#zdr

Quinn Comendantcom
2026-01-27

Two interesting points from ’s AI agent explainer:

’s compaction uses data from the ’s internal latent representation of the conversation rather than a text summary. Considering the computational expense of reprocessing tokens, this must a big efficiency win.
⁃ The client gets this data encrypted, with the keys held by OpenAI. This satisfies compliance, but I suspect also blocks model internals from reversing.

openai.com/index/unrolling-the

Screenshot of the blog post:

An early implementation of compaction required the user to manually invoke the `/compact` command, which would query the Responses API using the existing conversation plus custom instructions for. Codex used the resulting assistant message containing the summary for subsequent conversation turns.

Since then, the Responses API has evolved to support a special that performs compaction more efficiently. It returns that can be used in place of the previous `input` to continue the conversation while freeing up the context window. This list includes a special `type=compaction` item with an opaque `encrypted_content` item that preserves the model's latent understanding of the original conversation. Now, Codex automatically uses this endpoint to compact the conversation when the is exceeded.

[OpenAI Codex CLI 내부 동작 분석: 에이전트 루프와 프롬프트 캐싱 전략

OpenAI의 Codex CLI 내부 동작 분석 아티클이 공개되었습니다. 에이전트 루프(Agent Loop)의 구조, 프롬프트 구성 및 Responses API의 데이터 흐름, 성능 최적화 전략(프롬프트 캐싱, 대화 압축, 무상태 설계) 등을 심층적으로 다룹니다.

news.hada.io/topic?id=26089

#openai #codex #agentloop #promptcaching #zdr

Dank dem lieben Crazy Ape von twitch.tv/crazy_ape_tv habe ich nun die Möglichkeit, Star Wars FFG auch in Deutsch anzubieten! Er war so lieb sich den Aufwand zu machen, die Bücher die er verkaufen wollte in die Schweiz zu schicken.
#ttrpg #pnpde #StarWars #swffg #AoR #ZdR #ZeitalterDerRebellion

Client Info

Server: https://mastodon.social
Version: 2025.07
Repository: https://github.com/cyevgeniy/lmst