This release includes 2 security fixes for security teams reviewing exposed deployments.
Topics
+14 more
Affected surfaces
Summary
AI summaryUpdates New capabilities, Performance, and GAIA integrity family across a mixed release.
Full changelog
Two merged feature trains, driven by a 4-review audit (SOTA research + performance + architecture + security) implemented by six concurrent agents.
Performance
- CLI startup 5–10x faster — lazy ONNX embedder init removed a 450–2,800ms tax from every command (
--help0.50s → 0.09s, whole CLI back under the <500ms target) - Batch embeddings in memory import (200-entry import: 1220ms → 115ms), local-first ruvector resolution, metaharness skills now resolve from a pinned versioned cache (
score0.35s, zero registry traffic)
Security
- Killed third-party
npx metaharness@latest(supply-chain HIGH) — pinned versioned-cache resolution with graceful degradation intact - OTel floor-became-ceiling overrides lifted (root
npm audit --omit=dev: 30 → 0); grpc-js / hono / brace-expansion GHSAs cleared - CodeQL high fixed: polynomial ReDoS in A2A card slugify (remote input) → linear scan
GAIA integrity family (ADR-167 / 168 / 169)
- ADR-167: pre-submission exploit audit (AUD-1..10) + signed earning-integrity attestation — signing proves bytes; the audit proves the score was earned. Honest
harness_gapskips where the harness can't yet supply evidence - ADR-168: harness evidence recording (trajectories, tool names+args, split provenance, signed judge cache) — the forward contract that turns those skips enforceable (#2548)
- ADR-169: benchmark reporting integrity standard — strict-EM headlines, view-labeled scaling arms, disclosed no-work/empty rates, mandatory reproducibility blocks
New capabilities
hooks_model-verify— confidence-gated tier escalation ($0 structural verifier; route → generate → verify → escalate; feeds the routing bandit)- Temporal validity on hierarchical memory (validFrom/validUntil + supersede-on-conflict, Zep/Graphiti-style)
- A2A Agent Cards (spec 1.0.0) — federation discovery interop at
/.well-known/agent-card.json, ADR-166 bind posture preserved - Execution-state-tree trajectory retrieval prototype (MAGE-style) behind
strategy: "state-tree" metaharness evolve --diagnose— GEPA failure-class diagnosis- Browser session recording fixed end-to-end (trajectory flags were invalid on every prior ruvector version)
Validation
tsc clean · MCP contract 171/171 (15 metaharness tools) · plugin smoke 122/122 · memory 442/442 · federation 600/600 · root audit 0 vulnerabilities · CI fully green on merged main (incl. the #2552 lockfile hotfix that had 25 main jobs red)
PRs: #2543, #2547, #2552 · Issues: #2542, #2548 · Previous: v3.17.0 (metaharness learn + gepa)
Breaking Changes
- Removed lazy ONNX embedder init causing a 450–2800 ms startup tax for every command.
Security Fixes
- CVE/GHSA: Polynomial ReDoS in A2A card slugify fixed – linear scan implementation.
- Supply‑chain HIGH risk removed by killing third‑party `npx metaharness@latest`; now uses pinned versioned‑cache resolution.
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
Related context
Related tools
Beta — feedback welcome: [email protected]