fix: return fetched content to LLM in web_fetch tool

[lucamartinetti]

Summary

• WebFetchTool.Execute was setting ForLLM to a summary string ("Fetched N bytes from URL (extractor: ..., truncated: ...)") instead of the actual extracted text
• This was introduced in ca781d4 when the Tool interface was refactored from (string, error) to *ToolResult — the old code returned the full content as a single string, but the split into ForLLM/ForUser accidentally hid the content from the LLM
• The model never saw the fetched page content and could not answer questions based on web fetches

Fix

Return the extracted text in ForLLM (the field the LLM actually reads) instead of the metadata summary. ForUser continues to return the full JSON result.

Test plan

• Deployed in a sandboxed container and verified web_fetch now returns page content the LLM can use (tested with weather data queries)

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Fixes web_fetch so the LLM receives the actual extracted page content (instead of a metadata-only summary), restoring the tool’s usefulness for answering questions based on fetched URLs.

Changes:

• Update WebFetchTool.Execute to return extracted text in ToolResult.ForLLM.
• Keep ToolResult.ForUser as the full JSON payload (url/status/extractor/truncated/length/text).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

This change will break existing unit tests that assert WebFetchTool.ForLLM contains a summary string (see pkg/tools/web_test.go:40-43). Update the test expectations (and any other callers) to validate that ForLLM now contains the extracted text content instead of metadata, or include the needed metadata in ForLLM if the summary is still required by contract.

[afjcjsbx]

great bug you found! but instead of replacing ForLLM with text I would do something like this:

	return &ToolResult{
		ForLLM: string(resultJSON),
		ForUser: fmt.Sprintf(
			"Fetched %d bytes from %s (extractor: %s, truncated: %v)",
			len(text),
			urlStr,
			extractor,
			truncated,
		),
	}

there are also unit tests to fix, but nice PR!

[lucamartinetti]

Thanks @afjcjsbx, good call! Applied your suggestion — ForLLM now gets the full JSON result and ForUser gets the summary. Also updated the tests to match.

[afjcjsbx]

LGTM!

[PixelTux]

LGTM

[afjcjsbx]

@lxowalle PTAL

[afjcjsbx]

@xiaket PTAL

[xiaket]

LGTM, thanks for the fix!

[Orgmar]

@lucamartinetti Good catch on the web_fetch bug. Returning the summary metadata string instead of the actual fetched content to the LLM would make the tool basically useless for real web lookups. Subtle regression from the ToolResult refactor.

We have a PicoClaw Dev Group on Discord where contributors chat and collaborate. If you're interested, send an email to support@sipeed.com with the subject [Join PicoClaw Dev Group] lucamartinetti and we'll send you the invite link.