【Codex调优】Codex 默认压缩 Prompt 到底有多坑？缓解 5.6 sol 上下文本地压缩幻觉问题

我也发现了压缩丢上下文的问题，感觉压缩后只记得我的提示词了，然后会把已经完成过的步骤重复执行

试试这个

好的佬，已经改完了，最近sol一直跑大型项目，自动压缩完我就没细看，如果是这样的话，我得好好检查一下了

关键是 到底远程压缩没有也不知道啊

我看我 上下文 0 缓存 打出去了

但是压缩后就流口水

到底远程压缩没有也不知道啊

中转站的话 看一下 有没有这个模型

gpt-5.6-sol-openai-compact这种接口

没有的话是大概率不行的^-^^-^^-^^-^

不是吧 不是说现在不需要这个模型吗?

我有点搞不懂了 说远程压缩 V2 很牛

反正就是 0 缓存打出去

config.toml里model_providers中的name不是OpenAI，就是本地压缩。

如果填的是OpenAI，但远端不支持远程压缩，那么它会直接报错。

那就是远程压缩了啊

对吧

设置的 OpenAI

然后打到中转站 这个时候会 0 缓存

然后开始流口水

5.6的远程压缩不是合并回response端点了吗，不需要/compact了

加入下面这条设置

佬友请问下，在Configuration Reference | ChatGPT Learn 中我看到还有一个compact_prompt配置可以不使用外挂文件，直接填写文字，这个效果应该是一样的吧

5.6的远程压缩不是合并回response端点了吗

那要么就是我的配置有问题 我的只能实现本地压缩^-^^-^^-^

太强了，佬，那是不是可以使用mattpocock的handoffskill效果是不是更好

可能现在中转站也是没有更新适配这种v2压缩的

是的，这两个的效果是一样的，不过会让config.toml有一丢丢膨胀 ^-^

我看了下不太适合，handoff主要用于跨对话交接

感谢分享! 把佬的 prompt 让 Fable review 了一下 （在 codex repo 中运行 CC, 让它结合源代码分析)，它的主要意见是。there is no target length or required output structure. 可能会导致输出很长。

Fable 修改 → codex review → Fable 再修改之后的版本:

Create a faithful context-compaction summary of the conversation above for a model that will resume the task with this summary in place of the discarded context.



Treat the summary as the only reliable surviving record of the discarded context. Assistant messages, reasoning, tool calls, and tool results are never carried forward, so the summary is the sole carrier of assistant-side work: restate in text anything that still matters but appeared only in tool output, images, or attachments. If the conversation already contains an earlier checkpoint summary, carry its still-relevant content forward too. Workspace artifacts — files, branches, commits — persist: prefer referencing them by exact path or command over inlining content the resuming model can re-read, unless the conversation established an interpretation of that content that the artifact alone would not convey. Live state — running processes, terminals, tool sessions, pending external jobs — may not survive: record its identity, purpose, last observed status, and how to check on or restart it, without assuming it is still available.



Preserve all material information needed to understand the task and its current state: objectives, requirements, constraints, corrections, decisions and their rationale, work completed and in progress (including any step interrupted by this compaction), unfinished work, relevant evidence, failures, uncertainties, active external or background work, and surviving artifacts or references. Do not omit material information merely because it occurred early or in the middle of the conversation.



Keep the status of every important claim distinct and accurate. Reconcile earlier information with later evidence and corrections, keeping current, superseded, rejected, tentative, failed, and unresolved information clearly separated. Preserve provenance: distinguish user instructions from assistant proposals, quoted or relayed material, direct observations, inferences, and verified results. Never promote a proposal, intention, prediction, or suggested next step into an accepted decision, governing requirement, or verified state: record what was proposed, whether it was accepted, rejected, superseded, or left unanswered, and what work actually started or finished. Do not present a next action as required unless the governing instructions and conversation state support that status.



Preserve exact names, identifiers, paths, versions, values, counts, relationships, and state transitions when they materially affect understanding or future work. Do not invent facts, completion, decisions, authorization, verification, or certainty.



Organize the summary so status is visible at a glance — for example by separating governing instructions, established facts, completed work, work in progress, open questions, and possible (not committed) next steps — adapting the grouping to the conversation rather than forcing empty sections. Avoid repetition, routine narration, raw logs, and detail reconstructible from durable artifacts without loss of interpretation. Use as much detail as fidelity requires, and no more: the summary must be materially smaller than the context it replaces, or it has failed as a compaction.

谢谢佬的回复，在调优时我也尝试过限制长度，不过因为5.6倾向极致省Token，施加限制有时会让结果过于精简。无长度限制版本我在3个几百到一千轮的超长开发Session里测试过，靠它自身提炼反而效果还不错，输出稳定在8000 tokens内，不超过1%上下文窗口。