Server has 10 tools with basic schemas and descriptions, but significant quality gaps. Most tool descriptions are 20-80 chars (below the 194-char baseline for A+ tools). Parameter descriptions exist but are minimal. No error handling guidance, no output schema documentation, no tool annotations. Tool names lack consistent verb-noun patterns (saveSplat, startMCP, currPosRot are non-standard). No pagination support for list operations. Security concerns: no mention of permission gates or audit trails. Composition issues: tools like updatePosition, updateRotation, updateCamera operate on shared state without idempotency guarantees or confirmation patterns for destructive ops.

• Rename tools to verb_noun pattern: saveSplat → save_splat, currPosRot → get_current_position_rotation, startMCP → start_mcp_session, latestLog → get_latest_log. Use lowercase with underscores for consistency with 90% of production tools.
• Document output schemas for all tools. For each tool, add a 'Returns' section describing the response structure (e.g., 'Returns {id: string, xPos: float, yPos: float, zPos: float, ...}'). This enables LLMs to plan downstream calls.
• Add error recovery guidance. Replace bare HTTPException with structured errors: '{"error": "No poses found", "recovery": "Call updatePosition() first to create a pose, then retry."}' for currPosRot; similar for latestLog and trajectory.
• Implement idempotency for state mutations. Add an optional 'idempotency_key' parameter to updatePosition, updateRotation, updateCamera. Return the same result if called twice with the same key, preventing duplicate side effects on agent retries.
• Add tool annotations (readOnlyHint, destructiveHint, idempotentHint) to the MCP tool definitions. Mark updatePosition, updateRotation, updateCamera with destructiveHint=true to signal to agents that these modify state. Mark currPosRot, latestLog, trajectory with readOnlyHint=true.