The world model gets better every time someone chips in. You don't need to write code. You can contribute by running the harvester against public repos and sharing the results.

The easiest way: run seed-harvest against repos you care about.

# Clone this repo
git clone https://github.com/vystartasv/hermes-harness
cd hermes-harness

# Harvest public repos (or point it at your own)
bun run src/seed.ts harvest

# Export the model
bun run src/seed.ts export

# Share the output
# → ~/.hermes-harness/seed-world-model.publish.json

Open a PR with your exported model. The maintainer will merge it into the seed.

Repos with:

• 500+ commits (enough history for patterns)
• Conventional commits (fix:, feat:, revert:) — clearer signal
• Active issue/PR lifecycle — more reverts = more patterns
• Permissive license (MIT, Apache, BSD) — for redistribution

Repos to avoid:

• Personal dotfiles, config repos (no lifecycle data)
• Monorepos with auto-merge bots (noise, not signal)
• Private/proprietary code (we can't redistribute)

PRs are welcome for:

• New miners — Slack channel miner, calendar miner, browser history miner
• Better embeddings — improve the 128-dim word-count vectorizer
• Query UX — better CLI output, web interface
• Integrations — GitHub Actions, VS Code extension, MCP server

src/
├── config.ts         # Centralized paths (override via HARNESS_DATA_DIR)
├── harness.ts        # Shared utilities (LLM calls, embeddings)
├── cli.ts            # CLI entry point
├── seed.ts           # Public repo harvester (world creator)
├── world-model.ts    # Model training + prediction
├── repo-query.ts     # Cross-repo search
├── git-miner.ts      # Private git history miner
├── ticket-miner.ts   # ADO ticket lifecycle miner
├── post-session.ts   # Hermes session miner
├── wwa-issues.ts     # GitHub issue scanner
├── pr-creator.ts     # Auto-PR from fixes
└── self-improve.ts   # Full automation loop

MIT