PilotSwarm Hybrid Datastore

Use this skill when an app needs the production hybrid topology:

• Runtime store: stock PostgreSQL via DATABASE_URL for Duroxide state, CMS/session catalog, ordinary facts fallback, and default Docker/local runs.
• Enhanced facts store: HorizonDB via HORIZON_DATABASE_URL for lexical, semantic, and hybrid fact search.
• Graph store: HorizonDB via HORIZON_GRAPH_DATABASE_URL for the open knowledge graph and harvester crawl receipts.

This is an opt-in provider configuration. It does not require a custom PilotSwarm Docker image. The starter/local Docker path should continue to run on stock PostgreSQL by default unless the user explicitly asks for a HorizonDB environment.

Canonical References

• Configuration guide: https://github.com/microsoft/PilotSwarm/blob/main/docs/developer/reference/configuration.md#enhanced-facts--knowledge-graph-optional
• Horizon Harvester sample: https://github.com/microsoft/PilotSwarm/tree/main/examples/horizon-harvester
• Harvester deployment guide: https://github.com/microsoft/PilotSwarm/blob/main/docs/developer/deploy/harvester.md
• Horizon provider package: https://github.com/microsoft/PilotSwarm/tree/main/packages/horizon-store
• Environment examples: .env.example and .env.horizondb.example

Provider Slots

Keep the slots separate in app code, env templates, docs, and deployment assets:

DATABASE_URL remains mandatory for PilotSwarm runtime storage. HorizonDB does not replace it. When no HORIZON_* vars are present, the same app should still start with stock PostgreSQL and no enhanced facts/graph tools.

Worker, Client, And Management Wiring

For SDK apps, wire the same Horizon facts target into worker, client, and management client by spreading horizonConfigFromEnv() everywhere the runtime accepts storage/provider options:

import {
  PilotSwarmClient,
  PilotSwarmManagementClient,
  PilotSwarmWorker,
  horizonConfigFromEnv,
} from "pilotswarm-sdk";

const store = process.env.DATABASE_URL;
const horizon = horizonConfigFromEnv();

const worker = new PilotSwarmWorker({
  store,
  githubToken: process.env.GITHUB_TOKEN,
  ...horizon,
});

// Direct { store } construction is for co-located/dev/test processes only
// (like this single-process example). App code that talks to a shared
// deployment uses Web API mode instead — it needs no database URLs at all:
//   new PilotSwarmClient({ apiUrl })
//   new PilotSwarmManagementClient({ apiUrl })
// Enhanced facts/search/graph reach remote clients through the deployment's
// Web API data-plane (createWebFactStore / createWebGraphStore).
const client = new PilotSwarmClient({ store, ...horizon });
const management = new PilotSwarmManagementClient({ store, ...horizon });

The direct client/management paths ignore worker-only graph/embed settings safely, but they must agree on the facts target so cleanup, lineage reads, and management diagnostics do not hit the wrong database. Remote clients never see these settings — the portal's Web API brokers facts/search/graph for them.

Env Template Guidance

For normal local and starter-Docker scaffolds, keep .env.example stock-PG first:

DATABASE_URL=postgres://postgres:postgres@localhost:5432/pilotswarm
GITHUB_TOKEN=

Add a separate optional hybrid block or .env.horizondb.example for HorizonDB:

# Runtime store: stock PostgreSQL
DATABASE_URL=postgres://postgres:postgres@localhost:5432/pilotswarm

# Enhanced facts/search store: HorizonDB
HORIZON_DATABASE_URL=postgres://<user>:<pass>@<horizon-host>:5432/<db>?sslmode=require
HORIZON_FACTS_SCHEMA=horizon_facts

# Graph store: HorizonDB
HORIZON_GRAPH_DATABASE_URL=postgres://<user>:<pass>@<horizon-host>:5432/<db>?sslmode=require
HORIZON_GRAPH_SCHEMA=horizon_graph

# Optional durable embed loop
HORIZON_EMBED_ENDPOINT=
HORIZON_EMBED_MODEL=text-embedding-3-small
HORIZON_EMBED_DIM=1536
HORIZON_EMBED_API_KEY=

The facts and graph schemas must be distinct when they share one HorizonDB database. The defaults horizon_facts and horizon_graph are intentionally separate.

Harvester Agent Pattern

When the app needs durable source ingestion, consult pilotswarm-knowledge-harvester and generate a plugin/agents/<source>-harvester.agent.md with:

---
schemaVersion: 1
version: 1.1.0
name: source-harvester
description: Crawls source documents into durable facts and the graph.
crawler: true
---

Load-bearing rules:

• Harvester agents write raw source captures under corpus/<source>/..., not intake/*.
• Graph store is required for the harvester crawl queue and graph tools.
• Enhanced facts are recommended for facts_search / facts_similar, but graph crawl and reconciliation can run with base facts plus graph.
• Harvester cycles must mark crawl receipts with the exact { scopeKey, etag } values returned by facts_read_uncrawled.
• Deleted crawl rows (deletedAt) must call graph_remove_evidence(scopeKey, namespace) before marking crawled.
• Recurring harvests use cron / cron_at, never a hand-rolled polling loop.

Azure / AKS Guidance

For Azure deployments, document and validate two database targets:

1. Runtime PostgreSQL (DATABASE_URL): ordinary Azure Database for PostgreSQL or another stock PostgreSQL-compatible service.
2. HorizonDB (HORIZON_DATABASE_URL / HORIZON_GRAPH_DATABASE_URL): enhanced facts/search and graph. This may be one HorizonDB cluster with separate schemas, or separate clusters when the user wants stronger isolation.

Kubernetes secret guidance should include both runtime and optional Horizon vars, using kubectl create secret generic ... --from-env-file=.env.remote when values contain shell-significant delimiters. If HorizonDB is unavailable or its required extensions are not allow-listed, the worker can still deploy against stock PostgreSQL; enhanced facts, graph tools, and harvester agents remain disabled.

Docker Constraint

Do not change Dockerfiles or the starter image just to support the hybrid store. The shipped image should continue to boot with stock PostgreSQL. Hybrid mode is selected by env/provider configuration at runtime.

Checklist

• DATABASE_URL documented as runtime stock PostgreSQL.
• HORIZON_DATABASE_URL documented only for enhanced facts/search.
• HORIZON_GRAPH_DATABASE_URL documented only for graph/harvester crawl.
• Worker, client, and management client spread horizonConfigFromEnv().
• Harvester/crawler agent has crawler: true and source-specific corpus namespace.
• Azure secret docs include both runtime and Horizon vars.
• Docker/starter docs still default to stock PostgreSQL.