v2.7.158

What's Changed

• Add structured ask mode with intent routing and AI-generated commits by @codewithshinde in #1
• Enhance file caching, plan orchestration, and skill integration by @codewithshinde in #3
• Enhance Act orchestration with GitHub integration and UI improvements by @codewithshinde in #4
• Add support for AWS Bedrock, Azure OpenAI, OpenRouter, and new features by @codewithshinde in #6
• Enhance memory management, CLI functionality, and team services by @codewithshinde in #8
• Enhance file handling, tool functionality, and UI improvements by @codewithshinde in #9
• Migrate tools and enhance commit message generation with tests by @codewithshinde in #10

New Contributors

• @codewithshinde made their first contribution in #1

Full Changelog: v2.6.16-beta...v2.7.158

v2.6.16-beta

Release: The Mitii Rebrand & Core Intelligence Update (v2.16.beta)
We are incredibly excited to announce a major milestone for our local-first AI coding assistant. With this release, Thunder AI Agent has been officially rebranded to Mitii AI Agent.

Along with the new identity, this update introduces massive overhauls to the agent's core intelligence, prompt budgeting, and safety boundaries. We’ve replaced basic prompts with a strict, stateful task engine, ensuring Mitii behaves predictably, protects your context window, and executes complex plans without hallucinatory loops.

Here is what’s new in Mitii:

🌟 Major Highlights
The Mitii Rebrand: All internal and external references, configuration keys (now mitii.*), repository URLs, and documentation have been updated from Thunder to Mitii. The local storage folder is now .mitii/, and the database is mitii.sqlite.

Agent State Machine & Phase-Locking: We replaced manual string tracking with a strict, event-driven state machine (AgentTaskState). Mitii is now fully "phase-locked," meaning it is mathematically impossible for the agent to execute destructive write operations while in a read-only "Plan" phase.

Advanced Prompt Budgeting: Introduced fitChatRequestToBudget to eradicate "Lost in the Middle" syndrome. Mitii now dynamically compacts the conversation transcript, truncates older tool outputs, and shrinks codebase blocks to keep your prompt lean and fast before you hit your token limit.

🛠️ New Tools & Capabilities
Workspace Scripting: Added search_script_catalog and execute_workspace_script. Mitii can now discover and safely run pre-approved scripts directly from your workspace.

Task Saving (save_task_state): Mitii can now explicitly checkpoint its progress and state its next intended step before pausing for user approval, completely solving the issue of agents "forgetting" what they were doing after a long pause.

Context Expansion: Introduced expandContextQuery. Mitii now expands your intent before searching the SQLite FTS5 index, retrieving much higher-quality files. (Internal agent files are now strictly filtered out of context to reduce noise).

TaskAnalyzer: Automatically classifies user requests (Implementation, Audit, Question, Edit) to route the agent down the most efficient execution path. For example, Audit tasks now skip the Planner phase entirely for lightning-fast execution.

🛡️ Safety & Validation
Post-Write Validation: Mitii now supports immediate post-write validation (linting, type-checking). If it writes bad code, it catches the error and reflects on it before continuing.

Bulletproof MDX Handling: Added a dedicated mdxValidation.ts module to prevent common Docusaurus build failures (like raw TypeScript generics in tables or broken LiveCodeBlocks) during write_file and apply_patch operations.

Configurable Checkpoints: Added support for multiple safety checkpoint strategies before writes: file-copy, git-stash, and shadow-git.

Tool Guidance: Added ASK_TOOL_GUIDANCE to give the agent better contextual nudges, preventing redundant diagnostic loops (e.g., calling eslint over and over) and caching previous tool outputs as "soft blocks."

💻 UI & Developer Experience
Premium Webview UI: Added beautiful framer-motion animations to the approval/denial controls, and integrated ShikiCodeBlock for native, VS Code-accurate syntax highlighting in the chat.

Context Debugger UI: Added a new debugger view displaying the exact token consumption of the last request and your current context window limits.

Non-Blocking Memory: MemoryExtractor has been moved to an asynchronous background worker. Extracting long-term memory at the end of a task no longer blocks the UI thread.

Expanded Model Support: Improved routing and handling for a wider array of providers (OpenAI, Anthropic, Gemini, etc.) and added custom base URL configurations.

Intensity Modes (Preview): Laid the architectural groundwork and CSS for upcoming Agent Intensity Levels (Low, Medium, High, Rocket), which will allow users to dial the agent's autonomy and token budgets up or down.

Upgrading:
If you are upgrading from an older version, please note that your .thunder configuration folders should be renamed to .mitii, and any VS Code settings.json keys starting with thunder. must be updated to mitii..

Full Changelog: v2.6.7-beta...v2.6.16-beta

v2.6.7-beta

Full Changelog: https://github.com/Mitii-dev/Mitii/commits/v2.6.7-beta

Features

Context and indexing

• Workspace scanner respects .gitignore and .thunderignore; auto-indexes on folder open
• FTS5 full-text search with ripgrep fallback for unindexed paths
• Symbol extraction for TypeScript, JavaScript, Python, Java, Go
• PageRank repo map — surfaces the files that matter structurally, not just lexically
• Vector search — local MiniLM embeddings via @xenova/transformers, hash fallback if unavailable; SQLite or LanceDB backend
• Hybrid retriever + reranker — merges search, vectors, mentioned files, git state, and LSP diagnostics into one context pack

Agent workflow