Autonomous agent orchestration for open-source contribution, technical writing, and ML research.
Beatless is a hybrid control plane: a lightweight scheduler watches for useful work, then routes deep execution to Codex, Gemini, GitHub, Zotero, and local experiment workspaces.
| Area | Purpose |
|---|---|
| GitHub response | Watch open PRs and surface maintainer comments that need action. |
| GitHub PR pipeline | Discover issues, evaluate repositories, implement fixes, review, and prepare PRs. |
| Research automation | Resume or halt experiment workspaces based on recorded state. |
| Paper workflow | Harvest papers, deduplicate against Zotero, and sync metadata into notes. |
| Dashboard | Show agents, pipelines, experiment status, GPU state, and recent activity. |
| CLI bridges | Route execution through local Codex and Gemini CLIs. |
Beatless separates scheduling from execution.
- Hermes handles cron, wake gates, lightweight status checks, and routing.
- Codex is the primary long-running executor for reasoning, code edits, command execution, feasibility checks, and review.
- Gemini focuses on literature grounding, large-context review, and critique.
- Claude Code remains an optional legacy/manual fallback, not the default wake-gate executor.
- Zotero and Obsidian hold research inputs and reading outputs.
- The dashboard reads JSON state from local collectors and renders it through a decoupled frontend.
See docs/CODEX_PRIMARY_ARCHITECTURE.md for the current executor split.
| Module | Description |
|---|---|
commands/exp |
Slash commands for experiment status, init, discovery, run, and review. |
commands/agents |
Legacy bridge definitions for Codex CLI and Gemini CLI. |
hermes-scripts |
Wake-gate scripts for GitHub, Zotero, research, blog, and preflight checks. |
dashboard |
FastAPI backend, SSE stream, and Vite frontend. |
pipelines |
Pipeline behavior specs and operating rules. |
docs, design, plan |
Architecture notes, migration status, and design records. |
Create local configuration:
cp .env.example .env.localFill only the variables you need. Keep real keys in .env.local or your private runtime environment. Do not commit secrets.
Run the local preflight:
python3 hermes-scripts/preflight.pyRun safe dry-runs:
python3 hermes-scripts/auto-research.py --dry-run
python3 hermes-scripts/github-response.py --dry-run
python3 hermes-scripts/github-pr.py --dry-run --issue-limit 1 --approved-limit 1 --per-query-limit 1 --skip-closed-pr-history
python3 hermes-scripts/paper-harvest.py --dry-run --max-new 1Start the local dashboard:
cd dashboard
./start.shDefault endpoints:
- UI:
http://127.0.0.1:3720 - API:
http://127.0.0.1:3721/api/status - SSE:
http://127.0.0.1:3721/api/events
The dashboard is intentionally decoupled:
- backend collectors produce JSON only;
- the frontend consumes the
/api/*contract; - SSE pushes full state every 10 seconds;
- the default host is local-only.
| Command | Role |
|---|---|
/exp-status |
Check workspace readiness, runtime state, and integration availability. |
/exp-init |
Initialize planning files, branch state, and baseline expectations. |
/exp-discover |
Generate research hypotheses unless the workspace is already halted. |
/exp-run |
Execute or resume an experiment loop with halt/rollback guards. |
/exp-review |
Review the latest round and choose continue, pivot, rollback, or halt. |
Smoke workspaces halt after one verified run. Real experiment workspaces should provide a substantive program.md or Task.md.
This repository is designed to keep machine-specific state out of Git.
Ignored local-only files include:
.env,.env.local, and.env.*.local.mcp.json- local dependency folders and Python caches
- local GSD clones or scratch links
- local runtime archives
Use .env.example as the public template and keep provider keys, Zotero IDs, GitHub tokens, and local paths in private configuration.
- Python with
uv - Node.js and npm
- GitHub CLI (
gh) - Codex CLI (
codex) - Gemini CLI (
gemini) - Optional: Claude Code CLI (
claude) for manual legacy fallback, Hermes Agent, Zotero API access, NVIDIA tooling for GPU experiments
MIT
