dym-ruflo-route
When to use
- Someone suggests installing ruflo/claude-flow into a project that uses overstack.
- You need to orchestrate multiple agents and aren't sure which layer to use.
Steps
- Determine which layer the task belongs to — these 4 layers do NOT replace each other:
| layer |
task |
tool |
| within ONE Claude Code session |
fan-out subagent, shared memory, self-learning hook |
ruflo (or bare Task tool) |
| multiple CLI SESSIONS on one machine |
one pane per agent, know which pane is stuck/done |
herdr |
| assign work to any CLI via Orca, know when it's done |
sentinel shell + poll terminal |
orca-dispatch.py (ours) |
| just need to run in parallel then gather |
nothing more needed |
Task tool / claude -p |
- If you intend to use ruflo, run
/dym-ruflo init FIRST to know what it overwrites.
- Verdict for the overstack project: DO NOT TAKE it wholesale.
| reason |
measured evidence |
Takes up .claude/ |
init writes 241 files, 13 hooks / 10 types, replaces statusLine · permissions · model · env — exactly the things overstack already owns |
| Not an independent runtime |
"Never use MCP tools alone for execution — Task tool agents do the actual work" (its own CLAUDE.md) — drop Claude Code and nothing runs anymore |
| Installation cost |
1.5 GB prefix + writes .claude-flow/ and Library/ into $HOME |
| Surface too broad to maintain |
2280 source files, 25 workspace packages, 17 command groups, hooks.ts 211 KB · doctor.ts 113 KB |
Worth learning from, not worth swallowing (DISSOLVE mode, copy the Idea not the code):
autopilot — the "don't stop until all tasks are done" loop, plug into the Stop hook. We already have a
Stop hook; this is a cheap pattern to add.hooks route / hooks explain — route task→agent using a learned pattern, and be able to
explain the decision. The explain part is the good bit.swarm compress-message — compress messages between agents to a token budget.
Rules
- Don't run
ruflo init in a worktree that already has overstack. If you want to try it, set up a separate empty
project, then delete it.
- Don't read "60+ agent, consensus, neural, vector memory" as verified capability —
this measurement round only proves install · init · doctor · status · help. The swarm/neural/
embeddings/ruvector part not yet verified.
- Same product, two names: npm
ruflo = claude-flow, errors still print claude-flow.
