What's this about?
A pull request in OpenAI's Codex repository quietly reduced the context window of the Codex CLI from 372k to 272k tokens — a 27% cut, discovered by users reading the diff rather than any announcement. The same commit adds new safety language to Codex's system prompt: before any destructive action, resolve the exact targets with read-only checks and don't touch `$HOME`.
What exactly is a "context window"?
The context window is how much text the model can hold in working memory at once — your instructions, the code it read, the conversation so far — measured in tokens (roughly word-fragments; 272k tokens ≈ a 600-page book). When a session outgrows the window, the tool must compact: summarize older history to make room, losing detail in the process. A smaller window means compaction kicks in sooner, which is exactly what users are complaining about.
What's the community saying?
- Compaction loses too much — the top thread: summaries drop the design discussion details that long coding sessions depend on; "the level of detail you lose across compaction is wildly too much."
- The opposite camp: big contexts were an illusion anyway — several argue models get noticeably dumber past ~300k tokens and the cut just codifies reality; some Codex users say they never notice the limit at all.
- Why is this undocumented? — frustration that context sizes ship in unannounced code changes, discoverable only via X posts and Reddit threads.
- The guardrail addition — the new destructive-action rules ("resolve exact targets with read-only checks first") got their own sub-thread as a sign of agents being hardened against rm-rf-style accidents.
- Cost speculation — with compute strained industry-wide, a 100k-token haircut also reads as an inference-cost decision, with pointers to DeepSeek's cheap KV-cache tricks as the alternative.