ADR-154 — @metaharness/kernel per-platform NAPI-RS binaries: WASM-only at runtime, native fast-path unreached

Status: Accepted (WASM-only path is functional; native fast-path tracked upstream) Date: 2026-06-17 Related: ADR-150 (MetaHarness integration surfaces), ADR-153 (Darwin Mode integration), ADR-124 (optional native dependencies), ADR-148 (cost-optimal router lifecycle), ADR-152 (genome similarity) Upstream: ruvnet/agent-harness-generator/packages/kernel-js Affects: [email protected]+, @claude-flow/[email protected]+, [email protected]+ Affected packages: @metaharness/kernel@~0.1.0 (already in optionalDependencies per ADR-150)

Context

ADR-150 integrated @metaharness/kernel@~0.1.0 as an optional peer of @claude-flow/cli (and transitively ruflo). The kernel exposes loadKernel(), kernelDiagnostics(), ToolDispatcher (claims-checked), SelfEvolvingRouter, TrajectoryStore, and rankWithDecay to ruflo's intelligence layer.

The upstream design intent (per packages/kernel-js/README.md) is:

At load time, loadKernel() prefers the per-platform native package; falls back to wasm. The native packages (@metaharness/kernel-darwin-arm64, -linux-x64-gnu, etc.) are declared as optionalDependencies — npm installs only the one for your platform.

@metaharness/[email protected]'s package.json declares five per-platform NAPI-RS native binaries as optionalDependencies:

{
  "optionalDependencies": {
    "@metaharness/kernel-darwin-arm64": "0.1.0",
    "@metaharness/kernel-darwin-x64":   "0.1.0",
    "@metaharness/kernel-linux-x64-gnu":   "0.1.0",
    "@metaharness/kernel-linux-arm64-gnu": "0.1.0",
    "@metaharness/kernel-win32-x64-msvc":  "0.1.0"
  }
}

Current state (verified 2026-06-17)

None of the five per-platform binaries are published to npm:

@metaharness/kernel-darwin-arm64         NOT_PUBLISHED
@metaharness/kernel-darwin-x64           NOT_PUBLISHED
@metaharness/kernel-linux-x64-gnu        NOT_PUBLISHED
@metaharness/kernel-linux-arm64-gnu      NOT_PUBLISHED
@metaharness/kernel-win32-x64-msvc       NOT_PUBLISHED

Runtime evidence — fresh install of [email protected] on a darwin-arm64 host (Mac mini M4 Pro, the platform the upstream-intended @metaharness/kernel-darwin-arm64 should serve):

$ npm install [email protected]
added 1163 packages

$ find node_modules/@metaharness/kernel-* -maxdepth 0 2>/dev/null
(no matches — none of the platform packages were resolved during install)

$ node -e "(await import('@metaharness/kernel')).loadKernel().then(k => console.log(k.backend))"
wasm

loadKernel() falls back to WASM cleanly. Functionality is intact; performance is the WASM tier, not the native tier the README advertises.

Decision

Accept the WASM-only runtime path as the supported configuration for ruflo 3.12.3+. Do not block on the upstream native-binary publish gap. Specifically:

1. No code change in ruflo. @metaharness/kernel stays in optionalDependencies as it is today. loadKernel() already does the right thing — prefers native, falls back to WASM. The fallback is reached because the optional native packages don't resolve, which is the documented graceful path.
2. Treat WASM as the steady-state baseline for benchmarks, smoke tests, and CI gates. Any "native vs WASM" performance comparison in ruflo's perf docs must mark the native column as "unverified — binaries not published" rather than copying upstream's claimed numbers.
3. Add a smoke gate in plugins/ruflo-metaharness/scripts/smoke.sh that asserts loadKernel().backend === 'wasm' is reachable. This is the inverse of a regression check: if the upstream ever ships natives, the smoke alerts so we can re-baseline.
4. Per ADR-150 architectural constraint #3 (graceful degradation) the kernel-absent case is already handled — ruflo continues to work if @metaharness/kernel itself is uninstalled. The platform-binary absence is a softer degradation: kernel API surface is fully available, only the hot-path acceleration is missing.

Why not block on upstream

Three reasons:

1. WASM is functionally complete. loadKernel().backend === 'wasm' exposes the same API as the native backend. The router's SelfEvolvingRouter parallel-logging gate from ADR-150 Phase 2 works against the WASM kernel; ADR-152 §3.1 genome similarity is pure-TS and doesn't touch the kernel at all. No ruflo feature is blocked by the missing natives.
2. The performance ceiling isn't binding yet. ruflo's measured router path is dominated by the LLM call latency (seconds), not kernel dispatch (microseconds). Even if the native path were 10× faster than WASM, the user-visible improvement on a real query would be sub-percent. We'd capture that improvement later, when we have the natives, without architectural changes.
3. The fix is upstream-only. Publishing five NAPI-RS binaries requires release-engineering on ruvnet/agent-harness-generator (CI matrix → upload to npm). Ruflo can't ship the natives ourselves without forking — which would violate the "MetaHarness is first-party" framing from ADR-150.

Consequences

Positive

• No regression risk. This ADR is descriptive — it changes the documentation, not the code. Today's behavior is the future behavior.
• Honest perf surface. The smoke gate prevents accidentally claiming native speedups in benchmarks.
• Self-correcting. When the upstream publishes the natives, loadKernel() will silently pick them up on the next install. The smoke gate fires, we audit, and decide whether to re-baseline ruflo's perf docs.
• Aligns with ADR-124 — same pattern as better-sqlite3, hnswlib-node, sharp et al: declare optional natives, ship a WASM/JS fallback, never crash when the native is absent.

Negative

• The README claim "wasm primary, NAPI-RS native fallback" is currently inverted in practice — wasm IS the fallback, but it's also the only thing that exists. Cross-link this ADR from any user-facing doc that mentions kernel performance.
• Cold-start cost. WASM init is reportedly slower than NAPI-RS on first call. We accept this cost (and the user accepts it via npx cold cache anyway).
• One more upstream-coupling risk to monitor. If @metaharness/[email protected] introduces a new API that's native-only and silently no-ops in WASM, the smoke gate catches reachability but not behavioral parity. ADR-150 Phase 3 §3.4 (Capability Graph) is the right place to enforce per-API parity.

Neutral

• Tracking. The upstream gap is recorded in this ADR + the linked issue. No internal sprint commitment to chase the publish.

Implementation notes

Smoke gate (one-time addition)

Add to plugins/ruflo-metaharness/scripts/smoke.sh:

step "18a. @metaharness/kernel loadKernel returns wasm backend (ADR-153)"
miss=""
# Only run if kernel is installed (graceful per ADR-150 constraint #3)
if node -e "import('@metaharness/kernel').then(() => process.exit(0)).catch(() => process.exit(1))" 2>/dev/null; then
  BACKEND=$(node --input-type=module -e "
    const { loadKernel } = await import('@metaharness/kernel');
    const k = await loadKernel();
    process.stdout.write(k.backend || 'unknown');
  " 2>/dev/null)
  case "$BACKEND" in
    wasm)
      : # expected — see ADR-153
      ;;
    native|napi|napi-rs)
      # Upstream published the natives — re-baseline ruflo's perf docs
      miss="$miss kernel-now-native-rebaseline-perf-docs"
      ;;
    *)
      miss="$miss kernel-unknown-backend:$BACKEND"
      ;;
  esac
fi
[[ -z "$miss" ]] && ok || bad "$miss"

Doctor surface

npx ruflo doctor --component metaharness should already report the kernel backend; verify that the JSON includes a kernelBackend: 'wasm' | 'native' | 'absent' field. If not, add it as part of the next doctor revision.

Benchmark docstring

Wherever ruflo's perf docs cite kernel dispatch latency, add a one-liner:

Measured against @metaharness/kernel's WASM path. NAPI-RS native binaries (@metaharness/kernel-{darwin-arm64,linux-x64-gnu,…}) are declared but not yet published upstream — see ADR-153.

Cross-references

• ADR-150 — MetaHarness integration surfaces (where @metaharness/kernel was first added as optionalDependency)
• ADR-124 — Optional native dependencies pattern (the precedent this follows)
• ADR-148 — Cost-optimal router (consumes @metaharness/router, parallel-logs via the kernel's SelfEvolvingRouter)
• Upstream tracking: TBD — file an issue on ruvnet/agent-harness-generator for the per-platform publish (PR welcome, since the build matrix lives upstream)

Open questions

1. When to re-baseline perf docs. Trigger: smoke gate fires kernel-now-native-rebaseline-perf-docs. Action: re-run the relevant benchmarks, compare WASM vs NAPI-RS, update the docs with side-by-side numbers, drop the ADR's "currently inverted" caveat.
2. Whether ruflo should publish its own NAPI-RS binaries as a fallback fallback. No — out of scope. ruflo is a downstream consumer of @metaharness/kernel, not a native-binary publisher. If upstream stays unpublished for >1 quarter, revisit this decision.
3. Whether the kernel API surface differs at all between WASM and NAPI-RS. Spot-checked the WASM path on darwin-arm64; the exports match the TS declarations. ADR-150 Phase 3 §3.4 (Capability Graph) is the systematic answer for behavioral parity at API level — that's where to add a row per kernel method.