Critical gaps across all dimensions. Descriptions exist but are vague and lack actionable context. No parameter descriptions, no output schemas, no error handling guidance. Tools appear to be inferred from hook code rather than explicitly registered with proper MCP definitions. This server would not pass code review.

• Define complete input schemas for all tools. forge__start_workflow must specify what parameters it accepts (e.g., workflow_name, input_data, timeout_ms) with types and descriptions. Currently the schema is an empty object.
• Add output schemas. Document what each tool returns: forge__start_workflow should return {workflow_id, status, created_at}; forge__update_state should return {checkpoint_id, next_step}. This enables LLM planning.
• Expand tool descriptions to 80 - 150 characters with WHEN/WHY context. E.g., 'Initiates a Forge workflow execution. Call this to start a new workflow with specified parameters. Returns workflow_id for tracking. Blocks until workflow reaches first checkpoint or timeout (30s).'
• Add parameter descriptions for all properties. E.g., duration_ms: 'Active engineering time in milliseconds since last checkpoint (0 - 3600000, required for billing accuracy)'; skill_invocations: 'Array of skill names invoked since last flush (e.g., ["code_review", "test_runner"])'.
• Implement error handling guidance. For each WRITE tool, document: (1) what can fail (e.g., 'workflow not found', 'invalid parameters'), (2) whether to retry, (3) what the LLM should do next (e.g., 'Call list_workflows() to find valid workflow names').