Hablas CLI

Single-agent terminal AI engineering runtime.

Hablas is being rebuilt from scratch around one strict product decision:

one visible agent, one execution surface, full tools, zero internal theatre

This stable branch no longer treats multi-agent choreography as the product. The product is now Hablas only.

Product principles

1. Single visible agent
2. Fast direct path for trivial turns
3. Full tool execution for real engineering work
4. No hidden routing leaks in the UX
5. No handoff theatre, no mode maze, no orchestration drama
6. Read before edit, verify before claim
7. Direct technical tone: no emojis, no hype, no filler
8. Professional shell UX matters as much as model quality

Installation

git clone https://github.com/Monopoly63/mobile-bot.git
cd mobile-bot/flowagent
npm install
npm run build
npm link

Requirements:

• Node.js 20+
• One provider:
  • Ollama
  • NVIDIA NIM
  • any OpenAI-compatible API

Usage

First run / setup

On first launch, Hablas now offers the restored classic setup flow:

1. Hablas Integrated Engine

   • ready-made
   • preconfigured NVIDIA-backed endpoint
   • selected model already configured
   • user just chooses option 1

2. Custom Provider

   • Ollama
   • OpenAI
   • Groq
   • OpenRouter
   • Together
   • DeepSeek
   • custom OpenAI-compatible API
   • NVIDIA NIM

You can also run setup manually anytime:

hablas --setup

Interactive

hablas

Interactive with automatic tool confirmations where allowed

hablas --auto

One-shot

hablas run "read package.json and tell me the version"

Web and asset workflow

Hablas now keeps the professional web tool chain active:

• web_search
• search_image_candidates
• inspect_image
• scrape_url
• extract_links
• download_asset

This means Hablas can:

• search the web for references and assets
• rank trusted image page candidates before downloading
• inspect a remote or local image candidate before adoption
• resolve a page candidate into a direct image candidate with an explicit verdict
• download real images/files into the project when appropriate
• preserve full file reads in-context so long files are not silently cut into tiny fragments that force rereads

Image pipeline stance

For image replacement work, the intended flow is:

1. search_image_candidates
2. inspect_image
3. download_asset

The runtime should stay inside this image toolchain. It should not fall back to shell hacks (curl, powershell, guessed public image IDs) for public asset discovery unless the user explicitly asks for low-level debugging.

Current command surface

CLI commands

• hablas
• hablas run "..."
• hablas doctor
• hablas info
• hablas security
• hablas stats
• hablas --setup

Slash commands

• /help
• /status
• /model [name]
• /models [query]
• /provider [ollama|nvidia|custom|test]
• /history [n]
• /clear
• /workspace
• /doctor
• /version
• /about
• /exit

Shell UX

• tab completion for slash commands
• tab completion for common file paths after verbs like read, write, edit, patch
• #file reference completion
• persisted command history
• rebuilt Apple-slate terminal styling
• visible thinking state while the model is working
• direct-turn streaming when the provider supports it

Execution model

Hablas analyzes a request into one of four internal task kinds:

• casual
• read
• analysis
• implementation

The runtime uses a lightweight local task profiler so first-turn latency stays clean and there is no hidden pre-turn model call before Hablas starts responding.

Then it decides whether the turn should be:

• direct text execution
• tool-based execution

There is no public build/design/ask mode system anymore. There is no visible team routing layer.

Examples

Simple direct turn

hi

Expected behavior:

• no tool calls
• short direct answer

Project inspection

what is the current version of this project

Expected behavior:

• inspect the minimum relevant files
• answer from evidence

Architecture request

design the architecture for auth in this repository

Expected behavior:

• structured analysis
• inspect files only if needed
• answer as one agent

Continuation request

continue the unfinished project

Expected behavior:

• continue from existing work
• do not restart from zero

One-step delivery request

just give me the finished project in one step no questions

Expected behavior:

• aggressive execution posture
• no unnecessary follow-up questions
• verify before claiming done

Architecture docs

See:

• docs/ARCHITECTURE.md
• docs/CLI.md
• docs/ROADMAP.md
• docs/STATUS.md
• docs/PUBLISH.md

Development

npm install
npm run build
npm test
npm run dev -- run "your prompt"

Current automated checks cover:

• shared execution engine smoke behavior
• analyzer behavior
• session continuity behavior
• image pipeline helper behavior

Memory and state

Hablas now uses larger defaults for the rebuilt runtime and also keeps a project-local state folder:

• larger message history budget
• larger context budget
• project-local .hablas/session.json persistence for the current workspace

This is separate from the user-home config directory used for global config/history.

Rebuild status

This branch is under an intentional full rebuild of the stable product surface. Dead code from the previous multi-agent product is being physically removed instead of patched.

License

MIT