Components

The Coleo system consists of five major components that work together.

Brain (Coordinator)

The Brain is the central nervous system of Coleo.

┌─────────────────────────────────────────────────────────────┐
│                         BRAIN                                │
├─────────────────────────────────────────────────────────────┤
│  Responsibilities:                                           │
│  ├── Arm Lifecycle      - Spawn, monitor, kill arms         │
│  ├── Conflict Resolution - Mediate ownership disputes        │
│  ├── Governance         - Process proposals, enforce rules   │
│  ├── Human Interface    - Route approvals, notifications     │
│  ├── State Management   - Persist system state               │
│  └── Misbehavior Detection - Identify and stop bad actors   │
├─────────────────────────────────────────────────────────────┤
│  Powers:                                                     │
│  ├── PAUSE arm          - Temporarily halt an arm            │
│  ├── KILL arm           - Terminate destructive arm          │
│  ├── VETO proposal      - Override arm consensus             │
│  └── ESCALATE to human  - Require human decision             │
└─────────────────────────────────────────────────────────────┘

Misbehavior Detection

The Brain monitors arms for problematic behavior:

Behavior	Detection	Response
Touching files outside task scope/claims	Pattern matching on file paths vs. claims	WARN, then PAUSE
Ignoring consensus without override	Proposal tracking	WARN, reputation penalty
Destructive changes	Pattern matching (rm -rf, DROP, etc.)	KILL immediately
Resource exhaustion	Token/API call counters	PAUSE, notify human
Stuck in loop	Action repetition detection	PAUSE with exponential backoff

Loop Detection & Backoff Throttling (Design)

When an arm gets stuck in a loop (repeating the same actions, hitting the same errors, or consuming tokens without progress), the brain intervenes with an escalating backoff strategy:

interface LoopDetection {
  armId: string;
  detectedAt: Date;
  loopType: "action_repeat" | "error_loop" | "token_burn" | "thrashing";
  consecutiveLoops: number;      // How many times we've caught this arm looping
  backoffMinutes: number;        // Current pause duration
}

const BACKOFF_SCHEDULE = [
  1,    // First loop: 1 minute pause
  5,    // Second: 5 minutes
  15,   // Third: 15 minutes
  30,   // Fourth: 30 minutes
  60,   // Fifth+: 1 hour
];

Loop Response Protocol:

1. Detect: Brain notices arm repeating actions or burning tokens without progress
2. Pause: Arm is immediately paused (cannot consume more tokens)
3. Instruct: Brain sends message: "You appear stuck. Compact your session and reassess."
4. Wait: Arm remains paused for backoff duration
5. Resume: After backoff, arm is resumed with instruction to compact context and retry
6. Check Relevance: If original task is no longer relevant, arm is reassigned

interface LoopRecoveryMessage {
  type: "loop_recovery";
  armId: string;
  instruction: string;
  actions: ("compact_session" | "reassess_task" | "request_help")[];
  originalTaskStillRelevant: boolean;
  pauseDuration: number;
}

Token Budget Protection (Design):

interface TokenThrottle {
  windowMinutes: number;         // Rolling window (default: 10)
  maxTokensPerWindow: number;    // Hard cap (default: 50000)
  warningThreshold: number;      // Warn at 70%
}

Brain State

interface BrainState {
  status: "running" | "stopped" | "paused";
  startedAt: Date;
  lastPollAt: Date;
  pollIntervalMs: number;
  activeArms: string[];
  pendingProposals: number;
  pendingApprovals: number;
}

Brain Logic

The brain runs a polling cycle every 30 seconds that orchestrates arm lifecycle and task assignment. It includes event-window based arm health monitoring and automatic intervention capabilities.

Poll Cycle

flowchart TD
    classDef process fill:#e1f5fe,stroke:#01579b,stroke-width:2px
    classDef decision fill:#fff3e0,stroke:#e65100,stroke-width:2px,shape:rhombus

    subgraph POLL["Poll Cycle - Runs every 30s"]
        A[Start Poll] --> B[scanForRunningArms]
        B --> C{API Server Available?}
        C -->|Yes| D[Get Idle Arms]
        C -->|No| E[Skip API Arms]
        D --> F[promptIdleArms]
        F --> G[checkIdleArmStuckLoops]
        G --> H[End Poll]
    end

    class A,B,D,F,G,H process
    class C decision

Event-Window Based Health Monitoring

The brain continuously monitors arm health using an event-window based system that analyzes recent activity patterns to detect issues before they become critical.

Event Window Analysis

The health monitoring system fetches event windows for each arm from JetStream, grouping events by type and analyzing patterns to classify arm states:

interface ArmActivityState {
  productive: "actively doing useful work";
  idle: "waiting for work";
  waiting_permission: "blocked on a permission request";
  looping: "stuck in a repetitive pattern";
  silent: "no events for an extended period";
  error: "encountered an error state";
  starting: "in startup grace period";
}

Health Monitoring Components

1. BrainEventWindow: Centralized JetStream event window fetcher that retrieves event slices per arm
2. ArmActivityAnalyzer: Classifies arm states based on event windows using pattern recognition
3. ArmHealthMonitor: Coordinates health checks and automatic interventions

Automatic Intervention Capabilities

When issues are detected, the health monitoring system can automatically intervene:

• Prompting: Send messages to arms that appear stuck or idle
• Interrupting: Send /compact commands to arms stuck in loops
• Killing: Terminate arms that are consistently problematic
• Escalating: Notify humans for permission requests that timeout
• Recovering: Restart arms that crash or become unresponsive

Configuration Options

The health monitoring system is highly configurable:

interface HealthMonitorConfig {
  checkIntervalMs: number;      // How often to run health checks
  eventWindowMs: number;        // Size of event window to analyze
  autoInterventionEnabled: boolean; // Whether automatic interventions are enabled
  silentThresholdMs: number;    // Time before arm considered silent
  loopRepetitionThreshold: number; // Repetitions before considering looping
  permissionEscalationMs: number; // Time before escalating permission requests
}

File Reading During Poll

The brain reads markdown files during its poll cycle to sync tasks from human-editable sources:

Poll Cycle File Reading:
├── Step 8: syncPlanTasks()
│   └── Reads .project/plan.md
│   └── Extracts tasks from `- [ ]` checkbox items
│   └── Creates/updates tasks in SQLite
│
├── Step 8a: processInbox()
│   └── Reads .project/inbox.md
│   └── Parses ## headers and - [ ] items as new tasks
│   └── Deduplicates against existing tasks (by title similarity)
│   └── Clears inbox after processing
│
├── Step 8b: checkDocUpdateTrigger()
│   └── Checks if documentation needs updating
│
└── Step 8c: reEvaluatePlanProgress()
    └── Creates verification tasks for issues

Files the brain reads:

• .project/plan.md - Main plan with phases and deliverables
• .project/inbox.md - Quick task input (cleared after processing)
• **/*.plan.md - Any file ending in .plan.md
• **/plans/*.md - Files in plans/ directories

Arm State Machine

Arms follow a formal state machine with 7 states:

flowchart LR
    classDef state fill:#f3e5f5,stroke:#4a148c,stroke-width:2px

    subgraph LIFECYCLE["Arm State Machine"]
        direction LR
        disconnected["disconnected<br/>(not tracked)"] -->|scan finds process| idle["idle<br/>(ready for work)"]
        idle -->|brain assigns task| task_assigned["task_assigned<br/>(task assigned, awaiting ack)"]
        task_assigned -->|arm acknowledges| working["working<br/>(has task, processing)"]
        working -->|complete_task| idle
        idle -->|no activity after prompt| stuck["stuck detection"]
        stuck -->|productive activity detected| working
        stuck -->|confirmed stuck| intervention["intervention"]
        intervention -->|recover| idle
        working -->|process dies| stopped["stopped"]
        idle -->|process dies| stopped
        task_assigned -->|process dies| stopped
        stopped -->|restart| disconnected
    end

    class disconnected,idle,task_assigned,working,stuck,intervention,stopped state

State Transition Truth Table

From State	Event	To State	Notes
spawning	PROCESS_STARTED	starting	Process is running
starting	HARNESS_CONNECTED	idle	Harness ready
idle	TASK_ASSIGNED	task_assigned	Brain assigns task
task_assigned	TASK_ACKNOWLEDGED	working	Arm accepts task
task_assigned	TIMEOUT (3min)	idle	Arm didn't ack, release task
working	TASK_COMPLETED	idle	Task done
working	TASK_FAILED	idle	Task failed
idle/working	CONNECTION_LOST	disconnected	Network issue
disconnected	CONNECTION_RESTORED	previous	Reconnected
any	STOP	stopped	Intentional stop

Task Reordering and Management

Tasks now include a sort_order field that allows manual reordering:

POST /api/tasks/reorder
{
  "taskId": "task-123",
  "toSortOrder": 0  // 0=top, -1=bottom
}

Tasks can also be removed directly from plan.md files:

POST /api/tasks/:id/remove-from-plan

This links plan.md lines to tasks via plan_line_uid and removes both the line and the database entry.

Task Assignment Flow

sequenceDiagram
    participant Arm
    participant Brain
    participant DB[SQLite Database]
    participant State[State Machine]

    Note over Arm,State: Arm is IDLE and waiting

    Arm->>Brain: heartbeat (status: idle)
    Brain->>Arm: "You have tasks, call get_full_briefing"

    Note over Arm: Arm calls get_full_briefing

    Arm->>Brain: get_full_briefing()
    Brain->>Arm: Task context bundle

    Note over Arm: Arm decides to claim task
    Arm->>Brain: claim_task(task-id)

    Brain->>State: TASK_ASSIGNED event
    State->>State: idle → task_assigned
    Brain->>DB: UPDATE tasks SET status='claimed', assigned_to=arm
    Brain->>DB: UPDATE arms SET status='busy'

    Brain->>Arm: task_assignment message (via queue)
    Arm->>Brain: acknowledge_task(task-id)

    Brain->>State: TASK_ACKNOWLEDGED event
    State->>State: task_assigned → working
    Brain->>DB: UPDATE tasks SET status='in_progress'

    Note over Arm,State: Arm is now WORKING on the task

Key insight: The brain assigns tasks to arms in idle state, transitioning them to task_assigned. The arm must then explicitly acknowledge to transition to working. This two-step process ensures both brain and arm agree on task ownership.

Grace Period for Autonomous Arms

When the brain starts up and finds arms that were working autonomously (before the brain was running), it protects them from being interrupted:

sequenceDiagram
    participant Brain
    participant DB[SQLite Database]
    participant Arm

    Note over Brain,DB: Brain starts up - finds arm working autonomously

    Brain->>DB: scanForRunningArms()
    DB->>Brain: arm "rand" exists, status=busy
    Brain->>Arm: Check if process alive (kill(pid, 0))
    Arm-->>Brain: Process alive!
    Brain->>DB: UPDATE arms SET status='idle'
    Brain->>DB: armDetectionTimes.set("rand", now)

    Note over Brain: Grace period - no prompting for configured time

    Brain->>DB: checkIdleArmStuckLoops()
    Brain->>DB: Query productive activity in last 60 min
    DB-->>Brain: Found: complete_task at 14:30
    Brain->>Brain: lastProductiveAt = 14:30<br/>skip stuck detection<br/>arm was working autonomously!

    Note over Brain,Arm: Arm continues working without interruption

Grace period behavior:

• Arms detected during scanForRunningArms() are not prompted for a configurable grace period
• Default: 5 minutes
• Configurable via brain.arm_grace_period_minutes in config.toml
• The brain also checks for recent productive activity (heartbeat, claim_task, complete_task) before marking an arm as stuck

Stuck Loop Detection

The brain detects when arms repeatedly respond with "idle" without productive work:

1. Idle arm prompting: Brain prompts idle arms to check for work
2. Activity tracking: Recent activity is analyzed for prompt-response patterns
3. Productive actions: heartbeat, claim_task, acknowledge_task, complete_task, file_changed, tool_call
4. Escalation: If arm doesn't respond productively:
   • Level 0: Interrupt + different prompt
   • Level 1: Force context compaction
   • Level 2: Kill and respawn arm
   • Level 3: Notify human via email

---

Arms (General-Purpose Agents)

Each arm is a semi-autonomous general-purpose AI agent. Its behavior is determined by the task classification it is executing (architect, development, QA, documentation, etc.), not by a permanently assigned domain.

Arm Profile

interface ArmProfile {
  id: string;
  name: string;
  agent: "opencode-api" | "custom";

  // Task execution
  supportedClassifications: string[]; // e.g., ["architect", "development", "qa"]

  // Context Management
  contextBudget: ContextBudget;
  currentContext: ContextSnapshot;

  // Ownership
  claims: FileClaim[];         // Files/dirs this arm is tending

  // Governance
  reputation: number;          // 0-100, affects persuasion weight
  activeProposals: Proposal[];

  // State
  status: "idle" | "working" | "blocked" | "proposing" | "paused" | "dead";
  currentTask?: Task;
}

Arms can be used for different task classifications over time. The same arm might run an architect task for one assignment, then a development or QA task for the next.

MCP Server Catalog

Arms access the garden through MCP servers. Each provides specialized capabilities:

MCP Server	Purpose	Key Tools
git-mcp	Version control	commit, push, branch, diff, log
env-mcp	Environment variables	get, set, list (filtered for secrets)
docs-mcp	Library documentation	search, fetch, summarize
devtools-mcp	Browser automation	screenshot, console, network, lighthouse
deploy-mcp	Deployment operations	request, status, rollback, logs
db-mcp	Database operations	query, migrate, seed, backup
pkg-mcp	Package management	install, update, audit, outdated
observability-mcp	Logs & metrics	logs.search, metrics.query, traces.find
alerts-mcp	Alert management	list, ack, silence, escalate

Observability MCP Server

For production operations, arms need visibility into running systems:

interface ObservabilityMCP {
  // Log operations
  "logs.search": (params: {
    service: string;
    environment: string;
    query: string;
    timeRange: { start: Date; end: Date };
    limit?: number;
  }) => LogEntry[];
  
  "logs.tail": (params: {
    service: string;
    environment: string;
    follow: boolean;
  }) => AsyncIterable<LogEntry>;
  
  // Metrics
  "metrics.query": (params: {
    query: string;              // PromQL or similar
    timeRange: { start: Date; end: Date };
    step?: string;             // e.g., "1m", "5m"
  }) => MetricSeries[];
  
  "metrics.dashboard": (params: {
    name: string;
  }) => DashboardSnapshot;
  
  // Traces (distributed tracing)
  "traces.find": (params: {
    service?: string;
    traceId?: string;
    minDuration?: number;
    error?: boolean;
    limit?: number;
  }) => Trace[];
  
  // Health
  "health.check": (params: {
    service: string;
    environment: string;
  }) => HealthStatus;
}

interface LogEntry {
  timestamp: Date;
  level: "debug" | "info" | "warn" | "error";
  service: string;
  message: string;
  metadata: Record<string, unknown>;
}

interface HealthStatus {
  status: "healthy" | "degraded" | "unhealthy";
  checks: { name: string; status: string; message?: string }[];
  lastCheck: Date;
}

Alerts MCP Server

For incident response and on-call operations:

interface AlertsMCP {
  "alerts.list": (params: {
    status?: "firing" | "resolved" | "silenced";
    severity?: "critical" | "warning" | "info";
    service?: string;
  }) => Alert[];
  
  "alerts.ack": (params: {
    alertId: string;
    message: string;
  }) => void;
  
  "alerts.silence": (params: {
    matchers: { label: string; value: string }[];
    duration: string;         // e.g., "2h", "1d"
    reason: string;
  }) => Silence;
  
  "alerts.escalate": (params: {
    alertId: string;
    reason: string;
  }) => void;
  
  "runbook.fetch": (params: {
    alertName: string;
  }) => Runbook;
}

interface Alert {
  id: string;
  name: string;
  severity: "critical" | "warning" | "info";
  status: "firing" | "resolved" | "silenced";
  service: string;
  summary: string;
  firedAt: Date;
  resolvedAt?: Date;
  labels: Record<string, string>;
}

Legacy: Arm Domain Definition

interface ArmDomain {
  name: string;
  description: string;
  defaultPatterns: string[];   // Glob patterns for auto-claiming
  mcpServers: string[];        // Which MCP servers this arm can use
}

// Note: ArmDomain reflects an earlier design that relied on static domains.
// In the current design, arms are general-purpose and behavior is primarily
// guided by task classifications, task history, and configuration templates.

---

Garden (Shared Environment)

The Garden is the workspace that arms tend. It's represented as a 3D space for visualization.

Layers

┌─────────────────────────────────────────────────────────────┐
│                       THE GARDEN                             │
├─────────────────────────────────────────────────────────────┤
│  Physical Layer (files, dirs, repos):                        │
│  ├── Source code                                             │
│  ├── Configuration                                           │
│  ├── Documentation                                           │
│  ├── Tests                                                   │
│  └── Build artifacts                                         │
├─────────────────────────────────────────────────────────────┤
│  Logical Layer (exposed via MCP):                            │
│  ├── git-mcp        - VCS operations                         │
│  ├── env-mcp        - Environment variables                  │
│  ├── runtime-mcp    - Node/Python/Bun versions               │
│  ├── docs-mcp       - Library documentation                  │
│  ├── nx-mcp         - Monorepo orchestration                 │
│  ├── devtools-mcp   - Browser automation                     │
│  ├── deploy-mcp     - Deployment per environment             │
│  └── pkg-mcp        - Package management                     │
├─────────────────────────────────────────────────────────────┤
│  Ownership Layer:                                            │
│  ├── Who owns what (claims)                                  │
│  ├── Who touched what recently (activity)                    │
│  └── Conflict zones (multiple claims)                        │
└─────────────────────────────────────────────────────────────┘

3D Coordinate System (Radial)

Files are positioned in a radial 3D space where distance from center indicates activity - frequently touched files appear closer to the center, making the visualization naturally focus attention on what's actively being worked on.

interface GardenCoordinate {
  // Radial coordinates
  category: number;    // Angle in degrees (0-360) - which "slice" of the pie
  activity: number;    // Distance from center (0-100) - 0=hot, 100=cold
  depth: number;       // Vertical position (0-100) - stack layer
}

interface GardenNode {
  path: string;
  type: "file" | "directory";
  coords: GardenCoordinate;
  owner?: string;           // Arm ID
  lastTouchedBy?: string;   // Arm ID
  lastTouchedAt?: Date;
  conflictZone: boolean;
}

Axis	Heuristic	Meaning
Category (angle)	File type/category	Each file category gets a slice: UI (0-60°), API (60-120°), DB (120-180°), Infra (180-240°), Tests (240-300°), Docs (300-360°)
Activity (radius)	Recency & frequency	0=center=very active, 100=edge=dormant. Based on touches in last 7 days
Depth (vertical)	Stack layer	0=frontend/surface, 100=infrastructure/deep

Activity Calculation

function calculateActivity(path: string, touches: Touch[]): number {
  const now = Date.now();
  const weekMs = 7 * 24 * 60 * 60 * 1000;
  
  // Only consider touches in last 7 days
  const recentTouches = touches.filter(t => 
    now - t.timestamp.getTime() < weekMs
  );
  
  if (recentTouches.length === 0) return 100; // Edge - dormant
  
  // Score based on recency and frequency
  let score = 0;
  for (const touch of recentTouches) {
    const ageMs = now - touch.timestamp.getTime();
    const recencyWeight = 1 - (ageMs / weekMs); // 1.0 for now, 0.0 for week ago
    score += recencyWeight;
  }
  
  // Normalize: more touches + more recent = closer to center
  // Cap at 20 touches for max activity
  const normalized = Math.min(score / 20, 1);
  return Math.round((1 - normalized) * 100); // Invert: 0=active, 100=dormant
}

Visual Effect

In the 3D Garden view:

• Center cluster: Files being actively worked on right now
• Middle ring: Recently touched, part of ongoing work
• Outer ring: Stable files, not recently modified
• Colored by owner: Each arm has a color, files glow with owner's color
• Pulsing: Files with active claims pulse gently
• Conflict zones: Red highlight when multiple arms are contending

---

Observatory (Web UI + API)

The Observatory is how humans observe and control the system.

Server Components

┌─────────────────────────────────────────────────────────────┐
│                      OBSERVATORY                             │
├─────────────────────────────────────────────────────────────┤
│  Web Server (Hono):                                          │
│  ├── REST API          - CRUD operations, queries            │
│  ├── WebSocket         - Real-time updates                   │
│  ├── Static files      - React SPA                           │
│  └── Push endpoint     - Browser notifications               │
└─────────────────────────────────────────────────────────────┘

UI Views

View	Purpose
Dashboard	System overview, arm status at a glance
Garden	3D visualization of workspace with ownership
Arms	Arm details, context, activity log
Proposals	Active debates, arguments, signals
Approvals	Pending human decisions
Activity	Timeline of all system actions
Config	System settings, arm configuration

Available Actions

• Spawn/Kill arms
• Approve/Reject proposals
• Override arm decisions
• Configure context budgets
• Reassign file ownership
• Trigger deployments

---

Model Fallback System

The model fallback system ensures arms can spawn successfully even when configured with unavailable models by automatically resolving to alternative models based on availability and cost.

Model Resolution Process

1. Exact Match: Try the requested provider/model combination
2. Provider Fallback: If model not found, find cheapest model from same provider
3. Cross-Provider: If provider not available, find same model from another provider
4. Last Resort: Use first available model from any connected provider

Model Pricing Information

The resolver includes known model pricing to make cost-based decisions when multiple fallback options are available:

interface ModelPricing {
  input: number;   // Price per million input tokens
  output: number;  // Price per million output tokens
}

// Examples of known model pricing:
// Claude Sonnet 4: { input: 3, output: 15 }
// GPT-4o: { input: 2.5, output: 10 }
// Gemini 2.5 Flash: { input: 0.075, output: 0.3 }

Model Resolution Response

interface ResolvedModel {
  providerId: string;
  modelId: string;
  providerName: string;
  modelName: string;
  fallback: boolean;
  fallbackReason?: string;
}

When a fallback occurs, the system logs the reason for debugging and transparency.

Provider Availability Checking

The system can validate model availability before spawning arms:

async function isModelAvailable(
  providerId: string,
  modelId: string,
  apiUrl: string
): Promise<boolean>

Cost-Based Model Selection

For scenarios where cost optimization is important, the system can provide a list of all available models sorted by cost:

async function getAvailableModelsByCost(
  apiUrl: string
): Promise<Array<{ providerId: string; modelId: string; cost: number }>>

---

Nerve System (Communication Layer)

All communication flows through the Nerve System using NATS for distributed messaging and an internal queue for brain↔arm messages.

Message Flow

Human ◄──────► Observatory ◄──────► Brain ◄──────► Arms
         │                    │              │
         │ WebSocket          │ NATS         │ NATS/Queue
         │ Push Notifications │ JetStream    │ MCP Tools
         │ REST API           │              │

NATS Event Types (Arm Lifecycle)

Events published via NATS for distributed arm management:

Event Type	Direction	Description
arm.spawned	Agent → All	Arm process started
arm.killed	Agent → All	Arm process terminated
arm.recovered	Agent → All	Arm recovered from error
arm.status_changed	Agent → All	Arm status transition (idle→busy, etc.)
arm.activity	Agent → Brain	Arm performed an action
arm.log	Agent → Brain	Log message from arm
agent.connected	Agent → All	Agent host came online
agent.disconnected	Agent → All	Agent host went offline

Queue Message Types (Brain ↔ Arm)

Messages passed through the brain's message queue:

Message Type	Direction	Description
task_assignment	Brain → Arm	Assign task to arm
task_complete	Arm → Brain	Task finished successfully
task_failed	Arm → Brain	Task failed with error
discovery	Arm → Brain	Arm found something noteworthy
dependency_discovery	Arm → Brain	Arm found a dependency
status_report	Arm → Brain	Progress report on current task
status_update	Arm → Brain	General status update
heartbeat	Arm → Brain	Arm is alive and working
approval_request	Arm → Brain	Arm needs approval for action
approval_response	Brain → Arm	Approval granted/denied
human_message	Human → Brain	Message from human
share_note	Arm → All	Note shared between arms
tool_discovery	Arm → Brain	Arm found useful tool
doc_update	Brain → Arm	Documentation needs updating
file_subscription	Arm → Brain	Arm watching a file
file_change	Brain → Arm	Watched file changed
claim_transfer	Brain → Arm	File claim transferred
bug_report	Arm → Brain	Bug discovered
bug_assignment	Brain → Arm	Bug assigned for fixing
context_compression	Arm → Brain	Context was compressed
dev_server_restart_request	Arm → Brain	Request to restart dev server

Activity Event Types (JetStream)

Events tracked in JetStream for activity analysis:

Category	Event Types
Session	session.status, session.idle, session.error, session.updated, session.diff
Message	message.updated, message.removed, message.part.updated, message.part.removed
Permission	permission.asked, permission.replied
Todo	todo.updated
File	file.edited, file.watcher.updated
Command	command.executed
Arm Lifecycle	arm.spawned, arm.status_changed, arm.heartbeat, arm.killed, arm.stopped
Task	task.created, task.assigned, task.claimed, task.completed, task.blocked, task.failed
Brain	arm_prompted, event-status, started, stopped