Change8

v1.37.0

Goose

Summary

This release introduces a significant expansion of supported AI providers, including xAI SuperGrok, Perplexity, Alibaba (Qwen), Databricks, NEAR AI Cloud, and more, alongside new declarative provider support. Key usability enhancements include the introduction of a TUI command, local code review (`goose review`), and deep linking for session management. Numerous bug fixes address stability, configuration preservation, and provider compatibility across the platform.

New Features

  • Added xAI SuperGrok OAuth subscription provider.
  • Enabled replay of ACP images upon session load.
  • Exposed raw provider supported models over ACP.
  • Introduced ACP session system prompt setter.
  • Added Scholar Sidekick MCP extension.
  • Added `tui` feature flag to gate the TUI command.
  • Implemented CLI to list skills along with their token counts.
  • Added Perplexity as a declarative OpenAI-compatible provider.
  • Added Alibaba (Qwen via DashScope) declarative provider.
  • Enabled configurable tool output size limit via GOOSE_MAX_TOOL_RESPONSE_SIZE environment variable.
  • Added `/model` slash command to CLI for switching the session model.
  • Added Databricks AI Gateway provider.
  • Added Russian language support.
  • Added Turkish desktop locale.
  • Implemented TUI command on goose-cli.
  • Enabled configuration of MAX_CODE_BLOCK_LINES via environment variables.
  • Added NEAR AI Cloud provider.
  • Introduced `goose://resume` session deep link.
  • Added Scaleway provider.
  • Enabled support for GOOSE_FAST_MODEL environment variable in ModelConfig::with_fast.
  • Added Harbor eval runner.
  • Enabled slash commands (built-in, skill, recipe) in ACP server.
  • Added `/goal` command for agent self-evaluation before finishing.
  • Implemented summon subagent instructions.
  • Generalized Open-plugins and added skills support.
  • Added PreToolUse denial hooks.
  • Implemented ACP session list pagination.
  • Introduced `goose review` local code review command.
  • Added TUI diff viewer.
  • Introduced `goose://new-session` deep link to open a fresh chat.
  • Added Atomic Chat as a declarative OpenAI-compatible provider.
  • Enabled GOOSE_OAUTH_CALLBACK_PORT for stable OAuth redirect_uri.
  • Enabled proactive OAuth token refresh to avoid re-authentication on every session.
  • Added GPT-5.5 to ChatGPT Codex known models list.
  • Enabled opt-in to Vercel AI gateway leaderboard.
  • Enabled ACP to pass session cwd parameter to providers.
  • Added Routstr as a declarative provider.
  • Added FuturMix provider.
  • Added oMLX declarative provider.
  • Enabled stripping of chain-of-thought markers from custom provider output.
  • Implemented encrypted Nostr session sharing.
  • Added Simplified Chinese (zh-CN) translation.
  • Added Nushell terminal and completion support.
  • Enabled optional `--parameters` for scheduled recipes.
  • Added GitHub Copilot /responses API.
  • Added Vercel AI Gateway provider.
  • Introduced Hooks system.
  • Enabled Projects as backend sources with system prompt injection.
  • Added Linux Vulkan support for local inference.
  • Enabled ACP streamable HTTP spec compliance.
  • Enabled grouping of consecutive tool calls into one summarized chain card.
  • Added Quarterly option for scheduler.

Bug Fixes

  • Fixed forwarding of custom headers through the OAuth connect path.
  • Fixed issue with scheduled recipe session parameters.
  • Fixed skipping of OTLP signals when protocol=grpc to prevent background-thread panic.
  • Fixed refreshing of the provider list in the Switch Models picker.
  • Fixed starting a new chat in the current window from the recipe parameter modal.
  • Made the Azure api-version query parameter optional.
  • Fixed preservation of pending environment variables in the Add Extension form.
  • Fixed handling of missing responses output.
  • Fixed building of linux x86_64 standard inside manylinux_2_28 for glibc 2.28+ compatibility.
  • Fixed the main window growing taller on every launch.
  • Fixed desktop chat search session limiting.
  • Fixed serialization of per-session agent creation to prevent duplicate MCP initialization.
  • Fixed forwarding of ACP server context window size to clients.
  • Fixed using context limit from /model/info for custom LiteLLM models.
  • Fixed preservation of thinking content for provider context.
  • Fixed mentioning configurable timeout environment variables in Ollama stream stall errors.
  • Fixed inclusion of full recipe parameter details in load/discovery output.
  • Removed unused fetch-metadata IPC handler (SSRF vulnerability).
  • Fixed respecting GOOSE_MAX_TURNS in gateway sessions.
  • Fixed sending an empty object instead of null for Anthropic tool_use input.
  • Fixed preservation of the selected branch across project chats.
  • Fixed using current_exe() instead of PATH lookup when spawning goose.
  • Fixed setting TCP_USER_TIMEOUT on streamable HTTP clients.
  • Fixed activation of a custom provider after adding it via configuration.
  • Fixed reduction of excessive MISSING_TRANSLATION warnings for fallback locales.
  • Eliminated cross-window deep link contamination.
  • Improved Telegram gateway error reporting and connection reliability.

Goose Documentation