Single tool with clear naming (get_second_opinion) and comprehensive input schema. Description is adequate (194 chars, within baseline 34-392 range). All 5 parameters have type definitions and descriptions. However, output schema is undocumented, LLMs cannot predict response structure. Error handling returns text responses but lacks actionable recovery guidance. No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Parameter descriptions are functional but lack format constraints, ranges, or dependency hints. Tool calls external APIs (Gemini, Perplexity, Stack Overflow) but no timeout, rate-limit, or partial-failure handling documented.

• Document output schema explicitly: {content: [{type: 'text', text: string}], isError?: boolean, markdownPath?: string}. Include example response.
• Add readOnlyHint: true annotation to get_second_opinion in tool definition (line 47 of server.ts). This signals to agents the call is safe to retry.
• Enhance error messages with recovery guidance. Replace 'Gemini API error: ...' with 'Gemini API error: {message}. Try again in 30s or reduce code context size.' Categorize errors as retryable vs user-fixable.
• Add parameter constraints to descriptions: filePath should specify format (relative/absolute path, must exist); code and solutionsTried should note max length (e.g., 10000 chars) to prevent context explosion.
• Implement graceful degradation: if Perplexity or Stack Overflow fails, return Gemini analysis alone with a note: 'Perplexity insights unavailable; proceeding with Gemini analysis.' Avoid all-or-nothing failures.