Google Antigravity

• Improved context compaction to spread its user-request budget across all captured prompts, so a long initial instruction is no longer truncated to a small fixed slice when the other prompts in the conversation are short
• Improved context compaction to size summary and truncation budgets from the model's full context window instead of the compaction trigger threshold, so compacted summaries and background-task lists are no longer prematurely cut short
• Fixed PDF and audio tool outputs and attachments failing with unsupported mime type errors on custom models configured in settings.json; custom models now accept PDFs by default and honor the modelFeatures media flags for images, video, PDF, and audio
• Fixed pressing Ctrl+G to edit the prompt in a full-screen terminal editor such as vim hanging with Vim: Warning: Output is not to a terminal; external editors now run against a real terminal stdout
• Fixed voice dictation sessions longer than 4 minutes failing with a deadline error and erasing the drafted transcription; recordings now automatically stop and finalize at 3 minutes 30 seconds

0.1.18 September 21, 2026 Local model support with LiteRT and LocalOpenAI, standardized evaluation presets, and custom subagent models This release announces official Antigravity SDK local model support with first-class LiteRTAgentConfig and LocalOpenAIAgentConfig configurations, introduces a standardized evaluation preset ( AgentConfig.eval() ) for core coding evaluations, enables custom subagent model overrides, turns on the schedule tool by default across standard tool groups, and pairs background task management with command execution. Improvements (13) Antigravity SDK + Local Models Support : Official local model support is now ready to use with first-class LiteRTAgentConfig and LocalOpenAIAgentConfig configurations. Developers can execute on-device models with automated lightweight presets or integrate with local OpenAI-compatible endpoints. Standardized Evaluation Preset ( AgentConfig.eval() ) : Adds a standardized, benchmark-ready preset that configures autonomous permissions, benchmark retry behavior, daemon execution in commands, and strips image generation and subagents to focus on core coding evaluations. Custom Subagent Models : Allows developers to assign specific

• Added inline Kitty graphics rendering for LaTeX math equations ($$...$$ and fenced math, latex, and tex blocks) and Mermaid flowcharts and sequence diagrams in the artifact detail viewer, with Unicode ASCII fallback, Left/Right arrow horizontal panning for wide diagrams, and m to cycle between Image, ASCII, and Raw views
• Improved the ask_question interactive prompt with Left/Right arrow navigation between questions, inline previews of confirmed write-in answers, check marks on answered options, and single-step submission on the final question
• Improved the CLI startup banner for enterprise accounts to display the active Google Cloud Project ID beneath the signed-in account and plan tier
• Improved model API retry responsiveness by capping per-attempt retry backoff at 30 seconds instead of waiting up to 4 minutes between attempts
• Improved customization token budgeting by giving user and workspace rules a dedicated 20,000-token budget (cutting oversized rules on newline boundaries and listing over-budget rules by path and description) so large rule sets no longer evict skills, workflows, subagents, or MCP tools

• Added Remote Control (start a connection via --remote-control startup flag or /remote-control slash command) to create a session-scoped remote connection for following and controlling your active terminal session from another device. Typing /remote-control off or closing the session automatically tears down the tunnel and unregisters the device from the active Remote Control session list.
• Changed the default timeout for headless (-p / --prompt) runs from 5 minutes to unlimited so long-running agent turns run until the response completes unless --print-timeout is passed explicitly, and enabled daemon background commands in headless GEMINI_API_KEY sessions so background servers stay running after the turn finishes.
• Improved headless (-p / --prompt) error reporting when a turn terminates on an agent or model API failure: the CLI now prints a structured AGY_ERROR: {...} JSON line on stderr with canonical status, HTTP or gRPC error code, retryability, and error ID (including HTTP status mapping for GEMINI_API_KEY SDK errors) and exits with code 3 instead of 1.

• Added automatic subagent guidance for custom agents defined in Markdown: agents that list invoke_subagent in their tools now receive the roster of available subagents and usage instructions in their system prompt, so they can actually delegate to the subagents they declared.
• Improved background task naming so a task keeps the name it was given when sent to the background; completion notifications and task lists now show that name instead of a generic auto-derived one.
• Fixed stale credentials lingering after the sign-in server definitively rejects them (for example a revoked or invalidated enterprise account); the CLI now signs you out and clears the stale tokens so you can sign in again cleanly, while temporary network failures, server errors, and rate limits never cost you your session.
• Fixed commands killed by an outside signal (for example when the whole session shuts down mid-run) being recorded as successful runs with exit code 0; such commands now finish as canceled while keeping the output collected before the interruption.

• Added a /skills reload subcommand to asynchronously reload discovered skills and slash commands without restarting the session or blocking user input.
• Added an inline argument autocomplete dropdown when typing /model and a live fuzzy search bar inside the interactive /model picker to filter models by name or ID.
• Improved authentication error messages when OAuth tokens lack required scopes by including explicit instructions to run /logout and /login.
• Fixed mid-conversation model switching to Gemini thinking models failing with thought_signature validation errors when prior turns were generated by non-Gemini models.
• Fixed web search (search_web) failing with "no summary returned from GenerateContent" when the summarization model emits leading thought parts.
• Fixed agent turns terminating prematurely with NO_TOOL_CALL when a tool call fails schema validation, allowing the agent loop to return the validation error and self-correct.
• Fixed MCP tool schema augmentation mutating tool definitions in place, preventing strict argument validation failures.

• Added the /copy btw slash subcommand to copy the full text of the active /btw side-question response to the clipboard even when the response card is collapsed or scrolled, 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 the /hooks command and hook inspection utilities omitting hooks bundled inside enabled plugins when listing active hooks.json configurations.
• Fixed custom subagents created with enable_mcp_tools: true receiving an empty MCP server list instead of inheriting the parent agent's configured MCP servers, and fixed declarative subagent configurations failing to resolve relative config file paths (relative_path_to_config) against the parent agent's directory.
• Fixed proactive feedback prompts triggering prematurely during the first few turns of a session or on transient, recovered tool errors.

0.1.17 September 14, 2026 Conversation compaction controls, forward-looking budget scopes, and tool output token truncation This release introduces first-class conversation compaction controls via CompactionConfig , delta and forward-looking budget scopes for session resumption, tool output token truncation limits, and expanded arithmetic operations on UsageMetadata . It also delivers OS-level terminal sandboxing examples and automated tool wrapper reflection preservation. Improvements (7) Compaction Configuration : Adds CompactionConfig on AgentConfig to govern sliding-window conversation history compaction through an explicit token ceiling dial ( token_threshold ), deprecating CapabilitiesConfig.compaction_threshold . Forward-Looking Budget Scope : Adds BudgetScope.FORWARD_LOOKING to BudgetConfig to enforce model call and token budgets across newly resumed execution turns without counting previous historical usage. Tool Output Token Truncation : Exposes tool_output_truncation_config across agent configurations to cap token output volume initially for run_command executions in localharness. UsageMetadata Arithmetic : Supports standard Python arithmetic protocols on UsageMetadata

• Improved the startup warning for deprecated unsandboxed permission rules across CLI, shared, and project configuration files to list each affected file path, up to five offending rules, and step-by-step instructions for migrating them to command rules.
• Improved /resume startup responsiveness when opening a large conversation history with a cold or stale summary cache.
• Fixed MCP servers bundled inside plugins colliding with each other or with user-configured servers in mcp_config.json when they shared the same server name by automatically namespacing plugin MCP servers as <plugin>_<server>.
• Fixed a memory leak where opening or scanning conversations left background step-cache eviction goroutines running for the rest of the session, significantly reducing memory usage after opening /resume or switching conversations.
• Fixed view_file attempting to parse non-UTF-8 binary files as text or loading files larger than 100 MB into the model context; unsupported binary formats and oversized files are now rejected with a clear error before overflowing the context window.

• Added support for excludeDefaultComponents: true in custom agent Markdown frontmatter, allowing custom agents to opt out of default prompt sections and built-in tools while preserving post-invocation hooks.
• Improved model API error resilience and diagnostics: transient genai.APIError failures (502, 503, 504, per-minute 429 rate limits, and mid-stream interruptions) automatically retry in-process with exponential backoff while preserving completed tool call outputs, and unrecovered 503 and 429 responses surface clear user-facing error messages.
• Improved MCP and provider tool schema validation to preserve open object schemas (such as {"type": "object"} or explicit additionalProperties: true) instead of rejecting undeclared arguments on schemas that allow them.
• Improved per-turn responsiveness and reduced memory usage when opening large conversations.
• Fixed --continue starting a brand-new conversation when launched from a subdirectory, after a crash, or while another session is open in the same workspace; it now falls back to the most recent non-empty conversation in the current workspace or its parent/child directories.

• Added the remote-control start, remote-control status, and remote-control stop subcommands to run the CLI as a background service registered with your operating system's service manager so your machine stays reachable across logouts and reboots; remote-control start --name <label> sets a custom machine label in the Remote Control instance list, and --session scopes the service to your active login session.
• Improved half-page scrolling across all scrollable views — including altscreen mode, the diff viewer, and read-only detail panels where ctrl+d no longer triggers the exit prompt — and set shift+up and shift+down as customizable default keybindings for half-page navigation (navigation.half_page_up and navigation.half_page_down).
• Fixed prompts or model responses blocked by content safety filters failing with a spurious "no candidate found" error, ending silently with an empty turn, or repeatedly retrying the rejected input; the CLI now surfaces a clear content-filter stop reason.

• Improved resilience to transient model API errors: the agent now retries errors such as 503 Unavailable for much longer with exponential backoff, so a brief service hiccup no longer aborts your session.
• Improved sign-in and startup speed by reading your signed-in identity from the stored credential instead of making a network request on every launch and after browser consent; a failed identity lookup no longer prevents sign-in from completing, and the displayed plan tier no longer briefly disappears during startup.
• Improved headless (-p) runs to exit promptly once the final answer is delivered: the CLI waits for running background tasks and scheduled timers to finish, bound by --print-timeout, and leaves daemon background tasks such as dev servers running instead of terminating them.
• Improved failure reporting in headless (-p) runs: fatal errors now appear on stderr with a stable error: marker, a note is printed when the response may be truncated, and runs that previously ended silently with no output now explain why.

• Added /model <name> <prompt>, which runs a single prompt on another model and then returns the session to the model it was using, so you can consult a different model mid-conversation without disturbing your saved default.
• Added a conversation_title field to the JSON payload passed to custom status line and window title scripts, so they can show the active conversation's name and pick up renames immediately.
• Added an agents list to custom agent Markdown frontmatter, letting an agent declare the subagents it depends on using the same workspace-relative, absolute, and agent-relative path rules as skills.
• Improved /model to show a [model] argument hint as you type, so the command's arguments are discoverable without opening help.
• Improved the /settings panel so each Verbosity option explains itself as you highlight it, describing whether tool calls, commands, and thoughts are shown in full or collapsed into a token metrics table.
• Improved the sign-in and Google Cloud setup screens to use the CLI's standard styled key hints, which follow your color scheme and keybindings and drop the navigation hint when only one option is available.

• Added half-page scrolling with Ctrl+D and Ctrl+U to the artifact viewer.
• Added the pickerGrouping setting to /config and settings.json to configure the default conversation list view in /resume (flat or grouped by workspace).
• Improved terminal ASCII diagram rendering for Mermaid flowcharts in the artifact viewer and conversation.
• Improved /logout execution time by short-circuiting token removal directly to file storage when keyring storage is bypassed or unreachable.
• Changed unselected model families in the interactive /model picker to default to medium reasoning effort instead of low.
• Fixed subagent tasks unexpectedly prompting for tool approval in always-proceed mode while viewing the subagent details panel.
• Fixed orphaned Git worktrees accumulating on disk under .system_generated/worktrees when killing subagents or deleting conversations.
• Fixed SQLite database WAL checkpointing on CLI exit so trailing session metadata updates are flushed to disk before shutdown.
• Fixed customization discovery logging spurious errors when accessing temporary directories or paths outside the active workspace.

• Added an opt-in workspace-grouped view to the /resume conversation picker, allowing users to toggle between a flat list and conversations grouped by directory with Ctrl+F.
• Added Gemini 3.8 Flash to the model catalog when connecting with a GEMINI_API_KEY.
• Changed custom agents defined in Markdown to inherit ambient skills, rules, and subagents by default, matching the configuration of default agents.
• Fixed MCP OAuth authentication failing when authorization servers return authorization codes longer than 1024 characters.
• Fixed skill path matching and grouping in the /skills panel on Windows where host path separators (``) caused global and workspace skills to be misclassified.
• Fixed relative path resolution in Markdown-defined custom agents so relative subagent and plugin paths resolve correctly against the agent definition's directory.
• Fixed duplicate permission grants accumulating in configuration settings across session reloads and subagent invocations.
• Fixed /boost failing at runtime by improving worker tool configurations.