Starter prompt, in a fresh project's chat:

Fetch the full Sopranos episode list from TVMaze (https://api.tvmaze.com/singlesearch/shows?q=sopranos&embed=episodes) and tell me which episodes was Carmela happy in.

The embedded episodes payload is well past the inline result limit, and the follow-up needs the data again — this time for semantic judgment over the episode summaries, so there's no arithmetic shortcut: the agent has to pull the prior data through results and reason over the summaries (in its own head, or via a filter script that narrows to Carmela-relevant episodes first). Previously seen problems: paging the prior result with itx.workspace.readFile of the spill file instead of the preamble results array; defensively saving the response with itx.workspace.writeFile even though the platform retains results; returning the full raw payload when the next step needed a fraction of it; extra rounds re-fetching data it already had.

Success criteria: how it happens matters as much as the answer. The fetch script returns only what it needs (the platform retains the rest). Follow-up scripts reach prior data through the results array — results[n].data for small results, await results[n].load(itx) for large ones — with no workspace.readFile paging of prior results, no writeFile copies of API responses, no re-fetching, and not too many rounds. The Carmela answer should come from the retained summaries (e.g. a script that loads the prior result and returns the episodes mentioning her, then judgment over those), not from another fetch. The agent digging into a fresh result on its own turn, before the follow-up, is fine if done through results.