[BUG] subagent can't read image （image blob or image file path）

Environment

• Platform (select one):
• [x] Anthropic API
• [ ] AWS Bedrock
• [ ] Google Vertex AI
• [ ] Other: <!-- specify -->
• Claude CLI version: v1.0.83
• Operating System: macos
• Terminal: iterm2

## Bug Description
The image-describer subagent fails to analyze the provided image content, regardless of whether the image is supplied as a direct data stream (e.g., pasting into the terminal) or as a file path. Instead of describing the actual image, it returns a detailed but completely fabricated and unrelated description. This behavior was observed multiple times, indicating a critical failure in the agent's image processing or analysis capabilities.

## Steps to Reproduce

1. Start an interactive session with Claude Code.
2. Provide an image to the assistant either by pasting it directly or by referencing a local file path (e.g., describe 'image.jpg').
3. Ask the assistant to describe the image, which triggers the Task tool with subagent_type: 'image-describer'.
4. Observe the returned description from the agent.
5. Repeat the process with the other input method to confirm the behavior is consistent.

## Expected Behavior
The image-describer subagent should analyze the visual contents of the provided image and return an accurate description of what is depicted. For the given screenshot, it should have described the UI elements like icons and buttons.

## Actual Behavior
The image-describer agent returned completely hallucinated descriptions that had no relation to the actual image content.

• On the first attempt (pasted image), it described a hand holding a strawberry.
• On the second attempt (same image referenced via file path), it described a close-up photo of a male lion.

The agent did not seem to process the image at all and instead generated a random, plausible-sounding image description.

## Additional Context
This issue makes the image-describer agent unusable and unreliable. The failure is consistent across different input methods; both direct image pasting and providing a file path resulted in the same hallucinatory behavior. The bug appears to be a severe case of model "hallucination," where the agent invents a response instead of performing its designated task.