`execution`¶

Runs your workflows for real and records the ComfyUI canvas as video while they run. The level that produces the results gallery.


Needs	`server` (provided by `registration`)
Default	yes
Fails the run	yes
Source	`orchestration/levels/execution.py`, `reporting/screenshot.py`

This is the level the whole tool exists for: not a green check, but a video of your nodes working on somebody else's machine.

How the video is made¶

capture_execution_frames() drives a real browser at the live server, loads the workflow, queues it, and captures at a fixed cadence of 5 fps while it executes. Frames are written as frame_%06d.png, then encoded by encode_mp4_timeline() into driver.mp4 -- timeline-accurate, meaning the mp4's playback timing matches real elapsed capture time, with the pre-run frame held across the validate/queue window. The loose frames are then dropped; the mp4 is the artifact the report plays, through a native <video> element with a real seek bar.

Before capture starts the view is prepared the same way static_capture does it: fit the graph, close panels and alerts, hide the unsaved dot. A high-quality PNG is taken after execution completes, once previews have rendered.

What lands on disk¶

Per workflow, under the run's output directory:

videos/<workflow>/driver.mp4          the canvas recording
videos/<workflow>/metadata.json       frame timings + log offsets
screenshots/<workflow>_executed.png   final frame, full quality
logs/<workflow>.log                   that workflow's log slice
logs/<workflow>_console.log           browser console
logs/<workflow>_resources.csv         RAM/VRAM/CPU samples

Per-workflow status, duration and RAM/VRAM peaks go into results.json. See what a run does for the full tree.

What it catches¶

Everything the earlier levels structurally cannot: shape and dtype errors, OOM under real allocation, models that fail to load, nodes that produce black images or empty outputs, and the actual wall-clock cost of your graph on each lane.

Zero workflows is an error¶

A pack that ships no workflows at all fails the run. It used to log one line and return PASSED with no results.json written -- a green badge for a level that executed nothing.

This is the discovery bug it was hiding: a pack whose example workflows live in a folder comfy-test does not scan contributes nothing and went green. Confirm the folder is one of the names ComfyUI itself recognises -- see what a pack looks like. execution now behaves like coverage, which has always refused to pass a vacuous 0/0.

Say it deliberately when you mean it:

skip_workflow = true under [test.<lane>] -- run the pipeline but not the workflows, on that lane only
drop execution from [test] levels -- stop asking for it at all

An empty per-accelerator selection (cpu = []) is not affected: this checks that workflows were discovered, not which ones a given lane selected.

Config¶

Key	Effect
`[test.workflows] cpu` / `cuda` / `rocm`	which workflows run on this backend
`[test.workflows] timeout`	per-workflow timeout, default 3600s
`[test] res`	capture resolution (viewport height), default 1080
`[test.<lane>] skip_workflow`	run the pipeline but not the workflows

In the default set, and a terminal level: a run ends in exactly one of static_capture, validation, execution_light and execution, chosen by what you list in [test] levels (ADR-0012).

execution¶