Cline mcp

All MCP Bug Fixes

• Fixed the scheduled-task report vanishing when a finished run's step collapsed.(desktop-v0.0.23)
• Fixed one wedged MCP server blocking the rest from shutting down, preventing process leaks.(desktop-v0.0.23)
• Fixed background Hub process memory ballooning during long sessions by changing session status updates to carry only state snapshots.(v4.1.17)
• Prevented hook scripts that fail to spawn from crashing the extension's core process and taking down the running task.(v4.1.17)
• Fixed a chat render crash caused by malformed `api_req` payloads.(v4.1.17)
• Ensured cost estimates do not appear in task history for subscription-billed tasks.(v4.1.17)
• Fixed an issue where pasted provider API keys were not stripped of invisible characters, preventing them from being corrupted and rejected.(v4.1.17)
• Improved error handling for OpenAI Codex sign-in when the callback port is occupied, now displaying a clear 'port in use' error.(v4.1.17)
• Fixed OAuth redirect errors during sign-in to be surfaced correctly instead of being reported as a missing authorization code.(v4.1.17)
• Prevented users from being signed out due to transient network failures while refreshing OpenAI Codex or OpenAI-compatible-account tokens.(v4.1.17)
• Fixed tool calling being silently disabled for Dify, SAP AI Core, opencode, and Codex CLI models.(v4.1.17)
• Fixed images being dropped from file reads on models with an empty capability list.(v4.1.17)
• Prevented restoring a checkpoint from silently removing commits made after it.(v4.1.17)
• Ensured `apply_patch` now preserves a file's existing CRLF line endings.(v4.1.17)
• Prevented an enabled but unreachable remote MCP server from stalling session startup by implementing a 10-second budget for remote connections.(v4.1.17)
• Ensured aborting a task now also cancels delegated subagents and teammates.(v4.1.17)
• Fixed Langfuse tracing to work in released builds by addressing issues with OpenTelemetry provider detection after minification.(v4.1.17)
• Ensured hook execution telemetry events are fired by correctly threading the task ID into hook runner creation.(v4.1.17)
• Fixed the CLI crashing when an enabled remote MCP server is unreachable by implementing a 10-second connection budget, preventing stalled session startups and session teardowns.(cli-v3.0.61)
• Resolved silent disabling of tool calling for Dify, SAP AI Core, opencode, and Codex CLI models by correcting how model capabilities are read.(cli-v3.0.61)
• Fixed images being dropped from file reads on models with empty capability lists.(cli-v3.0.61)
• Ensured Langfuse tracing now functions correctly in released builds by addressing issues caused by code minification renaming OpenTelemetry provider class names.(cli-v3.0.61)
• Prevented restoring a checkpoint from silently removing subsequent commits; it now refuses to run and informs the user.(cli-v3.0.61)
• Fixed `apply_patch` incorrectly altering line endings by preserving existing CRLF line endings.(cli-v3.0.61)
• Resolved OpenAI Codex (ChatGPT subscription) sign-in failures when port 1455 is occupied, now providing a clear 'port in use' error instead of an uncompletable browser flow.(cli-v3.0.61)
• Prevented users from being logged out due to transient network failures during token refreshes for Codex or OpenAI-compatible accounts.(cli-v3.0.61)
• Ensured aborting a session now correctly cancels delegated subagents and teammates, preventing their work from continuing.(cli-v3.0.61)
• Fixed scheduled tasks disappearing after a hub restart.(cli-v3.0.61)
• Corrected markdown flashing at the end of streamed responses.(cli-v3.0.61)
• Updated the message shown to the model when a tool call is rejected to clearly name the tool and indicate it's a user decision, not an error.(cli-v3.0.61)
• Fixed tool calling being silently disabled for gateway models that declared no capabilities, ensuring tool definitions are correctly included for Dify, SAP AI Core, opencode, and the Codex CLI.(sdk/sdk/v0.0.82)
• Fixed an empty capability list from incorrectly stripping image input, ensuring the file-read tool and model pickers correctly handle models with and without vision capabilities.(sdk/sdk/v0.0.82)
• Fixed Langfuse tracing to work in minified release builds by changing how the OpenTelemetry provider is identified.(sdk/sdk/v0.0.82)
• Fixed the hub being retired out from under running work by establishing a drain barrier before checking for activity, preventing sessions from being killed by hub replacements.(sdk/sdk/v0.0.82)
• Fixed restoring a checkpoint from running when HEAD has moved past the checkpoint's commit, preventing later commits from being lost.(sdk/sdk/v0.0.82)
• Fixed `apply_patch` to preserve a file's own CRLF line endings across updates.(sdk/sdk/v0.0.82)
• Fixed global rules not being discovered at ~/Cline/Rules on WSL and headless installs, ensuring rules written in the Documents folder are now correctly recognized.(sdk/sdk/v0.0.82)
• Fixed OpenAI Codex sign-in to fail fast with an actionable error when the callback port is in use, instead of leading to a broken sign-in flow.(sdk/sdk/v0.0.82)
• Fixed transient token-refresh failures from logging out Codex and OpenAI-compatible-account users, ensuring only rejected refresh tokens force re-authentication.(sdk/sdk/v0.0.82)
• Fixed aborting a session not propagating to spawned work, ensuring delegated subagents and teammates are now cancelled with the session.(sdk/sdk/v0.0.82)
• Fixed hub-managed schedules being wiped by cron reconciliation on hub restart.(sdk/sdk/v0.0.82)
• Fixed the hub creating filesystem watchers on agenda spec directories while the todo tool was disabled, preventing signature mismatch errors after editing specs on disk.(sdk/sdk/v0.0.82)
• Fixed scheduled tasks disappearing after app updates by ensuring hub-managed schedules are not wiped by cron reconciliation.(desktop-v0.0.20)
• Fixed OpenAI Codex (ChatGPT subscription) sign-in silently dead-ending when a specific callback port was in use, now providing an actionable error.(desktop-v0.0.20)
• Fixed Codex and OCA sign-ins from being dropped during token refresh when encountering transient network failures or server errors.(desktop-v0.0.20)
• Fixed checkpoint restore from silently removing commits made after the checkpoint, now refusing to reset the workspace in such cases.(desktop-v0.0.20)
• Fixed an enabled but offline remote MCP server from stalling session startup.(desktop-v0.0.20)
• Fixed an issue where the window title bar was not draggable across all views.(desktop-v0.0.20)
• Fixed the hub from watching agenda spec directories when the todo tool is disabled, preventing unnecessary OS watch handle usage.(desktop-v0.0.20)
• Fixed a crash in the codebase search tool that occurred on files with a single, very long line.(cli-v3.0.60)
• Resolved an issue where installing an MCP server would misinterpret '--' separators in install arguments as part of the server command.(cli-v3.0.60)
• Fixed the background hub process from ballooning in memory during long sessions by correcting how session status updates broadcast conversation transcripts.(cli-v3.0.60)
• Fixed cost estimates incorrectly showing for flat-rate subscription providers, preventing confusion about actual charges.(v4.1.16)
• Resolved an issue where signing back in would incorrectly change the last-used provider.(v4.1.16)
• Ensured hooks correctly resolve their workspace from the VS Code window, fixing issues with discovering `.clinerules/hooks` scripts and incorrect path resolution in multi-window scenarios.(v4.1.16)
• Fixed new files being created with incorrect line endings by ensuring native platform line endings are used.(v4.1.16)
• Prevented the codebase search tool from crashing when encountering files with a single, extremely long line.(v4.1.16)
• Ensured credentials embedded in git remote URLs are now redacted from workspace information sent to the model.(v4.1.16)
• Fixed an issue where installing an MCP server from the marketplace could misinterpret command separators.(v4.1.16)
• Prevented the hub's event log from growing indefinitely and filling the disk.(v4.1.16)
• Fixed scheduled sessions getting stuck on a thinking shimmer when opened during execution.(desktop-v0.0.18)
• Resolved an issue where installing plugins and MCP servers from the Marketplace failed due to missing executables in $PATH.(desktop-v0.0.18)
• Fixed the application beach-balling for several seconds upon quitting.(desktop-v0.0.18)
• Corrected hover cards that were flashing closed and reopening when clicked.(desktop-v0.0.18)
• Fixed an issue where internal `[SYSTEM]` steering messages were incorrectly displayed in scheduled and automation runs.(desktop-v0.0.18)
• Fixed the provider list being unscrollable when a provider detail panel was open.(desktop-v0.0.17)
• Resolved an issue where a failed settings save incorrectly indicated a provider was configured.(desktop-v0.0.17)
• Fixed Uninstall buttons that were collapsing into a broken square.(desktop-v0.0.17)
• Fixed unreadable selected text within input fields.(desktop-v0.0.17)
• Fixed the codebase search tool crashing the app on files with a single extremely long line.(desktop-v0.0.17)
• Prevented the hub's event log from growing indefinitely and filling the disk.(desktop-v0.0.17)
• Fixed an issue where the "Use MCP servers" toggle did not auto-approve all MCP tool calls. Previously, the toggle only applied to individually opted-in tools, making it seem ineffective. Now, it correctly governs all MCP tools automatically.(v4.1.15)
• Fixed tool calling being silently disabled for custom OpenAI-Compatible models when capabilities were inferred from convenience flags.(desktop-v0.0.16)
• Fixed enterprise MCP controls enforcement on the Customize marketplace, ensuring MCP entries are hidden when remote config disables the marketplace and limited to `allowedMCPServers` when an allowlist is configured.(v4.1.12)
• Restored tool calling for custom OpenAI-Compatible models that had an empty stored capability list.(v4.1.12)
• Fixed checkpoint restore from getting stuck permanently by ensuring sessions that were never prompted or only had persistence updates correctly report their status.(desktop-v0.0.15)
• Resolved an issue where the "No sessions found" message would flash briefly while session history was still loading.(desktop-v0.0.15)
• Corrected the work summary to accurately count elapsed time when a tool call was attached to the answer instead of the run.(desktop-v0.0.15)
• Fixed the settings gear to no longer maintain its hover state when the Account screen is open.(desktop-v0.0.15)
• Prevented ClinePass from being incorrectly identified as not OAuth-managed in the chat credential gate, which was causing it to request credentials it already possessed.(desktop-v0.0.15)
• Fixed an issue where copying a user message would include its internal envelope data.(desktop-v0.0.15)
• Resolved the problem of multi-line code blocks collapsing into a single line.(desktop-v0.0.15)
• Ensured that provider-executed tool activity is now shown, rather than being dropped.(desktop-v0.0.15)
• Fixed the `PreToolUse` hook `contextModification` to correctly reach the model and ensured `PostToolUse` hook output and `cancel` control are no longer discarded.(desktop-v0.0.15)
• Improved PowerShell command execution to fail fast on the first error, preventing excessive output and incorrect success reporting.(desktop-v0.0.15)
• Fixed an issue where turns would get stuck on a streaming shimmer without final output, only healing after sending another message.(desktop-v0.0.14)
• Corrected sessions being assigned the Yolo-mode system prompt when auto-approve was on, ensuring auto-approval is an independent tool policy.(desktop-v0.0.14)
• Resolved an issue where `/skill` and `/workflow` commands would dump the entire skill body into the chat.(desktop-v0.0.14)
• Fixed command execution breaking for entire sessions when a model emitted a full command line without separate arguments.(desktop-v0.0.14)
• Ensured restoring a checkpoint now trims the saved transcript to prevent showing reverted file changes.(desktop-v0.0.14)
• Fixed Gemini custom base URLs, including host-root and proxy values, which were previously failing.(desktop-v0.0.14)
• Restored correct LiteLLM input token limits reported by the server, preventing a 128K default override.(desktop-v0.0.14)
• Fixed misaligned columns in the Usage table.(desktop-v0.0.14)
• Fixed routine dialog dropdowns not responding to mouse clicks.(desktop-v0.0.14)
• Fixed an issue where Cline installations on different builds would cause Hub daemons to shut each other down, resulting in abnormal socket closes for live sessions.(sdk/sdk/v0.0.75)
• Resolved a problem where newer builds would prematurely replace active Hubs, causing sessions to die mid-handshake.(sdk/sdk/v0.0.75)
• Fixed an issue where idle plugin sandbox processes would linger unnecessarily for the life of the session.(sdk/sdk/v0.0.75)
• Corrected `cline doctor fix` to accurately report process statuses, including surviving processes, newly appeared processes, respawned daemons, and held startup locks.(sdk/sdk/v0.0.75)
• Fixed the Claude Code provider to correctly run native tools and anchor sessions on the workspace directory, resolving issues with tool definitions and permission rules.(cli-v3.0.54)
• Fixed truncated tool-call JSON from being silently repaired into incorrect arguments, now rejecting payloads with unterminated strings.(cli-v3.0.54)
• Fixed strict providers rejecting turns with only empty text parts in user messages.(cli-v3.0.54)
• Fixed a mid-turn crash occurring during streamed tool calls with non-zero or non-contiguous indexes when using LiteLLM's Anthropic passthrough.(cli-v3.0.54)
• Fixed the Hub daemon from logging an unhandled error and exiting non-zero when a client was still connected at shutdown.(cli-v3.0.54)
• Fixed per-task token totals from being inflated due to cache-heavy sessions by reporting token telemetry in disjoint buckets.(cli-v3.0.54)
• Resolved an issue where upgrading the CLI would cause it to communicate with a daemon running the previous release; the CLI now correctly retires and respawns the Hub daemon on the new code.(cli-v3.0.54)
• Fixed the Claude Code provider from being unusable for agentic work by correctly handling tool definitions and session anchoring.(sdk/sdk/v0.0.74)
• Fixed truncated tool-call JSON from being silently repaired into incorrect arguments; payloads with unterminated strings are now rejected.(sdk/sdk/v0.0.74)
• Fixed strict providers from rejecting turns with empty text parts in messages.(sdk/sdk/v0.0.74)
• Fixed a mid-turn crash on streamed tool calls with non-zero or non-contiguous indexes by updating AI SDK packages.(sdk/sdk/v0.0.74)
• Fixed the Hub daemon from logging an unhandled rejection and exiting non-zero on client disconnect during shutdown; shutdown is now clean.(sdk/sdk/v0.0.74)
• Fixed involuntary Cline logouts from being reported instead of silent credential clearing, and prevented transient network failures from being logged as logouts.(sdk/sdk/v0.0.74)
• Fixed git jargon display when opening non-git repos.(desktop-v0.0.11)
• Fixed folder picker to report failures instead of doing nothing.(desktop-v0.0.11)
• Fixed diff panel to resolve file paths against the session's working directory for correct diff opening.(desktop-v0.0.11)
• Fixed failed turns to surface errors in the transcript instead of leaving the chat blank.(desktop-v0.0.11)
• Fixed a startup script-load error.(desktop-v0.0.11)
• Fixed webview errors to be attributed to the source URL that caused them.(desktop-v0.0.11)
• Fixed signing out to be handled as a normal state instead of a command error.(desktop-v0.0.11)
• Fixed tool output to no longer nest its own scrollbar.(desktop-v0.0.11)
• Fixed hung MCP servers to no longer prevent session creation.(desktop-v0.0.11)
• Fixed stdio servers that were never configured from blocking indefinitely.(desktop-v0.0.11)
• Fixed LiteLLM requests to route through Chat Completions instead of the Responses API.(desktop-v0.0.11)
• Ensured checkpoints are picked up when git is initialized part-way through a session.(desktop-v0.0.11)
• Fixed an issue where prompts queued during an interrupted turn were silently dropped; they are now preserved and drained after the turn aborts, with the stop surfaced.(v4.1.7)
• Resolved an issue where session context was not durable across aborts and hub restarts, ensuring interrupted sessions now resume with their previous state.(v4.1.7)
• Fixed a bug where a mode switch aborting a running turn would not settle the turn phase.(v4.1.7)
• Corrected the reporting of queued-turn failures, which now appear as `run.failed` instead of completing silently.(v4.1.7)
• Prevented hung MCP servers from causing session creation to fail and provided stdio servers a 30-second initialization budget instead of blocking indefinitely.(v4.1.7)
• Ensured OAuth authorization for SSE MCP servers is now surfaced on a 401 error instead of causing an outright failure.(v4.1.7)
• Fixed requests against LiteLLM proxies by routing LiteLLM through Chat Completions instead of the Responses API.(v4.1.7)
• Resolved an issue where network interruptions mid-stream but before any model output caused turns to fail; these are now retried.(v4.1.7)
• Fixed Vertex ADC token refreshes to use the configured fetch, enabling them to work behind proxies and custom transports.(v4.1.7)
• Included untracked files in checkpoint diffs when a snapshot was taken and ensured checkpoints are picked up even when git is initialized part-way through a session.(v4.1.7)
• Fixed @-mention file search in empty windows to fall back to the session cwd or Desktop.(v4.1.7)
• Prevented foreign compiled plugin-sandbox bootstraps from running for a source host.(v4.1.7)
• Fixed issue where queued messages were unreadable on light-theme terminals due to poor color contrast.(cli-v3.0.52)
• Fixed rendering of MCP tool results to display as readable text instead of escaped (cli-v3.0.52)
• Fixed mangling of binary payloads when expanded in MCP tool results.(cli-v3.0.52)
• Fixed rendering issues caused by malformed tool input/output payloads, which now degrade gracefully instead of throwing errors.(cli-v3.0.52)
• Fixed issue where a hung MCP server would prevent session creation.(cli-v3.0.52)
• Fixed indefinitely blocking stdio servers that were never configured by providing a 30-second initialization budget.(cli-v3.0.52)
• Fixed LiteLLM requests to route through Chat Completions instead of the Responses API, resolving issues with calls against LiteLLM proxies.(cli-v3.0.52)
• Fixed network interruptions mid-stream (before model output) to retry instead of failing the turn.(cli-v3.0.52)
• Fixed Vertex ADC token refreshes to use the configured fetch, ensuring they work behind proxies and custom transports.(cli-v3.0.52)
• Fixed prompts queued during a turn from being lost when interrupted; they are now preserved across user-initiated aborts, drained after a turn aborts itself, and edits made to the queue inside the abort window are applied.(sdk/sdk/v0.0.72)
• Fixed session context from resetting after aborts and hub restarts; interrupted sessions now resume with their previous state.(sdk/sdk/v0.0.72)
• Fixed queued turns that fail from completing silently; they are now reported as `run.failed`.(sdk/sdk/v0.0.72)
• Prevented a hung MCP server from taking down session creation.(sdk/sdk/v0.0.72)
• Resolved issue where stdio servers that were never configured blocked indefinitely; they now get a 30-second initialize budget.(sdk/sdk/v0.0.72)
• Fixed remote SSE MCP servers from failing outright on a 401; they now surface an OAuth authorization prompt.(sdk/sdk/v0.0.72)
• Fixed LiteLLM requests routing through the Responses API; they now route through Chat Completions, resolving calls against LiteLLM proxies.(sdk/sdk/v0.0.72)
• Fixed network interruptions mid-stream (before model output) from failing the turn; they are now retried.(sdk/sdk/v0.0.72)
• Ensured Vertex ADC token refreshes use the configured fetch, enabling them to work behind proxies and custom transports.(sdk/sdk/v0.0.72)
• Fixed checkpoints from not picking up untracked files when the snapshot was taken; diffs now include these files.(sdk/sdk/v0.0.72)
• Fixed checkpoints from not being picked up when git was initialized part-way through a session.(sdk/sdk/v0.0.72)
• Resolved issue where a source host ran a foreign compiled plugin-sandbox bootstrap; this is no longer the case due to centralized plugin management.(sdk/sdk/v0.0.72)
• Fixed an issue where the first message of a chat and some queued messages were rendering twice.(desktop-v0.0.10)
• Resolved a bug where the composer could get stuck on "Agent is working…" after a turn had already finished.(desktop-v0.0.10)
• Fixed an issue with Vertex AI credential refreshes that prevented ADC authentication behind proxies and custom networking.(desktop-v0.0.10)
• Fixed an issue where turns completing with a plan were incorrectly treated as failed when a plan-blocked command was the only tool call, ensuring the turn correctly re-runs the presented plan instead of appearing to do nothing.(v4.1.4)
• Fixed tool paths to display relative to the workspace in the chat view instead of absolute paths.(v4.1.4)
• Fixed pending attachments to reset when starting a new task, preventing images from previous tasks from carrying over.(v4.1.4)
• Fixed a generic failure to surface a clear error when the selected provider has no API key configured.(v4.1.4)
• Fixed MCP tool and resource lists to refresh when a server sends a `list_changed` notification, instead of only showing a toast.(v4.1.4)
• Fixed installed plugins to display under their real package names instead of all appearing as "index".(v4.1.4)
• Corrected the Linux keybinding label in the Plan/Act mode tooltip.(v4.1.4)
• Fixed recovery from running out of context by compacting and retrying once, and providing explanations for genuinely unrecoverable cases.(v4.1.4)
• Fixed empty model responses to retry on every provider, not just Ollama, resolving hard "Model returned empty response" failures on OpenRouter, Cline, and OpenAI-compatible endpoints.(v4.1.4)
• Fixed Claude 4.6+ and 5.x models from being rejected with "thinking.type.enabled is not supported" when resolving from the offline catalog or a hand-typed model ID.(v4.1.4)
• Restored Bedrock prompt caching by ensuring the correct cache format is used and routing Bedrock foundation models through geo inference profiles.(v4.1.4)
• Fixed requests for reasoning models on OpenAI-compatible endpoints by sending `max_completion_tokens` and substituting image content for models without image support.(v4.1.4)
• Fixed MiniMax default model inheritance from models.dev and refreshed the bundled catalog to include Infomaniak and SCX.ai.(v4.1.4)
• Fixed error telemetry to report provider failures only once and rate-limit repeated failures from unattended retry loops.(v4.1.4)
• Fixed issue where MCP tool names exceeding 64 characters caused errors with OpenAI-compatible providers(cli-v3.0.31)
• Fixed marketplace edge cases related to refreshing MCP servers after installs, remote configuration disabling of the MCP Marketplace tab, hiding workflows, surfacing plugin-bundled skills, and uninstalling shared marketplace entries.(v4.0.0)
• Fixed chat submission during active turns by queuing new messages instead of dropping or racing them, ensuring pending/queued states display promptly, and removing delayed send behavior.(v4.0.0)
• Fixed editing previous user messages, ensuring Escape cancels editing locally and clarifying reset action labels.(v4.0.0)
• Fixed terminal reliability issues, including standalone Windows output capture, hardened PowerShell command handling, correct running-state display, and preservation of raw structured commands.(v4.0.0)
• Fixed SDK tool-result and provider-message budgeting by implementing default truncation for large tool outputs, capping assistant text, limiting ingestion of bash/file-read/search output, and bounding media budgets.(v4.0.0)
• Fixed login and feature-flag resolution by ensuring the correct user/account identity is used on startup and simplifying the login user experience.(v4.0.0)
• Fixed Cline model display names so they resolve correctly by model name(cli-v3.0.28)
• Fixed issue where the CLI would silently treat unknown commands or unquoted multi-word input as a prompt; the CLI now rejects these inputs with a clear error message.(cli-v3.0.27)
• Fixed apply_patch to fail correctly when a hunk is skipped(cli-v3.0.25)
• Fixed run_commands to return captured stdout on failure and correctly handle split heredocs(cli-v3.0.25)
• Fixed search tools to treat zero results as a successful operation(cli-v3.0.25)
• Fixed disabled-reasoning handling specifically for StepFun flash operations(cli-v3.0.25)
• Fixed history resume rendering isolation issues(cli-v3.0.25)
• Fixed incorrect Cline OAuth token formatting within the provider configuration(cli-v3.0.25)
• Fixed an issue where settings writes during MCP server delete/add flows caused the MCP server list to be emptied by the file watcher.(v3.88.0)
• Corrected Fireworks model metadata and cache pricing, and removed stale Fireworks AI models.(v3.88.0)
• Fixed issue where the MCP "Restart Server" button was enabled when a server was toggled off(v3.84.0)
• Fixed validation failures occurring for MCP servers that require an object.(v3.83.0)
• Resolved issue where `write_to_file` could not create or overwrite files containing empty content.(v3.83.0)
• Fixed unreliable streamable HTTP MCP reconnects after disconnects(v3.71.0)
• Fixed a crash related to reasoning delta when processing usage-only stream chunks(v3.67.0)
• Corrected an issue where OpenAI tool ID transformation was incorrectly restricted to the native provider only(v3.67.0)
• Fixed authentication check failure when operating in ACP mode(v3.67.0)
• Resolved an issue where the CLI yolo setting was being incorrectly persisted to disk(v3.67.0)
• Fixed the inline focus-chain slider to correctly operate within its feature row(v3.67.0)
• Ensured compatibility with Gemini 3.1 Pro(v3.67.0)
• Corrected authentication issues when using the Cline CLI with the ACP flag(v3.67.0)
• Fixed CLI to correctly handle stdin redirection in CI/headless environments.(v3.58.0)
• Fixed CLI to preserve OAuth callback paths during auth redirects.(v3.58.0)
• Fixed OAuth callback reliability in VS Code Web by generating callback URLs via `vscode.env.asExternalUri`.(v3.58.0)
• Fixed Terminal to surface command exit codes in results and improved long-running `execute_command` timeout behavior.(v3.58.0)
• Fixed UI rendering by adding a loading indicator and fixing `api_req_started` rendering.(v3.58.0)
• Prevented duplicate streamed text rows after task completion during task streaming.(v3.58.0)
• Fixed API to preserve the selected Vercel model when model metadata is missing.(v3.58.0)
• Ensured telemetry flushes on shutdown and routed PostHog networking through proxy-aware shared fetch.(v3.58.0)
• Increased Windows E2E test timeout to reduce flakiness in CI.(v3.58.0)
• Removed Minimax-2.1 from the free model list because its free trial concluded(v3.49.0)
• Fixed issue where expired tokens were being used in authenticated requests(v3.47.0)
• Fixed diff view to correctly exclude binary files that lack file extensions(v3.47.0)
• Fixed issue where file endings and trailing newlines were not being preserved(v3.47.0)
• Fixed rate limiting issues specific to the Cerebras provider(v3.47.0)
• Fixed Auto Compact functionality for the Claude Code provider(v3.47.0)
• Made Workspace and Favorites history filters operate independently(v3.47.0)
• Fixed remote MCP server connection failures, specifically handling 404 responses(v3.47.0)
• Disabled native tool calling for the Deepseek 3.2 speciale model(v3.47.0)
• Fixed the Baseten model selector interface(v3.47.0)
• Fixed a race condition in MCP settings where toggling auto-approve or changing timeout settings caused the UI to flash and revert.(v3.45.1)