Cline's Vertex integration stops filtering the model picker to a hardcoded global-endpoint list and lets you type any Vertex model ID, including Fable 5. The Yolo Mode toggle is gone — it was never read by the approval path — and existing Yolo/auto-approve-all setups are migrated to auto-approving every action so unattended runs keep working.
Adds cline mcp uninstall and lets schedules inherit saved provider settings. The bulk is reliability work inherited from SDK v0.0.72: interrupted turns no longer silently drop queued prompts, a hung MCP server no longer blocks session creation, and LiteLLM traffic moves to Chat Completions so proxy setups work.
Reliability release for the Cline SDK centred on interrupt and restart semantics: prompts queued during a turn survive an abort, session context persists across aborts and hub restarts, and failed queued turns now report run.failed instead of completing silently. A hung or unconfigured MCP server can no longer stall session creation, and LiteLLM traffic moves from the Responses API to Chat Completions.
Remote MCP servers can be authorized with OAuth from Settings → MCP, including servers that need a pre-registered client ID/secret rather than dynamic registration. Turns that fail — including sending with no model credentials — now show an error in the transcript instead of failing silently, and MCP errors are attributed to the individual server rather than the whole page.
Reasoning effort now resolves the same way across providers — including Ollama — instead of going through per-provider thinking overrides, and an explicit request to turn reasoning off is honoured everywhere. Also adds meta/muse-spark-1.2-contributor on the Cline provider and fixes error telemetry attribution. Matters if you set reasoning per provider today.
Reasoning effort levels and enable/disable flags now map onto the AI SDK's native reasoning setting across providers, which the notes describe as replacing the previous per-provider thinking overrides; an explicit disable takes priority. Also refreshes the model catalog from models.dev and fixes telemetry model attribution.
The model can no longer switch itself out of plan mode mid-turn — that transition is now user-driven only — and plan mode hard-blocks file-editing shell commands. Anyone whose workflow depended on the agent moving itself into act mode has to change how they drive it.
Ships one universal macOS binary for Apple Silicon and Intel, with existing per-architecture installs migrating on next update. Also surfaces context-window usage in the composer and inherits the SDK's plan-mode command blocking and provider retry fixes.