webp/gif read via built-in Read-image tool 400s on kimi-for-coding and permanently deadlocks the session (failed image re-sent every turn)

[fanzhe47]

Summary

A single webp (or gif) image that is pulled into the conversation by the built-in Read image file tool makes every subsequent LLM request fail with a non-retryable 400 ... failed to decode image, and because the failed image stays in history and is re-sent on every turn, the session is permanently deadlocked — no prompt can ever make progress again.

The image MIME is sniffed correctly (image/webp / image/gif), so this is not a mislabeling bug. The kimi-for-coding endpoint simply rejects these formats, and kimi-code has no guard, no transcode, and no way to drop the offending image from context.

Environment

• kimi-code 0.18.0 (observed). The 0.19.2 changelog does not appear to address the failed-image / history-resend path, but I have not re-verified on 0.19.2.
• Model: kimi-for-coding (Kimi-k2.6)
• macOS (arm64)
• Driven headlessly by an ACP client in --yolo mode (non-interactive), though the same mechanism applies to any harness.

What happens

The agent, during a normal task, calls the built-in Read image file tool on a local asset that happens to be a webp (saved with a misleading .png filename) and a small gif:

tool.result:
  <system>Read image file. Mime type: image/webp. Size: 83184 bytes ...</system>
  <image path=".../original_imgs/article2_img07.png">
  image_url: blobref:image/webp;<hash>

The very next LLM request fails:

WARN  llm request failed  turnStep=2.26 attempt=1/3 model=kimi-for-coding
      errorName=APIStatusError statusCode=400
      errorMessage="400 Invalid request: prepare image failed error, status code: 400,
                    message: failed to decode image: invalid or unsupported image format"
ERROR turn failed  turnId=2
WARN  acp: turn ended with failed reason  ... "retryable":false

From that point on, every turn fails immediately with the identical error:

turn failed  turnId=2   (step 2.26, when the image entered context)
turn failed  turnId=3
turn failed  turnId=4
turn failed  turnId=5   (sending "?" — anything — just triggers another doomed request)

The conversation contained 28 images total; 23 PNG + 2 JPEG decoded fine. The only failing inputs were 2 GIF + 1 WebP, all of them sniffed and labeled correctly as image/gif / image/webp.

Root cause (two compounding problems)

1. No format guard before sending to the vision endpoint. The Read-image tool forwards image/webp and image/gif to kimi-for-coding, which returns 400 failed to decode image. (Note: the general Kimi Vision docs list webp/gif as supported, but kimi-for-coding clearly does not accept them — an inconsistency worth confirming on the server side. Either way, kimi-code should transcode to png/jpeg, or refuse to attach, instead of poisoning the turn.)

2. A non-retryable image error permanently wedges the session. Because the failed image remains in history and is re-sent on every following request, turns 3/4/5/... can never succeed — there is no recovery short of starting a brand-new session or hand-editing wire.jsonl. This is the same family as #2279 (historical images re-sent) and Pasted images fail on text-only models — should offload to file + MCP path #905 (no offload fallback), but here it is triggered by the built-in Read-image tool on a server-rejected format, not by a text-only model or a web-session restore.

Expected behavior

Any one of these would break the deadlock:

• Before attaching, check the active model's accepted formats and transcode webp/gif → png/jpeg (or downscale) when needed.
• If the endpoint returns a non-retryable image-decode 400, drop the offending image from context (replace with a text note like [image omitted: unsupported format]) and let the turn proceed, instead of re-sending it forever.
• Surface the failure to the user/harness as a recoverable condition rather than an infinite identical-error loop.

Steps to reproduce

1. Put a .png-named file that is actually webp bytes (any web-downloaded webp) in the working dir, plus a small .gif.
2. Ask the agent to read/analyze those images with kimi-for-coding as the model.
3. Observe the 400 failed to decode image (non-retryable).
4. Send any further message → it fails immediately with the same error. The session never recovers.

Related

• Web 模式：恢复会话后历史图片被重复发送给 LLM || Web mode: historical images are repeatedly sent to LLM after restoring the session kimi-cli#2279 — historical images repeatedly re-sent after session restore
• Pasted images fail on text-only models — should offload to file + MCP path #905 — pasted images fail on text-only models; should offload to file + MCP