@agentxin-ai/plugin-middleware-reflection

Post-execution reflection middleware for AgentXinAI agents — evaluates agent performance, distills SOPs from high-scoring executions, and archives known errors from failures.

Features

• Quality Evaluation — LLM-based scoring across tool selection, process adherence, and resolution quality (1–5 scale)
• SOP Distillation — Automatically extracts Standard Operating Procedures from high-scoring executions into a knowledge base
• KEDB Archiving — Archives failure patterns into a Known Error Database for future retrieval
• Knowledge Retrieval Augmentation — Injects relevant SOPs/KEDBs into agent system prompts via wrapModelCall
• Custom KPIs — Add domain-specific evaluation metrics
• Blended Scoring — Combines LLM evaluation with optional user ratings

Installation

npm install @agentxin-ai/plugin-middleware-reflection

Activation

Add the plugin to the PLUGINS environment variable in your AgentXinAI host project:

PLUGINS=@agentxin-ai/plugin-middleware-reflection

For multiple plugins, separate with commas:

PLUGINS=@agentxin-ai/plugin-middleware-reflection,@agentxin-ai/plugin-agent-middlewares

See Plugin Publish & Use for full host-loading documentation.

Configuration

Once loaded, navigate to an Agent's middleware settings in AgentXin Studio and add Reflection Middleware. Available options:

| Field | Description | |---|---| | Evaluation Model | LLM used for reflection scoring, SOP distillation and KEDB archiving | | Domain Context | Business domain description for context-aware evaluation | | SOP Knowledge Base | Knowledge base ID for storing distilled SOPs | | SOP Distill Threshold | Minimum overall score (1–5) to trigger SOP distillation (default: 4) | | KEDB Knowledge Base | Knowledge base ID for archiving known errors | | Error Archive Threshold | Maximum overall score (1–5) to trigger KEDB archiving (default: 2) | | Custom KPIs | Domain-specific evaluation metrics (key, type, description) | | Reflection Prompt | Advanced: fully override the built-in evaluation prompt | | SOP Distill Prompt | Advanced: override the SOP distillation prompt | | KEDB Prompt | Advanced: override the KEDB archiving prompt |

How It Works

User ──► Agent executes ──► afterAgent hook fires
                                │
                          ┌─────▼──────┐
                          │  LLM Score  │  (Phase 1: evaluate execution)
                          └─────┬──────┘
                                │
                    ┌───────────┼───────────┐
                    ▼           │           ▼
              Score ≥ 4         │      Score ≤ 2
           ┌────────┐          │      ┌────────┐
           │Distill │          │      │Archive │
           │  SOP   │          │      │  KEDB  │
           └────────┘          │      └────────┘
                               ▼
                     (no write for 3)

On the next conversation turn, wrapModelCall retrieves relevant SOPs/KEDBs and injects them into the agent's system prompt.

Development

# Build
npx nx build middleware-reflection

# Test
npx nx test middleware-reflection

# Publish
npx nx run @agentxin-ai/plugin-middleware-reflection:nx-release-publish --access public

License

BUSL-1.1