Open-source AI agent operating system for WordPress — capability-first, policy-aware, self-hosted agent orchestration layer with memory abstraction, policy enforcement, and WordPress/WooCommerce integration
WordPress Agent OS exposes 6 tools focused on approval workflows. All tools have descriptions (10-150 chars) and input schemas with typed parameters. However, critical gaps undermine production readiness: (1) No output schemas documented, LLMs cannot infer response structure for chaining; (2) Parameter descriptions are minimal (5-30 chars) and lack constraints, ranges, or format guidance; (3) No error handling guidance, tools return raw PHP errors with no recovery hints; (4) Missing idempotency markers and confirmation patterns for destructive operations (approve/reject); (5) Tool composition is narrow, approval workflow is isolated from WordPress capabilities it gates. Per-tool scores: request_approval 42, get_approval 35, approve 38, reject 38, list_approvals 40, list_pending 35. Average: 38.
Approve a pending approval request, with support for multi-operator approval workflows
Retrieve a specific approval request by ID
List all approval requests with optional limit parameter, sorted by creation date descending
List all pending approval requests (status 'pending' or 'pending_multi')
Reject a pending approval request with an optional reason
Request approval for a capability execution with trace ID, capability name, risk level, policy level, requested by, description, payload, rollback plan, and optional multi-operator requirement and dry-run requirement
No output schemas documented. LLMs cannot infer response structure (fields, types, pagination) for downstream tool chaining or data extraction. Approval responses must include approval_id, status, created_at, and operator fields for agent planning.
Parameter descriptions are too brief (5-30 chars) and lack actionable constraints. E.g., 'risk_level' has no enum values (low|medium|high|critical?), 'policy_level' lacks examples, 'limit' has no min/max bounds. LLMs cannot validate inputs without explicit constraints.
No error handling guidance. Tools return raw PHP errors (e.g., 'Approval not found') with no recovery hints. LLMs need actionable errors: 'Approval ID not found. Call list_pending() to see available approvals.' or 'Permission denied, only operators can approve.'
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 28 | <=2025-11-25 | v2 |
Destructive operations (approve, reject) lack confirmation/dry-run patterns. Agents can irreversibly approve/reject without a confirmation step. No idempotentHint or destructiveHint annotations present.
Approval workflow is isolated from WordPress capabilities. No tools to execute approved capabilities, query capability definitions, or link approvals to actual WordPress actions (post creation, user management, etc.). Approval gate exists but has no integration point.