GovForge defines 13 tools with consistent naming (verb_noun pattern: create_task, record_decision, etc.) and complete input schemas. All tools have descriptions (avg ~120 chars) and typed parameters with enums where appropriate (risk_level, verdict). However, output schemas are entirely undocumented, no tool specifies what fields are returned, forcing LLMs to guess downstream field names. Error handling is absent: no guidance on retryability, user-fixable vs fatal errors, or recovery paths. Parameter descriptions are minimal (avg ~40 chars) and lack format/constraint details. No tool declares permissions, audit trails, or idempotency guarantees. Composition is sound (each tool has one responsibility), but chaining is fragile without documented output schemas.

• Document output schemas for all 13 tools. Specify returned fields (e.g., create_task returns {task_id: string, created_at: ISO8601, status: string}). This enables LLMs to chain tools correctly.
• Add error handling guidance to tools that modify state (approve_decision, reject_decision, record_decision). Include recovery hints: 'If decision_id is invalid, call get_decision_context() with a known task_id to discover valid decisions.'
• Expand parameter descriptions to 50 - 100 chars. For 'focus' array in request_review, specify: 'Array of focus areas (e.g., ["security", "performance"]). Max 5 items. Optional.'
• Declare permissions for each tool. Add to descriptions: 'Requires: approve:decision' for approve_decision, 'Requires: read:decision' for get_decision_context. Enable least-privilege agent configs.
• Add pagination to list_open_reviews and list_pending_approvals. Accept limit (1 - 100, default 20) and offset/cursor parameters. Return total_count and next_cursor in response.