[1.257.0] — 2026-10-25
Two agent-fleet fixes: the in-app help assistant understands Devanagari and Tamil, and the usage page shows what Hermes actually costs.
Fixed
- Help assistant retrieval for hi/ta (CAR-103): the docs-assistant keyword tokenizer's word class was
[\p{L}\p{N}\u30fc]+— no combining marks. Every Devanagari matra (ा ै ्…) and Tamil vowel sign (ே ை ்…) cut a word into fragments; the length rule dropped short fragments (a Hindi query matched 0 sections) and longer words decapitated into garbage that collided with unrelated sections. The class keeps\p{M}now; the CJK bigram path is untouched. hi «मैं स्कैन कैसे चलाऊं?» and ta «ஸ்கேன் எப்படி இயக்குவது?» retrieve the right sections from the shipped bundles. - Usage cost for Hermes (CAR-106):
normalizeUsage()collapsed provider usage to{in, out}and silently dropped the gateway-reportedcost—#/usageshowed $0.0000 over 9.5 M Hermes tokens (its pricing-table row is 0/0). The usage record shape gains an optional backward-compatiblecost?; the aggregator prefers the provider-reported cost per record and falls back to the pricing-table estimate only when absent; garbage/negative values are dropped, never fabricated.
Notes
- Both fixes came from the agent fleet with red→green evidence (3 + 5 new cases; 14/14 and 10/10 green; neighbors 78/78 and 61/61). The first PRs (#470/#471) carried a stale worktree base that conflicted with the squash-merged v1.256.0 — superseded by clean cherry-picks (#472/#473).
- Deliberately not done: no retroactive re-pricing of historical
data/llm-usage.jsonllines (they carry no cost field; only new records benefit). - Tests 5713 (+8: docs-assistant ×3, usage ×5). No scanner sources touched (registry 121 = 116 EN + 5 RU · 116 adapters); no i18n keys.