ruflo Releases

This release (v3.10.24) validates the retrieval architecture's cross-repo generalization capability by pretraining on ruflo's history and testing on two unrelated corpora (agentdb and agentic-flow). Both cross-repo tests achieved exceptional results with nDCG@3 scores of 0.992-1.000, demonstrating the architecture generalizes to different commit conventions, vocabularies, and project scales.

May 30, 2026

This release introduces a joint re-grid optimization for the rerank path, discovering that it requires different hybrid sub-parameters (subjectWeight=3.0) than the non-rerank path (subjectWeight=2.0). This yielded a significant improvement: rerank nDCG@3 jumped from 0.900 to 0.963, bringing both paths to the same corpus ceiling.

May 30, 2026

This release (v3.10.22) introduces grid-search-tuned retrieval defaults against the ADR-081 labelled corpus, replacing the previous defaults that were tuned against a misleading regex proxy. The new hyperparameters (α=0.5, subjectWeight=2.0, mmrLambda=0.7) achieved a 7% improvement in nDCG@3 (0.900→0.963) and 100% top-3 accuracy, with a trade-off of +13ms latency.

May 30, 2026

This release introduces honest measurement for the retrieval system by replacing the regex-based relevance proxy with a labelled held-out corpus and proper nDCG/precision metrics. The finding revealed that the regex proxy was both over- and under-reporting relevance, with hybrid config showing 90% top-1 (vs 80% regex) and rerank showing 80% (vs 90% regex).

May 30, 2026

Release v3.10.20 introduces an optional cross-encoder reranker that significantly improves search relevance, pushing top-1 hit rate from 80% to 90% and top-3 hit rate from 80% to 100%. The feature uses a lazy-loaded MS-MARCO model combined with the existing hybrid approach via a weighted combination (default 0.5 hybrid + 0.5 cross-encoder), though it comes with ~25× latency cost (984ms vs 39ms).

May 30, 2026

This release introduces multi-field BM25 search with a 3:1 subject-to-body weight ratio and an opt-in type penalty for meta-commits, delivering a dramatic improvement from 0% to 80% top-1 hit rate. The ablation study found that adding type penalty actually hurts performance with current commit conventions, so it defaults to OFF but remains available as an opt-in feature.

May 30, 2026

This release introduces hybrid retrieval (BM25 + cosine + MMR) to address relevance issues in the pretrain harvester, achieving a dramatic improvement from 0% to 50% top-1 hit rate. It also adds outcome signal detection for identifying reverted and hotfixed commits, with new search parameters and Pattern.content field persistence.

May 30, 2026

Release v3.10.17 introduces pretrain self-learning from GitHub history (ADR-077), solving the day-one empty state problem where fresh ruflo installs started with zero patterns and zero trajectories. The script harvests commits and issues to seed the neural store through the same code paths real-time learning uses, achieving 100% retrieval match rate at ~3ms per item.

May 30, 2026

This release (v3.10.16) combines two major updates: Round B completes the trajectory pipeline wiring for hooks (post-edit and post-command now feed the trajectory pipeline, and trajectory-end now updates globalStats), while Round C introduces structured distillation with a new 4-field schema module, a 30-pair trajectory corpus, and an MRR benchmark harness that shows a 41.8% improvement in MRR (0.0964 to 0.1367).

May 30, 2026

Resolves the 'four contradictory stat sources' issue from ADR-074 by aggregating the view rather than the store. The four sources (globalStats, sonaCoordinator, memory-bridge, neural-patterns) measure different layers and now provide a unified stats interface with cross-store drift detection.

May 30, 2026

Wires up the self-learning subsystem that was reporting success but persisting nothing (issue #2245). Three CLI-side wirings are fixed plus honest multi-path output and adversarial hardening.

May 30, 2026

Bumps bundled agentdb to 3.0.0-alpha.16 with security fixes, new MCP surface, and honest measurement. Closes three security findings and adds real recall@k benchmark with verified 0.912 recall@10.

May 30, 2026

Bumps bundled agentdb to 3.0.0-alpha.15 to enable better-sqlite3 in the MCP memory bridge. Part A of issue #2235 that was upstream-only.

May 29, 2026

Fixes four user-reported bugs from the same triage round: opus alias, orphan prevention, uptime calculation, and Q-encoder. Each has a regression test.

May 29, 2026

Adopts three high-confidence patterns from Hermes-Agent: prompt caching, reasoning-tag scrub, and tool-loop circuit breaker. All three packages updated to 3.10.10.

May 29, 2026

Completes the intelligence-audit punch-list with per-task bandit learning and honest labeling. Fixes genuinely-fixable items and documents honest ceilings for the rest.

May 29, 2026

Two routing-learning correctness fixes: stale route cache and --explore false flag parsing. Both verified with tests.

May 29, 2026

Intelligence self-learning audit with critical fixes and honest performance numbers. Fixes negative-reward inversion and removes fabricated metrics.

May 29, 2026

Memory, routing and statusline bug-fix release addressing five reproducible bugs reported by external contributors. Each verified and covered by regression tests.

May 29, 2026

QMD v2.5.3 adds line-numbered output defaults, new `:from:count` suffixes for `qmd get`, and standardizes full-path handling across search and retrieval commands.

May 29, 2026