diff --git a/docs/architecture/adr/001-react-context-state.md b/docs/architecture/adr/001-react-context-state.md
new file mode 100644
index 0000000..c495589
--- /dev/null
+++ b/docs/architecture/adr/001-react-context-state.md
@@ -0,0 +1,65 @@
+# ADR 001: React Context for Centralized Rook-Ceph State
+
+**Status**: Accepted
+
+**Date**: 2026-03-05
+
+**Deciders**: Development Team
+
+---
+
+## Context
+
+The Rook-Ceph plugin needs to fetch and share data from many sources:
+
+- **4 Ceph CRDs** under `ceph.rook.io/v1`: CephCluster, CephBlockPool, CephFilesystem, CephObjectStore
+- **Standard K8s resources**: StorageClasses, PersistentVolumes, PersistentVolumeClaims
+- **6 pod label selectors**: operator, mon, osd, mgr, CSI RBD, CSI CephFS
+
+This data is consumed by 7+ page views and 3 detail view sections. The context exposes 16+ fields.
+
+Data fetching uses a two-track strategy:
+
+1. **Headlamp's `K8s.ResourceClasses.*.useList()`** for standard resources (StorageClasses, PVs, PVCs)
+2. **`ApiProxy.request()` in `useEffect`** for CRDs and pods
+
+Each API call is wrapped in its own `try/catch` for independent failure isolation.
+
+---
+
+## Decision
+
+Use a single `RookCephDataProvider` React Context that centralizes all data fetching.
+
+- Standard K8s resources use Headlamp's reactive `useList()` hooks.
+- CRDs and pods use `ApiProxy.request()` in a single `useEffect` keyed on `refreshKey`.
+- Expose all data, loading, error, and refresh via context value.
+
+---
+
+## Consequences
+
+- ✅ Single fetch point avoids duplicate API calls across 7+ views
+- ✅ All views share consistent data snapshot
+- ✅ Error isolation per API call prevents one failure from blocking others
+- ✅ Refresh mechanism updates everything atomically via `refreshKey`
+- ⚠️ Large context (16+ fields) causes all consumers to re-render on any update
+- ⚠️ Monolithic provider is complex to maintain
+
+Mitigated by infrequent update cadence — data changes only on cluster state changes, not on user interaction.
+
+---
+
+## Alternatives Considered
+
+1. **Individual hooks per resource type** — Rejected. Would cause duplicate fetches across 7 pages, each independently calling the same APIs.
+
+2. **Multiple specialized contexts** (CephContext, StorageContext, PodContext) — Rejected. Adds provider nesting complexity, and the data is cross-referenced (e.g., PVC filtering depends on PV data).
+
+3. **Redux / Zustand** — Rejected. Not available as a plugin dependency; Headlamp does not expose external state management libraries.
+
+---
+
+## Changelog
+
+- 2026-03-05: Initial decision accepted
diff --git a/docs/architecture/adr/002-extract-json-data.md b/docs/architecture/adr/002-extract-json-data.md
new file mode 100644
index 0000000..dfa7b53
--- /dev/null
+++ b/docs/architecture/adr/002-extract-json-data.md
@@ -0,0 +1,50 @@
+# ADR 002: extractJsonData() Pattern for KubeObject Unwrapping
+
+**Status**: Accepted
+
+**Date**: 2026-03-05
+
+**Deciders**: Development Team
+
+---
+
+## Context
+
+Headlamp's `useList()` hooks return arrays of `KubeObject` class instances that wrap raw JSON under `.jsonData`. The plugin's type system defines plain TypeScript interfaces (e.g., `CephCluster`, `StorageClass`) matching the raw Kubernetes JSON structure.
+
+To use these typed interfaces, the `KubeObject` wrapper must be unwrapped. This pattern appears in every plugin that uses `useList()` hooks.
+
+---
+
+## Decision
+
+Implement an `extractJsonData()` utility function that takes a `KubeObject` instance and returns the unwrapped `.jsonData` property.
+
+- Apply this consistently to all `useList()` results before storing in context state.
+- All type guards (e.g., `isRookCephProvisioner()`, `isRookCephStorageClass()`) operate on the unwrapped plain objects, not on `KubeObject` wrappers.
+
+---
+
+## Consequences
+
+- ✅ Clean separation between Headlamp's class instances and the plugin's typed interfaces
+- ✅ Type guards work on plain objects, which are easier to test
+- ✅ Consistent unwrapping pattern across all resources
+- ⚠️ Extra mapping step on every `useList()` result
+- ⚠️ Runtime cost of mapping arrays (negligible for typical cluster sizes of tens to hundreds of resources)
+
+---
+
+## Alternatives Considered
+
+1. **Use `KubeObject` instances directly** — Rejected. Type guards and filters become harder to write and test with class wrappers.
+
+2. **Type assertion (`as CephCluster`)** — Rejected. Unsafe with no runtime validation; silently masks shape mismatches.
+
+3. **Custom hook wrapping `useList()` with auto-extraction** — Considered but `extractJsonData()` is simpler and more explicit. A wrapper hook would hide the unwrapping step, making the data flow less obvious.
+
+---
+
+## Changelog
+
+- 2026-03-05: Initial decision accepted
diff --git a/docs/architecture/adr/003-common-components-only.md b/docs/architecture/adr/003-common-components-only.md
new file mode 100644
index 0000000..2d99c35
--- /dev/null
+++ b/docs/architecture/adr/003-common-components-only.md
@@ -0,0 +1,57 @@
+# ADR 003: Strictly CommonComponents Only (No Direct MUI)
+
+**Status**: Accepted
+
+**Date**: 2026-03-05
+
+**Deciders**: Development Team
+
+---
+
+## Context
+
+Headlamp exports UI primitives through `@kinvolk/headlamp-plugin/lib/CommonComponents`:
+
+- `SectionBox`, `SimpleTable`, `StatusLabel`, `NameValueTable`, and others
+
+Headlamp also bundles MUI (`@mui/material`) as a shared external, making it technically accessible to plugins. Some plugins (e.g., polaris, sealed-secrets) directly use MUI components such as `Drawer`, `Alert`, and `useTheme`.
+
+The Rook plugin must decide whether to use CommonComponents exclusively or mix in direct MUI usage.
+
+---
+
+## Decision
+
+Use CommonComponents exclusively. No direct imports from `@mui/material`.
+
+- All tables use `SimpleTable`
+- All layout uses `SectionBox`
+- All status indicators use `StatusLabel`
+
+This creates a hard dependency only on Headlamp's public component API, not on MUI internals.
+
+---
+
+## Consequences
+
+- ✅ Insulated from MUI version changes in Headlamp (e.g., MUI v5 to v6 migration)
+- ✅ Consistent look-and-feel guaranteed by Headlamp's own components
+- ✅ Simpler imports with a smaller effective API surface to learn
+- ⚠️ Limited UI expressiveness — cannot use MUI `Drawer`, `Dialog`, `Stepper`, or other components not exposed by CommonComponents
+- ⚠️ Some layouts require workarounds when CommonComponents lack needed primitives
+
+Mitigated by the plugin's read-only nature, which reduces the need for complex interactive UI patterns (modals, steppers, drawers).
+
+---
+
+## Alternatives Considered
+
+1. **Mix CommonComponents with direct MUI** — Rejected for this plugin. Adds coupling risk to MUI internals, and the read-only UI does not need advanced MUI components.
+
+2. **Use only MUI directly (skip CommonComponents)** — Rejected. Would miss Headlamp's styled wrappers and risk visual inconsistency with the rest of the Headlamp UI.
+
+---
+
+## Changelog
+
+- 2026-03-05: Initial decision accepted
diff --git a/docs/architecture/adr/004-read-only-cluster-scope.md b/docs/architecture/adr/004-read-only-cluster-scope.md
new file mode 100644
index 0000000..45805da
--- /dev/null
+++ b/docs/architecture/adr/004-read-only-cluster-scope.md
@@ -0,0 +1,58 @@
+# ADR 004: Read-Only Plugin with Cluster-Wide RBAC Scope
+
+**Status**: Accepted
+
+**Date**: 2026-03-05
+
+**Deciders**: Development Team
+
+---
+
+## Context
+
+Rook-Ceph manages cluster-wide storage infrastructure. The plugin needs to display:
+
+- **Ceph CRDs**: CephClusters, CephBlockPools, CephFilesystems, CephObjectStores (all cluster-scoped or in the `rook-ceph` namespace)
+- **Cluster-scoped K8s resources**: StorageClasses, PersistentVolumes
+- **Namespace-spanning resources**: PersistentVolumeClaims (all namespaces)
+
+The plugin could offer write operations (create/delete storage classes, manage pools) or remain strictly read-only. RBAC must cover all namespaces for PVCs to show complete storage utilization.
+
+---
+
+## Decision
+
+The plugin is strictly read-only — no create, update, delete, or patch operations.
+
+- RBAC requires only `get` and `list` verbs across cluster scope.
+- PVCs are fetched with `{namespace: ''}` (all namespaces).
+- This minimizes the RBAC footprint while providing comprehensive visibility.
+
+---
+
+## Consequences
+
+- ✅ Minimal RBAC requirements (read-only `get` and `list` only)
+- ✅ No risk of accidental mutation of storage infrastructure
+- ✅ Safe for monitoring and observability use cases
+- ✅ Can be deployed in restrictive environments with minimal permissions
+- ⚠️ Users cannot manage Rook resources from the UI
+- ⚠️ Must use `kubectl` or the Rook toolbox for operational tasks
+
+Mitigated by the plugin's purpose being observability, not management. Storage infrastructure changes are high-risk and better suited to GitOps or controlled `kubectl` workflows.
+
+---
+
+## Alternatives Considered
+
+1. **Full CRUD operations** — Rejected. Storage infrastructure changes are high-risk and better suited to GitOps/kubectl workflows with proper review processes.
+
+2. **Read-only with namespace-scoped PVC filtering** — Rejected. Would miss cross-namespace storage utilization data, providing an incomplete picture of cluster storage usage.
+
+3. **Optional write mode via RBAC detection** — Rejected. Adds significant complexity (capability detection, conditional UI) for unclear benefit given the observability focus.
+
+---
+
+## Changelog
+
+- 2026-03-05: Initial decision accepted
diff --git a/docs/architecture/adr/README.md b/docs/architecture/adr/README.md
new file mode 100644
index 0000000..fea8c02
--- /dev/null
+++ b/docs/architecture/adr/README.md
@@ -0,0 +1,39 @@
+# Architecture Decision Records (ADRs)
+
+## What is an ADR?
+
+An Architecture Decision Record (ADR) captures an important architectural decision made along with its context and consequences. ADRs are immutable once accepted — if a decision is reversed, a new ADR is created that supersedes the original.
+
+## Format
+
+This project uses the [Nygard-style ADR format](https://cognitect.com/blog/2011/11/15/documenting-architecture-decisions):
+
+- **Title**: Short noun phrase describing the decision
+- **Status**: Proposed, Accepted, Deprecated, or Superseded
+- **Context**: Forces at play, including technical, political, and project-specific
+- **Decision**: The change being proposed or enacted
+- **Consequences**: What becomes easier or harder as a result
+
+## Index
+
+| ADR | Title | Status | Date |
+|-----|-------|--------|------|
+| [001](001-react-context-state.md) | React Context for Centralized Rook-Ceph State | Accepted | 2026-03-05 |
+| [002](002-extract-json-data.md) | extractJsonData() Pattern for KubeObject Unwrapping | Accepted | 2026-03-05 |
+| [003](003-common-components-only.md) | Strictly CommonComponents Only (No Direct MUI) | Accepted | 2026-03-05 |
+| [004](004-read-only-cluster-scope.md) | Read-Only Plugin with Cluster-Wide RBAC Scope | Accepted | 2026-03-05 |
+
+## Creating New ADRs
+
+1. Copy an existing ADR as a template.
+2. Assign the next sequential number (e.g., `005`).
+3. Fill in all sections: Context, Decision, Consequences, Alternatives Considered.
+4. Set the status to **Proposed** and submit a PR for review.
+5. Once merged, update the status to **Accepted** and add the entry to the index table above.
+
+Use the filename pattern `NNN-short-slug.md` (e.g., `005-new-decision.md`).
+
+## References
+
+- [Michael Nygard — Documenting Architecture Decisions](https://cognitect.com/blog/2011/11/15/documenting-architecture-decisions)
+- [ADR GitHub Organization](https://adr.github.io/)