feat: compress.reasoning — drop oversized thinking from closed-turn compress calls (#336)

[ranxianglei]

Fixes #336. Supersedes the closed #338 (kernel reasoningReplay approach) — reimplemented cleanly on master with exact semantic alignment to opencode-acp #377 (owner-locked design).

Problem

compress tool-call messages are hard-exempt from every compression selection (their tool results anchor the block summaries), so their thinking parts ride along every request — a monotonically growing, unreclaimable context floor (in the #336 storm session, ~53K tokens of anchor thinking in the outgoing view).

Design (mirrors opencode-acp #377)

Request-time pass (src/reasoning-drop.ts) that removes thinking parts from a message only when ALL gates hold:

1. Closed turn — strictly before the last genuine user message; the active round is never touched (some providers require replaying the active round's thinking).
2. Selector — the message carries a toolCall part with name compress (only compress; other protected tools would need their own explicit config).
3. Size — the message's total reasoning length (chars, summed across parts of that message, never across messages) strictly exceeds threshold. 0 drops any non-empty.

Config — nested under the existing compress three-level cascade, merged field-wise (deepest set field wins):

{
  "compress": {
    "reasoning": { "drop": true, "threshold": 2048 },
    "providers": {
      "openai": { "reasoning": { "drop": false } },
      "anthropic": { "models": { "claude-sonnet-4-5": { "reasoning": { "threshold": 8000 } } } }
    }
  }
}

Opaque-reasoning providers (OpenAI encrypted reasoning) opt out per-provider as shown. Persisted history is never modified — only the outgoing view (rebuilt from the session log each request) changes. Pure, idempotent, fail-safe.

Wired in the context transform right after view rebuild and before the nudge push, so a synthetic user-role nudge can never become the boundary.

Tests

638 total, 0 fail. 12 new (tests/reasoning-drop.test.ts): defaults, invalid-value fallbacks, all three gates (boundary / selector / size incl. 2048==kept, 2049 boundary, per-message sum, no cross-message accumulation), threshold-0 semantics, purity + idempotence, fail-safe, kill-switch, no-user case, toolCall/thoughtSignature preservation, field-wise three-level merge.

Real-session validation

Replayed the #336 storm session (01a07b3c, 998 messages): 602K chars of thinking on compress-call messages → ~7K after the pass (48 messages hit), idempotent.

Docs: CONFIGURATION.md + CONFIGURATION.zh-CN.md (compress.reasoning section, field table, sample config).

[github-actions]

📦 Built Extension Artifact

Branch: 2026-09-09_compress-reasoning-drop (52567ac)

Option A — Install from npm PR tag (recommended)

pi install npm:billion-context-pi@pr-339

Each push to this PR publishes a new version under the pr-339 npm tag.

Option B — Download artifact

1. Download the artifact from the Actions run
2. Extract the tarball and install:

tar xzf billion-context-pi-pr339.tgz
pi install ./package

---

This comment is automatically updated on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 PR #339 review 完成 — LGTM(一个非阻塞 nit)

查重结论:非重复。#339 正是 #338 floor 9 承诺的 clean reimplementation(owner-locked:"exact alignment with opencode-acp #377 … Clean reimplementation coming on a fresh branch; kernel stays untouched")。#338 已按 superseded 关闭;仓库内无其他同主题 open 项。

#336 审核:问题真实存在——compress 调用被硬排除在压缩之外(FORCE_COMPRESS_PROTECTED),其 thinking 部分每轮原样重发,构成单调上涨的不可回收地板。机制与代码路径已核实。

本地验证(PR commit 52567ac,base f2c5924,npm ci + 全套检查):

• typecheck ✅ / test ✅ 638 total, 635 pass, 0 fail, 3 skipped(与 PR 声称一致)/ build ✅ dist/index.js 730.76 KB
• 新增 tests/reasoning-drop.test.ts 12/12 pass(与"12 new"一致)
• git merge-tree 对当前 master(6d3ee9d,v0.1.60):干净,无冲突

代码 review(逐点核实):

1. src/reasoning-drop.ts — 三道门控与设计一致:closed turn(严格位于最后一条 user 消息之前)、selector(toolCall 且 name === "compress")、size(单消息内各 thinking 部分求和严格大于 threshold;0 丢弃任何非空)。纯函数(slice + spread,不改输入)、幂等(无变化时返回同一引用)、fail-safe(try/catch 原样返回)。
   • pi-ai part 形状已核实:{ type: "thinking", thinking: string, thinkingSignature? },isThinking 匹配;整块移除(含 signature)语义自洽(signature 只随块回传才有意义)。
   • pi 消息角色已核实:user | assistant | "toolResult"——工具结果有独立角色,"所有 user 消息都是 genuine"成立。
   • content 不会变空:selector 要求存在 compress toolCall 部分,该部分必然存活。
   • 前缀缓存:pass 对同一 log 状态确定;边界只前移;一旦丢弃永不恢复(单调),缓存友好。
2. src/config.ts:356-371 mergeCompress — 逐字段 model?.reasoning?.drop ?? provider?.reasoning?.drop ?? global?.reasoning?.drop,最深 set 字段赢 ✅。
3. src/runtime.ts reasoningDropFor — resolveCompress(adapterRef.compress, provider, modelId) 参数顺序已核对;走 live adapterRef,config reload 即时生效。
4. src/index.ts:399-412 — drop 在 view rebuild 之后、nudge push(第 496 行)之前 ✅;nudge 不落盘(每次事件从 session log 重建),边界永远由 genuine log 消息决定。coreOutToAgentMessages 仅一处构建出口视图,无绕过路径。
5. 文档:CONFIGURATION.md + zh-CN 均有 compress.reasoning 章节(字段表 + 示例配置)✅。

Nit(非阻塞):src/index.ts:410 — debug.event("reasoning-drop", { dropped: droppedThinking.length, ... }) 中 droppedThinking 是返回的完整数组,.length 是整个视图的消息总数,不是实际被剥离 thinking 的消息数,字段名有误导(仅 debug 事件,无功能影响)。建议让 dropCompressReasoning 额外返回改写计数,或改日志字段。

Consideration(owner 拍板项,无需动作):默认 drop: true 对全部 provider(含 OpenAI 加密 reasoning)生效,文档只提示按 provider 退出;被取代的 #338 里 GPT 系是默认关。新设计 owner-locked 且与已合入的 opencode-acp #377 一致,属有意为之——仅提示 OpenAI 用户需自行设置 providers.openai.reasoning.drop: false。

结论:实现与 owner-locked 设计及已合入的 opencode-acp #377 语义对齐,全套 CI 本地通过,与 master 无冲突,可人工合并(按仓库规则 PR 合并仅人工)。"Real-session validation"数字(602K→~7K chars,48 messages)合理,但沙箱内无风暴会话日志,无法独立复现。