Cline CLI v3.0.62 Adds Retry Logic, Fixes Checkpoint Stalls, Refreshes Model Catalog to 6,079 Entries

Cline shipped CLI v3.0.62 on September 15, fixing a set of reliability and performance bugs while expanding its bundled model catalog to 6,079 entries, up from 5,788.

What changed

The release addresses a run-ending bug where a single transient provider error — including a forwarded 429 from OpenRouter — would terminate a session with exit code 1. Model turns that die mid-stream are now retried up to three times with backoff. Turns that already streamed output are not retried, preventing duplication.

Streaming performance also improved. Previously, every streamed token was proxied through the Hub as a round trip carrying a full copy of the session state, with the agent loop blocking on each one. That throttling has been removed.

A separate fix targets checkpoint behavior in large repositories. Each turn was re-hashing every untracked file before calling the model, which on multi-gigabyte workspaces blocked messages for seconds to minutes. A persistent per-session snapshot index now lets git skip unchanged files.

Other fixes in this release:

• run_commands no longer hangs until timeout when a command backgrounds a child process that holds output pipes open after the parent finishes.
• Cline no longer gets OOM-killed when run from a home directory — an @ mention previously indexed and re-ranked every file under $HOME on each keystroke.
• apply_patch no longer silently overwrites existing files when a model uses "Add File" on a path that already exists.
• PowerShell commands wrapped by the model in a nested powershell -Command "..." call are no longer double-parsed, which had stripped $_ from pipelines.
• API keys pasted with invisible characters (BOM, zero-width space) are now cleaned before saving, eliminating a failure mode where corrupted keys produced indistinguishable 401 errors.

Agent Plugins are now managed through the Hub. Packages under ~/.agents/plugins/* are discovered and validated, with their skills exposed via the skills tool and MCP servers starting without touching cline_mcp_settings.json. Workspace-level .agents/plugins directories are deliberately not scanned, preventing a repo from implicitly starting its own MCP servers on open.

Scheduled runs no longer stall behind each other — parallelism limits are now enforced at claim time, and work resumed after system sleep is no longer duplicated.

The release also teases Cline Desktop, described by Cline as a native app for working with open-weight models, supporting task import from Claude Code and Codex, scheduled runs, web search, voice input, and a plugin/MCP/skills marketplace. The CLI now shows a one-time startup notice pointing to it, suppressible via CLINE_DISABLE_CLINE_PASS_NOTICE=1.

Model catalog refresh

The bundled catalog grew to 6,079 models with four new providers: Infer by Flow7, Melious, NaN, and Wallaby. According to Cline, resolved default models changed for 44 providers in this refresh, with most — including Hugging Face, Fireworks, Requesty, Nebius, Cortecs, CrossModel, DigitalOcean, Eden AI, and OpenCode Go — now defaulting to DeepSeek V4.1 Flash. Gemini and Vertex now resolve to Gemini 3.8 Flash, GitHub Copilot and Vivgrid to GPT-6 Astra, and NVIDIA to GLM 5.3 Flash. Cline warns that any provider used without a pinned model will see a different default after this update.

The release also patches a nested undici@5.29.0 dependency (CVE-2026-1525) up to the 7.x line, after an earlier remediation's override was silently ignored by Bun.

What this means

This is a maintenance release for the Cline CLI, not a new model. The most consequential change for existing users is the silent default-model reassignment across 44 providers — anyone running Cline without an explicit model pin will start hitting different models after upgrading, which can shift latency, cost, and output behavior without any visible configuration change. Teams relying on pinned defaults for reproducibility should audit their provider configs before this update lands in production workflows.