2.4.35

Fixed

• Remove numbered turn banners from the transcript, keeping the clean user
  heading and prompt rail.
• Label shell tool actions as Run, and keep failed tools compact with a red
  cross and a clickable numbered detail view. Show failure cards only in
  verbose mode while retaining recorded command, output, and exit metadata.
• Move the moving three-dot working indicator into the empty prompt box. Keep
  drafts editable, pause its timer while typing or unfocused, and restore the
  usual prompt hint after work or when an approval needs an answer.
  Use a bright theme accent for the label and random theme colors for the dots.
• Show the keyboard selection on Free, Budget, and other BYOK model rows,
  keeping the highlighted model visible when navigating across groups.
• Preserve tool-call IDs across Core harness turns, saved sessions, and context
  compaction. Repair matching legacy tool results before sending history so
  strict providers such as OpenRouter do not reject subsequent prompts.
• Show branded top and bottom animations for three seconds when a run starts,
  then in one-second bursts every five seconds, keeping live status visible
  between bursts. Pause animations when the terminal loses focus and stop burst
  timers when the run ends.
• Refresh catalogs once at startup instead of hourly, and load LiteLLM only
  when needed by a selected connection instead of preloading on every launch.
• Run file searches, previews, filesystem autocomplete, and Git diff collection
  away from the TUI event loop. Reuse short-lived completion metadata, discard
  stale results, and keep typing responsive while slow filesystem work finishes.
• Stream directory-search files and retain only the first 50 preview lines
  instead of allocating whole files for those operations.
• Keep local warmup enabled, clearly label questions waiting for model readiness,
  and retain queued questions when warmup fails. Use one Ollama availability check
  and disable reasoning only for its tiny warmup request.
• Pass Ollama context and GPU options through LiteLLM without a nested options
  object that the server ignores; preserve explicit caller settings.