The first tagged release. session-warden graduates from a session lifeguard to a full self-improvement harness for persistent OpenClaw agent fleets.
Fixed
- Recovery delivery restored (#16): OpenClaw 2026.6.x requires
--channelwith multi-channel gateways and--session-keyfor session keys; both were breaking every post-rotation recovery silently. Failure counters now reset on rotation completion, not delivery success — the mechanism that permanently wedged sessions in BACKOFF. - BACKOFF now escalates once via Telegram and throttles log spam to hourly (#19).
- Delivery errors surface stderr in logs; channel-less sessions skip
--deliver(#19). update-openclaw.shrecognizes retired dist patches on OpenClaw ≥ 2026.6.5 (#19).- Portable stat helpers for macOS/BSD (#19, closes the #2 proposal with GNU-first ordering).
Added
- Reflector (#17): nightly per-agent lesson distillation — Sonnet maker + independent Haiku skeptic, ACE-style append-only bullets staged for human approval,
apply-lessons.shmirrors approved lessons to GBrain with provenance. - Skill harvester (#18): weekly mining of repeated workflows into staged SKILL.md drafts +
promote-skill.sh. - Model scorecard (#20): weekly blind-judged benchmark across experimental agents, results to GBrain + Telegram.
- Memory evals (#20): monthly regression eval of each agent's memory quality with per-run deltas.
- Rotation summaries link GBrain page slugs instead of restating facts, and stage lesson candidates for the Reflector (#19).
- GBrain graceful degradation, weekly logrotate policy (#19).
Requires
bash, jq, flock, the claude CLI, an OpenClaw gateway, and optionally GBrain + a Telegram bot for notifications.