Cline v4.0.2 リリース Cline Releases v4.0.2

Cline, the open-source AI coding agent that runs inside Visual Studio Code, has shipped version 4.0.2, a maintenance update centered on better support for reasoning-capable language models. The release matters because it extends fine-grained control over how much "thinking" a model performs before responding, a setting that directly affects code quality, latency, and token cost for developers using agentic assistants on real projects.

The headline change is added support for a new reasoning effort tier, called xhigh, aimed at DeepSeek thinking models. Reasoning effort is a parameter that tells a model how much internal deliberation to allocate to a task, typically exposed as discrete levels such as low, medium, and high. Higher levels generally produce more thorough planning and fewer mistakes on complex tasks, but they consume more tokens and take longer to return. By introducing an xhigh option, Cline lets users push compatible DeepSeek models toward maximum deliberation when a problem is hard enough to justify the extra cost, while still keeping cheaper settings available for routine edits.

DeepSeek is a Chinese AI lab whose reasoning-oriented models, including the R1 series and its successors, gained attention for delivering competitive performance at relatively low prices, largely because they are open-weight and can be served by many providers. Their distinguishing feature is a visible chain-of-thought stage, where the model works through a problem step by step before producing a final answer. Tools like Cline benefit from surfacing or hiding that reasoning, so the alignment between Cline's effort tiers and DeepSeek's thinking behavior is a meaningful integration detail rather than a cosmetic one.

This release also improves the ClinePass provider, Cline's managed gateway for accessing models without requiring users to assemble individual API keys. The update is described as offering clearer reasoning controls and an improved model selection interface, which suggests the team is consolidating how users discover and configure these effort levels across different providers. A cleaner selection flow reduces the chance of mismatched settings, where a user might request high effort on a model that does not support it, and helps standardize the experience whether someone connects directly to a vendor or routes through ClinePass.

Alongside the new capabilities, v4.0.2 fixes a bug affecting how reasoning content is displayed. Rendering a model's intermediate thinking is technically tricky, because the agent must distinguish chain-of-thought tokens from final output and stream them appropriately. A display fix indicates that earlier versions likely showed reasoning incorrectly or inconsistently, and the patch appears intended to make that output predictable for users who rely on it to audit how the agent reached a decision.

Cline operates in a crowded field of AI coding tools, sitting alongside GitHub Copilot, Cursor, Aider, and Continue. Its open-source nature and bring-your-own-key model are common selling points, letting developers choose among providers such as Anthropic, OpenAI, Google, and open-weight options like DeepSeek and Qwen. Incremental releases like this one reflect a broader industry pattern in which agent frameworks compete less on a single model and more on how flexibly they orchestrate many models, expose tuning knobs, and manage cost. Reasoning effort has become a focal point because the newest model families increasingly trade compute for accuracy at inference time.

For practical context, reasoning effort is part of a wider shift sometimes called test-time scaling, where systems improve answers by spending more processing during a request rather than relying solely on training. Anthropic, OpenAI, and Google have all shipped models with configurable thinking budgets, so Cline aligning its interface around effort tiers is consistent with where the tooling ecosystem is heading. The xhigh addition is small in scope but signals that DeepSeek's reasoning models are being treated as first-class options.

Users on existing Cline installations should generally receive the update through the VS Code extension marketplace, and the changes are most relevant to those already running DeepSeek models or using ClinePass. As a point release, v4.0.2 is unlikely to alter overall workflows, but the combination of finer reasoning control and a display fix should make agentic sessions more transparent and easier to tune. Anyone weighing cost against output quality may want to test the new tier on representative tasks before committing it as a default.