Cline v4.1.21: a new default model for 19 providers
1 month ago
What changed in Cline v4.1.21: a new default model on 19 providers, the ai& provider, local models that recover from long replies, five fixes, a js-yaml security floor, and Linux for the desktop app. https://github.com/cline/cline/releases/tag/v4.1.21 https://github.com/cline/cline/blob/v4.1.21/CHANGELOG.md https://github.com/cline/cline/compare/v4.1.20...v4.1.21
Ask about this presentation
Chapters
0:00Cline v4.1.21
0:1201 Models and providers
0:15Claude Opus 5.5
0:25Model catalog
0:36ai& provider
0:4702 Local models
0:51Output cap
1:02Compact and retry
1:12Partial answer kept
1:2203 Tasks and commands
1:24Failed tasks
1:33Silent commands
1:43Cancel during retry
1:54Windows @-mentions
2:0404 Security
2:07js-yaml 4.3.2
2:1805 The desktop app
2:21Linux
2:32Diagnostics, reasoning, voice
Show transcriptHide transcript
In Cline 4.1.21, nineteen providers that don't pin a model now resolve to a different default. If you use one of them without choosing a model, your next task runs on a new one.
Which model answers, and who serves it.
Eleven of those nineteen providers land on Claude Opus 5.5. GitHub Copilot and Vertex are among them. If you want to stay on the model you had, pin it.
The built-in model catalog grew to about sixty-four hundred models from 209 providers. That's nearly a hundred and fifty more models than the previous release carried.
There's a new built-in provider called ai&. It's an OpenAI compatible endpoint that serves open-weight models from Japan, and you pick it like any other provider.
Long replies on llama.cpp, Ollama and LM Studio.
These local servers stop generating once the context window is full, whatever output budget you set. On a long session, that limit used to end the task in the middle of an answer.
Now, when a local model hits that output limit, Cline compacts the conversation to free up room and retries the reply once, so the task keeps running.
When there's nothing left to compact, the existing recovery takes over and asks the model to answer more concisely. Either way, you keep the partial answer it already wrote.
What the chat shows you.
Reopen a task that failed, and you now see its error along with the retry option. Before this release, a failed task came back looking like it had finished.
Run a command with no output, like git add on a clean tree, and the chat used to print a raw JSON record where the output goes. That record is gone.
When a model sends back an empty response, Cline waits and then retries. Cancelling during that wait now stops the request right away, instead of after the backoff finishes.
On Windows, the at-mention picker showed a subfolder's README as if it sat at the root. It now shows the full path, and an open file appears once instead of twice.
The parser behind your rules and skills.
The minimum js-yaml version is now 4.3.2. That brings in a security fix for the parser Cline uses on the frontmatter of your rule and skill files.
The Cline desktop app, version 0.0.35.
The desktop app adds Linux support. It ships as deb and RPM packages, and an update downloads in the background, then installs when you choose restart now.
Settings can now export a diagnostics bundle to attach to a bug report. Your reasoning effort is remembered per provider instead of resetting to low. And voice input with live transcription works again.


