AI Publish Pipeline Plugin

The AI publish pipeline plugin runs review, scanning, or blocking logic before an AI resource is published. It is useful for production release governance of Skills, Prompts, MCP Servers, AgentSpecs, and future AI resource types.

Pipeline belongs to AI resource governance. It can approve or reject a publish operation, but it does not change the namespace, resource name, version, or visibility model of the AI resource.

When To Use It

Enable a pipeline when:

• Skill packages need security scanning before production use.
• Prompts, AgentSpecs, or MCP Servers need format, compliance, or custom checks.
• Production release should be separated from draft editing.
• Administrators need a visible result and reason for every publish review.

If no pipeline is enabled, submitting a resource may publish it directly or move it to a publishable state. The exact behavior depends on the resource type and console flow.

Execution Model

Pipeline is an ordered chain of plugins. For each publish operation, Nacos selects nodes that support the target resource type and runs them by getPreferOrder() in ascending order.

submit publish
  -> create pipeline execution record
  -> run selected nodes in order
  -> all passed: continue publish
  -> any rejected: stop publish and keep the version unpublished

Notes:

• Only configured nodes that support the target resource type are selected.
• When one node rejects the publish, later nodes do not run.
• If pipeline is disabled or no node matches, the publish flow is not blocked by pipeline.
• Force publish bypasses pipeline validation. Use it for emergencies, not as the normal release path.

Unified plugin management can list loaded ai-pipeline plugins. Until the execution chain is fully wired to unified plugin enablement, the pipeline feature is controlled by its own configuration.

Enable the Built-in skill-scanner

The default Nacos plugin set provides a skill-scanner pipeline node. It can process scannable content in Skills, Prompts, and AgentSpecs. A common use case is calling an external Skill scanning tool.

Enable it in ${nacos.home}/conf/application.properties:

nacos.plugin.ai-pipeline.enabled=true
nacos.plugin.ai-pipeline.type=skill-scanner
nacos.plugin.ai-pipeline.skill-scanner.enabled=true
nacos.plugin.ai-pipeline.skill-scanner.command=/path/to/skill-scanner

In production, keep the same plugin versions and configuration on all Nacos Server nodes. If the scanner depends on an external executable, make sure every node can access it.

Develop a Custom Pipeline

Add the dependency:

<dependency>
    <groupId>com.alibaba.nacos</groupId>
    <artifactId>nacos-ai-plugin</artifactId>
    <version>${project.version}</version>
</dependency>

Implement com.alibaba.nacos.plugin.ai.pipeline.spi.PublishPipelineServiceBuilder and declare it with Java SPI:

META-INF/services/com.alibaba.nacos.plugin.ai.pipeline.spi.PublishPipelineServiceBuilder

Builder methods:

Service methods:

Development Advice

• Return deterministic results for the same resource version and input.
• Set timeouts when calling external systems.
• Return clear rejection reasons so resource authors can fix the problem.
• Do not modify resource content in pipeline. Use the draft editing flow for content changes.
• Do not write full Skill packages, Prompt content, keys, or credentials to logs.

Troubleshooting

Related reading: AI Resource Lifecycle and AI Resource Import Plugin.