on a 24GB card.",
+    "cost": [
+        {"device": "cuda", "name": "RTX 3090", "vram_gb": 24, "minutes": 10.1}
+    ],
     "variables": {
         "board_1_launch": "close-up at puddle level on a rain-wet cobblestone street: a child's small hand in a yellow raincoat sleeve setting a folded white paper boat onto a wide rain puddle, raindrops ringing the water, blurred umbrellas and shopfront lights reflected in it, overcast soft daylight, photographic, sharp focus on the boat",
         "board_2_gutter": "low angle photograph of a folded white paper boat racing along a rushing rain-swollen stone gutter between cobblestones, white water splashing around its bow, wet grey street and blurred passing legs beyond, rainy afternoon light, photographic, sharp focus on the boat",