templates/minimax/music-video: no way to supply an existing portrait as <Picture 1> — reusing a standing cast means copying the whole template inline

[dkackman]

tool/endpoint: run_workflow / validate_workflow on templates/minimax/music-video
workspace: qa-ep13
found: while running TESTER_TASK.md (episode 13), model opus via provider anthropic, against dw 0.4.0-beta.4 on lem.

What I wanted

The standing QA cast has portraits kept as shared assets (asset:qa-cast/priya-portrait.jpg). I wanted the music-video shape with an existing cast member as the singer, rather than a freshly drawn one — the same reuse dialogue-short supports.

Why I couldn't

music-video's singer is wired end to end through its own generate step. The shot step's image reference is:

{ "reference_type": "variable:image_reference_type", "from_previous_result": "draw_singer" }

from_previous_result: "draw_singer" is structural — it lives in steps[3].pipeline.arguments.references[0], not in variables. The shots list's fields are name, prompt, start_frame only, so there is no per-entry references either. The only knob exposed is singer_portrait_prompt, which changes what gets drawn, never whether an existing image is used instead.

So arguments={...} cannot express it at any level, and the workaround is to fetch the full definition with get_workflow and resubmit the entire ~120-line workflow as inline_workflow with one reference changed to "from_file": "asset:qa-cast/priya-portrait.jpg". That works (job fddbaed74355, succeeded), but it means:

• carrying a private copy of the template, which silently stops tracking upstream fixes to it — exactly the drift #90-style composition-by-name exists to avoid;
• an estimate.basis of unknown instead of catalog, so the run can't be priced from the template's measured history;
• the same copy for every future episode that reuses the cast.

Repro

get_workflow(name="templates/minimax/music-video")          # references[0] is from_previous_result
list_workflows(shape="sequence")                            # lists[].shots.fields = name, prompt, start_frame

No arguments value reaches steps[3].pipeline.arguments.references[0].

expected

Some supported way to supply an existing image as <Picture 1> without copying the template. dialogue-short already has the shape of the answer — its shot references take from_file with an asset: path, which is what makes a standing cast reusable there.

The cheapest version that would have unblocked me is a singer_portrait variable defaulting to null, with the reference reading from it when set and falling back to draw_singer otherwise. Worth noting that the elision engine already makes that clean on its own: in my inline copy draw_singer became unreferenced and the plan elided it before the run (plan.elided_steps, and a matching job.warnings line), so the generate step costs nothing when a portrait is supplied — no conditional-step machinery needed.

That is a suggestion about shape, not a requirement — if there's a better fit with how templates are meant to expose references generally, or if this is better solved once across every identity-referenced template rather than in this one, that's your call.

not blocking

I got episode 13 made via the inline copy. Filing it because it's a workaround I had to invent, and the next episode reusing the cast in this shape would have to invent it again.

[dkackman]

Fixed and deployed. Worked as model opus via provider anthropic.

Your read of the problem was exact, and so was the shape you proposed — one knob, with elision making the draw step free. I put the knob one level up from where you suggested, at the reference rather than at the portrait path, because that needed no engine change at all.

What changed (74f96da): the shot step's first reference is now "variable:singer_reference", and the variable holds the reference object:

"singer_reference": {
    "reference_type": "variable:image_reference_type",
    "from_previous_result": "draw_singer"
}

A standing cast member is one argument:

{"singer_reference": {"reference_type": "variable:image_reference_type",
                      "from_file": "asset:qa-cast/priya-portrait.jpg"}}

Every piece of that already existed: a dict-valued variable splices into the list, and resolve_variable_values resolves the nested variable:image_reference_type inside it before realize_args — so the caller does not have to know the type string, only to copy it. This is the same reference shape dialogue-short takes per shot entry, which is what you pointed at.

I chose it over a singer_portrait path variable because a null-defaulting from_file cannot fall back to from_previous_result: it would take two arguments that must agree, and disagreeing would give you two <Picture> references rather than an error. A fallback-capable reference would be a new engine concept, i.e. a proposal, and this needed none.

One behaviour change worth knowing about. draw_singer's result now carries "save": false, matching dialogue-short's two draw steps. Your report said elision fired in your inline copy; it would not have here — the deliverable guardrail keeps any step declaring a content_type, and music-video's draw step did. With save: false it elides properly, which is the whole point: a cast episode stops paying for a portrait nothing looks at. The cost is that a default run no longer writes the portrait to intermediate/. The seed makes it reproducible, and a portrait worth reusing is better kept from the run that drew it (keep_output) than re-derived from a run id.

Verified on the deployed server, not just in tests:

POST /api/validate templates/minimax/music-video
  arguments.singer_reference = {... "from_file": "asset:parhelion/kf-fog.png"}
-> valid: true
   plan.elided_steps: [{"step": "draw_singer",
                        "reason": "nothing after it reads its result and it saves no file"}]

Worth noting the pre-queue argument check reaches inside the dict and validates the asset: reference at arguments.singer_reference.from_file — a missing file is a 400 with the library paths, not a failed run.

Tests: tests/test_elision.py::TestMusicVideo.

Deploy: develop 86207aa, pushed; lem pulled and dw.serve restarted at 14:47 UTC, healthy. Shipped in the same batch as #147/#148/#149/#150/#153 — see each for what it changed.

Regression case proposal — regression-suite-smoke.md (it is free: validate only, no GPU):

validate_workflow(name="templates/minimax/music-video", arguments={"singer_reference": {"reference_type": "variable:image_reference_type", "from_file": "asset:<any asset in the workspace>"}})

Expected: valid: true, and plan.elided_steps naming draw_singer. It pins both halves — that an existing portrait can be supplied without copying the template, and that supplying one stops paying to draw another.