v1.40.0
GooseSummary
This release introduces extensive new capabilities, most notably adding support for numerous new declarative AI providers (including Together AI, Perplexity, and Alibaba) and expanding desktop localization to 10 new languages. Key functional additions include new CLI commands (/status, /model, /goal, `goose review`), enhanced extensibility via a new hooks system, and better configuration controls for tool output size and update behavior. Several critical bugs related to tool call argument handling, session syncing, and provider configuration have also been resolved.
New Features
- Added language selection capability in the desktop application.
- Introduced desktop locales for French, German, Italian, Portuguese, Indonesian, Malay, Vietnamese, Traditional Chinese, Korean, and Japanese.
- Added the /status slash command for use in the CLI.
- Added the /model slash command to allow switching the session model.
- Added the /goal command for agent self-evaluation before task completion.
- Introduced the `goose review` command for local code review.
- Enabled Peek mode for asynchronous background tasks during summon operations.
- Added support for context parameter and working_dir override for delegate tasks.
- Added support for the GLM-5.2 model.
- Added declarative providers for Together AI, EmpirioLabs (OpenAI-compatible), OrcaRouter (OpenAI-compatible), NEAR AI Cloud, Perplexity (OpenAI-compatible), Alibaba (Qwen via DashScope), Databricks AI Gateway, Scaleway, Vercel AI Gateway, and xAI SuperGrok (OAuth subscription).
- Enabled OpenAI GPT models access via AWS Bedrock mantle endpoint.
- Added support for Azure Entra ID bearer token authentication via AZURE_OPENAI_AD_TOKEN.
- Added support for MLX models for local inference provider.
- Enabled streamable HTTP support for the deeplink generator.
- Enabled loading global hints from ~/.agents/AGENTS.md.
- Unified thinking effort control across all providers.
- Introduced canonical thinking modes.
- Implemented a hooks system for extensibility.
- Added PreToolUse denial capability within hooks.
- Enabled honoring blocking Stop hook decisions.
- Added an option to disable automatic update downloads.
- Enabled tracking cache tokens for more accurate cost reporting.
- Added configuration for tool output size limit via GOOSE_MAX_TOOL_RESPONSE_SIZE environment variable.
- Added configuration for MAX_CODE_BLOCK_LINES via environment variables.
- Added support for the GOOSE_FAST_MODEL environment variable.
- Added Hugging Face OAuth support with a dedicated auth tab in settings.
- Enabled chat history search functionality in the navigation panel.
- Added the ability to import sessions.
- Added functionality to view model interactions.
- Added functionality to view JSON content.
- Added an image read tool.
- Introduced `goose://new-session` deep link for starting new sessions.
- Introduced `goose://resume` deep link for resuming sessions.
- Enabled initial prompt execution when using the `goose://new-session` deep link.
- Added support for MCP extensions in open plugins.
- Added Nushell terminal and completion support.
- Added the TUI command to `goose-cli`.
- Implemented a TUI diff viewer.
- Unified OTLP logging schema for improved cross-tool detection.
- Enabled prompt injection mitigation by default for internal users.
- Added encrypted Nostr session sharing.
- Added a quarterly option for the scheduler.
- Added optional `--parameters` support for scheduled recipes.
- Implemented a worktree-aware directory switcher.
- Added Harbor eval runner support.
- Enabled Linux musl CLI builds.
Bug Fixes
- Fixed Mac auto-update configuration seeding issues.
- Fixed OTLP export failure in `goose serve` mode by using an async reqwest client.
- Fixed panic when encountering non-object tool-call arguments by rejecting them instead.
- Fixed /edit prompts not being saved to history.
- Fixed incorrect listing of schedule sessions by using `session.message_count`.
- Fixed session store not syncing after provider and model updates.
- Fixed tool call buffering issues for DeepSeek/Kimi by buffering thinking across stream chunks.
- Fixed truncated tool-call arguments by surfacing actionable errors.
- Fixed extension disabling by showing the optimistic toggle state.
- Fixed passing of the temperature parameter to the ChatGPT Codex provider.
- Resolved Databricks v2 pricing and correctly surfaced cost in standard usage updates.
- Activated hermit on macOS to ensure STDIO extensions resolve node correctly.
- Fixed Autovisualiser to accept JSON-encoded strings for the data parameter.
- Fixed image URL fetches in the Developer/read_image tool by sending a User-Agent.
- Fixed starting a turn when /goal or /grind is set.
- Fixed passing of thinking configuration to Bedrock Anthropic models.
- Fixed preservation of Windows backslash paths in custom extension commands.
- Fixed custom OpenAI providers dropping the port number when the URL scheme was omitted.
- Resolved an ACP cancel race condition.
- Fixed window closing behavior where the window remained 'goosed' when closing.
- Made the context exceeded checker more precise.