v6.7.0

Added

• Real model context and output limit discovery across providers, including /api/show metadata inspection and output maximum learning on Ollama Cloud and local Ollama.
• Minimal-thinking configuration per provider so reasoning models reserve maximum token budget for the actual answer.

Changed

• Updated Together AI default model to openai/gpt-oss-120b and Perplexity default model to perplexity/sonar (Agent API), with automatic migration for saved legacy defaults.
• Filtered non-text models from Ollama Cloud discovery and removed retired models from bundled provider catalogs.
• Replaced ESLint with Oxlint and streamlined CI/pretest build steps to eliminate duplicate typechecks and packaging passes.

Fixed

• Fixed thinking models returning empty or truncated responses by using real reported or documented model limits and omitting adapter fallback caps when the API applies the model maximum.
• Classified prompt-too-long rejections as input-limit errors so oversized diffs automatically split and retry.
• Removed unsupported sampling parameters on Claude 4.7+ requests.

v6.5.0

[v6.5.0] - 2026-10-13

Added

• Added Inception Labs AI provider with model discovery, API key validation, and UI integration.
• Added custom API form with quick presets and real-time connection diagnostics.

Changed

• Redesigned settings screens to use responsive grid rows and SVG icons.
• Replaced unicode emojis with CSS spinners and SVG icons across the UI.

Fixed

• Fixed Inception model loading and updated the encryption report layout.

[v6.4.1] - 2026-10-07

Added

• Added Ollama Cloud provider to the provider list and documentation.
• Added model input limit discovery to adjust budgets for large diffs.
• Added structured data summarization for large JSON, CSV, TSV, and lockfiles.
• Added user-approved large-diff batching with modal confirmation before calls.

Changed

• Changed large diff processing to run approved requests sequentially and stop on rate limits.

v6.4.0 - 2026-10-05

Highlights

• Pro Large Diff Processing: Intelligent budgeted generation and ordered batching for oversized diffs, analyzing every source span without capping chunks or dropping diff headers.
• Media File Handling: Automatic detection for audio, video, and oversized source files (>1 MiB), representing them by filename in change context to avoid prompt token exhaustion while preserving change tracking.
• Model Metadata Discovery: Dynamic context discovery and model capability introspection for OpenRouter, Together AI, and NVIDIA NIM providers, with bounded trials for unknown models.
• Hardened Reliability & Security: Refined HTTP 429 quota classification vs. payment limits, isolated cancellation and timeouts per generation attempt, and applied a verified depth limit patch for braces (GHSA-vfj7-8cjw-p6xm).

Added

• Pro large-diff settings (gitmind.pro.largeDiffHandling.enabled, .chunkSize, .batchSize, .concurrency, and backward-compatible .maxChunks).
• Audio and video media detection to exclude multimedia content from AI prompts while preserving git status in change context.
• Model metadata discovery and context limit lookup for OpenRouter, Together AI, and NVIDIA NIM providers.
• Bounded direct trial execution for newly configured or unknown provider models.

Changed

• Intermediate large-diff analysis and compression prompts are capped at 120 words with bounded token limits, leaving final generation prompts uncapped.
• Token diagnostics are converted into non-blocking informational notices based on approximate prompt token counts.
• Applied verified braces depth limit patch for GHSA-vfj7-8cjw-p6xm.

Fixed

• Large diff coverage no longer drops chunks past legacy ceilings; diffs are partitioned into ordered batches that run with managed concurrency.
• HTTP 429 error messages now distinguish quota limits (rate limits) from billing or payment failures.
• Provider errors now preserve structured response bodies and HTTP status codes for rejected requests on OpenRouter and Together AI.
• Cancellation and timeout signals now isolate per-attempt resources so aborted operations do not pollute subsequent requests.

v6.3.0 - 2026-09-27

Highlights

• Comprehensive AI provider API modernization across all 23 providers, keeping all models, endpoints, protocols, and discovery aligned with late 2026 specifications.
• Restored Groq dynamic model discovery via the /models endpoint with text/chat filtering and modern default openai/gpt-oss-120b.
• Hardened key validation, error reporting, and model fallbacks for NVIDIA NIM, MiniMax, Z.ai, DeepSeek, OpenAI, and Cohere.

Changed

• Updated default Groq model to openai/gpt-oss-120b with max 1,000 token limit to match Groq free-tier rate boundaries.
• Replaced decommissioned meta/llama-3.3-70b-instruct in NVIDIA NIM validation and default configs with nvidia/nemotron-3-super-120b-a12b.
• Modernized MiniMax fallback and key validation to MiniMax-M3.
• Modernized Z.ai (GLM) fallback catalog and key validation probe to glm-5.3.
• OpenAI model discovery now sorts GPT-6 models (gpt-6-sol, gpt-6) at top priority, and default generation fallback is set to gpt-6-sol.

Fixed

• Cohere key validation now queries the v2 Chat API on https://api.cohere.com/v2/chat, eliminating deprecated calls to api.cohere.ai/v1/tokenize.
• DeepSeek key validation probes canonical deepseek-flash instead of deepseek-v4-flash.
• OpenAI-compatible endpoints (including LM Studio and Cloudflare Workers AI) and OpenRouter now parse structured JSON error payloads, displaying clear error messages instead of raw JSON strings.

v6.2.1

v6.2.1 - 2026-09-22

Highlights

• A clearer Marketplace and Open VSX listing: a new name, description, square icon, and README with a real demo, install steps for VS Code and editors that use Open VSX, and a Free vs Pro comparison that labels every Pro feature.
• The published package no longer includes internal development files.

Changed

• The extension is now listed as GitMind: AI Commit Message Generator, in the AI and SCM Providers categories.
• Documentation links point to the GitMind wiki.

Fixed

• Internal build artifacts (graphify-out/, plans/, wiki/) are excluded from the package, and the package audit rejects them.

v6.2.0 - 2026-09-22

Fixed

• GitHub Copilot no longer reports a missing or invalid API key. GitMind detects Copilot through the editor's language model API. When Copilot cannot be used, GitMind says why: not detected in this editor, not installed, access not allowed, or no models available. Each message comes with an action (Find GitHub Copilot, Retry, or Choose Another Provider).
• Copilot errors are identified by LanguageModelError.code, so permission, content-filter, and model-not-found failures show their own messages.
• Copilot requests no longer switch silently to another model after a failure. Model fallback follows the Pro Automatic Recovery setting, as for other providers.
• Provider errors are formatted once. Credential guidance matches how the provider signs in, so keyless providers (Copilot, Ollama, LM Studio, Bedrock, Vertex AI) never show API-key advice.
• Azure OpenAI and Cloudflare keys are now included in secure-storage migration and encryption status.
• Hugging Face accepts organization/model IDs from every organization (such as openai/... and mistralai/...).

Changed

• Perplexity uses the Agent API (/v1/agent) ahead of the Sonar Chat Completions shutdown on September 27, 2026.

• Hugging Face uses the Inference Providers router (router.huggingface.co/v1) for chat and model discovery.

• Copilot Test Connection sends one short request to the selected model instead of trying every model. Requests include a consent justification. The settings view is told when Copilot models change.

• Defaults now use each provider's latest recommended model:

  • Gemini gemini-3.8-flash
  • OpenAI gpt-6-sol
  • MiniMax MiniMax-M3
  • Z.ai glm-5.3

• Known-model tables were updated for Gemini 3.x, Claude Opus 5.5 / Fable 5.1, GPT-6, MiniMax M3, Grok 4.7, and GLM 5.3. Retired Gemini 2.x names map to current models.

• Claude Opus 5.5 and Fable 5.1 get room for their always-on thinking. NVIDIA Nemotron requests turn thinking off, and reasoning traces are removed from commit messages.

• Anthropic key validation no longer calls a retired model.

• Updated retired default models:

  • Mistral mistral-small-2603
  • Hugging Face meta-llama/Llama-3.3-70B-Instruct
  • OpenRouter openrouter/free
  • DeepSeek deepseek-flash
  • xAI grok-4.7
  • Groq llama-3.3-70b-versatile
  • NVIDIA nvidia/nemotron-3-super-120b-a12b

  A saved model equal to a retired default moves to the new default. Other saved choices are kept.

• Refreshed fallback model lists for all providers from current provider documentation.

• Provider names, sign-in types, default models, and model lists now come from a single provider catalog. Onboarding offers all 23 providers.

v6.1.2

• Add LM Studio, Azure OpenAI, Amazon Bedrock, Vertex AI, and Cloudflare Workers AI/Gateway
• Update Settings UI with new provider markers and robust icon fallback coverage
• Centralize provider model choices and remove duplicate stale manifest model lists

v6.1.1

Added

• Added unit tests to verify version parsing, header formatting, and repairs for the ChangelogService.

Changed

• Migrated dependency overrides from package.json to pnpm-workspace.yaml to ensure consistent resolution.
• Replaced npm with pnpm for package management across all workflows.
• Upgraded to @google/genai SDK and updated Gemini service implementation.
• Updated Node.js runtime to v24 and refreshed development dependencies.
• Updated ChangelogService to align with Keep a Changelog 1.1.0 specification.
• Pin @types/vscode to an exact version to ensure build consistency.
• Updated verification metadata and counts for v6.1.1 in handbook pages and manifests.

Removed

• Removed temporary update script.
• Removed package-lock=false from .npmrc.

v6.1.0

v6.1.0 - 2026-07-15

New Features (Pro)

• Added dedicated Pro workspace tab to group health and workflow tools.
• Implemented local commit hygiene scoring system for staged changes, evaluating scope, size, safety, and test coverage.
• Added gitmind.openHealthReport command and dedicated settings card for Pro users to manage health checks.
• Added device management, deactivation, and "+2 devices" add-on purchase flow.

Enhanced

• Implemented local history health analysis to provide insights without requiring AI providers.
• Decoupled health metrics from AI-generated commit messages to ensure objective, deterministic analysis.
• Centralized health scan recording into a reusable function within the health module.
• Replaced legacy debug logging with a structured diagnostic event system using operation correlation IDs.
• Updated workspace header and health report UI for improved readability.
• Improved license validation cadence from 24h to 6h.
• Updated configuration enum arrays for improved schema readability.

Bug Fixes

• Resolved persistent migration notice for active users by implementing self-healing logic for legacy entitlements.
• Ensured generation state is correctly reset after asynchronous operations.

Technical

• Added integration tests to verify local-only execution of health checks.
• Added structured error codes for device limits and account status.
• Enforced strict redaction of sensitive data in all diagnostic outputs.

v6.0.0 - 2026-07-14

Commit Intelligence

• Added an explicitly opt-in reviewed commit workspace with optional intent, issue/branch context, notes, reversible context selection, local noise classification, secret screening, and inline provider/model/host/token disclosure. Existing one-click generation remains the default.
• Added automatic, concise, and detailed output modes; deterministic policy validation and separately confirmed repair; and exact insertion of only a valid user-approved draft into the selected repository's SCM input.
• Added versioned .gitmind/commit-policy.json rules and safe static commitlint JSON import without executing JavaScript configuration.
• Added interactive QuickPick candidate selection (gitmind.draftChoices) with live multiline preview in the top placeholder bar, displaying complete candidate subject and body text as the user navigates choices with arrow keys.
• Prompted AI candidate engine to generate three distinct archetypes: Candidate 1 (Concise 1-liner), Candidate 2 (Detailed breakdown), and Candidate 3 (Intent-focused summary).
• Overhauled the Reviewed Commit Workspace into a guided 3-step dashboard with branch pill badges, vector stroke SVG line icons, file line diff stats (+24/-10), noise classification indicators, and bulk context action buttons (Include All, Summarize All, Exclude All).
• Added real-time Commit Health Score visual gauge meter (0–100) with line-item conventional compliance status checks.
• Added dedicated Pre-Commit Code Review Panel rendering findings with color-coded severity tags (ERROR, WARNING, INFO), details, and impacted file atom links when review.enabled is active.
• Added global colon sanitization across all 16 AI providers, prompt envelopes, and response processors, ensuring prefix formatting never creates double colons like feat(commit)::.
• Added Commit Intelligence status indicator (Intelligence: Active / Off) to the top settings configuration banner.
• Redesigned Changelog Statistics and Git History Statistics popups with metric KPI cards, rating badges, vector line icons, and zero emojis.

Safety And Compatibility

• Replaced shell-interpolated Git paths with argument-based process execution and treated repository/user context as delimited untrusted data.
• Preserved all 18 providers, 12 commit styles, existing command aliases, licensing, settings, changelog, recovery, multi-repository SCM targeting, and sanitized support workflows.
• Eliminated extension-host disposable-leak warnings during clean test shutdown.

Privacy

• Removed all GitMind product telemetry, analytics settings, Application Insights runtime dependencies, and deployment assets.
• Added automatic cleanup for retired telemetry settings and locally persisted telemetry state.
• Removed the production raw-debug setting and command; development builds retain local developer logging only.
• Added Pro-only, local sanitized support sessions with a strict allowlist, privacy validation, explicit review, and manual file sharing.

Fixed

• Reworked Automatic Recovery so normal HTTP 429/model-limit failures can use the configured fallback, while account/billing quota failures remain excluded.
• Applied the same bounded recovery behavior to normal generation, custom prompts, commit-history learning, changelog generation, and large-diff chunk/merge requests.
• Normalized provider errors into meaningful user-safe categories without exposing raw response bodies or credentials.

Changed

• Centralized provider defaults and curated model choices, refreshed current OpenAI, Gemini, Anthropic, Cohere, MiniMax, Copilot, Perplexity, DeepSeek, and xAI integrations, and corrected xAI model discovery.
• Updated the extension baseline to VS Code 1.101 and Node.js 22, with npm as the declared package manager.
• Updated the handbook, generated product surface, Wiki mirror, and UI copy to match the current implementation.

GitMind Pro v5.1.0

GitMind Pro v5.1.0

Installation

Download the .vsix file below and install it manually:

code --install-extension ai-commit-assistant-5.1.0.vsix

Or install directly from the VS Code Marketplace.

See CHANGELOG.md for full release notes.

v5.0.6

• Included lock files in diff output when no other non-lock files are staged.
• Updated @types/node to 25.9.3 and @types/vscode to 1.120.0.
• Upgraded eslint to 10.5.0 for improved linting capabilities.
• Updated sanitize-html to 2.17.5 for security patches.

v5.0.4

New Features

• Added gemini-3.1-flash as the default model and updated model aliases.
• Added MiniMax-Text-01 support and pruned deprecated MiniMax models.
• Added gitmind.commit.excludeFiles configuration to ignore lock files and build artifacts.
• Implemented automatic issue ID extraction from branch names via regex with configurable placement options.
• Added VSIX release pipeline to automate extension packaging and GitHub Release creation on tag pushes.

Enhanced

• Simplified prompt generation by removing redundant generateCommitPrompt calls.
• Updated subscription validation to use absolute time difference for lock checks.
• Improved error handling by scrubbing sensitive data from API logs.
• Increased TIMEOUT_DURATION to 120s to prevent premature request failures.
• Redirected users with valid license keys to settings instead of the portal.
• Standardized Node.js version to 20 LTS and switched CI caching to pnpm.

Technical

• Configured a unique user-data-dir in the VS Code test environment.
• Synchronized model lists across configuration types and webview constants.
• Cleaned up duplicate entries in MiniMax integration test assertions.
• Updated @types/vscode to ^1.96.0 to align with targeted VS Code API versions.

v5.0.5

Enhanced

• Refactored configuration updates to use updateSingleSetting for improved resilience against unregistered configuration errors.
• Standardized asynchronous configuration updates by replacing Thenable with Promise.
• Added warning notifications to provide user feedback when settings fail to register.

Technical

• Updated GitHub Actions to Node.js 24 by setting FORCE_JAVASCRIPT_ACTIONS_TO_NODE24.
• Removed deprecated enablement property from actions/configure-pages.
• Standardized environment configuration across documentation workflows.
• Updated wiki deployment to use GITHUB_TOKEN for authentication and synchronization.
• Configured dynamic repository references and write permissions for wiki deployment jobs.
• Removed redundant token validation steps in CI workflows.
• Updated .gitignore to enable version tracking for lock files.