Local Idea Studio v0.2.1 - Agent Mode path fix

Agent Mode workspace-path compatibility fix

Fixed

• Exact Windows paths pasted from Explorer now work when they are inside the selected Agent Mode workspace.
• Absolute in-workspace paths are converted to workspace-relative paths before the permission dialog and execution.
• Smaller local models may use safe aliases such as read, inspect_file, or open_file without producing an unsupported-operation error.
• Agent requests with a harmless extra closing > or a missing display reason are recovered safely; the app generates a clear permission reason.
• Automatic continuation suppresses repeat operations on the same completed target even when the model rewrites equivalent content differently.

Security preserved

• Every operation still requires a native Allow once confirmation.
• Absolute paths outside the selected workspace, UNC/device paths, traversal, linked-folder/file escapes, binary editing, deletion, moves, and shell commands remain blocked.

Verified

• Type checks and all automated reasoning, thinking-mode, Agent Mode path, parser, and continuation tests passed.
• A real Qwen3.5-4B-IQ4_NL GGUF test on Vulkan successfully proposed reading D:\ai test\number_guessing_game.py from the selected workspace.

The Windows installer is unsigned. Download only from this GitHub release or the official Local Idea Studio website and verify its SHA-256.

SHA-256: 31cf3633f208291e1e7d4d717e8e3c1ff9accf2fe16d0cf8868ccdb6473b23cd

Local Idea Studio v0.2.0 - Workspace Code Agent

Workspace Code Agent

Local Idea Studio can now inspect and safely edit code inside one user-selected workspace.

New

• Inspect folders with permission-gated directory listing
• Read bounded sections of text and code files (up to 400 lines per action)
• Understand existing code before proposing an edit
• Apply exact, unique code-block replacements
• Automatically continue the same AI task after every approved inspection or edit
• Create a local backup before replacing existing code

Safety

• A native Allow once confirmation is required before every inspection or edit
• Workspace boundaries, drive paths, traversal, linked-folder escapes, and binary files are blocked
• Ambiguous replacements are rejected instead of changing multiple locations
• File contents are treated as untrusted workspace data
• Shell commands, program execution, deletion, moving files, and outside-workspace access remain unavailable

Verified

• TypeScript checks passed
• Reasoning and thinking-mode tests passed
• Agent parsing, continuation, workspace-boundary, folder inspection, code reading, backup, exact-edit, and ambiguity tests passed
• Windows x64 installer built successfully

The Windows installer is currently unsigned. Windows SmartScreen may display a warning. Download it only from this GitHub release or the official Local Idea Studio website and verify its SHA-256.

SHA-256: da41c3d88f69363afc5e310662d223feb1b254ff19ce1ca6bda81230f897afb7

Local Idea Studio v0.1.9 - Agent Mode continuation fix

Agent Mode continuation fix

• After Allow once, the real file-operation result is saved into the same local conversation.
• The model automatically resumes the original request and confirms the result instead of stopping at the permission card.
• Identical completed, declined, or failed actions are suppressed during automatic continuation, preventing repeat-action loops.
• A later retry still works when the user explicitly asks for it.
• The hidden action-result context stores only the workspace-relative path, not the user's absolute workspace path.

Verified

• Full TypeScript, reasoning, thinking-mode, Agent workspace, continuation, and duplicate-action test suite passed.
• Real Qwen3.5 4B Vulkan two-turn test proposed hello.py, received a simulated successful approval result, continued with a completion response, and did not propose the action again.

The Windows installer is currently unsigned, so Microsoft Defender SmartScreen may display a warning. Download it only from this GitHub release or the official Local Idea Studio website and verify its SHA-256.

SHA-256: 803ed3f88e05e8d82a25d9543420b53b283d06a80b6f2057251bd600d91c8419

Local Idea Studio v0.1.8 - Agent paths and response-token controls

Fixed

• Agent Mode now treats model-generated /hello.txt as the workspace-relative hello.txt instead of failing.
• Windows drive paths, UNC paths, traversal attempts, and linked-file escapes remain blocked.
• Requested response tokens now support an exact user-entered value from 32 to 32768.
• Increasing the response request automatically raises that model's configured context cap when its trained context allows it.
• The loaded runtime still safely caps output to available context and hardware memory; reload the model after changing context.

Verified

• Full automated test suite, TypeScript checks, lint, and production build passed.
• Real Qwen3.5 4B GGUF Agent Mode proposal test passed on Vulkan without executing a file operation.

The Windows installer is currently unsigned, so Microsoft Defender SmartScreen may display a warning. Download it only from this release or the official Local Idea Studio website and verify its SHA-256.

SHA-256: 737bfe7ad7e4512ef4abb31a91453e9ed4b69cc407c4fa20cc9ca21f9fe9683f

Local Idea Studio v0.1.7 - Thinking modes and permission-gated Agent Mode

What changed

• Added Minimal, Standard, and Max thinking modes.
• Enforced a real thought-token budget so reasoning models reserve space for a final answer.
• Added automatic final-answer recovery when a model still reaches its response limit while reasoning.
• Added a permission-gated Agent Mode preview for workspace text-file and folder operations.
• Every proposed file operation shows its path and reason, is revalidated in the main process, and requires a native Allow once confirmation.
• Declining permission makes no change. Absolute paths, path traversal, linked-folder escapes, shell commands, deletion, and paths outside the selected workspace are blocked.

Verified

• Qwen3.5 4B on Vulkan returned visible reasoning and a complete final answer with a 384-token response budget.
• Minimal mode returned a direct final answer.
• A real GGUF Agent Mode test generated a valid hello.py proposal without executing a file action.
• CPU -> Vulkan -> CPU backend switching passed.
• Reasoning parser, settings, portable paths, and Agent workspace validation tests passed.

The Windows installer is currently unsigned, so Microsoft Defender SmartScreen may display a warning. Download it only from this release or the official website and verify its SHA-256.

SHA-256: 81c0d623170540c071685b8ed7c03843229e70f01898aeed9e5303cd05d88d58

Local Idea Studio v0.1.6 - Performance and visible reasoning

0825-Local-Idea-Studio-YouTube.mp4

Performance and visible reasoning

• Auto mode now uses every logical CPU thread available to the app.
• Accelerated backends use a 2048 prompt batch and request full GPU layer offload when memory permits.
• Runtime info reports actual CPU threads, batch size, backend, GPU layers, and offload percentage.
• Thinking-capable GGUF chat templates stream their real model-generated reasoning in a collapsible live panel.
• Standard models show only their answer and never display a fake empty Thinking panel.
• Backend switching tests cover CPU -> Vulkan -> CPU without stale state.

Verified on the development system

• Ryzen 5 8600G: 12/12 logical CPU threads available.
• Radeon 760M Vulkan: all 29/29 Qwen3 1.7B Q2_K model layers offloaded.
• Measured generation: about 46.8 tokens/second with the tested model and settings.
• Qwen reasoning output and TinyLlama non-reasoning behavior both passed.

Performance varies with model, context, quantization, memory bandwidth, drivers, temperatures, and hardware. A constant 100% on both CPU and GPU is not guaranteed or necessarily optimal.

The Windows installer is currently unsigned. Verify its SHA-256 before running it:

235f81dd38efd91600175e2a4b28d7f0a1d3ae40c5f64ac848a6778a44233f45

Local Idea Studio v0.1.5 — Migration-safe rename

Local Idea Studio v0.1.5

This is the completed brand-rename release.

Fixed

• Existing models in Documents\Lumen Studio\models are detected automatically after upgrading.
• Existing legacy D:\LLM AI installations remain supported.
• New installations use Documents\Local Idea Studio\models.
• Custom model folders continue to work.

Included

• Full Local Idea Studio app, installer, shortcut, legal identity, and installer artwork.
• CPU, NVIDIA CUDA, and Vulkan local inference.

Verify before installing

The Windows installer is currently unsigned. Download it only from this release or the official website and compare its SHA-256 with SHA256SUMS.txt.

SHA-256: b72baf72db177fd003be86cc250f5bbaf47eb7c68e17330fac9d2fd48e1d812b

Local Idea Studio v0.1.4 — Brand rename

Local Idea Studio v0.1.4

This release completes the product rename from Lumen Studio to Local Idea Studio.

Changed

• New app, installer, shortcut, package, and legal identity: Local Idea Studio.
• New default model folder: Documents\Local Idea Studio\models.
• Existing installations keep their current selected model folder, so downloaded GGUF files remain visible.
• CPU, NVIDIA CUDA, and Vulkan runtime choices are preserved.
• Installer artwork and Windows metadata use the new brand.

Verify before installing

The Windows installer is currently unsigned. Download it only from this release or the official website and compare its SHA-256 with SHA256SUMS.txt.

SHA-256: d8a18addda51e60cc26a46bee805a7d82d919185fa3fd7c7af21995753ae6cf7

Lumen Studio v0.1.3 - GPU backend switching fix

GPU backend switching fix

This patch fixes GPU inference and fast response streaming when changing backends.

Fixed

• Vulkan and CUDA responses no longer appear frozen or stop printing when chunks arrive quickly.
• Changing CPU, Vulkan, CUDA, or Auto now safely stops any active generation and automatically reloads the same model.
• The loaded-model panel and confirmation message now report the backend actually in use.
• Failed backend switches restore the previous working backend instead of leaving stale UI state.
• Chat completion and error events stay attached to the correct conversation.

Verified

• Automated CPU -> Vulkan -> CPU switching test passed.
• CPU mode used 0 GPU layers.
• Vulkan mode offloaded all 23 layers of the test model and streamed 28 chunks successfully.
• Windows package includes CPU, Vulkan, and CUDA x64 runtimes.

The Windows installer is currently unsigned, so Microsoft Defender SmartScreen may display a warning. Download it only from this GitHub release or the official Lumen Studio website and verify its SHA-256.

SHA-256: dd16bb840c5b11e0bc4caf59a3e4faeb0928cf638588991f28d00e52ba83f4a9

Lumen Studio v0.1.2 - NVIDIA CUDA controls

What changed

• Auto-detect now clearly states that CUDA is preferred on supported NVIDIA RTX/GTX hardware.
• Added an explicit CUDA backend option alongside Vulkan and CPU.
• GPU layer guidance now covers CUDA and Vulkan.
• Runtime labels and documentation accurately describe NVIDIA acceleration.
• Existing portable model storage behavior from v0.1.1 is preserved.

Verification

The Windows installer is currently unsigned. Verify its SHA-256 with the attached SHA256SUMS.txt before running it.

SHA-256: 8f48c3dced35af46e574558dacd07416fa4533da2b3b912eab9249e670319e1b