• v1.2.6 b6c373d3fa

    v1.2.6
    Some checks failed
    CI / build-and-test (push) Failing after 14s
    CI / docker-build (push) Successful in 22s
    Stable

    joeyr released this 2026-07-12 22:42:38 -05:00 | 32 commits to main since this release

    Bug-fix release: unbreaks LoL live-game predictions and the /aufhocker chat command against Sonnet 5, records prediction outcomes again, and paces Riot API calls more conservatively.

    Fixed

    • Stop sending temperature to Claude. The current Claude families (Sonnet 5 and newer) reject temperature with a 400 (temperature is deprecated for this model), which was failing live-game predictions, lane detection, and — whenever invoked — the /aufhocker chat command itself. Dropped the parameter from all three call sites and removed the now-dead CLAUDE_TEMPERATURE config field, its validation, and the env templates.
    • Record prediction outcomes again. The recordResult UPDATE bound $3 to both a varchar assignment (actual_winner) and a varchar comparison (predicted_winner); with node-pg's untyped parameter, Postgres deduced two conflicting types and rejected the query (42P08: inconsistent types deduced for parameter $3 — text versus character varying). Every result write failed, so accuracy stats stayed empty. Cast $3::varchar in both spots.

    Changed

    • More conservative Riot API pacing. A game detection fans out ~80 Riot calls (rank + mastery + up to 60 match-v5); the shared client throttle sat at 90 req/2min — barely under a personal key's ~100/2min ceiling, which is also shared with the lol-companion project — so the match-v5 burst tripped 429s (absorbed by retry, logged as warnings). Lowered the throttle and made it env-tunable via LOL_RIOT_MAX_REQUESTS_PER_2MIN (75), LOL_RIOT_MAX_REQUESTS_PER_SEC (15), and LOL_RIOT_MIN_SPACING_MS (350).
    Downloads