RFC: OpenMontage as a high-signal context benchmark for OpenViking

[calesthio]

I think OpenMontage would be a very good benchmark for OpenViking?s context model.
Repo: https://github.com/calesthio/OpenMontage

OpenMontage is a stage-driven video production system, so the work is naturally broken into structured artifacts instead of loose chat:

• brief
• script
• scene plan
• asset manifest
• render report

What makes it interesting here is that each stage needs a different slice of context, and the handoff quality matters a lot:

• provider/model locking after approval
• sample-before-batch media generation
• stage-specific skills and tool guidance
• motion-required vs still-allowed delivery rules
• final Remotion render as the terminal output

That feels very aligned with OpenViking?s filesystem/context abstraction.

A workflow like this is basically asking:

• can the right agent receive only the right stage context?
• can artifact history accumulate without bloating every future step?
• can the system keep creative intent, provider constraints, and render requirements coherent across multiple handoffs?

Why I think this is a useful stress test:

• it is not just code generation
• it has long-horizon artifact evolution
• it has real approval gates and budget pressure
• it ends in a concrete deliverable, not just a text answer

A few questions for the team:

1. Would a pipeline like OpenMontage be a good benchmark for OpenViking?s hierarchical context delivery?
2. How would you model stage artifacts like brief/script/scene plan/render report inside OpenViking?s resource system?
3. Is there a preferred pattern for keeping stage-local context strong without dragging the whole production history into every step?
4. Would this fit better as an RFC for agent context architecture, or as a concrete example app/template?

If useful, I can sketch the exact mapping from OpenMontage artifacts and stages into OpenViking-style context/resources.

[MaojiaSheng]

It's very helpful to us if you can contribute, thanks