GitHub: CLI v3.0.50

• Added user-selectable color themes to the interactive TUI. Pick one with /theme, the command palette, or the Theme row in /settings — the picker previews each theme live. Built-in themes are Auto (terminal-adaptive, the default), Cline Dark, Cline Light, Tokyo Night, Gruvbox Dark, Nord, Dracula, Catppuccin Mocha, One Dark, Solarized Dark, and Solarized Light. Named themes paint the background, foreground, accents, syntax highlighting, and diff colors, and CLINE_THEME overrides the persisted choice at startup
• The git branch shown below the prompt now updates when you switch branches from another terminal or your editor, instead of showing whatever was checked out when the TUI started
• Telegram slash commands such as /clear now reach the connector command host — the Telegram library was intercepting them and they were silently dropped
• Racing connector launches no longer collide: an instance is claimed before it opens socket mode, the hub supervises connector processes, and doctor/connect skip connectors that are already starting. Connector tools are also enabled by default, and the Slack greeting is no longer replayed on reconnect
• Auto-approval settings are now honored over ACP
• Plan mode now hard-blocks file-editing shell commands instead of relying on prompting alone — run_commands stays available for read-only investigation, but file-manipulation commands, in-place editors (sed -i, perl -i), redirection to files, mutating git subcommands, package installs, and nested command strings (sh -c, eval, sudo) are rejected, on Windows and PowerShell too (from SDK v0.0.70)
• A turn that ends with a completed plan is no longer rendered as a failed turn when a plan-blocked command was its only tool call
• Running out of context is now recovered from instead of failing with a raw provider error: the run force-compacts and retries once, and the cases that genuinely cannot be recovered report why (from SDK v0.0.70)
• Empty model responses are now retried on every provider, not just Ollama — OpenRouter, Cline, and OpenAI-compatible endpoints previously failed the task outright with "Model returned empty response" (from SDK v0.0.70)
• Claude 4.6+ and 5.x models are no longer rejected with "thinking.type.enabled is not supported" when they resolve from the offline catalog or from a hand-typed model id (from SDK v0.0.70)
• Bedrock prompt caching works again — the provider was sending a cache format Bedrock silently discards, so cache reads and writes were always 0 — and Bedrock foundation models are now routed through geo inference profiles (from SDK v0.0.70)
• Reasoning models on OpenAI-compatible endpoints now receive max_completion_tokens instead of the rejected max_tokens, and requests to models without image support substitute the image content instead of failing (from SDK v0.0.70)
• MiniMax now inherits its default model from models.dev, and the model catalog picked up two new providers, Infomaniak and SCX.ai (from SDK v0.0.70)
• Upgraded the model layer to AI SDK 7 and switched Ollama to the native AI SDK provider (from SDK v0.0.70)
• Error telemetry no longer reports the same provider failure twice, and repeated failures from unattended retry loops are rate-limited (from SDK v0.0.70)

Full Changelog: cli-v3.0.49...cli-v3.0.50