v0.preview.91

v0.preview.91 - 2026-08-30

Added

• Anthropic xhigh reasoning effort — available on Claude Sonnet 5 and Opus 5, Opus 4.7/4.8, and always on Fable/Mythos 5; Opus/Sonnet 4.6 keep their four-level adaptive ladder
• gpt-5.4-mini, gpt-5.4-nano, and glm-5.3-flash in the Antix catalog
• Completed reasoning blocks show how long the model thought — Thought for <duration>, derived from protocol event timestamps so replays report the same number

Changed

• Compaction tuning: manual /compact folds more history than automatic compaction, so running it twice in a row no longer re-folds a fresh summary; an overflow retry stops once it stops making progress instead of spending more provider calls; and a single oversized tool result is trimmed rather than costing a rejected request
• GLM 5.3 defaults to max reasoning effort when an OpenAI-compatible profile supplies no effort, and Anthropic stream errors carrying the invalid_request provider alias are classified instead of surfacing raw
• Telemetry uses the standard OTEL_EXPORTER_OTLP_ENDPOINT and OTEL_EXPORTER_OTLP_HEADERS environment variables

Fixed

• Markdown link URLs render in the design steel blue instead of hardcoded ANSI blue, which was nearly unreadable on dark backgrounds