@hotmeshio/long-tail

Long Tail

Durable workflow for humans in the loop.

npm install @hotmeshio/long-tail

How it works

Route work to machines or people interchangeably with ease.

import { Durable } from '@hotmeshio/hotmesh';
import type { LTEnvelope } from '@hotmeshio/long-tail';
import * as activities from './activities';

const { analyzeContent } = Durable.workflow.proxyActivities<typeof activities>({ activities });

export async function reviewContent(envelope: LTEnvelope) {
  // method calls are checkpointed and crash safe
  const analysis = await analyzeContent(envelope.data.content);

  if (analysis.confidence >= 0.85) {
    return { data: { approved: true, analysis } };
  }

  //role-based escalations are baked-in. create HITL escalations with one call
  const { workflowId } = Durable.workflow.workflowInfo();
  const decision = await Durable.workflow.condition<{ approved: boolean; notes?: string }>(
    `review-${workflowId}`,
    {
      role: 'reviewer',
      type: 'content-review',
      priority: 2,
      description: `Confidence ${analysis.confidence} — needs a human`,
      metadata: { contentId: envelope.data.contentId },
      envelope: { data: envelope.data, analysis },
    },
  );

  return { data: { approved: decision.approved, analysis } };
}

Two surfaces, one model. A proxyActivity targets a machine: call it, get a result. condition() targets the external world — a reviewer, an operator, a factory cell. It suspends the workflow and writes a single escalation row carrying everything needed to route the work: the role that should act, its type and priority, and any metadata to display or filter on. People work that row through an RBAC-scoped surface — find it, claim it, resolve it — from the dashboard, the API, or MCP, and resolving it resumes the workflow exactly where it paused.

Activities are plain functions:

export async function analyzeContent(content: string) {
  const result = await llm.classify(content);
  return { confidence: result.confidence, flags: result.flags };
}

Start

Point at Postgres. Everything else is optional.

import { start } from '@hotmeshio/long-tail';

const lt = await start({
  database: { host: 'localhost', port: 5432, user: 'postgres', password: 'password', database: 'mydb' },
  workers: [{ taskQueue: 'default', workflow: reviewContent }],
});

Dashboard at http://localhost:3000. The boilerplate has a working project with workflows, MCP servers, and MinIO.

The pattern

Four moves, from a plan you write to a plan that keeps up with reality.

Step 1 — Author a durable workflow. Your function checkpoints to Postgres. It can sleep, branch, call child workflows, wait for signals. Standard durable execution.

Step 2 — Certify it. Promotion to certified adds interceptor guarantees: failures escalate instead of throwing, escalation chains route through RBAC-scoped roles, and every error is either handled or surfaced. It cannot silently fail.

curl -X PUT http://localhost:3000/api/workflows/reviewContent/config \
  -H "Authorization: Bearer $TOKEN" \
  -d '{ "invocable": true, "task_queue": "default", "default_role": "reviewer" }'

Step 3 — React to events. Signals are reality reporting back. Workflows publish topics; agents subscribe. When activity.failed fires, an automation can re-run the step, notify a team, or trigger a different workflow. The choreography is dynamic — add subscribers through the dashboard without changing code.

Step 4 — Compile what repeats. The same workflow has two forms. What you wrote in Step 1 is the procedural form — readable, Temporal-like, emulated atop the graph: cheap to maintain, heavier to run. The Designer compiles a working execution into the graph form — the same durable workflow as a deterministic DAG: no LLM at runtime, no replay overhead, typed in and out, roughly 3x faster. Every procedural pattern has a graph equivalent and the reverse; you pick readability or speed without giving up durability, escalation, or transactional guarantees. It deploys as a reusable tool that any workflow or API call can invoke.

Over time, the system accumulates compiled tools. Problems that once required a human, then required AI reasoning, eventually require neither.

These four steps map to how the dashboard is organized. React is the reactive side (Step 3) — topics, subscriptions, automations. Orchestrate is the orchestrated side (Steps 1 and 4) — procedural and graph flows side by side, both durable and pull-based under the hood. Design is the optional bridge: with an ANTHROPIC_API_KEY it turns a description or a tool run into a graph flow; without one, choreography and orchestration stand on their own, no tradeoff.

Register MCP tools

Long Tail connects to any MCP server. Registered tools become durable activities and are available to the Pipeline Designer.

Existing package — no code:

curl -X POST http://localhost:3000/api/mcp/servers \
  -H 'Content-Type: application/json' \
  -d '{
    "name": "filesystem",
    "transport_type": "stdio",
    "transport_config": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem", "/data"] },
    "tags": ["files", "storage"],
    "auto_connect": true
  }'

Remote server — point at a URL:

curl -X POST http://localhost:3000/api/mcp/servers \
  -d '{ "name": "my-python-server", "transport_type": "sse", "transport_config": { "url": "http://python-service:8000/mcp" } }'

In-process — write your own:

import { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js';
import { z } from 'zod';
import { registerMcpTool } from '@hotmeshio/long-tail';

export function createImageToolsServer(): McpServer {
  const server = new McpServer({ name: 'image-tools', version: '1.0.0' });

  registerMcpTool(server, 'resize_image', 'Resize an image.', {
    path: z.string().describe('Path to the image'),
    width: z.number().optional(),
    height: z.number().optional(),
  }, async (args: any) => ({
    content: [{ type: 'text', text: JSON.stringify(await resize(args)) }],
  }));

  return server;
}

const lt = await start({
  // ...
  mcp: { serverFactories: { 'image-tools': createImageToolsServer } },
});

All three paths produce the same outcome: tools callable as durable activities. See the MCP guide.

Compile workflows

The ltc compiler scans TypeScript workflow files and compiles them to YAML DAGs — like tsc for workflows.

export ANTHROPIC_API_KEY=sk-ant-...
npx ltc compile workflows/

The source is the spec. The compiled YAML is the optimized execution. Both live in the repo. See the Compiler Guide.

Register a graph flow by hand

graphWorkflows is the graph-form peer of workers: hand-author the HotMesh YAML and it's created, deployed, and activated at startup. The same human surface is declarative here — a hook with an escalation: block. When the flow reaches it, the workflow suspends and writes the escalation row; a person resolves it and the flow continues:

const lt = await start({
  database: { connectionString: process.env.DATABASE_URL },
  graphWorkflows: [{
    name: 'order_approval',
    namespace: 'graph',
    inputSchema: { type: 'object', properties: { orderId: { type: 'string' }, region: { type: 'string' } }, required: ['orderId'] },
    yaml: `
app:
  id: graph
  version: '1'
  graphs:
    - subscribes: order_approval
      publishes: order_approval.done
      input:  { schema: { type: object, properties: { orderId: { type: string }, region: { type: string } } } }
      output: { schema: { type: object, properties: { approved: { type: boolean } } } }
      activities:
        trigger:
          type: trigger
        review:
          type: hook
          escalation:
            role: approver
            type: order-approval
            priority: 2
            description: Approve order for dispatch
            metadata:
              orderId: '{trigger.output.data.orderId}'
              region: '{trigger.output.data.region}'
            envelope:
              instructions: Review and approve or reject
          job:
            maps:
              approved: '{review.hook.data.approved}'
      transitions:
        trigger:
          - to: review
      hooks:
        order_approval.approve:
          - to: review
            conditions:
              match:
                - expected: '{$job.metadata.jid}'
                  actual: '{$self.hook.data.id}'
`,
  }],
});

It appears under Orchestrate › Graph and routes to a person exactly like the procedural condition() above — the same claim-and-resolve surface, declared in YAML. The metadata expressions resolve against the live job at suspension time, so the row carries the real order values.

Full configuration

const lt = await start({
  database: { connectionString: process.env.DATABASE_URL },
  workers: [{ taskQueue: 'default', workflow: reviewContent }],

  // Everything below is optional
  graphWorkflows: [{ name: 'hello_world', namespace: 'graph', yaml: helloWorldYaml }],
  seed: { admin: { externalId: 'admin', password: process.env.ADMIN_PASSWORD } },
  mcp: { server: { enabled: true }, serverFactories: { 'my-tools': createMyToolsServer } },
  escalation: { strategy: 'mcp' },
  auth: { secret: process.env.JWT_SECRET },
  telemetry: { honeycomb: { apiKey: process.env.HNY } },
  logging: { pino: { level: 'info' } },
  maintenance: true,
});

Embed in an existing app

Long Tail runs as an embedded package inside NestJS, Express, or any Node.js application. No extra HTTP server, no extra ports.

import { start, createClient } from '@hotmeshio/long-tail';

const lt = await start({
  database: { connectionString: process.env.DATABASE_URL },
  server: { enabled: false },
  seed: { admin: { externalId: 'system' } },
  workers: [{ taskQueue: 'default', workflow: reviewContent }],
});

const client = createClient({ auth: { userId: lt.adminUserId } });

const tasks = await client.tasks.list({ status: 'completed', limit: 10 });
const result = await client.escalations.claim({ id: 'esc_123', durationMinutes: 30 });

Mount the dashboard at a subpath:

import { LTExpressAdapter } from '@hotmeshio/long-tail';

const adapter = new LTExpressAdapter();
adapter.setBasePath('/admin/longtail');
app.use('/admin/longtail', adapter.getRouter());

Subscribe to events with callbacks:

client.events.on('task.completed', (event) => console.log('done:', event.workflowId));
client.events.on('escalation.*', (event) => notifyTeam(event));

Every SDK call returns an LTApiResult — same status codes, same validation, same RBAC. See the SDK guide.

Deployment

Three modes from the same codebase:

// Standalone — dashboard + API + workers
await start({ database: { connectionString: process.env.DATABASE_URL } });

// Worker-only — no HTTP server
await start({ database: { connectionString: process.env.DATABASE_URL }, server: { enabled: false }, workers: [...] });

// Embedded — inside your app, SDK calls only
await start({ database: { connectionString: process.env.DATABASE_URL }, server: { enabled: false } });
const lt = createClient({ auth: { userId: 'service' } });

All modes share PostgreSQL and scale independently. See Cloud Deployment.

Docs

Guide	What it covers
The Long Tail Story	Why this exists, what accumulates over time
Workflows	Activities, interceptor, escalation lifecycle, composition
IAM	Identity propagation, service accounts, credential exchange
Dashboard	Navigation, key pages, event feed
MCP	Server registration, tool calls, human queue
Compilation	Dynamic to deterministic pipeline wizard
Compiler	ltc compile — durable TypeScript to YAML DAGs
CLI	ltc — terminal access to workflows, escalations, knowledge, MCP
Escalation Strategies	Default, MCP triage, custom handlers
SDK	Embedded usage, createClient, event subscriptions
Architecture	Project structure, conventions, discovery
Cloud	AWS ECS, GCP Cloud Run, Docker
Data Model	Database schema

Adapters: Auth · Events · Telemetry · Logging · Maintenance · OAuth

HTTP API: Workflows · Tasks · Escalations · YAML Workflows · Users · Roles · Service Accounts · MCP Servers · Pipelines · Exports

SDK: Overview · Workflows · Tasks · Escalations · YAML Workflows · MCP · Events

Contributing

git clone https://github.com/hotmeshio/long-tail.git
cd long-tail
docker compose up -d --build

Open http://localhost:3000. Example workflows seed the dashboard.

User	Password	Role
superadmin	l0ngt@1l	superadmin
admin	l0ngt@1l	admin
engineer	l0ngt@1l	engineer
reviewer	l0ngt@1l	reviewer

See Contributing.

License

See LICENSE.