Google Antigravity

0.1.9 July 27, 2026 Model-call retry/backoff configuration, LiteRT warm-up scaling, and tool execution error handling Release 0.1.9 of the Google Antigravity Python SDK adds model-call retry/backoff configuration, improves tool registration for custom functions, adds missing BuiltinTools exports, fixes user audio payload processing, and adds connection DebugConfig support along with ToolExecutionError handling. Improvements (5) Model-Call Retry & Backoff Configuration : Exposes configurable retry and backoff parameters in LocalAgentConfig for model calls. LiteRT Warm-up Timeout Scaling : Dynamically scales LiteRT engine warm-up timeouts based on context size and synchronizes warm-up cleanup. Improved Tool Registration : Enhances automatic tool registration and docstring parsing for custom Python functions. BuiltinTools Typing Exports : Exports BuiltinTools at top-level package and typing boundaries for subagent tool configuration. Connection DebugConfig : Adds base DebugConfig options for enhanced connection debugging and logging. Fixes (2) Audio Payload Processing : Fixes processing for user audio payloads in interactive agent sessions. ToolExecutionError Handling : Introduces

• Improved permission prompts for compound shell commands so the full command is shown when any part of it needs approval.
• Fixed disabled plugins still running their hooks and contributing other customizations, which could keep a broken hook active and break file-editing tools even after the plugin was turned off.
• Fixed MCP OAuth against providers that do not strictly follow the spec (such as Salesforce and Atlassian) by relaxing issuer validation and including the refresh_token grant.
• Fixed /btw failing with a "parent conversation not found" error when used as the very first action in a fresh session.
• Fixed clipboard corruption of CJK and other non-ASCII text when copying on Windows.
• Fixed print mode (-p) sending a prompt before the account-eligibility check finished.

• Custom Agents (Markdown Format). Added support for defining custom agents using Markdown files (agent.md) with YAML frontmatter and H1-delimited system prompts. Markdown agents support mainAgent, subagent, hidden, inheritMcp, and commandExecutionPolicy frontmatter fields for fine-grained control over agent behavior. Dynamically defined subagents (via define_subagent) now also write Markdown format so they resolve correctly on external builds.
• Added an optional index argument to /copy so /copy <n> copies the n-th most recent response to the clipboard, while /copy and /copy 1 still copy the latest.
• Improved /codesearch to render results progressively as they stream in, showing a live count while loading and letting you cancel an in-flight search with Esc instead of blocking until the whole search finishes.
• Improved default file access by granting read access to the system temporary directory out of the box, resolved correctly per platform, so agents no longer trigger permission prompts when reading temporary files.

• Added a /effort command to view and change the current model's reasoning effort, with a left/right timeline-gauge picker and a direct /effort <level> form so you can trade latency for depth on the fly.
• Added an --effort flag to select a model's reasoning-effort variant when launching the CLI.
• Added stable, user-facing model slugs that appear in the /model picker and are accepted by --model, so you can pin a specific model reliably across sessions.
• Added a model option to custom agent frontmatter so an agent runs at a chosen model tier (such as flash or pro) when invoked as a subagent, defaulting to inherit (the parent's model).
• Redesigned the /model picker to group models by their base model and choose reasoning effort from a timeline gauge navigable with Left and Right, and added an effort badge to the status line for models that expose multiple effort variants.
• Improved the /settings (/config) panel by making it a bounded, scrollable list so it renders correctly in short terminals instead of overflowing, and stopped it from flickering when opening and closing dropdowns.

0.1.8 July 21, 2026 Gemini 3.6 Flash default model, custom subagent instructions, and Pydantic argument coercion Release 0.1.8 of the Google Antigravity Python SDK updates the default text model to Gemini 3.6 Flash, adds support for custom subagent instructions, and implements automatic Pydantic argument coercion for tool calls. It also brings defensive prompt sanitization, configurable tool retries, and comprehensive stability improvements for local agent execution. Improvements (7) Default Model Upgrade to Gemini 3.6 Flash : Upgrades the default generative text model in the Python SDK to gemini-3.6-flash . Pydantic TypeAdapter Tool Argument Coercion : Automatically coerces stringified numbers, booleans, and nested models returned by LLMs into strict Python types declared in tool signatures. Custom Subagent Instructions : Adds capability to specify custom system instructions and allowlisted tool configurations for spawned subagents in multi-agent workflows. Prompt Sanitization : Strips null bytes ( \x00 ) and non-printable control characters ( DEL , BEL , C1 ) from incoming user prompts at the wire boundary to prevent HTTP 400 errors and terminal corruption. Configurable Tool

Changelog

• Added support for stacking multiple leading slash commands in a single prompt, so a chain like /plan /grill-me <prompt> parses, activates, and renders every command in the order you typed them.
• Improved scrolling in the /diff viewer so paging through a diff no longer jitters or pushes the status line off the screen when lines wrap or comments expand.
• Fixed custom agents that declare subagent: false still appearing in the available-subagents list and being invocable as subagents.
• Fixed headless (-p / --print) runs so they now honor persisted settings.json policies, including permissions, file access, sandbox mode, auto-execution, and artifact review.
• Fixed /btw side-questions leaking into the conversation list as duplicate entries that carried the parent conversation's title.
• Fixed the prompt to honor a custom Enter binding to prompt.insert_newline, so a remapped Enter inserts a newline instead of submitting.
• Fixed eligibility error messages so the CLI shows the real reason again instead of defaulting to a generic "unknown reason".

• Added a /codesearch command (aliases /cs and /search) to interactively search code across your workspace, interpreting queries as regex by default with -F/--literal for exact matching and f:/file: globs to include or exclude paths.
• Added copy-on-select in no-flickering mode so dragging highlights text and releasing the mouse copies the ANSI-stripped selection to the clipboard, and hides the virtual scrollbar so it no longer interferes with copying multi-line output.
• Added an indicator at each context-compaction boundary so you can see where earlier compaction happened.
• Improved interactive startup latency by loading skills asynchronously so the CLI no longer blocks on a synchronous, filesystem-heavy skill-discovery pass during bring-up.
• Improved eligibility error handling by showing errors with a verification URL inline in the input loop instead of stacking them above the screen.
• Improved customization loading latency for skills, rules, agents, and hooks by consolidating directory walks and caching filesystem lookups to cut redundant I/O during discovery.
• Removed the padding spaces around inline code for tighter rendering.

0.1.7 July 14, 2026 Multi-threaded state handling, subprocess isolation, extra high thinking, and local model MCP Release 0.1.7 of the Google Antigravity Python SDK expands end-user control over agent execution environments, strengthens concurrency safety for stateful tools, and introduces deeper reasoning capabilities. Key highlights include atomic multi-threaded state handling for tools and hooks, customizable subprocess environment variable isolation, support for an "extra_high" thinking severity level, and full Model Context Protocol (MCP) and subagent support across local model backends. This release also resolves interactive console prompt clobbering and improves socket discovery under containerized setups. Improvements (9) Multi-threaded Hook & Tool State Handling : Developers can now safely read and mutate shared context variables across multi-threaded tools ( asyncio.to_thread or ThreadPoolExecutor ) using atomic updates and thread locking: Custom Subprocess Environment Variables : Developers can now pass custom environment variables directly to isolated agent instances via LocalAgentConfig , avoiding pollution of the global parent environment: Extra High Thinking

• Added an f (full diff) shortcut to the create-file tool review screen so new-file confirmations can open a full-screen diff view, matching the existing file-edit experience.
• Added support for pasting the OAuth authorization code in print mode (-p) via the controlling terminal (/dev/tty on POSIX and CONIN$ on Windows) when stdin is consumed by a piped prompt, and made truly headless runs fail fast with an actionable message instead of blocking.
• Improved responsiveness on large conversations (5000+ steps) in no flickering mode by switching hot-path line-count methods to pointer receivers, cutting the per-frame prefix-sum cost and eliminating sustained 99% CPU and keystroke lag.
• Fixed print mode silently downgrading to the default model when --model cannot be resolved by hard-failing with a non-zero exit and listing the available models, while interactive sessions keep the fallback-with-warning behavior.
• Fixed permission checks not respecting the allowlist for nested command substitutions, so a command like echo "$(dirname $(git rev-parse --show-toplevel))" now runs without prompting when echo and git are allowlisted, instead of double-counting the nested command and

• Added the --agent flag and agent/agents subcommand, allowing users to select a custom agent at launch and list available agents.
• Added in-file keyword search (/) and jump navigation (n/N) to the artifact detail viewer, allowing users to find and cycle through matches without disrupting terminal escape sequences or image grids.
• Fixed print mode (--print / -p) silently exiting with a success code and empty output when a request failed server-side, now writing the error to stderr and returning a non-zero exit code.
• Fixed agy -p hanging when run inside a shell script or subprocess by no longer reading stdin when a prompt is provided via a flag.
• Fixed a data race on the /btw cancellation function.
• Added support for displaying nested subagents (grandchild and deeper) and handling tool confirmation requests across all subagent depths by recursively relaying nested subtrajectory updates to the root conversation.
• Changed the default mode to respect write_file permissions allowlisted in settings.json under permission.allow, so pre-approved file writes no longer prompt for review.

0.1.6 July 9, 2026 LiteRT model configurations, dynamic GCP environment hydration, and performance updates This release expands local execution capabilities by broadening support for multimodal and web-connected tool workflows. Developers can now run Gemma models locally using LiteRT or integrate with local OpenAI-compatible APIs, natively return multimodal media from custom tools, and leverage a more robust, streamlined lifecycle hooks framework that is fully aligned with Model Context Protocol (MCP) workflows. Improvements (4) Local Model Connectivity : Introduced LiteRTAgentConfig and LiteRTConnectionStrategy for LiteRT-LM (supporting local Gemma execution), LocalOpenAIAgentConfig and LocalOpenAIConnectionStrategy for OpenAI-compatible APIs (supporting Ollama and LM Studio), and a background loopback HTTP translation server. Multimodal Tool Outputs : Enabled custom tools to return media assets ( Image , Document , Audio , Video ) directly via a single tool response without needing separate follow-up turns ( supplemental_media ). Built-in Web Fetch Tool : Integrated the read_url_content tool end-to-end for fetching structured web content natively with the ReadUrlContentResult

• Agent execution mode cycling is now publicly available: default -> accept-edits -> plan)
• Added request-review (default) mode as the default execution behavior: automatically pauses before file write operations to display an interactive, line-level diff preview (f shortcut) where users can review, accept, or reject individual code modifications before they are saved to disk.
• Added an Agent Mode option to the /settings panel so users can set and persist a default execution mode (default, accept-edits, plan) without manually editing settings.json or passing --mode on startup, with real-time synchronization so changes take effect immediately.
• Added a dedicated "Create file" confirmation preview for new file creations (write_to_file without overwrite): renders new content as an addition-only diff preview.
• Added /plan mode to replace legacy /planning, and removed /fast slash commands: consolidated and simplified execution mode switching around shift+tab mode cycling and the /plan mode prefix

• Improved the /tasks detail panel to automatically scroll to the bottom as new background task logs stream in, and default to the latest output when opened while preserving scroll position if scrolled up manually.
• Improved model generation resilience by adding automatic client-side retries when encountering transient errors.
• Fixed dynamically defined subagents by transitioning definitions from JSON to Markdown format, fixing an issue where dynamically created subagents failed to invoke.
• Fixed a crash occurring when executing background tasks or terminal commands that produce empty outputs (such as sleep).
• Fixed shutdown resource leaks by integrating the shared SQLite summary store for background synchronization and resolving goroutine and database connection leaks on CLI exit.
• Fixed a permission manager hook error by safely handling empty decision strings returned by pre-tool hooks instead of failing with an "unknown pre-tool hook decision" error.

• Introduced a new interactive status indicator below the input box that displays active subagents and background tasks in real-time, making it easy to monitor and navigate parallel workflows at a glance.
• Added ctrl+g on the artifact view to open $EDITOR. Also added a warning confirmation prompt before opening the editor in the artifact detail view if there are unsent comments, and ensured these comments are preserved upon reload if the artifact content was not modified.
• Added alt+v as an alternative paste shortcut on Windows to resolve issues where ctrl+v is intercepted by the terminal emulator, enabling reliable image pasting.
• Improved the /permissions panel to dynamically reload configurations from disk and prevent accidental overwrites.
• Increased the MCP connection timeout to 60 seconds to improve reliability for slow-starting custom MCP servers.
• Fixed a bug on Windows where print mode and other non-TUI command outputs were silently discarded when run in non-TTY environments (such as pipes or subprocesses).

• Allowed image pasting from the clipboard in local tmux sessions.
• Removed the max limit for the /goal command, allowing goals to run indefinitely until completed or cancelled.
• Enabled "always proceeds" mode for subagents to auto approve artifacts, preventing them from hanging when the parent is blocked.
• Fixed plugin import logic to copy the entire plugin directory, preventing it from stripping non-skill directories (like shared/).
• Fixed an MCP configuration path mismatch in the CLI and permission manager to ensure reliable custom MCP server loading.
• Fixed a TUI layout race condition caused by stale input state in the conversation model.
• Fixed a bug where the inline viewport was not properly reset after a conversation rewind.