v3.9.26: 10 Learning-Pipeline Fixes

What's New

Ten learning-pipeline fixes reported by @Jordi-Izquierdo-DDS against v3.9.24. Individually they look like small short-circuits in the post-task hook chain; together they were keeping the end-to-end Stream B/D/F learning loop pinned at zero — experiences tagged agent='unknown' and task='edit: ', Q-learning buckets all converging on one key, dream insights generated but never applied, patterns crossing the promotion thresholds but staying short-term forever, and stale sentinels inverting the routing-quality metric.

After this release, the learning loop actually converges on real per-agent signal instead of zeros.

Fixed

• #453 — post-edit now reads tool_input.file_path from the Claude Code event stdin when $TOOL_INPUT_file_path doesn't expand; fixes 88% of captured experiences losing their file path
• #454 — aqe hooks stats --json falls back to DB aggregation when in-memory counters are zero (hook subprocesses start fresh every invocation); live verify: routingRequests: 1904, learningOutcomes: 272, patternSuccessRate: 0.75
• #455 — bridge loop in persistTaskOutcome now promotes patterns crossing the successful_uses≥3 / success_rate≥0.7 / confidence≥0.6 thresholds; previously they stayed short-term forever
• #456 — hook-fired dream cycles now apply actionable high-confidence insights inline; reporter saw 378 unapplied vs 9 applied before fix
• #460 — Q-learning now falls back to the pre-task bridge's recommended agent when --agent env-var is unavailable; every rl_q_values row used to land on action_key='unknown'
• #461 — concurrent hooks no longer overlap into database is locked failures; dream_cycles peek before opening engine bails out with reason='already-running'
• #462 — RVF init purges orphan vectors when vectorCount > qe_patterns count; ghost vectors no longer route to non-existent IDs
• #463 — experience_applications.tokens_saved populates with round(qualityScore * 100) instead of hardcoded 0
• #464 — high-quality cli-hook experiences (post-edit, quality ~0.75) are no longer blanket-excluded from consolidation; the existing HAVING avg_quality ≥ 0.5 clause is the correct quality gate
• #465 — post-route sweeps stale routing_outcomes sentinels older than 5 minutes; reporter saw 122/149 rows stuck at -1, inverting AVG(quality_score) to -0.717

Closed as duplicates

• #457 (dup #454), #458 (dup #455), #459 (dup #456)

Security

• Patched 11 Dependabot alerts (#476): @protobufjs/utf8, OpenTelemetry bumps

Getting Started

npx agentic-qe init --auto

See CHANGELOG and docs/releases/v3.9.26.md for full details.

Thanks to @Jordi-Izquierdo-DDS for the meticulous bug reports and patch shapes.