长会话的上下文延续

较长会话中早期内容的处理方式。

Markdown 原文Raw markdown

在较长的会话中，模型每次能够参考的内容是有限的。为支持长任务持续进行，Bi 会自动处理历史内容：

最近几轮完整保留
更早的内容被整理成一段摘要，包含你的要点、已经得到的结论以及改动过的文件
过于久远的内容会被舍弃

整理摘要不会额外消耗模型额度。

这意味着什么

会话越长，越早的细节就越容易被概括掉。如果某段信息仍然重要，可以在后续会话中重新说明，或者新开一个会话专门处理它。

上下文预算

每轮提示词有硬上限。超出时，旧的工具结果会收缩成占位，避免上下文无限膨胀；这与上面的历史摘要是同一套预算机制的两面。取舍是明确的：少读无关内容，把预算留给当前任务。

是否需要设置

这一处理自动进行，默认设置对大多数任务都合适，通常无需手动调整。

Long conversations

How earlier content is handled automatically in a long session.

Markdown 原文Raw markdown

In a long conversation, the model can only reference a limited amount of content per turn. To keep the session going, Bi manages history automatically:

The most recent turns are kept intact
Earlier content is condensed into a summary that includes your key points, conclusions reached and files changed
Very old content is dropped

Generating the summary does not consume extra model quota.

What this means

The longer the conversation, the more likely early details are condensed away. If something still matters, restate it later or start a dedicated session for it.

Context budget

Every turn has a hard prompt cap. When it is exceeded, stale tool results contract into placeholders so context cannot grow without bound. This is the other side of the same budget mechanism as history summarization: read less irrelevant material and spend the budget on the current task.

Do you need to configure it?

The handling is automatic. Defaults suit most tasks and rarely need adjustment.