v4.1.4
ClineSummary
This release introduces Chutes as a new provider and enhances the slash command menu to show skills alongside workflows. It also brings significant improvements to plan-to-act switching, making it user-driven, and hard-blocks file-editing shell commands in plan mode for better security.
New Features
- Added recognition for Chutes as a provider.
- Introduced the display of skills alongside workflows in the slash command menu, with disambiguation for commands sharing names.
Bug Fixes
- Fixed an issue where turns completing with a plan were incorrectly treated as failed when a plan-blocked command was the only tool call, ensuring the turn correctly re-runs the presented plan instead of appearing to do nothing.
- Fixed tool paths to display relative to the workspace in the chat view instead of absolute paths.
- Fixed pending attachments to reset when starting a new task, preventing images from previous tasks from carrying over.
- Fixed a generic failure to surface a clear error when the selected provider has no API key configured.
- Fixed MCP tool and resource lists to refresh when a server sends a `list_changed` notification, instead of only showing a toast.
- Fixed installed plugins to display under their real package names instead of all appearing as "index".
- Corrected the Linux keybinding label in the Plan/Act mode tooltip.
- Fixed recovery from running out of context by compacting and retrying once, and providing explanations for genuinely unrecoverable cases.
- Fixed empty model responses to retry on every provider, not just Ollama, resolving hard "Model returned empty response" failures on OpenRouter, Cline, and OpenAI-compatible endpoints.
- Fixed Claude 4.6+ and 5.x models from being rejected with "thinking.type.enabled is not supported" when resolving from the offline catalog or a hand-typed model ID.
- Restored Bedrock prompt caching by ensuring the correct cache format is used and routing Bedrock foundation models through geo inference profiles.
- Fixed requests for reasoning models on OpenAI-compatible endpoints by sending `max_completion_tokens` and substituting image content for models without image support.
- Fixed MiniMax default model inheritance from models.dev and refreshed the bundled catalog to include Infomaniak and SCX.ai.
- Fixed error telemetry to report provider failures only once and rate-limit repeated failures from unattended retry loops.
Improvements
- Removed model-initiated plan-to-act switching; switching out of plan mode is now user-driven.
- Hard-blocked file-editing shell commands in plan mode for enhanced security, while still allowing read-only investigation.
- Bundled catalog refresh adds Infomaniak and SCX.ai.