[ranxianglei]

Fixes #489 (found while analyzing billion-context#1869)

What

rebuildMessages pinned the first user message unconditionally ahead of the covered check. Once a covered message became the view's first user (client trimmed earlier history before resending, or the fold targeted the opening message), its coverage was defeated forever: syncBlocks kept the block active because the base id was still present pre-prune, so the folded payload re-sent itself verbatim every turn while every fold reported success.

Behavior change (old → new)

• Old: first user message always survives prune, even when covered by an active block — coverage of that message is silently void for the life of the session.
• New: a covered first user drops when a plain user text core survives behind it (firstUserDropSafe: earliest non-covered, non-rendered-summary follower must be role user AND contentType text — the only shape the downstream orphan-stripping passes cannot remove). The rebuilt wire then opens with that surviving user (rendered summary anchored in front, re-voiced/stripped downstream per mode).
• Unchanged: shapes where no user would lead after the drop ([U1-covered, A1, …], sole-user sessions) keep today's pin — shipping a non-user-leading frame would reject strict providers, which outweighs the compression gain. Documented as a residual limitation in DESIGN.md §8.1.

This is a wire-shape change on purpose: frames lose only a previously-pinned covered message whose content an active block has summarized. No ref/id churn (byRef untouched, ids never re-issued), no summary-format change, no node-ordering change.

Why kernel-side

The producer of the bad bytes is the rebuild step; adapter-side detection could only deactivate the stale block (mask) while the payload still occupies context.

Pre-flight

• npm run typecheck clean
• Targeted tests green: all 11 test files importing src/prune.js (121 tests) + issue-regression family (21), incl. the rewritten/added regression tests:
  • tests/state-prune.test.ts: split into the degenerate-keep shape (assistant followers → pin holds), the new drop case (user follower → covered first user drops, with and without summary injection), and a stale-rendered-summary-copy case (host passes a previously-pruned view)
  • tests/orphan-fixes.test.ts: assertion wording corrected to the conditional semantics (behavior of that fixture unchanged — degenerate shape)
  • tests/issue1476-reid.test.ts: the old "pinned echo keeps id/ref" test now covers the pruned-across-passes path (#1869); added the degenerate-pin variant preserving the original id/ref-stability coverage (reconcile-live-ids (#459) re-numbers post-fold RESENDS as new instances: folded messages escape prune and re-inflate the wire (P0, breaks billion-context CI e2e on 0.0.96) #462)
• npm run build + npm run format:check green
• Full suite left to CI per local-testing discipline
• Second-agent review note: the oracle reviewer agent errored out twice in this environment (infra failure, 0s duration); the author ran an adversarial self-pass over anchor-placement, orphan-stripping, remint/ref-space interactions — human review carries the second-review requirement for this PR.

Sequencing

Kernel ships first (release v0.0.100 via the standard flow), then billion-context bumps its exact pin and lands the Responses WS lane regression test that reproduced this end-to-end.

[github-actions]

📦 Built Package Artifact

Branch: 2026-10-01_first-user-pin-yields-coverage (e593fa8)

Option A — Install from npm PR tag (recommended)

In your adapter project:

npm install acp-kernel@pr-490

Each push to this PR publishes a new version under the pr-490 npm tag.

Option B — Download artifact

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

tar xzf acp-kernel-pr490.tgz
npm install ./package

---

This comment is automatically updated on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 PR #490 review 完成 — 审的是最终 head e593fa8(注意:head 在我初审后从 ac4571d amend 过一次——firstUserDropSafe 的判定从 role === "user" 收紧为 role === "user" && contentType === "text",并新增 stale-summary 回归测试。amend 是改进,方向正确,见下)。CI 在 e593fa8 上全绿(pr-validation / test(22) / test(24) / build-artifact)。本地定向验证:typecheck 干净、160/160(11 个 import src/prune 的测试文件 + issue-regression family + issue1476)、build + format:check 绿、工作树干净。

结构化七问:

1. 是什么问题:fold 覆盖会话首条用户消息后,该消息被无条件钉在 wire 上逐轮原样重发,而 block 保持 active、每次 fold 都报成功——压缩静默失效(prune: unconditional first-user pin defeats fold coverage when a covered message becomes the view's first user #489,billion-context#1869 端到端复现)。

2. 什么原因导致的:rebuildMessages 里 pin 检查排在 covered 检查之前且无条件触发。pin 的本意是属性保证("输入有 user ⇒ 输出至少一个 user",strict provider 拒绝 0-user 帧),实现却过度保证为"钉住那条具体的首消息"。一旦某条 covered 消息变成视图的首 user(client 裁剪历史后重发,或 fold 直接打开场消息),它的 coverage 被永久击败。病因定位准确:修的是机制(pin 语义过强),不是症状。

3. 怎么修复的:pin 改为条件让位——covered 的 pin 仅当其后最早的非 covered、非 rendered-summary 跟随者是纯 user text core(firstUserDropSafe)时放行丢弃,重建 wire 由该存活 user 领衔(rendered summary 锚在其前)。退化形态(后面没有 user 能领衔)保持现状 pin,记入 DESIGN.md §8.1 residual limitation。e593fa8 的收紧是对的:若只判 role,一个 user-role 的 tool/reasoning core 理论上可被下游 orphan-stripping 再删掉,导致帧失去 user 领衔——只有 plain text user 是不可再删的形状。我核对了全部 codec:anthropic/openai/responses/google 的 tool-result core 一律 role: "tool",现有 wire 形状下首 user 不可能是 tool result;收紧属于对 host 构造 CoreMessage 的防御性加固,失败方向保守(宁可不压,不产出非法帧)。

4. 是否有回归问题:对 prune 触碰路径的既有行为逐项清点——

   • 非首 user 的 covered 丢弃:不变(同一 covered-skip 路径)。
   • 未 covered 的首 user pin:不变(!isCovered 短路)。
   • covered 但无 user 可领衔:不变,pin 保持——state-prune 退化用例 + orphan-fixes fixture([u0,a0,u1,a1],block 覆盖 u0/a0/u1,后续仅 a1 assistant → pin 保持)双重钉住;orphan-fixes 本次仅改断言措辞,fixture 行为未变,已核实。
   • summary 锚定(collectSummaryAnchors/pairSafeAnchorIndex)、stale summary 去重:不变,新测试把"stale 副本夹在 pin 与存活 user 之间 → 单份新渲染 summary + m3"钉死。
   • 工具对原子性(§7.3 fix: hard-enforce protection of recent messages + last user message #5):保持——新丢弃路径与既有 covered 丢弃走完全相同的下游清理(stripOrphanedToolCalls/Results/Reasoning),无新失败类别;更严的门进一步缩小风险面。
   • 无其他代码/测试依赖旧无条件 pin:CI 全套绿。
     唯一行为变化即修复本身(PR 正文已披露 old→new),未发现静默漂移。无回归。

5. 是否破坏缓存:否。byRef 不动、id 永不重发(issue1476 两遍 processTurn 断言 m00001 ref 稳定)、存活消息的 token 快照与 summary 格式不变。受影响帧变短(不再重发 pinned payload)是预期收益,不是失效churn。

6. 是否改了设计方向:否。管线、节点序、block 生命周期均未动(syncBlocks 仍按 pre-prune 存在性保活 block——现在 prune 真的删掉 payload,这才是正确闭环)。kernel 侧修复层次判断成立:坏字节的产地就是 rebuild,adapter 侧去活 block 只是掩盖,payload 仍占 context。

7. 是否引入了新的配置:不涉及。

测试有效性:新/改写测试断言有牙(deepEqual 精确 id 序列;两遍 processTurn e2e 走完整管线含 syncBlocks,断言 payload 文本消失 + ref 稳定,正是 #1869 的复现路径;退化变体保住 #462 的 id/ref 稳定性覆盖)。diff 整洁:相对当前 master 仅 5 个文件(DESIGN.md、src/prune.ts、3 个测试文件),无版本/lock 连带变更。

判定:可合(按规矩由人合并)。Sequencing 提醒:kernel 先发 v0.0.100,之后 billion-context 升精确 pin 并落 Responses WS lane 回归测试。

[ranxianglei]

设计内的
不需要修复
bug 本身：会话的第一条消息被设计成"永远保留在请求头位置"（这是故意的，为了前缀缓存稳定，本身没错）。但麻烦在于——当这条消息恰好被压缩折叠了之后，代码仍然把它的全文硬塞进每次发给模型的请求体
里。也就是说：这次压缩等于白干——摘要生成了、块也折叠了，可原文一个 token 都没省下来，而且全程不报错不提示，纯静默白干。