llm-agent

Standalone prompt agent powered by OpenAI SDK.

Environment

• HUB_ENDPOINT required, agent hub websocket endpoint.
• AGENT_SERVER_ID required, agent identity for auth (example: agent:summary).
• OPENAI_API_KEY optional, required only when mapped provider is OpenAI.
• GEMINI_API_KEY optional, required when mapped provider is Gemini.
• QWEN_API_TOKEN optional, required when mapped provider is Qwen.
• QWEN_BASE_URL optional, required when mapped provider is Qwen.
• PMATE_TOKEN optional.
• PROXY_URL optional, default http://localhost:7001.
  • OpenAI base URL is derived as ${PROXY_URL}/openai/v1.
  • Gemini base URL is derived as ${PROXY_URL}/gemini.
• HEARTBEAT_INTERVAL_MS optional, default 30000.
• APP_LOG_ENABLED optional, default true.
• APP_LOG_NAMESPACE optional, default pmate.
• APP_LOG_APP optional, default llm-agent.
• APP_LOG_INPUT_PREVIEW_CHARS optional, default 200.
• APP_LOG_OUTPUT_PREVIEW_CHARS optional, default 300.
• BLOCKCHAIN_CHAIN_ID / BLOCKCHAIN_BASE_URL / INDEXER_BASE_URL required when APP_LOG_ENABLED=true.

Model routing is fixed in /Users/ramroll/dev/pmate/pmate-mono/apps/llm-agent/src/model.def.ts:

• model + responseType -> AgentModel[]
• runtime tries candidates in order with unhealthy-route skipping and same-tier fallback.

Input Payload

client.prompt({ agentId, payload }) payload is variable values only. The agent template (instruction and prompt) uses {{var_name}} placeholders. payload must provide values for those variables.

Record<string, unknown>

Output shape:

{
  type: "text" | "json"
  content: any
}

Run

pnpm --filter @pmate/llm-agent dev

C2C (Vitest)

pnpm --filter @pmate/llm-agent test:c2c

PM2

pnpm --filter @pmate/llm-agent pm2:start
pnpm --filter @pmate/llm-agent pm2:logs