Simon Willison lost 27 minutes of agent work because compaction ate the code.

He had ChatGPT Work with GPT-6 Astra (Max) generate 5K and 10K running 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. Then he asked for the Python. The interface had never shown the code it ran, the thread had been compacted, and it was gone. His design demand applies to every agent harness: preserve pre-compaction text and expose it through a tool call. I've hit the same wall in my own long Claude Code sessions, and the workaround is writing artifacts to disk during the run instead of trusting the transcript.