claude-code - 💡(How to fix) Fix [FEATURE] optional raw/verbatim mode for WebFetch

Preflight Checklist

• I have searched existing requests and this feature hasn't been requested yet
• This is a single feature request (not multiple features)

Problem Statement

Summary

WebFetch routes fetched pages through a smaller extractor model before content reaches the main model. The extractor is tuned to summarize and paraphrase, and is reluctant to reproduce large blocks of user-generated content verbatim. This breaks WebFetch for any task where exact wording matters.

Requesting an optional parameter that bypasses the extractor and returns raw page content to the main model.

Reproduction

1. Point Claude Code at a forum thread with substantive user-generated content. Example: https://forum.freeciv.org/f/viewtopic.php?t=38 (a 2013 design proposal with follow-up discussion across two pages).
2. Ask it to read the thread and extract a structured spec preserving what each author actually wrote.
3. WebFetch returns paraphrased summaries of the posts. Claude Code reports back something along the lines of: "the WebFetch tool's processing layer declined to return full post bodies verbatim."

The failure is silent in the sense that something is returned — just not the content needed for the task. Downstream reasoning degrades without an obvious cause.

Impact

Any task where exact wording matters is currently better served by wget + local Read than by WebFetch. That includes:

• Extracting design specs or RFCs from forum threads and mailing list archives.
• Building transcripts from multi-post discussions.
• Verbatim quoting for citation or documentation.
• Structured-data scraping where field text must be preserved.

For these cases the extractor's compression isn't a win — the bytes are fetched regardless, the page is substantive end-to-end with no chrome worth stripping, and the user ends up paying for two model calls to get lower fidelity than one direct fetch would have provided.

Proposed Solution

Proposed fix

Minimal: add an optional boolean parameter (e.g. raw: true) to WebFetch that bypasses the extractor and returns page content directly to the main model, subject to the usual context-budget limits.

Slightly richer: a mode parameter — "summarize" | "extract" | "raw" — so the default behavior stays optimized for the common case while explicit overrides are available when fidelity matters.

Either way, the caller (or the main model on the caller's behalf) decides when fidelity matters more than compression, rather than that decision being made unilaterally by the extractor.

Alternative Solutions

No response

Priority

Medium - Would be very helpful

Feature Category

Performance and speed

Use Case Example

No response

Additional Context

Notes

• Discoverability is part of the problem: the failure mode isn't obvious from the tool's output, so users can spend significant time not realizing why downstream reasoning is going sideways.
• The fix is conceptually small and additive — existing behavior is preserved as the default.