[Bug]: v1.37.0 上下文压缩失败，上下文窗口突然爆满溢出

[erichekai]

Version line

v2 — Go rewrite (1.x), main-v2 (active development)

Exact version

1.37.0

What happened?

执行任务过程当中，上下文窗口超过设置的压缩阈值，没有自动进行压缩，最终上下文窗口占满溢出

Steps to reproduce

无复现，两个对话窗口都出现了占满溢出

OS / platform

win11

Relevant logs or output

[SivanCola]

我来看下

[SivanCola]

请问压缩之前大概占了多少？

[Linearl]

溢出以后，使用/compact不能正常压缩么？

[SivanCola]

溢出以后，使用/compact不能正常压缩么？

一般都是80%或者还没上下文满的时候压缩，溢出了就没法压缩了

[erichekai]

压缩不了，我没有盯着任务，但是两个对话貌似都是突然间溢出的，而且溢出很多，不能使用/compact命令进行压缩了，因为超出了1M，软件界面显示已经1.6M了，输入压缩命令提示：摘要失败 · 已停重试
context summary failed: deepseek: status 400: {"error":{"message":"This model's maximum context length is 1048576 tokens. However, you requested 1116302 tokens (1108110 in the messages, 8192 in the completion). Please reduce the length of the messages or completion.","type":"invalid_request_error","param":null,"code":"invalid_request_error"}}

[Linearl]

溢出以后，使用/compact不能正常压缩么？

一般都是80%或者还没上下文满的时候压缩，溢出了就没法压缩了

之前提过分段压缩再合并的PR，能解决这个问题，之前也合并了，但是不够完善（今天碰到一个2M的异常会话出现了压缩失败）
我已经本地修复并提交了改进方案

[SivanCola]

Fixed by #9879, merged into main-v2 as 3289e23 (on top of #9882 / 6361ba2).

What changed

• When the summary request itself is rejected by the provider (the 1108110 in the messages, 8192 in the completion 400 from this report), Reasonix now feeds the real prompt count back into its estimator, re-plans a smaller summary, then falls back to a bounded transcript form and, for /compact, the fragment path. The same request is never re-sent byte-identical, and summary planning keeps 5% of the window as headroom.
• At the hard ceiling, when no summary can form at all, a lossy truncate projection elides the oldest tool results and drops the oldest turns behind an explicit marker so the session can continue; /compact no longer dead-ends with "摘要失败 · 已停重试". The canonical transcript is never modified.
• Long autonomous turns: an overflow rescue may fold the turn's completed tool rounds (newest two kept), and a failed automatic attempt no longer silences maintenance for the whole turn.
• DeepSeek-style adapters now count replayed reasoning_content in the context estimate, closing the gap between "estimated" and what the API counted.

Verification

• internal/boot/effect_compaction_test.go: through the real build stack, a DeepSeek-format 400 on the summary request makes the next summary request strictly smaller and a 30-round tool loop completes.
• internal/agent/compact_summary_limit_test.go and updated ceiling tests; go vet, make lint, and the agent/boot/provider/control suites pass; PR CI 24/24 green.

Release availability: not in a released build yet (latest is v1.38.1). It ships with the next release from main-v2.

---

中文摘要：摘要请求自身超窗后会用 provider 返回的真实 token 数重新校准并缩小重试，再退到有界转录形式与分片；硬上限下摘要无法形成时改为有损截断投影而不是卡死。修复已合入 main-v2，尚未进入已发布版本（最新 v1.38.1），随下一版发布。