[Bug][0.2.0-rc.2] read_image passes only the first frame of an animated GIF, with no notice

[avpanasyuk]

Summary

read_image accepts animated GIFs but hands the model only the first frame, with no notice that frames were dropped. A model shown a moving clip concludes it is a still image.

Environment

• dsh 0.2.0-rc.2, @deepseek-ai/dsh-tool-fs (read_image), Linux host, OpenRouter provider declaring input: [text, image]

Reproduction

1. Make a 3-frame animated GIF whose frames are solid red, green and blue.
2. Ask the agent to read_image it and describe each frame.

Actual: the model receives one image, image/webp, solid red, and answers that the image is a single red square. Nothing in the tool result mentions that the source had 3 frames.

Expected: either the frames are delivered (or sampled) or the tool says that only frame 1 of N was passed on.

Where

• dsh-tool-fs lib/index.js:831 sniffs GIF87a/GIF89a and accepts the file as image/gif; the README lists "PNG/JPEG/WebP/GIF" with no mention of frames.
• Attachment normalization then re-encodes to a single still (the delivered block is image/webp), so the frame count is lost before the model sees it.

Suggestion

Cheapest fix: when the source has more than one frame, append a line to the tool result such as animated GIF: showing frame 1 of 3. Better: refuse multi-frame input with a message pointing to frame extraction, or deliver a contact sheet.

Workaround until then: extract frames (or a tiled contact sheet) with ffmpeg and read those instead.

---

Found and verified against the installed 0.2.0-rc.2 sources with the help of an AI coding agent.

[PerryLink]

模块文档确实接受 GIF，而且对帧数只字未提——这正是你这条的着力点

1. 文档侧的证据

packages/fs/tool-fs/src/read-image.ts 开头：

The model-facing read_image tool commits a PNG/JPEG/WebP/GIF file. A path…

⇒ 两件事同时成立：

1. GIF 是被接受的类型（不是"不支持所以拒绝"）；
2. 文档完全没提帧——既没说"只取首帧"，也没说会截断。

⇒ 你报告的"只传首帧且没有任何提示"与文档的沉默吻合：行为上有截断，说明上没有告知。"无提示"这一半因此是可证成的（文档承诺的只是"commit 这个文件"）。

2. 为什么严重度不止"保真度"（建议这样写）

你写"模型看了会认定这是静态图"——这一点比"丢了帧"重要得多：它让模型对内容作出错误判断（例如把动图当静帧、把"变化过程"当"单一状态"），而用户与模型都不知道信息被截断。⇒ 这是正确性问题，不是保真度问题。

建议的诉求（两条，第二条成本极低）：

1. 要么把全部帧传给模型（若协议允许多图）；
2. 要么在工具结果里显式说明"仅使用首帧"（并可在工具描述里也写一句）——让模型知道它看到的是第一帧，从而不据此下结论。

第 2 条几乎零成本，且能独立成立——即使最终决定不做多帧传输，这一条也该做。

3. 请补两样

1. read_image 实际交给模型的载荷（是单张图？还是"图 + 文本说明"）——这一条决定第 2 条能否落地；
2. 版本基线：你写 0.2.0-rc.2。当前最新是 0.2.1-alpha.1（714 个提交，packages/client 是重灾区）⇒ 请确认 read_image 的帧处理在新版未变。我这边核到的是文档里仍写着接受 GIF、且仍不提帧。

4. 一条相关线索

同日 #8793 这类"模型看到的与真实内容不一致"的报告还有一条相邻的：#8829（会话日志的 usage 与厂商计费差约 26×，作者已自行更正为"输入 cache-miss"）。两条同属"观测值与实际不符"，但一条是内容、一条是计量——建议只互相引用、不要混为一条。

一条边界

我确认的是模块文档接受 GIF 且不提帧这一点（足以支撑"无提示"）。**"实际只传首帧"**以你的复现为准——我没有在运行时检查 read_image 的载荷。

[avpanasyuk]

Thanks. Answers to both asks, plus a correction to my own report.

Correction: the first-frame behaviour is documented, just not where the model or a read_image caller looks. @deepseek-ai/dsh-attachment-local README, 0.2.0-rc.2 and still in dsh-v0.2.1-alpha.1:

Animated GIF becomes static — normalization retains only the first frame; animation is outside the version-one image contract.

So this is a deliberate contract, and the report narrows to: the model is never told. read_image's description says only "Read a PNG/JPEG/WebP/GIF file and return the image itself".

1. What read_image hands the model (0.2.0-rc.2, dsh-tool-fs lib/index.js:1035-1047): one image reference and nothing else.

return { path, image: { attachmentId, mediaType, bytes, width, height, name?, originalDimensions? } }

There is no text part and no frame count. originalDimensions already exists to tell the model that the image was resized, so a frame count would fit naturally beside it (e.g. originalFrames: 3), with a line in the tool description.

2. Version: unchanged in dsh-v0.2.1-alpha.1. attachment-local/src/image.ts still reduces pages > 1 to an animated flag, normalization.ts still forces a single-frame output, and the README line above is still there.

Smallest fix: when the source has more than one frame, put a frame count in the result (and one sentence in the tool description), so the model knows it is looking at frame 1 of N.