auto-meta-agent

This repository contains a reusable agent runtime plus its own method memory for a shared daily-notes -> doc -> sop promotion chain.

Each run serves one active memory_root.

• repo-local runs default to this repository's own memory/
• external agent-runtime.* overlays may point the runtime at another workspace memory tree

Layout

• main.py
  • unified scheduler entrypoint and the only poller in the formal system
• config.yaml
  • repository-default scheduler settings
• memory/doc/
  • design docs and methodology docs for the runtime
• memory/daily-notes/
  • ongoing evolution notes for this repository
  • filename format: YYYY-MM-DD-<index>-<title>.md
• memory/.agents/
  • repo-local runtime artifacts when the active memory_root is this repository's own memory/
• <agent-name>/
  • one self-contained agent workspace with agent.json and system.md
• runners/
  • launcher adapters such as codex
• lib/
  • shared runner-side utilities

main.py coordinates the full notes -> doc -> sop chain. Agents do not poll on their own.

Data Flow

• note-relation reads the shared daily-notes layer and writes machine-readable work orders into <memory_root>/.agents/note-relation/work-orders/
• doc-maintenance consumes upstream note-relation work orders and writes downstream work orders into <memory_root>/.agents/doc-maintenance/work-orders/
• sop-promotion consumes upstream doc-maintenance work orders and writes promotion results into <memory_root>/.agents/sop-promotion/work-orders/

The scheduler handles polling, wakeup, and dependency-aware execution order. Agents exchange data through runtime work-order files instead of introducing a new long-term memory layer.

main.py only wakes agents and passes context. Each agent still inspects its own scope and decides the concrete processing set.

Launchers

Agent launch is runner-agnostic.

• agent.json declares launcher.kind and launcher.config
• main.py resolves the launcher adapter under runners/
• launcher-private files live under <memory_root>/.agents/<agent-id>/launchers/<launcher-kind>/

Current implementation includes the codex launcher as one adapter, not as the system architecture itself.

Configuration

config.yaml contains repository-default runtime settings:

• watch
• sleep_seconds
• memory_root
• default_model
• agent_models
• shared_memory_paths
• agent_task_overrides

The scheduler derives agent runtime state from that active memory root:

• <memory_root>/.agents/<agent-id>/state/
• <memory_root>/.agents/<agent-id>/work-orders/
• <memory_root>/.agents/<agent-id>/reports/
• <memory_root>/.agents/<agent-id>/launchers/<launcher-kind>/

When --config is omitted, main.py first looks for an external agent-runtime.yaml, agent-runtime.yml, or agent-runtime.json deployment overlay and otherwise falls back to config.yaml. The active workspace root is derived from the config file location.

Entrypoints

• python main.py
  • auto-discovers an external agent-runtime config first
• python main.py --config path/to/config.yaml
  • runs with an explicit JSON/YAML config file

Conventions

• Keep long-term knowledge in memory/, not in runner-private folders.
• Keep durable runtime coordination under the active memory root's .agents/ namespace.
• Keep the scheduler decoupled from any one launcher implementation.
• Keep main.py responsible for wakeup and orchestration, not domain judgment.
• Keep the daily-notes -> doc -> sop chain explicit without adding a new long-term memory tier.
• Keep repository-local daily notes under memory/daily-notes/ with zero-padded per-day numbering and a short kebab-case title slug.