Autopilot cycle 29 ran overnight; small, honest fixes landed in the train and a tiny homepage tidy keeps the shop window truthful. Notes below — short, evidence-backed, and actionable.
Shipped
-
cs2-train: autopilot cycle 29 ran and produced a small set of fixes: a legacy plan_versions → canonical migration to avoid POST /adaptive/plan 500s, launcher exec-line changes that seed
sv_hibernate_when_empty 0and+bot_join_after_player 0, and an authority-state change so drills take exclusive bot authority during the window. The commits are in the repo: https://github.com/pkia/cs2-train. - radar: appended the 2026-09-13 run log; the board shows no new actionable Proposed items after the sweep. See https://github.com/pkia/radar for the entries.
- personal-website: small content and layout caps applied to project cards and the "Currently" panel so the homepage stays concise; this keeps the portfolio honest without churn. Repo: https://github.com/pkia/personal-website.
On the radar
-
Train — prove a human run end-to-end (S): get a real CS2 client to connect from the Pi, fire a shot and advance a route. Next step: implement a 09:00 checklist that asserts
placed 5/5holds for 60s and captures the server log and attempt row. - Train — map the corpus (M): reconcile scenario slugs with prefire profiles. Next step: add a build-time check that fails when a scenario references an unmapped route and make the drill refuse a missing profile explicitly.
- Autopilot — exercise supervisor v2 under failure (M): inject a simulated provider failure in CI and assert the supervisor pins a safe provider and recovers within the budget guard.
Interesting reads
- "Self‑Improving AI Agents: The Reflection Loop" (Taskade) — a practical write-up of the generate→critique→revise pattern and how to scope pass/fail criteria for small, auditable agent loops.
- sysward — lightweight Linux monitoring for edge devices (GitHub) — a minimal alternative to Prometheus/Grafana for Raspberry Pi and headless servers where the full stack is too heavy.