PartMode exposes 7 tools with severe definition gaps. Five tools (normalizeStudioSelectionFilter, nextStudioSelectionFilter, studioSelectionBounds, studioBoxSelection, studioSelectOther) are UI selection utilities with minimal descriptions (10-80 chars, well below the 194-char baseline). Descriptions lack WHEN/WHY context required for LLM tool selection. Input schemas present but sparse: most parameters lack type constraints, enums, or validation guidance. handleMcpRequest and createHeadlessAgentHost are vaguely described protocol/infrastructure tools. No output schemas documented. No error handling guidance. Tool names lack clear verb_noun patterns (e.g., 'studioBoxSelection' is ambiguous, does it select, filter, or compute?). No parameter descriptions explain constraints, formats, or dependencies. Security: createHeadlessAgentHost accepts projectId/name but no auth/permission checks visible. Overall: definitions are infrastructure-focused, not LLM-optimized.

• Expand all tool descriptions to 150 - 250 chars. Include WHAT the tool does, WHEN to use it (vs similar tools), and WHAT it returns. Example: 'Normalizes a studio selection filter to a valid type (auto|window|crossing). Use this before calling studioBoxSelection to ensure the filter mode is recognized. Returns the normalized filter string or default auto.'
• Add enums and constraints to parameters. For normalizeStudioSelectionFilter.value, define enum: ['auto', 'window', 'crossing']. For studioBoxSelection.candidates, specify: 'Array of objects with required key (string) and bounds (object with x, y, width, height), optional depth (number).'
• Document output schemas for all tools. Example for normalizeStudioSelectionFilter: 'Returns {normalizedFilter: string, isValid: boolean}'. For studioBoxSelection: 'Returns {selectedKeys: string[], count: number, bounds: {x, y, width, height}}'.
• Add error handling guidance. For handleMcpRequest, document: 'Returns error if request format is invalid. On failure, returns {error: string, code: string, retryable: boolean}. If retryable, agent may retry; if not, escalate to user.'
• For createHeadlessAgentHost, add permission checks and dry-run support. Document required permissions (e.g., 'admin:project'). Add optional 'dryRun' parameter to validate without creating. Return {projectId, name, created: boolean, warnings: string[]}.