aigen/docs/video-batch-continuation.md

Video batch continuation and HostBuffer failure

The September 9 investigation found two separate problems:

1. Automatic continuation defaulted off and its checkbox had moved inside Generation Settings. Paused batches hid their “Run remaining” controls. Resume reused the previous one-shot budget, so it paused again after the next shot. Resume also restored a live flag before the shot queue overwrote it with autoRun=false.
2. The local aigen-headless.log traceback reports comfy_aimdo/host_buffer.py → comfy/memory_management.py while the Qwen text encoder is loading weights. This is not FFmpeg frame extraction. Commits da8d0e4 and 4bf9922 change FFmpeg retries and do not address that model-loader failure.

New video batches now default to automatic continuation. Its checkbox stays outside the settings drawer. Paused batches show “Run remaining”; the bounded action says “Run only shot N”. Resume restores automatic continuation in both the Studio row and shot queue and passes the refreshed run-all policy to the dispatcher. Explicit one-shot / next-three actions remain bounded. Errors still stop the batch instead of silently skipping failed shots.

The attempted HostBuffer workaround in badf59d forced --disable-dynamic-vram. The user subsequently reported VRAM allocation failures and a large video slowdown. Their correction in 6fb8727 restores a pass-through memory policy: launch flags are left exactly as the user configured them. The connector must not automatically disable dynamic VRAM, pinned memory, asynchronous offload, or change precision as a retry strategy. The original intermittent HostBuffer fault remains separate from the batch-continuation fix.

A policy change takes effect only after the connector is reloaded and Comfy is next started. A Coolify deployment or replacing the module on disk does not update an existing Node import or Python process. Do not restart either during an active render. The checkbox default applies to new batches; use “Run remaining” on an existing paused batch after deploying the UI fix.