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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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 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.
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.
Compaction triggers on long tasks now use provider-reported token counts and double the summarizer's output budget to prevent running out of context on dense content.
Images attached to models that cannot read them now display a warning badge and an option to switch models instead of being silently discarded.
Opening a repository on Windows that contains executables like rg.exe or git.exe in the workspace directory no longer executes them in place of system tools.
Tasks encountering transient provider errors mid-stream now retry up to three times with backoff instead of failing the task.
Web search is now enabled by default outside YOLO mode for models that support it.
Custom agents defined in Markdown that list invoke_subagent now automatically receive the roster of available subagents and usage instructions in their system prompt.
Background tasks now retain their original names in completion notifications and task lists instead of reverting to generic auto-derived names.
Definitively rejected credentials are now automatically cleared to allow clean re-authentication without session loss from temporary network errors.
Commands terminated by outside signals are now marked as canceled instead of reporting exit code 0, preserving output collected before interruption.
Fixed in-progress thinking frames freezing permanently in conversation history after terminal resizing or session reloads.
Windows updates now use the Windows Restart Manager to close running backend processes, preventing "Error opening file for writing" installer failures.
Child processes on Windows now opt out of current-directory executable searching to prevent workspace binaries from running instead of system tools like git, rg, and powershell.
Session compaction now tracks provider-reported token counts alongside character estimates and doubles the summarizer's output budget to avoid context window exhaustion.
Fixed a bug where streaming tools continued displaying a running spinner after finishing when their completion event had no final output.
Updated app icons across the application and installer, refreshed CoreWeave documentation links, and added a direct link to the Weights & Biases API key page.
Sessions stuck in endless retry loops on "unexpected tool_use_id" 400 errors now self-heal corrupted transcripts where possible or exit cleanly with a /rewind suggestion.
A visible warning now alerts users when memory usage reaches critical levels, providing steps to safely free memory or restart.
/code-review now uses streamlined inline review prompts instead of spawning multiple subagents for models without tuned settings.
The new CLAUDE_CODE_MCP_STARTUP_WAIT_MS environment variable limits how long the first non-interactive turn waits for connecting MCP servers.
The VS Code extension automatically resumes steps interrupted by window reloads, which can be disabled via the Claude Code: Continue After Reload 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 after reload, and configuration read or validation errors no longer discard active plugins.
The status row now indicates it is waiting for background shells instead of displaying "Working" when a turn finishes while a background shell is still running.
/sandbox policy now reports local-network access using your configured setting.
Pressing Enter with an empty composer while the agent is running now steers the first queued message into the next turn boundary instead of waiting for the task to finish.
Steer failures now trigger an error toast notification rather than failing silently.
Mac Dock icons now include standard margins so the app icon size matches neighboring applications.
The session sidebar sort toggle now shows the icon for the mode it switches to rather than the currently active mode.
Improved editor rendering performance in syntax-highlighted files, especially when the minimap is enabled, and in 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 triggering language-server actions from inlay hints, and support for showDocument requests.
Added SuperGrok sign-in for using Grok models in the Agent Panel and added support for DeepSeek Flash 4.1.
Added shift-backspace to remove recent entries from the command palette alongside the command_palette.use_command_history setting to disable history-based ranking.
Sessions started with claude --remote-control or /remote-control can now be forked from the Claude app to run as background sessions on your computer.
Added optional LLM gateway request headers for request class, agent type, compaction state, and tool durations, enabled via CLAUDE_CODE_GATEWAY_HINT_HEADERS=1.
Fixed the context meter and auto-compact counting advisor-tool turns at roughly twice their real size, which previously triggered auto-compaction at half the actual window.
Improved responsiveness in long sessions by avoiding full-conversation reprocessing on every hook progress or sub-agent update.
Changed auto mode on Bedrock, Vertex, and Foundry to default to the local classifier unless CLAUDE_CODE_AUTO_MODE_SERVER=1 is set.
Transient provider errors that occur before any output streams are now retried up to three times with backoff instead of immediately failing the run.
Checkpoints now use a persistent per-session snapshot index to prevent message stalls on workspaces with large untracked directories.
Agent plugins under ~/.agents/plugins/* are now automatically discovered and managed through the Hub, exposing skills as plugin-name:skill-name without manual settings edits.
Web search is now enabled by default in non-yolo sessions on supported models.
The updated model catalog refreshes default models across 44 providers and adds Infer by Flow7, Melious, NaN, and Wallaby.
Agent plugins placed under ~/.agents/plugins/* on the hub host are now automatically discovered, validated, and loaded alongside their skills and MCP servers.
Model turns that fail mid-stream due to transient provider errors are now retried up to three times with exponential backoff before failing the run.
Added RemoteEnvironmentService to execute sessions, tools, and persistence remotely on an SSH host while keeping the client local.
Checkpoints now use a persistent per-session snapshot index to skip re-hashing unchanged untracked files on every turn.
Provider-native web search is now enabled by default in non-yolo sessions on supported provider and model combinations.
Added the /copy btw slash subcommand to copy the full text of the active /btw side-question response to the clipboard along with argument ghost hints when typing /copy.
Improved tool-calling fidelity by translating per-request tool-choice constraints (any, required, and named functions) into native backend function-calling configurations.
Fixed custom subagents with enable_mcp_tools: true receiving empty MCP server lists instead of inheriting parent servers, and fixed declarative subagents failing to resolve relative config paths against the parent agent directory.
Fixed the /hooks command and inspection utilities omitting hooks bundled inside enabled plugins when listing active hooks.json configurations.
Fixed proactive feedback prompts triggering prematurely during initial session turns or on transient, recovered tool errors.
Corrected the thinking configuration sent to adaptive-only Claude models so they remain adaptive instead of failing, lowering reasoning effort when thinking is turned off and capping reasoning effort at high when disabled.
Triggered sessionEnd hooks properly when using the /clear command to close an interactive session.
The /config command opens a sidebar configuration screen directly inside the CLI.
Worktree operations using /worktree, /move, and --worktree are now generally available without enabling experimental mode.
The /sandbox command now supports host allow and deny rules without replacing upstream proxies, as well as session-level disabling when permitted by organizational policy.
Windows executables now run without requiring the Visual C++ Redistributable.
Active scheduled prompts now appear in the CLI footer by default.
Raised the default open file descriptor soft limit at startup on Unix and macOS to prevent "Too many open files" errors in large workspaces.
Added the dev: Debug Filesystem Watching command to inspect local watcher events, watch roots, and scan exclusions, with support for exporting diagnostic captures as JSON.
Fixed an issue in SSH remote mode where reopening a newly created file at a previously renamed path could incorrectly reuse the buffer for the renamed file.
Fixed diagnostics batches stopping at paths without an associated worktree.
Fixed CRLF line breaks not being normalized in completion labels.
Replaced duplicate authentication failure messages with a single in-place error card featuring contextual fix buttons.
Fixed an issue where signing out of Cline or Cline Pass failed to persist due to credential re-importing from classic extension storage.
Restored OpenCode Go model support by sending the required x-opencode-session header and routing requests to model-specific endpoints.
Added Crusoe as a supported OpenAI-compatible provider.
Renamed the settings sidebar's Models section to API Providers and updated failed voice input to show an in-chat error toast instead of redirecting to settings.
view_file now rejects non-UTF-8 binary files and files larger than 100 MB with an error instead of loading them into the model context.
MCP servers bundled in plugins are now automatically namespaced as <plugin>_<server> to avoid name collisions with each other or mcp_config.json.
Startup warnings for deprecated unsandboxed permission rules now show affected file paths, up to five offending rules, and step-by-step migration instructions to command rules.
Fixed a memory leak caused by leftover background step-cache eviction goroutines when scanning or opening conversations.
Fixed Gemini API (GEMINI_API_KEY) sessions dropping model thinking blocks from prior turns and failing to propagate thought signatures.
Shell completions now generate from the CLI grammar, presenting root flags alongside subcommands and scoping options strictly to their respective subcommands.
Session and memory import commands are now available for the semantic JSONL interchange format.
Subagent launches now honor model, reasoning effort, and context tier preferences specified in global and custom instructions.
Commands failing with EPERM or EACCES errors on sandbox-blocked paths now offer an option to run outside the sandbox.
The /usage command now displays per-model AI Credit consumption in usage breakdowns.
Transient genai.APIError failures (502, 503, 504, 429 rate limits, and mid-stream interruptions) now automatically retry in-process with exponential backoff while preserving completed tool outputs.
Custom agent Markdown frontmatter now supports excludeDefaultComponents: true to opt out of default prompt sections and built-in tools while keeping post-invocation hooks.
Running with --continue now falls back to the most recent non-empty conversation in the current workspace hierarchy instead of creating a new conversation when launched from a subdirectory, after a crash, or alongside another session.
MCP and provider tool schema validation now preserves open object schemas such as {"type": "object"} or additionalProperties: true instead of rejecting undeclared arguments.
Fixed status line incorrectly reporting the sandbox as disabled when launched with the --sandbox flag, and fixed full-screen blank flashes in inline mode.
Added Projects in beta, accessible from the left-hand navigation to manage large, multi-task bodies of work.
Introduced coordinator agents that plan workflows, delegate implementation tasks to parallel subagents, and return finished work for review.
Moved execution to cloud agents that run continuously when the local machine is closed, with the ability to launch local agents when on-machine testing is required.
Maintained a shared context layer that synchronizes codebase knowledge, research, and workflow preferences across all cloud and local agents in a project.
Added subscriptions to trigger coordinator actions automatically based on schedules, pull requests, or connected Slack channels.