List run screenshots

Scope: runs:read. Free. The model-input frames for a run, oldest first — the exact images the agent saw before each decision, which is what you want when reconstructing why it did something. Captured for every run, managed and BYOK alike.

Metadata only by default: a single frame is several hundred KB of base64, so pass include_image=true to inline image_b64 (page size is clamped to 10 in that mode, and the response is Cache-Control: no-store because frames can show whatever was on screen).

Page with after_index. index is flat and monotonic across the WHOLE run and is the stable address for a frame: a run that was reaped and retried starts a new attempt whose step counter restarts at 1, so step alone is not unique — attempt tells you which try a frame belongs to. A frame whose stored bytes cannot be decoded is returned with image_unavailable: true rather than failing the page.