OpenMindAI v3.0.0 Release Notes

Release line: v3.0.0
Status: Release candidate until the signed production installer and updater artifacts pass the final release pipeline.

OpenMindAI v3.0.0 is the largest workflow-focused update since v2.0.0.
The release moves OpenMindAI beyond local chat and model management into a local AI workstation that can actively work with real projects, files, terminal commands, Git repositories, and connected services while keeping explicit permission and validation boundaries.

Between v2.0.0 and the v3.0.0 source line, the repository contains hundreds of changes across the desktop UI, Rust backend, local AI runtimes, project system, connected apps, media capabilities, CI, security, installer, updater, and release infrastructure.

Highlights

Project Work is now a first-class workspace

Projects are no longer limited to instructions, attached files, and linked chats.

Each Project now includes:

Overview tab for project instructions, linked chats, and reference/context files.
Work tab for active local development work.
Attached local filesystem browsing.
Local file editor.
Local terminal controls.
Direct handoff to a linked Project Agent conversation.
Explicit Full PC + Terminal permission boundary.
Live project/workspace context refresh after changes.

This turns a Project into a practical local workspace instead of a passive context container.

Open Folder as Project

A real local folder can now be opened directly as an OpenMindAI Project.

The flow can:

Select an existing folder.
Create the Project automatically.
Attach the folder as the Project workspace.
Create a linked Project Agent chat.
Link the Project and conversation durably.
Open the Project directly in the Work area.
Roll back incomplete Project/chat state if setup fails before the relationship is committed.
Keep filesystem access scoped to the attached folder unless broader terminal access is explicitly approved.

This creates the foundation for working on existing applications, repositories, research code, websites, scripts, or other local projects directly inside OpenMindAI.

Local Project Agent

v3.0.0 introduces the new autonomous Project Agent workflow.

The agent can work against an attached local project and can:

Inspect files and directories.
Read project source code.
Create files.
Edit existing files.
Rename files.
Delete files when required.
Use recent Project conversation context.
Refresh its workspace view after modifications.
Reason over tool failures.
Repair failed changes.
Re-run validation.
Report the final result and validation evidence.

The agent uses the local OpenMindAI model/runtime rather than requiring a paid cloud coding API.

Agent Edit → Terminal → Detect → Fix → Retest Loop

The Project Agent now supports a real iterative development loop:

Inspect the workspace.
Plan the required change.
Edit project files.
Run an applicable command.
Detect non-zero exits or command timeouts.
Feed the failure back into the next reasoning step.
Repair the implementation.
Re-run tests, lint, build, or another applicable validation.
Complete only after valid evidence is available.

Important reliability improvements include:

Non-zero terminal exits are treated as real failures.
Command timeouts are treated as failures.
Successful recovery resets the consecutive-failure counter.
Duplicate exact-action protection prevents infinite repetition.
Agent execution has a bounded safety budget.
Transient local-model failures can be retried.
Running agent work can be cancelled.
Validation cannot be silently bypassed for projects with recognizable build/test manifests.

Stronger Validation Rules

v3.0.0 makes completion evidence stricter.

After a workspace-changing action:

Previous validation is considered stale.
The workspace must be validated again.
Mutating terminal commands also invalidate previous validation.
Successful validation commands are tracked as completion evidence.
Chained terminal commands cannot hide an earlier failure.
A Project with a recognized build/test manifest cannot simply claim that validation is unnecessary.
If no meaningful automated validation exists, the agent may record a concrete reason for skipping it.

Recognized project validation targets include common ecosystems such as:

Node.js / JavaScript / TypeScript
Rust
Python
Go
Maven / Java
Gradle
PHP / Composer
Ruby
.NET projects and solutions

Live Workspace Context

The local agent no longer works from a stale snapshot after making changes.

v3.0.0 refreshes Project workspace context after:

Direct filesystem changes.
Successful terminal actions.
Other operations that may modify the attached project.

This makes the next agent step reason over the current state of the project rather than the state that existed before the edit or command.

Terminal Safety and Permission Boundary

Terminal access remains intentionally permission-gated.

Without explicit broader permission, OpenMindAI keeps Project Agent work scoped to the attached workspace.

When Full PC + Terminal access is enabled, the agent can use non-interactive local shell commands while still following safety restrictions.

The terminal workflow avoids:

Interactive editors.
Password prompts.
sudo requests.
Commands that indefinitely wait for user input.
Interactive pagers.
Unsafe root-level destructive commands.

Cross-platform terminal guidance is implemented for:

Windows: non-interactive PowerShell.
Linux: /bin/sh -lc.
macOS: /bin/sh -lc.

Git-Aware Project Work

Project Agent can inspect Git state when terminal permission is available.

New Git-aware capabilities include:

Structured Git status inspection.
Structured Git diff inspection.
Safer Git execution configuration.
External diff helpers disabled during inspection.
Textconv disabled during inspection.
fsmonitor disabled during inspection.
Recursive submodule behavior restricted during inspection.

Potentially mutating Git operations such as commits, resets, cleans, branch changes, installs, or arbitrary commands remain behind the terminal permission boundary and the user's task instruction.

Connected Apps

v3.0.0 adds a major Connected Apps architecture.

Connected services are treated as internal assistant tools rather than a separate raw API console.

Users configure connections in:

Settings → Apps

Then they can ask for the desired result naturally from Chat or Project Work.

Supported app families include:

Google Workspace

Gmail
Google Drive
Google Calendar
Google Contacts

GitHub

Repositories
Repository files
Issues
Pull requests
GitHub Actions
Releases

Microsoft 365

Outlook
OneDrive
Calendar
Contacts

Additional integrations

Slack
Notion
Dropbox
MCP servers

Connected-app actions are selected internally by the assistant instead of exposing provider/action JSON to the user.

Connected-App Safety

The new connected-service backend includes additional controls for remote data and mutations.

Implemented protections include:

OAuth/client configuration isolated to Settings.
Tokens and connection secrets kept out of chat history.
OS credential-store integration for sensitive credentials.
Explicit backend approval for mutating remote actions.
Provider-level permission checks.
Connected tool output treated as untrusted external data.
Input guards for provider parameters.
Path and repository validation.
Fail-closed mutation classification.
Narrow explicit confirmation phrases before remote writes.
Pending approval state with bounded lifetime.
Duplicate approval protection.

This allows natural-language connected-service workflows without treating external tool output as trusted instructions.

Chat and Conversation Improvements

The chat backend received substantial persistence and workflow hardening.

Improvements include:

Better conversation/model association.
Stronger message mutation handling.
Branch cleanup when earlier user turns are removed.
Improved streaming-state handling.
Better interrupted-message recovery after reopening the database.
Project/profile context separation.
Active model persistence.
Better multimodal attachment handling.
Improved internal tool routing support.
Project context synchronization before connected-agent operations.
Improved regeneration handling for connected-app conversations.

Project Data Integrity Improvements

The Project system received database and relationship hardening.

Changes include:

Stronger Project CRUD validation.
Safer Project/conversation linking.
Single-project ownership enforcement for a conversation.
Migration support for the updated relationship model.
Bounded Project context.
Unicode-safe context limits.
Better tracking of local project resources.
Workspace metadata synchronized into agent context.

Local Media Runtime Expansion

v3.0.0 significantly expands local multimodal infrastructure.

New or expanded backend areas include:

Local diffusion runtime.
Image-generation runtime preparation.
Hardware-aware media preflight checks.
Voice runtime.
Speech/audio runtime.
Video-related local runtime validation.
Vision batch processing.
PDF/OCR support.
Multimodal input handling.
Model package validation for multi-file model packages.

The goal remains local-first: OpenMindAI reports missing local dependencies instead of silently sending generation requests to paid cloud APIs.

Image Generation Improvements

The image-generation path is more complete than the v2 entry-point implementation.

v3 improvements include:

Dedicated diffusion runtime management.
Backend selection based on available hardware.
Runtime integrity metadata.
Model/runtime preflight validation.
Output validation.
Safer generated-file handling.
Memory-aware defaults.
Local preview/artifact integration.

Hardware profiles include compatibility handling for CPU and supported GPU backends.

Voice, Speech, and Audio

New local audio infrastructure includes:

Audio data validation.
WAV decoding.
Audio resampling.
PCM16 WAV generation.
Voice output validation.
Local soundscape/audio generation support.
Better generated media artifact handling.

Vision and PDF Workflows

v3 introduces additional local document and vision foundations:

Vision batch processing.
Image payload validation.
PDF-related processing support.
OCR integration infrastructure.
Improved multimodal chat attachments.
Safer ephemeral image handling.
Better separation between persisted chat text and attached binary/media data.

Document Generation

The document-generation backend was substantially reworked.

Improvements include:

Better structured block parsing.
Real DOCX generation improvements.
Real multi-section PDF generation.
Long-text wrapping without silently dropping content.
Safer generated artifact destination handling.
Collision-free generated filenames.

Artifact and Generated-File Hardening

Generated content handling now includes additional integrity rules.

Changes include:

Expanded media artifact kinds.
Safer artifact cleanup.
Generated-path boundary validation.
Cleanup protection against paths outside the generated-data root.
Better artifact status/error persistence.
Conversation deletion cleanup for generated files.
Migration support for updated artifact behavior.

Model System Improvements

The model subsystem continues to evolve beyond the v2 implementation.

v3 adds or improves:

Updated bundled model catalog.
Multi-file model package support.
Package integrity metadata.
Safer model package reuse.
GGUF path validation.
Improved interrupted-download handling.
Resume behavior for partial downloads.
Size/checksum validation.
Better detection of stale or tampered files.
Hardware compatibility checks.
Prevention of false model-update reporting.
Runtime/model integrity validation.

The default v3 setup continues to target Qwen3 4B Q4_K_M as the baseline local language model.

Runtime and Hardware Improvements

The local runtime layer received additional reliability work.

Updates include:

Improved runtime discovery.
Safer runtime path validation.
Architecture-aware runtime catalog logic.
Windows, Linux, and macOS runtime patterns.
CUDA preference for supported NVIDIA systems.
Vulkan fallback where appropriate.
CPU fallback for compatibility.
Dedicated VRAM-aware launch planning.
Better runtime manifest validation.
Runtime digest enforcement.
Safer reuse of previously installed runtime files.

Portable Storage and Data Preservation

OpenMindAI continues to keep application binaries separate from the user's AI data root.

v3 strengthens this design.

The configured OpenMindAI root can continue to hold:

Models
AI runtimes
SQLite history
Projects
Generated files
Workspaces
Knowledge data
Cache
Logs
Backups

Upgrade and reinstall flows are designed so that application replacement does not delete the user's selected AI data root.

Portable-root behavior also continues to avoid silently creating a new empty data location when an expected external drive is unavailable.

Database Reliability

SQLite persistence received additional migration and recovery work.

New database changes include:

Artifact cleanup/media-kind migration.
Single-project-per-conversation migration.
Migration tracking hardening.
WAL/foreign-key validation.
Backup validation.
Interrupted chat recovery.
Better cleanup of generated artifacts.
Improved open/reopen reliability for existing installations.

Settings → Apps

A dedicated Apps settings area is introduced for connected services.

The UX is intentionally separated:

Settings → Apps: connection state and provider configuration.
Chat / Project Work: natural-language interaction.

Users do not need to manually choose raw provider action names for normal connected workflows.

UI and Workspace Changes

The desktop UI receives substantial structural work.

Major additions include:

Project Overview / Work tabs.
New local Project workspace UI.
Attached folder/file tree.
Local editor.
Terminal permission controls.
Direct Project Agent launch.
Work workspace components.
Connected Apps settings.
Updated chat mode handling.
Updated composer behavior.
Improved artifact cards.
Improved message actions.
Improved chat search integration.
Multimodal UI styling.
Project workspace responsive styling.
Work-specific responsive layout.

Security Improvements

v3 adds a dedicated security validation pipeline.

Security checks include:

npm audit
cargo audit
GitHub CodeQL
Scheduled security scanning
Strict dependency validation
Tauri Content Security Policy
Local secret-store handling
Connected-provider input validation
Remote mutation approval gates
Path traversal protections
Safer Git inspection
Source-control exclusions for local databases, models, runtimes, generated data, logs, caches, and credentials

New repository security documentation includes:

SECURITY.md
CODE_OF_CONDUCT.md
CONTRIBUTING.md
GitHub issue templates
Pull request template
CODEOWNERS
Dependabot configuration

CI Improvements

v3 introduces a much stricter multi-platform CI system.

Frontend validation

Clean npm ci
Version consistency checks
Release/updater configuration checks
ESLint
TypeScript build
Vite production build

Rust validation

Runs on:

Ubuntu
Windows
macOS

Checks include:

rustfmt
strict Clippy with warnings denied
full Rust tests with all features enabled

Desktop build validation

Windows Tauri application build
Desktop executable verification

The release line is therefore validated across more than a single development environment.

Version Consistency Enforcement

v3.0.0 adds automated version synchronization checks across:

package.json
package lockfile
Rust Cargo.toml
Rust lockfile metadata
Tauri configuration
OpenMindAI marker/bootstrap metadata

Version drift now fails CI instead of being discovered during release packaging.

Release Readiness Pipeline

A dedicated release-readiness workflow now validates the release before the production tag workflow is used.

Checks include:

Version synchronization.
Updater configuration.
Release contract validation.
Bootstrap script syntax.
Windows NSIS packaging.
Clean Windows install smoke test.
Real v2.0.0 → v3.0.0 upgrade smoke test.
v3 clean-install verification.
Uninstall validation.

This is designed to catch installer/upgrade regressions before a public v3 release.

Signed Windows Release Pipeline

v3 adds a production Windows release pipeline built around both Windows code signing and Tauri updater signing.

The production tag workflow requires:

Windows Authenticode signing certificate.
Windows certificate password.
Tauri updater private signing key.
Tauri updater key password.

The pipeline is intentionally fail-closed.

If signing credentials are unavailable, OpenMindAI does not silently produce an unsigned build and present it as a trusted production release.

Updater Infrastructure

The v3 release path now prepares the assets required for the Tauri updater.

Production release validation checks:

Signed Windows installer.
Installer signature.
Updater signature.
latest.json.
Version consistency.
Updater URL consistency.
HTTPS release URL.
Expected installer filename.
SHA-256 checksums.

This provides the foundation for a trusted in-app v2 → v3 update path.

Release Asset Contract

The trusted Windows v3.0.0 release is expected to contain:

OpenMindAI_3.0.0_x64-setup.exe
OpenMindAI_3.0.0_x64-setup.exe.sig
latest.json
SHA256SUMS.txt
OpenMindAI-Setup.bat
OpenMindAI-Setup.command
openmindai-setup.sh
openmindai.marker

The final public release should only be published after the production pipeline has generated and verified these assets.

Bootstrap Improvements

The cross-platform bootstrap files continue to be maintained for:

Windows
Linux
macOS

v3 synchronizes bootstrap version metadata with the application release and validates bootstrap configuration as part of the release contract.

Repository Hygiene

The repository is more strictly source-focused in v3.

The source repository excludes machine-local data such as:

GGUF/SafeTensors model weights
AI runtimes
Local databases
Cache
Logs
Generated content
Project workspaces
Knowledge data
Backups
Signing certificates
Private keys
Local secret data

Old committed installer binaries are also removed from the normal source tree; release binaries belong in GitHub Releases.

Developer Experience

v3 improves development and contribution infrastructure with:

Node.js 24 CI baseline.
Stable Rust validation.
Cross-platform Tauri checks.
Dedicated release scripts.
Version consistency script.
Release-readiness script.
Dependabot.
Issue templates.
Pull request template.
CODEOWNERS.
Security policy.
Contribution guide.

Important Behavioral Changes from v2.0.0

Projects

v2.0.0:
Projects primarily organized chats, instructions, and attached local context.

v3.0.0:
Projects can become active local development workspaces with an attached folder, local filesystem, editor, terminal permission controls, Git inspection, and Project Agent execution.

Connected services

v2.0.0:
No unified internal connected-app agent architecture.

v3.0.0:
Google Workspace, GitHub, Microsoft 365, Slack, Notion, Dropbox, and MCP capabilities can be exposed internally to Chat/Project Work through permission-aware assistant tools.

Validation

v2.0.0:
No full autonomous edit → terminal → repair → retest project workflow.

v3.0.0:
Workspace mutations invalidate stale validation and the Project Agent is expected to test/build/lint before completing when applicable.

Release process

v2.0.0:
Public installer/bootstrap assets existed, but the release process was significantly simpler.

v3.0.0:
Dedicated CI, Security, Release Readiness, signed installer, signed updater metadata, v2→v3 smoke upgrade testing, and release-asset contract validation are implemented.

Upgrade from v2.0.0

The v3 application is designed to preserve the user's selected OpenMindAI data root.

The upgrade path must not intentionally delete:

Downloaded models.
Installed AI runtimes.
Chat history.
Project records.
Attached Project metadata.
Generated content.
User settings.
Backups.
Other files stored in the configured OpenMindAI root.

The production release pipeline includes a Windows v2.0.0 → v3.0.0 installer upgrade smoke test before the release is considered ready.

Installation

For a normal public v3 release, use the signed installer generated by the official GitHub Release pipeline.

Existing v2.0.0 users should be able to upgrade using the v3 installer while retaining their configured AI data root.

A clean installation remains supported as well.

Notes for Existing v2 Users

After upgrading, the biggest visible workflow change is inside Projects.

To use the new Project Agent workflow:

Open Projects.
Create a Project or use Open Folder as Project.
Attach/select the local project folder.
Open the Project's Work tab.
Start or open the linked Project Agent chat.
Enable Full PC + Terminal only when terminal/Git execution is required.
Ask the agent to inspect, modify, test, repair, or validate the project using natural language.

Connected services can be configured separately under Settings → Apps.

Compatibility

Primary target:

Windows x64

Source/build support and validation also cover:

Linux
macOS

Actual local inference and media performance depends on:

CPU
GPU
VRAM
System RAM
Driver support
Model size
Quantization
Runtime/backend availability

Known Release Requirement

v3.0.0 production publication requires the trusted signing pipeline to pass.

Required production secrets include:

Tauri updater signing private key
Tauri updater signing key password
Windows code-signing certificate
Windows certificate password

These private credentials must never be committed to the repository.

Summary

OpenMindAI v3.0.0 changes the product from a local AI chat application with Projects into a broader local AI work environment.

The defining v3 additions are:

Open Folder as Project
Project Overview / Work tabs
Autonomous local Project Agent
File read/write/rename/delete workflow
Permission-gated local terminal
Error detection and recovery loop
Required post-change validation
Live workspace context refresh
Git-aware project inspection
Connected Apps architecture
Google Workspace integration
GitHub integration
Microsoft 365 integration
Slack / Notion / Dropbox / MCP foundations
Local secret-store and mutation approval protections
Expanded local image/media runtime infrastructure
Voice and speech runtime work
Vision/PDF processing foundations
Stronger model/runtime integrity handling
Database and artifact hardening
Cross-platform Rust CI
Dependency auditing and CodeQL
Version consistency enforcement
Windows NSIS clean-install validation
Real v2.0.0 → v3.0.0 upgrade smoke testing
Authenticode signing workflow
Tauri updater signing workflow
Verified release asset contract
Improved repository governance and contribution tooling

Release Status

The v3.0.0 source line is prepared as a release candidate.

The final public release should be published only after the production signing credentials are configured and the complete signed Windows installer/updater pipeline passes.

OpenMindAI v3.0.0 — Local AI that can work with your project, not just talk about it.

v3.0.0