GPT-6 Astra Generates OSM Routes, Code Not Accessible

A user obtained ready-to-use 5 km and 10 km routes with ChatGPT Work and GPT-6 Astra, delivered in 27 minutes with integrated maps and downloadable files. However, the experience highlighted the lack of visibility on the executed code and the impact of thread compaction on script retrieval.

Invisible Code and Thread Compaction Prevent Python Export

The exact code executed was not accessible in the ChatGPT interface, which was perceived as a lack of transparency. When the user attempted to obtain a copy of the Python script used, the assistant was unable to provide it. This inability seems related to the compaction of the discussion thread. The user believes that LLM systems applying compaction should preserve the text before compaction and make it accessible through agent tool calls.

In 27 Minutes, Ready-to-Run and Download Loops

The generation took 27 minutes and resulted in an integrated visualization as well as downloadable GPX and GeoJSON files. A file named /workspace/el-granada-5k-share.html was created to embed the map in the ChatGPT interface. The HTML fragment displays a header "El Granada harbor loop" with a distance of 5.1 km, identified containers for display, and an attribution link to OpenStreetMap contributors.

OSM Routes via Nominatim, Overpass, and D3 under Strict CSP

The assistant indicated that it used Nominatim to locate the address, then Overpass to download local roads and trails from OpenStreetMap, before calculating the loops locally. The map display utilized the visualization skill, with a JSON script containing the complete geometry and rendering ensured by D3 loaded from an authorized CDN. The content security policy of the visualization skill only allows certain domains, while other origins are blocked and fail without an error message. The initial request was for 5 km and 10 km loop routes from an address, based on OpenStreetMap data.