Releases

Seventh open-beta build. Source Control is now a first-class app, chat transcripts can switch between Compact and Full views, and context usage is easier to read.

Highlights

• Source Control app — Dedicated rail app for changes, history, branches, worktrees, pull requests, and CI. Opens without clearing your Code chat.
• Compact and Full chat views (#1194) — Compact groups per-turn Working/Worked activity and shows edited-file summaries with inline diff review; Full keeps every transcript row visible.
• Context usage — Improved usage ring, stats strip, and web content extraction for clearer context budgeting.
• Default model configuration — New chats and the menubar model chip resolve the configured default provider/model consistently.
• Issues — Clearer closed-issue handling and list/board indicators.
• Orchestrate — Board baseline now returns seq for reliable reconnect sync.

Changes and fixes since v0.1.2

Source Control

• Released core app at #/app/source-control with the full git surface (changes, history, branches, stashes, worktrees, PRs, CI).
• Code left-pane Source Control toggle still works; the rail app opens the same capabilities in a dedicated view.

Chat and context

• Settings → Appearance → Chat view: Compact (default) or Full, persisted across restarts.
• Per-turn file-change summaries and inline diff review in Compact mode.
• Token/stat formatting handles trillion-scale counts.

Models and settings

• Default model binding management across Settings, menubar chip, and new chats.
• Full-width skill preview buttons; Impeccable skill sync updates.

What’s Changed

A broad update across model routing, browser automation, Issues, Orchestrate, chat reliability, and Git workflows. This release also improves tool loading and streaming, workspace handling, and the desktop build pipeline.

Models & providers

• Model pools — share capacity across models and providers, with automatic failover and clearer quota-exhaustion notifications.
• Multi-GPU GGUF loading — split GGUF models across selected GPUs from the Models inspector.
• OpenCode Go support — route requests with the correct provider identity and headers.
• Improved OpenAI-compatible response handling, provider discovery, and utility-model routing.

Agent Browser & tools

• Agent Browser — added browser automation with an integrated viewer, runtime controls, watchdog handling, and expanded test coverage.
• Lazy tool schemas — tool definitions can load on demand, reducing initial prompt weight while preserving list-only discovery.
• Added configurable context injection settings and improved tool streaming, output limits, slicing, caching, and transcript presentation.
• Improved Impeccable integration and built-in skill chips.

Issues

• Custom issue types now support colors and icons.
• Expanded label management, attachment handling, issue expansion, and run-target selection.
• Improved GitHub two-way synchronization, workspace handling, deletion behavior, and error reporting.
• Issue detail now includes linked chats and boards plus a collapsible files panel.

Orchestrate

• Added Reset and Rewind board actions.
• Added richer end-of-run reports, task evidence, notes, and file presentation.
• Hardened board recovery, merge queues, rebase seeds, worktree cleanup, and operational merge diagnostics.
• Improved task detail and attempt transcript handling.

Chat & context

• Improved streaming paint behavior, scrolling, thought persistence, and transcript rendering.
• Stopped turns keep their context when continued.
• Refined context trimming, screenshot token pricing, usage metrics, and the context ring.
• Improved per-chat question handling, mode refresh behavior, message actions, and queued activity.

Git, workspaces & desktop

• Improved Git feedback, operation errors, branch/worktree creation, and merge-failure diagnostics.
• Gitmoji shortcodes expand to Unicode in commit messages.
• Released apps can open in separate windows.
• Improved Electron workspace management, launch handling, and packaging validation.

UI, docs & maintenance

• Improved Markdown link handling and previews.
• Refined Settings layouts, model-pool controls, injection settings, and related app styling.
• Made generated skill metadata deterministic and updated built-in tool definitions, tests, documentation, and product-wiki metadata.

What’s Changed

A big update for the Issues tracker — label catalog, sub-issues, a redesigned issue detail view, and a much deeper GitHub two-way sync — plus Git Bash terminal support on Windows and server-backed appearance persistence.

Issues

• Label catalog — named labels with color swatches, right-click to recolor, up to three chips plus an overflow popover, and a + add flyout in list and board views.
• Sub-issues — one level of nesting: add/unparent children from the peek, or drop an issue onto another issue to make it a child.
• Issue detail redesign — sticky identity header, property chips, and workflow; resizable peek with full-screen sheet mode; secondary sections (Code, Attachments, Git, Plan, Related) collapse to add-rows when empty.
• Comments & activity — agent comments and activity render in issue detail.
• Chats section — issues link the chats/boards they were created from, with Running/Done state and unlink.
• Type & status icons — glyph picker in Settings → Issues; list and status chips render the glyphs.
• Drag & drop refinements — issue-on-issue parent drops, board column drops, and OS file / chat-capture drops into the detail view.
• New issue form — workspace picker (Scratch + recent workspaces) when the header scope is All workspaces.

GitHub sync

• Sync all button in two-way mirror mode, plus automatic sync: auto-push title/description/labels/closed-state after edits, create cards on GitHub on first edit of an unlinked issue, and a background sync every 5 minutes.
• Labels pushed to GitHub by name (missing labels are created first); dropped names surface in a toast when the repo forbids label creation.
• Conflict handling — Keep mine / Keep GitHub prompts; sync status chips in the peek (synced / needs push / open).
• Sync failures now show friendly local-backend copy instead of backend jargon.

Terminal

• Git Bash support for Windows terminal profiles, plus improved shell command handling (one-shot spawns, PTY environment, sandbox signal handling).

Desktop / windows

• Window retargeting — opening a plan or workspace reuses a window already on that folder instead of spawning a duplicate, with a guard against accidental workspace switches.

Appearance

• Server-backed appearance persistence — theme, fonts, and custom colors persist through the tool server instead of browser-local storage alone.

Settings & Models

• Settings → Models app navigation improvements; the Models app opens directly from settings sections.

Fixes

• Skip the corrupt image_url capability probe on loopback llama.cpp servers, avoiding spurious probe failures on local models. (#1145)

Docs

• v0.1.0 release notes, new plans (Git Bash terminal, issues GitHub auto-sync, sub-issues, labels, status icons, app windows), and manual updates for the Issues app and keyboard shortcuts.

Full Changelog: https://github.com/HenriGrimm/Minnow/compare/v0.1.0…v0.1.1

Sixth open-beta build — first 0.1.x. The board engine, session store, and workspace windows were rebuilt so the build loop holds up overnight, across restarts, and across more than one repo at a time.

Highlights

• Orchestrator V2 — Boards are a journaled server engine, not a chat that mutates a kanban. Pick a plan, start waves of Builder/Tester agents in isolated worktrees, merge when the ladder passes. Crash, reload, and workspace switch replay from the journal. V1 board tools (board_init and siblings) are gone.
• One folder per window — Open several workspaces at once as separate windows. Re-opening a folder focuses the window that already has it. Close asks before hiding to the tray. The workspace picker is a real welcome page.
• Sessions that survive restart — Chat state lives in sessions.db with rotating snapshots. A lazy boot no longer wipes transcripts. Failed turns offer Continue or Clear. Opening the app asks before resuming interrupted chats and boards.
• Shared agent runner — Product chat, boards, and sub-agents share one stream/tool loop. Sub-agent runs persist in their own journals and keep going when the parent chat is elsewhere. Context overflow summarizes and retries instead of hard-failing.
• Chat and Code polish — Queue a follow-up while a turn is running. Drag editor and browser tabs onto chat as durable link chips. Material file-type icons in the tree and tabs. In-app Review PR from Issues and Source Control. Expand an issue into a review overlay before saving.

Changes and fixes since v0.0.5

Orchestrate

• V2 Boards page is the only live orchestrator surface: plan picker, live kanban, task-detail overlay with chat-shaped attempt threads, follow-up chat from the finish report (#941, later V2 landings).
• Worktree slots are named from the board, wave, and task; orphan slots reclaim in bulk; Start fails closed if git init cannot run.
• Last-opened board resumes; running boards show an activity dot on Orchestrate; display-wake polling no longer kills live task chats (#949).
• Broken plan markdown offers Repair (plan-repairer sub-agent) instead of a dead parse.
• Skip the plan picker when a plan is already bound (#1027). Auto-fix illegal worktree and branch names (#958).

Workspaces and shell

• Multi-window: one workspace per view, restored geometry, close-workspace dialog, window-close vs tray preference.
• Welcome page / picker: Sandbox pinned, recents of folders you opened, Create project / Open folder.
• Ask before resuming chats and boards on app open (#953).
• Hardware acceleration setting; GPU work parks while the window is hidden or streaming.

Chat, context, and agents

• Queued follow-ups render in the transcript (#1032). Drag editor/browser tabs onto chat as link chips (#1034).
• Context overflow summarizes and retries; exhausted turns fail visibly with Continue / Clear (#948, #950, #959).
• In-flight tool rows stay visible across a chat switch (#944). Interleaved think tags stay off the reply (#957).
• Sub-agents adopt the shared runner; spawn cards re-anchor after reload; report nudges for boards and sub-agents.

Sessions and reliability

• Canonical store is SQLite sessions.db; rotating verified snapshots replace the JSON mirror.
• History-omitting writes cannot wipe transcripts; chat deletion is never inferred from absence.

Issues, git, and Code

• Sparkles Expand with a review overlay — nothing saved until Apply (#1033).
• In-app Review PR from Issues and Source Control.
• GitHub issue import no longer freezes the shell (#961).
• Material Icon Theme in the file tree and viewer tabs; Files / Browser / Source Control pane cluster (#1035).
• File-tree drag-and-drop move (#939). Word wrap on git commit review (#1029).

Models, terminal, and internals

• Vision probe on first model load (#960). MLX load and inference meters. Loading progress on Local Server.
• llama.cpp offload-slot handling; default reasoning options for openai-v1.
• Terminal: zsh history no longer shows ^A^K (#956); clipboard paste / copy shortcuts; Windows ancestor tracking.
• Dev servers: do not inject Vite --port into non-Vite CLIs.
• Appearance: Google Fonts catalog (selected pair only); wallpaper section removed (#1037).
• Est. cost metric cells no longer overlap (#937). Configurable tool-result size limits. Landlock path-list caps.
• Catalog is 105 built-in tools (V1 board tools removed).

Fifth open-beta build focused on Issues v2, local llama.cpp hosting, compact Code composer, vision and thinking models, and agent-loop reliability.

Highlights

• Issues v2 (#892) — Capture from anywhere (menubar, Alt+C, right-click, drag-and-drop), grouping/rank/triage, markdown-preserving editor, assign-to-agent via Orchestrate, and GitHub sync with explicit conflicts.
• Local llama.cpp (#891) — Auto-fit context/offload, crash/OOM diagnosis with retry, resumable downloads, split-GGUF shards, multi-serve LRU.
• Compact composer (#915) — Narrow columns: mode dropdown, overflow cog, model chip, context wheel. Cog is a two-page sheet (this-turn settings → Tools).
• Qwen3.8 + thinking (#878, #872) — Host Qwen3.8-27B; composer High → xhigh; thinking-budget cuts continue from prior reasoning, scoped per turn.
• Agent harness (#875) — Stop mid-batch no longer crashes; unpaired tool calls are repaired; shell/MCP writes bust the file cache; rate limits retry.
• Web search (#898) — Decoy Bing results filtered; DuckDuckGo + Stack Overflow in general; optional deep_read.
• Brain reindex (#855) — Full reindex completes on real repos; PageRank persists; UI polls a background job.

Changes and fixes since v0.0.4

Issues

• Issues app v2: saved views, grouping, rank, triage, sub-issues, inline editing, keyboard chrome, type icons in Settings (#892).
• Capture: menubar, Alt+C, right-click, drag-and-drop onto rail / list / board.
• Description editor round-trips untouched markdown blocks byte-for-byte.
• Assign to agent via Orchestrate; GitHub sync with conflict resolution (manual).

Models and hosting

• llama.cpp auto mode: --fit on, no -ngl, -c on a 32k-preferred ladder capped by train context (#891).
• Serve crashes/OOM visible and retryable; downloads resume; split GGUFs; multiple resident servers.
• Memory from attention geometry (GGUF header), not parameter count — sliding window and partial offload modelled (#854).
• Qwen3.8-27B catalog, llama.cpp b10448, --jinja + mmproj; High maps to xhigh (#878).
• llama.cpp /props reports live context window and vision.
• Models runtime log scrolls inside the OS stage instead of clipping (#882).

Chat, composer, and Code

• Compact composer + two-page overflow cog (#915).
• Jump chip and code-change strip no longer overlap; Commit / Create PR beside undo.
• Right pane keeps the file viewer when a preview tab stays open; preview stays painted beside stage views.
• Streaming bubbles paint incrementally instead of only at completion (#868).
• Drop folders from Explorer/Finder into the file tree (#886).
• Clipboard image paste no longer duplicates; composer spellcheck and auto-resize.

Agents, thinking, and vision

• Parallel Stop no longer throws; unpaired tool_calls repaired before send; cache invalidation for shell/MCP; 429/529 retry (#875).
• Thinking-budget continuation carries reasoning forward; budget is per turn (#872).
• Vision probe; Qwen/Harmony tool-call parsing; screenshots fed to vision models (#871).
• LM Studio tok/s trusts server stats when coherent; browser_screenshot capture bounded at 3s.
• Orchestrate boards: per-board reasoning effort and thinking mode.

Orchestrate, search, MCP, LSP

• AFK board chats recover after stall-stop; live tool output counts as progress (#894).
• Orchestrator git-init no longer loops on ask_question (#887).
• Web search relevance guard; safe_search; SearXNG secret key stable across launches; deep_read (#898).
• MCP snake_case dispatch fixed; MCP tools in Settings → Tools with persistable permissions (#881).
• Excel/Office read_file no longer dumps ZIP bytes (#884).
• tsserver restarts when Vite/Node project inputs change (#885).

Brain, preview, CI

• Brain reindex: streamed usage scan, framed worker IPC, 202 + poll, failed-file summary, PageRank write (#855).
• @emnapi lockfile restored for npm ci; issue v2 tools in capability catalog; wiki/budget refreshes.

Fourth open-beta build focused on Orchestrate board execution, model routing and MLX hosting, Brain code index, and memory editing.

Highlights

• Orchestrate and boards — Synthesis config, board chat metrics, board view polish, and task worktree sync for parallel execution.
• Model routing — Optional prompt expander binding in Settings → Model routing.
• Models picker — Streamlined Load / Unload, host-filter resolvers, and stronger provider resolution.
• MLX — HF hub cache created before mlx-lm spawn (#836).
• Brain code index — Workspace-scoped indexing and API improvements.
• Memory — Edit saved entries from the Memory UI.
• Product wiki — Preview hides behind the manual overlay; catalog guarded in CI.
• Chat — Clearer generation errors; improved stop-generation and agent activity handling.

Notable fixes since v0.0.3

• Composer model menus: host filter Load / Unload wired to correct resolvers.
• CI stall recovery; model picker CSS contract.
• Preview visibility when wiki overlay is open; screenshot links with session token support.
• @emnapi optional deps restored in lockfile for packaged Windows builds.

Third open-beta build centered on a workspace-first shell, Orchestrate library surfaces, polish across Models, Brain, and agent safety.

Highlights

• Workspace-first boot — Workspace picker at launch; app rail and Code align to the active folder; optional fresh chat when switching workspaces (#825).
• Slash skills — /plan-work, /orchestrate-plan (#824), and /fix-ci (#796).
• Agent shell sandbox (opt-in) — macOS Seatbelt / Linux Landlock for one-shot agent commands; default off (#810).
• Models — Auto-load before chat, stale-serve reconciliation, MLX context/log/shutdown fixes (#809, #823).
• Per-workspace Issues — Project keys follow the workspace (#783).
• Ctrl + scroll — Zoom the desktop shell (#790).

Notable fixes since v0.0.2

• Boot dual-gate loader and workspace gate polish; boot/streaming performance (#784).
• Enter stops chat instead of sending queued steer; Windows Open in Explorer.
• Terminal run tracking across eviction/restart; stable grep pagination.
• Custom GPU layers on hosting; gh CLI guidance for GitHub in agent chats.

Second open-beta build focused on the Code editor, Brain code map, LSP, and local models on Apple Silicon.

Highlights

• Code map follows your workspace — Brain code index and repo map injection respect the active workspace root.
• Packaged LSP works — Node-based language servers start correctly inside the Electron app.
• Go (gopls) — Installs via go install with correct upstream versioning (Go on PATH).
• Intent mode — Restored styling, idle auto-proposals, and correct accept/reindent behavior (#779).
• MLX on Apple Silicon — Optional mlx-lm runtime and Hugging Face catalog search (#780).
• My Models — Fixed Minnow Library provider binding (#778).

Fixes since v0.0.1

Code editor and Brain

• Code map / Brain code index use active workspaceRoot for repo maps and injection.
• Intent mode display, idle proposals, line-leave resolve, and accept reindent (#779).
• Packaging no longer kills unrelated Electron apps (e.g. Cursor) during npm run package.

Language servers

• Bundled Node LSPs spawn with Electron’s Node (ELECTRON_RUN_AS_NODE) (#768).
• gopls via go install; Settings install/remove for Go bundles.
• Bash LSP: ShellCheck requirement documented; improved workspace config resolution.

Models and providers

• MLX inference and Hugging Face search on Apple Silicon (#780).
• minnow-library provider id for hosted My Models (#778).
• Kimi / Moonshot: omit reasoning fields on tool-loop replay when the API rejects them.

Documentation and CI

• Docs and third-party notices aligned with current product (#775).
• CI stability fixes (grep fixtures, capability probe, LSP completion timeouts).

First public preview of Minnow: a free, AGPL local-first AI workspace for chat, code, orchestration, memory on the models you already run (LM Studio, Ollama, llama.cpp, and or any OpenAI-compatible API).

This is a beta. Expect rough edges. Feedback and bug reports are welcome on GitHub Issues.

Highlights

• Orchestrator boards — Turn a plan into a kanban line with Builder/Tester agents and isolated git worktrees.
• 114 built-in tools (106 in default builds), sub-agents, work agents, MCP, and per-tool permissions.
• Privacy by default — No Minnow account, no telemetry; data under ~/.minnow; loopback-only server until you opt into LAN access.