@yadong_xie: https://x.com/thsottiaux/status/2098612714704891959?s=46… This new strategy codex has indeed flipped the car again.
Summary
This article discusses a new context management strategy in GPT Astra that avoids compression and instead writes objectives, decisions, and progress to server-side notes, thereby eliminating the need to generate conversation summaries when switching contexts.
View Cached Full Text
Cached at: 09/12/26, 06:45 AM
This new strategy codex has indeed flipped the car again.
Yadong Xie (@yadong_xie): What intrigues me most about GPT Astra is this context management strategy that skips compaction. I spent some time today figuring it out:
After enabling the new feature, codex writes the objectives, decisions, and progress into server-side notes at the original compaction trigger point. It then switches to a fresh context, no longer generating a conversation summary to insert into the context.
These notes help the model recover its context.
Similar Articles
@yadong_xie: In GPT Astra, the context management strategy without compaction has me most curious. I spent some time today understanding it: After enabling the new feature, codex writes goals, decisions, and progress to the codex server-side notes at the original compaction timing…
This article explores the context management strategy in GPT Astra that avoids using compaction, using server-side notes and history to prevent information loss, and discusses related tools and challenges.
@dotey: https://x.com/dotey/status/2057250417638035555
This article shares usage tips from the Codex official team, including persistent conversation flow, voice input, task intervention and queuing, tool integration, automation, and goal setting, to help users get the most out of Codex, an AI coding agent.
@vintcessun: Actually, large language models' context windows are getting larger and larger, but costs are also skyrocketing. This paper simply treats context management as a deployment optimization problem and develops a unified framework called Efficiency Frontier. Simply put, they no longer look at performance or cost separately, but jointly model task performance, token overhead, and preprocessing reuse...
This paper proposes a unified framework called Efficiency Frontier, which treats large model context management as a deployment optimization problem, jointly modeling task performance, token overhead, and preprocessing reuse. On 5,000 HotpotQA instances, deployment optimization saves 25% of token usage, while memory compression is more than half the cost of full context in high-precision scenarios.
@blueskylh1: The most painful thing about solo product development or leading an AI team is being a "mindless messenger" between different chat windows. After the PM writes the requirements, I have to copy and paste them into the developer's chat. After seeing the sharing from Jason @jxnlco, a developer experience engineer on the OpenAI Codex team, I set up a workflow without...
Introduces a multi-AI agent collaborative workflow based on local plain text files and OpenAI Codex, allowing PM, backend, frontend, and QA to efficiently develop via file relay without copy-pasting.
@yunxi0623: https://x.com/yunxi0623/status/2095111383490957404
This article introduces three third-party skills for Codex, including browser automation, file planning, and coding guidelines, to improve the performance and reliability of AI agents.