David decided: Per-document lecture audio: listen-to-this-lecture feature

David specified a feature where each workspace document would have a 'listen to this lecture' button that triggers audio generation for that document only — distinct from the full-workspace podcast feature already built:

inside of a workspace document, near the top next to the test memo objectives buttons, i want a UI element button for 'listen to this lecture' where, if clicked, it generates the podcast just for that document

He also specified the player:

when the user clicks the button and the process starts, it shows an audio generating UI state and when its ready, it shows the UI in a ready state and that the player shows up in the bottom right of the screen but the player must be a specialised one wherein its different from the type of player used for the full podcast

When the agent presented an implementation plan, David caught that it hadn't addressed failure modes:

okay wait but your plan doesnt seem to account for race conditions, failure modes, recoveries, retrying, not having UI states become stale and not actually indicative of truth etc

He then sent 6 GPT-5.4 xhigh agents to audit the implementation for exactly those gaps before shipping.