Notch IDE 0.1.47

Web search in Notch IDE

Open Settings → Appearance → Web research to select DuckDuckGo, Tavily, Serper, or SerpApi. DuckDuckGo is the default and needs no key. The other providers each have their own API key field. Searches go directly to the selected provider.

Existing Tavily keys carry forward. Saved keys use system encryption when available. SerpApi searches that find zero results now report an empty result instead of a failure.

Windows download

Download notch-0.1.47-setup-x64.exe for Windows 10 or 11 on x64. The installer is unsigned. You can verify the download against SHA256SUMS.txt.

notch 0.1.46

Automation reliability and controls

• Fixed Stop during session preparation, schedule/run persistence after an interrupted write, timer recovery after storage errors, and long offline catch-up.
• Reused automation chats now apply each run's cancellation signal and limits.
• Added time/token limits, visible waiting status, next-run previews, timezone and missed-run controls, deduplicated notifications, reusable local presets and paged history.
• Kept the complete automation conversation list while paging its run history, including more than 1,000 saved sessions.
• Preserved original conversations through failed compaction and restart; token limits also cover summarization.
• Added explicit cleanup for unused, owned runtime versions while keeping the current/previous version, active engines, unknown or modified folders and downloaded models. Inactive runtime diagnostic logs are bounded.

Validation: 444 source tests passed, strict TypeScript checks passed, and the core was rebuilt. GPU benchmarks and model-weight downloads were skipped.

Packaged app verification passed all 18 automation scenarios, including preset restoration, early Stop, time/token limits and restart recovery. The installed 0.1.46 startup check used a copy of the saved profile and observed no visible console windows. Final comparisons preserved all 831 profile entries, 641 model file entries and runtime preferences.

Notch 0.1.45

Notch 0.1.45 improves local-model setup and switching, and introduces a diamond-shaped task menu with new icons and slim gray download progress bars.

• Colibri setup, validation, cleanup and queued model loads now share the same inference lock.
• Context changes finish checking memory and saving before a queued model can load.
• Strata folder changes cannot alter an active download after a delayed folder-picker response. Late cancellation restores the previous model selection while retaining downloaded files.
• Installers refresh their runtime dependencies from the tested build. Security updates cover fast-uri, undici and constrained Zod collection parsing; the legacy Zod 3 API is retained with a documented compatibility patch.

Verification uses the real Colibri 2.0.0 and Strata Control 0.1.47 processes with tiny test engines, including launch, eject, remount, switching, failed startup and retry. No large model downloads or GPU benchmarks are required. Full-weight GPU inference across every hardware configuration is outside this release's verification scope.

The update keeps the existing application identity, settings, chats and model locations. Windows x64 installer; unsigned. See the repository's release-readiness document for the audit findings and verification limits.

Final verification: 418 source tests passed, strict typecheck and core build passed, packaged UI checks passed in both themes, and the installed app passed Colibri and Strata lifecycle checks with zero observed console windows. Settings and chats remain unchanged. No model downloads or benchmarks were run.

notch 0.1.44

Fixes the remaining Windows terminal flashes during local model discovery and loading. Managed Ollama now runs in a persistent hidden Windows pseudoconsole, including its GPU-discovery children. Official Ollama binaries, AMD/NVIDIA backend selection and context policies are retained.

Existing Notch-managed runtimes transition under the shared inference lock and restore resident models. The host survives normal app exit and lives outside installed program files so updates can replace the IDE. Startup failures release the coordination lock and remain retryable.

Validation: desktop regression tests and strict typecheck; native failure, nested-process, output-drain and persistence checks; existing Qwen load/eject/remount without generation; packaged and installed startup/loading/failure/retry checks with continuous console observation. Settings, chats and downloaded models retain the same locations and application identity.

Benchmarks and new model downloads were skipped. Live Strata weight loading remains unverified on this PC. This Windows x64 installer is unsigned.

Installer SHA-256: d9a20771614bb4049bef78d1b30218b6e320e8619eccb45992aa49f09d145800.

Notch 0.1.43

Notch 0.1.43 improves local-model recovery and retains the download/mount/eject fixes from 0.1.42.

• Show loading dots above the composer during model discovery, mounting and ejecting; reconcile the catalog before clearing loading, and retain real failure messages. The composer controls stay mounted when status appears or disappears.
• Count shared GPU allocations above physical VRAM as reclaimable system RAM when switching models, including integrated Radeon setups.
• Improve Colibri/Strata text spacing, backend explanations, wrapping, model-row layout and keyboard navigation.
• Cancelled Colibri and Strata setup/download jobs stay cancelled even if their underlying operation finishes at the same moment. A new attempt can complete normally.
• Local CLI operations have a bounded deadline. Discovery uses a one-minute limit; confirmed file downloads have up to 24 hours. Timeout errors retain downloaded files and support retry.
• Regression checks cover failed startup and remount, rejected overlapping changes, failed eject, release-before-load ordering, cancellation with inherited helper output, and timeout/retry.

Validation: 397 automated tests and strict type checking pass. Release checks include the exact packaged modules, renderer loading/error/retry/eject states, installed Ollama model load/eject/remount/switch using two entries that share existing weights, profile/model preservation, Windows console visibility, and public installer/updater checksums.

Settings, chats, downloaded models and application identity are preserved. No large Colibri/Strata models are downloaded for testing, and no speed benchmarks are run. The official Strata catalog recognizes the Radeon 8060S through AMD HIP, but no offered model fits the current free memory on the test PC, so real loaded Strata eject/switch remains unverified. Live NVIDIA/macOS/Linux execution remains unverified.

Installer SHA-256: e8900e7394dfc695ca68eadf7da8ab1d79587b15cc0baeb73a4bed53a5049e5a.

Notch 0.1.42

Notch 0.1.42 is the next stable Windows release of the local-model and console-window fixes validated in 0.1.41.

• Download Ollama/MLX files without mounting them or failing because free RAM changes. Retained downloads are reused; memory and context are checked when loading.
• Keep mount/eject/remount operations coordinated. Reject failed Strata stops and prevent stale refreshes or delayed replies from restoring incorrect Loaded labels.
• Wait for model-selection persistence before sending the next message. Keep named loading feedback inside the selector.
• Detect Colibri initialization failures and fall back to another supported backend.
• Avoid detached Windows consoles for background servers and shell supervisors.

The app, executable metadata, installer and updater manifest use version 0.1.42. Settings, chats and downloaded model locations retain the same application identity.

Validation: 381 automated tests and strict type checking pass, plus packaged artifact and runtime checks. A Windows temporary-folder cleanup retry makes the test harness more reliable.

Live benchmarks remain skipped. Colibri previously passed real load/eject checks on the Radeon 8060S using CPU fallback after Vulkan initialization failed. No Strata model weights are installed on the test PC, so live Strata weight loading and live NVIDIA/macOS/Linux execution remain unverified.

Installer SHA-256: bba87922a2cbb0bf57c9b6b1947d4c60c92d040bb67a804c5e4d10a39c339fc5.

Notch 0.1.41

Notch 0.1.41 fixes local model loading, ejection and downloads.

• Downloading an Ollama or MLX model no longer mounts it or fails because free RAM changes. Existing downloads are reused. Memory and context are checked when loading.
• Mount/eject/remount operations recheck active responses after acquiring the inference lock. Stale refreshes and delayed replies cannot restore incorrect Loaded labels or select an ejected model.
• New messages wait for model-selection persistence. Loading status names the chosen model and clears after failures.
• Colibri detects engines that exit during initialization and falls back promptly. Strata refuses to switch when its control runtime reports an unsuccessful stop.
• Background Windows servers and shell supervisors avoid detached consoles, while background commands retain output and survive closing the IDE.

Validation: 381 automated tests and strict type checking pass. Packaged download UI checks cover light/dark themes, low-memory warnings, progress, cancellation and retries. Real Colibri load/eject checks on Radeon 8060S pass using CPU fallback after Vulkan initialization fails. The existing Nimbus 4B download was recovered from cache without generation.

GPU speed benchmarks were skipped. Strata has no installed weights on the test PC; its hardware catalog and fixture-based switching/stop tests pass, while live weight loading remains unverified. Existing chats, settings and model files are retained.

Installer SHA-256: e2967c1af607c581cbb9694e7f62c98912910547f847b70af5cd7295bba0eca5.

notch 0.1.40

Opening a task now clears its red failure badge while keeping the error in the conversation. Acknowledgement survives restarting Notch, and a new failure can show a new badge.

• Show which local models are loaded in the model picker.
• Click a loaded local model again to eject it, or use the eject icon in the picker or Manage models. Ejecting releases memory and keeps downloaded files and the chat's selected model. Sending another message loads that model again.
• Selecting another local model releases the previous model before loading the next, including switches between Ollama, Colibri, Strata and managed MLX. Duplicate copies on separate local Ollama daemons are released too.
• Keep an active local response intact: finish or interrupt it before a manual model change. Queued inference retains cancellation, and failed loads report an error without claiming the new model is ready.
• Keep existing settings, conversations and downloaded models in place.

Validation: 365 desktop tests and the strict native-harness typecheck pass. Regression fixtures cover badge acknowledgement/restart, new failures, preserved error history, picker and management eject controls, release-before-load ordering across all four local providers, duplicate daemons, active-response protection, failed loads and IPC restrictions. The packaged archive matches the verified source and compiled harness.

Live GPU benchmarks and macOS/Linux inference were not run. Colibri and Strata remain experimental. Includes public CLI 0.7.34 with the updated desktop-native harness.

Windows 10/11 x64 installer; unsigned. Existing stable installations receive the update through GitHub Releases. SHA256SUMS.txt and latest.yml accompany the download.

Installer SHA-256: 8789e7ede85e72d2a4f808b648c6adaad18a2755250e5ece88aca7c46d6c3764.

notch 0.1.39

Local follow-up conversations now work when a model is already loaded and available memory is low. Wikipedia and other retrieved or cited pages appear in the Sources panel, including in previously saved chats.

• Scale built-in prompts and tool descriptions to the selected context: shorter at 8K, with more detail at 16K and 32K+. Preserve custom instructions and explicit context selections.
• Add hardware-aware experimental Colibri and Strata options. Select supported backends and context sizes automatically, including detection of Radeon 8060S configurations. Unsupported hardware stays visible with an explanation. Engine and model downloads remain separate from the installer.
• Automatically open model management when no usable local model is found. Accept a drive root for Strata model storage by creating a dedicated subfolder.
• Suppress background Windows command windows and correct executable branding. Remove remaining focus rings while retaining keyboard navigation.
• Remove duplicate built-in content-moderation prompt instructions while keeping tool permissions and project access checks.
• Preserve existing settings, conversations and downloaded models. A missing app-owned temporary task folder can be recreated when resuming its chat.

Validation: 351 desktop tests, 785 CLI tests and both strict typechecks pass. The actual installed Windows app completes a question and verbose follow-up using a real local Gemma model at 8K; Wikipedia is visible in both new and older Sources panels. Installation and post-update checks confirm existing chats, settings and models are preserved.

Colibri and Strata remain experimental. Live GPU benchmarks, live Strata inference, paid cloud inference and macOS/Linux were not verified. Includes public CLI 0.7.34 and the updated desktop-native harness.

Windows 10/11 x64 installer; unsigned. Existing stable installations receive the update through GitHub Releases. SHA256SUMS.txt and latest.yml accompany the installer.

Installer SHA-256: 95b530de10e8fbcfbb5f4c81e65a4b38e50ae19ed38c3f29dc3b4ae86b33d951.

notch 0.1.35

Windows Origin sign-in, repository browsing and Git operations now run natively. WSL is optional.

• Sign in to Cursor Origin from the repository browser, continue through your browser, then choose a namespace and repository. HTTPS is the default; native SSH remains available. Previously cloned repositories show Open and remain remembered after restarting.
• The CLI and IDE share a Windows user-protected local login. Git receives short-lived Origin credentials through an Origin-only helper, including in terminals opened before sign-in. No FreeSyntax account or hosted backend is involved.
• Improve English navigation and product messages while preserving code, terminal output, compiler diagnostics and structured responses. Keep provider error details out of ordinary error displays.
• Fix background command startup reporting and final-message save failures. Includes Notch CLI 0.7.34.

Validation includes native Windows credential protection/restart/cancellation, the official SDK sign-in URL, Git credential-helper pipes, and real local Git clone/pull/push fixtures over HTTPS/SSH URL handling. Completing live Origin sign-in and Cursor-hosted writes still requires an Origin-enabled account and was not verified during this release.

Windows x64 installer; unsigned. Download the installer and compare its SHA-256 with SHA256SUMS.txt. Existing stable installations receive this update through GitHub Releases.

Installer SHA-256: 1ff454690c03a1e4094ef7dd5ecd1308134c7b05de8ce2035bdab5a671075a9f.