Cline

Cline proposes cline-bench, reproducible agentic-coding environments drawn from real open source tasks where models stumble, with repo snapshots, authentic problem statements and automated verification, plus a funding commitment for maintainers.

Seven days of Cline data: Claude Sonnet 4 holds 42.6% of Plan mode and 46.6% of Act mode usage. The most common cross-mode pairing is Opus 4.1 for planning and Sonnet 4 for acting. Includes OpenRouter pricing context.

Cline reports Nemotron 3.5 Super 120B scoring 61.8% on Terminal-Bench 2.1 in its harness with thinking off, ahead of three anonymized open-source harnesses. Covers plan-before-code prompting and tuning for open-weight models, with caveats on attribution.

Cline's engineering blog separates workflows, which inject instructions into one message on demand, from clinerules, which load into every system prompt. It gives a decision framework and a PR-review example showing when each saves tokens.

Cline 3.48.0 adds Skills compatibility and web search. Skills are directories with a SKILL.md that Cline loads progressively, exposing only name and description until a task matches, unlike always-on rules.

Compares Kimi K3, DeepSeek V4 Pro and Flash, GLM 5.2, and MiniMax M3 for coding on reported Terminal Bench 2.1 scores, API pricing, and total versus active parameters, with guidance on where each fits.

Cline announces Kimi K2-0905 availability across Groq, Moonshot, OpenRouter, Fireworks and Cline providers. The post covers the doubled context window, tool-calling reliability, diff-edit failure rates, Groq throughput and pricing, with a recommendation to use it in Act mode.

Cline 3.39 adds Explain Changes, which gives plain-language explanations alongside AI-generated diffs. The post argues comprehension, not authoring, is now the limiting step and that context switching drives superficial reviews.

Argues infrastructure runbooks should become executable playbooks: hardened prompts stored as .clinerules files, version-controlled with the code, composable, and runnable by engineers who did not write them.

Cline's Saoud Rizwan explains why bigger context windows do not stop coding agents from forgetting earlier decisions and introduces Focus Chain, which anchors every step to the goal over very long tasks.

Cline announces Kanban, a CLI-agnostic app that shows coding agents as live task cards so you can see which are running, blocked, or done, and chain tasks across parallel agents. Evidence reviewed is the announcement text only.

Cline open-sources its agent SDK and rebuilds its products on it. The runtime keeps the agent loop stateless while sessions persist and move across VS Code, JetBrains, and CLI, with revised prompts and context handling and published Terminal-Bench numbers.

Cline adds DeepSeek V3.2 and V3.2-Speciale. The post covers thinking during tool use with traces retained across calls, DeepSeek Sparse Attention, the agent training mix, and per-million-token pricing.

Cline CLI 2.0 rebuilds the agent for the terminal with Plan and Act modes, slash commands, mid-session model switching, headless and piped use, and parallel sessions, positioning coding agents as scriptable infrastructure.

Cline ran eight models on IMO 2026 problems with anonymized grading by two model graders and a tie-breaker. DeepSeek V4 Flash scored 30/42 against a 29 gold cutoff for $0.12, while GPT-5.6 Sol scored 42/42. Traces published; harness fixes noted.

Cline's field notes on GPT-5.1 as a coding agent: it maps whole codebases before acting, thrives on specs and phased workflows, and stays on task when a todo list returns to context every six turns.

Cline CLI preview ships as open source, built on Cline Core, a standalone service exposing a gRPC API. It supports headless scripting, multi-instance orchestration, and experimental subagents that IDE Cline can delegate to.

Summarizes AMD's testing of over 20 local models with Cline and LM Studio, finding only a few reliable for coding. Explains RAM checks, GGUF versus MLX formats, and 4-bit versus 8-bit quantization tradeoffs.

Setup guide for supervising Cline agents remotely: run Kanban with its host set to 0.0.0.0, join phone and Mac to one Tailscale tailnet, and open the board in a mobile browser to review diffs, approve changes and start tasks.

Cline explains how to use Claude Sonnet 4's 1M-token context: deeper Plan-mode sessions and longer build-test-debug cycles. It lists the price tiers above 200K tokens, the Tier 4 requirement and prompt caching as cost levers.

Cline details moving its 11-million-user VS Code extension from a monolithic private core to its shared SDK harness: a rollback after the first attempt, a homemade A/B rollout, and a reported tenfold drop in agent failures.

Cline v3.26.6 adds Grok Code Fast, a compact-prompt local setup using LM Studio and Qwen3 Coder 30B, and a Qwen Code OAuth provider with 2,000 free requests a day. It also includes GPT-5 auto-compact fixes and clearer retry feedback.

Cline's walkthrough of plugins and lifecycle hooks in its SDK-based harness. Shows how callbacks at fixed points in the agent loop add logging and block unwanted actions, and why that differs from MCP tools the model chooses to call.

Cline adds an OpenAI Codex provider using OAuth sign-in. Tokens refresh automatically, no API keys are needed, and gpt-5.2-codex and other models are available per subscription tier at flat-rate pricing.