ChatGPT bug: final prompt at context limit silently overwrites previous state, corrupting handoffs

xtraa · reddit · 2026-09-25

A user documented a ChatGPT Web bug: when a long conversation hits the context limit, the final prompt appears to replace rather than append to the previous turn, so the model loses the latest task result. The resulting handoff doc contained a stale Git commit, and the follow-up session repeatedly failed a fail-closed HEAD check against the real repo. Author proposes explicit rejection or compaction at the context boundary.

Original post →

More from coding & agent

coding & agent channel →