diff --git a/docs/superpowers/specs/2026-04-22-zfs-pool-detail-design.md b/docs/superpowers/specs/2026-04-22-zfs-pool-detail-design.md
new file mode 100644
index 0000000..2b87092
--- /dev/null
+++ b/docs/superpowers/specs/2026-04-22-zfs-pool-detail-design.md
@@ -0,0 +1,159 @@
+# ZFS Pool Detail — Design
+
+Date: 2026-04-22
+Status: approved (pending implementation)
+
+## Problem
+
+The host detail view renders a compact row per ZFS pool today (`server/lib/server_web/live/host_detail_live.ex:67`):
+
+```
+rpool [ONLINE]
+cap 0% · frag 0% · err 0 · vdevs 4 (deg 0)                          scrub never
+```
+
+This hides information the user needs at first glance:
+
+- Total / used / free size (bytes are already collected but never rendered).
+- Pool layout (mirror / raidz1 / raidz2 / stripe / mixed) — not collected.
+- Scan state — only `end_time` is kept, so an in-progress scrub looks like a finished one.
+
+The original concept doc calls for "Health, Capacity-Bar, Fragmentation, Error-Counters, Scrub-Info, vdev-Liste" per pool (`proxmox-monitor-konzept.md:227`). We never finished that.
+
+## Goal
+
+One compact block per pool that answers at a glance: *is it healthy, what layout is it, how full is it, is a scrub running*. No drill-down yet.
+
+## Scope
+
+In scope:
+
+1. Agent collector enrichment — derive `pool_type`, keep vdev summary list, keep scan function/state. No new shell-outs; `zpool status -j --json-flat-vdevs --json-int` already returns all of this.
+2. Host detail LiveView — replace the current single-line pool row with a richer compact block (see layout below).
+3. Capacity bar styling in `assets/css/app.css`.
+4. Tests — extend `agent/test/proxmox_agent/collectors/zfs_test.exs` fixtures and assertions for the new fields.
+
+Out of scope (YAGNI):
+
+- Drill-down view with per-vdev disk state, resilver progress bars, or scan history.
+- Persistence schema changes — payload is stored as JSON blob; adding keys is additive.
+- Storage/dataset/VM panel changes — separate conversation.
+
+## Agent changes
+
+### Collector output
+
+Extend `ProxmoxAgent.Collectors.Zfs.pool_summary` with three fields:
+
+```elixir
+%{
+  # existing fields unchanged:
+  name:, health:, size_bytes:, allocated_bytes:, free_bytes:,
+  fragmentation_percent:, capacity_percent:, error_count:,
+  vdev_count:, degraded_vdev_count:, last_scrub_end:,
+
+  # new:
+  pool_type: String.t(),            # "mirror" | "raidz1" | "raidz2" | "raidz3" | "stripe" | "mixed"
+  scan_function: String.t() | nil,  # "scrub" | "resilver" | nil
+  scan_state: String.t() | nil,     # "SCANNING" | "FINISHED" | "CANCELED" | nil
+  vdevs: [%{name: String.t(), type: String.t(), state: String.t(),
+            read_errors: non_neg_integer(), write_errors: non_neg_integer(),
+            checksum_errors: non_neg_integer()}]
+}
+```
+
+### Derivation rules
+
+`pool_type` is derived from the set of `vdev_type` values across top-level vdevs:
+
+- All vdevs the same type → that type (`"mirror"`, `"raidz1"`, `"raidz2"`, `"raidz3"`).
+- All vdevs are `disk` (plain top-level disk with no redundancy) → `"stripe"`.
+- Anything else → `"mixed"`.
+
+Special vdev types (`log`, `cache`, `spare`, `dedup`, `special`) are ignored for layout classification — they don't change the data redundancy story. They are still included in the `vdevs` list.
+
+`scan_function` / `scan_state` read `get_in(status_info, ["scan", "function" | "state"])`.
+
+Per-vdev numeric fields (`read_errors`, `write_errors`, `checksum_errors`) are parsed the same way `error_count` already is (string or int tolerant).
+
+### Tests
+
+`agent/test/fixtures/zfs/zpool_status.json` already has a mirror and a raidz2 pool; extend assertions in `zfs_test.exs`:
+
+- `rpool.pool_type == "mirror"`
+- `tank.pool_type == "raidz2"`
+- `rpool.scan_state == "FINISHED"`
+- `rpool.vdevs` has length 1 with `type: "mirror"`, `state: "ONLINE"`
+
+Add one new fixture-free unit test covering the `"stripe"` and `"mixed"` branches by injecting a synthetic runner.
+
+## Server changes
+
+None in the collector pipeline. The channel handler already stores the whole `zfs_pools.pools` list as JSON (`server/lib/server_web/channels/host_channel.ex` — to confirm in plan) and the LiveView reads it with `get_in/2`. New keys flow through automatically.
+
+## UI changes
+
+### Layout
+
+Replace the current `.pool-row` flex block in `host_detail_live.ex:69-86` with a per-pool compact block:
+
+```
+rpool  mirror                                               [ONLINE]
+████████░░░░░░░░░░░░░░░░░░░░░░░░░░░░  40%
+used 200.0 GB  ·  free 300.0 GB  ·  total 500.0 GB
+frag 17%  ·  err 0  ·  vdevs 1 (deg 0)  ·  scrub finished 2026-04-19
+```
+
+Element mapping:
+
+- Line 1: pool name (bright mono, bold) · pool_type (muted) · health badge (right).
+- Line 2: capacity bar (div with width % + background color keyed to capacity thresholds).
+- Line 3: used / free / total — rendered with the existing `format_bytes/1` helper.
+- Line 4: the existing compact details line, plus scrub state — `scrub scanning` / `scrub finished <date>` / `scrub never`.
+
+### Capacity bar
+
+CSS in `server/assets/css/app.css`:
+
+```css
+.capbar {
+  height: 4px; background: var(--panel-2); border-radius: 2px;
+  overflow: hidden; margin: 0.25rem 0;
+}
+.capbar > span { display: block; height: 100%; background: var(--ok); }
+.capbar[data-level="warn"] > span { background: var(--warn); }
+.capbar[data-level="crit"] > span { background: var(--crit); }
+```
+
+Thresholds (matching the concept doc's thresholds at `proxmox-monitor-konzept.md:218-219`):
+
+- `cap >= 90` → `data-level="crit"`
+- `cap >= 80` → `data-level="warn"`
+- else → default (ok green).
+
+### Degraded pool callout
+
+For ONLINE pools with `degraded_vdev_count == 0`, do not render per-vdev detail — keep it simple. For anything else, render one line per non-ONLINE vdev below the detail line:
+
+```
+!  mirror-1  DEGRADED  r=0 w=0 cksum=12
+```
+
+Styled with the existing `.callout.err` class.
+
+### Scrub rendering
+
+- `scan_state == "SCANNING"` → `"scrub scanning"` (no date).
+- `scan_state == "FINISHED"` and `last_scrub_end` present → `"scrub #{format_date(last_scrub_end)}"`.
+- Otherwise → `"scrub never"`.
+
+`last_scrub_end` is a string like `"Sat Apr 19 02:00:00 2026"` — keep as-is or reformat to `YYYY-MM-DD` with a tiny helper (strptime isn't stdlib-trivial in Elixir; simplest: split on whitespace and reorder). Accept "as-is" if reformatting is ugly.
+
+## Risks
+
+- ZFS JSON output has changed shape between OpenZFS releases. The concept doc requires `OpenZFS 2.3+`. Agent code tolerates missing keys via `Map.get/3` defaults — keep that discipline.
+- `zpool status --json-flat-vdevs` flattens nested mirrors-of-mirrors. Top-level vdevs are keyed by name; pool_type derivation inspects only top-level entries (no child vdev walking needed in the flat form).
+
+## Rollout
+
+Additive collector changes + additive UI. No DB migration, no breaking payload change. Old agents without the new fields render the "graceful degraded" path: `pool_type` shows as `—`, scrub line falls back to `never`, capacity bar still renders from existing bytes.