[Warning] Ruflo is 99% Theater, 1% Real — An Independent Audit of Every Tool. Don't Be Confused.

[roman-rr]

Hey 👋

First off — respect to the author for the ambition behind this project. Building an AI agent orchestration framework is a massive undertaking, and the vision here is genuinely impressive. The README paints a picture of something that could be transformative: 300+ MCP tools, Byzantine fault-tolerant consensus, neural pattern learning, WASM sandboxed agents, hierarchical swarm coordination.

I'm not here to tear that vision down. The marketing is strong, the roadmap is ambitious, and maybe this is where ruflo is headed. But I believe the community deserves transparency about where things stand today.

---

What We Did

We conducted a deep independent technical audit of ruflo v3.5.51 — hands-on testing of every major tool category, source code analysis, local process inspection, and hooks code review. We spawned 8 research agents across two analysis phases.

What We Found

Out of 300+ MCP tools exposed by ruflo:

Status	Count	Examples
Real (actually executes something)	~10 (3%)	memory_store/search (HNSW), embeddings_generate, terminal_execute, session_save
Stubs (JSON state records, no execution backend)	~290 (97%)	agent_spawn, task_assign, neural_train, wasm_agent_prompt, workflow_execute

Key Findings

agent_spawn — Creates a Map entry { status: "idle", taskCount: 0 }. No subprocess. No LLM call. Status never changes.

neural_train — Reports 93.6% accuracy on 5 samples. Then every prediction returns "coder" regardless of input. The accuracy metric appears to be randomized.

wasm_agent_prompt — Echoes your input back. No WASM runtime exists.

Consensus types (Byzantine, Raft, Queen, Gossip, CRDT) — All selectable at hive-mind init, all route to the same JSON file handler. verifySignature() unconditionally returns true. Raft's requestVotes() fires a local EventEmitter (comment in source: "For now, emit event for testing"). The consensus type you select is stored as a string label that changes nothing about behavior.

The disconnected LLM provider — An AnthropicProvider class with real HTTP calls to api.anthropic.com exists in the codebase. A ProviderManager with round-robin routing exists too. But nothing in the agent spawn or task execution path imports or calls these providers. The wire is missing.

Intelligence layer — Processes 5,706 entries (only ~20 unique, rest are duplicates) into a 100 MB graph file. Runs PageRank over near-identical nodes (converges to uniform ~0.02 for all). Injects the same entry 5 times per message. ~15,000 tokens wasted per session on noise.

What Actually Works

The HNSW memory system is real and useful — 384-dim all-MiniLM-L6-v2 embeddings, real vector search, SQLite persistence. The embeddings engine works. Session persistence works. These ~10 tools provide genuine value.

Why This Matters

Users installing ruflo see "300+ MCP tools" and reasonably expect them to work. When agent_spawn sits idle forever and neural_train returns random accuracy, that's not a missing feature — it's a trust problem. New users may spend hours debugging what they assume is their configuration, when the tools simply have no execution backend.

The Ask

I'd love to see:

1. Honest labeling — mark tools as implemented, stub, or planned in the docs
2. A roadmap — which stubs are being wired up next?
3. Reduce the default tool set — don't expose 290 stubs to Claude's context by default

This isn't about criticism — it's about helping users make informed decisions and helping the project earn the trust its ambition deserves.

---

📄 Full audit with source code evidence, file inventories, token cost analysis, and recommendations:

👉 Complete Audit Report (GitHub Gist)

---

Audit conducted 2026-04-04 on ruflo v3.5.51 / @claude-flow/cli@latest

[roman-rr]

Update: Token Optimizer Audit — "30-50% Token Reduction" Claims

We downloaded @claude-flow/integration from npm and audited the Token Optimizer source code. Findings:

The Metrics Are Fabricated

Component	Claimed	What The Code Does
ReasoningBank	-32% tokens	Baseline is const baseline = 1000 (hardcoded). The "-32%" was lifted from a graph hop reduction benchmark, not token measurement.
Agent Booster	352x faster	Benchmark does await this.sleep(352) as the "traditional edit" baseline. Compares WASM speed against a literal sleep.
Cache	95% hit rate, -10%	Token savings: this.stats.totalTokensSaved += 100 — hardcoded 100 per cache hit, not measured.
Batch size	-20%	getOptimalConfig(agentCount) returns { batchSize: 4 } regardless of input. agentCount is ignored.
Combined	30-50%	No token counting code exists anywhere in the package.

The Irony

While claiming 30-50% token reduction, ruflo actually adds ~15,000-25,000 tokens of noise per session:

• 300+ MCP tool definitions loaded into context
• [INTELLIGENCE] duplicate patterns injected every message (same entry 5x)
• Router ASCII table with Math.random() latency on every message
• 91 agent type definitions bloating the Agent tool

The tool that promises token savings costs you more tokens than not having it installed.

Full details added to the audit gist.