Inspect — Logic Artifact Inspection
The entry point for AI Agent defect discovery. Systematically scans System Prompts, Skills, and Tool Schemas for quality defects and consistency breaks before deployment.
Why Inspect?
The reliability of an AI Agent depends on the quality of its Logic Artifacts — System Prompts, Skill descriptions, and Tool Schemas. These artifacts are highly susceptible to hidden defects introduced during the design phase:
- Contradictory instructions, scope ambiguity, and implicit assumptions in Prompts
- Unbounded execution, resource runaway, and permission overflow in Skills
- Missing parameter constraints, semantic ambiguity, and unprotected high-risk operations in Tool Schemas
- Authorization mismatches, trigger condition conflicts, and broken constraint chains across artifacts
Inspect serves as the first static quality defense line — before the Agent enters Risk scanning and Quality assessment.
Five Sub-Specifications
Inspect Prompt
Static defect inspection framework for System Prompts.
Covers contradiction, mismatch, scope definition, redundancy, implicit assumptions, gray zones, and other defect types. Supports L1/L2/L3 tiered inspection and P0/P1/P2 defect severity classification.
Numbering system: QD-P-x.y
Inspect Skill
Quality defect inspection standard for Skills.
Entry gate (QD-S-0) and five defect categories: Resource Runaway (QD-S-1), Unbounded Directive (QD-S-2), Ambiguity-Induced Inference (QD-S-3), Failure Escalation (QD-S-4), Permission Overflow (QD-S-5). Includes the six-dimension Review Schema analysis framework.
Numbering system: QD-S-x.y
Inspect Tool
Defect inspection specification for Tool Schemas.
Covers Structural Validity (QD-T-1.x), Parameter Constraints (QD-T-2.x), High-Risk Operations (QD-T-3.x), Semantic Clarity (QD-T-4.x), and Schema Change and Compatibility (QD-T-5.x).
Numbering system: QD-T-x.y
Inspect Cross
Cross-Artifact consistency inspection standard.
The Cross sub-specification currently contains six groups of inspection relationships: Prompt → Skill (QD-PS), Prompt → Tool (QD-PT), Skill ↔ Tool (QD-ST), Skill ↔ Skill (QD-SS), Tool ↔ Tool (QD-TT), and circular invocation (QD-CYC), with 29 inspection items that uncover hidden defects undetectable by single-artifact inspection.
Numbering system: QD-PS/PT/ST/SS/TT/CYC-n.m
Inspect Permission
Permission-responsibility proportionality inspection standard.
Six inspection groups (QD-PM-1~6): alignment, scope, granularity, propagation, boundary, and auditability.
Numbering system: QD-PM-x.y
43 inspection items across six dimensions.
Defect Severity Levels
| Level | Meaning | Disposition |
|---|---|---|
| P0 | Blocking: prevents Agent functionality or introduces security risk | MUST be fixed |
| P1 | Warning: may cause behavioral instability or functional defects | SHOULD be fixed |
| P2 | Advisory: affects readability or efficiency, does not impact functionality | MAY be fixed |
Tiered Inspection Mechanism
Different Agent risk levels apply different inspection intensity:
| Agent Level | Typical Characteristics | Inspection Strategy |
|---|---|---|
| AC-L1 | Text processing only, no tool invocations | Minimum rule set |
| AC-L2 | Resource access, medium permissions | Standard rule set |
| AC-L3 | Sensitive operations, high permissions | Full rule set |
Position Within the SanityOps Framework
Inspect (Defect Inspection) ── You are here
│
▼
Risk (Risk Scanning) ── Explicit + Implicit
│
▼
Quality (Service Quality) ── RAG Agent + Tool Agent
│
▼
Relevance (Relevance Assessment)Inspect is the first gate of SanityOps. Discover defects → classify severity → recommend fixes, ensuring that only quality-verified Logic Artifacts proceed to downstream stages.
Quick Start
- Read the Overview for the full SanityOps framework picture
- Navigate to the sub-specification matching your artifact type and begin inspection
- After completing single-artifact inspection, run Inspect Cross for cross-artifact consistency inspection
- Finally, run Inspect Permission for permission-responsibility proportionality inspection