Skip to content

Releases

November 14

  • Agents can now use the run_workflow tool to run bundled Python workflow scripts that coordinate multiple subagents in parallel or stages.
  • Managed admin settings introduce a pluginControls section with pluginsDisabled and marketplaceDisabled settings for organizations not using Gemini Enterprise.
  • The /resume command now accepts any unique conversation ID prefix and no longer stalls by loading full histories up front in the picker.
  • Custom HTTP headers configured for MCP servers, such as API keys, are no longer forwarded if the server responds with a redirect to a different host.
  • Windows fixes ensure ! shell commands preserve double quotes in PowerShell, /copy reaches local clipboards over SSH, and scheduled automations receive multi-line prompts.
  • Use each model's recommended context window by default while preserving explicit context configurations.
  • Enforce session-wide AI credit limits configured via --max-ai-credits across subagents and resumed sessions.
  • /copy now preserves formatting, paragraphs, tables, lists, and quotes when pasted into Word, Outlook, or Teams.
  • /chronicle cost-tips now incorporates Copilot App sessions and recommends cheaper models and reasoning settings based on recorded usage.
  • Tool previews, timeline rows, and MCP system prompt displays now strip terminal escape sequences and control characters to prevent UI corruption.

November 13

  • Sandboxed Bash commands now follow the normal permission flow by default; set sandbox.autoAllowBashIfSandboxed: true to keep running them without prompts.
  • Scheduled tasks and /loop now ask for confirmation before running a skill that declares allowed-tools or hooks unless Skill(name) is added to permissions.allow.
  • Prompt suggestions are now available in VS Code terminals and can be accepted with Tab or disabled by setting promptSuggestionEnabled: false in settings.json.
  • Fixed a security issue in how zstd-compressed responses are decoded.
  • Fixed subagents with their own model being switched to the plan-mode model when the model setting is opusplan or haiku.
  • The CLI now uses each model's recommended context window by default while preserving explicit context choices.
  • /chronicle cost-tips now includes Copilot App sessions, recommends cheaper models and reasoning settings based on usage, and reports earlier when local usage is available.
  • Subagents and resumed sessions now properly respect the session-wide AI credit limit configured with --max-ai-credits.
  • Startup handling is improved on macOS Ventura 13.5 or later.
  • Enterprise-managed Shell, Bash, and PowerShell permission rules now support regular expressions.
  • The new --host-allow-session-creation flag allows remote clients to create sessions on CLI hosts for a single launch without modifying saved settings.
  • Permission prompts, questions, and confirmations now appear directly inside the sub-agent view so they can be answered without leaving it.
  • Sessions started with /new or from the Sessions tab can now be controlled from GitHub when remote sessions are enabled.
  • Follow-up prompts typed while an agent is active now run strictly in the order they were submitted.
  • Subagents now run in the background by default and return their session ID immediately, reporting results back to the parent session whenever they become idle.
  • Transcript verbosity in the TUI now defaults to low, summarizing runs of tool calls and thoughts into a single collapsible activity row.
  • The desktop app adds a Jump to message command palette action (Cmd+Alt+G or Ctrl+Alt+G) to search and jump to prompts across the current session.
  • Providers now support mutual TLS client certificates configured via settings.tls with cert, key, and optional passphrase.
  • Local model discovery for Ollama, LM Studio, and vLLM now waits up to 15 seconds per request instead of 1 second to better handle remote or tunnelled endpoints.
  • Added Gemini 4 Hydrogen and Gemini 4 Carbon model variants to /model and --model across API key, ADC, and enterprise gateway authentications.
  • Added click-to-expand in full screen rendering mode, allowing individual tool groups, thoughts, prompt images, and multi-line command outputs to be toggled directly.
  • Fixed security issues where URL path traversal could bypass terminal sandbox network rules and Windows shell quoting could bypass always-allow command rules.
  • Fixed files with UTF-8 byte order marks (BOM) causing settings.json, hooks.json, and custom agent files to be silently ignored or rejected on Windows.
  • Added live context window usage updates to agy acp sessions for integration with external editors.
  • Enterprise-managed Shell, Bash, and PowerShell permission rules now support regular expressions.
  • The new --host-allow-session-creation flag lets remote clients create sessions on CLI hosts for a single launch without changing saved settings.
  • Follow-up prompts typed while the agent is working now run strictly in the order they were submitted.
  • Scratch files under /tmp on macOS are allowed without path permission prompts unless --disallow-temp-dir is set.
  • Interactive /plugin marketplace commands now detect repository-local marketplaces when launched from a subdirectory.

November 12

  • Code Mode is available as an experimental opt-in feature where Claude writes JavaScript to execute tool calls.
  • Unrecognized model IDs now default to an assumed 1M token context window instead of 200K, which can be reverted using CLAUDE_CODE_DISABLE_1M_CONTEXT=1.
  • Plugins can now declare supported version ranges via engines["claude-code"] in plugin.json, blocking incompatible installations and updates.
  • The claude plugin validate command now type-checks TypeScript hook modules and warns on errors unless --strict is specified.
  • In VS Code, untrusted workspace folders now trigger a trust prompt before starting a session, and running shell commands stream live output directly under their chat cards.
  • An experimental masked dialog can answer git, ssh, and sudo password prompts from agent commands when enabled with /experimental on.
  • Changes to memory, URL rules, hooks, and bypass permissions mode in /settings now apply immediately without restarting the CLI.
  • /compact respects the session AI credit limit, and sessions now offer a credit limit top-up at 90% usage before access to tools is blocked.
  • MCP commands and plugin settings preserve invalid configuration files and report detailed syntax errors instead of overwriting existing entries.
  • Legacy remote steering has been removed in favor of AHP environments for cloud sessions and remote connections.
  • Sessions without a saved mode now start in Auto mode where available, while setting Tool Permission to strict still requires approval for file edits.
  • Fixed multi-second UI freezes, streaming stalls, and window resizing delays when commands print very long lines without spaces.
  • Fixed agent replies stalling while tool call token counts finish animating.
  • Fixed backgrounded commands silently stalling when waiting for input such as passwords or confirmation prompts.
  • Fixed Gemini 4 Carbon requests through enterprise model gateways failing with HTTP 400 thinking budget errors.
  • Avoid unnecessary error popups during Microsoft Entra sign-in when an organization account is available.
  • Wrap `/usage` model rows at the current terminal width instead of truncating at 100 columns.
  • Preserve existing MCP configuration and display a save error when `/mcp add` or `/mcp edit` encounters an invalid configuration file.
  • Prevent the `/` command picker from matching, listing, or preselecting skills based on words hidden inside terminal escape sequences in descriptions.
  • Update screen reader announcements to identify the current session and report when sessions in the list are idle or need attention.

November 11

  • copilot -p now warns on stderr and in JSON output when MCP servers fail to start rather than continuing silently.
  • /compact now respects session AI credit limits, preventing extra charges once the limit is reached.
  • Starting the CLI from a deleted working directory now displays a recovery message instead of crashing with a Node stack trace, and -C <directory> works from that state.
  • Plugin and marketplace updates no longer overwrite other settings when settings.json contains an invalid value.
  • Memory and URL rule updates in /settings take effect immediately without requiring a CLI restart.
  • Auto mode prompts less frequently for safe reads, skipping repository age and popularity checks on fetches or single-file downloads, allowing public SSH key reads, and recognizing hostnames starting with digits.
  • Pay-as-you-go Google Cloud accounts using Application Default Credentials via AGY_ADC_AUTH can now use Auto mode.
  • The /swarm board interface adds visual status indicators for active members, supporting animated cards in Kitty graphics terminals and pixel art in standard terminals.
  • Plugin installations and imports now perform atomic updates to prevent broken states on failure, and agy plugin install now loads skills declared in Claude Code plugin manifests.
  • Undo and redo actions in the prompt now return the cursor to where the edit occurred instead of jumping to the end of the input.

November 10

  • Mark sessions resumed with copilot --resume as in use immediately to prevent secondary windows or background processes from writing to them while idle.
  • Handle lone UTF-16 surrogate escapes in local MCP server replies with a replacement character instead of hanging tool calls or timing out after 60 seconds.
  • Show the sign-in prompt immediately after /logout removes the last saved account with invalid credentials, without requiring an application restart.
  • Preserve plugin preferences and settings files on startup when settings.json contains invalid values or fails to parse.
  • Make copilot mcp add retain existing invalid server configurations and prompt for removal before reusing their names.
  • Agent commands can now answer git, ssh, and sudo password prompts in a masked dialog when experimental mode is enabled with /experimental on.
  • Sessions now offer an AI-credit limit top-up at 90% usage before the agent loses access to tools.
  • Expanded reasoning sections can now be collapsed by clicking anywhere within them instead of scrolling back to the header.
  • Tool calls no longer fail when MCP tool results include top-level keys such as skills, tools, or messages.
  • The /diff command on feature branches now includes committed files containing non-ASCII characters, quotes, backslashes, or tabs.
  • Added preview support for swarms via /swarm <task>, enabling a coordinator agent to launch parallel worker agents across models that collaborate over a shared message board and monitor stalled tasks.
  • Added /swarm board and an active-items tray indicator (alt+s or alt+x to dismiss) to inspect member rosters, view live post feeds, and render Markdown messages.
  • Added Gemini 4 Carbon (Low, Medium, and High) model options for /model and --model when using Application Default Credentials (AGY_ADC_AUTH).
  • Subagents and swarm workers now preserve staged, unstaged, and untracked workspace changes when creating isolated worktrees or directory copies.
  • Added the ANTIGRAVITY_DISABLE_EMBEDDED_RIPGREP environment variable to bypass the bundled ripgrep binary in restricted environments and fall back to PATH or built-in search.
  • Support IPv4 and IPv6 CIDR notation for sandbox host allow and deny rules in settings.
  • Add support for Claude Fable 5.2 in the model picker under New models.
  • Pass text typed after /init to the model as extra guidance for generated instruction files instead of silently dropping it.
  • Allow URL permission prompts to approve every subdomain for the current session.
  • Automatically delete orphaned search cache indexes from deleted repositories and worktrees.
  • The model picker now supports Claude Fable 5.2 under New models.
  • Sandbox host allow and deny settings now accept IPv4 and IPv6 CIDR notation.
  • Sandbox path settings now support environment variables while preserving authored expressions.
  • Prompt mode with `--output-format json` now emits a final result record with an exit status even when a run fails early.
  • Session history now retains every turn when multiple windows work on the same session simultaneously.
  • Editors can now switch agy acp sessions to Auto mode to automatically approve workspace edits and sandboxed terminal commands.
  • Hook commands now receive hookEventName, cwd, and toolCall.id in their JSON input to identify working directories and match PreToolUse to PostToolUse calls.
  • Auto mode reviews commands that exit the terminal sandbox faster and with fewer timeouts by switching to a lighter Gemini Flash model.
  • Fixed subagent crashes and freezes in narrow terminals, along with bugs where subagent questions closed prematurely or opening a subagent view hung the CLI.
  • Fixed agy acp sessions dropping their permission mode upon reload, turns hanging on secondary permission requests, and stored sign-ins overflowing OS credential stores.
  • Added fast mode on Vertex in the global, us, and eu regions.
  • Requests automatically fall back to standard speed if Vertex refuses fast mode.
  • pi auth login <provider> and pi auth logout <provider> allow logging in and out of providers directly from the shell without starting a session, supporting --api-key and --no-browser options.
  • cacheWarming now defaults to "idle" to keep eligible prompt caches warm for up to 30 minutes between agent runs when cost-effective.
  • Anthropic requests now use message threads by default, sending only new messages and falling back to full requests when earlier context changes.
  • pi update now lists each updated package's npm version or git commit change and reports whether updates were unchanged, skipped, or warned.
  • Default models across multiple providers were updated to newer releases, including Claude Opus 5.5 for Anthropic and Amazon Bedrock, and GPT-6.1 Sol for Azure.

November 9

  • Text typed after /init now passes to the model as extra guidance for the generated instructions file rather than being dropped.
  • URL permission prompts now allow approving all subdomains for the active session.
  • The CLI now asks for confirmation before /share or /export follows symlinks when saving local files.
  • Running /research again on the same topic saves a new report instead of attempting to overwrite the earlier file.
  • Custom agents added to .github/agents/ during an active session can be used immediately without reloading.
  • Added Claude Fable 5.2 (claude-fable-5-2) with a 1M token context window as the default Fable model.
  • Added the --settings <file> flag to claude remote-control to apply a specified settings configuration to every started session.
  • Fixed MCP OAuth requests leaking secrets such as refresh tokens to external hosts when redirected by a server.
  • Fixed NotebookEdit and Edit tools corrupting files, replacing whole files, or deleting notebook cells on network paths such as WSL folders and UNC shares.
  • Fixed Bash permission checks running read-only commands outside the sandbox without prompts and evaluating write permissions from the wrong folder.
  • Add /index commands and enable managed shared tgrep indexing by default for eligible repositories.
  • Add sandbox.userPolicy.filesystem.enumeratePaths in settings.json to configure enumerate-only directory grants on Windows.
  • Restart local MCP servers in sandboxed sessions after /cd to apply the new working directory's denied paths.
  • Allow Ctrl+C to cancel a running /compact operation instead of arming the exit prompt and closing the CLI on a second press.
  • Resolve HTTP 421 Misdirected Request errors during /model by re-resolving the account endpoint rather than displaying a model list failure.

November 7

  • Subagent rows above the prompt, the /agents view, and the status line JSON now display each subagent's model name.
  • Plugin hook commands now receive PLUGIN_ROOT and PLUGIN_DATA environment variables pointing to the plugin folder and its persistent storage directory.
  • Plugin management commands such as agy plugin uninstall, enable, and disable now work correctly with symlinked plugins or plugin names containing spaces and symbols.
  • Agent turns that fail with HTTP/2 header timeouts when a model takes more than 60 seconds to start responding are now automatically retried.
  • File tools now correctly read UTF-16 and legacy code-page files, preserve existing byte order marks and encodings on edits, and write PowerShell scripts with UTF-8 BOMs for Windows PowerShell 5.1 compatibility.
  • PDF attachments on Grok 4.7 and Grok 4.6 now fall back to file references rather than failing the turn.
  • Running local MCP servers now restart after /cd when the sandbox is active, ensuring the server applies the new directory's denied paths.
  • Added /index commands and enabled managed shared tgrep indexing by default for eligible repositories.
  • Added the sandbox.userPolicy.filesystem.enumeratePaths setting in settings.json for Windows enumerate-only directory grants.
  • Successful command output and file diffs now remain collapsed in concise transcript mode until expanded.
  • Autopilot with Assisted Permissions now attempts alternative approaches and reports blocked actions instead of stopping at permission prompts.
  • The grep and glob tools now return discovered matches when unreadable files are encountered instead of failing entirely on permission errors.
  • Support shared MCP server connection pooling across projects using pool: true to reduce process overhead.
  • opencode run adds -w (--worktree) to execute sessions in a new git worktree and --archived to keep automated runs out of session lists.
  • opencode run now waits for background shell commands, subagents, and image jobs to finish so the model can respond to their results before exiting.
  • GUI plugins can now register custom color themes, add actions or custom rendering to code blocks, and use native UI components.
  • Mobile composer actions on phones and tablets now sit in a dedicated touch dock above the keyboard.

November 6

  • Add the /code command to open the current session in VS Code.
  • Open a unified dashboard in /settings with global search across configuration settings and local resources.
  • Make /exit close only the current session instead of quitting the entire CLI when the Sessions sidebar is open.
  • Report interactive CLI program status to supporting terminals, including indicators when user attention is needed.
  • Render inline images in Rex terminals and herdr panes using Kitty graphics direct placements without manual setup.
  • Realtime two-way voice chat lets you converse with agents, hear live progress updates, and steer coding tasks in the Agents Window, the IDE, on web, and in Cursor for iOS.
  • Voice calls can be initiated using the waveform "Start voice chat" button when the chat input is empty, or on iOS by saying "Hey Siri, call Cursor" with Lock Screen controls via Live Activities.
  • Voice chat is powered by Grok Voice, supports multiple voice selections (including Ara, Eve, Leo, Rex, and Sal), and provides zero data retention under Privacy Mode.
  • Two-way voice chat operates separately from text dictation, which remains available via Cmd+Shift+Space.
  • Auto mode with telemetry or feature flags disabled now prompts for or denies actions that the server cannot verify instead of falling back to the built-in classifier, unless CLAUDE_CODE_AUTO_MODE_SERVER=0 is set.
  • Terminal sessions now support MCP tasks by automatically executing tool calls in the background and returning results to the model upon completion.
  • Added the permissions.denyInsteadOfPrompting setting to automatically deny tool calls that would otherwise pause for a permission prompt and continue the turn.
  • Added the /claude-api latency-optimize command to locate latency bottlenecks in Claude API code and evaluate faster configuration settings.
  • Added an "Earlier in this session" section to /tasks to inspect, reopen, and message finished subagents after restarts or removal from the active task list.
  • Skills from cloud plugins connected with `/plugin` are now available in `--remote-agent` sessions and appear in the `/` menu and `/skills` list.
  • macOS Option-key shortcuts like `alt+m` now work out of the box without configuring Option as Alt in terminal settings, configurable via `AGY_CLI_MAC_OPTION_AS_ALT`.
  • `--dangerously-skip-permissions` no longer auto-approves built-in protected commands or actions restricted by Gemini Enterprise administrator policies.
  • File access prompts for locations outside the workspace now allow or deny the specific path rather than granting access to everything outside the workspace.
  • Plugin management commands and repository-level plugin configs now enforce strict path validation to prevent writing or loading outside designated workspace folders.
  • Images removed from prompts with Ctrl+U, Ctrl+K, or Ctrl+W, or replaced via history recall, are no longer mistakenly sent with messages.
  • Switching sessions while other sessions are busy no longer crashes the CLI and leaves a blank terminal.
  • Running `copilot --resume <id>` now times out after 10 seconds instead of hanging indefinitely if the Copilot task service fails to respond.
  • The model picker now displays context window sizes for flat, unpriced, and single-tier models.
  • Pasted text with Unicode line or paragraph separators now splits across separate prompt lines, and DEL or C1 control characters are automatically stripped.
  • /settings now opens a unified dashboard for all users with global search across settings and local resources.
  • Add the /code command to open the active session directly in VS Code.
  • /exit now closes the current session instead of exiting the entire CLI when the Sessions sidebar is open.
  • Add GPT-6.1 Terra as a selectable model under New models when available to your account.
  • Add a Sandbox extensions setting to run extension processes inside the sandbox environment.

November 5

  • Cloud agents can now scaffold web applications, provision hosting, spin up Postgres databases, and configure authentication behind single sign-on from a text prompt.
  • Publishing deploys applications to a chosen unique domain from new or existing repositories across Cursor Origin and GitHub.
  • Automated deterministic and agentic security checks run before publishing to block exposed credentials or secrets in application files.
  • Deployed applications can be shared with specific individuals, groups, whole teams, or publicly on individual plans, and support provisioning API keys for programmatic access.
  • Team administrators gain controls to inspect team applications and secrets, add custom security review rules, approve blocked deployments, and unpublish applications.
  • Autopilot no longer starts implementing recommendations after requests that only ask for a review or analysis.
  • Requests that return an empty 404 under load are now retried automatically instead of ending the turn with an execution failure.
  • OneAuth is now the default Microsoft Entra sign-in path on Windows and macOS.
  • Assistant replies now preserve non-HTML angle-bracket placeholders and types such as Vec<String> instead of silently removing them.
  • agy acp now offers all models available to your account to every ACP client, lifting the previous restriction where non-Gemini models were limited to specific editors.
  • Subagent rows above the prompt now name the specific tool a subagent is waiting on instead of showing a generic processing state.
  • Terminal commands with background processes still holding output streams now release the agent after at most 1 second instead of 5.
  • File edits on Windows now preserve and match CRLF line endings instead of failing with target content not found errors.
  • Commands run by the agent on Windows now use UTF-8 by default for Python and PowerShell to prevent garbled output on non-ASCII characters.
  • Delegates UX verification tasks to a subagent.
  • Preserves delegated tool approval policies across agents.
  • Preserves worker identity and proxied execution mode.
  • Preserves child identity across Hub event projections.

November 4

  • Long conversations now recover from oversized requests by trimming historical reasoning and retrying.
  • Rewinding and resubmitting a prompt now preserves image and file attachments.
  • Messages referencing a path with @path while the agent is working are no longer rejected when an earlier message is pending.
  • copilot config now warns when an invalid value type causes settings.json to be ignored instead of treating all settings as unset.
  • Markdown output no longer converts colon-delimited strings such as file paths and code identifiers into emoji.
  • Run the CLI as an Agent Client Protocol (ACP) server over stdio using agy acp, enabling editor integrations with session modes, session resumption, and terminal Google sign-in via agy acp --login.
  • Run conversations on remote agents using the new --remote-agent flag, supporting remote file searches in @-mentions and project selection via --project.
  • Manage and switch between local projects using the /projects command.
  • Allow /btw side questions to execute read and search tools live while automatically denying any tool calls that require user permission.
  • Ensure headless -p runs without --print-timeout wait for background tasks to finish rather than exiting after a 30-minute limit.
  • Windows local model downloads now install llama.cpp automatically and display engine installation progress.
  • Model loading now fails faster on unreachable network endpoints and provides connectivity guidance.
  • Absolute file paths pasted into messages and multiple pasted files sharing the same name from different folders are now attached correctly.
  • Hook configuration files saved with a UTF-8 byte-order mark are now loaded instead of being skipped.
  • The /fork command no longer fails when a session database cannot be read, creating the fork without the todo list and avoiding incomplete session directories.
  • Fixed --continue losing previous conversation history after retry attempts for a malformed tool call ran out.
  • Fixed permission prompts being inadvertently answered by keys typed immediately after sending a message or before a slow prompt rendered.
  • Fixed macOS sessions stalling and consuming excessive memory when running many Bash commands.
  • Improved configuration, setting, and login persistence on shared network home directories to prevent dropped settings or file corruption.
  • Stopped automatically installing or updating the IDE extension when launched inside an IDE terminal, requiring the /ide command or editor extension view instead.
  • Add the /code command to open the current session in VS Code.
  • Allow the --yolo and --allow-all flags to take effect when resuming a local session with --resume or --continue.
  • Enable background Autopilot sessions to auto-answer ask_user prompts without requiring the session to be foregrounded.
  • Render inline images in Rex terminals using Kitty direct placements.
  • Sanitize control characters and escape sequences across diffs, footers, approval prompts, and tool error messages.

November 3

  • /settings now opens a unified dashboard for all users with global search across settings and local resources.
  • The interactive CLI reports program status to supporting terminals, including notifications when attention is needed.
  • MCP content addressed only to the user is no longer sent to the model, and displays images inline or as saved files.
  • SQL tools now abort long-running queries after 60 seconds instead of hanging the session.
  • New sessions no longer display startup tips under the header.
  • Added a confirmation prompt before a web or shell tool call sends your account email address to the network.
  • Added onFailure: "block" for Agent SDK hook callbacks so a failing callback blocks the action instead of allowing it to proceed.
  • Added a warning to /schedule when a new cloud scheduled task receives unrequested connectors, including instructions on how to remove them.
  • Fixed Ctrl+C stopping all running background agents when pressed to clear typed text at the prompt.
  • Removed the Read tool's stub message for unchanged files so re-reading a file returns its content again.
  • File mentions written inside parentheses, brackets, braces, quotes, or backticks now attach the target file and expand in shell commands instead of being silently ignored.
  • Expanding a tool call with Ctrl+O now displays the full contents of object and array arguments rather than placeholder tokens.
  • Remote HTTP MCP servers now support connecting through proxies that require Kerberos (Negotiate) authentication.
  • A preToolUse hook returning "modifiedArgs": null now leaves tool arguments unchanged instead of failing with a JSON parsing error.
  • Command hooks now receive COPILOT_HOOK_DEPTH to allow hook-launched sessions to prevent recursive agentStop loops.
  • Setting includeCoAuthoredBy: false is now honored in ACP mode (copilot --acp), preventing the agent from adding a Co-authored-by trailer to commits.
  • Paths starting with ~/ in skillDirectories, COPILOT_SKILLS_DIRS, and COPILOT_CUSTOM_INSTRUCTIONS_DIRS now resolve to your home directory.
  • Gemini models automatically defer tools behind search when requests exceed the 512-function limit.
  • In --autopilot, background task completions that arrive after a run stops are delivered with your next message instead of dropped.
  • Command-mode drafts are now preserved after accepted commands so you can continue typing immediately.

November 2

  • Add opt-in session management tools and @mentions for opted-in local sessions.
  • Assistant replies, rendered markdown, file paths, and tool outputs no longer pass terminal escape sequences and control characters that could restyle or corrupt the interface.
  • /add-dir and /cd now accept directory paths wrapped in quotes.
  • Resuming a remote session locally no longer requires a clean working tree, preserving uncommitted changes unless a checkout conflict occurs.
  • Valid saved MCP OAuth credentials are now reused for anonymous connections and initial challenges without requiring another sign-in.
  • Extension packages can now display changelogs on the first launch after an update, with support for the collapseChangelog setting to condense notices.
  • Streaming performance is improved by parsing tool-call deltas incrementally and reusing syntax highlighting and rendered Markdown across redraws.
  • Added support for custom self-update commands through PI_SELF_UPDATE_COMMAND and the global selfUpdateCommand setting when running pi update.
  • Added GPT-6.1 Terra support across OpenAI, Azure OpenAI Responses, and OpenAI Codex, and updated the default OpenAI model to gpt-6.1-sol.
  • Tool argument validation now uses @earendil-works/pi-schema, providing uniform coercion for MCP and JSON Schema tools while dropping TypeBox-specific extensions outside JSON Schema.

November 1

  • Running with --screen-reader now prefixes failed or rejected tool calls, errors like unknown commands, and warnings with "Error:" or "Warning:".
  • The /resume command now foregrounds an active background session instead of duplicating it on exit.
  • Unresponsive HTTP MCP servers that accept connections now delay startup only once for their connection timeout instead of three times.
  • Saving MCP server modifications preserves symlinked mcp-config.json files and updates the symlink target.
  • The /skills add, remove, and reload commands now stay in sync with session skills and immediately report refresh failures.

October 31

  • Multiline text pasted after /clear, /new, /ask, /delegate, or /diagnose now preserves its line breaks and indentation instead of collapsing into a single line.
  • apply_patch preserves original line endings for unchanged lines in files with mixed LF and CRLF endings.
  • COPILOT_PROVIDER_GHES_HOST now accepts full browser URLs such as https://ghes.example.com/ without creating malformed URLs, and provides clear errors for invalid host values.
  • web_fetch now detects declared character encodings from Content-Type headers or meta tags to prevent character corruption on non-UTF-8 pages.
  • Reopening a recently closed local session no longer stalls for up to 30 seconds.
  • Added CLAUDE_CODE_DISABLE_ULTRATHINK to turn off the ultrathink keyword and CLAUDE_CODE_SUBAGENT_EFFORT to set default effort levels for subagents.
  • Fixed background shell commands remaining running on macOS and Linux when exiting automatically due to refused credentials or archived sessions.
  • Fixed prompts containing invisible characters requiring a second Enter press to send.
  • Fixed the Bash tool garbling or losing functions starting with a dash as well as multiline aliases.
  • Added support for signing in to a Claude apps gateway directly from the VS Code chat panel.

October 30

  • New keyboard shortcuts allow choosing a worktree (⌘⌥O / Ctrl+Alt+O), choosing its base branch (⌘⌥B / Ctrl+Alt+B), and toggling shell mode in the composer (⌘⇧X / Ctrl+Shift+X).
  • The compaction.buffer setting in opencode.json now accepts a percentage of the context window, such as "20%", alongside absolute token counts.
  • Quotes and comments can be sent directly from the quote popup by holding ⌘ or Ctrl, or by pressing ⌘Enter or Ctrl+Enter.
  • Tool calls to bash automatically run as shell with the same permissions instead of failing when no plugin tool named bash is present.
  • MCP servers sending stdio messages larger than 10 MiB no longer trigger disconnects, and large message processing speed is improved.
  • Added "optional": true for API key providers in a plugin's auth.json, allowing commands to run unauthenticated requests at free rate limits instead of failing with 401 errors when no key is set.
  • Fixed Markdown tables in long streaming responses leaving misaligned borders and repeated text in terminal scrollback by holding rows in the live view until the table completes.
  • Desktop terminals now persist across server restarts with scrollback intact, stay synchronized with other windows and the TUI, and retain custom names.
  • Setting "experimental": { "artifacts": true } in opencode.json allows agents to publish, update, and serve HTML or static artifacts.
  • Selecting Claude Agent in the provider picker now provides guided installation, authentication, and connection verification.
  • Added a Choose base branch shortcut (⌘⌥. on macOS, Ctrl+Alt+. on Windows and Linux) and command palette action during worktree creation.
  • User, project, and remote configuration can now override Console-supplied provider names, limits, variants, settings, and headers.
  • Added a built-in Claude Agent provider in /connect that uses local Claude Code authentication and models, with an optional OPENCODE_CLAUDE_PATH executable override.
  • Reduced rendering overhead and improved responsiveness during streaming replies and long conversations in the Desktop app.
  • Fixed multiple Desktop bugs, including unclosable side-panel tabs, repeated "Session not found" toasts, and opening side-panel files over HTTP.
  • Split managed policy rules so integration.use no longer applies to MCP servers, skills, or plugins, introducing distinct mcp.use, skill.use, and plugin.use actions.
  • Added a spellcheck toggle under Settings → General in Desktop with system-language dictionary support on Windows and Linux.
  • Users can now choose between a Gemini subscription and Pay As You Go billing after signing in with Google Cloud.
  • PreInvocation hooks can now return terminationBehavior: "terminate" to cancel turns before invoking the model and display an errorMessage without sending it to the model.
  • Runs using --output-format json or stream-json now include a finished_reason field when a turn ends early due to print_timeout, interrupted, or error.
  • The /plugin Discover tab now limits long shelves to six items by default, adding Ctrl+E to toggle expansion and [ and ] to jump between shelves.
  • Fixed agy plugin import claude to properly detect marketplace plugins, retain remote HTTP MCP servers, and resolve ${CLAUDE_PLUGIN_ROOT} paths.
  • Added the permissions.blockWritesOutsideWorkingDirectories setting so Write, Edit, and NotebookEdit refuse paths outside the working directory in all permission modes.
  • Added idle compaction in VS Code and Claude Desktop to summarize conversations over 200K tokens left idle after prompt cache expiration.
  • Fixed claude -p exiting with code 0 and silent failures on startup errors; it now outputs the error and exits with code 1.
  • Fixed repeated session failures when an organization admin caps effort level; Claude Code now automatically lowers effort and retries.
  • Fixed session failures and refusals across Amazon Bedrock, Microsoft Foundry, and custom gateways by adding fallbacks for missing deployments, unsupported beta headers, and image-heavy contexts.

October 29

  • Enabled experimental_image_generate by default when an image-capable provider is connected, with model selection configured through experimental.image_generate.model.
  • Clicking the Git branch on the new-session screen now opens the project's worktree picker directly.
  • Unified timeline quotes, terminal quotes, and browser annotations around a shared optional-comment editor.
  • Restored classic question cards by default in the desktop app and placed the redesigned dock under Settings → Experimental → New question UI.
  • Preserved media attachments across session forks and imports.
  • Run /goal <objective> in the TUI or Desktop to let sessions continue executing and reviewing progress until an objective is met.
  • Enable experimental.image_generate to let agents create and edit images with the new image_generate tool.
  • Inspect live output and stop background shells directly from the Desktop activity pill.
  • Search favorites, recent models, providers, and variants in Desktop using the redesigned default model picker.
  • Add /reload in Desktop to refresh configuration without restarting the application.
  • Agent reply streaming now scrolls lines into terminal scrollback instead of repainting the entire screen and batches wrapped text, significantly reducing terminal output over slow SSH connections.
  • Images in conversations are now capped at 2000 pixels per side, large PNGs are converted to JPEGs when smaller, and mismatched image format labels are tolerated to prevent exceeding model request size limits.
  • About 100 MB of startup memory is automatically returned to the operating system 15 seconds after opening the interactive UI.
  • Context compaction and /compact now preserve a subagent's parent task, parent ID, and the parent's 10 most recent messages.
  • The /browser subagent now properly receives page snapshots when targeting isolated browser contexts or Incognito windows.

October 28

  • Added Pyrefly as an optional built-in Python language server enabled through the language_servers setting.
  • Debugger hovers now include an expandable, keyboard-accessible variable tree for inspecting nested values.
  • Added DWARF debugging to the JavaScript debugger to support WebAssembly sources.
  • Added Display P3 color space support on macOS using the use_display_p3 setting.
  • Fixed a bug where discarding unstaged changes in the Git Panel also discarded staged changes.
  • Removed support for macOS Catalina and Big Sur, raising the minimum macOS requirement to 12.0 (Monterey).
  • Removed agent checkpoints and the "Restore Checkpoint" button to prevent slowdowns and hangs in large Git repositories.
  • AI commit message generation now references the three most recent repository commit messages as style examples.
  • Added an option to open README files by default by setting on_new_window to "readme".
  • Added syntax highlighting for Go 1.27 generic methods.
  • Terminal image pasting is now supported via standard shortcuts like Cmd+V or Ctrl+Shift+V, dragging files into the prompt, and over SSH with X11 or Wayland forwarding.
  • The /plugin Discover tab now groups marketplace plugins into category sections rather than a single flat list.
  • Approving generic Google domains in execute_url permissions no longer auto-approves sensitive services such as Gmail, Docs, Drive, or Cloud Console.
  • Fixed always-allow permission rules for multi-line commands failing to apply and repeatedly prompting for confirmation.
  • Fixed conversation failure loops caused by corrupted image headers, unmatched tool calls after history trimming, and outdated thinking blocks.

October 27

  • Large CLAUDE.md, rules, and included files can no longer be accidentally overwritten after partial reads, as Write now requires reading the full file or editing in place.
  • Unrecognized model names and deployment ARNs on Bedrock, Vertex, Foundry, and Mantle now use the shorter system prompt by default unless CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=0 is set.
  • Clicking the "to send now" hint under a queued message in fullscreen mode now sends it immediately.
  • Added the oidc.user_id_claim setting to the Claude apps gateway to support custom id_token claims such as Microsoft Entra ID's oid.
  • Prompt and agent hooks are now limited to a maximum of 24 concurrent executions per session, plus up to 12 hooks waiting past their timeout.
  • /diff now loads faster in large repositories by rendering tracked changes and recent turns immediately while loading untracked files and older turns in the background, skipping loading untracked files over 1 MB into memory.
  • The Per Turn tab in /diff adds [ and ] shortcuts to jump to the first file of the previous or next turn, and key-hint footers now wrap on narrow terminals.
  • Failed, canceled, or nonexistent tool calls are now returned to the agent as errors rather than standard results, allowing the agent to recover on the next step.
  • Deleting a conversation now stops and cleans up running subagents, history, and workspaces, and forked conversations or /btw side questions no longer inherit active /goal states.
  • agy plugin install outputs installed component types and target paths, while agy plugin validate checks author, homepage, and repository fields in plugin.json.
  • Assisted Permissions are now available without enabling experimental mode.
  • Interactive and SDK Autopilot sessions continue waiting for attached background tasks instead of stopping after COPILOT_TASK_WAIT_TIMEOUT_SECONDS, though -p and --autopilot still honor the timeout.
  • Sessions using the managed model "fusion" now start on Auto and apply fusionTier settings.
  • Auto falls back to its default preference instead of failing each turn when a saved preference is no longer available.

October 26

  • Standard-context sessions that exceed 200K tokens now compact the conversation and continue rather than halting with a context credit error.
  • Fixed freezes at startup and during file read/edit operations when CLAUDE.md contains a very long list item.
  • Added --hold-idle-file and --max-hold-min flags to self-hosted-runner to maintain idle sessions while host processes continue running.
  • The ultrathink keyword in the interactive terminal now sets the turn's reasoning effort level to xhigh on supported 5.5 and 5.1 models.
  • The VS Code extension now provides syntax highlighting for code blocks in replies across light, dark, and high contrast themes.
  • Scaffold and test agents locally with an interactive playground using `npx @cursor/agent-sdk init` and `agent-sdk dev`, with starter templates available via `--template`.
  • Deploy agents to hosted infrastructure with `agent-sdk deploy` to receive a live URL and manage secrets and integrations from the dashboard.
  • Structure agents as code directories with `instructions.md` for prompts, `tools/` for typed actions, `skills/` for on-demand procedures, and `mcp-connections/` for Model Context Protocol servers.
  • Trigger agent runs via GitHub, GitLab, Bitbucket, Slack, custom webhooks, or cron schedules while preserving persistent conversation sessions across thread follow-ups.
  • Run automated local and CI tests using `agent-sdk eval`, or simulate live pull request webhook payloads with `agent-sdk github replay`.
  • Settings > API Configuration adds a Local | Cloud toggle to configure models for cloud sessions separately from local models, and cloud sandboxes now display a progress card during startup.
  • Starting a command no longer moves keyboard focus to the terminal, preventing chat input from accidentally typing into the shell.
  • Toggling Plan/Act mode while a task starts no longer drops the initial prompt or leaves the session configured with the wrong mode's tools.
  • Requests for GPT-6 models on OpenAI and Azure OpenAI endpoints now send max_completion_tokens instead of max_tokens, and the gemini provider no longer sends native Google Search requests that triggered 429 errors.
  • Refreshed the model catalog by adding CommandCode, Reflection, Liquid Inference, and Munito AI providers, updating default models across providers, and dropping CrofAI, iFlow, and Llama.
  • Added a Changes button and pane in the session header to review file diffs and line counts across the entire session or within the latest turn.
  • Added a project files side panel with a file tree, tabbed viewer, git status indicators, and file-versus-diff toggles for local and SSH workspaces.
  • Added an integrated terminal panel at the bottom of the window, toggled with Ctrl+`.
  • Added a Local Models settings page to manage Ollama and LM Studio models, local storage, and real-time model memory usage.
  • Added configurable keybindings in Settings, including an optional hold duration for quitting the app on macOS.
  • Opening a task already running in another Cline app now attaches to the existing run instead of launching a duplicate instance.
  • Running cline hub status and cline doctor now times out after 3 seconds instead of hanging if the Hub is frozen.
  • Turns that drop connections after streaming only reasoning are now retried automatically, while the maximum retry attempts for model requests has decreased from 5 to 3.
  • The gemini provider no longer sends native Google Search requests, resolving 429 quota errors on free-tier keys and dropped tools on Gemini 2.x.
  • Refreshed the model catalog by adding Liquid Inference and Munito AI providers, updating default models across 56 providers, and switching ClinePass default to Kimi K3.1 Pro.
  • Starting a session that already has a run in progress now reattaches to the live session instead of replacing its runtime, and rejects new prompts while a run is active.
  • Post-compaction size estimation now evaluates messages as actually sent with truncated tool results, preventing repeated compaction loops caused by earlier overestimates.
  • Checkpoint diffs now read file contents through a single git cat-file --batch process instead of spawning up to two git show processes per file.
  • The gemini provider no longer sends native Google Search to prevent 429 quota errors on Gemini 3+ and dropped tool calls on Gemini 2.x.
  • SSH remotes now support ProxyJump aliases and increase the connect budget to 30s, command timeout to 60s, and tunnel readiness limit to 40s.

October 24

  • Auto mode asks for approval less often on routine tasks such as cleaning up temporary files, stopping local processes, and running code formatters.
  • Hook commands now receive lastUserInput in their stdin JSON payload to inspect the latest user prompt.
  • Headless -p runs use significantly less CPU by skipping search index generation, transcript rendering, and custom status or title commands.
  • On-demand MCP tools now return output payloads up to 10 KB directly to the agent instead of offloading responses over 4 KB to disk.
  • Fixed skills, rules, custom agents, plugins, and hooks failing to load when authenticating with Application Default Credentials (AGY_ADC_AUTH).
  • Added the Monitor tool and plugin monitors on Bedrock, Vertex, and Foundry, as well as when telemetry or nonessential traffic is disabled.
  • Retries requests automatically without the effort-2025-11-24 beta header or mid-conversation system messages when behind gateways or servers that reject them.
  • Fixed compaction retaining too much dense text like base64 or SVG after a "Prompt is too long" error.
  • Fixed background cleanup deleting recent files such as pasted images when cleanupPeriodDays was set to a large value.
  • Added Claude Code cloud sessions and Remote Control permissions to Enterprise custom roles.
  • Interactive and SDK Autopilot sessions now continue waiting for attached background tasks instead of stopping after COPILOT_TASK_WAIT_TIMEOUT_SECONDS, while -p and --autopilot still honor the timeout.
  • Managed model "fusion" now starts sessions on Auto and applies fusionTier settings.

October 23

  • Assisted Permissions are now available without enabling experimental mode.
  • Auto now falls back to its default preference instead of failing each turn when a saved preference is no longer available.
  • Installed plugins in /plugin now include a "Configure MCP servers..." option to update server variables at any time without disabling and re-enabling the plugin.
  • Terminal sandboxes now block access to ~/.docker, ~/.kube, and ~/_netrc by default, requiring explicit permissions to use Docker registry or Kubernetes cluster credentials.
  • Context compaction, including /compact, now preserves attached images, PDFs, and files in conversation summaries.
  • Allowing a permission prompt for a conversation now applies across both the main agent and all subagents without prompting again.
  • The /diff command now highlights changed words within modified lines across all color schemes.

October 22

  • Fixed multiple sandbox and permission vulnerabilities, including sandboxed Bash commands executing startup files outside the sandbox and macOS denyWrite globs failing on symlinked folders.
  • Changed sed -i to run directly via standard Bash prompts instead of simulating a file-edit preview, removing it from /rewind restores.
  • Updated claude ultrareview to require workspace trust, prompting in terminals or requiring --trust-workspace / CLAUDE_CODE_DISABLE_ULTRAREVIEW_TRUST_CHECK=1 in automated scripts.
  • Decreased the post-edit diagnostics wait time in the VS Code chat panel from up to 1.5 seconds down to 0.75 seconds for faster write and edit workflows.
  • Fixed empty "Untitled" sessions cluttering VS Code session history when starting a session without sending a message.
  • Clicking a completed response code block now copies its original code when enabled.
  • ACP sessions honor saved sandbox settings and the --sandbox and --no-sandbox flags, including on session resume.
  • Legacy --remote and --no-remote startup flags are retired in favor of updated hosting and connection flags.
  • The sessions browser now displays a session's recorded title instead of a generic Session N fallback when available.
  • Cloud replies keep full text when reconnecting after a dropped connection.
  • Pausing screen refreshes when nothing changes roughly halves idle CPU usage, alongside reduced CPU use while streaming responses and faster loading for long conversations.
  • Recurring cron schedules created via the schedule tool now run at a consistent per-schedule offset within their period instead of precisely on the cron boundary to prevent traffic spikes.
  • Collapsed thought blocks now display only their header once a response begins and are separated from the response text by a blank line.
  • Search with / in the /diff view now counts matches inside folded unchanged lines, and the Commit Tree view adopts standard selection cursors.
  • Fixed the agent silently and infinitely retrying when responses are blocked by content safety filters, and resolved plugin loading failures when skills is defined as a single path string in plugin.json.
  • Clicking a completed response code block copies its original code when enabled.
  • Remote launch options now use clearer hosting and connection flags, retiring the legacy --remote and --no-remote startup flags.
  • Cached local MCP tool snapshots are now reused across registered Git worktrees within the same repository.
  • ACP sessions now honor saved sandbox settings and --sandbox or --no-sandbox flags, including when resuming sessions.
  • The sessions browser now displays a session's recorded title instead of a generic fallback.

October 21

  • Compaction and branch summaries now use the model's full output limit instead of a fixed cap, preventing failures when reasoning uses most of the token budget.
  • Codemode scripts and subagents now queue for up to 60 seconds when workers are busy instead of failing immediately, with the worker limit configurable via PI_CODEMODE_MAX_WORKERS.
  • Codemode models.classify() now accepts image inputs for TypeSafe's Jev and Cloudflare's Clef classifier models.
  • Added Rex terminal support with OSC 8 hyperlinks, truecolor, and inline images.
  • Updated default models for GitHub Copilot (gpt-6.1-sol), NVIDIA (nvidia/nemotron-3.5-super-vl-120b-a12b), and Xiaomi (mimo-v2.6-pro).
  • Tool calls that fail to parse are now retried up to twice before ending the turn with an error.
  • Added /claude-api api-version to update Claude API code to version 2026-10-21.
  • Interrupted HTTP MCP tool calls are now only retried if they are read-only or idempotent, avoiding duplicate executions.
  • Fixed requests failing with a model-not-found error when CLAUDE_CODE_DISABLE_1M_CONTEXT is set with [1m] model aliases such as opus[1m].
  • Reduced binary size by approximately 33% on macOS, 28% on Windows, and 7% on Linux.
  • Added support for applying inlay hint text edits by double-clicking a hint or running the editor: accept inlay hint command.
  • Folded code sections now automatically unfold when navigating to locations via go to definitions, symbol pickers, search results, or navigation history.
  • Added the tab_switcher.show_file_preview setting to control editor file previews while cycling tabs, and stopped the tab switcher from changing selection on mouse hover.
  • Fixed a security issue where environment variables and secrets passed to external agents and MCP servers in remote projects appeared in the local machine's process list.
  • Added Pyrefly as an optional built-in Python language server configurable through the language_servers setting.
  • Added an expandable, keyboard-accessible variable tree to debugger hovers and DWARF debugging for WebAssembly in the JavaScript debugger.
  • Added macOS support for the Display P3 color space via the use_display_p3 setting.
  • Added an extension option to project_panel.sort_order to group files by file extension.
  • Fixed discarding unstaged changes in the Git Panel also discarding staged changes.
  • Prompt mode now honors managed Auto model selection for -p and piped stdin runs.
  • Subagents now inherit custom and MCP tool approvals to prevent duplicate permission prompts.
  • Long-running sandboxed commands now continue running reliably beyond 10 seconds.
  • Replies interrupted by provider overload errors now retry with backoff and display the provider message and request ID.
  • Sandboxed network traffic now respects NO_PROXY and ALL_PROXY when inheriting host proxy settings.
  • Added Auto mode on Windows 11 version 24H2 and later for personal Google Accounts, selectable via --mode auto, Agent Mode in /config, Shift+Tab, or Remote Control Settings.
  • Added whole-file toggling (f) and hunk folding (Ctrl+O) to Commit Tree and Per Turn views in /diff, and consolidated multiple per-turn edits into a single combined diff.
  • Added _meta["antigravity.google/root_conversation_id"] to MCP tools/call requests so MCP servers receive the top-level conversation ID during nested subagent calls.
  • Fixed subagents losing plugins inherited from the parent session, as well as duplicate loading of rules and SKILL.md skills across standalone customization folders and plugins.
  • Fixed prompt editing glitches where Ctrl+Z undo corrupted pasted text or emojis, the prompt box remained mid-screen after closing panels, and blank gaps appeared in Medium verbosity mode.
  • Prompt mode now honors managed Auto model selection for -p and piped stdin runs.
  • Subagents now honor custom and MCP tool approvals, preventing duplicate permission prompts and allowing inherited workflows to run.
  • Fixed API errors being incorrectly replaced with "prompt is too long" and triggering unnecessary compaction in sessions with many MCP tools.
  • Fixed claude -p inside cloud sessions reporting "Not logged in" when started without session environment variables.
  • Sandboxed network traffic now honors NO_PROXY and ALL_PROXY when inheriting host proxy settings.
  • Long-running sandboxed commands continue running reliably past 10 seconds.
  • Replies cut off by provider overload errors now retry with backoff and display the provider's message and request ID.
  • Organization plugin preparation no longer blocks sending chat messages.
  • Plugin marketplaces now accept git-subdir sources and long plugin descriptions.

October 20

  • Fixed requests failing on Vertex, Bedrock, and other providers or gateways that reject or omit tool search by automatically retrying without it.
  • Added sandbox.allowNestedUserNamespaces to allow sandboxed commands on Linux, such as Chromium or rootless podman, to create user namespaces.
  • Added CLAUDE_CODE_RETRY_WATCHDOG_USAGE_LIMIT_RECHECK_MS so unattended retry mode notices when a usage limit is lifted before its scheduled reset.
  • Fixed /effort and /autocompact settings failing to apply when a model name differed from modelOverrides by letter case or a [1m] tag.
  • Improved /usage warnings to show one-line fixes naming their source, along with an f shortcut to apply them.
  • Dynamic workflow subagents now support native structured output on compatible models.
  • The new assistedApprovalInstructions setting guides Assisted Permissions confirmation decisions.
  • Prompts submitted without a selected model at startup no longer hang silently, prompting instead for a choice via /model while keeping entered text.
  • Shell commands that create and immediately enter a directory now request path approval instead of being denied early.
  • Plugins assigned to teams can now be installed from team-defined GitHub marketplaces.
  • Resuming a session now displays the conversation sooner without waiting for placeholder environment setup.
  • Permission approval prompts now explain escalation causes and display a request-specific diagnosis command.
  • Starting a session with sandboxed MCP servers avoids unnecessary restarts and responds faster.
  • The /config and /settings commands can now set and unset showReasoning.
  • Shell commands that create and then enter a directory now request path approval instead of being denied early.
  • Added `/` search to the `/diff` file list and file view with live match highlighting, match counts, and `n`/`N` navigation across files.
  • Expanded `agy plugin validate` to inspect hooks, sidecars, and component lists, reporting errors for invalid regex matchers, unparsed JSON files, and missing paths.
  • Fixed subagents and custom agents in Auto mode prompting for every command that leaves the terminal sandbox by routing them to an automated reviewer first.
  • Fixed cursor positioning and line wrapping issues caused by emoji, CJK characters, combining marks, terminal resizes, and Vim visual mode.
  • Fixed empty file creation failing for files like `__init__.py` or `.gitkeep`, and prevented failed `PostToolUse` hooks from marking successful tool calls as failed.

October 19

  • Added /auto-mode-setup, a guided setup that analyzes your project and recent sessions to configure auto mode for your environment, including on Bedrock, Vertex, and Foundry.
  • Added a warning on Pro and Max plans when resuming an inactive conversation would consume a significant portion of your 5-hour usage limit.
  • Fixed claude --worktree removing the worktree and its branch without asking at exit when HEAD was detached or on another branch.
  • Changed tool search to be enabled by default behind a custom ANTHROPIC_BASE_URL, with automatic retries without it if the gateway returns HTTP 400 or 422.
  • Changed background sub-agents in VS Code to show as a single status line under their Agent card instead of streaming each step into the chat panel.
  • The first prompt no longer hangs silently when starting without a selected model, prompting to select one via /model while keeping the entered text.
  • Switching from GPT models to Gemini no longer fails when earlier turns include custom tool calls.
  • Added native structured output for dynamic workflow subagents on supported models.
  • Added the assistedApprovalInstructions setting to guide Assisted Permissions confirmation decisions.
  • MCP PNG images with ICC color profiles are now preserved instead of omitted.
  • Pass `--disable-client-telemetry` to turn off client telemetry uploads, or `--trust-folder` for one-off prompt-mode sessions started with `-p` or piped input.
  • Instruction files located in `.github/instructions` are now applied automatically when reading or editing matching files.
  • Semantic code search is now enabled by default for online sessions.
  • Sandboxed environments now allow local network access by default so commands can reach localhost and private-network services.
  • Install verified skills and plugins directly from GitHub install links using `copilot plugin install "ghapp://air/install?..."`.

October 17

  • The command agy plugin validate now verifies plugin skills, subagents, and rules for invalid frontmatter, missing properties, size limits, and naming conflicts.
  • Inline diff previews for file edits and creation now match the /diff view with syntax highlighting, color-scheme-aware background tints, and untinted line-number gutters.
  • Socket disconnects and network drops now retry automatically instead of immediately terminating the turn with an execution error.
  • Windows systems lacking terminal sandbox support now fall back to asking before running commands rather than failing to start them.
  • Rendering performance and CPU usage while typing or running agent turns are reduced by batching prompt draws and caching line measurements across frames.
  • Fixed requests to GPT-6 models (such as gpt-6-sol and gpt-6.1-sol) on OpenAI and Azure OpenAI-compatible endpoints by sending max_completion_tokens instead of max_tokens.
  • Restored Claude Sonnet 5.5 as the default model and top recommended option for the Cline provider.
  • Added the CommandCode and Reflection providers to the model catalog.
  • Updated default models across multiple providers, including 302.AI (Gemini 4 Argon), Eden AI and Prime Inference (Muse Spark Open), CoreWeave (Reflection Beam 1.0), Pioneer (Nemotron 3.5 Super VL), Scaleway (DeepSeek V4.1 Flash), and NanoGPT.
  • Refreshed the model catalog to 7,041 models, added CommandCode and Reflection providers, returned the default model for the cline provider to Claude Sonnet 5.5, and updated defaults across several providers.
  • OpenAI-compatible requests for GPT-6 through GPT-9 model IDs now transmit max_completion_tokens instead of max_tokens to prevent OpenAI and Azure API rejections.
  • The hub daemon now records client launcher identity via CLINE_HUB_LAUNCHER_CLIENT, CLINE_HUB_LAUNCHER_VERSION, and CLINE_HUB_LAUNCHER_TRANSPORT and publishes them as OpenTelemetry resource attributes.
  • Telemetry self-check reports now include traces_exporter.
  • The unified Sessions browser is now used by default for /sessions, /session, and /resume.
  • Session browsing now caps account discovery at 500 records and respects Retry-After headers for automatic refresh intervals.
  • Sandbox host-rule traffic now respects HTTPS_PROXY and HTTP_PROXY environment variables when no sandbox proxy is configured.
  • Voice dictation no longer inserts language markers into transcript text.

October 16

  • Fork subagents (subagent_type: "fork") are now enabled by default in claude -p, the Agent SDK, the desktop app, and VS Code.
  • The CLAUDE_CODE_AUTO_MODE_MODEL environment variable configures which model the client-side auto mode classifier uses.
  • An Idle compaction setting in /config allows turning idle conversation compaction off or back on.
  • The Read tool now refuses image sources larger than 128 MB instead of loading them entirely into memory.
  • Startup time is reduced when multiple plugins are enabled by decreasing file system calls on plugin commands and agent files.
  • Cloud agents can now generate interactive canvases that can be shared with team members using a link.
  • Cloud agent environments now support up to 300 repositories and load large environments five times faster.
  • The Environments API is available in beta for programmatically creating and managing cloud agent environments.
  • Sessions no longer stall for minutes when an MCP server hangs while loading tools.
  • A prompt now asks to update the sandbox proxy certificate authority when masked hosts change.
  • Managed plugins now update their activation state in existing sessions when managed policies change.
  • Gemini tool calls no longer fail when array parameters define enum values.
  • Gemini chats no longer fail when prior tool results contain no readable content.
  • Added the `--disable-client-telemetry` flag to disable client telemetry uploads directly from the CLI.
  • Sandboxed shell commands now explain credential-proxy certificate trust failures and can offer an approved rerun outside the sandbox.
  • Saved `/diff` review comments are now retained when reopening an unchanged diff in the same session, alongside an explicit discard action.
  • Reasoning-only max-token responses no longer loop indefinitely and now terminate with an explicit error when no progress is made.
  • Plugin enable, disable, update, and uninstall actions now route reliably after plugin inventory refreshes.
  • Model reasoning options can now be configured to generate reasoning variants, replacing catalog variants.
  • Configured variant overrides are preserved when generating reasoning variants.
  • Terminal status reporting over OSC 7501 now updates supported terminal tabs or windows with working, waiting, done, or failed states, configurable via AGY_CLI_OSC7501 and AGY_CLI_OSC7501_MESSAGES.
  • The alt+m shortcut (Option+m on macOS) sends a message using the opposite of the current Queued Messages configuration.
  • agy plugin validate now applies runtime loading rules, reports line-level errors and warnings, and adds --strict and --json flags.
  • Startup eligibility checks now retry transient service errors and rate limits instead of locking the session with Eligibility check failed.
  • Plugin MCP server headers now expand environment variables such as ${TOKEN} instead of sending literal placeholders.
  • Plugins now run in a dedicated worker process instead of inside the hub, and CLINE_PLUGIN_MODE=in_process is no longer supported.
  • Mid-stream response interruptions caused by provider errors are now automatically retried up to three times before failing the turn.
  • cline hub ensure and cline hub start add --json, --allow-port-fallback, --no-connectors, and --login-shell-path flags.
  • The default fallback model for the Cline provider is now Gemini 4 Argon, alongside catalog updates that add Boundless and Standard Compute providers.
  • Failed model requests now log the provider request ID for easier troubleshooting.
  • Removed in_process plugin execution, requiring plugins to run in worker threads (CLINE_PLUGIN_MODE=worker) or the subprocess sandbox (sandbox).
  • Retried provider errors that occur after a stream has started up to three times without triggering whole-turn agent retries, and added provider request IDs to failure logs.
  • Updated the default model for the cline provider to Gemini 4 Argon and expanded the model catalog with Boundless and Standard Compute providers.
  • Added setHubDaemonLauncher and a beforeSpawn hook to run hub daemons through the bundled CLI, and removed the standalone SSH remote helper entrypoint.
  • Allowed AskQuestionExecutor to return text and image blocks so attachments sent with an answer are forwarded to the model.
  • Adds Cline Cloud sessions to run tasks in hosted sandboxes on GitHub repositories and branches, keeping tasks running after closing the editor.
  • Updates model catalog defaults across multiple providers, setting Gemini 4 Argon as the default for the Cline provider when no model is chosen.
  • Prevents long requests to local models in LM Studio and Ollama from timing out with BodyTimeoutError after 5 minutes.
  • Uses OpenAI Compatible base URLs exactly as entered without automatically appending /v1.
  • Fixes dropped image and file attachments when answering prompts, and adds automatic retries up to 3 times when a provider interrupts mid-stream.
  • Update checks now only look for the latest npm opencode-ai version to keep 1.x installs on the 1.x release line instead of offering 2.x builds as regular updates.
  • Upgrading to version 2 now requires an explicit command via the 2.0 offer in /update or opencode upgrade --major.
  • Fixed upgrades for installations managed through yarn.
  • Update checks now skip 1.x updates for Homebrew Core installs tracking version 2, and verify tap and Chocolatey package availability before offering updates.
  • The install script now refuses to install 2.x versions.
  • Semantic code search is now enabled by default for online sessions.
  • Pressing Ctrl+C twice now exits immediately even while the CLI interface is busy.
  • The interface displays a clear notice when administrator-managed settings override the selected model.
  • Native chats can now run and finish while optional managed plugin downloads proceed in the background.
  • Linux sandboxed commands start more reliably by avoiding intermittent loopback setup failures.

October 15

  • Added LaTeX math rendering for inline and block equations in VS Code responses.
  • Added CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS to stop unattended retry mode from stalling on distant usage limit resets.
  • Added the % motion to vim mode to jump between matching brackets.
  • Fixed quitting removing worktrees created by a WorktreeCreate hook without confirmation when they may contain unpushed commits.
  • Fixed the Read tool parsing binary file formats such as .xlsm, .parquet, .epub, and .heic as plain text.
  • Discontinues session sharing, making existing share links and session-sharing commands no longer available.
  • Removes support for importing sessions from share URLs.
  • Added the sessionShareDestination setting to route /share uploads to GitHub gist, Radius, or auto.
  • Switching models now drops thinking blocks from earlier models rather than sending them as plain text.
  • Resumed sessions now restore their previously active tools, with --tools +name/-name modifiers applied on top.
  • Changing the thinking level mid-session on OpenAI GPT-6 models now preserves the prompt cache.
  • Updated the kimi-coding provider to use the OpenAI Chat Completions endpoint, allowing tools added mid-conversation to preserve the prompt cache.
  • Added Claude Opus 5.5 and Claude Sonnet 5.5 support for Application Default Credentials (AGY_ADC_AUTH) sign-in, filtering them in /model based on Google Cloud project access.
  • Improved Auto mode terminal sandbox reviews to run faster without timing out, automatically permit workspace format, lint, and fix commands, and block sensitive data uploads.
  • Fixed agent hangs by canceling and retrying requests after 5 minutes of stalled data, and by setting GIT_EDITOR, HGEDITOR, and JJ_EDITOR to false so commands requiring an interactive editor fail immediately instead of waiting.
  • Fixed lost subagent and workflow questions by showing an alert with the question and enabling a question picker via Ctrl+K or a.
  • Improved the /diff Commit Tree view with distinct colors for commit hashes, change IDs, descriptions, and the uncommitted changes working copy row.
  • Added the --trust-folder flag for one-off prompt-mode sessions started with -p or piped stdin.
  • Added support for installing verified Skills and plugins directly from GitHub install links via copilot plugin install "ghapp://air/install?...".
  • Improved Microsoft Entra /login by allowing users to select a cached account or sign in with a different account.
  • Allowed sandboxed git on github.com to use the configured GITHUB_TOKEN from credentials.envVars.
  • Fixed custom agents declaring search, web, todo, or github-mcp-server/web_search tool aliases so they correctly receive matching tools.
  • Running /share without arguments opens a full-screen action menu for sharing the active session.
  • Sandboxes allow local network access by default so commands can reach localhost and private-network services.
  • Path-specific instruction files in .github/instructions apply automatically when reading or editing matching files.
  • Trusted repository plugins now install and update automatically before chat runs.
  • The built-in browser plugin is renamed to browser-use while maintaining existing user preferences.

October 14

  • Added support for Gemini 4 Argon.
  • Added Ctrl+E to expand /ask answers, Ctrl+Y to copy the full answer, and /ask or /btw to reopen the session's last usable completed answer.
  • MCP status and /mcp commands now stay synchronized with the active session across split-view switches.
  • Characters typed during startup are now preserved instead of being dropped while the input box loads.
  • Resuming a session now drops abandoned pending-permission requests and continues only active ones.
  • Added the /advisor command on Claude Platform on AWS, configurable through advisorModel, --advisor, or CLAUDE_CODE_DISABLE_ADVISOR_TOOL=1.
  • Added new mod features including virtualText prompt decorations, $.ui.view for viewing agent transcripts, right-to-left text support, and prompt.change and agent.spawn hooks.
  • Added OpenTelemetry GenAI semantic-convention attributes and separated cache creation token metrics by 5-minute and 1-hour lifetimes.
  • Fixed multiple --worktree issues involving reftable ref storage, bare repository fallbacks, and plugin lifecycle hooks.
  • Added the claudeCode.attachSelection setting in VS Code to stop text selected in the editor from being automatically attached to messages.
  • Added a `git: create tag at head` command and a `Create Tag…` context menu action in Git history and Git Graph to create local tags.
  • Terminal tabs can now be dragged directly into the Agent Panel.
  • The File Finder now automatically pre-fills its search query with text selected in the editor or terminal, which can be disabled with `file_finder.prefill_query_from_selection`.
  • Added a side-by-side comparison view for directory diffs opened with `zed --diff`.
  • Added Gitiles and self-hosted Gerrit support for permalinks and Git blame diff navigation.
  • Folded code sections now unfold automatically when jumping to a location via symbol navigation, outline pickers, navigation history, or go-to commands.
  • Inlay hint text edits can now be applied by double-clicking the hint or running editor: accept inlay hint with the cursor on it.
  • Binary sizes are reduced by roughly 33% on macOS, 28% on Windows, and 7% on Linux.
  • The tab_switcher.show_file_preview setting controls whether files preview in the editor while cycling tabs, and the tab switcher no longer changes selection on hover.
  • Fixed a security issue where environment variables and secrets passed to external agents and MCP servers in remote projects appeared in the local process list.
  • Expand /ask answers with Ctrl+E, copy full answers with Ctrl+Y, and reopen the session's last usable completed answer using /ask or /btw.
  • Command hooks now complete reliably when child-exit signals are lost under Node.
  • Added Auto mode (`--mode auto`, `/config`, or `Shift+Tab`) on Linux and macOS, which approves file edits and safe commands automatically while prompting before irreversible actions and reviewing unsandboxed commands with a separate agent.
  • Added the `/compact` command to summarize the conversation on demand and free context window space.
  • Added syntax highlighting, background tinting, and folding controls (`Ctrl+O` to fold/unfold and `f` to view the full file) to the `/diff` view.
  • Improved `/diff` performance by rendering only visible rows on screen, disabling fold expansion for files exceeding 100,000 lines or 10 MB.
  • Fixed color degradation to 256 colors in iTerm2 over SSH by enabling full 24-bit color when `LC_TERMINAL=iTerm2` is present.
  • Conversations with over 200,000 tokens left idle for roughly 54 minutes are now automatically compacted before prompt cache expiration, configurable via CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS or disabled with "idleCompaction": false.
  • Added the CLAUDE_CODE_DISABLE_FILE_MENTIONS environment variable, causing @ file mentions to pass file paths to file tools rather than injecting their raw contents.
  • Added /effort and /fast commands in the agent view to set defaults for new sessions, and entering bare /model, /effort, or /fast commands now opens an interactive picker.
  • Fixed multiple sandbox vulnerabilities, including sandboxed commands writing to --plugin-dir folders and symlink traversals bypassing denied paths on macOS.
  • Resume commands printed at exit now identify sessions by ID instead of session name to prevent unintended execution when pasted.
  • Adds support for Gemini 4 Argon.
  • Maintains MCP status and /mcp command synchronization with the active session across split-view session switches.
  • Resumes sessions by continuing only active permission requests while respecting abandoned pending-permission work.
  • Preserves characters typed while the input box loads during startup instead of dropping them.

October 13

  • Inline images now render by default on supported terminals and can be turned off with /settings inlineImages false.
  • Custom agents now support agent-scoped declarative hooks.
  • Invalid entries in config.json no longer trigger a full reset of saved settings, trusted folders, and sign-ins.
  • Emergency memory compaction in long sessions can bypass cooldowns to prevent JavaScript heap out-of-memory crashes.
  • Skills and custom agents loaded via --plugin-dir now reliably appear in autocomplete and invocation.
  • Introduced the App Marketplace featuring over 20 integrations across deployments, CI, code review, security, monitoring, and project management.
  • Opened the marketplace to public submissions, allowing developers to build apps using the API, authenticate via JWT and access tokens, handle webhooks, and submit them for review.
  • Added an Apps plugin that answers questions using live API documentation and helps plan migrations for existing apps.
  • Remote plugin marketplace and plugin fetches now retry once using your signed-in GitHub token if HTTPS authentication fails.
  • Streamed responses behind TLS-intercepting proxies now complete properly when close_notify is omitted.
  • Fixed parsing of GitHub SCP remotes containing absolute paths so repository owner and name resolve accurately.
  • A new outline along the left edge of conversations displays sent prompts, allowing you to hover for a preview and click to jump directly to a prompt.
  • Plugins that fail to load or error during a session now display a status message naming the plugin and error in chat instead of silently losing tools.
  • Scheduled routines now reliably start queued runs, display errors for failed model calls, and execute only the latest missed run after downtime.
  • Disabling reasoning now works across all providers and no longer triggers thinking-budget errors on level-only Gemini models.
  • Fixed an issue where images inflated token estimates and caused auto-compaction to run prematurely.
  • Defaults to Claude Sonnet 5.5 when no model is chosen for the Cline and OpenRouter providers, and updates unpinned default models across numerous other providers.
  • Refreshes the model catalog with Claude Haiku 5.5, Step 5 Preview, and new providers including AI/ML API, LowRouter, and Prime Inference.
  • Fixes 400 errors when turning off reasoning across all providers for supported models, as well as thinking-budget errors on level-only Gemini models.
  • Prevents screenshots and other images from inflating token estimates and triggering premature auto-compaction.
  • Resolves extension host crashes caused by MCP servers or hooks exiting before consuming all input, and enables discovery of skills in .claude/skills.
  • Plugins now run directly inside the hub instead of a separate Node or Bun subprocess, removing the requirement for a JavaScript runtime on PATH unless CLINE_PLUGIN_MODE=sandbox is set.
  • The new cline hub plugins command reports plugin statuses and errors, and supports reloading plugins using --reload <plugin> without restarting the hub.
  • The --thinking flag now works in ACP mode (--acp), where reasoning was previously always disabled.
  • Provider API keys set via environment variables like OPENROUTER_API_KEY are now recognized at startup, bypassing the TUI setup prompt.
  • Screenshots and images no longer inflate token estimates, preventing auto-compaction from running prematurely.
  • Plugins load in-process through a single registry per process by default instead of a subprocess, removing the need for a runtime on PATH while preserving subprocess isolation with CLINE_PLUGIN_MODE=sandbox.
  • Disabling reasoning now works across all providers instead of a fixed list, with provider-specific adjustments for Gemini and Vertex models.
  • Token estimation now charges a flat 1,600 tokens per image payload instead of pricing base64 text tokens, preventing premature transcript auto-compaction.
  • The default model for CLINE_DEFAULT_MODEL_ID and the OpenRouter provider is now Claude Sonnet 5.5, with the Cline provider also defaulting to Claude Sonnet 5.5.
  • Skills authored in .claude/skills are now discovered and made accessible alongside native skills.
  • Specifying --model or /model without --effort now automatically selects the default thinking level instead of failing with an error.
  • agy plugin list adds a --json output flag and now lists all installed marketplace and imported plugins along with their version and enabled status.
  • Remote Control browser sessions now show CLI settings directly in the browser Settings modal and synchronize changes with settings.json.
  • Automatic retries after repeated model request failures now cap delays at 30 seconds instead of up to 4 minutes.
  • Passing an unknown name to --agent now exits with status 1 in headless runs and displays a warning in interactive mode instead of silently using the default agent.
  • Custom agent discovery now detects all agents in repositories that define more than 100 agents.
  • Custom agent files that are nested deeply within a repository are now properly discovered.
  • OpenRouter integrations now fetch and list only the models available to the configured key, including accurate context sizes, output limits, and pricing.
  • Official JSON configuration schemas are now published for settings, models, keybindings, and themes to enable editor validation and autocompletion.
  • Setting PI_CLIPBOARD_OSC52=1 sends copied text to the terminal via OSC 52 sequences when Pi cannot automatically detect a remote session or multiplexer.
  • Codemode tool declarations now surface input constraints and defaults, and tool rows display live elapsed and total execution times.
  • Extension runs initiated by pi.sendMessage now trigger before_agent_start events and mark the session as busy during startup.

October 12

  • Interactive sandbox settings now suggest possible environment secrets and allow adding masking hosts before saving.
  • The timeline now displays whether each permission decision was made by the user, Assisted Permissions, policy, or unattended fallback.
  • The `/add-dir` command now grants sandbox access to added directories for the active session.
  • Model IDs in `/model` and `/config model` are now case-insensitive and save as canonical IDs.
  • The `/user list` command now displays the authentication source for each signed-in account.
  • Inline images now render by default on supported terminals without needing experimental mode, and can be disabled with /settings inlineImages false.
  • Custom agents now support agent-scoped declarative hooks.
  • Corrupted or invalid entries in config.json no longer cause all saved settings, trusted folders, and sign-ins to reset on launch.
  • Long sessions can bypass memory relief cooldowns to compact history and prevent JavaScript heap out-of-memory crashes.
  • Resuming suspended sessions now reports interrupted tool calls, avoids replaying canceled writes, and preserves active plugin hooks properly.

October 11

  • Updates are no longer installed automatically in the TUI; use the /update command to install updates or upgrade to OpenCode 2.
  • Run opencode upgrade --major to upgrade to OpenCode 2.
  • OpenAI's ultrafast service tier is now allowed.
  • PTY connection origins and tickets are now validated before checking whether a PTY exists.

October 10

  • Added a pinned prompt header in no flickering mode that shows the active response's prompt at the top of the transcript and jumps back to it when clicked.
  • Changed the macOS terminal sandbox to block sandboxed commands from reading ~/Library/Application Support by default unless explicitly allowed.
  • Added support for a root-level agents.json file in plugins to declare subagents alongside skills.json and rules.json.
  • Added support for MCP servers to mark tools as always loaded using _meta: {"antigravity/alwaysLoad": true} in tools/list responses.
  • Fixed subagents configured with a branch or share workspace in multi-repository setups to use isolated repository copies rather than working in the real checkout.
  • Interactive sandbox settings now suggest possible environment secrets and allow adding masking hosts before saving.
  • The /allow-all command remains available during startup while enterprise policy is still resolving.

October 9

  • The Read tool adds an allow_large option that lets file reads exceed standard size limits in a single call when context space allows.
  • Subagent definitions and frontmatter now support autoCompactWindow to trigger auto-compaction earlier than the main conversation window.
  • The CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL environment variable allows running all workflow agents on a specified model while other subagents keep their defaults.
  • Default character limits for MCP tool descriptions and MCP server instructions have been increased from 2,048 to 4,096 characters.
  • Edit and NotebookEdit operations now refuse modifications to non-UTF-8 files rather than replacing non-ASCII characters.
  • Interactive sessions in git repositories reach the input prompt faster.
  • The timeline now displays whether you, Assisted Permissions, policy, or unattended fallback made each permission decision.
  • The /add-dir command now grants sandbox access to added directories for the active session.
  • The /user list command now shows the authentication source for each signed-in account.
  • Retry progress messages now display before recovered responses after transient model failures.
  • macOS now uses native Microsoft Entra broker authentication when available, with fallback to the browser.
  • The --context flag now applies to both new and resumed ACP sessions instead of silently using the default or previously saved context tier.
  • copilot config supports sandbox credential injectHosts keys, including shell completion in Bash, Zsh, and Fish.
  • Managed plugin setup retries hourly or following policy changes rather than on every message failure.
  • Supported sandbox credential injectHosts keys in copilot config.
  • Added key autocompletion for injectHosts settings across Bash, Zsh, and Fish shells.
  • Added the /plugin interactive manager and agy plugin shell command to browse, install, configure, and manage plugins bundling skills, MCP servers, subagents, rules, and hooks.
  • Changed --dangerously-skip-permissions to no longer auto-approve implementation plans in /plan mode, limiting the flag to skipping tool permission prompts.
  • Reduced the maximum scaling limit for attached or viewed images from 4096 to 2576 pixels per side to prevent Claude models from failing in image-heavy conversations.
  • Fixed Tab, Left Arrow, and Right Arrow in the /diff Commit Tree view moving the cursor instead of switching views, and added Shift+Tab to return to the previous view.
  • Fixed high CPU usage caused by scheduled wake-up condition timers and repeated UTF-8 rescanning during long-running shell command output.
  • The --context flag now applies to new and resumed ACP sessions instead of silently using the default or previously saved context tier.
  • Managed plugin setup now retries hourly or after policy changes instead of on every message failure.

October 8

  • Added onFailure: "block" for command and HTTP hooks to block an action instead of letting it through when a hook fails to start, times out, or exits unexpectedly.
  • Added Program Status Protocol (OSC 7501) support so compatible terminals can display whether the session is working, waiting on input, or finished.
  • Fixed terminal freezes triggered by syntax highlighting very long lines, long runs of blank lines, unclosed heredocs, or outputs spanning tens of thousands of lines.
  • Added a dedicated chat row in the VS Code extension for received files so they can be opened directly in the editor.
  • Fixed request failures on 1M context models when an upstream provider rejects the context-1m beta by automatically retrying without the flag.
  • Claude Haiku 5.5 is now available in model selection and --model completions.
  • Assisted permissions now send visible shell code directly to the permission judge to avoid unnecessary manual approval prompts.
  • copilot mcp add now recovers cleanly if MCP configuration initialization is interrupted.
  • MCP servers can now be enabled or disabled before server discovery without starting the servers.
  • Clicking a row in the Sessions sidebar now reliably switches sessions during split-view reconciliation.

October 2

  • Added the /compact slash command to summarize the conversation and free up context using the session's provider settings.
  • Made Connectors the default first tab in Customize with a new two-column searchable Browse catalog and dedicated search boxes for installed and browsable connectors.
  • Added brand logos to connectors across Customize and Marketplace rows.
  • Updated the Connectors tab to display explanatory states when unavailable, including a sign-in card when logged out and a beta rollout notice with a Check again button.
  • Agent teams no longer re-save the entire team state on every streamed chunk, preventing performance degradation and database bloat in ~/.cline/data/db/teams.db.
  • Existing team data is compacted automatically the first time it is opened.
  • Default models are updated across several providers, including 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).
  • Cloud sessions now support Cline Usage-Billing, ClinePass, and ClineFree models, and wait for the model to load before sending requests.
  • Unsent prompts are now saved as drafts when navigating away from a session, with New in the sidebar restoring the draft and Cmd/Ctrl+N creating a blank session.
  • Customize → Connectors now features search, a refresh button, an Installed section, and suggested connector packages with one-click installation.
  • Marketplace entries now indicate their item type to help differentiate between MCP servers and connectors with identical names, with category filters consolidated into one dropdown.
  • Default models have been updated 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).
  • Prevents agent team performance degradation by batching changed entity writes every ~300 ms, storing summaries instead of full transcripts in run records, capping team_events, and migrating SQLite storage to schema v2.
  • Failed team writes are now automatically retried instead of dropped, and SqliteTeamStore.vacuum() is available to reclaim freed disk space.
  • Standalone provider requests outside a session can now resolve surface headers, making sessionId optional in resolveProviderRequestHeaders and omitting empty X-Task-ID headers.
  • Updates default models 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).

October 1

  • The TUI now runs in fullscreen mode by default, which can be reverted by setting tuiMode to "regular" or passing --tui-mode regular.
  • Codemode prompt overhead has been reduced by approximately 40%, and error messages now provide actionable recovery hints.
  • Codemode scripts and extensions can now generate images using session credentials via models.generateImages().
  • Anthropic /login adds a copy-code authentication flow for headless environments where the browser runs on a separate machine.
  • The /login command adds top-level Radius authentication with automatic MCP server configuration in mcp.json.
  • Opus 4.7+ and Fable now default to a 1M context window on Bedrock, Vertex, Foundry, and the Claude apps gateway without requiring the [1m] suffix, configurable via CLAUDE_CODE_DISABLE_1M_CONTEXT=1.
  • Added Claude Mods to let plugins modify deeper behavior, along with an opt-in cc-plugin-you-should-know@builtin plugin that runs a side agent to flag missed issues.
  • Added an n:<text> filter to the agents view to match session names and tasks, including inside collapsed sections.
  • Added support for URL prompts from MCP servers implementing the 2025-11-25 protocol, such as for sign-in flows.
  • Added a "Run in background" action in VS Code to background running commands or sub-agents and view their output in agent map cards.
  • Added copilot sandbox ca commands to check, create, trust, rotate, and remove proxy CA trust, replacing /sandbox ca install with create and trust.
  • Allowed complete, statically analyzable read-only shell pipelines to enter execution-evidence review, while incomplete or unbound pipelines require explicit approval.
  • Enabled sandboxed commands to run on Windows versions lacking filesystem enumeration support, issuing a warning about potential PowerShell current location mismatches.
  • Added a sandbox network bypass prompt for Node/npm EACCES socket denials on Windows.
  • Fixed session timelines to properly clear busy status after interrupted turns finish.
  • Added `copilot sandbox ca` commands to check, create, trust, rotate, and remove proxy CA trust, replacing `/sandbox ca install` with separate `create` and `trust` commands.
  • Enabled sandboxed command execution on Windows versions lacking filesystem enumeration support, with a warning regarding potential PowerShell location inaccuracies.
  • Fixed session timelines to clear busy status after interrupted turns complete.
  • Fixed footer text selection to remain on the same visible line when the footer height changes.
  • CLI shutdown now flushes pending telemetry before exiting with a bounded delay if telemetry is still initializing.
  • macOS CLI release binaries are now signed with an official Developer ID.
  • Model requests now send namespaced session and parent-session identity headers.
  • Locally compiled binaries are re-signed to run reliably on macOS 27+.
  • Fixed plugin name extraction in the /status dialog by using path.sep.

September 30

  • Adds support for GPT-6.1 Sol in model selection.
  • Adds the --mcp-github-auth flag to scope GitHub account authentication to approved MCP server origins.
  • Adds session-scoped read-only directory approvals to path access prompts.
  • Allows MCP tools to recover after transient discovery failures without restarting the session.
  • Evaluates messages typed while the agent is actively working for auto-approval rules.
  • Complete, statically analyzable read-only shell pipelines can now enter execution-evidence review, while incomplete or unbound pipelines require explicit approval.
  • A sandbox network bypass option is now offered for Node/npm EACCES socket denials on Windows.
  • Oversized MCP and Composio tool results are now cached in memory up to 16 MiB per session and can be paged through by line range using read_files with cline://cache/... URIs, with custom tools opting in via resultPolicy: "cache-oversized" on createTool.
  • Custom providers defined in providers.json and models.json now work on the agent path instead of failing with an unknown provider error.
  • Provider settings now save reliably even if model discovery fails, serialize mutations, restore prior settings on failed catalog writes, and make saveLocalProviderSettings async.
  • The model catalog updates defaults for 302.AI, NanoGPT, Vivgrid, and Vultr, and notes renamed Vultr model IDs that may require re-selecting pinned models.
  • MCP servers with default codemode exposure now connect in the background without blocking the first prompt, and are discovered dynamically via searchTools() and describeNamespace().
  • Added an oauth.clientName setting for MCP servers and support for authenticating HTTP MCP servers using existing provider login tokens via "auth": { "provider": "<provider>" }.
  • Added Anthropic workload identity federation support configured through the ANTHROPIC_FEDERATION_RULE_ID, ANTHROPIC_ORGANIZATION_ID, and ANTHROPIC_IDENTITY_TOKEN_FILE environment variables.
  • The /reload command now activates tools newly added to the defaultTools setting during an active session.
  • Fixed prompt submission slowdowns in long sessions caused by redundant model catalog lookups per assistant message.
  • Resuming or continuing sessions with claude --resume or --continue no longer drops turns following a batch of parallel tool calls after a crash.
  • Permission prompts now display a progress count when multiple permission requests stack up.
  • The --bare flag now connects only explicitly named MCP servers, sends no system reminders, starts no background tasks, and terminates timed-out commands instead of moving them to the background.
  • Typing /compact, /clear, or /rewind while viewing a background agent's transcript now confirms the target via dialog instead of silently modifying the main conversation.
  • The VS Code extension adds response bookmarks in a side panel and displays previous question cards along with your selected answers directly in the conversation.
  • Added support for GPT-6.1 Sol in model selection.
  • Clicking anywhere on expanded tool calls in the compact timeline now collapses them.
  • Permission prompts remain answerable after resuming interrupted sessions.
  • Auto-approval now accounts for messages typed while the agent is actively working.
  • MCP tools now recover after transient discovery failures without requiring a session restart.
  • Disables WebKitGTK's DMA-BUF renderer automatically when proprietary NVIDIA drivers are detected on Linux to prevent blank app windows, with an override available via WEBKIT_DISABLE_DMABUF_RENDERER=0.
  • Adds app metadata to Linux .deb and .rpm packages so GNOME Software and KDE Discover display the icon, description, and installed status.
  • Added the "Queued Messages" setting to /config and settings.json ("queuedMessages": "send-immediately") to either hold follow-up messages until a turn ends or immediately interrupt the agent.
  • Allowed remote-control start to run as a background process on Linux environments lacking systemd, such as containers, with remote-control status showing its PID.
  • Accelerated loading and resuming long conversations with thousands of steps by no longer scanning every step up front.
  • Updated --json-schema to fail at startup with exit code 1 when given plain text, bare type names, missing files, or schemas without a root "type": "object".
  • Fixed agent hangs when viewing pipes, sockets, or device files, and prevented conversations from breaking on truncated video or audio recordings.
  • Linux on ARM is now supported with auto-updating arm64 .deb and .rpm packages.
  • Cloud agent sessions now retain their files, history, and reasoning settings when suspended and resumed.
  • Pressing Stop immediately after sending a prompt now cancels the turn even while it is still being initialized.
  • Web search settings moved to Customize → Tools, showing whether the active provider supports search.
  • Refreshed the model catalog defaults across multiple providers and added Bee and Pareto as providers.
  • Refreshed default models across numerous providers, setting GPT-6.1 Sol as the default for OpenAI, OpenRouter, GitHub Copilot, and several gateways.
  • Added support for two new providers: Bee (by HEOSSI) and Pareto Inference via PARETO_API_KEY.
  • Enabled typing and sending a new message or slash command (/compact, /smol, /newtask) following an API error instead of only offering a Retry button.
  • Improved Anthropic error handling with server-side refusal fallback, upstream provider failovers on OpenRouter and Cline, and explicit messages for content-filter blocks.
  • Fixed task cancellation so canceling immediately after sending a message reliably stops the task instead of running.
  • Requests for Anthropic models now configure API refusal fallbacks and provider fallbacks to automatically fail over to upstream providers instead of failing outright.
  • Fixed an issue where Stop commands issued while preparing a turn were dropped, which allowed the turn to keep running despite showing as stopped.
  • Fixed a regression from 0.0.87 that blocked resuming existing sessions with a duplicate-start error.
  • Content-filtered turns are now surfaced with a dedicated content-filter finish reason and an explanatory error message instead of appearing as empty responses.
  • Refreshed the model catalog with updated default models across multiple providers and added new model options.

September 29

  • Resolved an issue where "No supported model available" appeared on launch or in the model picker when a configured provider already supplied a model.
  • MCP tool calls now complete properly even when servers continue sending progress updates after responding.
  • Added the claude --desktop command to open the Claude desktop app on the current directory or a specific session.
  • Added the claude plugin configure <plugin> command to inspect and modify plugin configuration options.
  • Added the CLAUDE_CODE_DISABLE_WEB_FETCH environment variable to disable the WebFetch tool.
  • Changed background Bash and PowerShell commands to automatically terminate after a timeout limit, defaulting to 30 minutes with a 2-hour maximum.
  • Added an allowedProviders managed setting to restrict which API providers a machine can access.
  • Added built-in extensions for codemode QuickJS sandbox execution, dynamic tool search, and MCP server support managed through /mcp and pi mcp commands.
  • Made the new system theme the default, automatically deriving UI colors from the terminal's reported palette and light/dark modes.
  • Added interactive Sign in with ChatGPT to the OpenAI provider via /login openai to use ChatGPT subscriptions directly.
  • Added experimental virtual models via pi.registerVirtualModel(), enabling extensions to route requests dynamically to physical models and thinking levels.
  • Added classifier model support for all llama.cpp chat models alongside inherited Jev classifiers on OpenRouter, Cloudflare Workers AI, and Vercel AI Gateway.
  • Added support for GPT-6.1 Sol (gpt-6.1-sol) across OpenAI, Azure OpenAI Responses, and OpenAI Codex providers.
  • Changed the default OpenAI Codex model to GPT-6.1 Sol (gpt-6.1-sol).
  • Fixed an issue where /login with OpenAI failed in the bundled release due to a missing openai-chatgpt.js module error.
  • Adds the --mcp-github-auth flag to scope GitHub account authentication to approved MCP server origins.
  • Adds session-scoped read-only directory approvals to path access prompts.
  • Waits for the server-specified retry delay instead of a fixed 5 seconds on API rate limits, and stops immediately without retrying if the delay exceeds 30 seconds or if hitting a daily or billing quota cap.
  • Reduces CPU usage and memory churn during conversation view redraws, including during fast scrolling or while the agent is running.
  • Updates the artifact viewer's m shortcut hint to specify the content type and destination mode, such as diagram ASCII, diagram source, or math image.
  • Fixes text clipping on narrow terminals (~40 columns) across the workspace trust dialog, sign-in and MCP authentication screens, and the /help panel by wrapping lines and compacting the tab bar.
  • Fixed hub startup failures on Windows systems configured with proxies by automatically bypassing loopback addresses in NO_PROXY and no_proxy variables.
  • Corrected token usage calculations so reasoning tokens are no longer double-counted in session totals and telemetry alongside output tokens.
  • Added shared cloud handoff support to ./cloud for moving sessions between local and cloud agents, along with persistent cloud workspaces that can suspend and resume.
  • Expanded fallback codebase search to index PHP source and template files while excluding the Composer vendor directory.
  • Updated Amazon Bedrock integration to route OpenAI GPT model IDs through inference profiles and pass reasoning effort parameters properly.
  • MCP OAuth sign-in to servers such as Datadog now reuses still-valid cached tokens.
  • Withdrawn running prompts now remain removed after resuming a session.

September 28

  • Added Claude Sonnet 5.5 (claude-sonnet-5-5) as the default Sonnet model on the Anthropic API with a 1M context window.
  • Interactive terminal and VS Code sessions now start in auto mode by default across all plans and providers when no permission mode is configured.
  • Added the /mcp reconnect all command in the interactive terminal to retry all disconnected or unauthenticated MCP servers at once.
  • Converted Ultracode into an independent toggle in /effort that can remain active at any effort level rather than forcing xhigh effort.
  • Added a "Yes, but ask again next time" prompt option in auto mode before allowing a read outside working directories.
  • Add support for Claude Code rule files in .claude/rules as custom instructions.
  • Add GPT-6 Sol, GPT-6 Luna, and Claude Opus 5.5 to the model picker.
  • Pull request creation now follows repository pull request templates, preserving required sections and checklist structures.
  • Direct plugin installs can now be toggled with commands like copilot plugin enable.
  • Pressing Escape twice in an empty chat input during local sessions removes a prompt whose turn has not started answering yet.
  • Pull request creation now uses repository pull request templates to preserve required sections and checklist structures.
  • Automatic indexed search activation can now be configured using the TGREP_FILE_COUNT_THRESHOLD environment variable.
  • Tool filters using server/tool and server/* now match MCP tool names containing slashes.
  • Shell output no longer displays trailing command completion metadata.
  • Timeline output no longer auto-scrolls while text selection is active.
  • Cloudflare AI Gateway models now honor provider response and stream timeouts.
  • Gemini thinking defaults and effort options now match supported controls across model generations.
  • Debug configuration output now redacts credentials and sensitive headers.
  • MCP browser launch failures are now reported when the launcher exits immediately.

September 27

  • Bursts of mouse-wheel or trackpad events are now coalesced before redrawing in the full-screen view, reducing CPU use and scroll lag in long conversations and the artifact viewer.
  • Sessions using GEMINI_API_KEY now stop immediately when encountering exhausted daily quotas, spend caps, or depleted credits instead of retrying for several minutes.
  • Resuming long conversations in medium or low verbosity no longer re-animates earlier finished tool groups or lags on keypresses during transcript replay.
  • The CLI no longer sends image-support probes through GNU screen, tmux, or Zellij, preventing terminal and tab titles from corrupting on startup.
  • Vim mode fixes restore non-ASCII character support after r, f, F, t, and T, correct fv and fV behavior in Visual mode, and stop D, C, or Visual ~/u/U from creating empty undo steps when nothing changes.
  • Added support for Claude Code rule files in .claude/rules as custom instructions.
  • Sidebar sessions now display a blue dot when they finish a turn that has not been opened yet.
  • Left-clicking supported ask_user and elicitation form inputs now focuses them and sets the cursor at the clicked position.
  • Fixed Git failures on Git 2.36 or newer in applications launched from CLI shells when empty environment variables are dropped.
  • Fixed extensions failing to load under enterprise managed settings while managed MCP policy is being applied.

September 26

  • Settings adds an About page that displays the current version and channel, provides update and restart controls, links to recent release notes and changelogs, and includes an issue reporter.
  • A one-time What's new dialog now displays after updates to summarize recent features, and can be revisited using the Highlights button on the About page.
  • Amazon Bedrock requests to OpenAI models through inference profiles no longer fail with an unknown reasoningConfig parameter error when reasoning effort is set.
  • On macOS, the Help menu adds an Export Diagnostics… item to open diagnostic export directly.

September 25

  • Added availableModelsMatch and deniedModels managed settings to control exact model version matching and explicitly block specified models.
  • Added /doctor prompt-audit (and /checkup prompt-audit) to scan CLAUDE.md files, skills, agents, and custom commands for prompting patterns written for older models.
  • Added x-claude-code-prompt-id to gateway hint headers when CLAUDE_CODE_GATEWAY_HINT_HEADERS=1 is set.
  • Added MCP tool, WebFetch, and WebSearch outputs to OpenTelemetry span events when OTEL_LOG_TOOL_CONTENT=1 is enabled.
  • Reverted the reservation of the claude-ai namespace, allowing skills, commands, workflows, and MCP server items using the prefix to load again.
  • Auto mode now suggests a routing tier that can be switched with a shortcut or click, and prompts for feedback after switching away to a manual model.
  • Direct plugin installs can now be enabled and disabled, with disabled plugins stopped from loading and re-enabled via copilot plugin enable.
  • Auto tier selection in /model drops the unsupported Fast profile, automatically falling back to Balance.
  • Wrapped commands such as timeout 60 gh ... now run correctly inside sandboxed sessions.
  • Gemini models no longer fail with a 400 invalid request body error when an MCP server uses anyOf alongside type or properties.
  • Local connection checks now bypass system HTTP(S) proxies, preventing startup failures caused by routing internal hub connections through proxies.
  • The main window now automatically shrinks to fit the screen and opens centered on small or scaled displays.
  • Plugin slash commands now load when a workspace opens and retry in the background, preventing them from disappearing from the menu or sending as plain text.
  • The settings tab is renamed to Providers, and long model lists now scroll in place inside the panel.
  • OpenAI models on Amazon Bedrock no longer require cross-region inference, and India regions (ap-south-1/2) now use the in. inference profile.
  • Improved reasoning effort level selection for supported models via the --effort flag or through the /effort and /model commands.
  • Fixed Mermaid diagrams and LaTeX rendering garbled characters in WezTerm and the VS Code integrated terminal by falling back to ASCII instead of assuming Kitty graphics support.
  • Fixed project custom agents in .agents/agents/ not being found or selectable in Desktop App workspaces, trusted workspaces, headless runs with -p or --prompt, --agent executions, and the /agents panel.
  • Fixed artifact viewer text copying when Copy on Select is disabled in /config, releasing mouse capture so native terminal selection works and Ctrl+C interrupts or exits.
  • Changed manually added plugins in ~/.gemini/config/plugins to start disabled if their MCP server requires configuration variables.

September 24

  • Added a maxProseWidth setting that caps the width of prose in wide terminals while keeping tables and code blocks at full width.
  • Changed auto mode to use the server-side classifier by default on direct Anthropic API connections when telemetry is off, with CLAUDE_CODE_AUTO_MODE_SERVER=0 available to opt out.
  • Added the allowClaudeInChromeWithManagedMcp managed setting to allow claude --chrome to run alongside an exclusive managed-mcp.json.
  • Fixed multiple session resumption and continuation bugs that caused API errors or dropped earlier extended thinking blocks.
  • Fixed requests failing with 400 errors when conversation history contained web search results from third-party gateways that the API could not decrypt.
  • Pressing Esc Esc in an empty chat input during local sessions takes back an unstarted prompt and removes it from the conversation.
  • Listing shells now displays running shells first instead of burying them behind finished ones.
  • Sandboxed commands on supported Windows versions can access localhost when Local network access is turned on.
  • Model IDs are now autocompleted in /model and /model plan argument pickers.
  • MCP configuration loading now preserves valid workspace servers when sibling entries are invalid.
  • Added support for ai&, an OpenAI-compatible endpoint offering open-weight models from Japan.
  • Refreshed the model catalog, changing the default model for 19 unpinned providers to Claude Opus 5.5 (including GitHub Copilot and Vertex).
  • Improved local model handling (llama.cpp, Ollama, LM Studio) by compacting context and retrying once when responses hit token limits instead of terminating the task.
  • Reopening a failed task now correctly displays the error and retry option rather than marking the task as completed.
  • Fixed Windows @-mention file search rendering incorrect nested paths and duplicate open files, and stopped silent terminal commands from printing raw JSON blobs.
  • Added a medium verbosity mode to /config that groups related tool calls and thoughts into concise summaries while keeping commands and responses visible.
  • Updated terminal sandbox logic so sandboxed commands can read and write internal artifact and scratch directories without requesting a bypass, and always attempt sandboxed execution first.
  • Improved Mermaid diagram and LaTeX rendering in the artifact viewer on Kitty-compatible terminals with background pre-rendering, smaller image uploads, and flicker-free zooming.
  • Changed directory scanning in skills.json, rules.json, agents.json, and plugins.json to load only direct child items instead of scanning recursively unless specified in include_only.
  • Fixed headless runs (-p/--prompt) exiting with code 0 when encountering errors mid-stream; they now exit with code 3, print an AGY_ERROR line, and include partial responses in JSON output.
  • Local models running on llama.cpp, Ollama, and LM Studio no longer fail mid-answer when context runs low, automatically compacting conversations and retrying before preserving partial responses.
  • Yolo mode (-y) now prompts models with tighter output rules to reduce narration, shorten plans, skip preambles, and write edits directly into tool calls.
  • Error messages from failed runs are now retained in the session transcript when resuming, without being sent back to the model or counted toward compaction.
  • Hub startup failures now display the underlying error cause, and the launch timeout has been increased from 8 seconds to 15 seconds.
  • Added the ai& provider for OpenAI-compatible endpoints using the AIAND_API_KEY environment variable.
  • Text-only turns truncated at the output token limit now attempt context compaction before falling back to nudge-and-retry recovery, preventing truncated responses on local servers such as Ollama and LM Studio.
  • Hub startup failures now display the underlying cause rather than a generic error message, and the client startup timeout is extended from 8 to 15 seconds.
  • Session renames performed through the hub now persist across relaunches without discarding other session metadata like pinned state.
  • Aborting a request now cancels immediately during empty-response backoff intervals instead of waiting for the delay to elapse.
  • Added built-in support for the aiand provider and updated the yolo-mode prompt to encourage shorter plans and fewer preambles.

September 23

  • Pressing Send Now (Ctrl+Enter or Ctrl+X Ctrl+S) now moves running tools to the background instead of cancelling the turn.
  • Added "attribution": false in settings.json to hide all commit and PR attribution.
  • Fixed multiple session resumption bugs where reconnecting MCP servers, pending permission prompts, or interrupted tool calls could drop prior reasoning history or break the prompt cache.
  • Auto mode and --dangerously-skip-permissions now prompt before executing recursive rm commands targeting command-substitution outputs or sensitive shell variable paths.
  • Added support for IAM role assumption via STS and Bedrock guardrails on Claude apps gateway Bedrock upstreams.
  • Rollouts generates monitoring plans on pull requests, tracks telemetry across separate deploy environments, and flags regressions or opens revert PRs when issues occur.
  • Security Review analyzes pull requests in full repository context to detect vulnerabilities such as injections, auth bypasses, leaked secrets, SSRF, and unsafe deserialization with proposed fixes.
  • Custom team rules can be configured to enforce codebase-specific constraints during security reviews.
  • Both bots are available on Teams and Enterprise plans and can be enabled from the dashboard automations tab.
  • Adds GPT-6 Sol and GPT-6 Luna to the model picker when available.
  • Fixes the view tool to honor line ranges when providers send flattened view_range arguments.
  • Pressing Up in an empty chat input during local sessions now recalls the pending message and preserves queued prompts in the queue.
  • Improved rendering performance for syntax-highlighted files—especially with the minimap enabled—and Markdown code blocks across panels and previews.
  • Added the agent.prevent_idle_sleep setting, enabled by default, to prevent system sleep while agent threads are running.
  • Added a language server command picker, support for language servers to open files and URLs via showDocument requests, and direct action triggers from actionable inlay hints.
  • Added SuperGrok sign-in alongside BYOK support for Claude Opus 5.5, GPT-6 Astra, GPT-6 Sol, and GPT-6 Luna.
  • Added the comment_empty_lines parameter to the editor::ToggleComments action to control whether blank lines in multiline selections get commented.
  • Added BYOK support for Claude Opus 5.5 with an Anthropic API key, as well as GPT-6 Astra, Sol, and Luna with an OpenAI API key.
  • Added an optional model parameter to spawn_agent to override the configured or parent subagent model, and enabled compaction for subagents.
  • Changed the shortcut to close the active dock to ctrl-alt-w across all platforms so cmd-w and ctrl-w close the active pane item instead.
  • Added a unified/split toggle to the diff view opened with zed --diff <old> <new>.
  • Added support for "..." in settings such as edit_predictions.disabled_globs, terminal.path_hyperlink_regexes, and file_scan_inclusions to extend inherited configurations.
  • Added `@<subagent> <message>` prompt syntax to send messages directly to subagent conversations with autocomplete for running and completed subagents.
  • Added Vim numeric count multipliers to operators, motions, actions, and text objects in Normal and Visual modes.
  • Fixed headless runs (`-p`/`--prompt`) to properly terminate daemon background processes on exit and wait for background tasks until the `--print-timeout` deadline.
  • Fixed a database corruption risk where checking write access on conversation-history files silently dropped locks held by concurrent CLI processes.
  • Updated `/rewind` to show relative timestamps for each step and focus the most recent step when opened.

September 22

  • Sub-agents spawned in the same step now execute their tool calls concurrently rather than one after another.
  • Unsent text, quote context, images, and files in the composer are now preserved when triggering Retry.
  • UserPromptSubmit and TaskStart hooks now reliably inject context into task runs instead of dropping context modifications.
  • Output from background commands now streams into the UI in real time rather than appearing only after the command finishes.
  • Models with large output token limits now receive a default output budget of 30% of their maximum limit instead of a flat 32,000 tokens.
  • Adds support for the claude-opus-5.5 model.
  • Shows managed Connector consent progress alongside a copyable authorization URL during connect and reconnect.
  • Added support for Claude Opus 5.5, GPT-6 Sol, and GPT-6 Luna across GitHub Copilot, Anthropic (with adaptive thinking and a 1M context window), and OpenAI keys or Codex subscriptions.
  • Changed the default model for new xAI sessions to Grok 4.7.
  • Fixed missing or invalid --mode values being silently ignored instead of exiting with an error and a nonzero status.
  • Fixed split-turn compaction summaries being rejected by Claude Fable 5.1.
  • Fixed image-only user messages failing on some OpenAI-compatible providers due to an included empty text part.
  • Run /fork during active turns to branch work without waiting for completion.
  • Dismiss rows in the Sessions tab by pressing x twice to permanently delete local sessions or close server-backed sessions.
  • Enterprise managed settings now apply to sessions opened in ACP mode (copilot --acp), by AHP hosts (copilot --ahp-host), and via --server.
  • Pressing Enter in freeform ask_user prompts now inserts a newline, while Ctrl+Enter or Ctrl+S submits.
  • Optional OSC 777 terminal notifications are now supported for direct Ghostty and WezTerm sessions.
  • Added support for Claude Opus 5.5 as the default Opus model, and switched the default model on Pro and Team Standard plans from Sonnet to Opus.
  • Added the CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH environment variable to adjust the 2,048-character cap on MCP tool descriptions and server instructions across a session.
  • Prevented auto mode, allow rules, and acceptEdits from approving writes through symlinks that resolve outside the project tree.
  • Added interactive dialogs in the VS Code extension for /status, /sandbox, /chrome, /export, /skills, and /plan mode.
  • Fixed multiple terminal dialog bugs, including keybinding conflicts in text inputs and double Ctrl+C or Ctrl+D inputs quitting the CLI instead of closing the dialog.
  • Turns that hit the output-token limit without producing a usable tool call now automatically retry up to three times with a conciseness prompt instead of failing the run.
  • Default output-token allowances now scale up to 30% of a model's maximum limit for models supporting more than ~106,667 tokens when no explicit limit is configured.
  • Added cline-pass/mimo-v2.6-flash and cline-pass/mimo-v2.6-pro to the Cline Pass tier and expanded the catalog to 6,237 models.
  • Updated unpinned default models across 10 providers, moving several defaults to MiMo V2.6 Flash, Grok 4.7, DeepSeek V4.1 Flash, and Gemini 3.8 Flash.
  • Context compaction now allocates user-request budgets across all prompts and sizes summary budgets from the model's full context window rather than the compaction trigger threshold.
  • Custom models configured in settings.json now support PDF tool outputs and attachments by default and respect modelFeatures flags for media handling.
  • Pressing Ctrl+G to edit prompts in external terminal editors now runs against standard terminal output without hanging.
  • Voice dictation sessions automatically stop and finalize at 3 minutes 30 seconds to prevent deadline failures and lost transcriptions.
  • Exiting the CLI now cancels open streaming connections immediately instead of waiting for a 5-second forced timeout.
  • Context compaction now triggers based on actual provider token usage instead of character estimates and keeps current OAuth credentials and model selections during long sessions.
  • Subagents now run tool calls concurrently and no longer prompt for individual tool approvals once delegation is authorized.
  • Startup no longer blocks on feature-flag network requests, refreshing flags in the background instead.
  • On Windows, the working directory is no longer prioritized over PATH when spawning internal executables like rg, git, and powershell.
  • Non-interactive commands like cline history --json no longer initialize the full interactive runtime.
  • Subagent tool calls now run concurrently by default while sequential tools remain ordering boundaries.
  • Run-start hooks can now inject context into the initial model request via an append-context channel, with opt-in support for blocking execution via blockingRunStartHooks.
  • Context compaction now triggers using provider-reported input-token counts alongside character estimates, and the default summarizer output budget increases to 8192 tokens.
  • Windows process execution now disables current-directory executable searches to prevent untrusted binaries planted in a workspace from running during indexing.
  • Concurrent feature-flag polls for the same account now share a single in-flight request instead of returning default disabled values during cache races.
  • Preserves /allow-all permissions when managed-settings refresh failures occur.
  • Remembers exact session approvals for missing paths without granting their parent directory, exposing exact grants in /list-dirs and clearing them with /reset-allowed-tools.
  • Shows bypass guidance when sandboxed network denials occur from proxy tunnel failures.
  • Added Grok 4.7 support to Zen and Go.
  • Restricted Bedrock image attachment hoisting to Claude, Nova, and Llama 4 models.
  • Fixed streaming usage reporting for Together AI.

September 21

  • Run /fork during active turns to branch work without waiting for the turn to finish.
  • Pressing Enter in freeform ask_user prompts now inserts a newline, while Ctrl+Enter or Ctrl+S submits the input.
  • Add optional OSC 777 terminal notifications for direct Ghostty and WezTerm sessions.
  • In the Sessions tab, dismissable rows now require pressing x twice to confirm permanent deletion for local sessions or closing for server-backed sessions.
  • Enterprise managed settings now apply to sessions opened in ACP mode (copilot --acp), AHP host mode (copilot --ahp-host), and server mode (--server).
  • Consecutive steering prompts in the same mode combine into a single pending message that can be recalled for editing by pressing Up in an empty chat input.
  • The new worktreePathTemplate setting configures directory paths created by /worktree, /move, /new, and --worktree.
  • User and managed startup defaults can now enforce or set organization policies for the Auto routing tier.
  • Sandbox proxies now work on Windows, and authenticated proxies with usernames and passwords are supported across all platforms.
  • copilot mcp list and copilot mcp get now display the built-in github-mcp-server when authenticated.

September 20

  • Added Meta Muse provider support via /login meta with automatic API key refresh, as well as META_API_KEY support for Muse Spark models.
  • Enabled Node's persistent compile cache before loading the bundled CLI runtime to reduce repeat launch times.
  • Restored OSC 52 clipboard fallback when no display is available and added a Windows clipboard backend for WSL to fix copy failures in containers and WSL environments.
  • Fixed /bug descriptions dropping line breaks from pasted diagnostics, and stopped /bug hints from triggering on user cancellations and retryable provider failures.
  • Fixed inherited Cerebras models causing HTTP 400 errors when mixing strict and non-strict tools, and recognized inherited z.ai "Prompt too long" errors as context overflow.
  • Added cost-aware prompt cache warming to keep caches alive during long tool runs and idle periods.
  • Added the /bug command to submit bug reports containing redacted diagnostics, metadata, and optional session transcripts.
  • Preserved mid-conversation system prompt and tool updates across session resume and branch navigation while retaining cached prefixes.
  • Added per-model compaction overrides for reserveTokens and keepRecentTokens under compaction.modelOverrides.
  • Enabled strict-prefer JSON-schema sampling by default for built-in read, bash, powershell, edit, and write tools.

September 19

  • Defaults auto mode to the server-side classifier for Claude API and Enterprise users, as well as Bedrock, Vertex, Foundry, and gateways, eliminating charges for classifier overhead.
  • Supports opting out on Bedrock, Vertex, Foundry, and gateways by setting CLAUDE_CODE_AUTO_MODE_SERVER=0, and adds a warning when falling back to billed classification.
  • Adds an Auto mode server row to /status to show whether the session's auto mode classifier runs on the server.
  • Added inline Kitty graphics rendering for LaTeX equations and Mermaid diagrams in the artifact detail viewer, with ASCII fallback, horizontal panning, and an 'm' shortcut to cycle views.
  • Improved the ask_question prompt with arrow-key navigation between questions, inline previews for write-in answers, check marks for answered options, and single-step final submission.
  • Allocated a dedicated 20,000-token budget for user and workspace rules to prevent large rule sets from evicting skills, workflows, subagents, or MCP tools.
  • Capped model API retry backoff at 30 seconds per attempt rather than waiting up to 4 minutes.
  • Retired legacy find_by_name, grep_search, and list_dir tools from the default agent and subagent toolsets while keeping them available for custom configurations.

September 18

  • Consecutive steering prompts in the same mode combine into a single pending message that can be recalled and edited by pressing Up in an empty chat input.
  • A new worktreePathTemplate setting configures directory structures for worktrees created through /worktree, /move, /new, and --worktree.
  • User and managed startup defaults now support strict and overridable organization policies for the Auto routing tier.
  • Per-server MCP slow-connection warning thresholds can be customized using the slowConnectionThresholdMs setting.
  • Sandbox proxies now work on Windows, and authenticated proxies with usernames and passwords work across all platforms.
  • Support for AGENTS.md is now available when a project lacks a CLAUDE.md file, configurable under "Project instructions" in /config.
  • Headless and SDK sessions (claude -p) now report errors and exit with code 1 rather than hanging indefinitely after an internal error.
  • Fixed conversations failing every request with a "text content blocks must be non-empty" error when an earlier turn contained an empty text block.
  • Invisible Unicode formatting and tag characters in prompts are now stripped and displayed for review before the prompt is sent.
  • The deprecated TaskOutput tool has been removed in favor of reading background task output files with Read.
  • Added Remote Control support via the --remote-control startup flag or /remote-control slash command to follow and control an active terminal session from another device.
  • Changed the default timeout for headless (-p / --prompt) runs from 5 minutes to unlimited unless --print-timeout is passed explicitly, and enabled daemon background commands to stay running after headless GEMINI_API_KEY turns.
  • Improved headless error reporting on agent or model API failures to print a structured AGY_ERROR JSON line to stderr and exit with code 3 instead of 1.
  • Updated prompt editor word-wise navigation and deletion shortcuts to stop at punctuation boundaries instead of only whitespace.
  • Fixed /model <name> <prompt> commands submitted during an active turn so the model switch waits until the queued prompt starts executing instead of switching mid-turn.
  • Fixed a regression from 2.1.275 where all requests failed with a 400 error mentioning input tag advisor_20260301.
  • Restored support for proxy and gateway endpoints configured through ANTHROPIC_BASE_URL.
  • Fixed a startup failure where the app would crash before reaching the workspace picker with "desktop backend exited before publishing its endpoint" or "No compatible hub runtime is available" errors.
  • Synchronized runtime code between the app and the background Hub to eliminate build identity mismatches during launch.

September 17

  • Added a send-now key combination (ctrl+enter or ctrl+x ctrl+s) that interrupts the active turn and immediately sends all queued messages.
  • Terminal sessions automatically sync skills and plugins enabled on your claude.ai account, with opt-out available via syncClaudeAiSkills: false or syncClaudeAiPlugins: false.
  • Added /plugin install <plugin> --marketplace <source>, which offers to register a marketplace source before installing a plugin.
  • Plugins installed from npm are now fetched with npm pack --ignore-scripts and integrity-verified, preventing package install scripts from executing.
  • The VS Code extension adds accept and reject buttons under individual changes in the proposed-diff view, memory management in the Memory dialog, and a claudeCode.scrollToBottomOnSend setting.
  • Custom agents can opt into repository instruction files like AGENTS.md, copilot-instructions.md, and CLAUDE.md by setting include-custom-instructions: true in their frontmatter.
  • Resuming an active session preserves marketplace plugins and skills across reloads, and configuration read or validation errors no longer discard active plugins.
  • Autopilot now reliably stops after an accepted task is completed instead of continuing unexpectedly.
  • Sessions can now resume even if transcript files contain recoverable corruption.
  • The status row now indicates it is waiting for background shells instead of displaying 'Working' when a turn ends with attached background processes still active.

End of the releases.

177 older releases are outside this window.