Cline

• Direct Anthropic requests now ask the Claude API for its recommended server-side refusal fallback (anthropic.fallbacks: "default"), and OpenRouter and Cline requests for Anthropic models set provider.allow_fallbacks so a request can fail over to another upstream provider instead of failing outright
• A Stop issued while a turn is still being prepared is no longer dropped. The agent runtime ignores abort() with no run in flight, so a Stop that landed between runTurn() entry and the run start (session persistence, the git metadata refresh, credential sync) was lost and the turn kept running while the host showed it as stopped. LocalRuntimeHost now re-sends the abort once the run exists, and clears a leftover between-turns abort flag when a new turn starts so it can't cancel the next prompt
• Content-filtered turns are no longer reported as empty responses. content-filter is a new AgentModelFinishReason, mapped by both the AI SDK and ApiHandler adapters (including the compat round trip); the run fails with a message saying the response was blocked by a content filter and suggesting a rephrase, and the run-failed sdk.error event carries the turn's finish reason

• The hub now starts reliably on Windows machines with a system proxy. Bun's fetch sends loopback requests through HTTP(S)_PROXY with no localhost bypass, so on machines that export those variables (Clash, v2ray, corporate proxies) every hub discovery probe went to the proxy, the healthy daemon looked unreachable, and startup failed with "No compatible hub runtime is available". New ensureLoopbackProxyBypass() adds the loopback addresses to NO_PROXY/no_proxy whenever a proxy variable is set, and the hub client calls it before every probe, drain, and shutdown request. Compiled Bun entry detection also recognizes the Windows B:\~BUN embedded path, so Windows daemons are spawned with the --cline-hub-daemon marker instead of a dead entry path
• The hub daemon log now records one line per socket close with the close code and reason, whether the heartbeat sweep dropped it for a missed pong, and whether the server was already shutting down. When the daemon dies with EADDRINUSE, hub.daemon.startup telemetry and the daemon log now say what holds the port (PID, command line, whether the instance lock was held, and whether the occupant answers the hub health probe)

• Settings has a new About page. It shows your version and channel, has Check for updates and Restart to update buttons, and lists the release notes for recent versions with links to each GitHub release and the full changelog. Report an issue is there too
• After an update, a one-time What's new dialog now catches you up on recent features. The first one covers SSH remotes, worktrees, pull request status in the composer, and parallel sub-agents. To see it again, use Highlights on the About page
• On macOS, Help → Export Diagnostics… now opens the diagnostics export directly, so you don't have to find it in Settings
• On Amazon Bedrock, OpenAI models reached through inference profiles (us.openai.…, global.openai.…) no longer fail with "Unknown parameter: 'reasoningConfig'" when reasoning effort is set

Full Changelog: https://github.com/cline/cline/compare/desktop-v0.0.36...desktop-v0.0.37

• The app now starts on machines that set a system HTTP(S) proxy, such as Clash, v2ray, or a corporate proxy. Before, the backend sent its local connection checks through the proxy, so it couldn't find its own hub and failed with "No compatible hub runtime is available." Local connections now skip the proxy, and any proxy exemptions you already had are kept
• On small or scaled displays, the main window now fits on screen. A 1080p laptop at 150% scaling used to open the window larger than the screen, which pushed the settings and account controls out of reach. An oversized window now shrinks to fit the screen (not counting the taskbar) and opens centered
• Plugin slash commands like /goal no longer vanish from the slash menu the first time you use them. Plugin commands now load as soon as a workspace opens instead of on your first /. If that first load is slow or fails, the app retries in the background, so the command no longer gets sent to the model as plain text
• The Providers settings (renamed from API Providers) now keep the model list inside the panel. Long model lists scroll in place, and the model controls stay reachable in short windows

Added

• New provider: ai&, an OpenAI-compatible endpoint serving open-weight models from Japan.

Changed

• Refreshed the model catalog to 6,386 models across 209 providers. The resolved default model changes for 19 providers that do not pin one, 11 of them to Claude Opus 5.5 (including GitHub Copilot and Vertex). If you use one of those providers without pinning a model, expect a different default.
• Raised the minimum js-yaml version to 4.3.2 to pick up a security fix in the parser used to read rule and skill frontmatter.

Fixed

• Long replies on local models (llama.cpp, Ollama, LM Studio) that hit the output-token limit now compact the conversation and retry once instead of ending the task. These servers cap generation at whatever context is left, regardless of the output budget you set. If compaction cannot help, the existing concise-retry recovery still runs, and the partial answer is kept.
• Reopening a task that failed now shows the error with the retry option, instead of presenting it as a completed task.
• A command that prints nothing no longer shows a raw JSON blob like [{"query":"git add -A","result":"","success":true}] as its output.

• Cline Desktop now runs on Linux. Each release ships x64 .deb and .rpm packages alongside the macOS and Windows builds. Open folder… uses the native GTK picker, and updates download in the background and install when you choose Restart now, so you are never hit with a surprise password prompt. There is no AppImage for now
• Plugin slash commands now work in the desktop app. Commands a plugin registers, like /goal, used to be sent to the model as plain text; they now run the plugin's handler, show its reply, and start a turn only when the command asks for one. Enabled plugin commands appear in the slash menu, skills and workflows in that menu now come from the conversation's own workspace (including worktrees), and a broken plugin no longer makes every slash prompt fail
• Settings has a new Diagnostics row. Export… writes a single text file to your Downloads folder with the app version, OS, settings, recent sidecar and hub logs, and the manifests of the sessions you pick. API keys, credential-shaped values, your prompts, and your home directory path are stripped, so the file is safe to attach to a GitHub issue

• Long sessions on local models no longer die mid-answer at the output-token limit. llama.cpp, Ollama, and LM Studio cap generation at whatever context is left, whatever output budget you set, so a text-only reply could be cut off and fail the run. The CLI now compacts the conversation and retries that turn once before falling back to the existing concise-retry recovery, and the partial answer is kept if nothing helps
• When the hub fails to start, the error now says why instead of only "No compatible hub runtime is available". The CLI also waits up to 15 seconds for a freshly started hub instead of 8, since the first launch after an install or update can take 8 to 13 seconds on Windows
• Error messages in a session are now kept when you resume it. A failure shown during a run, including one reported after retries ran out, used to disappear once you left the session; it now reappears in the transcript on resume without being sent to the model or counted by compaction
• cline history update --title and --prompt now persist for hub-managed sessions. The new title only changed in memory and reverted on the next launch, and updating metadata could drop other keys such as pinned

• A text-only turn truncated at the output-token limit now gets one compact-and-retry attempt per run before the existing nudge-and-retry recovery. Local OpenAI-compatible servers (llama.cpp, Ollama, LM Studio) cap generation at whatever context remains regardless of the requested output budget, so long sessions died mid-answer even with sensible limits. The runtime now routes that finish through the same forced-compaction path prepareTurn uses for context-window overflow; if compaction has nothing to remove or the retry truncates again, the loop's nudge-and-retry takes over and the original max-tokens error still surfaces with the partial answer persisted. Turns that produced tool calls, or provider-executed tool activity, are never replayed. A new task.max_tokens_recovery lifecycle event (started, retried, failed) measures how often it helps
• Hub startup failures now say why. ensureCompatibleLocalHubUrl swallowed the error from ensureDetachedHubServer, so every failure read as "No compatible hub runtime is available"; the cause is now appended to that message and attached as cause, and the daemon reports its own startup failure (hub.daemon.startup) before

• Composio connectors now load all their tools, not just the first 20. Google Calendar, for example, showed only 20 of its 47 tools, and the Installed view wrongly said "20/20." The full list is now fetched and the tool cache refreshes instead of staying stale forever
• You can now connect to a Mac as an SSH remote from a Mac. Picking a Mac host used to fail with "Remote target darwin/arm64 is unsupported in SSH" even though Test passed on the same profile. The app now uses its own signed backend as the helper on both Apple Silicon and Intel Mac hosts. Windows and Linux desktops still can't connect to a Mac out of the box
• Session errors now stay in the transcript when you leave a session or open it in another client. Before, a failed run's error disappeared once you went away and came back, or opened the session in the CLI. Failures after all retries run out are now recorded too, and these error-only entries are left out of compaction
• Stopping a run while it waits to retry an empty model response now takes effect right away, instead of after the backoff finishes

Full Changelog: https://github.com/cline/cline/compare/desktop-v0.0.33...desktop-v0.0.34

Changed

• Sub-agents spawned in the same step now run their tool calls at the same time rather than one after another. Tools that must run in order still do, and the parent still waits for every result before its next turn.
• Models that advertise a large output limit now get a bigger default output budget — 30% of the limit rather than a flat 32,000 tokens, whichever is larger. Nothing changes for models under roughly 107k output tokens; longer responses can mean higher per-turn cost and latency.
• Refreshed the model catalog: 203 to 209 providers and 6,079 to 6,237 models. Kimi For Coding splits into separate kimi.com and kimi.ai providers, and AI21 Labs, ainetcafe, Inco, OCI Generative AI, Tempr, and Vispark are new. The resolved default model changes for 36 providers that do not pin one — most landing on DeepSeek V4.1 Flash, GLM 5.3 Flash, or MiMo V2.6 Flash. If you use one of those providers without pinning a model, expect a different default.

Fixed

• UserPromptSubmit and TaskStart hooks can inject context again. What those hooks returned as contextModification was being dropped — only cancel survived — so a hook meant to add repository facts or house rules to

• Start a task in its own git worktree. The welcome screen's “Work in” switch (next to the folder and branch chips) now offers Local or Worktree; pick Worktree and the first prompt of a new thread cuts a fresh cline/<id> branch off the current one, creates a worktree under ~/.cline/worktrees/, and runs the task there, so the agent never touches your working tree. The conversation chip shows <repo> / cline/<id> with the full path in its tooltip. Deleting the task removes its worktree and branch (discarding uncommitted changes in it), unless another session still lives there. Only new threads are affected — follow-ups and reopened sessions stay where they are
• Long sessions in the desktop app now compact automatically. Auto-compaction had never actually run here: the sidecar opted sessions into checkpoints but never into compaction, so the 90% trigger was never installed and long conversations simply ran out of room. This was not a recent regression — the gap dated to compaction becoming opt-in in core back in April

• Runs no longer die when a model turn hits its output-token limit before making a tool call. A reasoning-heavy or oversized response could spend the whole output allowance and end the run with an output-token-limit error; the turn is now retried up to three times with a reminder to be concise and split the work across tool calls, and only fails after that
• Models with a large advertised output limit get a bigger default output budget — 30% of the limit rather than a flat 32,000 tokens, whichever is larger. Nothing changes for models under roughly 107k output tokens. Longer responses can mean higher per-turn cost and latency on those models
• Refreshed the model catalog: 6,188 to 6,237 models across the same 209 providers. Cline Pass gains MiMo V2.6 Flash and MiMo V2.6 Pro. The default model changes for 10 providers — six land on MiMo V2.6 Flash (Xiaomi and its three token plans, Empirio Labs, NanoGPT), Eden AI moves to Grok 4.7, CoreWeave to DeepSeek V4.1 Flash, CoralBricks to DeepSeek V4.1 Flash Fast, and Tempr to Gemini 3.8 Flash. If you use one of those without pinning a model, expect a different default

• A turn that hits the model's output-token limit without producing a usable tool call is now retried instead of failing the run. Reasoning or a large response could consume the whole output allowance before the agent acted, and a single max-tokens finish ended the run with the output-token-limit error. The runtime now retries up to three consecutive times, appending a reminder to respond concisely and split large work across tool calls; the counter resets on tool-call progress and at the start of each run, and exhaustion still fails with the existing error. Empty max-tokens responses — turns that spent the allowance with no visible output — take the same recovery path, and turn-finished is emitted before each recovery iteration so iteration events stay paired
• The default output allowance now scales with the model. When neither a request limit nor a caller default is supplied and the model advertises an output limit, the gateway uses max(32000, floor(maxOutputTokens * 0.3)) instead of a flat 32,000. The fraction can only raise the default, never lower it, so it changes nothing below roughly a 106,667-token model limit (a 128,000-token model goes from 32,000 to 38,400); the

• Context compaction no longer silently degrades to truncation partway through a long session on OAuth providers. The summarizer kept the access token resolved when the session started while the main agent loop refreshed its own each turn, so once the token rotated the summarizer's request failed with 401 and the failure was swallowed by the truncation fallback — you got a chopped transcript instead of a summary, most visibly on cline-free/* models. The summarizer also stayed on the original model after a mid-session model switch. Both now follow the session's current credentials and model
• Startup no longer waits on a feature-flag network round trip. The CLI awaited a PostHog poll before the session was ready; flags now refresh in the background, and a concurrent poll no longer reads a stale default while the first request is still in flight
• Context injected by a hook is no longer mistaken for something you typed. The prompt_submit dispatch and the audit logger both fired on injected hook-context messages, which carry the user role but a system display role. Run-start hook control (cancel and context injection from agent_start/agent_resume scripts) remains inert in

• Concurrent feature-flag polls for the same account now share one in-flight request. The cache was marked fresh before its network request finished, so a caller that raced the initial poll returned immediately with the default value — a flag that was actually enabled read as disabled until something triggered a later refresh. Failures now propagate so a subsequent call can retry, and switching accounts does not wait on the previous account's request
• Run-start hooks can now inject context. beforeRun returns a new AgentRunStartResult carrying an appendContext channel alongside stop/reason — AgentStopControl stays assignable, so existing implementors compile unchanged — and the runtime injects what it collects as a single <hook_context source="RunStart"> message placed after the run's input messages, so the model sees it on the run's first request. A resumed session seeded with a trailing unresolved tool_use gets the block immediately before that call instead, keeping the tool_use/tool_result pair adjacent, and a stopping hook's context is never injected. At the hook-file and subprocess layers, agent_start and agent_resume had always spawned detached with