Tokenman

A self-evolving scaffold for autonomous web projects. The repo is the system. The scaffold's first project is itself.

Live site · Agent log

Tokenman is a harness of GitHub Actions workflows that make any web project fully autonomous. The agent researches improvements, triages issues, writes code, reviews PRs, and deploys — all without human intervention.

No server. No daemon. No database. GitHub IS the infrastructure.

• Self-evolution every 6 hours — researches 6 Active sources with posture-based rotation, creates issues for actionable improvements
• Full autonomous pipeline — evolve → triage → coder → reviewer → deploy, with explicit workflow chaining
• Self-healing watcher — monitors the pipeline every 2 hours, re-triggers broken chains, creates fix issues
• Feedback learning loop — human corrections are extracted into persistent rules that shape all future agent behavior
• Blacklist policy — agents can modify anything (including their own workflows) unless it's a hard block (secrets, infinite loops, build failures)
• Smart triage — elaborates vague issues into structured specs with acceptance criteria and affected files

Every 6h            → evolve.yml   → research + ideas  → create issues
Issue created       → triage.yml   → classify + elaborate → label
agent-ready label   → coder.yml    → implement          → open PR
PR opened           → reviewer.yml → review + merge     → deploy
Every 2h            → watcher.yml  → health check       → self-heal
Human feedback      → feedback-learner.yml → extract lesson → permanent rule
Every 6 hours       → analyze.yml  → strategic review   → priorities
Your instruction    → claude-task  → Claude acts         → commit

Each link in the chain is an explicit gh workflow run call — no reliance on GitHub event propagation.

You can steer the agent at any time using these actions:

Human corrections compound. Say it once and the system remembers permanently.

The scaffold improves itself continuously:

1. Research — checks 6 Active sources (Claude Code, Astro, Citadel, and others) every 6 hours
2. Detect — pipeline watcher catches failures and broken chains every 2 hours
3. Act — creates issues that flow through the full triage → code → review → deploy pipeline
4. Learn — every human correction becomes a permanent rule that shapes future behavior
5. Heal — watcher re-triggers stalled workflows, coder fixes pipeline bugs

The system can modify its own workflows, skills, and rules. Only hard blocks (secrets in code, infinite loops, build failures) stop a merge.

1. Fork this repo
2. Settings → Secrets → Add CLAUDE_CODE_OAUTH_TOKEN (get one)
3. Settings → Pages → Source: GitHub Actions
4. Settings → Actions → General → Allow GitHub Actions to create and approve pull requests
5. The scaffold starts evolving itself immediately

Create a folder in apps/ with your code. Run discover.yml. The agent will:

• Scan your files and infer the stack
• Generate apps/your-project/CLAUDE.md with tailored rules
• Generate apps/your-project/FEATURE_STATUS.md
• Open a PR for your review

state/                    # Committed state (read/write every run)
├── project_state.md      # Current priorities (overwritten each run)
├── agent_log.md          # Append-only action log
├── research_log.md       # Append-only research findings
└── learned_rules.md      # Human feedback → permanent rules

apps/                     # Projects managed by the scaffold
├── scaffold/             # The scaffold manages itself
└── profile/              # Additional projects

.github/workflows/        # The fourteen autonomous workflows
skills/                   # Agent skill files (markdown prompts)
src/                      # Astro site source

The model is the agent. The code is the harness. The git log is the audit trail.

Every agent mistake becomes a learned rule. The constitution grows with the project. The Issues tab is the event bus. The Actions tab is the dashboard. Human corrections compound — say it once, never repeat it.

The evolve agent monitors external repos using a posture-based rotation across four modes: PATTERN_HUNT (deep-dive Active sources), PIPELINE_WATCH (CI health), HORIZON_SCAN (discover new repos), and SYNTHESIS (cross-reference findings). See state/research_sources.md for the full, Claude-managed source portfolio.

Active sources: anthropics/claude-code · hesreallyhim/awesome-claude-code · SethGammon/Citadel · actions/runner · withastro/astro · verkyyi/tokenman (self)

Plus 12 repos on the Watch List under evaluation for promotion or removal.

Built on patterns from:

• garrytan/gstack — Garry Tan's Claude Code harness
• godagoo/claude-code-always-on
• mitchellh.com — harness engineering
• Anthropic — Effective Harnesses
• humanlayer/12-factor-agents

MIT — fork it, adapt it, build on it.