Do Son 
The Codex project recently merged multiple changes. These updates indicate that Codex is redesigning the context mechanism. This applies to agents handling ultra-long tasks. Currently, most agents utilize a context compression mechanism. Compression inherently means losing some critical information. Consequently, compressed sessions might decrease processing efficiency. They might even deviate from the original task direction.
Replacing Compression with New Windows
The changes implemented by the Codex team reveal a new direction. Codex will completely abandon the compression mechanism later. Instead, it will directly open entirely new windows. The model can open a new window when session content becomes too long. Then, it restores the previous state using history and note tools. This approach better suits agent tasks lasting several hours or longer.
Introducing Direct Model Invocation Tools
The pull request #27488 introduces a critical modification. Codex added new content tools strictly for direct model invocation. The model evaluates when the current context loses retention value. It can then actively request a fresh context window. This new window does not generate old conversation summaries. It also abandons previous user, assistant, or tool invocation records. Instead, it injects the initial environment, permissions, and task configurations again. The agent then resumes execution. This avoids common issues with repeatedly compressing long tasks. It prevents summary bloat, detail distortion, and token occupation.
Updating the Compression Workflows
The subsequent PR #29743 connects this specific logic. It integrates with manual and automatic compression workflows. Activating the token budget mode triggers a compression event. This event actually switches directly to a new window. It no longer asks the server to summarize history. Meanwhile, it retains the original Compact Hook and lifecycle events. This preserves crucial client compatibility.
Achieving Cross-Window Memory Storage
Directly discarding historical content introduces obvious problems. The new session window forgets previously completed actions. Therefore, the changes submitted in PR #39827 introduce new tools. The history tool lists previous windows and messages. It reads specific records and searches historical content. The note tool provides independent, persistent note-taking capabilities. It can query, append, or rewrite the current task state.
Preserving Task Continuity
The model saves goals, decisions, and progress before switching windows. It also saves subsequent planned operations. The new window then reads these persistent notes. The agent might need specific detail content later. It can call the history tool to query exact previous contents.
However, these modifications currently remain in the code merging phase. Therefore, they are temporarily unavailable for public use. These functions likely require further code submissions for completion. The exact official release date remains unknown.
Support Our Threat Intelligence
Find our zero-day alerts and CVE reports helpful? Support our work today and unlock a 100% ad-free reading experience!
Buy Me a Coffee
PayPal
