Summary

The Brave LLM Context API integration (tools.web.search.brave.mode: "llm-context") silently discards all extracted page content. The web_search tool returns results with empty snippets: [] arrays despite the Brave API returning rich pre-extracted text.

Root Cause

In src/agents/tools/web-search.ts, the runBraveLlmContextSearch function parses snippets as objects with a .text property:

snippets: (entry.snippets ?? []).map((s) => s.text ?? "").filter(Boolean)

But the Brave LLM Context API (/res/v1/llm/context) returns snippets as plain strings:

{
  "grounding": {
    "generic": [
      {
        "url": "https://example.com/article",
        "title": "Article Title",
        "snippets": [
          "# Heading\nActual extracted page content as a plain string...",
          "Another text chunk from the same page..."
        ]
      }
    ]
  }
}

Since "some string".text evaluates to undefined, every snippet falls through to "" and is removed by .filter(Boolean). The result is that all extracted content is silently thrown away.

Verified Against Live API

Direct curl to the Brave LLM Context endpoint confirms:

curl -s "https://api.search.brave.com/res/v1/llm/context?q=best+practices+for+React+hooks" \
  -H "X-Subscription-Token: $BRAVE_API_KEY" | jq ".grounding.generic[0].snippets[0] | type"
# Returns: "string"

A test query returned 9 URLs with 19 total snippets of actual page content — all discarded by the current parsing logic.

Suggested Fix

Handle both strings and objects (in case Brave changes their API in the future):

snippets: (entry.snippets ?? []).map((s) => typeof s === "string" ? s : (s.text ?? "")).filter(Boolean)

Environment

• OpenClaw 2026.3.8
• Brave Search API with LLM Context endpoint (/res/v1/llm/context)
• Config: tools.web.search.brave.mode: "llm-context"

Impact

The feature appears to work (no errors, results returned with titles and URLs) but provides zero value over standard web search mode since all extracted content is lost. Users have no indication anything is wrong.