Goose shell

All SHELL Bug Fixes

• Fixed issue where Databricks parallel tool image responses did not interleave tool results correctly.(v1.35.0)
• Stopped killing `goosed` process when a window is closed.(v1.35.0)
• Eliminated cross-window deep link contamination issues.(v1.35.0)
• Fixed file fallback check when the keyring has no entry.(v1.35.0)
• Ensured custom providers are activated immediately after being added via configure.(v1.35.0)
• Fixed spawning of `goose` by using `current_exe()` instead of PATH lookup.(v1.35.0)
• Set TCP_USER_TIMEOUT on streamable HTTP clients.(v1.35.0)
• Reduced excessive `MISSING_TRANSLATION` warnings for fallback locales.(v1.35.0)
• Zeroed out cost calculation for local providers (ollama, local).(v1.35.0)
• Fixed session cost persistence, ensuring accumulated cost survives session reloads.(v1.35.0)
• Prevented tool-use marker leakage in toolshim output.(v1.35.0)
• Fixed re-application of canonical limits when a delegate overrides the model.(v1.35.0)
• Changed CLI prompt from goose emoji to plain `> ` for better compatibility.(v1.35.0)
• Added flag to handle login shell PATH correctly.(v1.35.0)
• Enabled VT processing on Windows Console Host.(v1.35.0)
• Resolved Azure CLI invocation on Windows by correctly using `az.cmd`.(v1.35.0)
• Handled non-interactive terminal scenarios during `goose configure` on Windows.(v1.35.0)
• Fixed emitting `trace_output` as a span attribute instead of an event.(v1.35.0)
• Ensured reliable flushing of OTLP traces on exit with configurable timeout.(v1.35.0)
• Improved Telegram gateway error reporting and connection reliability.(v1.35.0)
• Fixed error message styling in the ScheduleModal.(v1.35.0)
• Aligned the sidebar hamburger icon correctly in macOS fullscreen mode.(v1.35.0)
• Coalesced streaming Thinking deltas and now list available tools when a tool is not found.(v1.35.0)
• Ensured the system honors `Retry-After` headers on 429 responses.(v1.35.0)
• Replaced the Venice custom provider implementation with declarative configuration.(v1.35.0)
• Used `mkdir -p` for self-test workspace initialization.(v1.35.0)
• Fixed preservation of user-set session names when chats are initiated via recipes.(v1.35.0)
• Enforced expiration independently of `MAX_TOKEN_AGE_SECONDS` in the OIDC proxy.(v1.35.0)
• Improved inline code contrast in the light theme.(v1.35.0)
• Omitted `max_tokens` for OpenAI-compatible requests when it is unset.(v1.35.0)
• Set correct `mcp-app` host capabilities.(v1.35.0)
• Normalized nullable schemas for Vertex Gemini compatibility.(v1.35.0)
• Isolated GitHub recipe temporary paths.(v1.35.0)
• Set the `X-Initiator` header on GitHub Copilot requests.(v1.35.0)
• Ensured dynamic_models: false is honored in declarative provider configurations.(v1.35.0)
• Cached the trailing message for stable prefix consistency across Bedrock agent turns.(v1.35.0)
• Implemented case-insensitive model name lookup for context_limit.(v1.35.0)
• Fixed returning 400 status instead of panicking on invalid CSP header values.(v1.35.0)
• Fixed insertion of tool pair summaries to appear at the chronological position in the conversation.(v1.34.0)
• Enabled SQLite foreign key enforcement in the session pool.(v1.34.0)
• Fixed issue where deleted chat sessions still appeared in the sidebar list.(v1.34.0)
• Reverted unexpected change from ui/goose2 work.(v1.34.0)
• Fixed requirement of extension_name on read_resource.(v1.34.0)
• Fixed queued messages being lost when clicking send immediately.(v1.34.0)
• Fixed model picker and session config staying out of sync.(v1.34.0)
• Fixed reading prompted file parameters in recipes.(v1.34.0)
• Enabled acceptance of null tool_call arguments in OpenAI streaming chunks.(v1.34.0)
• Fixed developer extension instructions to use python3 for macOS/Linux compatibility.(v1.34.0)
• Fixed SIGCHLD race condition by synchronously reaping ACP child.(v1.34.0)
• Fixed goose2 small-window chat and settings layouts.(v1.34.0)
• Resolved elicitation issues.(v1.34.0)
• Fixed SACP not notifying clients of generated session names.(v1.34.0)
• Fixed unscheduling a recipe incorrectly deleting it.(v1.34.0)
• Fixed seeding provider handoff history in ACP.(v1.34.0)
• Fixed goose2 not respecting interface density settings.(v1.34.0)
• Fixed goose2 not respecting accent color.(v1.34.0)
• Fixed preservation of replay message timestamps.(v1.34.0)
• Fixed goose2 window minimum sizing constraints.(v1.34.0)
• Fixed transform-rasterized dialog text appearing in goose2.(v1.34.0)
• Fixed renamed skills not staying open in the detail view.(v1.34.0)
• Fixed Gemini OAuth onboarding failure.(v1.34.0)
• Improved copy and content in goose2.(v1.34.0)
• Fixed model picker remaining usable during provider loading.(v1.34.0)
• Enabled handling of ACP requests concurrently.(v1.34.0)
• Fixed CRT linkage in Windows CUDA build.(v1.34.0)
• Fixed exclusion of preprompt from session title generation.(v1.34.0)
• Prevented login-shell PATH probe from suspending goose on startup.(v1.34.0)
• Removed defaults yaml.(v1.34.0)
• Fixed slow startup on first launch by caching TLS certs to disk.(v1.32.0)
• Fixed preprompt showing incorrectly after session loading.(v1.32.0)
• Fixed handling of streaming chunks missing the delta field.(v1.32.0)
• Fixed OpenAI provider accepting streaming chunks with both reasoning fields.(v1.32.0)
• Corrected model handling for Responses API tools and gpt-5.4 using `strict:false`.(v1.32.0)
• Corrected grammar in the apps view description.(v1.32.0)
• Fixed issue where links in chat could not be opened.(v1.32.0)
• Fixed shell tool execution to run under bash/sh regardless of the login shell.(v1.32.0)
• Fixed appending /chat/completions for prefixed v1 base URLs.(v1.32.0)
• Aligned Kimicode model metadata with the live Kimi Code API.(v1.32.0)
• Fixed an idle 300ms re-render loop that could cause TUI OutOfMemory errors.(v1.32.0)
• Fixed chat loading-state model placeholder display.(v1.32.0)
• Fixed tool calls to expand by default when Response Style is Detailed.(v1.32.0)
• Fixed creation of logs directory before writing LLM request logs.(v1.32.0)
• Enabled token usage tracking and configurable stream timeout for the Ollama provider.(v1.32.0)
• Fixed thread timestamps by using sqlx chrono decode instead of manual parsing.(v1.32.0)
• Removed the double border artifact on content in the chat view.(v1.32.0)
• Fixed acceptance of self-signed certs from configured external goosed hosts.(v1.32.0)
• Fixed sidebar sessions not syncing after chat history deletion.(v1.32.0)
• Prevented session hang when tool approval was required in headless mode.(v1.32.0)
• Eliminated a ~5s delay when opening the Switch Models panel with a local provider.(v1.32.0)
• Removed hardcoded dependency of goose-acp-macro.(v1.32.0)
• Fixed stopping the load of openai fast model for openapi compatible custom endpoints.(v1.32.0)
• Fixed empty tool results originating from resource content (e.g., auto visualiser).(v1.29.0)
• Implemented cleanup of the OAuth token cache upon provider deletion.(v1.29.0)
• Updated SSE parsers to correctly accept optional space after the data: prefix.(v1.29.0)
• Fixed hard-coded tool call ID in the code mode callback.(v1.29.0)
• Configured GOOSE_MAX_BACKGROUND_TASKS setting.(v1.29.0)
• Fixed issue where the goosed child process's listener was not being unregistered.(v1.29.0)
• Ensured token usage is included in Databricks streaming responses.(v1.29.0)
• Fixed overwriting of the deprecated googledrive extension configuration.(v1.29.0)
• Implemented emitting a user-facing message upon compaction failure.(v1.29.0)
• Updated to use the Responses API for gpt-5.4.(v1.29.0)
• Fixed auto-focusing the chat input upon Hub mount.(v1.29.0)
• Fixed Apps extension to coerce string arguments from inner LLM responses.(v1.29.0)
• Implemented handling of mid-stream error events in OpenAI SSE streaming.(v1.29.0)
• Fixed fullscreen header bar and always-visible close controls behavior.(v1.29.0)
• Implemented using the default provider and model when the provider in the session no longer exists.(v1.29.0)
• Added support for Gemini models via Databricks.(v1.29.0)
• Fixed handling of reasoning content blocks in the OpenAI-compat streaming parser.(v1.29.0)
• Removed the configured marker when deleting OAuth provider configuration.(v1.29.0)
• Fixed issues related to the Download manager.(v1.29.0)
• Fixed macOS Intel signing.(v1.29.0)
• Removed incorrect agent_visible filter on user message in claude-code.(v1.28.0)
• Implemented retry logic on authentication failure with credential refresh.(v1.28.0)
• Bumped pctx_code_mode to 0.3.0 to fix iterator type checking.(v1.28.0)
• Passed session_id when loading extensions to ensure skills are discovered.(v1.28.0)
• Added tool choice and parallel_tool_calls support for the chatgpt_codex provider.(v1.28.0)
• Fixed tool confirmation handling for multiple requests.(v1.28.0)
• Resolved tokio::sync::Mutex deadlock in the recipe retry path.(v1.28.0)
• Stopped writing without error when hitting a broken pipe for goose session list.(v1.28.0)
• Skipped upgrade-insecure-requests CSP for external HTTP backends.(v1.28.0)
• Prevented hang when a command backgrounds a child process.(v1.28.0)
• Added exit from the agent loop when tool call JSON fails to parse.(v1.28.0)
• Prevented SSE connection drops from silently truncating responses.(v1.28.0)
• Prevented SageMaker TGI from being marked configured when only Bedrock keys are set.(v1.28.0)
• Disabled computercontroller functionality when no $DISPLAY is detected.(v1.28.0)
• Registered MCP extensions when resuming a session.(v1.28.0)
• Loaded .gitignore in prompt_manager for hint file filtering.(v1.28.0)
• Remapped max_completion_tokens to max_tokens for OpenAI-compatible providers.(v1.28.0)
• Preserved OpenAI Responses API tool call/output linkage.(v1.28.0)
• Returned ContextLengthExceeded when the prompt exceeds the effective KV cache size.(v1.28.0)
• Used \"includeThoughts/part.thought\" for Google thinking handling.(v1.28.0)
• Fixed model selector showing the wrong model in tabs.(v1.28.0)
• Avoided word splitting by space for Windows shell commands.(v1.28.0)
• Unified path resolution around GOOSE_PATH_ROOT.(v1.28.0)
• Passed OAuth scopes to DCR and extracted granted_scopes from the token response.(v1.28.0)
• Ensured writing to the real file if config.yaml is a symlink.(v1.28.0)
• Preserved pairings when stopping the gateway.(v1.28.0)
• Sanitized streamable HTTP extension names derived from URLs.(v1.28.0)
• Improved detection of keyring availability errors.(v1.28.0)
• Removed insecure default secret from GOOSE_EXTERNAL_BACKEND.(v1.28.0)
• Added #[serde(default)] to the description field on ExtensionConfig variants.(v1.28.0)
• Restored smart-approve mode.(v1.28.0)
• Made TLS configurable in the goosed agent via the GOOSE_TLS environment variable.(v1.28.0)
• Added max turns configuration.(v1.28.0)
• Implemented comparison of extension configs before skipping add_extension.(v1.28.0)
• Ensured only viewable channels are added to Slack bot context.(v1.28.0)
• Verified platform binary existence.(v1.28.0)
• Replaced the global static output buffer with a per-instance TempDir.(v1.28.0)
• Resolved parameters in the initial message with autosubmit.(v1.28.0)
• Added analyzer extension in the recipe to maintain backwards compatibility.(v1.28.0)
• Scoped empty session reuse to the current window to prevent session mixing.(v1.28.0)
• Prevented abort in local inference.(v1.28.0)
• Restored provider and extensions for LRU-evicted sessions.(v1.28.0)
• Removed double dash in pnpm command.(v1.28.0)
• Fixed an issue where saving and running a recipe was not working.(v1.24.0)
• Fixed local speech/whisper functionality.(v1.24.0)
• Fixed keyring fallback to handle it as a success.(v1.24.0)
• Replaced a panic with proper error handling in get_tokenizer.(v1.24.0)
• Fixed the broken text editor view.(v1.24.0)
• Fixed configuration of add provider custom headers.(v1.24.0)
• Fixed MCP issue where resource content was not being decoded.(v1.24.0)
• Fixed the Flatpak icon.(v1.24.0)
• Fixed missing content from Langfuse traces.(v1.24.0)
• Fixed extension selector syncing issues.(v1.24.0)
• Improved defensive coding for model switching in Claude Code.(v1.24.0)
• Fixed duplicated output in Code Mode caused by filtering content by audience.(v1.24.0)
• Fixed ACP per-session Agent for model isolation and load_session restore.(v1.24.0)
• Fixed isolation of Claude Code sessions via stream-json session_id.(v1.24.0)
• Fixed Windows bundle random failure and optimized the canary build.(v1.24.0)
• Fixed Goose Desktop missing Calendar and Reminders entitlements.(v1.24.0)
• Fixed stripping of the empty extensions array when using a deeplink.(v1.24.0)
• Fixed contrast issue for text-muted elements.(v1.24.0)
• Fixed an issue where the provider engine type was not preserved when editing custom providers.(v1.20.0)
• Improved the output style of Goose review Pull Requests to better match contributor standards.(v1.20.0)
• Fixed an issue where KaTeX was incorrectly treating underscores as subscripts in plain text.(v1.20.0)
• Corrected the skills directory path to use ".config/agents" (plural).(v1.20.0)
• Fixed an issue where extensions were being loaded upfront instead of when the agent starts.(v1.20.0)
• Ensured platform-specific keyboard shortcuts are correctly displayed in the UI.(v1.20.0)
• Prevented an infinite loop of tool-input notifications occurring in MCP Apps.(v1.20.0)
• Updated the MCP Apps _meta.ui.resourceUri to use the nested format.(v1.20.0)
• Prevented duplicate rendering when a tool returns both mcp-ui and mcp-apps resources.(v1.20.0)
• Handled special Claude model responses within the GitHub Copilot provider.(v1.20.0)
• Fixed a ResultsFormat error that occurred when loading old sessions.(v1.20.0)
• Fixed an issue where invalid call tool results were not being validated or debugged properly.(v1.20.0)
• Fixed an issue in CI where the openapi-schema-check job was missing dependencies.(v1.20.0)
• Fixed Google provider logic to treat signed text as regular content during streaming.(v1.20.0)
• Fixed Azure OpenAI API key being required when it should be optional to match documentation.(v1.19.0)
• Fixed code execution to now support bracket notation and multiple import styles.(v1.19.0)
• Correctly handled interleaved streaming chunks in OpenAI tool calls.(v1.19.0)
• Fixed handling of tool names that do not contain underscores.(v1.19.0)
• Fixed session history order from being modified when viewing a session in the desktop application.(v1.19.0)
• Fixed the MCP UI to send the prompt message when an element is clicked.(v1.19.0)
• Added minute-level granularity for MOIM timestamps.(v1.19.0)
• Fixed the install script to work correctly when piped into bash.(v1.19.0)
• Corrected Bash syntax in terminal integration functions.(v1.19.0)
• Added more open models to the provider list.(v1.19.0)
• Fixed code mode to use MCP structuredContent for proper JS object access.(v1.18.0)
• Improved tool signatures for LLM discovery in code mode.(v1.18.0)
• Fixed the Goose reviewer to produce less sycophantic responses.(v1.18.0)
• Improved the overall quality of the Goose reviewer.(v1.18.0)
• Corrected the parameter name in the `read_module` handler.(v1.18.0)
• Fixed platform extensions to respect the `default_enabled` value.(v1.18.0)
• Fixed conditional configuration for `GOOSE_BIN_DIR` in PATH.(v1.18.0)
• Resolved 400 Bad Request errors when using Gemini 3 thought signatures with Google.(v1.18.0)
• Fixed the tool count warning from showing in code mode.(v1.18.0)
• Added a workaround for cmd's multiline prompt limitations when using Gemini-cli on Windows.(v1.18.0)
• Fixed MCP UI rendering issues caused by CallToolResult structure changes.(v1.18.0)
• Fixed shell output displaying as static text instead of a spinner.(v1.18.0)
• Fixed custom providers with empty API keys from incorrectly showing as configured in the desktop application.(v1.18.0)
• Corrected instructions used for system prompt and prompt for user message in subagents.(v1.18.0)
• Fixed compaction loop issues occurring with small models or large inputs.(v1.18.0)
• Fixed an issue with MCP-hermit cleanup path expansion.(v1.18.0)
• Fixed YAML serialization for recipes containing special characters.(v1.18.0)
• Resolved TokenState loading issues on new sessions.(v1.18.0)
• Fixed ephemeral extensions from being persisted when resuming sessions.(v1.18.0)
• Fixed subagent tools from being incorrectly hidden when running as a subagent.(v1.18.0)
• Resolved a keyboard shortcut conflict for Focus Goose Window.(v1.18.0)
• Fixed filtering of tools in Ollama streaming when chat mode is enabled.(v1.18.0)
• Reverted `/reply` behavior to its previous state (replacing session history) when a full conversation is provided.(v1.18.0)
• Fixed issue where parameters were not being substituted correctly in activities.(v1.16.0)
• Resolved an error from Anthropic regarding "final assistant content cannot end with trailing whitespace".(v1.16.0)
• Fixed configuration issues specifically for the Groq provider.(v1.16.0)
• Fixed an issue where the model was being read from config instead of removing a prompt change.(v1.16.0)
• Used a lock mechanism to ensure the tunnel only runs once, preventing issues from multiple concurrent runs.(v1.16.0)
• Updated the acceptable issues list for moim injection checks to include "Merged consecutive assistant messages".(v1.16.0)
• Fixed Anthropic provider model fetching logic.(v1.16.0)
• Added CSP (Content Security Policy) for images to the proxy HTML in MCP-UI.(v1.16.0)
• Corrected a typo in a blog post reference (AIMDOEL -> AIMODEL).(v1.16.0)
• Adjusted strange spacing issues found in agent.rs.(v1.16.0)
• Reverted a change that prevented loading active extensions when no extensions were present in the recipe.(v1.16.0)
• Fixed multi-tool calling functionality.(v1.16.0)
• Improved handling for multiple content chunks and images.(v1.16.0)
• Fixed an issue where GitHub broke runners for macOS environments.(v1.16.0)
• Added generic checks for Gemini compatibility.(v1.16.0)
• Fixed sessions returning with empty messages.(v1.16.0)
• Added backward compatibility for the conversationCompacted message type.(v1.16.0)
• Improved handling for 404 errors.(v1.16.0)
• Removed setx calls to prevent permanent modification of the Windows shell PATH.(v1.16.0)
• Fixed parsing of maas models for the GCP Vertex provider.(v1.16.0)
• Added support for Gemini 3's thought signatures.(v1.16.0)
• Added error catching for cron jobs.(v1.16.0)
• Fixed reading paths from interactive and login shells.(v1.16.0)
• Applied a quick fix for the community all stars script.(v1.16.0)
• Fixed Linux hermit locking issues(v1.15.0)
• Corrected Windows node path retrieval from registry(v1.15.0)
• Fixed Windows CLI build issues(v1.15.0)
• Resolved CredentialStore issues for authentication(v1.15.0)
• Fixed tool call chain rendering problems(v1.15.0)
• Fixed GitHub Copilot image processing errors(v1.15.0)
• Fixed Developer MCP gitignore handling(v1.15.0)
• Fixed action icons overlapping in chat history(v1.15.0)
• Fixed context progress bar failing to reset after /clear command(v1.15.0)
• Improved Cron error catching(v1.15.0)
• Fixed tray icon not updating with GOOSE_VERSION set(v1.15.0)
• Ensured Recipes respect the quiet flag(v1.15.0)
• Fixed visibility issue for Autocompact threshold UI(v1.15.0)
• Fixed an issue that caused runaway subagent chains by preventing infinite recursion.(v1.14.0)
• Resolved repeated 404 errors that occurred when accessing sessions that had already been deleted.(v1.14.0)
• Ensured that "goose recipe list" now returns only unique entries, eliminating duplicates.(v1.14.0)
• Fixed an issue preventing proper AWS Bedrock credential rotation and refresh.(v1.14.0)
• Set the default mode to Auto for better integration with the Claude Code provider.(v1.14.0)
• Replaced a broken default model configuration with haiku-4.5 for Tetrate.(v1.14.0)
• Corrected the handling of goosehints/AGENTS.md processing, moving it from the developer extension to core goose.(v1.14.0)
• Fixed an issue where the right-click context menu was appearing empty; standard menu items now display correctly.(v1.14.0)
• Ensured the token count now displays correctly upon load and after message completion.(v1.14.0)
• Switched the Gemini model used for MCP smoke tests from Flash to Pro for improved test reliability.(v1.14.0)
• Improved Claude command-line interface integration.(v1.14.0)
• Prevented minijinja parsing errors when processing customized recipes.(v1.14.0)
• Implemented schema-aware numeric coercion for proper type handling of MCP tool arguments.(v1.14.0)
• Corrected the Nix flake configuration for the main branch.(v1.14.0)