Release Notes

AI Code Sherlock
Agent Constructor v3.0

The most ambitious update in Sherlock's history. A full visual AI workflow designer — drag-and-drop agents, browser automation, multi-threaded project execution, and a runtime engine with self-healing pipelines.

📅 April 3, 2026
🏷️ v3.0.0
🔬 Major Release
✅ MIT License
40+
Node Types
∞
Workflow Complexity
N
Parallel Projects
7
Browser Actions
3
Routing Modes

🚀 Headline Features

🧩
Visual Agent Constructor
Build AI pipelines with drag-and-drop. Connect AI agents, snippets, conditions, loops, and browser nodes on an infinite canvas with real-time minimap navigation.
New Module
🌐
Browser Automation Suite
7 dedicated browser nodes: Launch, Action, Close, Screenshot, Click-by-Image, Profile Management, and AI Browser Agent with DOM comprehension.
Browser
⚡
Multi-Threaded Project Runner
Run multiple workflow projects in parallel. Per-project thread control, stop conditions, cooldown timers, and a live dashboard showing all execution states.
Major
🔧
Self-Healing Runtime Engine
Exponential backoff retries, auto-prompt improvement on failure, dynamic skill injection, and automatic code patching when output validation fails.
Engine
🧠
LLM-Based Dynamic Routing
Three routing modes: Sequential, Conditional (expression eval), and LLM Router — where an AI model decides which node executes next at runtime.
Engine
🎯
Skill Registry
10 built-in reusable skills (Code Gen, Patching, Debugging, Vision, etc.) plus user-defined skills stored in JSON. Dynamic skill selection via LLM suggestion.
New
🐛
Step-By-Step Debugger
Pause, step, resume any running workflow. Set breakpoints on individual nodes. Inspect execution context, variable state, and path history at any step.
New
↩️
Full Undo/Redo History
50-step undo/redo stack for all canvas edits — node moves, property changes, connections. Ctrl+Z / Ctrl+Y / Ctrl+Shift+Z across the entire constructor.
UI

🗂️ Complete Node Library (40+ Types)

💻Code Writer
🔍Code Reviewer
🧪Tester
🏗️Planner
🎨Image Gen
👁️Image Analyst
📁File Manager
▶️Script Runner
✅Verifier
🎯Orchestrator
🩹Patcher
📜Code Snippet
❓If Condition
🔁Loop
🔀Switch
📝Variable Set
🌐HTTP Request
⏳Delay
📋Log Message
🔔Notification
🟨JS Snippet
⚙️Program Launch
📃List Operation
📊Table Operation
📄File Operation
📁Dir Operation
✂️Text Processing
🔣JSON/XML
🔧Variable Proc
🎲Random Gen
🌐Browser Launch
🖱Browser Action
🔴Browser Close
🌐🧠Browser Agent
🖼Click by Image
📸Screenshot
🪪Profile Op
✅Good End
🛑Bad End
📌Note
▶Project Start

📋 Full Changelog

✨ Agent Constructor — Brand-new visual workflow editor with PyQt6 QGraphicsScene canvas, node drag-drop, bezier edge rendering, zoom/pan, and minimap. New
✨ WorkflowRuntimeEngine — Full async execution engine: retries with exponential backoff, auto-skill loading, output validation, self-patching, and LLM router. Engine
✨ ProjectExecutionManager — Thread pool for running N projects simultaneously. Per-project start/stop, cooldown, stop conditions, and dashboard status widget. New
✨ Browser Automation Module — 7-node browser suite: launch profiles, click, type, screenshot, click-by-template-image, close, and full AI DOM agent. Browser
✨ SkillRegistry — Centralized skill store: 10 built-in skills, user JSON skills, LLM-powered skill recommendation, prompt injection, project-scoped skills. New
✨ WorkflowDebugger — Step-by-step executor: pause/step/resume, breakpoints on nodes, path history, variable inspection, and logger callbacks. New
✨ HistoryManager + Command pattern — Full undo/redo with 50-item stack, legacy tuple format support, and recursion guard via _is_undoing flag. UI
✨ WorkflowView — Ctrl+Scroll zoom, Shift+Scroll horizontal pan, Middle-click drag pan, edge-zone auto-scroll during node drag. UI
✨ AgentWorkflow serialization — Complete JSON save/load for workflows including all node properties, edges with priority/conditions, and execution metadata. Core
✨ Planner preprocessing mode — Browser Agent can auto-split AI plan output into sequential subtasks with configurable separators (numbered, bullet, header, regex). Browser
✨ Verification system — Three modes: self-check, another model, custom agent. Configurable strictness (stop-on-fail or warn-and-continue). Engine
🔧 AgentNode extended fields — Added reasoning_model_id, fast_model_id, auto_improve, auto_test, backoff_strategy, conditional_branches, and more. API
🔧 EdgeCondition enum — Four edge conditions: ALWAYS, ON_SUCCESS, ON_FAILURE, ON_CONDITION. Edges carry priority for deterministic multi-path sorting. Core
🔧 Cycle detection — add_edge() validates node existence, duplicate conditions, and self-loops. Intentional cycles (ScriptRunner↔Patcher) are explicitly allowed. Logic
🔧 Autosave — Canvas state auto-saved to temp file on changes; recovered on next launch. Prevents work loss on crash or accidental close. UX
Ready to automate everything?

Download v3.0 and build your first agent workflow in minutes.

⬇ Download v3.0 ⭐ Star on GitHub 📖 Read the Docs