How to correctly leverage Ruflo for successive commands?

[darrendavid]

I successfully spun up Ruflo with a swarm using claude-flow and much work was performed, however, subsequent commands in the same Claude Code window don't seem to be engaging a swarm or multiple agents - what are the best practices for successive commands? Do I need to quit out of the Claude code terminal and spin up a new swarm from the command line, or is there's something I should be adding in my prompt to make the most of Ruflo? It does seem like the session is aware of Ruflo, given the status bar:

 ▊ RuFlo V3.5 ● user  │  Opus 4.6 (1M context)  │  ⏱ 26m34s  │  ● 13% ctx  │  $1.44     
  ─────────────────────────────────────────────────────
  🏗️  DDD Domains    [○○○○○]  0/5    ⚡ target: 150x-12500x
  🤖 Swarm  ○ [ 0/15]  👥 0    🪝 16/16    🔴 CVE 0/0    💾 3MB    🧠   1%
  🔧 Architecture    ADRs ●0/0  │  DDD ●  0%  │  Security ●PENDING
  📊 AgentDB    Vectors ●23  │  Size 16KB  │  Tests ●0 (~0 cases)  │  MCP ●1/1

[roman-rr]

The reason subsequent commands don't engage the swarm is that the swarm was never actually running in the first place.

Here's what happens:

1. Your first command with swarm init writes a config JSON and 'spawns' agents → these are Map entries with {status: 'idle'}, not real processes
2. The 'much work performed' was likely Claude Code's own native execution, not the ruflo swarm
3. Subsequent commands in the same window don't see a swarm because the MCP swarm_status reports agentCount: 0 — it doesn't track its own spawned agents

We tested this extensively in our audit: spawned 5 agents via MCP, checked status after 30 minutes — all remained idle, taskCount: 0, lastResult: null.

What to do instead:
For successive multi-agent tasks, use Claude Code's native Agent tool with run_in_background: true. Each agent is a real subprocess with its own context. You can spawn multiple in one message and they run in parallel.

Ruflo's memory tools (memory_store/search) do work and can share context between those native agents — that's the one genuinely useful integration.

Full audit: Independent Audit