Governance mesh for AI agent tool calls — MCP server that enforces policy-based access control, approval workflows, and temporal grants for tool execution
flux7-mesh defines 10 tools with mixed quality. Most tools have basic descriptions (10-50 chars), but lack depth on WHEN to use them or what they return. Input schemas are present but sparse: many tools have minimal parameters with no type constraints or enums. Critical gaps: no output schemas documented, no error handling guidance, no parameter validation rules stated. Tool names are reasonable (verb_noun style) but descriptions are too terse. The governance-focused tools (approval.*, grant.*) are better documented than basic utilities (echo, risky_tool). No evidence of idempotency markers, permission gates, or audit trail patterns.
List all pending approval requests
Approve or deny a pending approval request
Echo back input
Read a file
Get an order
Create a temporal grant (supervisor-only)
List all active temporal grants
risky_tool has no description (15 chars), no input schema, and a generic name that violates verb_noun convention. Agents cannot determine when or why to call it.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., what does approval.pending return? Is it an array? What fields per item?).
Tool descriptions are too terse (10-60 chars). Baseline for A+ tools is 50-200 chars. Descriptions lack WHEN to use, prerequisites, or what the tool returns. E.g., 'List all pending approval requests' does not explain the response structure or how to use it with approval.resolve.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 49 | <=2025-11-25 | v2 |
Revoke a temporal grant (supervisor-only)
List all available tools with their policies
Risky
Parameter descriptions are missing or minimal. E.g., grant.create has 'Agent ID', 'Tool pattern (glob)', 'Duration (e.g. 1h, 30m)' but no explanation of what 'Agent ID' format is expected, whether glob patterns are required, or what happens if duration is invalid.
No error handling guidance. Tools like fs.read_file and get_order have no documented error cases, recovery steps, or actionable error messages. Agents cannot self-correct on failure.