Cline 4.1.21 switches 19 providers to Claude Opus 5.5 by default, and retries local models that hit their output cap
Cline 4.1.21 flips the default model for 19 providers that do not pin one, 11 of them to Claude Opus 5.5, and now compacts and retries local model responses that hit the output-token cap.
Why it mattersA Cline user who never picked a model just changed models, and any team on llama.cpp, Ollama or LM Studio will get fewer failed tasks and less lost context on long replies.
A Cline user pointing at GitHub Copilot or Vertex without a pinned model now runs Claude Opus 5.5 by default. Cline 4.1.21 shipped this afternoon and moved the resolved default for 19 providers that leave the choice to Cline, 11 of them to Claude Opus 5.5.
Cline says GitHub Copilot and Vertex are in that group of 11. The release notes list the change as a refreshed catalog of 6,386 models across 209 providers, so any user on one of the 19 providers without an explicit model in their config will see a different default the next time they open a task. Cline does not name the other eight providers in the notes.
What that means for local models
The bigger fix in 4.1.21 is for developers running Cline against llama.cpp, Ollama or LM Studio. Cline says those servers cap generation at whatever context is left, regardless of the output-token budget set in the client. A long reply that runs into that cap used to end the task.
Cline now compacts the conversation and retries once when a local model hits the limit. If compaction cannot help, the existing concise-retry recovery still runs, and the partial answer is kept. That trades a possibly long completion for a shorter one, but it is a completion rather than a dropped task.
Smaller changes worth checking
Cline raised the minimum js-yaml version to 4.3.2 to pick up a parser security fix. The parser is what reads rule and skill frontmatter, so that patch reaches anyone shipping Cline skills.
Reopening a task that failed now shows the error and the retry option, rather than presenting the task as complete. On Windows, at-mention search results now show the correct file name, so a subfolder README no longer renders as /README.md or appears twice.
A command that prints nothing used to render a raw JSON blob as its output, such as [{"query":"git add -A","result":"","success":true}]. That is gone. Cancelling a request while it waits to retry an empty model response now takes effect straight away rather than after the backoff finishes.
A new provider, ai&, is added: an OpenAI-compatible endpoint serving open-weight models from Japan. That widens the list of places Cline users can reach open-weight models from, without changing anything for people not using it.
The consequence
Anyone on Cline who never picked a model just got a different one. If a workflow was tuned against a specific Copilot or Vertex default, that tuning may not hold from this release. Pin the model in the Cline settings if the change is unwanted. The local-model retry is the fix most teams will actually feel: a workflow that failed on the third or fourth long reply of a session should now complete, with the last message compacted rather than lost.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
