Highlights
- GitHub release update check. Agents now notice when a newer release is published. One-line banner on
elophanto chat startup and a row in elophanto doctor. Cached daily — single unauthenticated GitHub API call per machine per day. Opt-out via telemetry.update_check: false. No metrics sent, only fetched.
- Memory leak fixes for long-running agents. Three unbounded accumulators were growing 1:1 with LLM call volume — overnight sessions ballooned RAM and the per-tick CPU walk made the terminal sluggish. All three now bounded:
CostTracker.calls → deque(maxlen=1000) + running token totals (kills the O(n) status-bar refresh).CostTracker._pending_records → deque(maxlen=1000) (flush() was dead code; llm_usage table empty in prod).ProviderTracker._events → deque(maxlen=100) (only last 20 ever read).
- Scheduler
update_schedule action. Lets the daily review loop close: agents can now amend a bad cron or rewrite a goal via schedule_list action=update without losing the schedule_id or run history that proves the fix worked. Previously you had to delete + recreate, breaking the audit trail.
- SQLite InterfaceError fix. The 2026-05-07 resource-typed concurrency rewrite enabled parallel
_run_one tasks; the read path on the shared connection wasn't locked, so concurrent cursors crashed with sqlite3.InterfaceError: bad parameter or other API misuse and took the scheduler down. Read path now locked via _conn_lock (renamed from _write_lock). 3 new pinning tests for 50/40/6 concurrent ops.
- Polymarket risk engine + safe-compounder + analyzer + mark-to-market. New tools and analytics for the Polymarket integration.
Upgrade
git pull && pip install -e .
Full diff
v2026.05.07...v2026.05.09
