feat: Review kind filter + polish; audit LOW items — goals 0002 and 0025 close

Review queue (goal 0002 item 4, its last): a kind Select over the four
pending kinds (policy ask / human review / breakpoint / MCP write —
discriminated off the same fields each row's own icon and badge already
key on, wording reused verbatim from row prose, fixed option order),
shown only at 2+ kinds present (the single-option-select-is-noise
rule); Blankslate empty state + Spinner loading replace the bare div/
nothing. E2e: kind-filter narrowing + emptied-queue Blankslate
(mcpTestClient.ts now the shared real-MCP-client helper across three
specs).

Audit LOW items (goal 0025 item 9, its last): WorkflowHoverPreview
clears its hover timer on unmount; unused PageContainerVariant export
removed; authoring-validation's single-node canvas click hardened with
the sibling specs' candidate-point pattern.

Both goals archived (0002's badge item was 0005's build; every other
box checked). Rides along per the intake rule: SPEC §4 gains the
owner-named communication-preset design input (Discord/Telegram/
Twilio/ntfy — goal 0023's generic forward already reaches them all;
a preset earns a build on real friction).

Gates: tsc/eslint/loc clean, 178 vitest, suite 153 passed (the new
kind-filter test was flaky-then-passed once — first observation,
logged against ADR-0034's quarantine threshold, watched).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FYwojT8GdUbYSoggbvEFft

Author: alicoding

docs/SPEC.md

@@ -2864,6 +2864,12 @@ choice doesn't).
   UI — Phase B's own bigger, separate design surface.
 - Jira/Confluence as a first-class example: still `OPEN`, unbuilt — the
   generic request type is real, but no named-vendor preset exists yet.
+  Same future-preset category, owner-named 2026-08-11 after goal 0023's
+  generic forward shipped: **communication presets** (Discord webhook /
+  Telegram bot / Twilio SMS / ntfy push — n8n's communication node
+  category) — the generic HTTPRequest already reaches all of them; a
+  preset earns a build when real wiring hits real friction, not
+  speculatively.
 - Whether connectors are built-in or a plugin surface: still `OPEN`.
 - **Backend CRUD + Configure UI — `LOCKED` and built, end-to-end.**
   `ConfigureService` (`configureservice.go`) is the Wails-bound service —

docs/goals/0002-review-queue-maturation.md

@@ -34,7 +34,7 @@ this pipeline and on this code)**
 1. [ ] [0024 — CI/CD target architecture + operating model](0024-cicd-target-architecture.md)
    — IN FLIGHT: catch-up pushed, e2e triage + target-architecture
    build land next, then the ruleset on a green main (ADR-0034)
-2. [ ] [0025 — Substance hardening](0025-substance-hardening.md) —
+2. [x] [0025 — Substance hardening](archive/0025-substance-hardening.md) — DELIVERED 2026-08-12 (both audit waves + LOW items) —
    fix the audited green-but-wrong class (12 silent persistence
    sites, unfailable test fake, uncovered safety-badge source);
    frontend half already largely clean
@@ -58,7 +58,7 @@ live-review material, interleaved during owner reviews, not a lane.**
 
 **Group C — Attention layer**
 5. [x] [0005 — Pending-attention model](0005-pending-attention-model.md) — core delivered 2026-08-10 (unified guardrail event + sidebar badge + traceless-timeout fix; OS-notification future named)
-6. [ ] [0002 — Review queue maturation](0002-review-queue-maturation.md) (remainder: pending badge built on 0005's eventing; kind filter/polish)
+6. [x] [0002 — Review queue maturation](archive/0002-review-queue-maturation.md) — DELIVERED 2026-08-12 (kind filter over four pending kinds, Blankslate/loading polish; badge came via 0005)
 
 **Unscheduled (reorder into a group when prioritized)**
 7. [x] [0012 — Authoring hot-exit](archive/0012-authoring-hot-exit.md) — canvas half delivered 2026-08-10 (scratch persistence + restored-unsaved banner + dirty dots; Configure forms recorded-remaining in the archived file)

docs/goals/BACKLOG.md

@@ -0,0 +1,65 @@
+# 0002 — Review queue maturation
+
+## Goal
+The Review queue works as the case-management-style inbox from
+`docs/SPEC.md` §3.2/§8 (ADR-0023) for real daily use — without
+crossing the no-Camunda line (no assignment/SLA/notes).
+
+## Plan
+1. [x] Workflow filter over pending + resolved (delivered 2026-08-10).
+2. [x] Recently-resolved visibility — RunSummary.Resolution read from
+   the same park event after resolution; Review shows the last 10
+   with approved/denied/timed-out labels (delivered 2026-08-10,
+   e2e-covered: deny → resolved section → filter).
+3. [x] Pending-count sidebar badge — delivered alongside 0005's unified
+   event, same build as planned (`guardrail-pending-changed`, consumed
+   in `AppSidebar.tsx` via `review-pending-count`, summed with the
+   mcp-write-approval channel, no polling); e2e-covered
+   (`guardrail.spec.ts`'s "Sidebar Review badge..." test). Checkbox was
+   stale — found already-built while working item 4, corrected here
+   rather than left inaccurate.
+4. [x] Kind filter + empty/loading polish (delivered 2026-08-11):
+   `ReviewView.tsx` gained a "Filter by kind" Select (policy ask /
+   human review checkpoint / debug park / MCP write request,
+   discriminated off `pending.source`/`pending.nodeTypeID` plus the
+   separate pendingWrites list), shown only when 2+ kinds are
+   present (SPEC §3.5's single-option-select-is-noise rule); a
+   Human review checkpoint also got its own leading icon
+   (`PersonIcon`, distinct from the ambient-ask `ShieldIcon`) since
+   filtering it as a separate kind needed a real recognition cue, not
+   just a new Select bucket. Empty state is now the shared Primer
+   `Blankslate` (`InboxIcon` + one heading line, same pattern as
+   `HomeView.tsx`/`InventoryList.tsx`); loading (`pending === null`)
+   now shows `HomeView.tsx`'s own centered-`Spinner`-under-the-Heading
+   treatment instead of rendering nothing. E2e-covered end-to-end
+   (`guardrail.spec.ts`): all three kinds parked at once (a policy
+   ask, a human-review checkpoint, and a real MCP write request over
+   a live MCP client), the Select narrows to each, and the Blankslate
+   shows once every kind clears back to zero. The MCP-client test
+   helpers (`connectMCPClient`/`findWorkflowIdByLabel`/
+   `exportWorkflowViaMCP`/`enableMCPWritesWithApprovalRequired`/
+   `restoreMCPWriteDefaults`) were promoted out of
+   `mcp-write-approval.spec.ts` into a shared `e2e/mcpTestClient.ts`
+   module so this new test didn't hand-roll a second copy.
+5. [x] Row drill-down (delivered 2026-08-10): every Review row opens
+   its run in the work-tab shell at the workflow's Runs tab, run
+   detail preselected (consumed-once `pendingRunFocus` store seam);
+   Approve/Deny stopPropagation-protected; e2e ×3. The "1-12-31"
+   timestamp root cause was deeper than display: DBOS only writes
+   StartedAt on queue-dequeue, which Mill never uses — EVERY run had a
+   zero start; `summaryFromStatus` now falls back to CreatedAt
+   (regression-tested), fixing Review, Runs, Activity, and MCP
+   `get_run`/`list_runs` responses all at once.
+6. [ ] The fuller case-management growth path (durable Case entity,
+   Statuses/Queues/Checklists/Automations, the MCP-written AI summary)
+   is recorded as design input in SPEC §3.2.1 — future research, not
+   this goal's scope.
+
+This goal's Acceptance is now fully met (items 1-5 delivered; item 6
+is recorded future design input, not build scope) — left for the
+completing commit to archive per CLAUDE.md's own convention.
+
+## Acceptance
+Run the seeded review example plus a guardrail ask; both flows are
+filterable, resolutions are visible after the fact, and the sidebar
+shows pending count — judged live by the owner.

docs/goals/archive/0002-review-queue-maturation.md

@@ -34,9 +34,19 @@ Frontend wave (audit came back largely clean; remainder):
 8. [x] Vitest for `canvasScratch.normalize/draftsEqual` (hot-exit
    dirty-detection — data-loss-adjacent pure logic) and
    `requestDraft.authConfigFrom/joseConfigFrom`.
-9. [ ] LOW items: WorkflowHoverPreview unmount timer, unused
-   PageContainerVariant export, authoring-validation.spec's unhardened
-   canvas click.
+9. [x] LOW items (delivered 2026-08-11): `WorkflowHoverPreview.tsx`'s
+   `scheduleOpen` timeout now clears on unmount (a real `useEffect`
+   cleanup, alongside the existing mouseleave-triggered `cancelOpen`);
+   `PageContainer.tsx`'s `PageContainerVariant` type is no longer
+   exported (verified zero references outside the file first — it was
+   dead public surface, kept as an internal type alias);
+   `authoring-validation.spec.ts`'s `deleteStarterNode` now uses the
+   same candidate-point/`elementFromPoint` verification the sibling
+   specs (`composition-canvas-interactions.spec.ts`,
+   `child-workflow.spec.ts`) already use for canvas-node clicks,
+   rather than a plain `.click()` that could land on React Flow's own
+   Controls/MiniMap chrome — this file's own local copy, per the
+   suite's per-file-helper convention.
 
 ## Acceptance
 Every accepted finding fixed with its committed repro (a persist

docs/goals/0025-substance-hardening.md …oals/archive/0025-substance-hardening.mddocs/goals/0025-substance-hardening.md renamed to docs/goals/archive/0025-substance-hardening.md docs/goals/0025-substance-hardening.md renamed to docs/goals/archive/0025-substance-hardening.md

@@ -50,9 +50,38 @@ async function dragPaletteItemToCanvas(page: import('@playwright/test').Page, no
 
 // Removes the pre-populated starter node -- test (a) needs an exact,
 // known single-node graph (a lone Capture, no Trigger at all), not
-// "the starter plus whatever I added."
+// "the starter plus whatever I added." Selects it by clicking a point
+// PROVEN to land inside its own card, not a fixed offset -- see
+// composition-canvas-interactions.spec.ts's/child-workflow.spec.ts's own
+// copies of this pattern for the full MiniMap/Controls-overlap reasoning
+// (a plain `.click()` targets the card's center, which React Flow's own
+// Controls/MiniMap chrome can sit under depending on layout); this
+// file's own copy, since the helper is deliberately per-file. There's
+// only ever one node at this point (a brand-new workflow's starter), so
+// no label filter is needed.
 async function deleteStarterNode(page: import('@playwright/test').Page) {
-  await activePanel(page).locator('.react-flow__node').click()
+  const node = activePanel(page).locator('.react-flow__node').first()
+  const box = await node.boundingBox()
+  if (!box) throw new Error('deleteStarterNode: starter node has no bounding box')
+  const candidates = [
+    { x: box.x + 10, y: box.y + 10 },
+    { x: box.x + box.width - 10, y: box.y + 10 },
+    { x: box.x + box.width / 2, y: box.y + box.height / 2 },
+    { x: box.x + 10, y: box.y + box.height - 10 },
+  ]
+  let clicked = false
+  for (const point of candidates) {
+    const insideNode = await page.evaluate(({ x, y }) => {
+      const el = document.elementFromPoint(x, y)
+      return !!el?.closest('.react-flow__node')
+    }, point)
+    if (insideNode) {
+      await page.mouse.click(point.x, point.y)
+      clicked = true
+      break
+    }
+  }
+  if (!clicked) throw new Error('deleteStarterNode: no candidate point resolved inside the starter node\'s own card')
   await activePanel(page).getByRole('button', { name: 'Delete selected' }).click()
   await expect(activePanel(page).locator('.react-flow__node')).toHaveCount(0)
 }

frontend/e2e/authoring-validation.spec.ts

@@ -1,4 +1,9 @@
 import { test, expect } from './fixtures/server'
+import { clickRowAction } from './inventoryRow'
+import {
+  connectMCPClient, exportWorkflowViaMCP, findWorkflowIdByLabel,
+  enableMCPWritesWithApprovalRequired, restoreMCPWriteDefaults,
+} from './mcpTestClient'
 
 // The guardrail execution gate end-to-end in the live app (docs/SPEC.md
 // §8, ADR-0019/0022), driven through the seeded "Example:
@@ -240,3 +245,98 @@ test('Review row drill-down: pending-row Approve/Deny still resolve in place, wi
   await expect(page.getByRole('tab', { name: 'Runs' })).toHaveCount(0)
   await expect(page.getByTestId('review-item').filter({ hasText: GUARDED })).toHaveCount(0, { timeout: 10_000 })
 })
+
+// Kind filter + empty/loading polish (docs/goals/0002-review-queue-
+// maturation.md item 4): three real pending kinds parked at once
+// (a policy ask, a human-review checkpoint, and an MCP write request --
+// docs/adr/0032), proving the Select appears only once 2+ kinds are
+// present, narrows correctly per kind, and the calm Blankslate empty
+// state shows once every kind is cleared back to zero.
+test('Review kind filter narrows pending rows by kind, and the Blankslate empty state shows once cleared', async ({ page }, testInfo) => {
+  await page.goto('/')
+  await page.getByRole('link', { name: 'Workflows' }).click()
+
+  // Kind 1: a policy ask.
+  const askRow = page.locator('[data-testid="inventory-row"][data-entity="workflow"]').filter({ has: page.getByText(GUARDED, { exact: true }) })
+  await askRow.getByRole('button', { name: 'Run' }).click()
+
+  // Kind 2: a human-review checkpoint.
+  const reviewSeed = 'Example: Human review with input'
+  const reviewRow = page.locator('[data-testid="inventory-row"][data-entity="workflow"]').filter({ has: page.getByText(reviewSeed, { exact: true }) })
+  await reviewRow.getByRole('button', { name: 'Run' }).click()
+  const dialog = page.getByRole('dialog')
+  await dialog.getByLabel('Note').fill('')
+  await dialog.getByRole('button', { name: 'Run' }).click()
+
+  // Kind 3: an MCP write request -- export/re-import a throwaway
+  // workflow over a real MCP client, same shape mcp-write-approval.spec.ts
+  // uses (mcpTestClient.ts is the shared helper both now import).
+  await enableMCPWritesWithApprovalRequired(page)
+  await page.getByRole('link', { name: 'Workflows' }).click()
+  await page.getByTestId('new-workflow').click()
+  await page.locator('[role="tabpanel"]:not([hidden])').last().getByLabel('Label').fill('E2E kind-filter MCP source')
+  await page.locator('[role="tabpanel"]:not([hidden])').last().getByTestId('save-workflow').click()
+  const mcpSourceRow = page.locator('[data-testid="inventory-row"][data-entity="workflow"]').filter({ has: page.getByText('E2E kind-filter MCP source', { exact: true }) })
+  await expect(mcpSourceRow).toBeVisible()
+
+  const client = await connectMCPClient(testInfo.parallelIndex)
+  const sourceId = await findWorkflowIdByLabel(client, 'E2E kind-filter MCP source')
+  const exported = await exportWorkflowViaMCP(client, sourceId)
+  const importResultPromise = client.callTool({ name: 'import_workflow', arguments: { json: exported } })
+
+  // All three kinds pending at once -- the Select appears (2+ kinds).
+  await page.getByRole('link', { name: 'Review' }).click()
+  const kindSelect = page.getByTestId('review-kind-filter')
+  await expect(kindSelect).toBeVisible({ timeout: 15_000 })
+
+  // "Awaiting approval" narrows to just the guardrail ask.
+  await kindSelect.selectOption({ label: 'Awaiting approval' })
+  await expect(page.getByTestId('review-item').filter({ hasText: GUARDED })).toBeVisible()
+  await expect(page.getByTestId('review-item').filter({ hasText: reviewSeed })).toHaveCount(0)
+  await expect(page.getByTestId('review-mcp-write-item')).toHaveCount(0)
+
+  // "Human review" narrows to just the checkpoint.
+  await kindSelect.selectOption({ label: 'Human review' })
+  await expect(page.getByTestId('review-item').filter({ hasText: reviewSeed })).toBeVisible()
+  await expect(page.getByTestId('review-item').filter({ hasText: GUARDED })).toHaveCount(0)
+  await expect(page.getByTestId('review-mcp-write-item')).toHaveCount(0)
+
+  // "MCP write request" narrows to just the pending write.
+  await kindSelect.selectOption({ label: 'MCP write request' })
+  await expect(page.getByTestId('review-mcp-write-item')).toBeVisible()
+  await expect(page.getByTestId('review-item')).toHaveCount(0)
+
+  // Back to "All kinds": every row is visible again.
+  await kindSelect.selectOption({ label: 'All kinds' })
+  await expect(page.getByTestId('review-item').filter({ hasText: GUARDED })).toBeVisible()
+  await expect(page.getByTestId('review-item').filter({ hasText: reviewSeed })).toBeVisible()
+  await expect(page.getByTestId('review-mcp-write-item')).toBeVisible()
+
+  // Clear every kind back to zero: deny both runs, approve the write.
+  await page.getByTestId('review-item').filter({ hasText: GUARDED }).getByTestId('review-deny').click()
+  await expect(page.getByTestId('review-item').filter({ hasText: GUARDED })).toHaveCount(0, { timeout: 10_000 })
+  await page.getByTestId('review-item').filter({ hasText: reviewSeed }).getByTestId('review-deny').click()
+  await expect(page.getByTestId('review-item').filter({ hasText: reviewSeed })).toHaveCount(0, { timeout: 10_000 })
+  await page.getByTestId('review-mcp-write-item').getByTestId('review-mcp-write-approve').click()
+  await expect(page.getByTestId('review-mcp-write-item')).toHaveCount(0, { timeout: 10_000 })
+
+  const result = await importResultPromise
+  await client.close()
+  if (result.isError) throw new Error(`import_workflow ultimately errored: ${JSON.stringify(result.content)}`)
+
+  // Nothing pending: the kind Select disappears (fewer than 2 kinds
+  // present) and the calm Blankslate empty state shows.
+  await expect(page.getByTestId('review-kind-filter')).toHaveCount(0)
+  await expect(page.getByTestId('review-empty')).toBeVisible({ timeout: 10_000 })
+
+  // Cleanup: both minted workflows (import always mints a new ID), and
+  // the MCP-write settings toggle.
+  await page.getByRole('link', { name: 'Workflows' }).click()
+  let remaining = await mcpSourceRow.count()
+  while (remaining > 0) {
+    await clickRowAction(page, mcpSourceRow.first(), 'Delete')
+    remaining -= 1
+    await expect(mcpSourceRow).toHaveCount(remaining)
+  }
+  await restoreMCPWriteDefaults(page)
+})