Repository navigation
Protected-message exemption inflates the incompressible context floor: per-compression reasoning is never reclaimable (~83.5% of measured residual) #651
Description
参考: opencode-acp#368 (ranxianglei/opencode-acp#368) —— 同一问题，按 owner 要求在此仓库同步提交。
问题
在长会话中，compress/skill 携带的 assistant 消息上的 reasoning 部分构成一个永久不可压缩的上下文地板。
保护是消息粒度的：filterProtectedToolMessages 把整条消息——包括 reasoning——排除出所有压缩，因此这些消息每轮都被原样重发。每次压缩轮次增加约 9 KB 无法回收的 reasoning。在真实长会话中测量，这占"从未被覆盖的残留"的 ~83.5%，是一个单调递增的反馈回路，会持续恶化长会话的可用性。
机制（代码路径）
- 保护是消息粒度：整条消息（含 reasoning）被排除出压缩选择。
- 被排除的消息不进入
byMessageId，因此prune每轮都重发它们（reasoning 一并重发）。 compress/skill是默认的compress.protectedTools；compress在FORCE_COMPRESS_PROTECTED中（强制追加，用户无法关闭）。- 真实 usage 计算包含 reasoning（
input + cacheRead + cacheWrite + output + reasoning）。
影响
- 长会话的上下文占用单调上涨，且这部分无法通过压缩回收。
- 显示 /
acp_status的占用估算甚至不包含 reasoning，进一步掩盖了这个问题。
建议
在请求时增加一个 pass，从已关闭历史轮次的受保护豁免消息上剥离 reasoning 部分（保留当前开放轮次以兼容需要 reasoning 重放的 provider），并加 kill-switch 与大小阈值。opencode-acp 侧的修复见 PR #370。
Activity
ranxianglei commentedon Sep 8, 2026
[bot] 🏷 收到，已接手。正在按 opencode-acp#368 / PR #370 的思路在本仓库定位等价代码路径（acp-kernel 的
isMessageProtected/prune/reasoning配对），确认 billion-context 侧的 reasoning 载体形态后给出修复方案。稍后回复。docs: name the kill-switch use case — round-trip-requiring thinking m…
fix(#651): close rounds on tool-result evidence, not user messages (#348
ranxianglei commentedon Sep 9, 2026
Follow-up found while observing the feature live (twin of billion-context-pi #348): the closure gate "compress call followed by a genuine user message" is unreachable in long agentic sessions — the whole session has no user messages after the opening prompt, so every compress round stays
activeforever and its thinking is retained as a permanent floor (observed: 0 drop events; 20612 / 8411 / 10581 chars retained across three compress rounds).Fixed in PR #667 (commit
8471882, rebased on the docs commit): a round now closes on round evidence — the compress call'stool-result(contentType: "tool-result", matchingtoolCallId") exists at a later index and at least one message follows it. In-flight rounds (result missing, or result still the last message) stay untouched; the per-providerreasoning.drop=falseescape hatch for round-trip-requiring thinking models (DeepSeek / GLM thinking / Qwen-QwQ) is preserved. Tests 1266/1266 green.Merge pull request #667 from ranxianglei/fix/compress-reasoning-floor
fix(ranxianglei#651): drop oversized reasoning from closed compress t…
fix(ranxianglei#684): reasoning-drop self-disables on strict-echo ups…
Merge pull request #1668 from ranxianglei/2026-09-29_anthropic-thinki…
fix(ranxianglei#1658): gate ranxianglei#651 reasoning-drop on signed …
fix(server): redacted_thinking arms the ranxianglei#651 reasoning-dro…
fix: recognize mcp__bili__compress in reasoning drop and plan it befo…
merge master into fix/redacted-thinking-reasoning-drop-gate
Merge pull request #3196 from sorinirimies/fix/redacted-thinking-reas…
feat: wire google lane into snapshot render (#3764)
test: pin #3922 mechanism — #651 reasoning-drop creates the 47/169 vi…
test: pin #3922 mechanism — #651 reasoning-drop creates the 47/169 vi…