Skip to content

claude-flow

v3.9.0 Feature

This release adds 3 notable features for engineering teams evaluating rollout.

Published 10d AI Agents & Assistants
✓ No known CVEs patched
Read the diff → Tool health → What is this tool? →

✓ No known CVEs patched in this version

Topics

agentic-ai agentic-framework agentic-rag agentic-workflow agents ai-agents
+14 more
ai-assistant ai-coding ai-skills autonomous-agents claude-code codex mcp-server multi-agent multi-agent-systems npm skills swarm swarm-intelligence typescript

Summary

AI summary

Unified Knowledge Graph Backend introduces shared graph_edges table, new MCP tools, SONA hooks, and benchmarked performance.

Changes in this release

Feature Medium

Adds `graph_edges` table with temporal columns and PQ-encoded embeddings.

Adds `graph_edges` table with temporal columns and PQ-encoded embeddings.

Source: llm_adapter@2026-05-25

Confidence: high

 —
Feature Medium

Adds `embedding-quantization.ts` implementing Int8 global-scalar PQ (400 bytes per 384‑dim embedding).

Adds `embedding-quantization.ts` implementing Int8 global-scalar PQ (400 bytes per 384‑dim embedding).

Source: llm_adapter@2026-05-25

Confidence: high

 —
Feature Medium

Adds `graph-edge-writer.ts` providing fire‑and‑forget SQL.js accessor for edge writes.

Adds `graph-edge-writer.ts` providing fire‑and‑forget SQL.js accessor for edge writes.

Source: llm_adapter@2026-05-25

Confidence: high

 —
Feature Medium

Adds `agentdb_graph-query` MCP tool with k‑hop CTE traversal, Personalized PageRank, and semantic cosine ranking.

Adds `agentdb_graph-query` MCP tool with k‑hop CTE traversal, Personalized PageRank, and semantic cosine ranking.

Source: llm_adapter@2026-05-25

Confidence: high

 —
Feature Medium

Adds SONA trajectory hooks `hooks_intelligence_trajectory-step` and `hooks_post-task` to write edges fire‑and‑forget.

Adds SONA trajectory hooks `hooks_intelligence_trajectory-step` and `hooks_post-task` to write edges fire‑and‑forget.

Source: llm_adapter@2026-05-25

Confidence: high

 —
Feature Medium

Adds `GraphEdgesSource` class and `createAutoGraphAdapter()` for plugin adapter contract in `plugin.json`.

Adds `GraphEdgesSource` class and `createAutoGraphAdapter()` for plugin adapter contract in `plugin.json`.

Source: llm_adapter@2026-05-25

Confidence: high

 —
Feature Medium

Adds `agentdb_graph-pathfinder` MCP tool with six algorithm variants and depth clamping.

Adds `agentdb_graph-pathfinder` MCP tool with six algorithm variants and depth clamping.

Source: llm_adapter@2026-05-25

Confidence: low

 —
Feature Medium

Enforces complexity budget in `agentdb_graph-query` via `maxNodesVisited`, `maxDepth`, and `maxMillis`.

Enforces complexity budget in `agentdb_graph-query` via `maxNodesVisited`, `maxDepth`, and `maxMillis`.

Source: granite4.1:30b@2026-05-25-audit

Confidence: low

 —
Feature Medium

Adds `agentdb_graph-pathfinder` MCP tool with six algorithm variants (`personalized-pagerank`, `dynamic-mincut`, `spectral-sparsify`, `temporal-centrality`, `connected-component-churn`, `witness-chain-divergence`) and clamps depth > 5.

Adds `agentdb_graph-pathfinder` MCP tool with six algorithm variants (`personalized-pagerank`, `dynamic-mincut`, `spectral-sparsify`, `temporal-centrality`, `connected-component-churn`, `witness-chain-divergence`) and clamps depth > 5.

Source: granite4.1:30b@2026-05-25-audit

Confidence: low

 —
Performance Medium

Benchmark shows 2345 ops/sec write throughput (≥500 target) and 578 bytes/edge SQLite footprint (≤1024 target).

Benchmark shows 2345 ops/sec write throughput (≥500 target) and 578 bytes/edge SQLite footprint (≤1024 target).

Source: llm_adapter@2026-05-25

Confidence: high

 —
Performance Medium

k‑hop depth = 1 p99 latency is 4.9 ms (<10 ms target).

k‑hop depth = 1 p99 latency is 4.9 ms (<10 ms target).

Source: llm_adapter@2026-05-25

Confidence: high

 —
Performance Medium

PQ encode p99 latency is 0.063 ms (<1 ms target).

PQ encode p99 latency is 0.063 ms (<1 ms target).

Source: llm_adapter@2026-05-25

Confidence: high

 —
Full changelog

ADR-130 — Unified Knowledge Graph Backend (All 6 Phases)

Gives all 4 existing graph layers (graph-node, AgentDB, ruflo-knowledge-graph, ruflo-graph-intelligence) a shared graph_edges sql.js table with PQ-encoded embeddings, two new MCP tools, SONA trajectory hooks, and a plugin adapter contract.

New features

Phase 1 — graph_edges schema + PQ encoder

  • graph_edges table with temporal columns: confidence, decay_rate, last_reinforced, witness_id, embedding_ref
  • embedding-quantization.ts: Int8 global-scalar PQ (400 bytes/384-dim); inlineCosine() for zero-decode similarity
  • graph-edge-writer.ts: thin sql.js accessor with fire-and-forget writes

Phase 2 — agentdb_graph-query MCP tool

  • k-hop traversal via recursive CTE (sql-cte backend)
  • Personalized PageRank (PPR) power iteration
  • Semantic cosine ranking on inline PQ embeddings
  • complexityBudget enforcement: maxNodesVisited, maxDepth, maxMillis

Phase 3 — SONA trajectory-to-graph hooks

  • hooks_intelligence_trajectory-step: writes trajectory-caused edges fire-and-forget
  • hooks_post-task (success=true): writes reinforced-by edges fire-and-forget
  • Neither write blocks tool response (<200ms latency preserved)

Phase 4 — Plugin adapter contract

  • GraphEdgesSource class: default KnowledgeGraphSource reading from graph_edges
  • createAutoGraphAdapter(): zero-boilerplate autoRegister path
  • graph_adapter field in plugin.json schema (documented in ruflo-plugin-creator SKILL.md)

Phase 5 — agentdb_graph-pathfinder MCP tool

  • 6 algorithm variants: personalized-pagerank, dynamic-mincut, spectral-sparsify, temporal-centrality, connected-component-churn, witness-chain-divergence
  • Depth > 5 clamped; non-existent seed returns empty paths not error

Phase 6 — Benchmark + CI

  • benchmark-graph.mjs: 6/6 targets met (2345 ops/sec write, 578 bytes/edge, k-hop depth=1 p99=4.9ms)
  • 5 new CI jobs in v3-ci.yml gating all phases

Benchmark

| Metric | Result | Target |
|--------|--------|--------|
| Write throughput | 2345 ops/sec | ≥500 ✓ |
| SQLite footprint | 578 bytes/edge | ≤1024 ✓ |
| k-hop depth=1 p99 | 4.9ms | <10ms ✓ |
| k-hop depth=3 p99 | 0.1ms | <50ms ✓ |
| PQ encode p99 | 0.063ms | <1ms ✓ |
| PQ decode p99 | 0.031ms | <0.5ms ✓ |

Test baseline

1999 passed | 46 skipped — no regression

PR: #2129 | Tracking: #2128

View diff on GitHub

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

Share on X Share on Bluesky

Track claude-flow

Get notified when new releases ship.

Sign up free

About claude-flow

Deploy multi-agent swarms with coordinated workflows.

All releases →

Related context

Beta — feedback welcome: [email protected]