Server has 7 well-named tools with clear descriptions (avg 150 chars, within baseline 194). All tools start with action verbs (get_, query_, check_, impact_). Input schemas are present and typed. However, output schemas are NOT documented in the visible code, critical gap for LLM planning. Parameter descriptions are present but sparse (avg ~40 chars vs baseline 72). No error handling guidance visible. No tool annotations (readOnlyHint, destructiveHint). Missing pagination support on list-like tools (query_graph, get_decisions could return many results). No confirmation/dry-run pattern for drift detection or impact analysis.

• Document output schemas for all 7 tools. Include field names, types, and what each field represents. E.g., get_context should return {context: string, token_count: int, truncated: bool, files_included: [string], decisions: [object]}.
• Add pagination to query_graph and get_decisions: accept limit (default 10, max 50) and offset/cursor parameters. Return total_count and next_cursor in response.
• Add tool annotations to all tools: set readOnlyHint=true for all 7 (they are all read-only). This signals to LLMs that these tools are safe to call without confirmation.
• Expand parameter descriptions. E.g., 'include_depth: How many levels of dependencies to include (1=direct only, 2=direct+transitive, etc.; default: 1)'. Baseline is 72 chars; current avg is ~40.
• Add error recovery guidance. When initialization fails, return structured error: {error: 'not_initialized', message: '...', recovery_steps: ['Run: codebase-intel init <project_root>', 'Then retry your request']}.