Skip to content

v3.7.0-alpha.22 — Discoverable, Verifiable, Networked

Pre-release
Pre-release

Choose a tag to compare

View all tags
@ruvnet ruvnet released this 11 May 04:31
v3.7.0-alpha.22
7415f83

Ruflo v3.7.0-alpha.22 — "Discoverable, Verifiable, Networked"

Three big wins in this release: every MCP tool now tells Claude when to use it, the WireGuard mesh layer ships for federation peers who want trust changes to follow at the packet layer, and all 102 documented fixes are cryptographically verified end-to-end. Plus a handful of audit responses + bug fixes.

Install

npx ruflo@latest                  # 3.7.0-alpha.22
npx claude-flow@latest            # 3.7.0-alpha.22
npx @claude-flow/cli@latest       # 3.7.0-alpha.22
npx @claude-flow/plugin-agent-federation@latest   # 1.0.0-alpha.15

The headline change: every Ruflo tool now answers "when should Claude actually use this?"

Before this release, 279 of 285 MCP tools shipped descriptions that just said what they do (e.g. "Spawn an agent", "Store data in memory"). When Claude sees that next to native Task or Write, it has no signal to pick Ruflo over native — and it usually doesn't.

After: 285 of 285 tools have an explicit "Use when native [X] is wrong because [concrete value-add]; for [inverse case], native is fine" clause. Example:

agent_spawn: Spawn a Ruflo-tracked agent with cost attribution + memory persistence + swarm coordination. Use when native Task tool is wrong because you need cost tracking per agent / cross-session learning via patterns / coordination with other agents in a swarm topology. For one-shot subtasks with no learning loop, native Task is fine.

Same pattern across all 32 tool categories. ADR-112 documents the rule + the CI guard that keeps it enforced (any new tool shipping without guidance fails the build).

User-visible effect: Claude will actually call Ruflo tools where they add value — cost-tracked agents, persistent memory, swarm coordination — instead of falling through to native every time.

ADR-111 — WireGuard mesh for federation peers (opt-in)

@claude-flow/plugin-agent-federation@1.0.0-alpha.15 ships an opt-in WireGuard layer that hooks into the existing federation trust ladder + circuit breaker. When the breaker SUSPENDs or EVICTs a peer, the same transition propagates to WG AllowedIPs — a peer the trust layer just blocked can't reach the rest of the mesh at L3 either.

Phases shipped: manifest extension + key generation, WgMeshService, breaker integration, firewall projection (nftables/pf), witness attestation chain, and three operator MCP tools (federation_wg_status, _attest, _keyrotate).

Phase 7 (the actual wg-quick up) stays operator-mediated — staged configs land in /tmp/adr-111-stage/, operator reviews + activates manually. See docs/federation/phase7-mesh-bringup.md for the bringup procedure.

Audit response (#1896)

External audit by AlphaSignal AI named six gaps in v3.6.30. Re-measured on current main with the v3.7.0-alpha work landed:

Gap Article claim (v3.6.30) Current main Status
G1 — agent_spawn no LLM In-memory Map, no subprocess agent_execute calls Anthropic API directly ✅ remediated
G3 — workflow_execute broken Always returns "not found" Real step executor + variable interpolation + pause/resume ✅ remediated
G4 — WASM agent echoes No LLM, returns "echo: X" Detects stub + routes through Anthropic when ANTHROPIC_API_KEY set ✅ remediated
G6 — auto-memory bloat 5,706 entries / 20 unique per message Single routing call; 8 backend entries observed; no Trigram/Jaccard code ✅ refuted
#1748 — tool discoverability 237/300 tools lack guidance 0/285 (this release fixes it) ✅ resolved
Benchmark integrity simulate_benchmarks.py + "84.8% SWE" Both removed from main ✅ cleaned

See #1896 for the full file-path-receipted writeup + ADR-095 status update.

Smaller fixes

  • #1892 — UI version banner no longer hardcoded. The statusline now reads @claude-flow/cli's package.json at runtime so the UI matches ruflo doctor output.
  • Security — peer-content injection vector found during ADR-111 security audit and fixed (readSafePeerWgFields validator). A compromised federation peer with a valid Ed25519 key could otherwise inject extra [Peer] blocks via newline-laden wgEndpoint. Closed before any user activated WG mesh.

Verification — 102/102 fixes cryptographically attested

Every documented fix in the witness manifest is now Ed25519-signed and verifiable end-to-end via node plugins/ruflo-core/scripts/witness/verify.mjs. New entries this release: ADR-111, ADR-112, #1892.

What this release is NOT

  • A swarm-runtime rewrite. ADR-095 G2 (multi-process consensus) is still single-process for hive-mind; the federation plugin's transport is the only multi-host wire that's hardened.
  • A WASM-runtime rebuild. G4 is "echo + LLM fallback", not "full WASM sandbox with deterministic replay" — that's a separate ADR.
  • A benchmark publication. The fake simulate_benchmarks.py is gone but Ruflo still doesn't appear on the official SWE-bench leaderboard — by design until a real run is published.

Upgrade

If you're on any v3.7.0-alpha, npm i ruflo@latest is enough. If you're on v3.6.x:

# Snapshot first (sane practice):
ruflo memory list > backup-memory.txt
# Then:
npx ruflo@latest init --upgrade

Federation users wanting the opt-in WireGuard layer: npx @claude-flow/plugin-agent-federation@latest, then follow docs/federation/README.md.

Links

Assets 2
Loading