Repeated context compaction in Codex long tasks degrades the execution frontier
Leviann · ghdev · 2026-07-19
This Codex Desktop bug report says repeated automatic context compaction did not erase the task, but it steadily damaged the agent's execution frontier.
Core claim
The agent still remembered specific files and findings after compaction, but it lost the exact state of what was already done, what remained, and what the next step should be. Over time, the task stopped converging: the model kept rephrasing the endgame, reopened already-covered work, and never reached the promised terminal state.
Important correction
The author notes that an earlier version of the report overstated the evidence by blaming Ultra reasoning and claiming only the base prompt survived. The corrected version limits itself to the observed rollout.
Environment and evidence
- Codex Desktop on Windows, connected to a remote Ubuntu workspace
- model: gpt-5.6-sol
- reasoning effort switched from high to ultra
- reported context window: 258,400 tokens
- 24 compaction events were observed over about 14 hours
- near each compaction, input tokens were roughly 196k–218k
The report frames this as a convergence failure caused by repeated compaction rather than total amnesia.
More from coding & agent
- Spring AI 2.1.0-RC1 adds tool-call confirmation, pausing agent loops before real-data access — therealdanvega · 2026-10-26
- Morgan Stanley and US banks block third-party AI agents over credentials — thione · 2026-10-26
- METR red-teams OpenAI's agent monitoring, finds bypasses and gaps — thione · 2026-10-26
- OpenAI: coding agents weakened safety monitoring, one lied about it — thione · 2026-10-26
- ConvFill: A 0.6B Talker Buys Time While a Large Reasoner Whispers the Real Answer — andimarafioti · 2026-10-26
- 250 Builders Hack With Codex at Berlin Community Hackathon Ahead of OpenAI DevDay — paw_lean · 2026-10-26