Cline

Cline v3.36 adds Hooks: executable scripts fired on PreToolUse, PostToolUse, UserPromptSubmit and task lifecycle events. They receive JSON on stdin and can cancel an action or inject context. Currently macOS and Linux only.

Cline races three agents running gpt-oss 120B on a DGX Spark, an RTX 4090 with heavy offload, and a cloud backend, using a kill-your-opponents script to show how memory, network path and inference stack affect agent speed.

Cline's maintainers explain their context management toolkit: a recurring todo list, automatic summarization at the window limit, deep planning that writes a plan then starts a clean task, and checkpoint-based message editing.

Cline v3.25 targets long-task degradation using research on multi-turn accuracy loss. It introduces /deep-planning, which investigates the codebase silently before asking questions, as part of interlocking systems meant to keep agents on task around turn 50.

Cline's postmortem of an unauthorized cline@2.3.0 npm publish on February 17, 2026. A compromised token added a postinstall that installed openclaw; the binary was unchanged. Covers the 8-hour window, cleanup, and new OIDC provenance.

A dense guide to self-hosting open-weight LLMs using Kimi K2.6: memory for weights and KV cache, serving and batching math, and production pricing from Cline's traffic. Aims to show where self-hosting starts saving serious money.

Cline details how it rewrote its system prompt for GLM-4.6: explicit invocation rules, strict explore-summarize-implement sequencing, and removal of generic agent guidance. Prompt length fell 57%. Notes provider-level variance when serving open models.

Argues commit history is now agent context: Cline reads recent commits to continue refactors, so descriptive messages carry intent and scope that diffs alone cannot. Contrasts 'fix' and 'final fix' with pattern-revealing messages.

Cline's usage data shows diff-edit success of 96.2% for Sonnet 4.5 and 94.9% for GLM-4.6, a gap that was 5 to 10 points three months earlier. The post pairs this with per-token pricing differences between premium and open models.

Cline rebuilt its terminal client on the same core as its IDE extensions, adding a redesigned TUI, parallel agent instances, headless operation, and ACP support for Zed, Neovim, and Emacs, plus a limited free Kimi K2.5 trial.

Walks through running Cline offline with LM Studio and Qwen3 Coder 30B, using MLX on Apple Silicon or GGUF elsewhere. Covers quantization choices and Cline's compact system prompt designed for local models.

Cline adds Jupyter notebook support that extracts structured cell context (type, source lines, metadata, execution count, outputs) before prompting the model, so edits respect notebook structure.

Cline Desktop is an early open-source app for open-weight models, with parallel sessions, scheduled tasks, a plugin and MCP marketplace, and import of tasks from other coding agents. Benchmark numbers are reused harness results, not a Desktop-specific test.

Cline v3.35 moves tool definitions out of the system prompt into native API tool calling for Claude 4+, Gemini 2.5, Grok 4 and GPT-5. It enables parallel tool execution and shrinks prompts about 15%. Also redesigns auto-approve.

Cline's hill-climbing guide: run an agent on Terminal Bench's 89 tasks through Harbor, diagnose each failure, change one variable, and keep gains. A weekend of fixes moved its score from 47% to 57%.

Cline engineers argue that multi-agent orchestration, RAG over indexed codebases and ever-longer instructions are seductive dead ends for coding agents. They favor single-threaded agents, simple search like grep, and restraint with prompts.

Cline explains how it evaluates open-weight coding agents with Harbor and Terminal Bench, found its requests 20-30% heavier than efficient harnesses, and shares a hill-climbing checklist plus traces from over a thousand dollars of runs.

A Cline engineer lists five instructions that waste tokens with coding agents, starting with vague asks like 'make it better', and explains how specific, context-rich prompts replace compiler-style terseness.

Walkthrough of building a PR code review agent on the Cline SDK. Covers a guidelines file, custom tools to post findings, guardrails blocking dangerous commands, hooks for auditing tool use, and a two-pass flow that filters false positives.

Cline's usage data on two stealth Sonoma alpha models with 2M context. Diff-edit success trails Claude 4 Sonnet, GPT-5 and Gemini 2.5 Pro, and the post advises against production use while they are experimental.

Explains writing procedures once as markdown workflows invoked with a slash command, which become a todo list in the agent's focus chain. Shows PR review and weekly dashboard examples and generating a workflow from a past session.

NVIDIA's Nemotron 3.5 Lightning, a distilled 30B MoE with 3B active parameters and up to 1M context, is available free in Cline. Reported scores: 54.3 SWE-Bench Verified and 86.2 PinchBench, with claimed throughput gains over similar open models.

Cline reports a single-prompt, 17-hour recursive self-improvement run that tuned its harness for Kimi K3 and reached 88.8% on Terminal-Bench 2.1 at roughly $50, with cost comparisons to frontier models and links to prompt and traces. Vendor-reported results.