aaif-goose/goose v1.46.0

on GitHub

✨ Features

• Unrolled agent loop #9574
• Cache-safe request assembly with append-only turn context and declared cache semantics #11022
• Streaming shell output while commands run #10808
• Per-message usage stats UI (tokens, cost, TTFT, tok/s) #10210
• Per-message usage/cost tracking with derived session totals #10172
• Track cache tokens for accurate cost reporting #9752
• Markdown format option to session export #10703
• Resizable sidebar with drag handle and persisted width #10795
• Interactive menu for single select elicitations #10327
• Session metadata on sidebar chat hover #10663
• Configurable LOCAL_WHISPER_LANGUAGE for multilingual local transcription #10634
• Allow disabling built-in skills #10600
• Structured summary output with template rendering for compaction #10471
• Pass working_dir to the Stop hook context #10296
• Surface output-token-limit info #10831
• Tab completion and switch provider for model command #10585
• Collapse const-union enums in tool schemas #10577
• GenAI semantic convention attributes for OTEL #10700
• Enrich root span with gen_ai attributes and improve OTEL output format #10816
• Forward images and MCP embedded-resource blobs in Anthropic and Google formats #10340
• ACP session/new supports Hidden session via _meta.hidden #10811
• Stable agent event message identity #10716
• OAuth logo_uri in client ID metadata documents #10727
• Termux detection in installer to select musl portable build #10568
• Hooks feature with PreToolUse denial #9093, #9304
• /goal command for agent self-evaluation before finishing #9069
• /status slash command #9845
• /model slash command to switch session model #8747
• --edit session flag to edit conversation before forking #9799
• goose review local code review command #9114
• Open plugins generalization with skills support #9112
• Summon extension: peek mode for async background tasks #9519
• Summon extension: context and working_dir parameters for delegate tasks #9518, #9520
• Load global hints from ~/.agents/AGENTS.md #9736
• Proactive OAuth token refresh to avoid re-auth on every session #8386
• GOOSE_OAUTH_CALLBACK_PORT for stable OAuth redirect_uri #9209
• Unified thinking effort control across all providers #9242
• Projects as backend sources with system prompt injection #8739
• Group chat sessions by project in nav panel #10212
• Collapsible project group headers in nav panel #10278
• Worktree-aware directory switcher #8450
• Chat history search feature in navigation panel #8448
• Search filter to provider grid #10437
• Option to disable automatic update downloads #9872
• TUI diff viewer #9260
• MAX_CODE_BLOCK_LINES configurable via env vars #9301
• GOOSE_MAX_TOOL_RESPONSE_SIZE to configure tool output size limit #9256
• GOOSE_FAST_MODEL env var in ModelConfig::with_fast #9296
• GOOSE_DOCS_ROOT for air-gapped docs access #10294
• Quarterly option for scheduler #9076
• Harbor eval runner #9138
• goose://new-session and goose://resume deep links #9196, #9343
• Initial prompt with goose://new-session #9427
• Encrypted Nostr session sharing #8922
• Import sessions #9474
• Hugging Face OAuth support with auth tab in settings #9552
• MLX models support in local inference provider #9154
• Linux Vulkan support for local inference #9038
• Reconnect desktop ACP sessions after sleep and connection loss #10411
• TLS support for ACP serve #10088
• Group consecutive tool calls into one summarized chain card #8995
• Bind MCP apps to trusted ownership metadata #10747
• MCP extensions in open plugins #9471
• Skills to the chat composer #8881
• Dynamically refresh skill instructions each turn #9217
• CLI to list skills with token counts #9326
• Image read tool #9607
• View model interactions #9806
• UniFFI SDK for cross-language bindings #9593
• Provider bindings MVP to goose-sdk with Python wheel publishing #10208
• Kotlin FFI with build and Maven publish #10375
• Create source archive as release artifact #10786

New Providers

• Celeris provider #10714
• Friendli declarative provider #10762
• Azure AI Foundry multi-LLM provider #10622
• OllamaCloud provider with dynamic model discovery #10264
• Sakana AI provider for OpenAI-compatible Fugu API #10357
• iFlytek Spark and Astron MaaS providers #9837
• Fireworks AI declarative provider #9990
• OpenRouter request parameters support #9276
• Together AI declarative provider #9545
• OrcaRouter declarative provider #9515
• Perplexity declarative provider #9324
• Alibaba (Qwen via DashScope) declarative provider #9443
• Databricks AI Gateway provider #9274
• Scaleway provider #9254
• NEAR AI Cloud provider #9352
• EmpirioLabs declarative provider #9771
• xAI SuperGrok OAuth subscription provider #9420
• Vercel AI Gateway #9144
• Routstr declarative provider #9175
• FuturMix provider #8840
• oMLX declarative provider #9177
• Atomic Chat declarative provider #9210
• Muse Spark 1.1 via Meta Models API #10432
• AWS Bedrock support for OpenAI GPT models via mantle endpoint #9707
• Azure AD (Entra ID) bearer token auth via AZURE_OPENAI_AD_TOKEN #9716

New Models

• GPT-5.6 support #10384, #10349
• GPT-5.5 to ChatGPT Codex known models #9292
• GLM-5.2 support #9967
• Opus5 support with adaptive thinking #10686
• Claude-sonnet-5 and claude-fable-5 to known models #10865, #10183
• Cursor composer-2.5 models #10618
• Cursor-agent models from CLI #10367
• Latest Gemini models #10630
• Gemini 3.x known locations and KNOWN_MODELS entries #9142
• MiniMax-M3 and M2.7 variants #10431
• OVHcloud Qwen3.6-27B model #10339

Internationalization

• French, German, Italian, Portuguese, Indonesian, Malay, Vietnamese, zh-TW desktop locales #10072
• Korean localization #9856, #10667
• Spanish (es) locale #9833
• Japanese locale #9768
• Hindi Desktop locale #9733
• Russian language support #9406
• Turkish desktop locale #9392
• Simplified Chinese (zh-CN) translation #8765
• Language selection feature #9405

🐛 Bug Fixes

• Reject invalid subrecipe content #10994
• Keep turn-context in place for OpenAI Responses-stack models for prompt caching #10993
• Byte-bound the shell truncation preview #10992
• Stop silently ignoring thinking effort off for OpenRouter #10991
• Dispatch edited queued messages #10933
• Contain recipe template paths #10930
• Make shell approval titles faithful #10986
• Block MCP app form submissions #10985
• Sanitize Bedrock tool errors #10934
• Implicit OAuth during model discovery #10929
• Keep ACP session naming out of live conversations #10963
• Sanitize Unicode tags in MCP resources #10746
• Preserve RFC 9207 iss from MCP OAuth callback #10678
• Shell ACP providers on desktop #10907
• Stop killing streaming responses at the total request timeout #10620
• Restore final_output_tool when agent is recreated after LRU eviction #10793
• Avoid O(n²) cloning during session load #10665
• Prefix child stderr log target so it passes the default goose=info filter #10901
• Fall back to static model list only for non-models payloads #10189
• Stdio extensions silently skipped when name missing or env: used in config #10773
• Start scheduler at startup instead of first ACP connection #10766
• Respect disabled builtin extensions (developer) at session start #10223
• Parse PyPI requirements for OSV checks #10510
• Contain voice file extensions in telegram #10456
• Preserve denied tool request precedence #10612
• Contain subdirectory hint discovery #10545
• Clear stale validation error when reopening the schedule modal #10627
• Sanitize shell/subprocess call in linux.rs #10748
• Unlink destination before copying binaries #10705
• Honor timeout_seconds on the Anthropic declarative provider #10440
• Allow removing images when editing user messages #9979
• Preserve thinking inline for Cerebras models that reject reasoning_content #10774
• Make CLI streaming render incremental instead of O(n^2) #10487
• Together provider: parse model list response #10576
• DeepSeek model names to v4 API #10729
• Forward Claude Code and Pi model picker selection to the adapter #10675, #10707
• Preserve enabled flag when migrating platform extensions #10690
• Pass session working_dir to on_call_tool #10689
• Preserve adversary tool arguments #10529
• Keep XML snippets inside JSON as data for local-inference #10606
• Sync generated chat title in header #10578
• Prompt for tool approval in Telegram gateway sessions #10613
• Keep CLI provider prompts out of process arguments #10547
• Ignore stale inventory for static model lists #10591
• Preserve featured model size on delete and backfill missing sizes #10422
• Create inventory tables atomically with schema version #10586
• Rewrite oneOf to anyOf in tool schemas for OpenAI-compatible backends #10571
• Allow custom model as default for non-local providers #10438
• Require absolute goose path roots #10454
• Scope smart approval by request #10457
• Preserve fixed subrecipe values #10452
• Bundle git so hermit can clone its package registry in Flatpak #10511
• Forward image content chunks to client during live ACP session #10485
• Dispatch Code Mode callbacks on main runtime to prevent hangs #10484
• Protect OAuth token caches #10516
• Preserve content audience boundaries #10493
• Bound scheduled recipe validation #10509
• Return JSON-RPC errors from dispatch handlers instead of killing the connection #10499
• Avoid local inference Metal teardown crash #10508
• Enable indexmap serde support for musl builds #10507
• Scroll chat to the latest reply on session resume #10501
• Secure large response spill files #10482
• Register LiteLLM with refresh_only inventory so model picker lists models #10489
• Limit provider selector height in CLI #10420
• Tighten existing secrets file permissions #10453
• Classify image access capabilities in developer extension #10451
• Preserve one-time permission scope #10450
• Register goose:// scheme handler in Flatpak desktop entry #10461
• Reject duplicate recipe parameter keys #10448
• Synchronously kill orphaned goose serve backends on GUI close #10292
• Break self-referential $ref cycles before Code Mode schema #10469
• Stop multiline paste auto-submitting line-by-line on Windows #10326
• Hide nested config values in Configuration Editor #10390
• Stop resolving TUI scripts from cwd #10449
• Exclude MCP App tools from Code Mode callbacks #10290
• Apply the default extension timeout to developer shell tool calls #10348
• Preserve reasoning_content for DeepSeek thinking mode on multi-turn tool calls #10366
• Make prompt timestamps timezone-explicit #10209
• Compact disabled during approval #10089
• Treat empty-string finish_reason as non-terminal in streaming #10258
• Gemini OAuth onboarding #10363, #8905
• Retry empty provider turns instead of stopping silently #10360
• Guard llama.cpp backend init against SIGILL on x86_64 CPUs without FMA/AVX2 #10105
• Sanitize responses function call names #10344
• Pick session mode from the ids the agent actually offers #10320
• Dedupe duplicate signed thinking blocks to fix Anthropic 400 #10083
• Resolve provider default model instead of leaking sentinel "current" #10305
• Return effective context limit from /model-info #10165
• Gate rcgen aws_lc_rs feature behind rustls-tls #10116
• Tolerate partial Responses output items #10108
• Always use OpenAI-compatible "parameters" for tools in Databricks #10101
• Unescape shell-escaped image paths #10098
• Seed mac auto-update config #10132
• Send inference config (max_tokens, temperature) on Bedrock Converse #9889
• Dedupe Nostr session deep link imports #9918
• Don't retry deterministically-permanent 400s (thinking-block immutability) #10005
• Use async reqwest client so OTLP export works in goose serve mode #10100
• Reject non-object tool-call arguments instead of panicking #9832
• Save /edit prompts to history #10011
• Use session.message_count for schedule sessions listing #10026
• Support streamed tool calls without indexes #10023
• Disable extended thinking for fast-model operations #9815
• Buffer thinking across stream chunks for DeepSeek/Kimi tool calls #9704
• Surface actionable errors for truncated tool-call arguments #9946
• Show optimistic toggle state when disabling an extension #9882
• Resolve databricks_v2 pricing and surface cost in standard usage update #9925
• Publish mac updater metadata #9945
• Activate hermit on macOS so STDIO extensions resolve node #9482
• Accept JSON-encoded string for autovisualiser data parameter #9838
• Send User-Agent on URL image fetches in developer/read_image #9927
• Do not pass temperature to ChatGPT Codex provider #9931
• Sanitize extension environment maps #9884
• GCP: use refreshed gcloud token after reauth (retry Vertex AI on 401/403) #9849
• Start a turn when /goal or /grind is set #9801
• Pass thinking config to Bedrock Anthropic models #9794
• Keep extensions icon visible after switching sessions #9787
• Skip non-recipe project config files in summon #9808
• Dedup duplicate tool-call ids within a turn #9792
• Resolve bundled extensions from discovery #9759
• Custom OpenAI provider dropping port when URL scheme is omitted #9730
• Preserve Windows backslash paths in custom extension command #9741
• Include agentInfo in ACP initialize response #9765
• ACP cancel race condition #9804
• Inherit login-shell PATH in spawned subagents #9737
• Accept string values for GOOSE_CONTEXT_LIMIT #9738
• Clear rejected OAuth credentials after refresh #9694
• Keep extended thinking within the Anthropic output cap #9814
• Disable RPM build-id links #9671
• Preserve custom API path in base_url derivation for OpenAI #9649
• Handle resume deep links #9298
• Repair PATH for plugin hook commands #9615
• Restore Cmd/Ctrl+T new chat shortcut #9614
• Lead with extension description, shorten HTTP/SSE transport label #9653
• Mention menu icon to open the sidebar #9643
• Dedupe React to prevent multiple instances in renderer #9631
• Add hermit cleanup to uvx and jbang scripts #9616
• Use blocking OTLP HTTP exporter #9599
• Fall back to default provider when resuming session with unavailable provider #9547
• Provide Tokio runtime for OTLP HTTP exporter batch thread #9541
• Read meta.n_ctx from /v1/models to fix context limit for local servers #9530
• Align session history scrollbar with window edge #9601
• Refresh session name in sidebar after rename #9543
• Show resolved skill supporting file paths #9584
• Implement real Bedrock streaming via ConverseStream API #9579
• Record OpenAI-native cached_tokens in usage metering #9829
• Add option for overriding API URL for Moonshot provider #9819
• Detect image paths with spaces #9387
• Restore new chat shortcut navigation #9528
• Evict stale completed tasks to prevent unbounded memory growth in summon #9514
• OAuth: time out the MCP callback wait instead of hanging in WSL #9536
• Warn on unmatched extension names in summon delegate #9513
• Deleting a recently created session and then creating a new shows the deleted session's content #9351
• Propagate elicitation decline and cancel actions end-to-end #9437
• Close modal when canceling recipe parameter form #9195
• Ensure tool request timestamp precedes tool response timestamp #9462
• Correctly map ollama_cloud to canonical provider #9639
• Incorrect max tokens values for groq provider #9790
• Mark stream decode errors retryable #9723
• Classify Bedrock ValidationException as ExecutionError #9735
• Page through all Databricks AI Gateway v2 endpoints when listing models #9753
• Preserve unparseable extension entries during config refresh #9439
• Sync slash help with builtins in CLI #10173
• Treat unicode punctuation as image path terminators #10106
• Don't bleed provider-specific request_params into delegate() sessions #9906
• Purge empty phantom sessions on startup #9651
• Include structuredContent from MCP tool response #10128
• Propagate agent cleanup errors on session delete #10112
• Prevent login-shell PATH probe from suspending goose on startup #8804
• Exclude preprompt from session title generation #8793
• Unscheduling a recipe should not delete them #8978
• Seed provider handoff history in ACP #8941
• Respect goose2 interface density settings #8953
• Make goose2 respect accent color #8952
• Preserve replay message timestamps #8942
• Goose2 window minimum sizing #8946
• Avoid transform-rasterized dialog text in goose2 #8937
• Keep renamed skills open in detail view #8935
• Model picker stays usable during provider loading #8900
• Handle ACP requests concurrently #8781
• Apply hermit env directly in node shims so a fish login shell doesn't break MCP startup #10028
• Preserve command arguments through extension forms #10527
• Enforce manual approval for code mode #10528
• Confine app file operations #10481
• Preserve recursive schema types with pctx 0.4.1 in code-mode #10572
• Use shell working directory on initial launch #10655
• Enable inventory refresh for OpenRouter model picker #10641
• Preserve read lines and reload tool output in ACP #10662
• Surface tool call output in CLI and Desktop #10654
• Overlong function names in provider requests #10659
• Preserve ACP tool call update fields #10653
• Omit disabled OpenRouter reasoning #10517
• Clear stale pending ACP connection after terminal recovery failure #10552
• Confirm pending steer message on queuedSteer notification #10532
• Restore dynamic model discovery and current Grok support #10756
• Reuse parent provider for summon delegates #10754
• Skip permission mode negotiation for Pi #10701
• Freeze turn-context timestamp at turn start to preserve prefix cache #10734
• Refresh Moonshot model list with current Kimi lineup #10696
• Update Kimi-code known models metadata with current platform lineup #10713
• Avoid double loading full conversation on session open #10794
• Recover malformed tool calls from GLM/Minimax models instead of rejecting them #10230
• Restore cursor when configure exits #4889

🔧 Improvements

• Keep turn-context out of the OpenAI-compatible prefix cache #10937
• Keep turn-context out of the Anthropic prefix cache #10030
• Parallelize provider resolution and eagerly init SQLite pool #8899
• Deduplicate _goose/providers/list RPC call at startup #8873
• Compact tool schema JSON #10409
• Drop debug info for dependencies in the dev profile #9843
• Use LRU cache for token counting #9586
• Mergeable configs with cleanup #8378
• Structured per-provider config block with non-destructive provider switching #8977
• Linux musl CLI builds #9240
• Windows CUDA release artifacts #8750
• Linux desktop Vulkan packages #9323
• Migrate desktop routing to React Router 8.3.0 #10971
• Upgrade to rmcp 3.0 #10789
• Handle MCP tool list change notifications #e7088ee
• Redesign Extensions page #8940
• Redesign skills library #8868
• Dark mode inline code styling for message bubbles #10861
• Refine dark theme message surfaces #10693
• Inline code contrast improvement in light theme #9058
• Polish sidebar navigation and project icons #8896
• Simplify UI customization #9353
• Better parsing of pasted HTML as markdown so agents understand #9190
• Remove CLI project support #10838
• Deprecate and remove ui/text TUI #10799
• Remove MCP sampling support #10087
• Make context exceeded checker more precise #9831
• Update canonical model inventory #10352, #9709, #9551, #9331
• Add canonical thinking modes #9743
• Nushell terminal and completion support #8628
• Migrate diagnostics to JSON report #9964
• Route MCP elicitations through tool streams #9943
• Allow unlisted models in search #9933
• ACP paginate session list #9199
• Flush OTLP traces reliably on exit with configurable timeout #9228
• Increase /mcp-app-guest body limit for larger single-file MCP apps #10324
• Use current_exe() instead of PATH lookup when spawning goose #9236
• Set TCP_USER_TIMEOUT on streamable HTTP clients in extension_manager #9207
• Fail closed on unverifiable update provenance #10168
• Trim macOS desktop entitlements #10490
• Consolidate and harden release workflows #10447
• Update Rust toolchain and raise recursion limit #10303
• Enhance the uniffi API layer #10427
• Bound code-mode execution with timeout and cancellation #10214
• Switch local inference MLX backend to use safemlx crate #10304
• Preserve thinking content for providers that require it #8857, #9314
• Build summon instructions per turn #9329
• Refactor local inference around backends #9137
• Stricter GGUF requirements with auto detection of tool calling support and fixed thinking output parsing #9442
• Prefer goose aliases for Databricks v2 inventory #9430
• ACP migration: UI connect to ACP directly instead of goosed #10081
• Unified OTLP logging schema for cross-tool detection #9713
• Pattern-based detection updates for prompt injection #9350, #9198, #9690, #8546
• Enable prompt injection mitigation by default for internal users #9612
• OpenAI Responses API store param configurable #10040
• GitHub Copilot /responses API #9043
• Check for responses API support in Databricks #9347
• Honor Retry-After on 429 responses #9161
• Coalesce streaming Thinking deltas and list available tools on not-found #9162
• Strip chain-of-thought markers from custom provider output #8635
• Coalesce streaming chunks under one message id in ACP #8788

📚 Documentation

• Document auto-injected GOOSE_SHELL flags #10759
• Clarify macOS sandbox feature was removed in post-v1.25.0 cleanup #10900
• Update Discord invite #10863
• Update environment and config references #10203
• Document hook deny contract, matcher regex, tool_input keys, and provider config #10467
• Add Code of Conduct #10424
• Adopt issue-first contribution workflow #10819
• Update docs for ACP clients #9772
• Hooks guide #9288
• Custom agents guide #9293
• Guide for connecting goose Desktop to a remote goosed server #9275
• Document summon extension requirement for delegate and load tools #9231
• Refresh Ophis extension page (tool list, chains, skills link) #10681
• Clean up stale documentation audit findings #10114
• Reorganize docs #9310
• Blog: Goose with Peekaboo #8884
• Blog: Built-in Local Inference #8808
• Blog: Orchestrating with Goose #9104
• Blog: Self-improving agents #9846
• Blog: AAIF migration #9866
• Blog: MiniMax and Office QA benchmark findings #8984