Features Overview

Flightdeck is a multi-agent orchestration platform. This page covers everything it can do, organized by what matters to you as a user.

Crew Orchestration

At its core, Flightdeck turns a single task into a coordinated team effort.

Project Lead

Every session starts with a Project Lead agent. You describe what you want built, and the lead:

• Analyzes your task and breaks it into subtasks
• Creates a task DAG (directed acyclic graph) with dependencies
• Assembles a crew of specialist agents
• Delegates work, coordinates progress, and synthesizes results

14 Specialized Roles

Each agent has a role with a purpose-built system prompt and recommended model:

Role	Focus
Project Lead 👑	Orchestration, delegation, crew coordination
Developer 💻	Code implementation, tests, bug fixes
Architect 🏗️	System design, architecture decisions. Can also delegate.
Code Reviewer 📖	Readability, maintainability, code patterns
Critical Reviewer 🛡️	Security, performance, edge cases
Readability Reviewer 📐	Naming, organization, documentation quality
Product Manager 🎯	User needs, product quality, UX review
Technical Writer 📝	Documentation, API design, developer experience
Designer 🎨	UI/UX, interaction design, accessibility
QA Tester 🧪	Test strategy, quality assurance, coverage
Generalist 🔧	Cross-disciplinary problem solving
Radical Thinker 🚀	Challenge assumptions, propose bold alternatives
Secretary 📋	Plan tracking, status reports, session summaries
Agent ⚙️	Neutral general-purpose agent

Custom roles can be created via Settings → Roles with your own system prompts, colors, and icons.

Task DAG

The lead creates a dependency graph that controls execution order. Tasks flow through states: pending → ready → running → done (or failed, paused, skipped). The EagerScheduler automatically assigns ready tasks to idle agents as soon as dependencies resolve.

Three views in the dashboard:

• Graph view — Interactive dependency graph (ReactFlow) with status colors and critical path highlighting
• Kanban board — Drag-and-drop columns by status with context menus
• Gantt chart — Timeline view with dependencies

→ Auto-DAG Guide

Parallel Execution

Multiple agents work simultaneously — a developer writes code while a reviewer checks it, an architect designs the system, and a secretary tracks progress. Up to 50 concurrent agents (configurable).

Multi-Provider Support

Flightdeck is provider-agnostic. All agents communicate through the Agent Client Protocol (ACP) — a standardized message-passing interface that abstracts away provider differences — so you can mix and match providers in the same crew.

Supported Providers

Provider	Binary	Auth
GitHub Copilot 🐙	copilot	GitHub CLI (gh auth login)
Claude Code 🟠	claude-agent-acp	ANTHROPIC_API_KEY
Google Gemini CLI 💎	gemini	GEMINI_API_KEY
Codex 🤖	codex-acp	OPENAI_API_KEY
Cursor ↗️	agent	CURSOR_API_KEY
OpenCode 🔓	opencode	(manages own keys)
Kimi CLI 🌙	kimi	Kimi auth
Qwen Code 🔮	qwen	Qwen auth

Model Resolution

Each role has a default model, but you can override per-role or per-agent. The ModelResolver handles cross-provider translation — request claude-opus-4.6 from a Gemini agent and it maps to the closest equivalent. Standard/fast/premium tier aliases work across all providers.

Provider Ranking

Set a preference order for providers. If the default is unavailable, Flightdeck falls through to the next in the ranking. Configure via flightdeck.config.yaml or Settings → Providers in the dashboard.

→ Provider Guide

Real-Time Dashboard

The web dashboard is your control center, updating in real-time over WebSocket.

Home Dashboard

At-a-glance view across all projects:

• Action Required — Pending decisions and permission requests
• Active Work — What agents are doing right now, grouped by project
• Decisions Made — Recent decisions for awareness
• Recent Activity — Latest events across all projects
• Progress — Per-project DAG summaries

Lead Dashboard

The main working interface for an active session:

• Chat panel — Send messages to the Project Lead (queue or interrupt mode)
• Decision panel — Approve, reject, or dismiss pending decisions with optional comments
• Sidebar tabs — Seven tabs (reorderable): Crew, Comms, Groups, DAG, Models, Costs, Timers
• Catch-up banner — "While you were away" summary of what happened

Timeline

Swim-lane Gantt chart showing agent activity over time:

• One lane per agent with color-coded events
• Communication links between agents
• Zoom, pan, and keyboard navigation (←→ pan, +/- zoom)
• Session replay scrubber for reviewing past sessions
• Live auto-scroll mode during active sessions

→ Timeline Guide

Analytics

Session-level analytics and cost tracking:

• Token usage trends — Input/output tokens over time
• Cost breakdowns — Per-agent and per-task token attribution from provider data
• Session comparison — Side-by-side comparison of two sessions
• Auto-generated insights — Efficiency observations

Overview Dashboard

Per-project overview with:

• Quick status bar (running/stopped, agent count, task progress, duration)
• Attention items (failed agents, blocked tasks, pending decisions)
• Decision feed and activity feed
• Session history

Communication

Agents communicate through structured messaging channels.

Direct Messaging

Agents message each other by ID. Messages can be:

• Queued — Delivered when the recipient is ready (non-blocking)
• Interrupt — Breaks into the agent's current work immediately (priority)

Group Chat

Create groups by member ID or role. Groups are auto-created when 3+ agents work on the same feature and auto-archived when all members finish. The dashboard shows a dedicated Groups tab for following multi-agent conversations.

Broadcasts

Send a message to every active agent at once — useful for announcing decisions or sharing context that affects the whole crew.

@Mentions

Type @ in the chat to autocomplete agent names. Mentioned agents receive the message.

→ Agent Communication · Chat Groups

Coordination & Safety

These features prevent agents from stepping on each other and keep you in control.

File Locking

Before editing a file, an agent must acquire a lock. Locks have:

• TTL — Expired locks are automatically cleaned up (no deadlocks from crashed agents)
• Glob support — Lock src/auth/* to claim an entire directory
• Conflict detection — Overlapping lock requests are rejected with clear error messages

Scoped Commits

When an agent commits, git add only stages files the agent has locked. Post-commit verification confirms the right files landed. This prevents git add -A from leaking other agents' uncommitted work.

Trust Dial (Oversight System)

Three levels of human oversight:

Level	Behavior
Supervised	Agents explain reasoning before acting. Significant actions require approval.
Balanced	Key decisions need approval; routine work proceeds automatically.
Autonomous	Agents work independently. Only critical failures require intervention.

Set globally or per-project. The AttentionBar at the top of the dashboard shows the current level with an escalation indicator (green/yellow/red).

→ Oversight Guide

Governance Pipeline Future Feature

The governance infrastructure exists in the codebase but is currently disabled (enabled: false). It will be activated in a future release.

The governance system is designed to route every agent command through ordered hooks before execution:

1. Security — Blocked patterns, path traversal checks
2. Permission — Role-based access control
3. Validation — Payload schema validation
4. Rate Limiting — Per-command throttling
5. Policy — Custom policy rules
6. Approval — Human approval gates (when configured)

Post-hooks will handle audit logging and metrics collection.

Decision Queue

When agents face architectural choices or need permission, they surface decisions to you:

• Review decisions in the Approval Queue (Shift+A shortcut)
• Approve with optional comment, reject with reason, or dismiss
• Auto-deny timer pauses while you're reviewing (no missed decisions)
• Decisions are categorized automatically (architecture, style, testing, etc.)

Security

• Prompt injection sanitization — 4-layer defense at write boundary
• Secret redaction — 12 regex pattern categories (AWS, GitHub, API keys, etc.) redacted from WS broadcasts, DB writes, logs
• CORS lockdown and rate limiting
• Path traversal validation for CWD and file operations
• Challenge-response auth for integrations

→ Coordination Guide

Knowledge & Persistence In Development

What agents learn carries across sessions. The core knowledge infrastructure exists and is functional — the system design is being refined for the best developer experience.

Knowledge Base

Four-category knowledge system:

Category	Purpose	Example
Core 🛡️	Project rules, identity	"Use factory pattern for services"
Procedural 🔧	Patterns, corrections, how-to	"Always run lint before commit"
Semantic 🗄️	Architecture, facts, relationships	"Auth module uses JWT with bcrypt"
Episodic 📜	Session summaries, recent events	"Session 5 refactored the API layer"

Browse, search (fuzzy search via Fuse.js), and manage entries from the Knowledge page in the dashboard.

Knowledge Injection

On agent spawn, the KnowledgeInjector automatically injects relevant knowledge into the agent's context:

• Token-budgeted to avoid context overflow
• Priority: Core (always) > Procedural > Semantic > Episodic
• Sanitized against prompt injection

Skills System

Drop Markdown files in .github/skills/ and they're hot-reloaded into agent prompts:

.github/skills/
├── testing-conventions/
│   └── SKILL.md
├── api-patterns/
│   └── SKILL.md
└── error-handling/
    └── SKILL.md

Each skill has YAML frontmatter with name and description to control when it's loaded.

Collective Memory

Cross-session knowledge that compounds over time. Agents can remember() facts and recall() them in future sessions. Knowledge is automatically extracted from completed sessions.

Session Management

• Session resume — Resume from a previous session ID with full context recovery
• Session replay — Scrub through past sessions with a timeline scrubber. Keyframes capture agent state, messages, and DAG changes. Adjustable playback speed. Shareable via tokenized links.
• Session history — Browse past sessions with metadata (duration, tasks, token usage)
• Persistent projects — Projects survive across sessions. Chat history, state, and knowledge auto-load on startup.

→ Session Management

Monitoring & Notifications

AttentionBar

Persistent status bar at the top of every page with three escalation states:

• 🟢 Green — All clear
• 🟡 Yellow — Needs attention (pending decisions, context pressure)
• 🔴 Red — Action required (failed agents, blocked tasks)

Updates via WebSocket push (<3s latency). Sensitivity adjusts with the Trust Dial level.

PulseStrip

Compact horizontal strip showing real-time crew health: active agents, token usage, context window pressure per agent, and pending decision count.

Notification Channels

Configure alerts for:

• Desktop notifications — Browser notifications with sound/preview options
• Telegram — Bot integration with batched delivery and challenge-response auth
• Slack — Webhooks with optional thread-per-session
• Discord — Webhooks with optional thread-per-session

Configurable quiet hours (timezone-aware) and per-event-type preferences.

→ Telegram Integration

Navigation & Productivity

Command Palette

Press ⌘K (Mac) or Ctrl+K (Windows/Linux) to:

• Navigate to any page, agent, or setting
• Search across entities
• Execute natural language commands (27 NL commands across 4 categories — no LLM required)

→ Command Palette Guide

Keyboard Shortcuts

Shortcut	Action
⌘K / Ctrl+K	Command palette
Shift+A	Approval queue
Alt+1–5	Switch project tabs
Escape	Close modals/panels

Global Search

Search across messages, tasks, decisions, and activity from the search dialog (⌘+Shift+K).

Historical Data

All pages load from the REST API when no live agents are present — no empty states for existing projects. You can browse any past project's timeline, tasks, decisions, and analytics even when nothing is running.

Onboarding

Contextual Coach

Behavior-triggered tips appear as toasts when specific conditions are met — for example, after your first approval, after repeated manual approvals (suggesting you adjust the Trust Dial), or on first agent crash. Tips appear once per trigger and are tracked in localStorage.

QuickStart Templates

Pre-configured project templates let new users launch their first crew in seconds. Five built-in templates cover common workflows:

• Code Review — Lead + 2 developers + reviewer
• Bug Fix — Architect + 2 developers + QA
• Quick Fix — Lead + developer (recommended for first-time users)
• Docs Blitz — Writer + developer
• Full Build — Lead + architect + 3 developers + reviewer + QA

Spotlight Tour

A 6-step interactive guided tour that highlights key UI elements — the pulse strip, agent cards, approval queue, chat panel, sidebar, and command palette. Each step spotlights one element with a tooltip explanation. Progress is tracked and the tour only appears once.

Progressive Route Disclosure

The sidebar starts minimal and reveals more navigation items as you gain experience:

Tier	Visible Routes	Unlocked When
Starter	Lead, Overview, Crews, Settings	Always
Active	+ Tasks, Timeline	2+ agents spawned or any tasks created
Collaboration	+ Mission Control	3+ agents in a session
Power	+ Analytics, Groups, Org Chart, Database	3+ sessions or manual expansion

All routes remain accessible via the command palette and direct URLs regardless of tier.

Data Management

Purge old session data from Settings → Data Management. Preview shows exact record counts before deletion. Configurable retention period (7 days to 1 year).

→ Settings Guide · Data Management