v3.18.1

ruvnet/ruflov3.18.1Jul 3, 2026by ruvnet

AI Summary

Patch release fixing a bug where the native ruvllm training backend capability was incorrectly reported as 'Unavailable' in the neural status command.

Key Highlights

Fixed neural status reporting for native training backend
Corrected capability probe logic using resolveTrainingBackend()
Removed incorrect claim about ruvllm checkpoints being enabled

Full Release Notes

Patch release fixing #2549 (reported by @pacphi — exemplary report with dist-verified line numbers and a capability-vs-environment proof).

## Fixed
- `neural status` reported the bundled @ruvector/ruvllm training path as **Unavailable — Install @ruvector/ruvllm** even though the module was installed and functional. Two defects: a dead `_trainingBackend` variable (declared, returned, never assigned) and contrastive availability read only from an in-process global a fresh status process never populates.
- Backend now comes from a capability probe (`resolveTrainingBackend()` — module resolution, pipeline stays lazy); Contrastive Trainer has three honest states (Active with session counts / Available — ready on demand / genuinely Unavailable); the Install hint only shows on true resolution failure.
- Dropped the 'ruvllm checkpoints enabled' claim — upstream `saveCheckpoint()` (@ruvector/ruvllm 2.5.6) verifiably writes no file; tracked separately.

## Not in this release (follow-ups from #2549 triage)
- Routing `neural train` through the native TrainingPipeline
- Upstream checkpoint persistence fix

PR: #2554 · Regression test pins the capability contract in both resolve/no-resolve environments.