compression can fail on images for kimi

[acoliver]

What happened?

✕ [API Error: Compression strategy "middle-out" failed: LLM provider call failed: 400 {"error":{"message":"litellm.BadRequestError: Fireworks_aiException - Error code: 400 - {'error': {'object': 'error', 'type': 'invalid_request_error', 'code': 'invalid_request_error', 'message': 'Cannot decode or download image: Incorrect padding'}}","type":null,"param":null,"code":"400"}}]

What did you expect to happen?

I think before compresison the images should just be removed unless they are in the presevered area. etiehr way this shouldn't happen

Client information

About LLxprt Code

CLI Version 0.10.0-nightly.260406.b4e83ea07
Git Commit b4e83ea
Model anthropic:accounts/fireworks/routers/kimi-k2p5-turbo
Provider anthropic
Base URL https://api.fireworks.ai/inference/
Sandbox no sandbox
OS darwin

Login information

No response

Anything else we need to know?

No response

[acoliver]

@CodeRabbit look into this

[coderabbitai]

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
gh issue view 1889 --repo vybestack/llxprt-code

Length of output: 1486

---

🏁 Script executed:

#!/bin/bash
# Find middle-out compression related files
rg -rn "middle-out" --type ts -l
echo "---"
rg -rn "middle.out" --type ts -l
echo "---"
fd -e ts -e js | xargs grep -l "middleOut\|middle_out\|MiddleOut" 2>/dev/null

Length of output: 3020

---

🏁 Script executed:

#!/bin/bash
cat packages/core/src/core/compression/MiddleOutStrategy.ts

Length of output: 14178

---

🏁 Script executed:

#!/bin/bash
cat packages/core/src/core/compression/utils.ts

Length of output: 12745

---

🏁 Script executed:

#!/bin/bash
cat packages/core/src/services/history/IContent.ts
echo "---"
# Also check for any existing image stripping logic
rg -n "image\|media\|inlineData\|base64" --type ts packages/core/src/core/compression/ -A2 -B2

Length of output: 7216

---

@acoliver, I've traced the root cause. Here's what's happening:

---

Root Cause

In packages/core/src/core/compression/utils.ts, the sanitizeHistoryForCompression function converts tool_call and tool_response blocks to plain text (to avoid Anthropic's strict tool pairing validation), but media blocks (images) pass through completely unchanged:

// In sanitizeHistoryForCompression — the fallthrough at the end of the map:
return block; // ← MediaBlocks fall here, raw base64/URL data sent as-is to the compression LLM

When toCompress (the middle section) contains messages with MediaBlock entries, those blocks — including raw base64-encoded image data — are forwarded verbatim to the compression provider (Fireworks/Kimi). The provider then fails with:

Cannot decode or download image: Incorrect padding

The toKeepTop and toKeepBottom sections are never passed through sanitizeHistoryForCompression, so only the middle (compressed) messages are affected — consistent with the reporter's observation.

---

Suggested Fix

In sanitizeHistoryForCompression, add a case to replace media blocks with a descriptive text placeholder so the compression LLM understands an image was present without receiving any binary/base64 data:

// In sanitizeHistoryForCompression, inside the .map() on blocks:
if (block.type === 'media') {
  const label =
    block.caption ?? block.filename ?? block.mimeType ?? 'image';
  return {
    type: 'text',
    text: `[Media attachment: ${label}]`,
  } as TextBlock;
}

This also means the hasToolBlocks guard at the top of the function needs to be extended to detect media blocks, so messages containing only images are also sanitized:

// Change this:
const hasToolBlocks = msg.blocks.some(
  (b) => b.type === 'tool_call' || b.type === 'tool_response',
);
if (!hasToolBlocks && msg.speaker !== 'tool') {
  return msg;
}

// To this:
const needsSanitization = msg.blocks.some(
  (b) =>
    b.type === 'tool_call' ||
    b.type === 'tool_response' ||
    b.type === 'media',
);
if (!needsSanitization && msg.speaker !== 'tool') {
  return msg;
}

Would you like me to plan this as a code change?