Six tools with clear action verbs and detailed descriptions (avg 180 chars, within baseline 194). All have input schemas with typed parameters. However, critical gaps: (1) No output schemas documented, LLMs cannot plan downstream calls or extract structured data; (2) Parameters lack descriptions in schema, only docstring prose exists; (3) No error handling guidance, tools return raw model output with no recovery hints; (4) Missing parameter constraints (enums, ranges, formats); (5) No pagination support despite tools returning unbounded text. Tool names are strong (local_* prefix, clear intent), but schema completeness and error patterns are weak.

• Add JSON Schema 'description' field to every input parameter. Move constraint info from docstring prose into schema (e.g., 'text: string, max 100000 chars, required').
• Document output schema for each tool. local_review_diff should declare: {findings: [{severity: enum, category: enum, message: string, file?: string, line?: number}], summary: string}. Others should declare return type (string) and any structure (e.g., 'code only, no markdown').
• Implement error handling wrapper. Catch backend exceptions, classify as retryable/user-fixable/fatal, and return structured error: {error: string, code: string, recovery_hint: string}. E.g., 'Model timeout. Retry with a smaller input or try a faster model.'
• Add parameter constraints to schema. Set maxLength on text/spec/diff/source (e.g., 100000). Add enum to focus: ['security', 'performance', 'correctness', 'tests'] or similar.
• Add tool annotations (readOnlyHint, idempotentHint) to signal safety. All six tools are read-only and idempotent, declare this via MCP tool annotations so agents know they are safe to retry.