✨ Features
- Automatic thinking level operation for improved model reasoning #12237
- Resume failed or cancelled delegate subagents #13015
- Subagent session link in UI with view-only subagent sessions #12998
- Limit concurrent subagents with
GOOSE_MAX_CONCURRENT_SUBAGENTS#12946 - Report session status via OSC 7501 in CLI #12928
- Windows NSIS installer for desktop #12869
- IO Intelligence provider #12795
- Provider-defined cache prices support #12944
- Dangerous include history option #13051
- Replace PCTX code mode with Boa JavaScript runtime #12952
🐛 Bug Fixes
- Windows build error #13081
- Omit stream_options for Databricks Gemini model services #12675
- Stale models issue #13054
- Show model changes when provider reports no resolved model #12914
- Show why an update install failed in desktop #12628
- Keep
!in npx arguments on Windows #12680 - Inherit active chat folder for New Chat #12941
- Bound long conversation rendering in desktop #12976
- Drop ActionRequired and ToolConfirmationRequest blocks for Bedrock instead of sending blank text #12843
- Missing GLM models in OpenCode Go and Zen #12057
- Load the live model catalog in goose-acp #12962
- Load recipe extensions for slash commands and delegation #12938
- Use the context length OpenRouter reports #12853
- Context freeze preserved message must stay user-visible during compaction #12692
- Do not present compaction countdowns to models #12902
- Keep the agent's model list order in ACP #12823
- Send session id as prompt_cache_key on OpenAI Responses requests #12420
- Map Moonshot provider name to moonshotai for canonical model lookup #12041
- Tolerate missing created_at on OpenAI Responses stream events #12241
🔧 Improvements
- MCP tools now have provider-safe public names #12875
- Use global turn limit as subagent turn limit fallback #12992
- Optimise list session database query #12940
- Skills discovery moved into GDK #12935
- Run foreground subagents in the GDK #12792
- Leave the login-shell PATH to whoever launches goose #12799