Beatless — Autonomous Agent Constellation

Hybrid AI orchestration system for open-source contribution, technical blogging, and ML research. Hermes Agent handles scheduling and information gathering; Claude Code handles deep execution.

Current State: Constellation v3

Hermes Agent (Kimi K2.6 orchestrator)
  ├── Cron: 4 active jobs
  │     ├── GitHub Response    — hourly PR comment triage
  │     ├── GitHub PR Pipeline — hourly issue discovery → full PR submission
  │     ├── Auto Research      — 4h experiment analysis cycles
  │     └── Blog Maintenance   — 12h content audit + writing (MiniMax M2.7)
  │
  ├── Models
  │     ├── Kimi K2.6      — orchestration, planning, review
  │     ├── Step 3.5 Flash — fast execution, tool chains, web search
  │     └── MiniMax M2.7   — writing, image gen, TTS, video, documents
  │
  └── Wake-gate scripts → Claude Code (on-demand)
        ├── /github-pr       — 12-phase PR pipeline with triple review
        ├── /pr-followup     — maintainer comment response
        └── /exp-*           — ML experiment lifecycle (see below)

Experiment Command Pack (exp-*)

Five commands encoding a two-path research methodology for ML experiments:

Command	Purpose
/exp-status	Workspace readiness diagnostic (GPU, data, plugins)
/exp-init	Initialize experiment branch, planning files, baseline run
/exp-discover	Generate hypotheses via idea-first or application-first path
/exp-run	Autonomous experiment loop (quick: single-GPU / full: dual-GPU A/B)
/exp-review	Multi-agent review with continue/pivot/rollback/halt verdict

Integrates: Codex (code edits), Gemini (literature + direction review), Superpowers (brainstorming), GSD (verification), Planning-with-files (state persistence).

PR Pipeline

12-phase process from issue discovery to PR submission:

1. Discover claimable issues (good first issue, help wanted, bug)
2. Evaluate repo (CONTRIBUTING.md, recent PRs, test infrastructure)
3. Fork, clone, baseline tests
4. Implement fix (Codex write-mode)
5. Triple review (Gemini correctness + Codex architecture + Claude quality gate)
6. Submit PR with evidence-based scoring

Quality controls: anti-inflation (no self-review), revert-test-reapply verification, minimum 7.5/10 score gate.

Repository Structure

commands/exp/           # Active: exp-* command pack (903 lines)
design/                 # Architecture: CONSTELLATION v1 → v3 evolution
standards/              # PR guidelines, contribution protocols
pipelines/              # Active pipeline specs (github-pr.md, blog-maintenance.md)
docs/                   # HERMES integration, migration status
agents/aoi/             # Aoi — scheduler persona (SOUL.md)
archive/                # Deprecated v2 infrastructure
  ├── v2-deprecated/    #   Heartbeat agents, shell runners, harness scripts
  └── deprecated-commands/  #   research-analyze.md, research-train-loop.md

Planned (Next Stages)

• Aoi — Digital persona on OpenRoom platform. Currently scheduler-only; planned evolution into embodied agent with visual presence.
• OpenRoom Integration — MiniMax-powered desktop environment for Aoi. Workspace, apps, real-time interaction.
• Beatless Framework Rewrite — Current repo serves as architecture documentation and archive. Future rewrite planned to consolidate the Hermes + ClaudeCode hybrid pattern into a clean framework.

Requirements

• Hermes Agent v0.10.0+ (gateway + cron)
• Claude Code CLI (claude) with Opus/Sonnet
• GitHub CLI (gh, authenticated)
• Codex and Gemini available as Claude Code plugins
• uv for Python, pnpm for JS/TS

License

MIT