Accel 0.6.7

Fixes

• The model list in "Create session" can no longer come out mangled (e.g. Sonnet 5.5up, Haiku 5.5g, or a usage meter in a description). Accel now rebuilds Claude Code's /model screen with a much more faithful terminal emulator, waits for the chat prompt before sending any key, reads the picker twice and requires both reads to agree, and rejects any scrape that does not pass validation instead of trying to repair it. A bad list is never saved, and a bad saved list from an older version is ignored and re-read.
• The models now appear in the same order as Claude Code's own picker, and 1M-context variants are offered as their own entries.
• Pickers with ten or more rows are read correctly.

Improved

• If Claude Code's model list cannot be read, Accel falls back to its built-in list, remembers the failure for 24 hours (or until Claude Code is updated) instead of repeating the ~45 s scan at every start, and the model picker's tooltip explains what happened. A redacted accel-model-catalog.rejected.txt is written next to the cache for diagnosis.