SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsruflo

ruflo

Deployment & operations

View repository

Updated 10-29 04:36

About this skill

Ruflo repository operator

When to use this skill

Use this skill for source-grounded, approval-gated operation of ruvnet/ruflo: establish what is installed without invoking it, inspect a pinned checkout, preview initialization, choose swarm or hive topology and consensus, register MCP without overwriting existing entries, inspect memory/state, and clean up only Ruflo-owned artifacts. The helper is offline-only and never runs Ruflo, Claude Code, npm, npx, servers, hooks, or network calls.

Route agent-company planning and artifacts to paperclip; route local tmux process/session lifecycle to agent-manager; route graph-based provider-CLI runs and verification to zeroshot; route ambiguity reduction and immutable spec loops to ooo; route recursive agent-harness optimization to harness2.

Instructions

  1. Run the offline gate first. From the repository root run python3 .agent-skills/ruflo/scripts/preflight.py --json; from an installed skill root run python3 "$SKILLS_ROOT/ruflo/scripts/preflight.py" --json (--help lists options). It only discovers tools and inspects an explicitly supplied checkout; it never invokes upstream executables.
  2. For a source audit, use --repo PATH --expect-commit a00b2855e5b8204aec61133d5fbc53060e1726b5; confirm the Git root, commit, and public markers before using any command from the reference.
  3. Use ruflo --help, ruflo --version, or equivalent package metadata as read-only discovery. Do not infer runtime readiness from package presence, and do not run npx ruflo or claude-flow init during preflight.
  4. Before init, show the exact target paths and generated components. Approval is mandatory because init can create .claude/ directories, .claude/settings.json hooks and permissions, .mcp.json, helpers, skills, commands, agents, runtime config, statusline data, .swarm/memory.db, and CLAUDE.md; do not overwrite existing config without explicit approval.
  5. For swarm or hive runs, record the requested topology and strategy first. Swarm supports hierarchical, mesh, ring, star, hybrid, hierarchical-mesh, and pheromone-adaptive; hive-mind also offers adaptive and consensus choices including byzantine, raft, gossip, crdt, and quorum. Approval is mandatory before spawning agents, starting servers, opening ports, or using credentials or remote providers.
  6. Treat MCP registration as a separate approval gate. Inspect existing .mcp.json, parent and global Claude registrations, and legacy claude-flow keys; never overwrite or create duplicate registrations without approval.
  7. Use memory peek for read-only store resolution. Treat memory init, store/delete/import/export, cleanup, and any command that writes .swarm/memory.db, .claude-flow/data, coordination, or session state as mutating and approval-gated.
  8. For cleanup, preview first. Remove only Ruflo artifact paths and .claude/helpers; preserve unrelated .claude content, memory databases, worktrees, and user settings. Approval is mandatory for --force, rollback, deletion, publishing, installs, downloads, telemetry or remote calls, and any destructive operation.
  9. When behavior is unclear, read the pinned source references and report observed versus unverified claims. Do not install dependencies, start a server, call a provider, or claim a live swarm result from static evidence.

Examples

python3 .agent-skills/ruflo/scripts/preflight.py --json
python3 .agent-skills/ruflo/scripts/preflight.py --json --repo /work/ruflo --expect-commit a00b2855e5b8204aec61133d5fbc53060e1726b5
# then preview init target paths and ask approval before any write

Best practices

  • Keep the ruflo and claude-flow names visible when selecting binaries, MCP keys, and documentation.
  • Prefer memory peek, status/list/help, and dry-run cleanup for evidence.
  • Treat package.json postinstall as a mutating install hook: it patches reachable agentdb trees; never run package installation implicitly.
  • Make backups and a file manifest before approved initialization or cleanup, and verify no unrelated settings changed afterward.
  • Label vector-memory, consensus, throughput, and learning claims as configured or estimated unless a separately approved runtime capture proves them.

References