/v3/docs/adr/ADR-108-native-quic-binding.md

ADR-108 — Native QUIC binding plan

• Status: Proposed — gated on upstream agentic-flow Phase-1
• Date: 2026-05-09
• Authors: claude (drafted with rUv)
• Related: ADR-104, ADR-105, upstream ruvnet/agentic-flow#15-21

Context

Federation transport today (alpha.9 + agentic-flow@2.0.12-fix.2) uses WebSocket. The loader pattern (ADR-104) lets us auto-upgrade to QUIC when a native binding is available — set AGENTIC_FLOW_QUIC_NATIVE=1 and the same code path picks up the upgrade. But the native binding doesn't ship.

What exists today in upstream ruvnet/agentic-flow:

• crates/agentic-flow-quic/ — Rust crate using quinn (the canonical Rust QUIC impl). Full client + server features behind compile-time client / server features.
• crates/agentic-flow-quic/src/wasm.rs — WASM bindings, but explicitly a stub ("WASM build is a stub since browsers don't support UDP/QUIC directly. For production QUIC, use native Node.js builds.")
• crates/agentic-flow-quic/wasm-pack-build.sh — build script for the WASM stub bundle (already runs)
• 7 open Phase-1 issues: foundation impl (#15), WASM build deps (#16), TS wrapper (#17), integration tests (#18), benchmark (#19), wasm-pack pipeline (#20), validation (#21)

What's missing for a real native build:

1. No N-API binding crate. We need a crates/agentic-flow-quic-node/ or similar that wraps the existing client/server in napi-rs for Node.js native modules
2. No per-platform binary distribution. Pattern that works for @ruvector/* packages: separate @agentic-flow/quic-native-darwin-arm64, @agentic-flow/quic-native-linux-x64-gnu, etc., resolved at install via optionalDependencies
3. No platform detection in loadQuicTransport. Today the env-var probe is a placeholder; it needs to detect the platform-specific package + try to load it
4. No CI matrix for cross-compiling Rust → multi-platform binaries. GitHub Actions Linux/macOS/Windows runners + cross compilation for ARM

Decision

Defer the native QUIC binding to a future iteration, gated on:

1. Upstream agentic-flow#15-21 progress. The Phase-1 milestone has the work this ADR depends on. We don't fork that — it's their roadmap.
2. A concrete federation use case where the WS ceiling matters. Today's federation traffic is human/agent-rate (≤100 RPS per peer). WebSocket handles that fine. QUIC's wins (0-RTT, multiplexed streams, mobility) only matter under load OR for mobile peers roaming networks.

When the native binding ships upstream:

1. Update agentic-flow dep range in @claude-flow/plugin-agent-federation/package.json to the version that includes native (likely ^2.1 after their Phase-1 closes)
2. Document the AGENTIC_FLOW_QUIC_NATIVE=1 env var in the federation operator runbook
3. Update doctor surface so --component federation reports selectedBackend=quic when native is loaded
4. Add a federation-side smoke that asserts getTransportCapabilities().selectedBackend === 'quic' when env is set + binding installed
5. Add to the 12h verification routine: probe both backends + confirm the same federation send round-trip works on each

What this ADR is NOT proposing

• We won't build the Rust→N-API binding ourselves. That's upstream work; doing it in our repo creates a fork burden we don't want.
• We won't switch the default to QUIC even after native ships. WebSocket is cheaper to debug, has fewer corners, works everywhere. QUIC stays opt-in via env var.
• We won't drop WebSocket fallback even after QUIC is solid. The fallback is critical for: browsers, environments where UDP egress is firewall-blocked, debugging cycles where you want a wire-shark-friendly transport.

Implementation plan (when triggered)

Phase 1 — Detection

In agentic-flow/src/transport/quic-loader.ts:

async function isRealQuicAvailable(): Promise<boolean> {
  if (process.env.AGENTIC_FLOW_QUIC_NATIVE !== '1') return false;
  try {
    // Try to load the platform-specific native module
    const platform = `${process.platform}-${process.arch}`;
    const nativeName = `@agentic-flow/quic-native-${platform}`;
    await import(nativeName);
    return true;
  } catch {
    return false;
  }
}

Phase 2 — Federation plugin upgrade

// In plugin.ts initialize():
const transport = await loadQuicTransport({
  serverName: nodeId,
  enable0Rtt: true,         // pays off only with native QUIC
  maxConcurrentStreams: 100, // ditto
  // ...
});

// At the doctor surface — capability probe:
const caps = await getTransportCapabilities();
context.logger.info(`Federation transport: ${caps.selectedBackend}` +
  (caps.selectedBackend === 'quic' ? ' (0-RTT, multiplexed streams)' : ' (fallback)'));

Phase 3 — Verification routine update

Add Check 9 to the 12h routine:

# With native QUIC available:
AGENTIC_FLOW_QUIC_NATIVE=1 node -e '
import("agentic-flow/transport/loader").then(async (m) => {
  const caps = await m.getTransportCapabilities();
  if (caps.selectedBackend !== "quic") {
    console.error("FAIL: native QUIC env set but loader fell back to WS");
    process.exit(1);
  }
  console.log("ok: native QUIC selected");
});
'

Anti-goals

• No QUIC-or-nothing. The fallback path stays first-class.
• No federation-side QUIC implementation. All transport code lives in agentic-flow. The plugin only consumes the loader.
• No protocol-version pinning. Whatever QUIC version quinn ships, we use. We won't constrain to a specific draft.

Implementation status

Step	Status
Loader-pattern transport (forward-compat)	Implemented (alpha.9 + ADR-104)
AGENTIC_FLOW_QUIC_NATIVE env-var probe	Implemented (placeholder returns false today)
Native binding (Rust→N-API)	Deferred — gated on upstream agentic-flow#15-21
Per-platform binary distribution	Deferred
Federation-side adoption (env var + doctor surface)	Deferred
Verification check	Deferred

Decision review trigger

Re-open when:

• Upstream agentic-flow ships a native binding (any of #15-21 closing with merged PRs)
• A federation user reports needing >100 RPS to a single peer (WS head-of-line blocking becomes a real problem)
• A federation user reports needing peer mobility (mobile device roaming networks — QUIC's connection ID survives IP changes)
• We add browser peers (rules QUIC out — browsers can't do raw UDP — confirms the WS-default decision)