Simon Willison's 27-minute GPT-6 Astra run produced GPX routes he cannot reproduce, because compaction ate the code

Simon Willison's Weblog·medium signalXBlueskyLinkedInCopy link

Willison had ChatGPT Work with GPT-6 Astra (Max) generate 5K and 10K loop routes from his address; it worked for 27 minutes using Nominatim for geocoding, Overpass to pull local OSM roads and trails, and the visualize skill, and returned an embedded map plus downloadable GPX and GeoJSON. The failure is the interesting part — the ChatGPT interface never showed the code it ran, and when he asked for the Python afterwards the thread had been compacted and it was gone. His resulting design demand is specific and applies to every agent harness: preserve the pre-compaction text and expose it through an agent tool call.