Size follows the model: reconcile dimension options on a model swap

[mikkel]

The bug

Swap an image node from nano-banana-2 to qwen-image-3 and the size dropdown keeps 2k — a size qwen-image-3 does not offer.

Both engines kept the old value and injected it as an option of the new model:

before fix, after the swap:  ["2k","auto","1024x1024","512x512","768x1024","576x1024","1024x768","1024x576"]  (selected: 2k)

The next run then posts size=2k to a model that rejects it or silently substitutes its own — a paid request for something the user did not ask for.

Same class of bug across every dimension, because each model names and values them differently:

model	dimension	real options
nano-banana-2	size	1k 2k 4k
qwen-image-3	size	auto 1024x1024 512x512 768x1024 …
Sora	length	seconds: 4 8 12
most video models	length	duration: 5 10
Sora / WAN	aspect	orientation: landscape portrait

The fix

On a model swap the stored value snaps to a real option of the new model:

1. same picture — aspect ratio first (a portrait never becomes a landscape), pixel count breaks the tie
2. else the nearest number (duration 5 → Sora's 4; 720p → 480p)
3. else the model's own default

1k → 1024x1024. 1024x1024 → 1k. 16:9 → landscape.

On hydration nothing changes: opening a saved or shared graph leaves the author's stored value exactly as chosen, because there the model did not change. Same in an exported app — the author's baked-in value survives the first render.

Disclosure, not silence

A changed dimension changes what the run buys, so it is announced, like a trimmed prompt is (#369):

⚠ qwen-image-3 doesn't offer that — switched size 2k → 1024x1024

Localized in all 5 languages. In an exported app the new value lands in the same <select> the user just used and the run-cost chip updates with it.

Verification

scripts/check-dim-reconcile.mjs (wired into pre-commit) pins, offline and with zero API spend:

• 15 mappings against real catalog shapes, in both engines
• editor ↔ play parity on every one of them
• every reconciled value is an option the new model actually lists
• the wiring: both model-swap sites reconcile, catalog arrival and first render do not
• the disclosure exists

Negative-tested — the guard fails when the fix is reverted.

Driven end-to-end in real Edge over CDP against the live catalog payload, zero API spend:

✓ nano-banana-2 offers only its 3 tiers — {"opts":["1k","2k","4k"],"value":"2k"}
✓ size options are exactly qwen-image-3's
✓ no stale 2k option survives
✓ size reconciled to 1024x1024
✓ the change is disclosed in a toast — qwen-image-3 doesn’t offer that — switched size 2k → 1024x1024
✓ the graph on disk carries the new size
✓ back on nano-banana-2 the options are the tiers again
✓ 1024x1024 reconciled to 1k

Full pre-commit suite passes.

🤖 Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	nanoodle	33ec51f	Commit Preview URL Branch Preview URL	Jul 27 2026, 02:07 PM