v3.18.2

ruvnet/ruflov3.18.2Jul 3, 2026by ruvnet

AI Summary

Completes the integration of native ruvllm checkpoints by fixing version floors, directory creation for checkpoints, and capability reporting.

Key Highlights

Native checkpoint persistence now active using @ruvector/ruvllm@2.5.7
Fixes 'neural train' checkpoint creation to prevent ENOENT errors
Version-gated capability reporting in 'neural status'
Removed stale dependency directories from dev-tree

New Features

Checkpoint persistence end-to-end integration
Automatic parent directory creation for checkpoints

Full Release Notes

Completes the #2549 arc: ruflo now actively uses the checkpoint persistence shipped in @ruvector/ruvllm@2.5.7 (RuVector#637).

## Changed
- **Version floors** `@ruvector/ruvllm >=2.5.7` in the ruflo wrapper overrides, root npm overrides, and v3 pnpm overrides — every install path is guaranteed the fixed version
- **`neural train` checkpoints actually land**: `LoRAAdapter.saveCheckpoint` now creates parent directories before writing — the JS fallback threw ENOENT on missing dirs and silently returned false, so the best-effort checkpoint never materialized even after the upstream fix
- **Version-gated capability reporting**: `neural status` shows 'native @ruvector/ruvllm pipeline + disk checkpoints' only when the resolved ruvllm is >=2.5.7 — never advertises persistence a stale install can't deliver
- Dev-tree hygiene: removed a stale npm-installed `@ruvector/ruvllm@0.2.4` directory that had shadowed pnpm resolution since April

## Proof
```
neural train -p coordination -e 2
→ .claude-flow/neural/lora-checkpoint-<ts>.json  (format: ruvllm-checkpoint v1, 48KB)
```

#2549 status: neural-status reporting fixed (3.18.1) → upstream persistence fixed (@ruvector/ruvllm 2.5.7) → ruflo wired end-to-end (this release). Remaining follow-up: routing `neural train` fully through the native TrainingPipeline.

PR: #2555 · Upstream: ruvnet/RuVector#637