agents: Build shared agent framework for authenticate/dequeue/process/report pattern

[rdwj]

Summary

Build a shared memoryhub-agent Python package that provides the common authenticate/dequeue/process/report lifecycle pattern used by all four curation agents. Agent-specific logic lives in separate modules as plugins.

Design reference

• planning/autonomous-curation-agents.md Section 13 Open Question 1 (shared framework vs independent codebases)

Scope

• Python package: memoryhub-agent with plugin architecture
• Common lifecycle: MCP authentication via register_session, Valkey queue dequeue, process loop, metrics reporting
• Per-agent configuration via environment variables: AGENT_TYPE, MH_MCP_URL, MH_API_KEY, LLM_ENDPOINT, LLM_MODEL, VALKEY_URL
• Leader election helper (for singleton agents: Curator, Statistician)
• Exponential backoff with jitter for MCP server unavailability (5 retries per tool call)
• Per-agent daily token budget enforcement via operator CRD config
• Container image: single memoryhub-agent:latest image, agent type selected by env var
• Agent-specific modules can be split out if coupling becomes a problem

Dependencies

• infra: Deploy Valkey to memoryhub-agents namespace #283 (Valkey deployment)
• Tracking: Autonomous Curation Agents (epic tracker) #281