Verification Engine direction
Roadmap · after v1.0Agent Governance
Planned for the Verification Engine. Let AI agents read engineering facts through MCP tools under role-based access — while every write still requires human-reviewed patch preview (Gate 2).
What is agent governance?
When coding agents act on your repo, governance answers: what may they read?, what may they propose?, and who must approve writes? ContextMint's model: agents get read tools on the engineering graph; mutations queue as proposals — never silent auto_applied on your main branch.
Planned MCP tools (read)
# Target tool surface (read-only)
get_blast_radius(target) → impacted services + evidence paths
get_feature(name) → feature registry entry
get_decision(id) → decision / ADR graph trace
get_trace(deployment) → CI/CD path (enterprise, later)
get_blast_radius(target) → impacted services + evidence paths
get_feature(name) → feature registry entry
get_decision(id) → decision / ADR graph trace
get_trace(deployment) → CI/CD path (enterprise, later)
MCP = Model Context Protocol — a standard for agents to call typed tools. See Glossary → MCP.
Policy engine
- One policy module for agents, assisted actions, and scorecards
- Custom rules (ownership, PCI boundaries, blast thresholds)
- Responses validate against AgentContextContract JSON schema
Safety defaults
- No auto-apply on main workspace until execution gates pass
- Graph proposals queue for human approval
- Audit trail for agent propose events
Govern agents before they govern your repo
Walk through MCP contracts and RBAC on a design-partner roadmap session.