Shadow exposes 53 tools but lacks visible input schemas and parameter documentation in source code. Tool definitions appear scattered across compiled assets (index-D6GUFT8I.js) and server.ts without explicit JSON Schema registration. Descriptions are present but minimal (10-30 chars), insufficient for LLM tool selection. No evidence of parameter type definitions, constraints, or error recovery guidance. Tools follow a consistent naming convention (shadow_*) but lack the depth required for production agent use.

• Add explicit JSON Schema definitions for all 53 tools. Each tool must declare inputSchema with type, properties, required array, and descriptions for every parameter. Example: shadow_memory_teach should specify { type: 'object', properties: { memory: { type: 'string', description: 'The memory content to teach...' }, category: { type: 'string', enum: ['personal', 'work', ...] } }, required: ['memory'] }.
• Expand tool descriptions to 50-200 characters. Include WHAT the tool does, WHEN to use it, and what it returns. Example: 'Create a new memory. Use when Claude learns something about the user. Returns memory_id and confirmation timestamp.' Current 'Teach a new memory' is too vague.
• Document every parameter with type, format, constraints, and examples. For shadow_memory_update, specify: 'memory_id (string, UUID format): The ID of the memory to update. memory_content (string, 1-5000 chars): New content. tags (array of strings, optional): Labels like ["important", "work"].'
• Add error recovery guidance to tool descriptions. For destructive tools, state: 'This operation is irreversible. If it fails, the agent should call shadow_memory_list to verify the memory was deleted.' For write tools, specify which errors are retryable (network timeout) vs user-fixable (invalid memory_id).
• Implement confirmation pattern for destructive operations. shadow_memory_forget, shadow_task_remove, and shadow_repo_remove should support a dry_run parameter or return a confirmation_required response with a confirmation_token that must be passed in a follow-up call.

Inferred effective spec: 2026-07-28+.

• Relies on Stateful initialize / Mcp-Session-Id (removed; protocol is stateless) - make each request self-contained