🌓

Features

Complete reference for every user-facing feature in MultiTerminal. The application is a Windows desktop host for multiple Claude Code terminal sessions with integrated project management, team coordination, and institutional memory.

UI Panels (11 panels + controls)

MultiTerminal uses a VS-style docking framework (DockPanelSuite). All panels can be toggled from the toolbar, dragged to dock at any edge, floated as windows, or auto-hidden as slide-out tabs.

Tasks Panel (Kanban Board)

Chat Panel

Activity Panel

Project Panel

Profile Panel

Inbox Panel

Office Panel

Agent Panel

File Preview Panel

Start Screen

Terminal Controls

Terminal Status Bar

Task HUD

Per-Terminal Browser Tabs

Task Lifecycle Board

Debug Panel

Terminal Management

ConPty Hosting

Each terminal tab hosts a ConPty (Windows Console Pseudo-Terminal) instance rendered inside a WebView2 control. This provides a full terminal emulator experience with proper ANSI escape code support, colors, cursor positioning, and resize handling.

Tab Creation & Management

Status Bar (3-Row Layout)

Each terminal has a status bar at the bottom rendered via WebView2:

• Row 1 (Folder): Current working directory path
• Row 2 (Model/Git): Active Claude model, git branch name, and dirty/clean indicator
• Row 3 (Context): Visual progress bar showing context window token usage

Session Persistence

Terminal sessions are automatically saved and restored across application restarts:

• Terminal identity (name), working directory, and Claude Code session ID are persisted
• On restart, terminals are recreated with their previous identity and can resume Claude Code sessions
• Session history is stored in SQLite for long-term access
• The docking layout (panel positions, sizes, visibility) is saved as an XML layout file

Start Screen

When launching with no saved session, the Start Screen panel displays registered projects. Click a project to open a terminal in that project's directory with its configuration pre-loaded.

Kanban Board

Task & Checklist Lifecycle

Tasks move through todo → in_progress → done, and each checklist item runs its own pending → coding → testing → done cycle (with coding ↔ testing looping when testing turns up a fix). Every transition carries a mandatory note that is delivered as an inbox notification. For the full state machines, the one-active-task rule, and the testing gate, the Kanban Workflow guide is the canonical reference.

Plans

Tasks can have a markdown plan field describing the implementation approach. Plans are written by agents during the planning phase and serve as the work specification.

Helpers

Additional agents can be assigned as "helpers" to a task. The helper system allows multiple agents to collaborate on a single task, each handling different aspects of the work.

Continuation Notes

Free-text field for session handoff context. When an agent's session ends mid-task, continuation notes capture the current state so the next session (or a different agent) can pick up where it left off.

Stale Task Tracking

Tasks with no recent activity are flagged as "stale" by the StaleTaskService. This helps identify blocked or abandoned work items that need attention.

Priority System

Tasks can be assigned priority levels that affect their visual presentation on the board. Higher-priority tasks are highlighted with color-coded badges.

Git Worktree Isolation

When worktree mode is enabled (MULTITERMINAL_WORKTREE_MODE=on), claiming a task automatically gives each agent its own git worktree, so multiple agents can work on different tasks — or different parts of the same task — without stepping on each other's files.

Messaging

Point-to-Point Chat

Any registered terminal can send a direct message to any other registered terminal. Messages are routed through the MessageBroker and delivered to the recipient's terminal via webhook injection.

Broadcasts

A terminal can broadcast a message to all other registered terminals simultaneously. Broadcasts are used for team-wide announcements.

Inbox Notifications

The inbox system delivers notifications for:

• New messages directed at your terminal
• Task assignment changes (claimed, status updated)
• Checklist transitions with notes
• Helper assignments
• System events (stale task warnings)

Reliable Delivery

Messages use a reliable delivery system backed by the messages.db database:

• Messages are queued in SQLite before delivery
• Failed deliveries are retried on a timer (every 2 seconds)
• Message deduplication prevents double-delivery
• Messages expire after a configurable TTL

Real-Time Delivery (Channels)

When a message arrives for a terminal, it is delivered through Claude Code's native Channels transport: the message appears in the session as a structured <channel> event the agent reacts to immediately, and the agent replies with the channel's reply/send tools — no polling required. This replaces the older [cm] ConPTY "check messages" nudge. See Channels (Messaging) for how it works and the setting it requires.

Project Management

Project Registration

Projects are registered in SQLite with metadata including:

• Name, description, and primary path
• Source path and deploy path
• Build command
• Git repository URL, default branch, and auto-commit preference
• Pinned status for quick access

Context Service

The ProjectContextService provides a single API call (GET /api/projects/{id}/context) that returns everything an agent needs to know about a project: core details, agents, MCP servers, specialist agents, named paths, prompts, and skills.

Agent Assignment

Projects can have named agents assigned with roles (lead, member) and preferred Claude models. This determines which agents are spawned when working on the project.

MCP Configuration

Each project can have MCP servers configured and enabled/disabled. The McpConfigService generates .mcp.json files from the registry for Claude Code consumption.

Specialist Agents

Projects can enable specialist agents with custom prompts:

• Verifier: Confirms work is genuinely complete after coding
• Devils Advocate: Scores plans 0-100, blocks bad approaches
• Test Designer: Generates acceptance criteria per checklist item
• Security Auditor: OWASP Top 10 scanning, injection detection
• Debugger: Root cause analysis when testing fails
• Code Reviewer: Quality, patterns, naming, and consistency review
• Session Distiller: Compresses session learnings into memory
• Session Summarizer: Generates session recaps from JSONL transcripts

New Project Wizard

The /new-project skill provides a 13-step conversational wizard for registering new projects. It supports 5 archetypes: Clarion COM, WebView2, Clarion App, MT Feature, and Generic C#.

Team & Agent System

Native Claude Code Teams

MultiTerminal supports Claude Code's native team agent system (requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1). Teams are created via TeamCreate, and agents are spawned as subagents that appear in the Agent Panel.

Spawned Agents (AgentProcess)

The TerminalSpawner service can launch Claude Code processes programmatically. Each spawned agent:

• Gets its own terminal with a unique identity
• Is registered in the messaging system automatically
• Can communicate with other terminals via MCP tools
• Has its transcript monitored by TeamWatcherService

TeamWatcherService

Monitors JSONL transcript files produced by native team agents. Streams transcript events to the Agent Panel in real-time using a file-tailing approach (TranscriptTailer).

Agent Naming Convention

System	Format	Example	Use For
MultiTerminal	Plain name	Alice, Diana	Planning/design (interactive)
Native Teams	"Agent " prefix	Agent Alice	Coding sprints (parallel)

Code Intelligence (Code Graph & Wiki)

MultiTerminal indexes your C# codebase so agents can reason about it structurally instead of grepping blindly.

Institutional Memory

MultiTerminal implements a host-controlled memory system that replaces the traditional agent-controlled approach to context management.

Why Host-Controlled?

Knowledge Base

The knowledge base stores reusable insights organized into 6 semantic categories:

Category	Purpose	Example
decision	Architectural and design decisions	"Chose SQLite over JSON files for project storage"
pattern	Established coding patterns	"All panels follow Initialize(broker) + ApplyTheme(isDark) pattern"
gotcha	Non-obvious pitfalls	"WebView2 init blocks UI thread if done in constructor"
anti_pattern	Things to avoid	"Never use $env:VARNAME in Bash tool - gets mangled"
debug_insight	Debugging discoveries	"SplitterMoved fires during restore, corrupting ratios"
preference	User/team preferences	"Always use PowerShell syntax in terminal commands"

Confidence Levels

Each knowledge entry has a confidence level:

• observed: Seen once, may need verification
• confirmed: Verified across multiple interactions
• deprecated: Superseded by newer knowledge

Attention Decay

Knowledge entries that are actively referenced stay prominent; stale entries fade out. Each entry tracks:

• reference_count — incremented every time the entry is returned by a search or injected at session start
• last_referenced — timestamp of most recent access

At session start, the session-status-hook.js ranks all knowledge entries using the decay formula:

score = (reference_count + 1) / (days_since_last_referenced + 1)

Injection is tiered by rank:

Tier	Rank	Injection
Full	Top 10	Title + 200-character content preview
Title-only	Next 5	Title only (use query_knowledge for details)
Available	Rest	Not injected — queryable on demand via MCP tool

This ensures frequently-used knowledge stays in context while rarely-accessed entries don't waste tokens. Old entries that are still actively referenced retain high scores.

Code Digests

Compressed summaries of codebase structure that agents can query to understand the project without reading every file. Digests capture file purposes, key classes, and relationships.

Session Lineage

Tracks the chain of Claude Code sessions for a terminal identity, enabling agents to understand the history of work done on a project across multiple sessions.

Full-Text Search

The knowledge database supports full-text search across all entries, enabling agents to query institutional memory with natural language. Searches automatically bump the reference_count and last_referenced timestamp of returned entries, feeding the attention decay system.

Settings & Theme

Dark / Light Theme

Toggle between dark and light themes using the sun/moon button in the toolbar. The theme applies to:

• Main form and toolbar
• All docking panels
• All WebView2 HTML panels (via ApplyTheme callback)
• Terminal background and text colors

Font Settings

The Settings dialog (WPF-based) allows configuring:

• Terminal font family and size
• Font changes propagate to all open terminals in real-time

Docking Layout

Panel positions, sizes, and visibility are saved automatically when the application closes and restored on next launch. The layout is stored as an XML file.

Grid Layout Presets

The Grid Layout dropdown provides preset arrangements for terminal tabs:

• 2x2, 2x3, 3x2 grids
• 2 or 3 horizontal splits
• 2 or 3 vertical splits
• Reset to Tabs (single tabbed view)

Toolbar

The main toolbar contains buttons for all panel toggles, layout controls, and settings:

New Terminal | Project | Chat | History | Help | Activity | Tasks | Profiles | Inbox | Office | Debug | File Preview | Recent | Grid Layout | Theme | Settings | About

Dialogs

Modal dialogs for specific operations:

Dialog	Purpose	Opened Via
Settings (WPF)	Configure terminal font, size, and application preferences	Toolbar gear button
New Project (WPF)	Register a new project with paths, build command, and git config	Project panel "New" button
Project Manager	Browse, edit, and manage all registered projects	Project panel
Edit Project	Edit details of a specific project	Project Manager dialog
Chat History	View historical chat messages with search and filtering	Toolbar "History" button
Session Viewer	Browse Claude Code session history and metadata	Terminal context menu
Identity Picker	Select or create a terminal identity (name) for session resumption	Terminal startup
Rename Tab	Rename a terminal tab (changes the display name)	Right-click terminal tab
Save Prompt	Save a reusable prompt to the project's prompt collection	Project panel
Owner Profile (WPF)	Manage owner identity, display name, avatar, and git configuration	Toolbar profile button
About	Application version, build info, and credits	Toolbar info button

Remote & Phone Access (Multi-Connect)

The current, recommended way to reach MultiTerminal from your phone is Multi-Connect — an in-app gateway published over your private Tailscale network. It serves a mobile PWA for task access without exposing anything to the public internet.

Mobile features include a kanban task board, terminal management, an interactive console, chat with voice input, and Web-Push notifications.

Mobile Features (Phone PWA)

Feature	Description
Task Board	View and manage kanban tasks with checklist status
Terminal Management	List running terminals and spawn new ones from your phone
Interactive Console	Full xterm.js terminal via WebSocket — type commands from your phone
Chat Messaging	Send messages to agents with voice input (Web Speech API)
Push Notifications	Native phone notifications for task updates, messages, and alerts (Web Push API + VAPID)
Live Commentary	Optional TTS commentary on agent activity via edge-tts
PWA Support	Add to home screen, works offline-first with service worker

Notifications flow from Claude Code hooks → NotificationsController → the active gateway → Web Push to subscribed phones.

Presence-Aware Routing

MultiTerminal can sense whether you're at your desk and route agent questions accordingly — to the desktop when you're present, or to phone-push when you've stepped away.