Cline releases SDK v0.2.0 and Desktop v0.0.49 with major stability fixes

Cline has rolled out a coordinated set of releases across its SDK, Desktop application, and CLI, marking a significant step toward stabilizing its core agent infrastructure. The primary release, SDK v0.2.0, is accompanied by Desktop v0.0.49 and CLI v3.0.74, all of which address persistent issues with session persistence, provider-specific API failures, and context management. According to the GitHub release notes, these updates aim to resolve long-standing complaints about "stale snapshots" and "hanging requests" that have plagued developers using Cline for complex, multi-turn coding tasks.

Session Management and Stability

The most critical change in SDK v0.2.0 is the reimplementation of session reattachment logic. Previously, starting a session that already had a run in progress would replace the runtime, leaving the original run unreachable and its transcript writes landing on a stale snapshot. Now, the system reattaches to the live session. If a start request includes a prompt while a run is in progress, it is rejected to prevent state corruption. This change is mirrored in Desktop v0.0.49 and CLI v3.0.74, where opening a task that is already running in another Cline app now attaches to that run instead of starting a second, unstoppable instance. Furthermore, requests that were waiting on an app that closed mid-task now fail with a clear message instead of hanging indefinitely.

The hub architecture has also been refined to track every client that registered a capability for a session. Requests are sent to the first client still connected; if none remain, the request fails quickly with an actionable message. This eliminates the "waiting forever" scenario that occurred when a client disconnected during a tool call.

Provider and Model Fixes

A major focus of this release cycle is fixing provider-specific bugs that caused total workflow failures. The Gemini provider no longer sends native Google Search. This was a critical fix because, on Gemini 3+, free-tier keys failed every turn with a 429 quota error, and on Gemini 2.x, the native search feature caused every Cline tool to be dropped. Vertex users retain native search capabilities.

For Kimi K3, the gateway no longer incorrectly maps "max reasoning" to "xhigh," a setting the model rejects. Now, Kimi K3 correctly receives "max" reasoning for Extra High and Max settings. Additionally, model requests now retry at most 3 times after the first attempt, down from 5. This reduces latency for failed requests while still providing resilience. Failed attempts are recorded in Langfuse traces, and cancelled requests are no longer reported as errors, improving observability.

Context and Compaction

Auto-compaction, a feature that manages context window limits, has been overhauled. It no longer runs when the model's input limit is unknown, preventing unnecessary and potentially harmful compaction. More importantly, the size estimation logic has been fixed. Previously, it overestimated context usage by about 7x, causing compaction to run again every few tool calls. Now, the estimate uses the messages as they are actually sent, with tool results truncated. This significantly reduces the frequency of compaction events, leading to smoother long-running sessions.

Desktop and CLI Enhancements

Desktop v0.0.49 introduces several new UI features. A new "Changes" button in the session header displays line counts and opens a pane listing files changed in the last turn or across the session, with diffs. A project files panel provides a file tree beside the conversation, with git status indicators and a marker on files the session changed. An integrated terminal panel is now available, toggleable with Ctrl+backtick. The Local Models settings page has been expanded to manage Ollama and LM Studio models, including setup guidance and memory usage indicators.

CLI v3.0.74 focuses on robustness. The cline hub status and cline doctor commands no longer hang on a frozen Hub; the health check now times out after 3 seconds. Compiled CLI binaries now load plugins and serve the dashboard from resources embedded in the binary, improving portability. Plugin slash commands now run in the active session, ensuring consistency.

What It Means for Developers

For developers using Cline, this release should result in a more predictable and stable experience. The session reattachment fix is particularly important for teams using multiple Cline instances or for those who frequently restart their IDEs. The Gemini and Kimi fixes are critical for users relying on these providers, as the previous bugs rendered the tools unusable in many cases. The improved compaction logic should reduce the "amnesia" effect in long sessions, where the agent loses context due to aggressive compaction.

Developers should update to the latest versions of the SDK, Desktop, and CLI to benefit from these fixes. Users of Gemini free-tier keys should verify that their tools are working correctly after the update. Those using Kimi K3 should check that their reasoning settings are applied as expected. The new Desktop UI features, particularly the project files panel and integrated terminal, offer a more integrated development environment, reducing the need to switch between Cline and external tools.

What to Watch

• Langfuse Integration: Monitor the new Langfuse trace recording for model request retries to ensure that failed attempts are correctly logged and that cancelled requests are not misreported as errors.
• Compaction Frequency: Observe the frequency of auto-compaction events in long sessions to confirm that the new size estimation logic is working as intended and that context is not being lost unnecessarily.
• Multi-Instance Stability: Test the session reattachment logic in multi-instance scenarios to ensure that no state corruption occurs when multiple Cline apps are connected to the same session.
• Provider Compatibility: Verify that the Gemini and Kimi fixes work across all supported versions and configurations, particularly for users with custom API endpoints or proxy setups.