[BUG] WebFetch returns raw page content instead of processing prompt

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

WebFetch returns the entire raw page content (~80KB) regardless of the prompt parameter. Even with a minimal prompt like "summarize this in one sentence", the tool returns the full page text without processing it through the summarizer model.

The tool description states:

• "Processes the content with the prompt using a small, fast model"
• "Results may be summarized if the content is very large"

Neither behavior occurs. The same ~80KB output is returned verbatim for both a detailed extraction prompt and a one-sentence summary prompt.

What Should Happen?

The internal small model should process the fetched content according to the prompt parameter and return a focused response — not the raw page content.

Error Messages/Logs

No error is raised. The tool returns successfully, but the output is the raw page content rather than a processed response. The output is flagged as Output too large (80.2KB) and persisted to disk.

Steps to Reproduce

1. Start a Claude Code session
2. Trigger a WebFetch call targeting a large documentation page, e.g.:

• URL: https://platform.claude.com/docs/en/docs/build-with-claude/prompt-caching
• Prompt: "Summarize this page in one sentence"

1. Observe that the output is ~80KB of raw markdown/HTML content starting with the page title
2. Repeat with a different prompt (e.g., "What is the cache TTL?") — the output is identical

The output is the full page content verbatim, with no evidence that the prompt was processed by the summarizer model.

Claude Model

Opus

Is this a regression?

I don't know

Claude Code Version

2.1.53

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

iTerm2

Additional Information

Tested against the Anthropic prompt caching documentation page (~1800 lines of MDX content with custom components like <Note>, <CodeGroup>, <Tip>, <section>). It's possible the issue is specific to large pages or pages with non-standard HTML structure, but the behavior (prompt completely ignored, raw content returned) suggests the summarizer model is being bypassed entirely.

✍️ Author: Claude Code with @carrotRakko (AI-written, human-approved)