The Reqursor Architecture
Reqursor is an AI platform with two distinct layers. The AI layer interprets natural language, generates store declarations, and proposes changes. The deterministic engine validates, executes, verifies, and recovers — guaranteeing that every AI-generated deployment works.
The boundary between these layers is the Git commit. Above it, everything is non-deterministic and human-gated. Below it, everything is deterministic, auditable, and rollbackable.
Core Pipeline
From Intent to Running Store
Every AI action follows the same path: generate → validate → commit → plan → apply → reconcile. Same inputs, same result, every time.
1. AI Generates
Agency provides natural language intent. AI produces a structured store declaration conforming to the adapter schema.
2. Validate
3-tier validation: stateless per-file, stateless cross-file, and stateful checks against platform + adapter schemas. Invalid output is rejected before any mutation.
3. Git Commit
Confirmed declarations are committed to Git — the boundary between non-deterministic AI and deterministic execution. Everything below is reproducible.
4. Plan → Apply
Runtime reads desired state at an explicit SHA, generates a deterministic operation graph, and executes it. No dynamic decisions during apply.
5. Reconcile
Observed state is compared against expected state. Convergence or divergence is reported as structured evidence. Drift is surfaced, not ignored.
Desired State as Code
Every store is defined by a set of versioned files in Git. The AI generates these files. The engine consumes them. The agency can review, diff, and rollback any change.
Store Definition Schema
Every store is a set of versioned files in Git: store-definition.yaml, pinned-refs.yaml, design-tokens.json, and pages/*.json. The schema is normative — unknown fields are rejected, null values are forbidden.
Immutable Version Pinning
Every artifact (runtime, adapter, container images) is referenced by SHA-256 digest. No floating tags. No `latest`. Every deployment is traceable to exact artifact versions.
Git as Source of Truth
Every change — AI-generated, human-authored, or migration-produced — is an auditable Git commit. Rollback is deploying a prior SHA. The Runtime never reads from anything other than an explicit revision.
store_id: "maison-loft" tenant_id: "nordic-digital" domain: "maison-loft.com" adapter_ref: name: "woocommerce" digest: "sha256:9f86d08..." runtime_ref: digest: "sha256:a1b2c3d..." resource_profile: "standard" secrets_ref: "vault://nordic-digital/maison-loft" # AI-generated, human-confirmed, Git-committed # Deterministic from here.
Deterministic Engine Guarantees
These guarantees are what make AI trustworthy for production stores. Without them, every AI output would be a suggestion requiring a developer to verify.
Validation Before Mutation
AI errors are caught before any infrastructure change occurs. The engine never applies broken state.
Per-Store Isolation
Dedicated network, volumes, containers, and resource limits per store. One store's failure cannot affect another. Structural, not configurable.
Continuous Drift Detection
Observed state compared against declared intent across topology, pinning, policy, and stack configuration. Silent degradation is surfaced automatically.
Safe Rollback
Any committed state is a valid rollback target. Managed config and artifacts revert; business data (orders, customers, media) is never affected.
Stack Adapter Model
One AI, one engine, many commerce stacks — via versioned adapters.
WooCommerce
Full adapter: schema declarations, topology (app/db/cache containers), plan operations, reconcile probes, and validation. AI reads the adapter schema to constrain its output.
- Design token injection via CSS custom properties
- Block Registry with parameterized page composition
- Managed config for wp-config.php, plugin settings
Magento / Adobe Commerce
Second adapter validates the adapter model with a fundamentally different stack architecture. Same AI layer, same Runtime — different adapter.
Shopify Headless · Laravel Commerce
Each new adapter extends platform reach without changing the AI layer, Control Plane, or Runtime.
Ready to stop hiring for work AI can do?
Migrate your existing stores for free. See the platform in action with your own data. No developer time required.