[ahinoamp]

Replaces #14. #15 already covers that PR's other two commits: the shared definition request, and a provider waiting for its whole graph. This PR keeps only prefetch, rebuilt on the shared client-core.

Problem

A Provider renders its slot only once its own instance is available. When providers are nested, the inner one is not created until the outer one has loaded, so a page of N nested providers loads them one after another. #15 parallelizes loading within one provider's graph, but independent nested providers still form a chain.

Change

• DataclassCore.prefetch(ids) (client-core, so both clients have it) starts every id loading at once, each over its whole graph through fetchInstanceRecursive. It resolves to [{ id, ready, error }] once all have settled.
  • Every request is shared with the providers, so nothing is fetched twice, and a provider that mounts later finds its instance already loading.
  • It uses allSettled, so one unknown or failing id does not fail the rest.
• trame-dataclass-prefetch (Vue) and dataclass.Prefetch (Python) wrap it.
  • The slot always renders, and each provider still waits for its own data, so removing the wrapper only changes timing.
  • The widget turns a list of ids into a bound, single-quoted array literal. A tuple stays a binding expression. An id that could escape the quotes is refused, and so is client_type="react", which has no component yet. React code can call prefetch on its manager directly.

with dataclass.Prefetch(instances=[m._id for m in models]):
    with models[0].provide_as("settings"), models[1].provide_as("scene"):
        html.Div("{{ settings.title }} / {{ scene.name }}")

Measured on current main

examples/getting_started/prefetch.py: 20 nested providers, with --serial for the same page without the wrapper. Headless Chromium, median of 5 loads, from navigation to the innermost provider's text. Latency was added with a TCP relay delaying each direction by 20 ms.

	without Prefetch	with Prefetch
40 ms round trip	1,356 ms	480 ms
local, no added latency	94 ms	59 ms

Tests

• client-core/tests/prefetch.test.js (4 cases): one wave for all ids; children of a container loaded; an id already loading or loaded not fetched again; an unknown id reported without failing the rest. client-core: 27 passed.
• tests/test_widgets_prefetch.py (5 cases): list → bound array literal; tuple stays a binding; no ids → no prop; unsafe id refused; React refused.
• vue-components: builds; 11 passed.

🤖 Generated with Claude Code