Cline

Changed

• With no model chosen, the Cline and OpenRouter providers now use Claude Sonnet 5.5. For Cline, that's the first model in Cline's recommended list.
• Selecting ai& shows its own API key and model picker, instead of the generic OpenAI-compatible form with base URL and custom headers.
• Refreshed the model catalog. Adds Claude Haiku 5.5 and the AI/ML API, LowRouter, and Prime Inference providers. The Cline free list adds Step 5 Preview. The v0 provider now offers only its default model, v0-1.5-md, and OpenCode Go uses its new https://opencode.ai/inference/go/v1 endpoint. Default models change for 302.AI, Cortecs, CrossModel, DevPass, DigitalOcean, Eden AI, GitHub Copilot, Google Vertex AI, LLM Gateway, Meta, Moark, NanoGPT, NEAR AI Cloud, Neuralwatt, Novita AI, Ofox, OpenCode Go, Requesty, Vivgrid, and a few others (most now Claude Haiku 5.5 or Muse Wonder 1.0). If you use one of those providers without pinning a model, expect a different default.

Fixed

• Turning reasoning off no longer fails with a 400 error on GPT-6 Astra, GPT-6.1 Sol, Claude Fable 5 or Claude Opus 5.5, and now works on every provider. Level-only Gemini models no longer reject the request with a thinking-budget error.
• An MCP server or hook that exits before reading all of its input no longer crashes the extension host about 3 seconds later.
• A 404 from a stale or retired model id is no longer shown as a sign-in problem. Cline shows the real error instead of the logged-out card.
• Errors in the Cline output log keep their details. Production builds used to drop them, so a line like "Cline OAuth login failed:" showed no reason.
• Deleting a slash command from the input keeps the spacing of the text around it.
• Screenshots and other images no longer inflate the token estimate, which made auto-compaction run long before the conversation was near its limit.
• Auto-compaction uses the selected model's context window when the catalog entry has none.
• Rule files that are symlinks now load, and skills in .claude/skills are now discovered.
• The editor tool no longer receives null in place of the text to replace from some models.
• Settings migrated from older versions no longer add an empty SAP AI Core provider.

Full Changelog: https://github.com/cline/cline/compare/v4.1.23...v4.1.24

• Plugins load in-process through one registry per process instead of a node/bun subprocess, so the compiled CLI and its hub no longer need a runtime on PATH, and plugins can use the Bun global under the Bun-built hub. CLINEPLUGINMODE=sandbox keeps the subprocess sandbox. The registry wraps import, setup, hooks, tools and commands with timeouts and error capture, and tracks each plugin's state (loading/ready/degraded/failed/disabled) with its last error. A failing plugin no longer silently drops its tools: session.create accepts a plugins policy and returns pluginIssues, later failures reach affected sessions as status notices, the hub answers plugins.list/plugins.status/plugins.reload and broadcasts plugin.status_changed, and the daemon marks a plugin failed instead of shutting down when an uncaught error comes from it. Each session gets its own copy of every plugin module, so plugins that keep per-session state at module level keep working
• Turning reasoning off now works on every provider, not just a fixed list. Models whose catalog entry has no off option are sent nothing and keep their default, and models without a reasoning capability are skipped. Gemini and Vertex send exact thinkingBudget values only to models that advertise a budget; level-only Gemini models get a thinking level instead (they rejected thinkingBudget). A request never carries both the portable reasoning option and a native provider control (effort, thinking, thinkingConfig, …); the native one wins
• The cline provider's default model is the first entry of the Cline recommended-models list, falling back to anthropic/claude-sonnet-5.5. CLINEDEFAULTMODEL_ID moves from Claude Sonnet 5 to Claude Sonnet 5.5, and the openrouter provider default is now anthropic/claude-sonnet-5.5 too
• estimateRequestInputTokens and createTokenEstimator no longer bill base64 image payloads at the text rate. A ~90 KB screenshot counted as ~30k tokens, so a few images triggered auto compaction long before the transcript was near budget. Each image now adds a flat ESTIMATEDTOKENSPER_IMAGE (1600). New createImageAwareReplacer export in @cline/shared
• Auto compaction uses the resolved model's context window. When the discovered catalog entry had no limit but the selected-model override did, compaction was handed the bare catalog entry and ran against the wrong budget
• Scheduled runs: queued runs are claimed correctly under bun:sqlite, which bound the store's named parameters as NULL so the desktop hub never claimed a queued run and it stayed pending with no error. Runs whose agent turn ends with a provider error (such as a retired model id) are now recorded as failed with the error message instead of succeeding silently, and the run report keeps usage and tool calls. Queued cron occurrences superseded by a newer occurrence of the same schedule are cancelled, so a backlog no longer replays every missed slot back to back
• Rule files that are symlinks are now loaded from config directories instead of skipped. Edits to a link's target are picked up on the next refresh or new session
• resolveSkillsConfigSearchPaths scans .claude/skills, so skills authored for Claude Code are visible to Cline
• The editor tool advertises old_text as a plain string. When the schema allowed null, some models sent null instead of the replacement text. Null is still accepted at runtime and treated as omitted
• Composio connectors: signed-in sessions register Composio's meta tools (search, manage connections, wait for connections, multi-execute), so the agent can find connector tools across toolkits, hand the user a Connect Link for an app that is not connected yet, and run what it found. Connector tools are discovered through meta search; saved desktop connection state no longer supplies model tools. Hosts can supply their own authentication through the new connectorsRequest option (ConnectorsRequest exported from @cline/core), which is also forwarded to scheduled sessions. resolveComposioToolsStatePath is renamed to resolveComposioConnectionStatePath
• The Cline cloud sign-in helpers in @cline/core accept an AbortSignal, so hosts can cancel a pending sign-in while it is polling
• Cloud session refreshes update open sessions with titles renamed on the web, and keep renames made locally while a refresh is in flight
• Refreshed the model catalog (6,572 → 6,862 models). Adds the AI/ML API, LowRouter, and Prime Inference providers. models.dev no longer lists v0, so the v0 provider ships with only its default model, v0-1.5-md. OpenCode Go's base URL moves from https://opencode.ai/zen/go/v1 to https://opencode.ai/inference/go/v1. The Cline free list adds Step 5 Preview. Default models change for 302.AI, CrossModel, Cortecs, DigitalOcean, NEAR AI Cloud, and Ofox (Claude Haiku 5.5), DevPass (LLM Gateway), LLM Gateway, and Meta (Muse Wonder 1.0), Aixy (GPT-6.1 Sol), CloudFerro Sherlock and Moark (GLM 5.3), Ambient, NanoGPT, Neuralwatt, and Novita AI

Full Changelog: https://github.com/cline/cline/compare/sdk/sdk/v0.0.92...sdk/sdk/v0.0.93

• A new outline along the left edge of the conversation lists each prompt you sent. Hover a mark to preview the prompt, and click it to jump there
• Plugins that fail to load or error during a session now show a status line in the chat naming the plugin and the error, instead of silently losing their tools
• You can cancel a Cline Cloud sign-in while it is waiting for the browser
• Session titles renamed on the web now update in the app
• The Cloud setup card stays visible on smaller windows when signed out
• Installing a skill without Git installed now says Git is required and how to fix it, instead of showing a raw clone error
• Scheduled routines: queued runs now start instead of staying pending forever, a run whose model call fails (for example, a retired model) is marked failed with the error, and after downtime only the latest missed occurrence runs instead of every one back to back. The routine editor now lists current models instead of retired ones
• With no model chosen, the Cline provider uses the first model in Cline's recommended list (Claude Sonnet 5.5)
• Turning reasoning off now works on every provider. Models that can't turn it off keep their default, and level-only Gemini models no longer reject the request with a thinking-budget error
• Screenshots and other images no longer inflate the token estimate, which made auto-compaction run long before the conversation was near its limit
• Auto-compaction uses the selected model's context window when the catalog entry has none
• Rule files that are symlinks now load
• Skills in .claude/skills are now discovered
• The editor tool no longer receives null in place of the text to replace from some models
• Refreshed the model catalog. Adds the AI/ML API, LowRouter, and Prime Inference providers. The v0 provider now offers only its default model, v0-1.5-md, and OpenCode Go uses its new https://opencode.ai/inference/go/v1 endpoint. The Cline free list adds Step 5 Preview. Default models change for 302.AI, CrossModel, Cortecs, DigitalOcean, NEAR AI Cloud, and Ofox (Claude Haiku 5.5), DevPass (LLM Gateway), LLM Gateway, and Meta (Muse Wonder 1.0), Aixy (GPT-6.1 Sol), CloudFerro Sherlock and Moark (GLM 5.3), Ambient, NanoGPT, Neuralwatt, and Novita AI

Full Changelog: https://github.com/cline/cline/compare/desktop-v0.0.45...desktop-v0.0.46

• Plugins now load inside the hub instead of a separate node or bun process, so the CLI no longer needs a JavaScript runtime on your PATH to run plugins. A plugin that fails to load no longer silently loses its tools: Cline prints a warning naming the plugin and the error, cline config tools --json lists it as a pluginfailure entry, and the new cline hub plugins shows each plugin's status and last error (--reload <plugin> re-imports one without restarting the hub). Set CLINEPLUGIN_MODE=sandbox to keep the old subprocess behavior
• --thinking now applies in ACP mode (--acp). Before, ACP sessions always ran with reasoning off
• cline hook accepts agent_error events. Before, it exited with an error and no response
• A provider API key set in the environment (such as OPENROUTERAPIKEY) now counts as configured at startup, for providers that read a single environment variable. Before, the TUI asked you to set up the provider anyway
• With no model chosen, the Cline provider now uses the first model in Cline's recommended list (Claude Sonnet 5.5)
• Turning reasoning off now works on every provider. Models that can't turn it off keep their default, and level-only Gemini models no longer reject the request with a thinking-budget error
• Screenshots and other images no longer inflate the token estimate, which made auto-compaction run long before the conversation was near its limit
• Auto-compaction uses the selected model's context window when the catalog entry has none
• Scheduled runs: queued runs now start instead of staying pending forever, a run whose model call fails (for example, a retired model) is marked failed with the error, and after downtime only the latest missed occurrence of a schedule runs instead of every one back to back
• Rule files that are symlinks now load
• Skills in .claude/skills are now discovered
• The editor tool no longer receives null in place of the text to replace from some models
• Refreshed the model catalog. Adds the AI/ML API, LowRouter, and Prime Inference providers. The v0 provider now offers only its default model, v0-1.5-md, and OpenCode Go uses its new https://opencode.ai/inference/go/v1 endpoint. The Cline free list adds Step 5 Preview. Default models change for 302.AI, CrossModel, Cortecs, DigitalOcean, NEAR AI Cloud, and Ofox (Claude Haiku 5.5), DevPass (LLM Gateway), LLM Gateway, and Meta (Muse Wonder 1.0), Aixy (GPT-6.1 Sol), CloudFerro Sherlock and Moark (GLM 5.3), Ambient, NanoGPT, Neuralwatt, and Novita AI

Full Changelog: https://github.com/cline/cline/compare/cli-v3.0.70...cli-v3.0.71

• The AI SDK moves together to one release line: ai 7.0.120 and every @ai-sdk/* package on @ai-sdk/provider-utils 5.0.50 (previously ai 7.0.79 with packages bumped one by one). Turning reasoning off no longer sends reasoning.effort: "none" to OpenAI models that can't turn it off, such as GPT-6 Astra and GPT-6.1 Sol (GPT-6 Astra answered with a 400, for example from commit message generation); GPT-5.x and GPT-6 Sol/Luna still get "none". Turning reasoning off also no longer sends thinking: {type: "disabled"} to Claude models that can't turn thinking off, such as Claude Fable 5 and Opus 5.5 (they answered with "thinking.type.disabled" is not supported for this model); other Claude models still get disabled
• Stdio MCP clients handle EPIPE when a server exits before reading all of its input. Previously that raised an uncaught exception on the host process. The SAP AI Core provider also no longer installs winston's process-wide uncaughtException handler, which called process.exit(1) three seconds after any uncaught error
• The legacy provider-settings migration no longer adds an empty sapaicore entry to providers.json. It used to do this for every migrated user, because the extension writes sapAiCoreUseOrchestrationMode as a default
• @cline/core exports parseMcpHeaders, used by the CLI's mcp add wizard to keep --header values
• Refreshed the model catalog. Adds Claude Haiku 5.5. Default models change for Google Vertex AI (Claude Sonnet 5.5 → Claude Haiku 5.5), Cortecs, DevPass (LLM Gateway), Eden AI, GitHub Copilot, LLM Gateway, NanoGPT, OpenCode Go, Requesty, and Vivgrid (most now Claude Haiku 5.5)

Full Changelog: https://github.com/cline/cline/compare/sdk/sdk/v0.0.91...sdk/sdk/v0.0.92

• SSH remotes start again. Since 0.0.38, connecting to an SSH environment failed with SyntaxError: Invalid character
• On SSH environments, the git branch label no longer opens two new SSH logins every 5 seconds. It checks every 30 seconds now, and switching branches from the picker still updates it right away
• Long model names in the composer show in full when the row has room, and only truncate when it runs out
• Turning reasoning off no longer fails with a 400 error on GPT-6 Astra, GPT-6.1 Sol, Claude Fable 5 or Claude Opus 5.5
• The background service no longer crashes about 3 seconds after an MCP server or hook quits before reading all of its input
• Settings migrated from the VS Code extension no longer add an empty SAP AI Core provider
• Refreshed the model catalog. Adds Claude Haiku 5.5. Default models change for Google Vertex AI (Claude Sonnet 5.5 → Claude Haiku 5.5), Cortecs, DevPass (LLM Gateway), Eden AI, GitHub Copilot, LLM Gateway, NanoGPT, OpenCode Go, Requesty, and Vivgrid (most now Claude Haiku 5.5)

Full Changelog: https://github.com/cline/cline/compare/desktop-v0.0.44...desktop-v0.0.45

• Commit message generation and tasks with reasoning off no longer fail with a 400 error on GPT-6 Astra, GPT-6.1 Sol, Claude Fable 5 or Claude Opus 5.5
• Cline no longer exits about 3 seconds after an MCP server or hook quits before reading all of its input
• Settings migrated from the VS Code extension no longer add an empty SAP AI Core provider
• cline mcp add keeps your --header values when it opens the setup wizard. Before, accepting the prefilled answers saved a remote server without its headers (for example, Authorization). The wizard now asks for one header per prompt and validates each one the same way as --yes
• cline dashboard is now cline hub dashboard. The old command still works as a hidden alias
• A bare cline hub prints its subcommands
• cline doctor log opens the file set by CLINELOGPATH when that variable is set
• cline --help is corrected: --retries shows its real default of 3 (it said 6), and --thinking explains that it falls back to the level picked in /model
• A bare --retries with no count is now rejected with an error instead of crashing
• Options that were accepted but never did anything are removed and now fail with "unknown option": cline history --page, and cline schedule create/update --autonomous, --no-autonomous, --idle-timeout, --poll-interval and --delivery-channel. Remove them from scripts. Existing schedules saved with them still load
• cline schedule create, update and import reject a delivery that could never be posted, such as --delivery-adapter without --delivery-thread or an adapter that isn't a known connector. Before, the schedule was saved and its results were silently never posted. Errors name the flag or --metadata-json path to fix
• Refreshed the model catalog. Adds Claude Haiku 5.5. Default models change for Google Vertex AI (Claude Sonnet 5.5 → Claude Haiku 5.5), Cortecs, DevPass (LLM Gateway), Eden AI, GitHub Copilot, LLM Gateway, NanoGPT, OpenCode Go, Requesty, and Vivgrid (most now Claude Haiku 5.5)

Full Changelog: https://github.com/cline/cline/compare/cli-v3.0.69...cli-v3.0.70

Changed

• If a model's response ends without a recognized finish reason, Cline asks it to continue once instead of treating the response as complete.
• Refreshed the model catalog. The Cline recommended list adds GPT-6.1 Sol, and the free list adds Solar Mini 4 and drops DeepSeek V4.1 Flash and space-bunny-alpha. Default models change for 302.AI, AKI.IO, Blue Claw, CoralBricks, CrossModel, DevPass, DigitalOcean, GMI Cloud, LLM Gateway, Mistral, NanoGPT, Neon, Nvidia, Ofox, Requesty, Subconscious, The Grid AI, Vivgrid, and Vultr. Vultr's model ids were renamed upstream, so a pinned Vultr model may need to be re-selected. If you use one of those providers without pinning a model, expect a different default.

Fixed

• Generate Commit Message now follows your .clinerules, so conventions like language, format, and ticket references apply to commit messages too.
• Claude through a custom Anthropic base URL (Azure AI Foundry, corporate gateways) no longer fails with a 400 error. 4.1.22 started sending an option only Anthropic's own API accepts.
• Kimi K3 and other models that only accept certain reasoning levels no longer reject requests. Cline picks the closest level the model supports.
• When an MCP tool returns more output than fits in context, Cline can now read the rest. It gets a preview plus a link it can page through with read_files, so the output past the cutoff is no longer lost.
• Approve and Reject on a tool approval work right away, even while the panel is still updating. A late update failure no longer rejects a later, unrelated approval.
• Old tasks no longer show leftover Focus Chain checklist rows when reopened.

Full Changelog: https://github.com/cline/cline/compare/v4.1.22...v4.1.23

• Responses that end with a missing or unrecognized finish reason are no longer treated as successful completions. AgentModelFinishReason gains unknown (the AI SDK adapter maps unified other and missing reasons to it instead of stop). Without tool activity, the agent keeps the partial response and continues once with a hidden user message; a second unknown finish fails the run. Queued user messages are now consumed at every request boundary, including the first iteration
• Stdio MCP servers get a 10s default initialize budget (was 3s). npx/uvx launchers through cmd.exe on Windows routinely took 3-6s and were silently dropped from the session. An explicit timeout still overrides it
• The Anthropic provider sends the server-side refusal fallbacks option only to api.anthropic.com. Custom endpoints such as Azure AI Foundry rejected it with a 400. New isOfficialAnthropicEndpoint helper
• The portable reasoning level for the cline and openai-compatible adapters snaps to the model's advertised effort levels. Kimi K3 (low/high/max) was sent medium and minimal, which it rejects
• OpenAI-compatible providers (generic, vercel-ai-gateway, user-defined ids) write providerOptions under the camelCase alias only, which stops the per-chunk providerOptions key 'openai-compatible' deprecation warning from @ai-sdk/openai-compatible
• New apiKeyOptional provider fact (GatewayProviderMetadata.apiKeyOptional, resolveProviderApiKeyOptional, ProviderListItem.auth.apiKeyOptional), declared on openai-compatible, ollama, lmstudio, litellm, vertex, bedrock, and user-added endpoints from providers.json/models.json
• Rules: new listEnabledRulesFromRecords / loadRulesForSystemPromptFromRecords exports. The combined user-instruction service now orders rules by name, matching the single-workspace path
• Opt-in Langfuse tracing for BYOK providers (CLINELANGFUSEALLPROVIDERS), plus operator tags, metadata and environment via CLINELANGFUSETAGS, CLINELANGFUSEMETADATA, LANGFUSETRACING_ENVIRONMENT
• Hook discovery telemetry handles Windows path separators
• Refreshed the model catalog. The Cline free list adds Solar Mini 4 and drops DeepSeek V4.1 Flash and space-bunny-alpha. Adds the Engy provider. Default models change for AKI.IO, Blue Claw, CoralBricks, CrossModel (GPT-6.1 Sol), DevPass, LLM Gateway, Mistral, NanoGPT, Requesty (Mistral Large 4), Neon (Claude Opus 5.5), Subconscious, and The Grid AI

Full Changelog: https://github.com/cline/cline/compare/sdk/sdk/v0.0.90...sdk/sdk/v0.0.91

• Mermaid diagrams in chat now render inline as interactive diagrams, with copy, download, fullscreen, and pan/zoom. Links inside a diagram go through the same confirmation dialog as other links, showing the real destination before anything opens
• If the Cline Hub connection drops mid-chat, the app now reconnects your session automatically for up to a minute instead of failing right away. Queued messages are sent once the session is back. If it can't reconnect, the turn ends with a notice so you can resend
• Local and self-hosted providers that don't need an API key (LM Studio, Ollama, vLLM, LiteLLM, and custom OpenAI-compatible endpoints) can start sessions again. They failed every turn with "Missing API key"
• MCP servers launched with npx or uvx on Windows now load. They often took longer than the 3-second startup limit and were silently dropped. The default is now 10 seconds
• Claude through a custom Anthropic base URL (Azure AI Foundry, corporate gateways) no longer fails with a 400 error
• Kimi K3 and other models that only accept certain reasoning levels no longer reject requests. The app picks the closest level the model supports
• If a model's response ends without a recognized finish reason, the agent asks it to continue once instead of treating the response as complete
• Refreshed the model catalog. The Cline free list adds Solar Mini 4 and drops DeepSeek V4.1 Flash and space-bunny-alpha. Default models change for AKI.IO, Blue Claw, CoralBricks, CrossModel, DevPass, LLM Gateway, Mistral, NanoGPT, Requesty, Neon, Subconscious, and The Grid AI

Full Changelog: https://github.com/cline/cline/compare/desktop-v0.0.43...desktop-v0.0.44

• MCP servers launched with npx or uvx on Windows now load. They often took longer than the 3-second startup limit and were silently dropped, so their tools never showed up. The default is now 10 seconds
• Claude through a custom Anthropic base URL (Azure AI Foundry, corporate gateways) no longer fails with a 400 error
• Kimi K3 and other models that only accept certain reasoning levels no longer reject requests. The CLI picks the closest level the model supports
• OpenAI-compatible providers no longer print a providerOptions key 'openai-compatible' deprecation warning on every response
• If a model's response ends without a recognized finish reason, the agent asks it to continue once instead of treating the response as complete
• cline config --json prints JSON again. It opened the interactive view instead, which failed outside a terminal; cline config --json mcp printed plain text
• -y/--yolo is now listed in cline --help, with a warning to use it only in sandboxed environments

Full Changelog: https://github.com/cline/cline/compare/cli-v3.0.68...cli-v3.0.69

• Agent teams no longer slow down as a session runs or while a teammate is streaming. Every streamed chunk and 2-second heartbeat used to re-save the whole team state, including every finished teammate's full transcript (one local teams.db reached 1.66 GB, with a single run row rewritten ~339k times). Stream chunks and heartbeats now go to live UIs only and are never persisted, only changed entities are written (batched into one transaction every ~300 ms), run records keep a summary instead of the transcript, and team_events is capped per team (2000 rows, 30 days). SQLite team storage moves to schema v2 with a one-time migration that compacts existing data; SqliteTeamStore.vacuum() returns the freed space when called explicitly. Failed team writes are retried instead of dropped
• Standalone provider requests made outside a session (such as commit message generation) can now resolve Cline surface headers. sessionId is optional in resolveProviderRequestHeaders, and X-Task-ID is omitted rather than sent empty
• Refreshed the model catalog. Default models change for DigitalOcean (GPT-6.1 Sol), GMI Cloud (Qwen 3.8 Flash), NanoGPT (Pareto 26.10 Preview), Nvidia (DeepSeek V4.1 Flash), and Ofox (GPT-6.1 Sol)

Full Changelog: https://github.com/cline/cline/compare/sdk/sdk/v0.0.89...sdk/sdk/v0.0.90

• New /compact slash command. It summarizes the conversation so far to free up context. The summary uses your session's provider settings and is not sent to the model as a prompt
• Connectors is now the first and default tab in Customize. A new Browse section lists the full connector catalog in two columns with search, so you no longer need the Marketplace to find a connector. Installed and Browse each have their own search box
• Connectors now show their brand logo, both in Customize and in Marketplace rows
• If connectors aren't available to you, the Connectors tab now says why: a sign-in card when you're signed out, or a rollout notice with Check again when your account doesn't have beta access yet

Full Changelog: https://github.com/cline/cline/compare/desktop-v0.0.41...desktop-v0.0.42

• Cloud sessions can now use the same models as Cline on the web. The model picker lists Cline Usage-Billing, ClinePass, and ClineFree models (organization accounts don't get ClinePass), instead of only Cline. A new Cloud session waits until a model is ready before sending, and you can retry if the model list fails to load
• An unsent prompt is no longer lost when you switch to another session and back. New in the sidebar returns to your unfinished draft; Cmd/Ctrl+N still starts a fresh session
• Sessions you start in the app are now highlighted in the sidebar while open, instead of only sessions reopened from history. They also no longer get marked unread by their own messages
• Customize → Connectors now matches the Skills, Plugins, and MCP tabs, with search, a refresh button, and an Installed section. A new Suggested section offers ready-made connector combos (such as incident root-cause analysis or organizing your day), each with an example prompt and one-click installs for its connectors
• Marketplace rows now show what type each item is, so same-named entries like the Figma MCP server and the Figma connector are easy to tell apart. Category filters moved into a single dropdown
• A connector's details now show how many tools it adds to new sessions, instead of the catalog's total tool count
• Agent teams stay fast in long sessions and while teammates are streaming. Existing team data is compacted automatically the first time it's opened
• Refreshed the model catalog. Default models change for DigitalOcean (GPT-6.1 Sol), GMI Cloud (Qwen 3.8 Flash), NanoGPT (Pareto 26.10 Preview), Nvidia (DeepSeek V4.1 Flash), and Ofox (GPT-6.1 Sol)

Full Changelog: https://github.com/cline/cline/compare/desktop-v0.0.40...desktop-v0.0.41

Added

• New providers: Bee (by HEOSSI) and Pareto Inference (PARETOAPIKEY).

Changed

• Direct Anthropic requests now use Anthropic's server-side refusal fallback. On OpenRouter and Cline, Anthropic models can fail over to another upstream provider instead of failing the request.
• When a response is blocked by a content filter, Cline now says so and suggests rephrasing, instead of "Model returned empty response".
• Refreshed the model catalog. GPT-6.1 Sol becomes the default model for OpenAI, OpenRouter, GitHub Copilot, Cortecs, Eden AI, Kilo Gateway, both LLM Gateway providers, NanoGPT, OpenCode Zen, and Requesty. Vercel AI Gateway moves to Ling 3.1 Flash, Tempr Gateway to MiMo V2.6 Flash, CrossModel and Ofox to Claude Sonnet 5.5, Pioneer to GLiNER 2.5 Decide, and Scaleway to Qwen 3.8 27B. The Cline recommended list adds Claude Sonnet 5.5 and Claude Opus 5.5. If you use one of those providers without pinning a model, expect a different default.

Fixed

• After an API error you can type and send a new message instead of only using Retry. A /compact (or /smol, /newtask) typed during recovery stays in the composer instead of being sent to the model as text.
• If a new task fails Cline sign-in, signing in and submitting a revised prompt now starts the task with that prompt. Before, nothing happened.
• Reset Code is now disabled when a message has no checkpoint, and no longer stays disabled for a message whose checkpoint exists (seen on slow disks). Changing the Checkpoints setting now applies to the active task, and follow-up messages typed while it applies are kept in order.
• Canceling right after sending a message now stops the task. A cancel that landed while the turn was still being set up was ignored, and the task kept running.
• Reasoning tokens are no longer counted twice in token usage. Cost is unchanged.
• Amazon Bedrock: GPT-6 and GPT-5.6 route through inference profiles, and India regions (ap-south-1, ap-south-2) resolve the in. profile. OpenAI models behind inference profiles get the reasoning effort setting they support. Nova 2 Lite with high reasoning, application inference-profile ARNs, and Nova Micro no longer get requests Bedrock rejects. A legacy AWS profile setting saved without the "use profile" flag now carries over as profile authentication instead of being dropped.
• Gateway models on providers that mix endpoints keep their own API protocol instead of falling back to the provider-wide default.

Full Changelog: https://github.com/cline/cline/compare/v4.1.21...v4.1.22

• Oversized MCP and Composio tool results can now be recovered in full. Core caches the oversized output in a per-session in-memory cache and sends the model a bounded preview plus a cline://cache/... URI that read_files can page through by line range. Custom tools can opt in with resultPolicy: "cache-oversized" on createTool. Entries expire after five model iterations without a read, the cache is capped at 16 MiB per session, and original output stays in history and tool events
• Custom providers from providers.json/models.json now run on the agent path. They appeared in pickers but failed with Unknown or disabled provider "<id>", because createAgentModelFromConfig only registered builtin providers with the gateway. @cline/llms now exports resolveGatewayProviderRegistration(Sync)
• Provider settings saves are more robust. Settings still save when model discovery fails, mutations are serialized, a failed catalog write restores the previous settings, and persistence errors propagate to the caller (saveLocalProviderSettings is now async). Model source requests are authenticated, and source catalogs refresh when credentials or endpoints change
• Refreshed the model catalog. The Cline recommended list adds GPT-6.1 Sol and drops the Pixel Canary stealth model. Default models change for 302.AI (Claude Sonnet 5.5), NanoGPT (Ling 3.1 Flash), Vivgrid (GPT-6.1 Sol), and Vultr (MiMo V2.6 Flash RL). Vultr's model ids were renamed upstream (for example zai-org/GLM-5.2-FP8 is now glm-5.2), so a pinned Vultr model may need to be re-selected

Full Changelog: https://github.com/cline/cline/compare/sdk/sdk/v0.0.88...sdk/sdk/v0.0.89

• 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
• Resuming an existing session works again. 0.0.87 added a duplicate-start guard that rejected starting an existing session ID with "session already exists", which blocked desktop and TUI resume; the guard is reverted, and the cloud controller keeps initial thinking and reasoning-effort preferences in session metadata instead
• NodeHubClient registers { version, pid } metadata (version defaults to the @cline/core version, overridable with clientVersion). hub.client.registered includes it, and the Cline Hub web app shows it for connected clients and in join/leave events
• @ai-sdk/amazon-bedrock moves to ^5.0.98, so Nova 2 Lite with high reasoning effort no longer returns a 400
• Refreshed the model catalog. GPT-6.1 Sol becomes the default model for OpenAI, OpenRouter, GitHub Copilot, Cortecs, Eden AI, Kilo Gateway, both LLM Gateway providers, NanoGPT, OpenCode Zen, and Requesty. Vercel AI Gateway moves to Ling 3.1 Flash, Tempr Gateway to MiMo V2.6 Flash, CrossModel and Ofox to Claude Sonnet 5.5, Pioneer to GLiNER 2.5 Decide, and Scaleway to Qwen 3.8 27B. New models include Grok 4.7 on Bedrock, DeepSeek V4.1 Flash on Nebius, and Qwen 3.8 27B on Cerebras and Groq; the Claude Opus 4 variants are dropped

Full Changelog: https://github.com/cline/cline/compare/sdk/sdk/v0.0.87...sdk/sdk/v0.0.88

• Custom providers added with Add Provider now work when you run a task. They showed up in the provider and model pickers but failed with Unknown or disabled provider
• New chats remember whether you last used Cloud or Local, and which Cloud model you picked, instead of always going back to Local. Switching a thread from Local to Cloud also opens on your remembered Cloud model
• Saving provider credentials no longer fails when the provider's model list can't be fetched, and the model list refreshes when you change a provider's API key or endpoint
• MCP settings always use the same file. With CLINEMCPSETTINGSPATH, CLINEDATADIR, or CLINEDIR set, reading and saving servers, MCP OAuth sign-in, and Open MCP settings now all use that path, where some could open a different file before
• Tool diffs now follow the app's font size setting instead of a fixed 13px
• When an MCP tool returns more output than fits in context, the agent can now read the rest. It gets a preview plus a link it can page through, so the output past the cutoff is no longer lost
• Refreshed the model catalog. The Cline recommended list adds GPT-6.1 Sol. Default models change for 302.AI (Claude Sonnet 5.5), NanoGPT (Ling 3.1 Flash), Vivgrid (GPT-6.1 Sol), and Vultr (MiMo V2.6 Flash RL). Vultr's model ids were renamed upstream, so a pinned Vultr model may need to be re-selected

Full Changelog: https://github.com/cline/cline/compare/desktop-v0.0.39...desktop-v0.0.40

• Linux on ARM is now supported: each release ships arm64 .deb and .rpm packages alongside the x64 ones, and they auto-update like the others
• Web search settings moved from General to Customize → Tools, next to the other built-in tools. That page shows whether your provider supports search, and links to provider settings when it doesn't
• Connectors are labeled Beta, and opening the marketplace from Connectors now shows only connectors
• Pressing Stop right after sending a prompt now stops the turn. Before, a stop that landed while the turn was still being set up was ignored: the agent kept editing in the background, and reverting failed with "Wait for all turns in this workspace to finish before restoring it"
• Inserting a slash command no longer deletes the text after your cursor
• On Windows, Export Diagnostics no longer fails when the output folder already exists
• Cloud agent sessions now keep their workspace. A suspended cloud session resumes with the same files, history, and reasoning settings
• When a response is blocked by a content filter, the app now says so and suggests rephrasing, instead of "Model returned empty response"
• Direct Anthropic requests now use Anthropic's refusal fallback. On OpenRouter and Cline, Anthropic models can fail over to another upstream provider instead of failing the request
• Reasoning tokens are no longer counted twice in token usage. Cost is unchanged
• On Amazon Bedrock, Nova 2 Lite with high reasoning, application inference-profile ARNs, and Nova Micro no longer get requests Bedrock rejects, and a legacy bare awsProfile setting now migrates as profile auth
• Gateway models keep their own API protocol instead of falling back to the provider-wide default
• Scheduled task specs saved with a UTF-8 BOM (as Windows Notepad does) now parse, and codebase search now includes PHP projects (Composer's vendor folder is skipped)
• On Windows, a nested pwsh -Command keeps your configured shell path instead of picking up a powershell.exe from the workspace
• Refreshed the model catalog. GPT-6.1 Sol becomes the default model for OpenAI, OpenRouter, GitHub Copilot, Cortecs, Eden AI, Kilo Gateway, both LLM Gateway providers, NanoGPT, OpenCode Zen, and Requesty. Vercel AI Gateway moves to Ling 3.1 Flash, Tempr Gateway to MiMo V2.6 Flash, CrossModel and Ofox to Claude Sonnet 5.5, Pioneer to GLiNER 2.5 Decide, and Scaleway to Qwen 3.8 27B. New providers are Bee and Pareto

Full Changelog: https://github.com/cline/cline/compare/desktop-v0.0.37...desktop-v0.0.38

• When an MCP tool returns more output than fits in context, the agent can now read the rest. It gets a preview plus a link it can page through with read_files, so the output past the cutoff is no longer lost
• Custom providers defined in providers.json/models.json now work when you run a task. They showed up in the provider and model pickers but failed with Unknown or disabled provider
• If saving provider credentials fails during onboarding or in the provider settings dialog, the error now shows in place instead of failing silently
• On Linux, the auto-approve indicator in the status bar uses a glyph that common monospace fonts include, instead of rendering as a blank box or nothing
• Refreshed the model catalog. The Cline recommended list adds GPT-6.1 Sol. Default models change for 302.AI (Claude Sonnet 5.5), NanoGPT (Ling 3.1 Flash), Vivgrid (GPT-6.1 Sol), and Vultr (MiMo V2.6 Flash RL). Vultr's model ids were renamed upstream, so a pinned Vultr model may need to be re-selected

Full Changelog: https://github.com/cline/cline/compare/cli-v3.0.66...cli-v3.0.67

• The CLI binaries are now built with Bun 1.4.2. macOS 27 no longer kills them at launch, and the x64 builds run on CPUs without AVX2
• The hub now starts on Windows machines with a system proxy (Clash, v2ray, corporate proxies). Loopback discovery requests were sent through the proxy, so a healthy hub looked unreachable and startup failed with "No compatible hub runtime is available"
• Pressing Esc right after sending a prompt now stops the turn. A stop that landed while the turn was still being set up was ignored, and the agent kept working in the background
• When a response is blocked by a content filter, the CLI now says so and suggests rephrasing, instead of "Model returned empty response" (which implied a retry would help)
• Direct Anthropic requests now use Anthropic's server-side refusal fallback. On OpenRouter and Cline, Anthropic models can fail over to another upstream provider instead of failing the request
• Reasoning tokens are no longer counted twice in output token totals. Cost is unchanged, since reasoning is billed at the output rate
• Bedrock fixes. GPT-6 and GPT-5.6 route through inference profiles, India regions (ap-south-1, ap-south-2) resolve the in. profile, and Nova 2 Lite with high reasoning, application inference-profile ARNs, and Nova Micro no longer get requests Bedrock rejects. A legacy bare awsProfile setting now migrates as Bedrock profile auth
• Gateway models on providers that mix endpoints keep their own API protocol instead of falling back to the provider-wide default
• Yolo mode (-y) no longer includes plan/act mode instructions in the system prompt
• Cron (.cline/cron/.md) and task (.task.md) specs saved with a UTF-8 BOM, as Windows Notepad does, now parse
• searchcodebase now includes PHP sources and templates (php, phtml, inc, twig), and Composer's vendor directory is excluded like nodemodules
• On Windows, a nested pwsh -Command wrapper keeps your configured shell path instead of resolving a bare powershell.exe, which could pick up an executable planted in the workspace
• Hub diagnostics. The hub daemon log records why each socket closed, a daemon that dies with EADDRINUSE reports what holds the port, and the hub web app shows each connected client's version and PID
• Refreshed the model catalog. New providers are Bee and Pareto (PARETOAPIKEY), and the Cline recommended list adds Claude Sonnet 5.5 and Claude Opus 5.5. GPT-6.1 Sol becomes the default model for OpenAI, OpenRouter, GitHub Copilot, Cortecs, Eden AI, Kilo Gateway, both LLM Gateway providers, NanoGPT, OpenCode Zen, and Requesty. Vercel AI Gateway moves to Ling 3.1 Flash, Tempr Gateway to MiMo V2.6 Flash, CrossModel and Ofox to Claude Sonnet 5.5, Pioneer to GLiNER 2.5 Decide, and Scaleway to Qwen 3.8 27B. If you use one of those without pinning a model, expect a different default

Full Changelog: https://github.com/cline/cline/compare/cli-v3.0.65...cli-v3.0.66

• 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 NOPROXY/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)
• Token usage no longer counts reasoning tokens twice. Providers report reasoning tokens as a subset of outputTokens, but normalizeUsage() returned the raw total alongside reasoningTokenCount, so session totals and telemetry counted them in both. outputTokens now excludes reasoning tokens, including for the AI SDK v4 outputTokens.reasoning shape. Cost is still computed from the full total, since reasoning is billed at the output rate
• Gateway models on providers that mix endpoints keep their per-model apiProtocol (responses, messages, gemini, or chat/completions). Configured models replace catalog entries and dropped the field, so those models fell back to the provider-wide default API
• Bedrock fixes. Bare OpenAI GPT-6 and GPT-5.6 ids now route through inference profiles like Claude, Nova, and DeepSeek already did, since Bedrock only serves them that way; gpt-oss stays on-demand, and India regions (ap-south-1, ap-south-2) now resolve the in. geo profile. @ai-sdk/amazon-bedrock moves to ^5.0.96: OpenAI models behind inference profiles get reasoning.effort instead of the rejected reasoningConfig, and models the adapter can't classify (application inference-profile ARNs, Nova Micro) no longer receive a reasoning config that Bedrock rejects with "Extra inputs are not permitted"
• Provider-settings migration now treats a bare legacy awsProfile (set with neither awsAuthentication: "profile" nor awsUseProfile) as Bedrock profile auth, instead of writing a Bedrock entry with no profile
• The yolo-mode system prompt no longer includes the plan/act mode instructions, since yolo sessions don't use that workflow. Plan and act prompts are unchanged
• Cron (.cline/cron/.md) and task (.task.md) specs saved as UTF-8 with BOM, as Windows Notepad does, now parse. Before, the BOM hid the opening frontmatter fence and the spec was rejected
• The fallback code search now includes PHP sources and templates (php, phtml, inc, twig), so Laravel and other PHP projects show up in searchcodebase. Composer's vendor directory is excluded from search and the workspace file index, matching nodemodules
• LocalRuntimeHost.updateSession returns the persisted metadata and refreshes the active session's in-memory metadata after an update, including a title change, so hosts no longer read stale metadata until the session is reloaded
• The ./cloud export gains shared cloud handoff support (moving a local session to a cloud agent and back) and persistent cloud workspaces that can be suspended and resumed. Saved reasoning settings survive a runtime restart and a resume, and automatic Git backups only run for ephemeral sandboxes
• Refreshed the model catalog: 209 to 211 providers and 6,386 to 6,447 models. New providers are bee (Bee by HEOSSI) and pareto (Pareto Inference, reads PARETOAPIKEY). The Cline recommended list adds Claude Sonnet 5.5, Claude Opus 5.5, and the stealth/pixel-canary stealth model, and drops Claude Opus 5

Full Changelog: https://github.com/cline/cline/compare/sdk/sdk/v0.0.86...sdk/sdk/v0.0.87

• 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
• In SSH settings, Save is disabled until you change a saved host, and the button for a new host now says Add
• On Amazon Bedrock, OpenAI GPT-6 and GPT-5.6 models now work without turning on cross-region inference. They used to fail with "on-demand throughput isn't supported." India regions (ap-south-1/2) now use the in. inference profile

Full Changelog: https://github.com/cline/cline/compare/desktop-v0.0.35...desktop-v0.0.36

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.
• On Windows, @-mention search results now show the right file names. Nested files such as a subfolder's README rendered as /README.md, and the same open file could appear twice.
• Canceling a request 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/v4.1.20...v4.1.21

• 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.maxtokensrecovery 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 flushing telemetry. The wait for a freshly spawned hub also goes from 8 to 15 seconds for every client, because the first launch after an install or update regularly takes 8 to 13 seconds on Windows while the new binary is scanned
• Session renames made through the hub now persist. HubRuntimeHost.updateSession folded the new title into metadata.title, but the persistence layer only treats an explicit title as a rename, so the name reverted on relaunch, and wholesale metadata replacement dropped other keys such as pinned state. session.update now carries prompt and title as their own fields and leaves untouched metadata alone
• Errors shown in a session now survive navigation and resume. Terminal failures, including those recorded after retries are exhausted, are persisted as display-only history entries that hosts render when a session is reopened, and those entries are excluded from compaction and from what the model sees
• Plugin slash commands are now a shared core service. createPluginCommandService in @cline/core owns plugin command discovery, loading, name and result normalization, and execution, and both the CLI and the desktop sidecar consume it. A plugin that fails to load (an invalid manifest, a sandbox startup timeout) no longer rejects the service: the failure is logged, the empty command set is cached for the current plugin set, and the load is retried after 30 seconds instead of spawning a sandbox on every prompt. Handler exceptions still propagate
• Model lists for endpoint-owned providers now report failures instead of returning an empty list. For private-catalog providers (LiteLLM, Baseten, Hicap, Poolside) and providers backed by a modelsSourceUrl (Ollama, LM Studio), an unreachable host, TLS rejection, or bad key now propagates from getLocalProviderModels when the caller asks for errors, so hosts can show the cause rather than "no models". The hub guards its initial model load so a failing default provider no longer aborts peer setup
• Aborting a request now takes effect during the empty-response backoff. The retry middleware slept through its backoff without watching the AbortSignal and then re-dialed with an already-aborted signal, so cancellation surfaced up to one backoff late. createGatewayApiHandlerAsync also ignored setAbortSignal() because it captured the signal at construction; it now reads the live one like the sync handler
• The yolo-mode system prompt replaces the vague "always show your planning process" rule with concrete output guidance: keep plans to one short paragraph, act once the next step is clear, put code and edit payloads in tool arguments rather than drafting them in text, skip preambles for routine tool calls, and resolve repeated uncertainty with a focused check instead of more speculation
• New built-in provider aiand (ai&), an OpenAI-compatible endpoint serving open-weight models from Japan. It reads AIANDAPIKEY and defaults to zai-org/glm-5.3
• Streaming transcription in @cline/llms now covers native OpenAI (gpt-realtime-whisper with a transcription-bound client secret), Vercel AI Gateway, and ElevenLabs, exposed through the new getBuiltinStreamingTranscriptionModels. Gateway and OpenAI-compatible batch transcription use AI SDK transcription with retries, cancellation, and providerOptions, detect the audio format from the bytes, and return segments and warnings alongside the text. Voice model discovery validates audio-to-text capability and keeps realtime-only models separate
• @cline/core has a new ./cloud subpath export with the shared cloud session runtime (API client, controller, snapshots, state) that hosts previously each implemented. The cron spec watcher also resolves its directory with realpathSync.native before watching, avoiding a fatal libuv assertion on Windows when the path contains a short-name segment like RUNNER~1
• @cline/shared exports toPosixSeparators, replacing four copies of the same helper across core and the VS Code extension. No behavior change
• Refreshed the model catalog: still 209 providers, 6,237 to 6,386 models. The bundled Cline catalog drops z-ai/glm-5.3-flash, cline-free/solar-pro4, and poolside/laguna-s-2.1:free, and adds cline-free/mimo-v2.6-flash. The resolved default model changes for 19 providers that do not pin one in builtins.ts: 11 land on Claude Opus 5.5 (Cortecs, CrossModel, DigitalOcean, Eden AI, GitHub Copilot, both LLM Gateway providers, Ofox, Requesty, Vertex, Vivgrid), both StepFun providers move to Step 5 Preview, Above to MiMo V2.6 Flash, Fireworks to Ember-1, Kenari to DeepSeek V4.1 Flash, NanoGPT to Aion 3.5, OpenCode Go to Space Bunny Free, and Pioneer to GLiNER 2.5 Multi. If you use one of those providers without pinning a model, expect a different default

Full Changelog: https://github.com/cline/cline/compare/sdk/sdk/v0.0.85...sdk/sdk/v0.0.86

• 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
• Voice input works again with provider-backed transcription, and it streams live. OpenAI, Vercel AI Gateway, and ElevenLabs transcribe as you speak; when the network drops mid-recording, the app falls back to the browser's recognizer and retries the provider next time. The model picker labels which voice models are realtime and which transcribe recordings
• Your reasoning effort choice is now remembered per provider. Reopening a session reset the thinking picker to Low, and that Low was sent with your next message; switching providers now applies the effort you last picked for that provider
• Model lists for LiteLLM, Baseten, Hicap, Poolside, Ollama, and LM Studio now show the actual error when your endpoint can't be reached, instead of an empty list or a placeholder model. The app also trusts your operating system's certificate store, so endpoints signed by a corporate CA stop failing with "unable to get local issuer certificate"
• When the backend's hub fails to start, the error now says why, and the app waits up to 15 seconds for it instead of 8. The first launch after an install or update can take 8 to 13 seconds on Windows while the new binary is scanned
• Renaming a session now sticks after a relaunch, and renaming no longer clears other state such as pinning
• CLI sessions that were opened and closed without a prompt no longer show up as empty entries in the sidebar
• Arrow-key navigation in the slash command and @-mention menus now scrolls the highlighted option into view
• Free models under the picker's Free header no longer carry a redundant FREE badge on every row
• Short session titles now have room to edit
• 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 failing the run
• New provider: ai&, an OpenAI-compatible endpoint serving open-weight models from Japan
• Refreshed the model catalog to 6,386 models. The default model changes for 19 providers, 11 of them to Claude Opus 5.5 (including GitHub Copilot and Vertex). If you use one of those providers without picking a model, expect a different default

Full Changelog: https://github.com/cline/cline/compare/desktop-v0.0.34...desktop-v0.0.35