v0.14.0 - Introduces the new dual-mode CLI architecture for multi-ai-cli.

What’s included

• add startup mode dispatch based on stdin.isatty()
• keep the existing stateful Interactive REPL mode intact
• add a new stateless Unix-style Filter mode (stdin -> AI -> stdout)
• add a dedicated filter_mode.py runner
• support Filter mode with exactly one AI agent plus -m and -r
• reject interactive-only features in Filter mode such as -w, -e, @sequence, ->, ||, and non-AI command adapters
• ensure Filter mode writes only the final AI result to stdout
• route diagnostics and validation errors to stderr
• suppress interactive auto-continue progress output during Filter mode execution
• extract shared reference-file loading for reuse across prompt builders
• add/update validation and checker scripts for the new behavior

Why

This change allows multi-ai-cli to be used in two distinct ways:

1. as a rich interactive REPL for multi-step AI workflows
2. as a shell-friendly filter for scripting and pipelines

Examples now supported:

• multi-ai
• echo "apple" | multi-ai @gpt -m "Translate this to Japanese only."
• cat spec.md | multi-ai @gpt -m "Write a design doc" -r existing_code.py > design.md

Notes

This PR intentionally keeps Filter mode minimal and focused:

• single AI agent only
• stateless execution
• stdin as primary input
• shell redirection for output files

It does not expand Filter mode to adapters or orchestration features.

v0.13.0 - Agent/Engine Separation and GitHub Adapter

v0.13.0 — Agent/Engine Separation and Read-Only GitHub Adapter

This release introduces two major upgrades to multi-ai-cli:

• Agent/Engine Separation
• Read-only GitHub Adapter

Together, these changes make the CLI significantly more flexible as a multi-agent development environment and expand its external interface surface beyond AI engines.

---

Highlights

Agent/Engine Separation

multi-ai-cli now separates:

• models ([MODELS])
• physical execution backends ([ENGINE.*])
• logical stateful agents ([AGENT.*])

This makes it possible to define reusable engine backends and bind them to multiple logical agents such as:

• @gpt
• @gpt.code
• @claude.review
• @gemini.chat
• @local.test

Key benefits:

• cleaner configuration structure
• reusable backend definitions
• agent-level conversation history separation
• clearer distinction between provider/backend and logical role
• better scalability for multi-agent orchestration

Legacy config format is still supported through a compatibility layer.

---

New GitHub Adapter (@github.*)

This release adds a new read-only GitHub adapter with native CLI commands for inspecting repositories directly from the terminal.

New commands:

• @github.repo
• @github.tree
• @github.file
• @github.issue
• @github.issues

Supported capabilities:

• repository metadata read
• directory/path listing
• file content read
• single issue read
• issue list read

Additional behavior:

• supports private repositories via token authentication
• filters pull requests out of @github.issues
• rejects pull requests in @github.issue (issue-only in v1)
• works naturally with -w for saving outputs into local artifacts

This makes GitHub a first-class external interface in multi-ai-cli, alongside the existing Shell and Figma adapters.

---

Configuration Changes

New Agent/Engine layout

The recommended config structure is now:

• [MODELS]
• [RUNTIME]
• [ENGINE.*]
• [AGENT.*]

This replaces the older provider-only mental model with a more explicit architecture.

New GitHub config section

Added support for:

[GITHUB]
token = ...
api_base_url = https://api.github.com

Environment variable overrides:

• GITHUB_TOKEN
• GITHUB_API_BASE_URL

---

Internal Improvements

• command dispatch updated for @github.*
• parser validation updated to recognize GitHub adapter commands
• welcome banner/help text updated for the new architecture
• registry reset behavior added for safer re-initialization and testing
• improved separation between adapter commands and agent commands

---

Example Workflows

Fetch a GitHub issue and use it as AI input:

@github.issue --repo ashiras/multi-ai-cli --number 40 -w issue40.md
@gpt.plan "Summarize the issue and propose an implementation plan." -r issue40.md

Fetch a file and review it with an AI agent:

@github.file --repo ashiras/multi-ai-cli --path src/multi_ai_cli/handlers.py -w handlers.py
@claude.review "Review this file and suggest improvements." -r handlers.py

---

Notes

• GitHub support in this release is read-only
• pull request support is not included in v0.13.0
• current implementation expects a GitHub token for @github.* commands
• legacy configuration remains supported, but the new Agent/Engine format is recommended

Thanks

This release marks an important architectural step for multi-ai-cli:
from a simple multi-model CLI into a more explicit multi-agent + multi-adapter orchestration environment.

As always, feedback, bug reports, and ideas are welcome.

Release v0.12.0: Figma Integration, Shell Orchestration & Auto-Continue

🚀 Release v0.12.0: Figma Integration, Shell Orchestration & Auto-Continue

Welcome to Multi-AI CLI v0.12.0! This major update bridges the gap between design workflows, local shell execution, and multi-agent AI orchestration. We've transformed the CLI into a true "strategic hub" for developers and designers alike.

✨ What's New in v0.12.0

🎨 1. Figma Adapter (@figma)

Seamlessly integrate your AI workflow with Figma design data. Bridge the gap between "Design as Code" and AI generation.

• @figma.pull: Fetch design files, specific nodes, or pages directly from Figma's REST API. Output as raw or AI-friendly normalized JSON.
• @figma.push: Send local content (Markdown, generated JSON specs) directly to a Figma-side plugin bridge, enabling AI-driven content updates.

⚙️ 2. Native Shell Orchestration (@sh)

Integrate directly with your local environment. Run commands, execute scripts, and feed the output right back into your AI sequence.

• Execute Scripts: Run .py, .sh, .ts, and more natively (@sh -r script.py).
• Capture Output: Save stdout/stderr as human-readable text or structured JSON (-w output.json).
• Complex Commands: Support for pipes and environment variables using the --shell flag.

🔄 3. Automatic Response Continuation

Never miss a word from your AI again.

• If an AI's response hits its maximum token limit (e.g., generating massive code blocks), the CLI now automatically detects it and instructs the AI to seamlessly continue from exactly where it stopped.
• Fully configurable via multi_ai_cli.ini (auto_continue_max_rounds, auto_continue_tail_chars).

---

💻 Example Workflow: The Power of v0.12.0

Combine the new features in @sequence (HAN Syntax) to build incredible pipelines:

# 1. Pull design data from Figma
@figma.pull --file "FILE_KEY" --node "NODE_ID" -w component.json

-> # 2. AI generates React code based on the design
   @claude "Generate a React component based on this design data." -r component.json -w:code Button.tsx

-> # 3. Run a local linter/formatter on the generated code
   @sh "npx prettier --write Button.tsx"

🍎 Installation & Upgrade

Download the pre-built binary for macOS / Linux:

Check the assets below to download the executable.

chmod +x multi-ai
sudo mv multi-ai /usr/local/bin/

For Python Developers (Source):

uv tool install multi-ai-cli  # Or your preferred uv run method

View the full documentation in the README.

Multi-AI CLI v0.11.0: The Local-Loop Evolution

This major update transforms the CLI from a simple AI client into a robust AI Orchestration Hub. By bridging Local LLMs and the Local Shell, we enable a fully private, autonomous development loop.

🚀 Key Updates:

• local Engine: Privacy-first AI integration via Ollama/LM Studio (OpenAI-compatible API).
• sh Shell Orchestration: Direct terminal command execution with Runner Map (Auto-detects Python, Bash, Ruby, etc.).
• Unified Dispatcher: Success/failure-aware command routing for reliable sequence pipelines.
• Improved I/O: Refined -w:raw and -w:code behavior for UNIX-like predictability.

Build the future: Think (Local) -> Write -> Execute -> Reflect.

Multi-AI CLI v0.9.1

This merge marks a major update to Multi-AI CLI, bringing it to version v0.9.1. The primary focus of this release is the introduction of complex multi-agent workflow orchestration via the HAN (Human-Agent-Network) Syntax and a complete overhaul of the file I/O system to follow a "predictable" UNIX-like philosophy.

Key Features & Changes

1. Workflow Orchestration (@sequence -e)

The @sequence command now supports an editor mode (-e), enabling users to define and execute sophisticated pipelines using the HAN Syntax.

• Sequential Execution (->): Define clear dependencies where one step follows another.
• Parallel Execution ([ ... || ... ]): Execute multiple AI tasks concurrently using ThreadPoolExecutor.
• Artifact Relay: Files written via -w in one step can be immediately consumed as input via -r in subsequent steps.
• Cascade Stop: Automatically halts the entire sequence if any individual task or step fails, preventing downstream errors.
• Thread-Safe UI: Uses threading.Lock to synchronize terminal output, preventing log "spaghetti" during parallel execution.

2. Refined File I/O (Raw-by-Default)

The I/O behavior has been redesigned to optimize both documentation and code generation workflows.

• Raw-by-Default (-w <file>): Saves the full AI response exactly as received. This ensures the integrity of READMEs, documentation, and context-heavy responses.
• Code Modifier (-w:code <file>): An explicit modifier to extract and concatenate fenced code blocks.
• Fallback Logic: If :code is specified but no code blocks are found, the system gracefully falls back to saving the full response.
• Multi-Input Support: Multiple -r flags can now be used to inject several local files into a single prompt context.

3. Fixed Prompt Construction Priority

To ensure stable inference, the final prompt sent to the AI is now assembled in a strictly defined order:

1. A1 (Title/Context)
2. Message (-m)
3. Editor (-e)
4. Files (-r)

4. Bug Fixes (Inception Bug)

• Markdown Parsing: Resolved the "Inception Bug" where the presence of triple backticks (```) inside the generated code would prematurely truncate the output. The parser now uses a robust line-by-line approach compliant with Markdown specifications.

Execution Sample (HAN Syntax)

# Step 1: Design
@gemini "Propose a system architecture" -w design.md
->
# Step 2: Parallel Review & Coding
[ 
  @gpt "Implement the code" -r design.md -w:code app.py 
  || 
  @claude "Review the specification" -r design.md -w review.txt 
]
->
# Step 3: Final Integration
@grok "Provide a final evaluation" -r app.py -r review.txt -w final.md

Tech Stack

• Python 3.x
• concurrent.futures.ThreadPoolExecutor (Concurrency)
• threading.Lock (UI Synchronization)
• re (Modifier Parsing)
• os.fsync (I/O Consistency)

Multi-AI CLI v0.5.4

Release Notes - v0.5.4 🚀

This release introduces a major architectural overhaul to improve stability, security, and scalability, along with several critical bug fixes.

🌟 Key Changes

🏗️ Architectural Overhaul (AIEngine Class)

• Refactored the engine management to a class-based architecture (AIEngine). This ensures robust state management across all providers (Gemini, GPT, Claude, Grok) and simplifies the addition of future models.

🐞 Bug Fixes

• Fixed @efficient Parsing: Resolved an issue where specifying the all target caused the command to misinterpret it as a filename.
• Gemini Session Reset: Fixed a scoping bug where @scrub gemini failed to correctly clear the conversation context.
• System Prompt Duplication: Eliminated an issue in GPT and Grok where system instructions were being duplicated in the conversation history.

🔒 Security & UX Improvements

• Directory Traversal Protection: Implemented strict path resolution to prevent unauthorized access outside of designated work directories.
• Environment Variable Support: API keys can now be loaded via environment variables (e.g., GEMINI_API_KEY), prioritizing them over INI file settings.
• HUD HUD Refresh Logic: Improved terminal width detection to ensure the "thinking..." status line is cleared accurately on all screen sizes.

📦 Installation & Update

1. Download the binary for your OS from the Releases page.
2. Replace your existing binary:

chmod +x multi-ai-macos
sudo mv multi-ai-macos /usr/local/bin/multi-ai

3. Verify the version:

multi-ai --version

Multi-AI CLI v0.4.1

• Quad-Engine support (Gemini, GPT, Claude, Grok)
• Real-time HUD logging (chat interface)
• Stealth Mode support
• Bug fixes for code block formatting

Multi-AI CLI First Beta

v0.1.0-beta

Add Google Generative AI and OpenAI to requirements