Ensemble Control API

The Ensemble Control API adds HTTP-based and WebSocket run management to the live dashboard. External systems (CI pipelines, orchestrators, custom UIs) can submit ensemble runs, query their status, and discover available capabilities — all without compiling Java code.

This guide covers Phases 1 and 2:

• Phase 1: Level 1 run submission (template + input variables), state queries, capabilities discovery.
• Phase 2: Task naming (Task.name), Level 2 per-task overrides, Level 3 dynamic task creation, WebSocket run_request message.

Phases 3-5 add run control (cancel, model switch), SSE streaming, and REST review decisions.

Overview

POST /api/runs          Submit a run with input variables
GET  /api/runs          List recent runs (filterable by status, tag)
GET  /api/runs/{runId}  Get full run detail (status, task outputs, metrics)
GET  /api/capabilities  List registered tools, models, preconfigured tasks

All endpoints are on the same Javalin server as the WebSocket dashboard — no new port or process required.

Setup

1. Configure the dashboard

Add toolCatalog, modelCatalog, and (optionally) concurrency limits to WebDashboard.builder():

ToolCatalog tools = ToolCatalog.builder()
    .tool("web_search", webSearchTool)
    .tool("calculator", calculatorTool)
    .build();

ModelCatalog models = ModelCatalog.builder()
    .model("sonnet", claudeSonnetModel)
    .model("haiku", claudeHaikuModel)
    .build();

WebDashboard dashboard = WebDashboard.builder()
    .port(7329)
    .toolCatalog(tools)
    .modelCatalog(models)
    .maxConcurrentRuns(5)          // default: 5
    .maxRetainedCompletedRuns(100) // default: 100
    .build();

2. Wire the ensemble

The template ensemble — whose tasks and model define what can run — is wired via Ensemble.builder().webDashboard(dashboard):

Ensemble.builder()
    .chatLanguageModel(claudeSonnetModel)
    .webDashboard(dashboard)
    .task(Task.builder()
        .description("Research {topic} focusing on recent developments in {year}")
        .tools(webSearchTool)
        .build())
    .task(Task.builder()
        .description("Write a concise executive summary of the research")
        .build())
    .build()
    .start(7329);

After start() (long-running mode) or run() (one-shot), the REST endpoints are available.

Submitting a run

Level 1: Template inputs

Send the input variables to substitute into {placeholder} patterns in task descriptions:

POST /api/runs
Content-Type: application/json

{
  "inputs": {
    "topic": "AI safety",
    "year": "2025"
  },
  "tags": {
    "triggeredBy": "ci-pipeline",
    "environment": "staging"
  }
}

Response (202 Accepted):

{
  "runId": "run-7f3a2b",
  "status": "ACCEPTED",
  "tasks": 2,
  "workflow": "SEQUENTIAL"
}

tags are arbitrary key-value metadata attached to the run for filtering and querying.

Empty body

An empty body (or {}) submits the template ensemble with no input substitution:

POST /api/runs
Content-Type: application/json
{}

Error responses

HTTP	Error code	Cause
400	BAD_REQUEST	Invalid or unparseable JSON body
429	CONCURRENCY_LIMIT	maxConcurrentRuns reached; includes retryAfterMs
503	NOT_CONFIGURED	No template ensemble wired via webDashboard()

Querying runs

List all runs

GET /api/runs

{
  "runs": [
    {
      "runId": "run-7f3a2b",
      "status": "RUNNING",
      "startedAt": "2025-03-15T10:30:00Z",
      "durationMs": null,
      "taskCount": 2,
      "completedTasks": 1,
      "workflow": "SEQUENTIAL",
      "tags": { "triggeredBy": "ci-pipeline" }
    }
  ],
  "total": 1
}

Filter by status

GET /api/runs?status=COMPLETED
GET /api/runs?status=RUNNING
GET /api/runs?status=FAILED

Valid statuses: ACCEPTED, RUNNING, COMPLETED, FAILED, CANCELLED, REJECTED.

Filter by tag

GET /api/runs?tag=triggeredBy:ci-pipeline
GET /api/runs?tag=environment:staging

Get run detail

GET /api/runs/{runId}

{
  "runId": "run-7f3a2b",
  "status": "COMPLETED",
  "startedAt": "2025-03-15T10:30:00Z",
  "completedAt": "2025-03-15T10:30:11Z",
  "durationMs": 11340,
  "workflow": "SEQUENTIAL",
  "inputs": { "topic": "AI safety", "year": "2025" },
  "tags": { "triggeredBy": "ci-pipeline" },
  "tasks": [
    {
      "taskDescription": "Research AI safety focusing on recent developments in 2025",
      "output": "## AI Safety Overview\n...",
      "durationMs": 8200,
      "tokenCount": 10150,
      "toolCallCount": 5
    },
    {
      "taskDescription": "Write a concise executive summary of the research",
      "output": "# Executive Summary\n...",
      "durationMs": 3100,
      "tokenCount": 5080,
      "toolCallCount": 0
    }
  ],
  "metrics": {
    "totalTokens": 15230,
    "totalToolCalls": 5
  }
}

HTTP	Error code	Cause
404	RUN_NOT_FOUND	Unknown run ID or run was evicted

Querying capabilities

GET /api/capabilities

Returns registered tools, models, and the preconfigured tasks from the template ensemble:

{
  "tools": [
    { "name": "web_search", "description": "Search the web using Google" },
    { "name": "calculator", "description": "Evaluate mathematical expressions" }
  ],
  "models": [
    { "alias": "sonnet", "provider": "anthropic" },
    { "alias": "haiku", "provider": "anthropic" }
  ],
  "preconfiguredTasks": [
    { "description": "Research {topic} focusing on recent developments in {year}" },
    { "description": "Write a concise executive summary of the research" }
  ]
}

If no catalog or template ensemble is configured, each list is empty ([]).

Concurrency

maxConcurrentRuns limits how many runs execute simultaneously. When the limit is reached, new submissions return HTTP 429 immediately. The retryAfterMs field in the response body gives a rough hint for when to retry.

WebDashboard.builder()
    .maxConcurrentRuns(10)         // allow 10 parallel runs
    .maxRetainedCompletedRuns(500) // keep 500 completed runs in memory
    .build();

Completed, failed, and cancelled runs are retained in memory for querying until the maxRetainedCompletedRuns cap is reached, at which point the oldest are evicted. Active runs (ACCEPTED and RUNNING) are never evicted.

Live events

While a run executes, the existing WebSocket dashboard continues to stream all events (task_started, tool_called, token, llm_iteration_started, etc.) to connected browsers. This is unchanged — the REST Control API and WebSocket dashboard work together transparently.

---

Task naming (Phase 2)

Give tasks an optional logical name so they can be referenced in Level 2 overrides and Level 3 context expressions:

Task.builder()
    .name("researcher")            // new -- optional but recommended for API use
    .description("Research {topic} developments in {year}")
    .expectedOutput("A detailed report")
    .tools(webSearchTool)
    .build()

• Names must be non-blank when set.
• GET /api/capabilities returns task names alongside descriptions.
• Level 2 override keys match by exact name first, then by description prefix (first 50 chars, case-insensitive).
• Level 3 context references use $name syntax.

---

Level 2: Per-task overrides (Phase 2)

Override specific fields of the template ensemble’s tasks at runtime — no recompilation needed:

POST /api/runs
{
  "inputs": { "topic": "AI safety" },
  "taskOverrides": {
    "researcher": {
      "description": "Research {topic} focusing on EU AI Act compliance",
      "expectedOutput": "A regulatory analysis report with citations",
      "model": "sonnet",
      "maxIterations": 15,
      "additionalContext": "The EU AI Act was formally adopted in March 2024.",
      "tools": {
        "add": ["web_search"],
        "remove": ["calculator"]
      }
    }
  }
}

The override key ("researcher") is matched against the template ensemble’s task names. If no task with that name exists, the request is rejected with 400.

Supported override fields:

Field	Type	Semantics
description	string	Replaces task description. Supports {variable} templates.
expectedOutput	string	Replaces expected output.
model	string	Model alias resolved from ModelCatalog. Replaces chatLanguageModel.
maxIterations	int	Replaces max tool-call iteration count.
additionalContext	string	Appended to the task description (not replacing).
tools.add	string[]	Tool names resolved from ToolCatalog and added to the task’s tool list.
tools.remove	string[]	Tool names resolved from ToolCatalog and removed from the task’s tool list.

Task matching rules:

1. Exact name match (case-insensitive).
2. Fallback: first 50 characters of task description compared case-insensitively.

The original task objects are never mutated — Task.toBuilder() creates modified copies.

---

Level 3: Dynamic task creation (Phase 2)

Define an entirely new task list at runtime without changing any Java code:

POST /api/runs
{
  "tasks": [
    {
      "name": "researcher",
      "description": "Research the competitive landscape for {product}",
      "expectedOutput": "A competitive analysis identifying 5 key competitors",
      "tools": ["web_search"],
      "model": "sonnet",
      "maxIterations": 20
    },
    {
      "name": "writer",
      "description": "Write an executive brief based on the research",
      "expectedOutput": "A 1-page executive summary suitable for C-suite",
      "context": ["$researcher"],
      "model": "sonnet"
    }
  ],
  "inputs": { "product": "AgentEnsemble" }
}

When tasks is provided, the template ensemble’s task list is replaced with the dynamic list. The template’s model, catalogs, and other settings (rate limits, workflow, etc.) are preserved.

Task definition fields:

Field	Required	Description
name	No	Logical name. Used for $name context references.
description	Yes	What the agent should do. Supports {variable} templates.
expectedOutput	No	Expected output. Defaults to Task.DEFAULT_EXPECTED_OUTPUT.
tools	No	Tool names resolved from ToolCatalog.
model	No	Model alias resolved from ModelCatalog. Falls back to template default.
maxIterations	No	Max ReAct loop iterations. Default: 25.
context	No	$name or $N (0-based index) references to predecessor tasks.
outputSchema	No	JSON Schema injected as structured output instructions into expectedOutput.
additionalContext	No	Extra text appended to the task description.

Context DAG (context field):

"context": ["$researcher"]   // by name
"context": ["$0"]            // by index (0-based position in the tasks array)

• Circular dependencies are detected and rejected (400).
• Unknown names or out-of-bounds indices are rejected (400).
• When context dependencies exist and no explicit workflow is set, PARALLEL (DAG-based) is automatically inferred.

---

WebSocket run submission (Phase 2)

As an alternative to REST, WebSocket clients can submit runs using the run_request message. This lets browser-based UIs and long-lived WS clients kick off runs without an additional HTTP round-trip.

Client sends:

{
  "type": "run_request",
  "requestId": "req-1",
  "inputs": { "topic": "AI safety" },
  "tasks": [...],
  "taskOverrides": {...},
  "options": { "maxToolOutputLength": 5000 },
  "tags": { "env": "staging" }
}

All fields except type are optional. Level detection: if tasks is present, Level 3 is used; else if taskOverrides is present, Level 2 is used; otherwise Level 1.

Server responds immediately (run_ack):

{
  "type": "run_ack",
  "requestId": "req-1",
  "runId": "run-7f3a2b",
  "status": "ACCEPTED",
  "tasks": 2,
  "workflow": "SEQUENTIAL"
}

On completion, the server sends run_result to the originating session only:

{
  "type": "run_result",
  "runId": "run-7f3a2b",
  "status": "COMPLETED",
  "outputs": [
    { "taskName": "researcher", "output": "...", "durationMs": 8200 },
    { "taskName": "writer", "output": "...", "durationMs": 3100 }
  ],
  "durationMs": 11340,
  "metrics": { "totalTokens": 15230, "totalToolCalls": 7 }
}

The run_result is targeted to the submitting session only. The existing ensemble_completed broadcast continues to go to all connected clients (backwards compatible).

When no template ensemble is configured (setEnsemble() has not been called), the server immediately responds with a run_ack carrying status: "REJECTED".

---

---

Phase 3: Run Control (Cancel + Model Switching)

Cancel a run

Cancel a running or accepted run at the next task boundary (cooperative cancellation). The current in-flight task completes normally; cancellation takes effect before the next task.

REST:

POST /api/runs/{runId}/cancel

Response:

{ "runId": "run-abc", "status": "CANCELLING" }

Possible status values: CANCELLING, REJECTED (already terminal), NOT_FOUND.

WebSocket:

{ "type": "run_control", "runId": "run-abc", "action": "cancel" }

Response: run_control_ack message.

Switch active model mid-run

Switch the LLM that subsequent tasks will use, without restarting the ensemble. Takes effect on the next LLM call; the current in-flight call completes with the previous model.

REST:

POST /api/runs/{runId}/model
Content-Type: application/json

{ "model": "haiku" }

Response:

{ "runId": "run-abc", "model": "haiku", "status": "APPLIED" }

The model field must be a registered alias in the configured ModelCatalog.

WebSocket:

{ "type": "run_control", "runId": "run-abc", "action": "switch_model", "model": "haiku" }

---

Phase 4: Event Subscription Filtering + SSE

Subscribe to a filtered event stream (WebSocket)

By default all events are broadcast to all connected sessions. Use subscribe to receive only the events you care about:

{ "type": "subscribe", "events": ["task_started", "task_completed", "run_result"] }

Optional run filter (only events from the specified run):

{ "type": "subscribe", "events": ["run_result"], "runId": "run-abc" }

Reset to all events:

{ "type": "subscribe", "events": ["*"] }

The server responds with a subscribe_ack message confirming the effective subscription.

SSE event stream (REST)

HTTP clients (curl, server-side systems, serverless functions) can receive live events without a WebSocket connection:

GET /api/runs/{runId}/events
Accept: text/event-stream

Query parameters:

Parameter	Description
events	Comma-separated event types to filter (* for all)
from	0-based index into stored task outputs for reconnection

For completed runs the stored events are replayed immediately and the connection closes. For in-progress runs the connection streams live events until the run completes.

---

Phase 5: REST Review + Context Injection + Tool Invocation

Submit a review decision via REST

Server-side systems (Slack bots, CI pipelines) can approve or reject review gates without a WebSocket connection:

POST /api/reviews/{reviewId}
Content-Type: application/json

{ "decision": "CONTINUE" }

Valid decisions: CONTINUE (or APPROVE), EDIT, EXIT_EARLY. For EDIT supply a revisedOutput field:

{ "decision": "EDIT", "revisedOutput": "Updated output..." }

Discover pending reviews:

GET /api/reviews
GET /api/reviews?runId=run-abc

Inject context into a running ensemble

Add a directive to an in-progress run’s DirectiveStore. The directive is picked up on the next LLM iteration of any agent in the ensemble.

POST /api/runs/{runId}/inject
Content-Type: application/json

{ "content": "Focus on EU AI Act compliance", "target": "researcher" }

Invoke a tool directly

Execute a registered tool from the ToolCatalog without running a full ensemble:

POST /api/tools/{toolName}/invoke
Content-Type: application/json

{ "input": "What is 42 * 17?" }

Response:

{ "tool": "calculator", "status": "SUCCESS", "output": "714", "durationMs": 2 }

Timeout is 30 seconds. The tool must be registered in the ToolCatalog configured on the WebDashboard.

---

See also

• Live Dashboard guide
• Long-Running Ensembles guide
• Ensemble Configuration reference
• Design doc 28: Ensemble Control API