m3taz-ahmed/aizee / skills/agent-governance-lord
agent-governance-lord
VerifiedLord skill for governing AI coding agents at runtime — gateway interception, agent/flow/model allowlisting, and MCP-server-as-securable. The control layer between an agent's decision and its real-world action.
Agent Governance Lord
[OBJ] Runtime governance for AI coding agents: intercept, allowlist, and audit every agent action.
Problem
AI agents can write, push, and deploy without human review at each step. Access control answers "can the agent do this?" — governance answers "should the agent do this?" Permission is necessary; it is not sufficient. An agent can be fully permitted and still wrong.
Rules
- [REQ] Gateway on the path. Every LLM request and MCP tool call MUST pass through
runtime/agent_gateway.py(pre-LLM + post-execution guardrails). No bypass. - [REQ] Three verdicts. Guardrails emit ALLOW / REDACT / BLOCK. Most severe wins (BLOCK > REDACT > ALLOW). Never silently drop a BLOCK.
- [REQ] Pre-LLM guardrails. Inspect outbound prompt + tool payloads BEFORE the model sees them. Catch: secret leakage, prompt injection in tool responses fed back, forbidden tool calls.
- [REQ] Post-execution guardrails. Inspect inbound response + generated code BEFORE the developer sees it. Catch: destructive commands, insecure patterns, hallucinated APIs.
- [REQ] Agent allowlist. Only agents registered in
runtime/agent_catalog.pywithstatus == ALLOWEDmay run. Unknown agent = BLOCK + alert. - [REQ] Flow allowlist. Each agent has an
allowed_flowslist. A flow not in the list = BLOCK. No blanket flow access. - [REQ] Model allowlist. Each agent has an
allowed_modelslist. A model not in the list = BLOCK. Prevents unauthorized model escalation. - [REQ] MCP-as-securable. MCP servers registered in
runtime/mcp_securable.pyas governed assets. Access via GRANT policies (USE/ADMIN/REGISTER). No ungoverned MCP server. - [REQ] Composite identity. Every agent action attributed to BOTH agent + human principal (
runtime/composite_identity.py). No anonymous agent actions. - [REQ] Audit every verdict. Every ALLOW/REDACT/BLOCK logged with agent_id, user_id, tool, reason. Tamper-evident via existing audit chain.
- [REQ] Rate limiting. Per-agent rate limits enforced at the gateway. Burst > threshold = BLOCK + cooldown.
- [REQ] Cost attribution. Per-agent cost tracked via
runtime/cost_attribution.py. Anomaly (spike/budget breach) = alert + optional throttle. - [REQ] Plan validation. Before edits, validate the agent's plan via
runtime/plan_diff_validator.py. Forbidden paths = BLOCK. File count > max = WARN. - [REQ] Diff validation. After edits, validate the git diff. Undeclared imports = WARN. Test gap = WARN. Unrelated refactor = WARN.
- [REQ] Supply-chain guard. New imports not in lockfile = WARN. External package not declared = BLOCK in strict mode.
- [REQ] Human-in-the-loop. Irreversible/destructive actions (rm, drop, force-push, deploy to prod) require explicit human approval. Gateway holds the action pending approval.
- [REQ] Rollback path. Every agent action that modifies state must have a documented rollback. No rollback = BLOCK.
- [PROHIBIT] Bypassing the gateway for "trusted" agents. All agents pass through.
- [PROHIBIT] Allowing an agent to call an ungoverned MCP server.
- [PROHIBIT] Executing a destructive action without composite identity attribution + human approval.
- [REQ] SARC enforcement sites. Implement 4 enforcement sites: Pre-Action Gate (block injection/PII/policy violations before LLM call), Action-Time Monitor (rate/cost/budget enforcement, circuit breakers), Post-Action Auditor (log, evaluate, score for reliability@k), Escalation Router (human approval, kill-switch, incident creation).
- [REQ] Agent SRE Governance v1.0. Implement SLOs, error budgets, circuit breakers, chaos engineering, trace replay, Ed25519 artifact signing, SBOMs, and OpenTelemetry for all agent deployments. Reference: Microsoft Agent Governance Toolkit.
- [REQ] Microsoft Agent Hooks. Support framework-neutral governance contracts — "deny means deny" enforceable across LangChain, CrewAI, OpenAI Agents SDK, etc.
- [REQ] MCP 2026-07-28 stateless governance. Update gateway for stateless MCP protocol. No session state assumptions. Per-request protocolVersion validation. Header-based routing (Mcp-Method, Mcp-Name).
- [REQ] Per-PR spend limits. Implement per-PR spend caps (soft block) and org-level ACU visibility. Devin enterprise pattern.
- [REQ] Enterprise MCP server allowlist. Maintain allowlist of approved MCP servers. Block ungoverned servers. Enterprise plugin governance (required/optional/forbidden plugins).
- [REQ] 5-layer control plane. Gateway Layer (auth, rate limiting, cost tracking) → Policy Layer (4-level hierarchy, fail-closed) → Observability Layer (tamper-evident audit) → Governance Layer (approvals, kill-switch, RBAC) → Integration Layer (A2A, MCP, REST, webhooks).
- [PROHIBIT] Deploying agents without SARC 4-site enforcement and Agent SRE Governance compliance.
Enforcement Stack
Agent request → AgentGateway.check_request (pre-LLM)
→ AgentCatalog.is_agent_allowed
→ AgentCatalog.is_flow_allowed_for_agent
→ AgentCatalog.is_model_allowed_for_agent
→ McpSecurableRegistry.check_permission (if MCP)
→ LLM call
→ AgentGateway.check_response (post-execution)
→ PlanDiffValidator.validate_diff (if code generated)
→ SupplyChainGuard.check_diff (if imports added)
→ CostAttribution.record
→ AuditLogger.log (composite identity)
References
- Fiddler AI: "controls only hold where they sit on the agent's request and response path."
- GitLab Duo: composite identity, agent/flow catalog, tool approval guardrails.
- Databricks Unity Catalog: MCP servers as securables with GRANT policies.
- repo-contract: negative-constraint enforcement (most effective for safe agent workflows).
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 471k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |