AI 程式開發工具
Cline 4.1.17 Stops Resending Full Conversations to Clients and Closes a Checkpoint Commit-Overwrite Risk
Cline fixes an issue that broadcast the full transcript with every state update during long-running sessions, causing Hub memory usage to balloon to tens of gigabytes, and now sends only state snapshots. The release also tightens checkpoint handling, subagent cancellation, and remote MCP connection behavior, but deployments that do not pin models may produce different results after major default-model updates.
Version 4.1.17 of the open-source coding agent Cline focuses on reliability issues affecting long-running sessions and multi-agent execution. The most technically significant bug was in the background Hub: every session state change copied and broadcast the full conversation transcript to all connected clients. After an agent accumulated large volumes of tool output, a single update could reach several megabytes, and repeated copying could drive process memory usage into the tens of gigabytes. The new release reduces these broadcasts to state-only snapshots, directly removing a data path whose overhead repeatedly grew with conversation length.
Several important fixes also address file and process boundaries. Before restoring a checkpoint, Cline now checks whether any subsequent Git commits exist and refuses the operation if they do, preventing commits from being silently moved off the current branch and recoverable only through the reflog. `apply_patch` now preserves existing CRLF line endings, and a hook that fails to start no longer crashes the extension core. Aborting a parent task also cancels the subagents and teammates it spawned, reducing cases in which users believe a task has stopped while background processes continue writing data or incurring charges. Unreachable SSE or streamable HTTP MCP servers now have a 10-second timeout budget instead of indefinitely delaying session startup.
On provider compatibility, an empty capability list was previously misinterpreted as “explicitly unsupported,” causing tool calls and some image inputs to be silently removed for models served through Dify, SAP AI Core, opencode, and Codex CLI. The new version corrects this behavior. OpenAI Codex OAuth now also distinguishes genuinely rejected tokens from temporary network errors and explains when local callback port 1455 is already in use.
The upgrade includes one easily overlooked behavioral change: the built-in model catalog adds ten providers, while the resolved default models for 57 providers have been updated. Anthropic’s default has also changed to Claude Fable 5.1. Teams that rely on unpinned model names should rerun regression tests for tool calling, cost, permissions, and outputs before deployment. The release notes do not yet provide a reproducible stress test comparing memory usage before and after the fix, so the tens-of-gigabytes figure should be treated as a known failure case rather than a general performance guarantee.