fix: preserve non-text blocks in Anthropic tool_result via sidecar

[ranxianglei]

Closes #366

问题

anthropicToCore() 的 tool_result 分支对结构化 content 数组只提取文本(src/wire/anthropic.ts:101),image block 没有 .text → 像素在解析期蒸发;tool_result 消息不挂 sidecar,重建侧只从 m.text 还原 content。宿主经 Anthropic 线发送带图工具结果时每一轮往返都静默丢图(与 billion-context#1205 同类,但发生在 tool_result 内嵌场景,新会话第一条就丢)。

修复(与 #365 同一原则)

1. 解析期挂 sidecar:content 数组含非 text block 时,把整个原始 tool_result block 存入既有 rawAnthropicBlock 字段(bili-message.ts:27-29 的文档早已预留该用途)。纯文本 tool_result 不挂 sidecar,保持旧形状字节稳定。
2. 重建期原样吐回:coreToAnthropic 的 tool-result 分支优先使用 sidecar verbatim(与 user image 同路径),content 顺序、is_error、cache_control 全部保留。
3. 折叠保护:hasMediaPayload(src/protected.ts)识别该载体——rawAnthropicBlock.type === "tool_result" 且 content 数组含非 text block 时为 true。由此自动接入 fix: preserve unknown OpenAI content parts and protect media payloads from folding #365 的三处机制:BLOCKED ref(assignRefsNode)、compressible-range gap(buildCompressibleRanges)、显式 range 的 media carve-out + warning(applySingleRange)。
4. 配对原子性:无需新增逻辑——media result 被 carve 后,turn-integrity gate(INV2)自动把配对 call 一并 withdrawn,call/result 永不拆散(有测试钉住)。

行为变更披露(old → new)

项	旧	新
带非 text block 的 tool_result 经 proxy 往返	image/document block 静默丢弃,content 退化为 join 字符串	原 block 数组逐字节恢复,顺序不变
带图 tool_result 的可压缩性	可被折叠(= 数据销毁)	BLOCKED ref,永不出现在 compressible range;显式 range 覆盖时被 carve out 并 warn
配对 tool-call(range 覆盖时)	可单独折叠留下孤儿 call	integrity gate 自动 withdraw,与 result 一起保持可见
AnthropicToolResult.content 类型	string | AnthropicTextBlock[]	string | (AnthropicTextBlock | AnthropicImage)[] —— 跨仓注意:adapter 若对 tool_result content 做 c.text 映射会收到编译错误,需处理 image 变体

不变的:text 提取逻辑与 id seed 逐字节一致(升级前后同一逻辑消息得到同一 id,ref 不漂移);纯文本 array / string content 的 tool_result 仍按旧路径重建为 join 字符串;OpenAI/Responses/Google 线不受影响。

验证

• issue 给出的最小复现形状:修复前 rebuilt content 退化为 "done\n"、image 消失;修复后 [text, image] 原序逐字节恢复,hasMediaPayload = true。
• 新增 7 个回归测试(4 codec round-trip + 3 保护/配对):全套件 1036 pass(typecheck/build green)。
• format:check 在本仓库 master 上即有 119 个文件的既有失败(无 .prettierrc,代码风格与 prettier 默认不符);CI 实际只跑 typecheck+test+build,本 PR 未引入任何新的格式违规,也未顺带 reformat(避免 mass whitespace diff)。

依赖

基于 #365 的 head(90588d1)stacked——hasMediaPayload 由 #365 引入。请先合 #365;合并后本 PR 可直接合,或我 rebase 到 master。

范围外(issue 中列出的其他盲区)

issue 提到的 anthropic 未知 top-level block 无 default 分支、google 线远程 fileData 引用未挂 sidecar——按「一个 issue 一个 scope」未在本 PR 处理,是否立项请 owner 定。

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-24_anthropic-tool-result-image-sidecar (035b489)

Option A — Install from npm PR tag (recommended)

In your adapter project:

npm install acp-kernel@pr-367

Each push to this PR publishes a new version under the pr-367 npm tag.

Option B — Download artifact

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

tar xzf acp-kernel-pr367.tgz
npm install ./package

---

This comment is automatically updated on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

收到。这是 #366 会话(ses_7faa...)已完成的 PR,我不重做,按 PR 审查规范走:① 核对原 issue(#366)方向与根因定位;② 拉取分支核对代码(注意本 PR stacked 在 #365 的 90588d1 上);③ 本地最小验证(typecheck/test/build);④ 回归清点 + 测试有效性 + diff 整洁度。有结论回帖。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Review complete on head 035b489 (stacked on 90588d1). Verdict per gate:

① Original issue — direction correct. Root cause verified on master b8b4d22 by code read: parse did b.content.map((c) => c.text).join("\n") (image blocks contribute undefined → "done\n"), attached no sidecar, and rebuild emitted content: m.text ?? "". Loss happens at parse time, every round-trip — #366's diagnosis is right, and this fix attacks the root cause (carrier + protection), not the symptom. The rawAnthropicBlock field doc (bili-message.ts:27-29) literally pre-reserved this use ("an image or a structured tool_result"); this PR implements the missing half.

② Regression inventory (existing behaviors on touched paths):

• string-content tool_result: untouched (typeof check first) — pinned by test.
• text-only array tool_result: no sidecar, still rebuilt as joined string — pinned by test.
• id seeds: (c.type === "text" ? c.text : "") is byte-identical to c.text under join (String(undefined)===""); test pins res.text === "done\n" → same message, same id across upgrade. ✓
• user-image path: early return before the new branch, unchanged.
• cache_control: map still populated; raw-block path re-emits the identical value.
• hasMediaPayload: behavior identical for all pre-existing carriers (bare/tool_result sidecars without non-text content fall through exactly as before); additive only.
• OpenAI/Responses/Google wires: untouched.
  All four disclosed old→new rows match what the code actually does. No unintended regressions found.

③ Code. Minimal (4 files, +244/−5), mirrors the existing user-image sidecar pattern. Mechanism chain verified end-to-end: hasMediaPayload → BLOCKED ref (compress.ts:605-610) → recommendation gap (recommend.ts:202-208) → explicit-range carve via filterProtectedToolMessages (compress.ts:1272-1298, pass 1 also shields the paired call's id) → INV2 pair gate (turn-integrity.ts:188-201) withdraws any call whose result was carved, so call/result never split. CCR/absorb already skip BLOCKED refs (ccr.ts:297/396, absorb.ts:163), so media results consistently bypass placeholder offload — no conflict between the two mechanisms.

④ Tests. Real teeth: round-trip tests pin the exact bug shape ([text, image] byte-for-byte, original order) and the legacy shapes (string / text-only array stay sidecar-free); protection tests pin both sides — advertised ranges never span the media result nor its call, and an explicit range keeps both visible together with both warnings. A regression of either the codec or the pairing would fail these.

⑤ Diff. Clean — every line serves this PR; no whitespace churn, no unrelated files. CI green on both heads (typecheck+test+build on Node 22/24); confirmed ci.yml does not run format:check, so the master's 119 pre-existing prettier failures are correctly out of scope here.

Non-blocking nits (leave or file later, your call):

1. The widened union string | (AnthropicTextBlock | AnthropicImage)[] omits document/unknown future types Anthropic allows inside tool_result content arrays. Runtime is fine — the sidecar path keys off any non-text type — but TS-typed hosts constructing such bodies would get compile errors against the declared type. Could widen later if it bites.
2. applySingleRange computes mediaExcluded (compress.ts:1002-1010) solely for warning text while the actual removal happens in filterProtectedToolMessages. Same predicate, same input set, cannot diverge — just redundant; fine as-is.
3. No round-trip test for a document-type block (pairs with nit 1).

Process notes: duplicate screening found no parallel work; the two out-of-scope blind spots (anthropic unknown top-level block, google remote fileData) have no issues yet — per loop guard they stay owner-decision items, as you noted. Per §7.4 this PR touches a §7.3 wire-artifact contract (rebuild output shape) plus a cross-repo type change, so it stays human-gated regardless.

Recommendation: mergeable once #365 lands (or rebase to master after that merge — either works, mergeable_state is clean). I can't merge — human-only operation: #367

中文摘要:审核了 PR #367(Anthropic tool_result 内嵌图片在解析期静默丢失的修复,sidecar 载体 + 折叠保护),已逐行核对根因、四处既有行为无回归、7 个新测试断言有牙且钉住精确 bug 形状,diff 干净、CI 全绿——可以合并(需先合 #365)。