Senparc.Xncf.AgentsManager

Senparc.Xncf.AgentsManager

Senparc.Xncf.AgentsManager is an NCF module for managing reusable AI-agent templates, chat groups, tasks, collaboration graphs, prompt optimization, and usage records.

Features

• Stores and manages AgentTemplate, ChatGroup, ChatGroupMember, ChatTask, and chat history data.
• Supports group execution, task streaming, agent collaboration graphs, and snapshot responses.
• Provides prompt optimization and prompt-catalyzer application services.
• Lets each AgentTemplate bind to one published KnowledgeBase without a cross-module database foreign key.
• Shows per-Agent completed conversation rounds/tasks, prompt/completion/total tokens, average response time, and last activity on the management page.
• Retrieves bounded KnowledgeBase context before model execution, labels it as untrusted external data, and falls back to the model when retrieval is unavailable.
• Exposes local application-service requests and DTOs suitable for NCF function rendering.
• Includes provider-specific EF Core contexts for NCF's multi-database model.
• Supports A2A remote agents as first-class ChatGroup members while preserving the existing local AgentTemplate and ChatGroupMember data path.
• Supports per-group context sharing: legacy full history (local-only), instruction plus bounded text conclusion (the secure mixed-group default), or instruction only.

Installation

<PackageReference Include="Senparc.Xncf.AgentsManager" Version="0.13.0-preview8" />

Key API

• AgentTemplateService and AgentTemplateAppService create, find, configure, and inspect agent templates.
• ChatGroupService, ChatGroupMemberService, and ChatGroupHistoryService manage group membership and conversation records.
• ChatTaskService, ChatTaskAppService, and ChatTaskStreamController coordinate task execution and streaming output.
• PromptOptimizationAppService and PromptOptimizationAgentBridge integrate prompt improvement workflows.
• AgentsTemplateService and the AgentGraph* DTOs support reusable collaboration graphs.

KnowledgeBase-first RAG policy

KnowledgeBase-first is implemented as retrieve-first, conditionally augment, not as a forced answer override. A bound Agent queries its published collection with the user's request, receives at most five chunks and 6,000 characters, and is instructed to use only relevant evidence and identify gaps or conflicts. Empty or failed retrieval is fail-open so an unavailable vector service does not stop the whole multi-Agent task.

This policy is appropriate because it improves domain grounding without pretending every query belongs to the knowledge base. It still needs production evaluation for similarity thresholds/reranking, document ACLs, prompt-injection filtering, and citation rendering. Token totals are based on provider usage metadata stored with completed chat history; providers that do not return usage cannot be estimated exactly.

A2A remote agent configuration

Use the 远程 A2A 智能体 entry on the Agent management page to register an Agent Card root URL or a /.well-known/agent-card.json URL, test card discovery, and add it to a ChatGroup alongside local agents. The current phase keeps the group owner and entry agent local; the remote agent participates in normal collaboration rounds.

Remote authentication is deliberately reference-only. The database stores AuthSecretKey, never a token value. Configure the real secret in the deployment configuration, for example:

{
  "A2A": {
    "Secrets": {
      "research-agent-token": "replace-with-deployment-secret"
    }
  }
}

Publish a local Agent as an A2A service

An existing local AgentTemplate can also be exposed to another NCF installation, or to any A2A-compatible client, without coupling the two systems. Open the local Agent editor and select 配置 / 发布为 A2A Agent. The publishing configuration is stored in the additive PublishedA2AAgent table; it does not change existing AgentTemplate, ChatGroup, or ChatGroupMember records.

When enabled, the standard discovery URL is:

https://your-public-host/a2a/{public-agent-key}/.well-known/agent-card.json

The Agent Card advertises the standard A2A JSON-RPC interface at https://your-public-host/a2a/{public-agent-key}. A second NCF system can paste the discovery URL into its 远程 A2A 智能体 configuration, test the card, then add that remote Agent to a ChatGroup.

For reverse-proxy deployments, configure the externally visible base address so the Agent Card never advertises an internal host name:

{
  "A2A": {
    "PublicBaseUrl": "https://your-public-host",
    "InboundSecrets": {
      "partner-ncf": "replace-with-deployment-secret"
    }
  }
}

Bearer Token and custom-header incoming authentication read only from A2A:InboundSecrets:{AuthSecretKey}. The secret body is not stored in the database, returned by the admin API, included in logs, or put into the Agent Card. Publishing is disabled by default. Function Calling and MCP tools are also disabled by default for A2A requests; enabling them is an explicit per-published-Agent decision for trusted callers only.

The A2A boundary instructs the local Agent to return shareable conclusions rather than hidden reasoning, system prompt text, secrets, or tool traces. It cannot prevent a model from intentionally placing sensitive content in a normal response, so public Agent descriptions, prompts, connected knowledge bases, and enabled tools still require a deliberate data-access review.

Choose research-agent-token as the remote agent's deployment key name. The default mixed-group policy forwards only the initial instruction and the current round's bounded text conclusion. It removes tool calls, raw provider representations, usage data, and earlier history from the participant broadcast. This reduces exposure but does not transform an agent's voluntarily emitted text into a guaranteed semantic summary; prompts therefore require a concise shared conclusion and prohibit chain-of-thought/private content.