cline sdk/sdk/v0.2.0 — SDK v0.2.0

Cline SDK v0.2.0 is mostly a reliability release for the agent runtime behind Cline, the open-source AI coding agent. Cline rebuilt its products on this SDK this year, so the fixes reach its VS Code extension and command-line tool, not just developers building on the library. Introducing Cline SDK explains how the runtime was extracted from the original extension, and the migration write-up describes moving the extension's users onto it. The v0.2.0 release notes list several failures that cost users time or money and now behave differently.

The clearest fix is in context compaction, where the agent summarises older conversation to stay inside the model's input limit. After the first compaction, Cline's estimate of how large the conversation was ran about seven times too high, so compaction kicked in again every few tool calls. The estimate now uses the messages as they are actually sent, with tool results truncated. Auto-compaction also no longer runs when the model's input limit is unknown, though manual compaction still works. A similar fix applies to checkpoints, the snapshots of the workspace that let a user roll back an agent's edits: diffs now read files through one git process instead of up to two per file, after workspaces with many untracked files spawned thousands of them.

Other changes address sessions and failed requests. Starting a session that already had a run in progress used to replace its runtime, leaving the original run going but unreachable and writing its transcript to a stale copy. It now reattaches to the live session. Model requests retry at most three times after the first attempt instead of five. A turn whose connection drops after the model has streamed only its reasoning is now retried, while turns that produced text, media or tool calls are still never retried. The gemini provider stops sending native Google Search, which had made every turn fail with a quota error on free-tier keys for Google's Gemini 3 and later and had dropped every Cline tool on Gemini 2.x. Vertex keeps native search.

The release also reshapes which models users get by default. The model catalog shrank from 7,041 to 6,457 entries. ClinePass, Cline's subscription for open-weight models, now defaults to Moonshot AI's Kimi K3.1 Pro, and default models changed for 56 other providers, mostly to Kimi K3.1 Pro, MiniMax's M3.1 Flash or DeepSeek's V4.1 Flash 1017. Anyone who relied on a provider's previous default without choosing a model explicitly should check which one they are now running.