v1.11.0

1.11.0 (2026-10-30)

Features

• benchmarks: LongMemEval runner on the official protocol (#318) (c22b293)
• gateway: WSL2 autostart and crash supervision without systemd (#333) (15bc061)
• memory: default evidence telemetry to shadow mode (#328) (04bc81f)
• memory: evidence telemetry and shadow decision traces (PLAN-59 Phase 0) (#322) (ecc86c0)
• memory: per-channel retrieval diagnostics and skill applicability profiles (PLAN-58) (#309) (6f704fc)
• memory: PLAN-59 Phase 1 pilot harness (forked consolidation arms) (#329) (f9ac780)
• memory: shadow surfacing trial eligibility and assignment (PLAN-59 Phase 2) (#331) (b6c09df)
• memory: weighted RRF channel fusion with sweep and real-query validation (#332) (fe51748)

Bug Fixes

• compute: local operator pools are local trust, startup reconcile waits for the coordinator, beta installs update to stable (#308) (733b741)
• gateway: restarted gateway outlives its parent's lock instead of dying at the 5 s timeout (#316) (fce458c)
• memory: attach compute episode and nudge sinks on every manager get (#292) (3e824c9)
• memory: bound post-fusion boosts to the RRF scale; held-out live-memory benchmark; evidence-level LongMemEval metrics (#315) (45e9cfe)
• memory: GCCRF and curiosity-target defects (#321) (098582e)
• memory: keyword channel matches natural-language queries; failed embedding batch no longer aborts sync (#312) (a22e2df)
• p2p: fail-closed skill ingest and publish gates (#320) (5ce4577)
• skills: lint src/agents/skills again, dedupe fit text, two-node mesh harness, startup log (#311) (ee51008)
• skills: resolve the receiver's primary model from agents.defaults.model for the applicability fit (#310) (d75ef15)