v1.13.0

1.13.0 (2026-10-02)

Features

• add bundled uv python (1ae0f94)
• benchmark and tune MLX prefill step size by Apple Silicon generation (585ae0c), closes #277
• update swift-transformers to aleroot/swift-tokenizers and adjust version (6a12ab5)

Bug Fixes

• isolate bundled uv from workspace and user configuration (b68269d), closes #279
• load attachment previews asynchronously (8f35436)
• parse DuckDuckGo results with SwiftSoup (8925259)
• preserve JSON nulls in MLX tool schemas (e9e42be)
• preserve skill accessibility and refresh UI tests (b751f9e)
• prevent numeric tool arguments from crashing history replay (74ceedd)
• remove unused diagnostics import (494e68f)
• update accessibility element behavior in AttachmentPreview and enhance UITests for clipboard image previews (cc99306)
• update Qwen model ID to the latest version for continuity test (5fe32b8)
• update swift-huggingface package to version 0.12.0 (e0c61bd)
• use authoritative MLX cache status for diagnostics (5262fcd)

v1.12.0

1.12.0 (2026-09-22)

Features

• add support for macOS 27 (#261) (6d2b4db)

Bug Fixes

• ci: avoid untrusted checkout when reading toolchain configuration (c459605)
• tests: temporarily exclude MLXChatSessionContinuationTests and MLXGuardedGenerationTests due to known data races in MLX scheduler/allocator (2e1bc72)
• update sanitizer check to include AddressSanitizer in MLXGuardedGenerationTests (aaf5c4a)

v1.11.0

1.11.0 (2026-09-11)

Features

• add read_document for workspace documents (09ffc4b), closes #245
• make workspace_diff structured, per-file bounded, and untracked-aware (#248) (1f6d76c)

Bug Fixes

• block duplicate run_command calls within one assistant response (95419bd), closes #199
• copy image from macos screenshot preview (7772fdd)
• enhance AgentFeature with connection configuration revision tracking and improve refresh logic (da93362)
• imported attachment files survive removal and chat/workspace deletion (fc1d44c), closes #246
• recover from MLX errors and drain generation before cleanup (8deb571)
• use compact spacing for transcript code and tool output (93594d0)

v1.10.0

1.10.0 (2026-09-05)

Features

• create a Personal workspace and first chat on initial launch (0eb4a68), closes #230
• preview persisted skills from chat history (87223f8)
• support full-document attachments and context budgeting (cc1e3a3)
• unify generation and reasoning indicators with Flow animation (e5e4b5d)

Bug Fixes

• add bounded process I/O for commands, diffs, and MCP stdio (de35967), closes #238
• ci: install pinned SwiftLint through SwiftPM (0be6616)
• ci: repair action pins and pinned just setup (0caeab5)
• remove application context cap from generation (7c3f039)
• treat blank list_files paths as workspace root (ab849b2)

v1.9.0

1.9.0 (2026-08-28)

Sumika v1.9.0 focuses on greater control, broader file support, and a smoother chat experience:

• Reasoning controls: Users can select the model’s reasoning effort to balance speed and answer quality.
• Improved skills: Skills can be referenced directly in messages and discovered across multiple configured locations.
• DOCX attachments: Word documents can now be attached and used as chat context.
• Better transcript interaction: Links and newly attached images are clickable, with improved readability and scrolling behavior.
• Performance and reliability: Transcript scrolling uses less CPU, images are correctly oriented for vision models.

Features

• add debug-gated MLX memory tracing (9fda3d0), closes #225
• add first-valid multi-root skill discovery (a064355)
• add selectable reasoning effort levels (935d73e)
• add support for skill mentions in user messages (c2d0c36)
• centralize model generation settings (f9fbb2b)
• enhance content classification for CSS and TypeScript, add new test cases (0b8e82c)
• implement link opening functionality in transcript views (56e25dd)
• set Qwen 3.8 reasoning effort to medium (ea22cf1)
• support DOCX chat attachments (#228) (98b3c22)

Bug Fixes

• add viewport width change handling and update streaming revision test for scroller styles (fde9901)
• correct formatting of FluidAudio package URL (c32613f)
• enhance download progress reporting in model downloader tests (019385c)
• enhance scroll view behavior and add viewport width handling (86b93e1)
• improve transcript readability (48969d4)
• make new image attachments clickable (81abfc0)
• normalize EXIF orientation before VLM image inference (a18e8b8), closes #224
• prevent Markdown responses from being misclassified as raw code (ccab9c3), closes #182
• restore notarized release builds (0e8939d)
• set explicit width constraint for standalone cell in transcript tests (d18f1e1)
• set max tokens to 32k for qwen models (125ef39)

Performance Improvements

• instrument streaming flush cadence (3c1fd15), closes #226
• isolate resource monitor updates to sidebar footer (2bfac22)
• reduce transcript scrolling CPU usage (4f0a3af)

v1.8.0

1.8.0 (2026-08-22)

Features

• add agent skill activation and resource access (2e69e30)
• add support for images in optiq models (153c78f)
• make workspace_diagnostics a format-agnostic command-output reader (58c52c4), closes #186
• remember interaction mode for new sessions (2d8de6e)

Bug Fixes

• clarify edit_file overlap contract (dfdf146)
• fail closed on rejected MLX tool calls (c8c98c1), closes #216
• increase maximum image file size limit to 20MB (8adcb82)
• keep the cache invalid when reasoning opens without emitted text (2983a77), closes #194
• recover from tool calls in final responses (a5f26ee)
• tests: update wait condition to check engine activity state (9e598f5)
• update scope access in ActivatedSkill and adjust tests for consistency (7859975)

v1.7.0

1.7.0 (2026-08-16)

Features

• add global tool approval mode in settings (fe45dbf)
• add Qwen 3.8 27B Optiq 4-bit model (7809241)
• add thinking budget (#211) (4b4721b)
• Implement model deletion functionality with confirmation alert (73b4ccb)
• implement session pagination in WorkspaceSidebar (24aae43)
• refactor model management to include group and recommendation attributes (906343c)
• update app icons with new designs (e9707fd)

Bug Fixes

• runtime: stabilize long MLX generations (aa482b9)
• tests: update error message for tool call crash in MCPClientTests (94e8ee6)
• update supportsImageInput to false in ManagedModelCatalog (3d8e74b)

v1.6.0

1.6.0 (2026-08-08)

Features

• add security policy and reporting guidelines (8124950)
• add total duration summary to chat messages and improve duration formatting (1c3a4b7)
• enhance argument name validation and repair for tool requests (d3f4f7f)
• respect .gitignore in workspace file discovery (d6a4575), closes #196

v1.5.0

1.5.0 (2026-08-02)

This release makes Agent mode faster, safer, and more reliable—especially during longer coding tasks with multiple tool calls.

Highlights

• Faster Agent workflows
  Sumika can reuse more of the MLX model context between tool calls, reducing the wait before follow-up responses.

• Better Qwen 3.6 continuity
  Qwen can retain completed reasoning across tool calls, reducing unnecessary repetition during longer tasks.

• Safer file editing
  Multiple edits to the same file can now be validated and applied atomically. If any edit is ambiguous or invalid, the file remains unchanged.

• More reliable task completion
  Agent mode provides clearer warnings near its tool limit and better explains when unfinished actions were rejected.

Additional improvements

• Markdown rendering in user messages
• Newest chats shown first in the sidebar
• Improved transcript spacing and scrollbar layout
• Support for Qwen 3.6 35B A3B OptiQ 4-bit
• File search can target an individual file
• Clearer line formatting in file-reading results
• Improved generation tracing and diagnostics
• Reduced unnecessary MCP server updates during active conversations
• More reliable application-update generation

Features

• add generation progress tracing and related settings to model management (087c7c9)
• allow search_files path to target a single file (9b50126), closes #179
• implement Qwen3.6 preserve_thinking correctly (2269c8b), closes #177
• render markdown in user bubble (94bb7a5)
• return a bounded applied-edit receipt from edit_file (c773a8e), closes #183
• show newest chats first (8f21771)
• support atomic same-file edit_file batches (8bd3583), closes #198

Bug Fixes

• adjust insets for user and assistant message cells to ensure consistent layout (b41506e)
• change build configuration to Debug in lint-analyze task (c961ff1)
• datamodel generation (#175) (76d81f2)
• don't fail tests if directory was not deleted (8f7df74)
• finalize transcript after output-limited tool batches (3b2356e)
• harden finish_task-only budget finalization and explain fallback cause (efa7588), closes #180
• implement tool budget warning and update maxToolLoopIterations to 20 in agent mode (fc62b2b)
• keep transcript scrollbar full height (7342e61)
• mlx cache reuse (#189) (56037cf)
• prevent unnecessary MCP server reconciliation during active conversation (37ed95d)
• remove unnecessary blank line in ManagedModelCatalog (bd89465)
• update maxToolLoopIterations to 15 and remove outdated tool loop budget test (3803cad)
• update maxToolLoopIterations to 18 for multiple ManagedModel instances (fa5d612)
• update source packages path handling in appcast generation scripts (bdab66a)
• use unambiguous colon-space gutters for read_file output (213af10), closes #191

Performance Improvements

• avoid generation progress tracing overhead when disabled (5b177ee), closes #181

v1.4.1

1.4.1 (2026-07-26)

Bug Fixes

• adjust minimum heights and vertical insets for transcript cell kinds (b362401)
• default legacy focused file completeness (b821945)
• ensure finish_task is called as the last tool (38bb9e0)
• format (6b73e0c)
• handle cancellation in model stream processing (681c10e)
• implement finish_task handling for tool budget exhaustion scenarios (1212513)
• trim whitespace for assistant message visibility in transcript (6c69187)
• truncate long tool headers in transcript (0af3006)
• update approval terminology for consistency and accessibility (d5b3cc5)
• update transcript cell insets and implement header action handling (817bc15)