Changelog

0.9.2 patch Alert composer retries, quieter nights

Most of this release is about what happens when things almost go wrong, rather than when they do. The composer now retries on LLM 429s with jitter, and quiet hours learned about holidays.

• add Composer retries on LLM 429 with exponential backoff (max 3 tries, jittered). If all fail, fall back to plain alert. #482
• add silence.holidays: ["DE", "EG"] reads ical feeds and silences severity=info on those days. #461
• fix Slack code blocks larger than 3 KB were truncated. Now split across messages with a continuation marker. #479
• fix Postgres replica-lag check would crash if pg_stat_wal_receiver returned NULL. Handle the nil. #477
• perf Dropped the SQLite WAL checkpoint from every 30s to every 5m. ~40% less disk churn on small hosts.
• chg Default LLM model bumped to gpt-4o-mini. Same quality on our eval set, ~70% cheaper.

--- a/internal/composer/client.go
+++ b/internal/composer/client.go
- if resp.StatusCode != 200 { return nil, ErrUpstream }
+ switch resp.StatusCode {
+ case 200:        return parse(resp), nil
+ case 429, 503:   return nil, ErrRetry   // new: retry path
+ default:         return nil, ErrUpstream
+ }

0.9.0 minor LLM checks: ask the model, not the metric

New check type: type: llm. Run a query, hand the result to a model, and ask a yes/no question. This is what we use to notice that our signup funnel is sad before any per-metric threshold fires.

• add type: llm check. Give it a SQL query and a question; the model decides if it's a problem.
• add tellr explain <name> writes a dry-run alert for any check without sending.
• add Config hot-reload. Save tellr.yaml, changes apply on the next tick. No restart.
• chg Alert payload renamed check_id → check.name for consistency. Old field still emitted, will be removed in 1.0.
• sec Signed releases with cosign. Checksums on every release page.