[ranxianglei]

fix(loop+wire): preserve redacted_thinking blocks on the Anthropic wire (#1960)

Fixes #1960

What was wrong

With Claude extended thinking, an in-turn compress left the session permanently
broken: the compress re-request 400'd with

thinking or redacted_thinking blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response.

and every later request in the session returned the same 400.

Root cause (verified by controlled repro driving the real runCompressLoop):
a redacted_thinking block arrives whole in content_block_start — no
delta ever carries its opaque payload — and nothing in either pipeline recorded
it:

1. Re-request path (this repo): the loop rebuilt the latest assistant
   message from recorded parts ([thinking, tool_use]) without the redacted
   block that sat between them → Anthropic rejected the modified message.
2. Steady path (acp-kernel): anthropicToCore had no case for
   redacted_thinking, so every client request echoing the round back silently
   dropped the block → the same modified shape went out forever. This is why
   the session could never recover.

The reporter's step-1 hypothesis (hideConsumedCompressCalls dropping the
current call) was checked and disproven: the executed compress call is kept
(its id is stamped onto the new active block, plus the newest-orphaned tail
scan), and the incident log has no hideConsumed hid line.

Changes

Loop (src/loop/core.ts, src/loop/adapter-anthropic.ts)

• The adapter captures the opaque redacted payload at content_block_start
  and emits a dedicated redacted_thinking event at block stop (instead of
  sealing an empty reasoning segment).
• One round's assistant parts (reasoning / text / redacted) are accumulated in
  stream order and emitted position-exact on the re-request; redacted
  parts ride the rawAnthropicBlock sidecar, which coreToAnthropic re-emits
  verbatim (same mechanism images already use).
• Recovery narrowing: a 400 matching cannot be modified skips the degraded
  strip-retry — that retry can only produce a second guaranteed 400 plus a
  swapped error — and surfaces the original upstream error immediately.

Steady path (acp-kernel pin 0.0.99 → 0.0.101)

• acp-kernel#500 adds the missing redacted_thinking case to
  anthropicToCore (sidecar-carried, verbatim on the way out).
• Also rides along from v0.0.100/v0.0.101 (release-note material):
  • 重启/会话分裂后上游 prompt cache 命中掉到 0% —— 每请求全量付费（成本回归） #499 prune/recommend shared per-turn index — ordinary requests in
    large sessions save hundreds of ms;
  • fix: land #454 (omp native compaction disable) on master — it was merged into its already-merged base branch #483 ref capacity widened to 9,999,999 (hybrid widths);
  • release v0.1.75 #484 case-insensitive protected-tool matching across clients;
  • fix: bound sessions storage growth (#407) #478 nudge cadence re-anchored when count drops below the shown reference;
  • fix: recognize sglang's overflow error format in context-window self-heal #493 language-preservation gated behind an opt-in flag, default off.
    All fixes or default-off — no default-behavior drift expected beyond what
    is listed above.

Tests

• tests/issue1960-loop-redacted.test.ts (new, 3 cases): byte-exact +
  position-exact redacted replay in the re-request; plain signed-thinking tail
  unchanged by the ordered-parts restructure; no-retry recovery on the
  cannot be modified 400 class.
• tests/issue1960-steady-redacted.test.ts (new): a client-echoed
  [thinking(sig), redacted(data), tool_use] survives prepare→outbound
  verbatim. Fails on acp-kernel 0.0.99 (verified), passes on 0.0.101.
• Preflight local: typecheck clean; targeted loop/thinking/reasoning suites
  186/186 green incl. both new [acp-loop] In-turn compress re-request edits the latest assistant message → Anthropic 400 "thinking blocks … cannot be modified" (session stays broken) #1960 files; build OK. Full suite via CI.

Behavior changes (flagged per policy)

1. Steady requests: previously SILENTLY DROPPED redacted_thinking blocks
   on every request (any session containing one was doomed to repeated 400s);
   now they are preserved byte-exact while their messages stay unmodified or
   are pruned with range. Old → new: dropped → preserved verbatim.
2. Compress re-request: the rebuilt latest assistant message now carries
   redacted blocks in stream position (previously lost — the defect). Plain
   signed-thinking shapes are byte-identical to before (regression-tested).
3. Degraded strip-retry: no longer fires for 400s matching
   cannot be modified (it guaranteed a second failure); the client sees the
   original upstream error text. Other 4xx classes retry exactly as before.

Config surface

None — no config field added, renamed, or re-semantics'd. No env vars, no CLI
flags.

[github-actions]

📦 Built Package Artifact

Branch: 2026-10-03_1960-redacted-thinking (00cf054)

Option A — Install this exact build (recommended)

npm install -g billion-context@0.0.2001-pr.1980.e5d91d9

This exact version is immutable and reproducible. The rolling pr tag tracks the newest test build across all PRs.

Option B — Download artifact

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

tar xzf billion-context-pr1980.tgz
npm install -g package

---

Install instructions are refreshed on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 收到,开始对 PR #1980 做独立 review(不重做,按 [PR] 流程核对代码/CI/回归)。已拉取分支 2026-10-03_1960-redacted-thinking(head 29f60d3c,两个线性 commit),正在本地装 acp-kernel 0.0.101 跑定向测试 + 核对 CI 状态,结论稍后按结构化七问给出。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Review 完成 — 结论:可合并(人工 merge;pin bump 属跨仓依赖,按 §7.4 必须留人)

独立复核(未重做):拉分支 2026-10-03_1960-redacted-thinking(head 29f60d3c,两个线性 commit),本地装 acp-kernel 0.0.101 跑定向验证 + CI 核对。

验证证据

• 定向批次(24 个 loop/thinking/reasoning 套件,含两个新 [acp-loop] In-turn compress re-request edits the latest assistant message → Anthropic 400 "thinking blocks … cannot be modified" (session stays broken) #1960 文件):235/235 绿;typecheck 干净。
• 测试有牙(我复验的):把 PR 的新 loop 测试文件放到 master 旧 loop 源码上跑(kernel 仍装 0.0.101 以隔离变量)——恰好钉住 loop 侧修复的两个 case(redacted 重放、no-retry recovery)失败,回归守卫 case 通过。非恒真断言。
• kernel codec 直接读了 0.0.101 dist 源码核实:入站新 redacted_thinking case 与 image case 同形(text:"[redacted_thinking]" + rawAnthropicBlock);出站对带 sidecar 的 text 消息只推原块、跳过占位文本——占位符不会漏上线变成多余 text block。
• GitHub CI 15/15 全绿(ubuntu/windows × node22/24、hermetic registry e2e 双 OS、real pi/codex/opencode/dsh lane、image billing e2e、build-artifact、version-guard)。PR mergeable,仅 behind v0.1.181 release merge(版本行,非阻塞)。

结构化七问

1. 是什么问题:Claude extended thinking 会话在一次 in-turn compress 后永久卡死——重发请求重建的最新 assistant 消息丢了其中的 redacted_thinking 块,Anthropic 400 "thinking blocks cannot be modified",之后每个请求重发同一畸形形状。
2. 什么原因导致的:redacted 块整体到达于 content_block_start,没有任何 delta 携带其不透明 payload,两条管线都无人记录——① loop 重建只回放已记录 parts(thinking+tool_use);② kernel anthropicToCore 无 redacted case,客户端(pi)忠实回显的历史也被每个稳态请求丢掉。一个机制、两个独立落点。用户 step-1 猜测(hideConsumedCompressCalls)在 [acp-loop] In-turn compress re-request edits the latest assistant message → Anthropic 400 "thinking blocks … cannot be modified" (session stays broken) #1960 triage 中已用证据排除。
3. 怎么修复的:流时捕获(adapter 在 block start 抓 payload)+ 一轮 parts 按流序累积、重放逐位还原(core,sidecar 携带、coreToAnthropic 原样吐出,复用 image 既有机制);稳态由 kernel fix: forward-once-then-learn for image-dominated payloads (#496) #500 补 codec case(pin 0.0.101);恢复收窄——命中 "cannot be modified" 的 400 跳过注定失败的 strip-retry,直接透传原始上游错误。
4. 是否有回归问题:逐项过完触碰路径的既有行为,无破坏——纯 signed-thinking 尾部形状字节一致(新测试钉住);unsigned/空 thinking 段仍按 偶尔模型在acp工具处停止执行 #221 丢弃;其他 4xx 类降级重试逻辑一字未动;客户端可见转发帧不变(redacted 走 thinking 同款 meta 路径);tool-call arguments 逐字([bug] 插件透传会剥掉工具调用参数里的 render tag（arguments 流挂了 tag-echo 过滤器） #1039 不变);Google/OpenAI lane 不受影响(新事件 kind 只有 Anthropic adapter 产出)。有意变更两处均已披露:稳态请求 redacted 块 dropped→preserved verbatim;"cannot be modified" 类不再 strip-retry(原错误直出,替代二次必败 400+错误被换)。
5. 是否破坏缓存:不涉及——前缀缓存、重试策略、增量游标、去重键均未触碰。
6. 是否改了设计方向:无转向——复用既有 rawAnthropicBlock sidecar 覆盖第三种块类型;loop 重建从固定布局改为流序保真(更忠实上游,常见形状输出不变)。
7. 是否引入新配置:无(config surface 零变更,PR 内已含该节)。

两点备注(不阻塞)

• 发布说明素材收窄:PR 正文 "rides along" 列了 重启/会话分裂后上游 prompt cache 命中掉到 0% —— 每请求全量付费（成本回归） #499/fix: land #454 (omp native compaction disable) on master — it was merged into its already-merged base branch #483/release v0.1.75 #484/fix: bound sessions storage growth (#407) #478/fix: recognize sglang's overflow error format in context-window self-heal #493,但经 compare API 核实 v0.0.100→v0.0.101 实际只有 fix: forward-once-then-learn for image-dominated payloads (#496) #500 + 重启/会话分裂后上游 prompt cache 命中掉到 0% —— 每请求全量付费（成本回归） #499(+prettier);fix: land #454 (omp native compaction disable) on master — it was merged into its already-merged base branch #483/release v0.1.75 #484/fix: bound sessions storage growth (#407) #478/fix: recognize sglang's overflow error format in context-window self-heal #493 已在 master 现 pin 0.0.100 里(其 pin commit 1a0c3b0a)。下个 release 写 release-notes entry 时窗口是 tag..HEAD,按实际 diff 列即可,别重复计数。
• wire-contract ledger([CI/testing] wire-contract anti-regression system: validation-parity fake upstreams + golden schema snapshots + constraint ledger + staged 4-protocol x 3-mode client smoke matrix #1304)判定:「cannot be modified」是跨请求历史保真约束,不是单请求 wire shape,ledger 的静态 gate/fake upstream 表达不了;两个新测试文件即永久守卫。判定 N/A,记录在此。

收工判定:代码、测试、CI、回归审查全部通过,无拦点。请人工合并 #1980(我不会代合);合并即闭环 #1960。