manul-engine 0.0.5.1

Resilient neuro-symbolic browser automation framework powered by Playwright and local LLMs (Ollama)

😼 ManulEngine — The Mastermind

ManulEngine is a relentless hybrid (neuro-symbolic) framework for browser automation and E2E testing.

Forget brittle CSS/XPath locators that break on every UI update — write tests in plain English. Stop paying for expensive cloud APIs — leverage local micro-LLMs via Ollama, entirely on your machine.

Manul combines the blazing speed of Playwright, powerful JavaScript DOM heuristics, and the reasoning of local neural networks. It is fast, private, and highly resilient to UI changes.

The Manul goes hunting and never returns without its prey.

---

✨ Key Features

⚡ Heuristics-First Architecture

95% of the heavy lifting (element finding, assertions, DOM parsing) is handled by ultra-fast JavaScript and Python heuristics. The AI steps in only when genuine ambiguity arises.

When the LLM picker is used, Manul passes the heuristic score as a prior (hint) by default — the model can override the ranking only with a clear, disqualifying reason.

🛡️ Unbreakable JS Fallbacks

Modern websites love to hide elements behind invisible overlays, custom dropdowns, and zero-pixel traps. Manul uses Playwright with force=True plus retries and self-healing; for Shadow DOM elements it falls back to direct JS helpers to keep execution moving.

🌑 Shadow DOM Awareness

The DOM snapshotter recursively inspects shadow roots and can interact with elements inside the shadow tree.

👻 Smart Anti-Phantom Guard & AI Rejection

Strict protection against LLM hallucinations. If the model is unsure, it returns {"id": null}; the engine treats that as a rejection and retries with self-healing.

🎛️ Adjustable AI Threshold

Control how aggressively Manul falls back to the local LLM via manul_engine_configuration.json (ai_threshold key) or the MANUL_AI_THRESHOLD environment variable. If not set, Manul auto-calculates it from the model size:

Model size	Auto threshold
< 1b	500
1b – 4b	750
5b – 9b	1000
10b – 19b	1500
20b+	2000

Set MANUL_AI_THRESHOLD=0 to disable the LLM entirely and run fully on deterministic heuristics.

🗂️ Persistent Controls Cache

Successful element resolutions are stored per-site and reused on subsequent runs — making repeated test flows dramatically faster.

---

🛠️ Installation

pip install manul-engine
playwright install chromium

Optional: Local LLM (Ollama)

Ollama is only needed for AI element-picker fallback or free-text mission planning.

ollama pull qwen2.5:0.5b
ollama serve

---

🚀 Quick Start

1. Create a hunt file

my_tests/hunt_smoke.hunt

@context: Demo smoke test
@blueprint: smoke

1. NAVIGATE to https://demoqa.com/text-box
2. Fill 'Full Name' field with 'Ghost Manul'
3. Click the 'Submit' button
4. VERIFY that 'Ghost Manul' is present.
5. DONE.

2. Run it

# Run a specific hunt file
manul my_tests/hunt_smoke.hunt

# Run all *.hunt files in a folder
manul my_tests/

# Run headless
manul my_tests/ --headless

# Run an inline one-liner
manul "1. NAVIGATE to https://example.com  2. Click the 'More' link  3. DONE."

3. Python API

import asyncio
from manul_engine import ManulEngine

async def main():
    manul = ManulEngine(headless=True)
    await manul.run_mission("""
        1. NAVIGATE to https://demoqa.com/text-box
        2. Fill 'Full Name' field with 'Ghost Manul'
        3. Click the 'Submit' button
        4. VERIFY that 'Ghost Manul' is present.
        5. DONE.
    """)

asyncio.run(main())

---

📜 Hunt File Format

Hunt files are plain-text test scenarios with a .hunt extension.

Headers (optional)

@context: Strategic context passed to the LLM planner
@blueprint: short-tag

Comments

Lines starting with # are ignored.

System Keywords

Keyword	Description
NAVIGATE to [URL]	Load a URL and wait for DOM settlement
WAIT [seconds]	Hard sleep
SCROLL DOWN	Scroll the main page down one viewport
EXTRACT [target] into {var}	Extract text into a memory variable
VERIFY that [target] is present	Assert text/element is visible
VERIFY that [target] is NOT present	Assert absence
VERIFY that [target] is DISABLED	Assert element state
VERIFY that [target] is checked	Assert checkbox state
DONE.	End the mission

Interaction Steps

# Clicking
Click the 'Login' button
DOUBLE CLICK the 'Image'

# Typing
Fill 'Email' field with '[e-mail]'
Type 'hello' into the 'Search' field

# Dropdowns
Select 'Option A' from the 'Language' dropdown

# Checkboxes / Radios
Check the checkbox for 'Terms'
Uncheck the checkbox for 'Newsletter'
Click the radio button for 'Male'

# Hover & Drag
HOVER over the 'Menu'
Drag the element "Item" and drop it into "Box"

# Optional steps (non-blocking)
Click 'Close Ad' if exists

Variables

EXTRACT the price of 'Laptop' into {price}
VERIFY that '{price}' is present.

---

⚙️ Configuration

Create manul_engine_configuration.json in your project root — all settings are optional:

{
  "model": "qwen2.5:0.5b",
  "headless": false,
  "timeout": 5000,
  "nav_timeout": 30000,

  "ai_always": false,
  "ai_policy": "prior",
  "ai_threshold": null,

  "controls_cache_enabled": true,
  "controls_cache_dir": "cache",

  "log_name_maxlen": 0,
  "log_thought_maxlen": 0
}

Set "model": null (or omit it) to disable AI entirely and run in heuristics-only mode.

Environment variables (MANUL_*) always override JSON values — useful for CI/CD:

export MANUL_HEADLESS=true
export MANUL_AI_THRESHOLD=0
export MANUL_MODEL=qwen2.5:0.5b

Key	Default	Description
model	null	Ollama model name. null = heuristics-only (no AI)
headless	false	Hide browser window
ai_threshold	auto	Score threshold before LLM fallback. null = auto by model size
ai_always	false	Always use LLM picker, bypass heuristic short-circuits
ai_policy	"prior"	"prior" (LLM may override score) or "strict" (enforce max-score)
controls_cache_enabled	true	Persistent per-site controls cache
controls_cache_dir	"cache"	Cache directory (relative to CWD or absolute)
timeout	5000	Default action timeout (ms)
nav_timeout	30000	Navigation timeout (ms)
log_name_maxlen	0	Truncate element names in logs (0 = no limit)
log_thought_maxlen	0	Truncate LLM thoughts in logs (0 = no limit)

---

📋 Available Commands

Category	Command Syntax
Navigation	NAVIGATE to [URL]
Input	Fill [Field] with [Text], Type [Text] into [Field]
Click	Click [Element], DOUBLE CLICK [Element]
Selection	Select [Option] from [Dropdown], Check [Checkbox], Uncheck [Checkbox]
Mouse Action	HOVER over [Element], Drag [Element] and drop it into [Target]
Data Extraction	EXTRACT [Target] into {variable_name}
Verification	VERIFY that [Text] is present/absent, VERIFY that [Element] is checked/disabled
Flow Control	WAIT [seconds], SCROLL DOWN
Finish	DONE.

Append if exists or optional to any step (outside quoted text) to make it non-blocking, e.g. Click 'Close Ad' if exists

---

🐾 Battle-Tested

ManulEngine is verified against 1200+ synthetic DOM tests covering:

• Shadow DOM, invisible overlays, zero-pixel honeypots
• Custom dropdowns, drag-and-drop, hover menus
• Legacy HTML (tables, fieldsets, unlabelled inputs)
• AI rejection & self-healing loops
• Persistent controls cache hit/miss cycles

---

Version: 0.0.5.1 · Status: Hunting...