Changelog
Track the evolution of GIDE. Every feature, improvement, and fix documented.
v2.2.0
July 2026
Local Multi-Model Support
GIDE now runs several on-device models beyond the default: Qwen, Ornith, Gemma, and DeepSeek in GGUF. You pick the family and point GIDE at the model file, and it auto-configures memory, context size, and tool-calling for that model.
Connect to Online Models
Bring your own API key to run Anthropic (Claude), OpenAI, or Google Gemini directly inside GIDE. Keys are stored locally in your system keychain and are only ever sent to the provider you choose — GIDE ships none of its own.
Skill Files
Author reusable skill files under .gide/skills to teach GIDE your project's conventions and procedures. They activate deterministically when you @mention them or when a matching file is open — no fuzzy guessing, so guidance is only applied when it truly fits.
OpenShell Integration
Connect OpenShell to GIDE over MCP and let the chat agent call its tools directly. This extends GIDE beyond its built-in tools to whatever your connected MCP servers expose.
Set Your Temperature
Every model now has a temperature slider in its settings, letting you dial responses from precise to more creative. GIDE shows a recommended value per model and applies the change on your next message.
GPU Offloading Setting
Control how many model layers run on the GPU with a per-model offload slider, recommended based on your VRAM and RAM. Adjusting it reloads the model so the change takes effect right away.
Upgraded Chat UI & Thinking
Refreshed the chat interface with a new toolbar, cleaner inline-diff controls, and a polished reasoning view. The model's thinking now streams and animates correctly for both local and online models.
Extension Compatibility Fix
Fixed an issue that blocked VS Code 1.x extensions from installing on GIDE 2.x. Gallery assets and chat anchors now resolve reliably, so the extension marketplace works as expected.
v2.1.0
June 2026
Subscription Tiers — Individual, Pro & Enterprise
Three capability tiers now power GIDE. Individual covers the Basic (Monthly) and Premium (Yearly) plans. Pro covers the Pro and Small Business (SMB) plans — both surface as "Pro". Enterprise is the top tier. Tiers are ordered so higher tiers automatically include everything below them.
Small Business (SMB) Plans
Small Business, Business, and Team plans are now recognized and mapped to the Pro tier, giving teams the full Pro feature set under a single shared plan.
Tier-Based Feature Gating
Features can now be unlocked by subscription tier. GIDE Wiki is available on Pro and Enterprise; Cloud Support is reserved for Enterprise. Free, trial, and Early-Access users get full access to everything during their access period.
Plan & Tier Visible in the App
Your current subscription plan and tier are now shown directly in the status bar and profile dialog, so you always know which features you have access to.
Usage Portal — Track Your Token Usage
A new usage portal shows your token usage in both the app and on the website, giving you a clear, real-time view of consumption across all your sessions.
Server-Authoritative Subscriptions
Your tier is now determined by your live subscription on the server. Access updates automatically and is correctly revoked when logged out or when a subscription expires.
Inline Command Gate
Shell commands run by the agent now pause for your explicit approval before executing. The exact command is shown inline in the chat so you can review it before allowing it to run.
Collapsible Thinking Block with Live Timer
The agent's reasoning is now shown in a clean, collapsible "Thinking" block with an accurate live timer. Thoughts are finalized before tool calls begin, so reasoning is no longer split or interleaved with actions.
Chat Status & Tool Chips
Dynamic status labels, accurate command chip labels, and removal of duplicate chips make chat activity easier to follow. Internal judge reasoning is hidden from the chat view.
Chat Stability
Fixed a chat crash, stabilized session restore, and hardened transcript serialization so messages with no visible content no longer break the conversation.
Auto-Logout Timing
Session expiry now follows your actual subscription end date instead of a fixed 7-day timer, preventing premature logouts.
Theme Override & Text Colors
Corrected theme override behavior and text color rendering across the chat view for consistent appearance across themes.
v2.0.0
May 2026
Private Cloud Mode
Toggle between local and cloud inference directly from the GIDE status bar. When enabled, model calls route to a remote OpenAI-compatible endpoint (Alibaba Cloud DashScope / qwen-plus) — no manual .env editing required. Includes bearer-token gating, graceful 503 fallback when unconfigured, and bundled-key support for packaged builds.
Write Gate — Confirm Every File Write
Every file write now pauses the agent and shows a confirmation popup: Accept, Reject, or Allow All (session-level bypass). Auto-rejects after 2 minutes of no response. The agent is told when a write is rejected and will not retry it.
Inline Command Gate
Shell commands via run_command now require explicit user approval before execution, mirroring the Write Gate. Shows the exact command and working directory. Running processes can be killed via a dedicated endpoint. Pending gates auto-reject on session disconnect.
Git Tool — Safe, Attribution-Aware Git Operations
Dedicated git tool replacing raw shell-based git access. Supports status, diff, log, add, commit, branch, checkout, init, push, remote management, and GitHub operations (repo create, issue create/list/get). All commits auto-append a Co-Authored-By: Gide trailer. Destructive operations (force push, hard reset, branch delete) are blocked unless explicitly allowed. Shell-based git/gh commands are now blocked — agents must use the git tool.
Agent Intelligence — Deadlock Recovery & Edit Judge Rewrite
New deadlock detector forces a strategy switch after 2 consecutive failed writes. The duplicate structure checker now auto-merges fixes instead of rejecting. Edit judge criteria rewritten: partial progress toward a goal is now accepted instead of blocked. Task planner emits unified replace steps when both new code and modifications target the same file.
Windows & Cross-Platform GPU Detection
GPU detection now works natively on Windows via nvidia-smi and Vulkan fallback. Fixed shell redirect syntax for Windows, replaced Linux-only pipe commands with JS equivalents, and changed GPU fallback from forced-CPU to auto-detect mode. Console window popups suppressed on Windows.
Chat Status & Tool Chips
Thinking blocks rendered as clean markdown blockquotes. Tool chips now show human-readable labels (e.g. "Writing filename.ts", "Searching codebase") with file language icons and LOC deltas (+47 / −12 lines). Judge reasoning hidden from chat UI. Delta values persist across workbench reloads.
GIDE Wiki v2 — Architecture Graph Rewrite
Complete rewrite of codebase visualization. Phase 2a now builds a real import dependency graph using Tree-sitter instead of directory-name heuristics. Enterprise repos clustered via Louvain community detection. Cluster names generated from actual file content. Architectural layers assigned by deterministic path rules. Extraction streams progress to the webview in real time via SSE. New scrollable layout with side-by-side diagram and system cards.
v1.106.0
May 2026
Next-Generation Agentic Core
Rebuilt request handling pipeline with a plan-gate architecture: plan management, gate enforcement, playbook engine, and 4 built-in playbooks covering common code issues such as missing packages, unused imports, incorrect import names, and wrong usage patterns.
Reasoning, Safety & Self-Correction Layer
Introduced convergence detection, edit judgment, error classification, recursive improvement, self-play, skill execution, task intent analysis, intent classification, smart apply, fix validation, post-edit verification, automated testing, sandboxed execution, and dependency guard systems.
Enhanced Task Planning & Model Orchestration
New task planning model and expanded model launcher delivering improved agent coordination and execution reliability.
Deprecated Agent Infrastructure
Retired the previous generation agent infrastructure including legacy route handlers, the old agent core, and the Python-based model runtime.
Secure Local Authentication
File-permission-protected local credential store with auth enforcement, protocol version validation, and a guard middleware chain. Includes a protocol handshake layer for verified client-server communication.
Telemetry & Event Logging
Structured telemetry event log with automatic log rotation for improved diagnostics and usage insights.
Web-to-Desktop Session Bridge Fix
Resolved an issue where the trusted-origin login flow was blocked by auth enforcement. Web companion sign-in now correctly establishes the desktop session without requiring a bearer token.
Extended Context Memory
Persistent semantic memory with recall and storage, file-level summary caching, and conversation archiving that extends AI context well beyond the in-session window limit.
Multi-Language Refactor Engine
Unified refactoring system supporting 21 languages through 9 core operations: insert, change value, inline, modify import, wrap call, add parameter, rename symbol, move to file, and extract function. Includes a dedicated scope analysis module.
Expanded Language Parsing Support (+12 Languages)
Added native parsing support for Bash, C, Dart, Dockerfile, HCL, HTML, Kotlin, PHP, R, Ruby, Swift, and YAML.
Expanded Agent Tool Library (30+)
New tools spanning user interaction, configuration, diagnostics, file operations, LSP integration, task management, context measurement, notebook editing, code search, web access, and structured output formatting. Expanded file-write capabilities.
Intelligent Code Indexer
Redesigned code indexing with semantic retrieval, structural fingerprinting, and symbol-level indexing for faster, more accurate codebase understanding.
Deprecated Retrieval Modules
Retired legacy experimental retrieval modules superseded by the new indexing architecture.
Upgraded AI Instruction Model
Introduced a new, significantly improved system prompt delivering more precise, context-aware AI responses across all coding tasks.
Deprecated AI Prompt Templates
Retired 12 legacy prompt templates now fully superseded by the unified instruction model.
Refactor Validation & Benchmarking Suite
Comprehensive test coverage for all 9 refactor operations (85 tests across 10 test files) and a 48-task benchmark harness with fixtures for JavaScript, TypeScript, Python, Java, Go, and Rust.
Editor Core Enhancements
15 new editor-layer modules covering agentic workflow patterns, context gathering, edit validation, memory, permissions, prompt construction, caching, and retry logic. Includes Gide service contracts, browser-side implementations, and an opt-in bug report bundler with automatic PII redaction.
Editor Stability & Compatibility Fixes
Resolved issues across AI chat, mention completions, error quick-fix, document handling, chat view, contribution registration, Electron runtime, and notebook serialization.
v1.105.4
March 2026
Linux platform support
GIDE is now available on Linux, extending offline AI-assisted development to the third major desktop platform.
Linux packaging
Distributed as AppImage, .deb, and .rpm for broad distro compatibility.
WSL2 compatibility
Validated build and runtime behavior on Ubuntu 24.04 under WSL2 environments.
v1.105.3
Early March 2026
Ornith 9B default model
Ornith 9B ships as the default bundled model across all platforms, delivering high-quality code generation out of the box.
Additional downloadable models
Added support for downloading Qwen, Gemma, and DeepSeek — plus any GGUF-format model — alongside the bundled Ornith 9B.
Model selection flow
Users can now switch between the bundled Ornith 9B and additional downloaded models directly from setup.
Multiple UI bug fixes
Resolved a batch of visual and interaction issues across the editor including icon rendering, context picker failures, and branding inconsistencies inherited from the VS Code fork.
v1.105.2
Late February 2026
VRAM detection bug
Fixed a critical issue on Windows where the system was reading total system RAM instead of available GPU VRAM to determine context window capacity.
Context memory sizing
Model context allocation is now correctly scoped to actual VRAM headroom, preventing OOM errors and incorrect memory budgeting.
GPU enumeration
Improved GPU detection logic to correctly identify and query dedicated graphics memory across different hardware configurations.
Memory headroom buffer
Added a conservative VRAM buffer to avoid instability on systems with shared or constrained GPU memory.
v1.105.1
Late February 2026
Backend crash on launch
Resolved a widespread failure where the AI backend failed to initialize correctly across a significant portion of user systems.
Startup reliability
Improved backend process lifecycle management to prevent silent failures on first run.
Error surfacing
Backend errors that previously failed silently are now surfaced to the user with actionable messaging.
v1.105.0
February 2026
Windows & macOS
First public release of GIDE shipped simultaneously for both platforms.
On-device AI inference
Fully offline, no cloud dependency. All model execution runs locally on user hardware.
VS Code foundation
Built on a fork of code-oss, preserving familiar editor UX with GIDE-native AI integrations.
Qwen model support
Bundled with Qwen model backend for local code generation and completion.