Measured the rerank hot path under cargo test --release:
| variant | latency (µs/q) | throughput |
|---|---|---|
| NoisyScore (no rerank, baseline) | ~0.15 | ~7.0M QPS |
| GnnDiffusion (+10.4pp recall) | ~400 | ~2.5K QPS |
The honest correction: the "~millions of QPS" associated with #479 is the
Turned the research win into a regression guard. A deterministic integration test reproduces the exact regime (N=5000, D=128, noise σ=0.40, seed=42) through the public reranker API and asserts GnnDiffusion beats the no-rerank baseline.
recall@10: noisy=0.280 gnn=0.384 delta=+0.104 (matches #479 exactly)
test result: ok. 1 passed
What: A 1-hop GNN score-diffusion reranker over approximate-ANN candidate sets that recovers recall lost to quantization. Measured in the source research (#479): recall@10 28.0% → 38.4% (+10.4pp), N=5K D=128, still ~millions of QPS.
Step 1 (implement/integrate):
ruvector-gnn-rerank into the ruvector workspace.feat/productionize-gnn-rerank.
npx ruvector its own "agent harness" — without adding a dependency"Could we improve the
npx ruvectorcommand-line tool by using the npm metaharness SDK?"
metaharness is a brand-new npm package (by the same author as ruvector) that helps you generate a custom AI-agent command-line tool from a repo. The idea sounded appealing: agent tooling, smart model routing, signed releases, memory. Should ruvector adopt it?
metaharness is a generator, not a library you plug in. It scaffolds a new tool; it doesn't upgrade an existing, mature one like ruvector.
Plain-language guide to what we built across 28 iterations of /loop work.
Companion to ADR-150 · Tracking issue #2399 · Upstream bug #9
MetaHarness mints a custom AI agent harness for any repo, on any host. You run one command, pick a host (Claude Code, Codex, OpenCode, …), and get a ready-to-use, npm-publishable harness with that host's native config files.
This release makes the multi-host promise actually true — and was hardened by installing and running each real host against generated harnesses.
Prepared: 2026-06-16
Companion artifacts: ADR-150 (decision) · tracking issue (phases) · this gist (research)
Research method: direct npm registry inspection, tarball extraction + .d.ts audit, live source code read across both repos. All claims graded HIGH / MED / LOW.
DRACO scores a cited research dossier on grounding (fraction of cited URLs that resolve), coverage, balance, cleanliness, and faithfulness, across 5 domains. Full investigation in ruvnet/agent-harness-generator (ADR-037→040), every number measured against live OpenRouter frontier models, key from GCP Secret Manager, nothing gamed.
Frontier tier (opus-4 / gpt-5 / gemini-2.5-pro), n=20:
| arm | Δ vs vanilla | verdict |
|---|
PR: ruvnet/symbolic-scribe#2 — Rust→WASM PromptOps compiler core + Optimizer UI
This gist records a full review pass: bug fixes, a native latency optimization, a
live OpenRouter-fusion benchmark wired to a GCP-stored key, and a publishable
agent harness.
A one-command scaffold for a non-interactive AI agent harness that runs on the GitHub Actions runner — the 9th MetaHarness host (ADR-033) and the only one built for CI/CD.
Every other MetaHarness host is interactive: a human starts a session and watches the agent work. GitHub Actions is different in kind. There's no human at the keyboard — execution is triggered by a webhook (a push, a pull request, an issue comment, or a schedule), the "operator" is the runner, and the harness must complete its task autonomously, emit structured output, and exit cleanly.
@metaharness/github-actions is a thin wrapper around the metaharness CLI. Running it is equivalent to npx metaharness my-bot --template minimal --host github-actions.