Four tools with complete input schemas and reasonable descriptions (avg 156 chars, baseline 194). All parameters typed and described. However, output schemas are undocumented, LLMs cannot predict response structure for downstream chaining. Tool naming follows verb_noun pattern (searxng_search, searxng_extract) but lacks clarity on distinctions (search vs search_and_extract). No error handling guidance, no recovery hints, no per-item success/failure for batch operations. Missing tool annotations (readOnlyHint, idempotentHint). Composition is reasonable but search_and_extract combines two concerns.

• Document output schemas for all tools. Example for searxng_search: {results: [{title: string, url: string, snippet: string, engine: string}], fallbackEngines?: string, fallbackAttempts?: string[]}. This enables LLM to extract fields and chain tools.
• Add error classification and recovery hints. Replace 'SearXNG error (500)' with 'SearXNG service unavailable (HTTP 500). Retry in 30s or check SEARXNG_URL env var.' Replace 'Page too large' with 'Page exceeds 3MB limit. Try a different URL or increase MAX_EXTRACT_BYTES.'
• Split searxng_search_and_extract into two tools: searxng_search (returns URLs) and searxng_extract (takes URLs). Let agent compose them. This follows single-responsibility and gives agent control over which results to extract.
• Add tool annotations to all tools: readOnlyHint=true, idempotentHint=true for all four. This signals to LLM that retries are safe and no state changes occur.
• Modify searxng_extract to return per-item results: {urls: [{url, success, content, error}]} instead of blanket success/failure. This lets agent retry only failed URLs and proceed with partial results.