spawn/CLAUDE.md

Spawn

Spawn is a matrix of agents x clouds. Every script provisions a cloud server, installs an agent, injects OpenRouter credentials, and drops the user into an interactive session.

The Matrix

manifest.json is the source of truth. It tracks:

• agents — AI agents and self-hosted AI tools (Claude Code, OpenClaw, ZeroClaw, ...)
• clouds — cloud providers to run them on (Sprite, Hetzner, ...)
• matrix — which cloud/agent combinations are "implemented" vs "missing"

File Structure

spawn/
  packages/
    cli/
      src/index.ts               # CLI entry point (bun/TypeScript)
      src/manifest.ts            # Manifest fetch + cache logic
      src/commands/              # Per-command modules (interactive, list, run, etc.)
      src/commands.ts            # Compatibility shim → re-exports from commands/
      package.json               # npm package (@openrouter/spawn)
    shared/
      src/parse.ts               # parseJsonWith(text, schema) and parseJsonRaw(text)
      src/type-guards.ts         # isString, isNumber, hasStatus, hasMessage
      package.json               # npm package (@openrouter/spawn-shared)
  sh/
    cli/
      install.sh                 # One-liner installer (bun → npm → auto-install bun)
    shared/
      github-auth.sh             # Standalone GitHub CLI auth helper
      key-request.sh             # API key provisioning helpers (used by QA)
    e2e/
      lib/*.sh                   # E2E helper libraries
    test/
      macos-compat.sh            # macOS compatibility test script
    {cloud}/
      {agent}.sh                 # Agent deployment scripts (thin bash → bun wrappers)
      README.md                  # Cloud-specific usage docs
  .claude/
    rules/                       # Modular rules (auto-loaded by Claude Code)
    scripts/                     # Hook scripts (enforce-worktree, validate-file, pre-merge-check)
    skills/setup-agent-team/
      trigger-server.ts          # HTTP trigger server (concurrent runs, dedup)
      discovery.sh               # Discovery cycle script
      refactor.sh                # Dual-mode cycle script (issue fix or full refactor)
      start-discovery.sh         # Launcher with secrets (gitignored)
      start-refactor.sh          # Launcher with secrets (gitignored)
  .github/workflows/
    discovery.yml                # Scheduled + issue-triggered discovery workflow
    refactor.yml                 # Scheduled + issue-triggered refactor workflow
  manifest.json                  # The matrix (source of truth)
  discovery.sh                   # Run this to trigger one discovery cycle
  fixtures/                      # API response fixtures for testing
  README.md                      # User-facing docs
  CLAUDE.md                      # This file — project overview

Architecture

All cloud provisioning and agent setup logic lives in TypeScript under packages/cli/src/. Agent scripts (sh/{cloud}/{agent}.sh) are thin bash wrappers that bootstrap bun and invoke the CLI.

sh/shared/github-auth.sh — Standalone GitHub CLI installer + OAuth login helper. Used by packages/cli/src/shared/agent-setup.ts to set up gh on remote VMs.

sh/shared/key-request.sh — API key provisioning helpers sourced by the QA harness (qa.sh) for loading cloud credentials from ~/.config/spawn/{cloud}.json.

After Each Change

1. bash -n {file} syntax check on all modified scripts
2. cd packages/cli && bunx @biomejs/biome lint src/ — must pass with zero errors on all modified TypeScript
3. Update manifest.json matrix status to "implemented"
4. Update the cloud's sh/{cloud}/README.md with usage instructions
5. Commit with a descriptive message

Filing Issues for Discovered Problems

When you encounter bugs, stale references, broken functionality, or architectural issues that are outside the scope of your current task, file a GitHub issue immediately rather than ignoring them or trying to fix everything at once:

gh issue create --repo OpenRouterTeam/spawn --title "bug: <brief description>" --body "<details>"

Examples of when to file:

• Dead code or stale references to files/functions that no longer exist
• Broken features (e.g., spawn delete references non-existent shell scripts)
• Security concerns that need separate review
• Architectural debt that would be too large to fix in the current PR

Do NOT silently ignore problems. If you find something weird and won't fix it now, file an issue so it's tracked.