cleveragents-core/.opencode/skills/auto-agents-system/references/agents

Agent Reference Docs

This directory documents internal agents used by the auto-agents system. These agents are hidden infrastructure — they are not invoked directly by users or supervisors. They are called by async-agent-util as part of its internal operations.

Health Evaluation Pipeline

The health operation in async-agent-util uses a three-agent pipeline to classify each session's health state without any single agent processing the full fleet's data.

async-agent-util (health op)
  └── session-health-util   (one per session, parallel batches of 10)
        ├── Tier 1: session-health-quick-util   (always)
        └── Tier 2: session-health-full-util    (only on escalate)

Agents

Agent	Docs	Role
session-health-util	docs	Two-tier coordinator; orchestrates quick → full escalation for one session
session-health-quick-util	docs	Fast classifier; commits only when obviously finished/errored, otherwise escalates
session-health-full-util	docs	Deep analyst; evaluates timing, tools, and content with full system context

Design Principles

• No heuristics. Every session receives LLM-based classification. The quick evaluator eliminates obvious cases cheaply; the full evaluator handles the rest.
• Contextual awareness. All three agents load the auto-agents-system skill, giving them knowledge of what each agent type is supposed to achieve and what healthy vs unhealthy operation looks like.
• Fail safe. The quick evaluator defaults to escalate when unsure. healthy is the default for the full evaluator when signals are mixed.
• No data files. All data flows through prompts and bash outputs. Nothing is written to disk as an intermediate step.