Repository navigation
stripHistoricalImages skips images nested in tool results (anthropic tool_result / Responses function_call_output / Gemini functionResponse) #436
Description
Problem
stripHistoricalImages (src/wire/strip-images.ts) only removes image parts at the top level of each message. Images inside tool results are never removed:
| Dialect | Where tool-result images live |
|---|---|
| anthropic | messages[].content[] → {type:"tool_result"}.content[] → {type:"image"} |
| responses | input[] → {type:"function_call_output" | "custom_tool_call_output"}.output[] → {type:"input_image"} |
contents[].parts[] → functionResponse.parts[] → {inlineData} / {fileData} |
In the affected session, historical images were nested in tool results (screenshots, file reads), so enabling stripImages did not remove them and the request body kept growing.
Repro (minimal)
stripHistoricalImages({ messages: [
{ role: "user", content: [{ type: "tool_result", tool_use_id: "t1",
content: [{ type: "image", source: { type: "base64", media_type: "image/png", data: "AAAA" } }] }] },
{ role: "assistant", content: [{ type: "text", text: "ok" }] },
]}, "anthropic", 1)
// → { removed: 0 }, the image is still thereImpact
Real pi session (billion-context-pi 0.1.77, Anthropic, 1701 messages, keepRecent: 3): replaying its request body locally through stripHistoricalImages gives 56.5 MB with removed: 0. Anthropic rejected it with 413 request_too_large (~32 MB limit) at only 22% of the token window, and every later request failed the same way.
Since #367, Anthropic tool-result images are protected from folding, so the opt-in strip needs to handle them explicitly.
Suggested fix
Strip images one level down inside tool-result parts (tool_result.content, functionResponse.parts), and walk Responses tool outputs through output instead of content. Emptied containers get the existing [image] placeholder, except Gemini functionResponse.parts, which only accepts media parts, so the key is dropped instead. Within Gemini functionResponse.parts, only explicit image/* MIME types are stripped; documents and parts without a MIME type are retained. These parts are left unchanged when response contains a $ref, to avoid dangling references. Top-level Gemini behaviour is unchanged. I have a PR ready.
Activity
ranxianglei commentedon Sep 25, 2026
[bot] 🏷 已确认问题属实:
stripHistoricalImages(src/wire/strip-images.ts)确实只处理每条消息的顶层 image part——anthropictool_result.content、Responsesfunction_call_output/custom_tool_call_output的.output、GeminifunctionResponse.parts里的图片全部漏掉。你提到的 PR #437 我已看到,正在做正式 review(代码 + 回归 + 测试有效性),结论稍后贴在这里和 PR 上。ranxianglei commentedon Sep 25, 2026
[bot] 🏷 Conclusion — bug confirmed (reproduced your minimal case on master:
removed: 0, image intact), and the fix is reviewed: PR #437 is mergeable.Full review posted on #437; highlights:
output; Gemini nested =image/*MIME only with emptiedpartskey dropped;$refskip; top-level Gemini unchanged).googleToCore— it never readsfunctionResponse.parts).anthropicToCorepairing check). Local on head: typecheck ✅ · 1131/1131 tests ✅ · build ✅.build-artifact) is a CI-permission defect affecting external fork PRs, unrelated to this code — filed as #440.Per repo rules the merge itself is human-side; the PR meets the §7.4 auto-merge gate, so it's ready whenever you are. Follow-up reminder from the PR body: after release, bump the
acp-kernelpin in billion-context-pi so the real 1701-message session actually gets the nested strip.