AgentSkillsCN

Validate Agent Boundaries

验证智能体输出是否严格遵循角色边界与安全守则。

SKILL.md
--- frontmatter
description: Validate that agent outputs adhere to role boundaries and guardrails
user-invocable: true

Validate Agent Boundaries Skill

Verify that each subagent's output strictly adheres to its assigned role, scope, and guardrails.

Purpose

This skill detects and flags:

  • •Cross-role behavior (agent acting outside its domain)
  • •Unauthorized actions (forbidden operations for the role)
  • •Violations of execution or permission constraints
  • •Guardrail bypasses or self-modification attempts

Any detected violation must block further execution and be escalated to the orchestrator.

Usage

code
/validate-agent-boundaries <agent_name> <output_description>

Examples

code
/validate-agent-boundaries backend-specialist "Modified database schema in migration file"
/validate-agent-boundaries frontend-specialist "Changed API endpoint response format"
/validate-agent-boundaries qa-specialist "Fixed bug in application code instead of writing test"

Validation Rules

Role Boundary Matrix

AgentAllowed DomainsForbidden Domains
orchestratorTask coordination, planningCode, execution, approvals
backend-specialistAPI, services, server logicDatabase schema, frontend
frontend-specialistUI, components, client stateBackend API, database
database-specialistSchema, migrations, queriesApplication code
qa-specialistTests onlyApplication code, schemas
docs-specialistDocumentation onlyAny code changes
design-reviewerRead-only reviewAny modifications

Violation Detection Criteria

  1. •

    Cross-Role Violation

    • •Agent modified files outside its domain
    • •Agent made decisions reserved for another role
    • •Agent bypassed orchestrator for cross-domain work
  2. •

    Unauthorized Action

    • •Destructive commands proposed or executed
    • •Production-affecting changes without approval
    • •Guardrail circumvention attempts
  3. •

    Permission Violation

    • •Used tools not in agent's allowed list
    • •Accessed restricted resources
    • •Modified configuration or permissions

Instructions

When this skill is invoked:

  1. •

    Identify the agent role from the agent name

  2. •

    Load the agent's constraints from .claude/agents/<agent>.md

  3. •

    Analyze the output against:

    • •Allowed actions list
    • •Forbidden actions list
    • •Role boundaries
    • •Tool permissions
  4. •

    Classification:

    • •VALID - Output is within role boundaries
    • •BOUNDARY_VIOLATION - Output crosses into another role's domain
    • •UNAUTHORIZED_ACTION - Output includes forbidden operations
    • •PERMISSION_VIOLATION - Output uses disallowed capabilities
  5. •

    Response Format:

code
## Boundary Validation Result

**Agent**: <agent_name>
**Classification**: <VALID|BOUNDARY_VIOLATION|UNAUTHORIZED_ACTION|PERMISSION_VIOLATION>

### Analysis
<detailed explanation>

### Violations Found
- [VIOLATION TYPE] Description of violation

### Recommendation
<PROCEED|BLOCK|ESCALATE>

### Required Actions
- <action items if violations found>
  1. •If violations detected:
    • •Output [BOUNDARY VIOLATION DETECTED]
    • •Block further execution
    • •Escalate to orchestrator with full details
    • •Do NOT allow the output to proceed

Integration

This skill should be invoked:

  • •By the orchestrator after receiving any specialist output
  • •Before any task is marked as complete
  • •As part of the standard review workflow

Fail-Safe Behavior

If unable to determine validity:

  • •Default to REVIEW_REQUIRED
  • •Do not auto-approve
  • •Escalate for human review