v3.10.2: Self-Learning Hardening (#508, #509)

What's New in v3.10.2

A self-learning hardening release that fixes two user-reported bugs which quietly broke pattern learning, plus broad improvements to forgetting, recall, and verification.

Critical fixes

• #508 — Task trajectories were silently recorded as failed. The generated post-task hook passed --success in a way that Commander parsed as failure, so every trajectory was stored with success=0 / agent=unknown and never became a pattern. Fixed the parser and the generated hook (real agent attribution). Re-run aqe init to pick up the corrected hook.
• #509 — Dream-scheduler readout was perpetually stale. lastDreamTime stayed null and pending-experiences grew without bound; the kernel scheduler now reconciles that state after each cycle.

Learning improvements

• EWC++ forgetting protection — proven, high-value patterns resist being crowded out by new low-value ones.
• Contradiction detection — conflicting "same context, opposite outcome" experiences are evicted from recall instead of merged.
• Routing-confidence de-dilution — recovers confidence on broad multi-domain tasks.
• Fixes for a dead failure-deprecation branch and an MCP unknown-tool error code.

New tooling

• Self-learning verification harness, real MCP protocol + tool-parity smokes, a resilient cross-platform hook shim (aqe init now installs and uses it), pretrain-from-history bootstrap, and a RaBitQ vector-memory utility.

Getting Started

npx agentic-qe init --auto

See the CHANGELOG and release notes for full details.