Recent ComfyUI builds (dynamic VRAM/RAM-pressure caching,
comfy_execution/caching.py RAMPressureCache) made the node outputs
cache's get() a coroutine. _ExecutionListProxy in capture.py called
it synchronously (no await), so every call silently returned an
un-awaited coroutine instead of the cached CacheEntry:
RuntimeWarning: coroutine 'RAMPressureCache.get' was never awaited
Downstream code then failed to read .outputs on the coroutine,
raising an exception that get_inputs() swallows per-node
(except Exception: continue), so metadata capture silently breaks
for essentially every node in the graph on affected ComfyUI
versions (embedded prompt/seed/sampler/model metadata ends up
missing or incomplete).
ComfyUI's own synchronous execution engine
(comfy_execution/graph.py ExecutionList.get_cache/cache_link)
never awaits this cache directly either -- it reads already
-computed outputs via the cache's synchronous get_local()
accessor. This change makes _ExecutionListProxy do the same,
falling back to the plain .get() for _FallbackOutputsCache
(a dict shim with no get_local).
Verified locally: warning disappears and Eagle-sent images
regain full metadata on ComfyUI v0.31.0.
00akii
347c9d90df