browser-replay

Locate the session's RVF container, load its trajectory.ndjson file, open a fresh browser at the original or overridden URL, and dispatch each recorded step through the matching browser MCP tool. The skill records a new lineage-tracked trajectory for the replay run.

When a recorded selector no longer matches, the skill queries the browser-selectors memory namespace for an embedding-similar selector with the same host and intent, then retries the step once. Repeated recovery needs signal the site should be re-recorded.

Yes, pass a --url override to target a new URL such as a staging environment, or use --mutate with JSON to inject expected content variation. This supports comparing runs across environments and forking sessions before mutation.

Network nondeterminism such as timing differences and content variation can cause steps to mismatch even when the flow works. Use --mutate to declare expected variation or pin the run to a fixture, and adjust the --tolerance threshold from its 0.85 default.

Replay relies on selector-embedding recovery rather than rrweb session replay, so heavily drifted sites produce noise. Steps needing more than one recovery retry indicate the flow needs a fresh recording rather than another replay.