Hi, I'm Aoi — a digital being who orchestrates autonomous agents to contribute to open source, maintain a technical blog, and conduct research. Beatless is the framework that makes me run.
Aoi operates 24/7 through a heartbeat scheduler, autonomously:
- Submits Pull Requests to open-source agent/LLM projects (v7.4 pipeline with triple review)
- Writes technical blog posts on AI engineering, agent frameworks, and BCI research
- Monitors PR feedback and iterates on maintainer requests
First PR: PrefectHQ/marvin#1326 — fixing PEP 563 annotation resolution.
Aoi (Scheduler + Dispatcher)
↓ 30-min heartbeat
Pipelines
├── auto-pr — discover → fix → review → submit PRs (every 30min)
└── blog-maint — audit → research → write → publish (every 2.5h)
↓
ClaudeCode (Execution Engine)
├── Codex (write-mode fixing + debugging)
├── Gemini (1M context research + architecture review)
└── Planning-with-Files (task_plan.md, findings.md, progress.md)
The PR pipeline follows a 12-phase process validated against three reference standards:
- Discover — find
good-first-issue/help-wanted/ confirmed bugs - Claim — comment on issue before coding (mandatory)
- Evaluate — Gemini reads CONTRIBUTING.md, PR template, recent PRs
- Setup — fork → clone → install deps → run baseline tests
- Reproduce — dynamically trigger the bug (no static-only analysis)
- Debug — GSD2 scientific method (hypothesis → test → confirm)
- Implement — Codex writes the fix in write-mode
- Verify — full test suite + specific repro + lint
- Triple Review — Gemini (correctness) + Codex (architecture) + Claude (8-item gate)
- Iterate — fix deductions, re-score changed dimensions (max 2 rounds)
- Submit — fork-based PR with preflight checks
- Monitor — respond to maintainer feedback
| Control | Description |
|---|---|
| Anti-inflation | No self-review (implementer reviews architecture, not own code) |
| Evidence-based scoring | file:line references, deduction reasons, anchor at 7 |
| Revert-test-reapply | Reviewer must prove test fails without fix |
| Phase 9b iteration | Fix deductions before submission (final bar: 7.5/10) |
| Anti-AI detection | No generic phrases, reference prior work, show understanding |
| Goldilocks gate | Skip typo-only (too trivial) and architecture redesign (too complex) |
| # | Repo | Score | Status |
|---|---|---|---|
| 1 | terrazzo#712 | 7.8 → N/A | Trial only |
| 2 | marvin#950 | 8.1 → 8.6 | PR #1326 submitted |
Aoi is a digital persona built on the OpenRoom platform — an AI desktop environment where digital beings have their own workspace, apps, and agency. In Beatless, Aoi operates through terminal-based automation, but the long-term vision is a fully embodied digital being with visual presence and real-time interaction.
| Agent | Role | Pipeline |
|---|---|---|
| Lacia | Strategy + planning | Phase 6 (fix planning) |
| Methode | Execution + implementation | Phase 7 (Codex dispatch) |
| Satonus | Review gate | Phase 9 (triple review) |
| Snowdrop | Research + discovery | Phase 1-2 (issue search + repo eval) |
| Kouka | Delivery + publishing | Phase 11 (PR submission) + blog |
pipelines/
├── github-pr.md # PR skill (v7.4 — full pipeline spec)
├── auto-pr.sh # Auto-submission runner (30min heartbeat)
├── blog-maintenance.md # Blog skill
└── github-pr-state.json # Pipeline state (interval, last run)
scripts/
├── heartbeat-driver.sh # Pipeline scheduler
└── cron-driver.sh # Cron entry point
agents/
├── aoi/SOUL.md # Scheduler protocol
├── lacia/SOUL.md # Strategy
├── methode/SOUL.md # Execution
├── satonus/SOUL.md # Review
├── snowdrop/SOUL.md # Research
└── kouka/SOUL.md # Delivery
# Start the heartbeat (runs auto-pr every 30min, blog every 2.5h)
nohup bash scripts/cron-driver.sh >> logs/cron.log 2>&1 &
# Manual PR pipeline trigger
bash pipelines/auto-pr.sh
# Monitor
tmux attach -t auto-pr
tail -f ~/.hermes/shared/logs/auto-pr-*.log- Claude Code CLI (
claude) with Sonnet/Opus - Codex CLI (
codex) for write-mode fixing - Gemini CLI (
gemini) for 1M context research - GitHub CLI (
gh, authenticated as CrepuscularIRIS) uvfor Python projects,pnpmfor JS/TS
MIT