Server has 8 tools with basic verb-noun naming (get_*, search_*) and descriptions present. However, critical gaps undermine quality: (1) All tools return unstructured markdown strings instead of structured JSON objects, breaking downstream tool chaining and forcing LLM text parsing. (2) No input schemas visible in source, tools use fastmcp decorators but parameter schemas are not explicitly defined in the provided code. (3) Descriptions lack actionable detail: 'Gets all projects assigned to me' doesn't explain when to use this vs search_projects, or what fields are returned. (4) No output schema documentation. (5) No error handling guidance, exceptions return raw error strings. (6) Parameters like 'state' and 'status' lack enum constraints in visible code. (7) No pagination support despite tools returning potentially large lists. Baseline: avg tool description ~60 chars (well below 194-char production baseline); no param descriptions visible; no structured output.

• Convert all tool responses from markdown strings to structured JSON objects. E.g., get_my_projects should return {projects: [{id: int, name: str, path: str, visibility: str, stars: int, forks: int, url: str}], total: int} instead of formatted text.
• Add explicit input schemas with JSON Schema types, enums, and descriptions for all parameters. E.g., state parameter should be {type: 'string', enum: ['opened', 'closed', 'merged', 'all'], description: 'Filter by merge request state...'}.
• Implement pagination: add limit (default 20, max 100) and offset/cursor parameters to list tools. Return total_count and next_cursor in response. Document in tool descriptions.
• Enhance error handling: categorize errors as retryable (network timeout, rate limit) vs user-fixable (invalid project_id) vs fatal. Return actionable messages: 'Project 999 not found. Try search_projects("project-name") first.'
• Document output schemas in tool descriptions. E.g., 'Returns a list of project objects with fields: id (integer), name (string), path (string), visibility (public|private|internal), stars (integer), forks (integer), url (string).' This enables LLM to plan downstream calls.