fix: address review feedback for PR #891

* feat(dev): add guided cleanup and reset commands

* fix: address cleanup reset review feedback

.github/workflows/eval-weekly.yml

@@ -14,7 +14,7 @@ on:
       config:
         description: 'Eval config file (relative to apps/eval/)'
         required: false
-        default: 'configs/browseros-agent-weekly.json'
+        default: 'configs/legacy/browseros-agent-weekly.json'
 
 permissions:
   contents: read
@@ -62,33 +62,27 @@ jobs:
           curl -sL -o /tmp/nopecha.zip https://github.com/NopeCHALLC/nopecha-extension/releases/latest/download/chromium_automation.zip
           unzip -qo /tmp/nopecha.zip -d extensions/nopecha
 
-      - name: Run eval
+      - name: Run eval and publish to R2
         working-directory: packages/browseros-agent/apps/eval
         env:
           FIREWORKS_API_KEY: ${{ secrets.FIREWORKS_API_KEY }}
           OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}
           CLAUDE_CODE_OAUTH_TOKEN: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
           NOPECHA_API_KEY: ${{ secrets.NOPECHA_API_KEY }}
-          BROWSEROS_BINARY: /usr/bin/browseros
-          WEBARENA_INFINITY_DIR: /tmp/webarena-infinity
-          EVAL_CONFIG: ${{ github.event.inputs.config || 'configs/browseros-agent-weekly.json' }}
-        run: |
-          echo "Running eval with config: $EVAL_CONFIG"
-          xvfb-run --auto-servernum --server-args="-screen 0 1440x900x24" bun run src/index.ts -c "$EVAL_CONFIG"
-
-      - name: Upload runs to R2
-        if: success()
-        working-directory: packages/browseros-agent/apps/eval
-        env:
           EVAL_R2_ACCOUNT_ID: ${{ secrets.EVAL_R2_ACCOUNT_ID }}
           EVAL_R2_ACCESS_KEY_ID: ${{ secrets.EVAL_R2_ACCESS_KEY_ID }}
           EVAL_R2_SECRET_ACCESS_KEY: ${{ secrets.EVAL_R2_SECRET_ACCESS_KEY }}
           EVAL_R2_BUCKET: ${{ secrets.EVAL_R2_BUCKET }}
           EVAL_R2_CDN_BASE_URL: ${{ secrets.EVAL_R2_CDN_BASE_URL }}
-          EVAL_CONFIG: ${{ github.event.inputs.config || 'configs/browseros-agent-weekly.json' }}
+          BROWSEROS_BINARY: /usr/bin/browseros
+          WEBARENA_INFINITY_DIR: /tmp/webarena-infinity
+          # OpenClaw container runtime is macOS-only; opt the Linux runner
+          # into the no-op stub so the server can boot and the eval can run.
+          BROWSEROS_SKIP_OPENCLAW: '1'
+          EVAL_CONFIG: ${{ github.event.inputs.config || 'configs/legacy/browseros-agent-weekly.json' }}
         run: |
-          CONFIG_NAME=$(basename "$EVAL_CONFIG" .json)
-          bun scripts/upload-run.ts "results/$CONFIG_NAME"
+          echo "Running eval with config: $EVAL_CONFIG"
+          xvfb-run --auto-servernum --server-args="-screen 0 1440x900x24" bun run src/index.ts suite --config "$EVAL_CONFIG" --publish r2
 
       - name: Generate trend report
         if: success()
@@ -109,3 +103,11 @@ jobs:
         with:
           name: eval-report-${{ github.run_id }}
           path: /tmp/eval-report.html
+
+      - name: Upload server stderr logs (for post-mortem on startup failures)
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: browseros-server-logs-${{ github.run_id }}
+          path: /tmp/browseros-server-logs/
+          if-no-files-found: ignore

.github/workflows/publish-vm-agent-cache.yml

@@ -1,167 +0,0 @@
-name: Publish VM Agent Cache
-
-on:
-  workflow_dispatch:
-    inputs:
-      agent:
-        description: "Agent name from bundle.json"
-        required: true
-        type: string
-        default: openclaw
-      publish:
-        description: "Upload to R2 and merge manifest slice"
-        required: false
-        default: false
-        type: boolean
-  pull_request:
-    paths:
-      - "packages/browseros-agent/packages/build-tools/**"
-      - ".github/workflows/publish-vm-agent-cache.yml"
-
-env:
-  BUN_VERSION: "1.3.6"
-  PKG_DIR: packages/browseros-agent/packages/build-tools
-
-permissions:
-  contents: read
-
-jobs:
-  check:
-    runs-on: ubuntu-24.04
-    steps:
-      - uses: actions/checkout@v4
-      - uses: oven-sh/setup-bun@v2
-        with:
-          bun-version: ${{ env.BUN_VERSION }}
-      - working-directory: packages/browseros-agent
-        run: bun install --frozen-lockfile
-      - working-directory: packages/browseros-agent
-        run: bun run --filter @browseros/build-tools typecheck
-      - working-directory: packages/browseros-agent
-        run: bun run --filter @browseros/build-tools test
-
-  build:
-    needs: check
-    strategy:
-      fail-fast: false
-      matrix:
-        include:
-          - arch: arm64
-            runner: ubuntu-24.04-arm
-          - arch: x64
-            runner: ubuntu-24.04
-    runs-on: ${{ matrix.runner }}
-    steps:
-      - uses: actions/checkout@v4
-      - uses: oven-sh/setup-bun@v2
-        with:
-          bun-version: ${{ env.BUN_VERSION }}
-      - name: Install podman
-        run: |
-          sudo apt-get update
-          sudo apt-get install -y podman
-      - working-directory: packages/browseros-agent
-        run: bun install --frozen-lockfile
-      - name: Build tarball
-        working-directory: ${{ env.PKG_DIR }}
-        env:
-          AGENT: ${{ inputs.agent || 'openclaw' }}
-          OUT: ${{ github.workspace }}/dist/images
-        run: bun run build:tarball -- --agent "$AGENT" --arch "${{ matrix.arch }}" --output-dir "$OUT"
-      - uses: actions/upload-artifact@v4
-        with:
-          name: tarball-${{ inputs.agent || 'openclaw' }}-${{ matrix.arch }}
-          path: dist/images/
-          retention-days: 7
-
-  smoke:
-    needs: build
-    strategy:
-      fail-fast: false
-      matrix:
-        include:
-          - arch: arm64
-            runner: ubuntu-24.04-arm
-          - arch: x64
-            runner: ubuntu-24.04
-    runs-on: ${{ matrix.runner }}
-    steps:
-      - uses: actions/checkout@v4
-      - uses: oven-sh/setup-bun@v2
-        with:
-          bun-version: ${{ env.BUN_VERSION }}
-      - uses: actions/download-artifact@v4
-        with:
-          name: tarball-${{ inputs.agent || 'openclaw' }}-${{ matrix.arch }}
-          path: dist/images
-      - name: Install podman
-        run: |
-          sudo apt-get update
-          sudo apt-get install -y podman
-      - working-directory: packages/browseros-agent
-        run: bun install --frozen-lockfile
-      - name: Smoke test tarball
-        working-directory: ${{ env.PKG_DIR }}
-        env:
-          AGENT: ${{ inputs.agent || 'openclaw' }}
-        run: |
-          set -euo pipefail
-          tarball="$(find "$GITHUB_WORKSPACE/dist/images" -name "${AGENT}-*-${{ matrix.arch }}.tar.gz" -print -quit)"
-          if [ -z "$tarball" ]; then
-            echo "missing ${{ matrix.arch }} tarball artifact for ${AGENT}" >&2
-            exit 1
-          fi
-          bun run smoke:tarball -- --agent "$AGENT" --arch "${{ matrix.arch }}" --tarball "$tarball"
-
-  publish:
-    needs: [build, smoke]
-    if: ${{ github.event_name == 'workflow_dispatch' && inputs.publish == true }}
-    runs-on: ubuntu-24.04
-    environment: release
-    concurrency:
-      group: r2-manifest-publish
-      cancel-in-progress: false
-    steps:
-      - uses: actions/checkout@v4
-      - uses: oven-sh/setup-bun@v2
-        with:
-          bun-version: ${{ env.BUN_VERSION }}
-      - uses: actions/download-artifact@v4
-        with:
-          pattern: tarball-*
-          path: dist/images
-          merge-multiple: true
-      - working-directory: packages/browseros-agent
-        run: bun install --frozen-lockfile
-      - name: Upload tarballs to R2
-        working-directory: ${{ env.PKG_DIR }}
-        env:
-          R2_ACCOUNT_ID: ${{ secrets.R2_ACCOUNT_ID }}
-          R2_ACCESS_KEY_ID: ${{ secrets.R2_ACCESS_KEY_ID }}
-          R2_SECRET_ACCESS_KEY: ${{ secrets.R2_SECRET_ACCESS_KEY }}
-          R2_BUCKET: ${{ secrets.R2_BUCKET }}
-        run: |
-          set -euo pipefail
-          for file in "$GITHUB_WORKSPACE"/dist/images/*.tar.gz; do
-            base="$(basename "$file")"
-            bun run upload -- --file "$file" --key "vm/images/$base" --content-type "application/gzip" --sidecar-sha
-          done
-      - name: Merge agent slice into manifest
-        working-directory: ${{ env.PKG_DIR }}
-        env:
-          AGENT: ${{ inputs.agent || 'openclaw' }}
-          R2_ACCOUNT_ID: ${{ secrets.R2_ACCOUNT_ID }}
-          R2_ACCESS_KEY_ID: ${{ secrets.R2_ACCESS_KEY_ID }}
-          R2_SECRET_ACCESS_KEY: ${{ secrets.R2_SECRET_ACCESS_KEY }}
-          R2_BUCKET: ${{ secrets.R2_BUCKET }}
-        run: |
-          set -euo pipefail
-          mkdir -p dist/images
-          cp -R "$GITHUB_WORKSPACE"/dist/images/* dist/images/
-          bun run download -- --key vm/manifest.json --out dist/baseline-manifest.json
-          bun run emit-manifest -- \
-            --slice "agents:${AGENT}" \
-            --dist-dir dist \
-            --merge-from dist/baseline-manifest.json \
-            --out dist/manifest.json
-          bun run upload -- --file dist/manifest.json --key vm/manifest.json --content-type "application/json"

.github/workflows/test.yml

@@ -63,15 +63,15 @@ jobs:
             junit_path: test-results/server-root.xml
             needs_browser: false
           - suite: agent
-            command: bun run test:agent
+            command: (cd apps/agent && bun run test)
             junit_path: test-results/agent.xml
             needs_browser: false
           - suite: eval
-            command: bun run test:eval
+            command: (cd apps/eval && bun run test)
             junit_path: test-results/eval.xml
             needs_browser: false
           - suite: build
-            command: bun run test:build
+            command: bun run ./scripts/run-bun-test.ts ./scripts/build
             junit_path: test-results/build.xml
             needs_browser: false
 

README.md

@@ -188,6 +188,21 @@ We'd love your help making BrowserOS better! See our [Contributing Guide](CONTRI
 - [ungoogled-chromium](https://github.com/ungoogled-software/ungoogled-chromium) — BrowserOS uses some patches for enhanced privacy. Thanks to everyone behind this project!
 - [The Chromium Project](https://www.chromium.org/) — at the core of BrowserOS, making it possible to exist in the first place.
 
+## Citation
+
+If you use BrowserOS in your research or project, please cite:
+
+```bibtex
+@software{browseros2025,
+  author = {Nithin Sonti and Nikhil Sonti and {BrowserOS-team}},
+  title = {BrowserOS: The open-source Agentic browser},
+  url = {https://github.com/browseros-ai/BrowserOS},
+  year = {2025},
+  publisher = {GitHub},
+  license = {AGPL-3.0},
+}
+```
+
 ## License
 
 BrowserOS is open source under the [AGPL-3.0 license](LICENSE).

packages/browseros-agent/README.md

@@ -79,14 +79,15 @@ cp apps/server/.env.example apps/server/.env.development
 cp apps/agent/.env.example apps/agent/.env.development
 cp apps/server/.env.production.example apps/server/.env.production
 
-# Install deps, generate agent code, and sync the VM cache
+# Install deps and generate agent code
 bun run dev:setup
 
 # Start the full dev environment
 bun run dev:watch
 ```
 
-`dev:watch` exits when the VM cache manifest is missing, but setup stays in `dev:setup`.
+`dev:watch` starts the server immediately. OpenClaw VM/image prewarm runs from
+the server startup path and pulls the configured GHCR image on demand.
 
 ### Environment Variables
 
@@ -156,9 +157,14 @@ bun run build:server          # Build production server resource artifacts and u
 bun run build:agent           # Build agent extension
 
 # Test
-bun run test                  # Run standard tests
-bun run test:cdp              # Run CDP-based tests
-bun run test:integration      # Run integration tests
+bun run test                  # Run all tests
+bun run test:all              # Run all tests
+bun run test:main             # Run key server tools and integration tests
+
+# App-specific test groups (from packages/browseros-agent)
+cd apps/server && bun run test:tools
+cd apps/server && bun run test:cdp
+cd apps/server && bun run test:integration
 
 # Quality
 bun run lint                  # Check with Biome

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/AgentCard.tsx

@@ -1,136 +0,0 @@
-import { Bot, Loader2, Wrench } from 'lucide-react'
-import type { FC } from 'react'
-import type { AgentCardData } from '@/lib/agent-conversations/types'
-import { cn } from '@/lib/utils'
-
-interface AgentCardProps {
-  agent: AgentCardData
-  onClick: () => void
-  active?: boolean
-}
-
-function formatTimestamp(timestamp?: number): string {
-  if (!timestamp) return 'No activity yet'
-  const diff = Date.now() - timestamp
-  const minutes = Math.floor(diff / 60000)
-  if (minutes < 1) return 'just now'
-  if (minutes < 60) return `${minutes}m ago`
-  const hours = Math.floor(minutes / 60)
-  if (hours < 24) return `${hours}h ago`
-  return `${Math.floor(hours / 24)}d ago`
-}
-
-function getStatusLabel(status: AgentCardData['status']): string {
-  if (status === 'working') return 'Working'
-  if (status === 'error') return 'Error'
-  return 'Ready'
-}
-
-function getStatusTone(status: AgentCardData['status']): string {
-  if (status === 'working') return 'bg-amber-500'
-  if (status === 'error') return 'bg-destructive'
-  return 'bg-emerald-500'
-}
-
-function formatCost(usd: number): string {
-  if (usd < 0.005) return `$${usd.toFixed(4)}`
-  return `$${usd.toFixed(2)}`
-}
-
-export const AgentCardExpanded: FC<AgentCardProps> = ({
-  agent,
-  onClick,
-  active,
-}) => (
-  <button
-    type="button"
-    onClick={onClick}
-    className={cn(
-      'group flex min-h-32 w-full min-w-0 flex-col rounded-2xl border p-4 text-left shadow-sm transition-all duration-200',
-      active
-        ? 'border-border/80 bg-card shadow-md ring-1 ring-[var(--accent-orange)]/20'
-        : 'border-border/60 bg-card/85 hover:border-border hover:bg-card hover:shadow-md',
-    )}
-  >
-    <div className="flex items-start justify-between gap-3">
-      <div className="flex min-w-0 items-center gap-3">
-        <div
-          className={cn(
-            'flex size-10 shrink-0 items-center justify-center rounded-xl',
-            active
-              ? 'bg-[var(--accent-orange)]/10 text-[var(--accent-orange)]'
-              : 'bg-muted text-muted-foreground',
-          )}
-        >
-          <Bot className="size-5" />
-        </div>
-        <div className="min-w-0">
-          <div className="truncate font-semibold text-sm">{agent.name}</div>
-          <div className="truncate text-muted-foreground text-xs">
-            {agent.model ?? 'OpenClaw agent'}
-          </div>
-        </div>
-      </div>
-      <div className="flex items-center gap-2 rounded-full border border-border/60 bg-background/70 px-2.5 py-1 text-[11px] text-muted-foreground">
-        <span
-          className={cn('size-2 rounded-full', getStatusTone(agent.status))}
-        />
-        <span>{getStatusLabel(agent.status)}</span>
-      </div>
-    </div>
-
-    <div className="mt-4 flex-1">
-      <p className="line-clamp-2 text-foreground/90 text-sm">
-        {agent.lastMessage ??
-          'Start a conversation to see recent work and summaries.'}
-      </p>
-    </div>
-
-    <div className="mt-4 space-y-1.5 text-muted-foreground text-xs">
-      <div className="flex items-center justify-between gap-3">
-        <span>{formatTimestamp(agent.lastMessageTimestamp)}</span>
-        {agent.costUsd ? (
-          <span className="tabular-nums opacity-70">
-            {formatCost(agent.costUsd)}
-          </span>
-        ) : null}
-      </div>
-      {agent.status === 'working' && agent.currentTool ? (
-        <div className="flex items-center gap-1.5 text-[var(--accent-orange)]/70">
-          <Loader2 className="size-3 shrink-0 animate-spin" />
-          <span className="truncate">{agent.currentTool}</span>
-        </div>
-      ) : agent.activitySummary ? (
-        <div className="flex items-center gap-1.5 text-muted-foreground/60">
-          <Wrench className="size-3 shrink-0" />
-          <span className="truncate">{agent.activitySummary}</span>
-        </div>
-      ) : null}
-    </div>
-  </button>
-)
-
-export const AgentCardCompact: FC<AgentCardProps> = ({
-  agent,
-  onClick,
-  active,
-}) => (
-  <button
-    type="button"
-    onClick={onClick}
-    className={cn(
-      'inline-flex items-center gap-2 rounded-full border px-3 py-2 text-sm transition-colors',
-      active
-        ? 'border-border bg-card shadow-sm ring-1 ring-[var(--accent-orange)]/20'
-        : 'border-border/60 bg-card/85 text-foreground hover:border-border hover:bg-card',
-    )}
-  >
-    <span
-      className={cn(
-        'size-2 rounded-full',
-        active ? 'bg-[var(--accent-orange)]' : getStatusTone(agent.status),
-      )}
-    />
-    <span className="truncate">{agent.name}</span>
-  </button>
-)

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/AgentCardDock.tsx

@@ -1,70 +1,71 @@
 import { Plus } from 'lucide-react'
 import type { FC } from 'react'
-import type { AgentCardData } from '@/lib/agent-conversations/types'
+import type {
+  HarnessAdapterDescriptor,
+  HarnessAdapterHealth,
+  HarnessAgent,
+  HarnessAgentAdapter,
+} from '@/entrypoints/app/agents/agent-harness-types'
 import { cn } from '@/lib/utils'
-import { AgentCardCompact, AgentCardExpanded } from './AgentCard'
+import { HomeAgentCard } from './HomeAgentCard'
 
 interface AgentCardDockProps {
-  agents: AgentCardData[]
+  agents: HarnessAgent[]
+  adapters: HarnessAdapterDescriptor[]
   activeAgentId?: string
   onSelectAgent: (agentId: string) => void
   onCreateAgent?: () => void
-  compact?: boolean
 }
 
-function CreateAgentButton({
-  compact,
-  onCreateAgent,
-}: {
-  compact?: boolean
-  onCreateAgent: () => void
-}) {
+function CreateAgentButton({ onCreateAgent }: { onCreateAgent: () => void }) {
   return (
     <button
       type="button"
       onClick={onCreateAgent}
       className={cn(
-        'flex shrink-0 items-center justify-center gap-2 border border-dashed text-muted-foreground transition-colors hover:border-[var(--accent-orange)] hover:text-[var(--accent-orange)]',
-        compact
-          ? 'rounded-full px-3 py-2 text-sm'
-          : 'min-h-32 rounded-2xl px-5 py-4',
+        'flex min-h-32 shrink-0 items-center justify-center gap-2 rounded-2xl border border-dashed px-5 py-4 text-muted-foreground transition-colors',
+        'hover:border-[var(--accent-orange)] hover:text-[var(--accent-orange)]',
       )}
     >
-      <Plus className={compact ? 'size-3.5' : 'size-5'} />
-      <span>{compact ? 'New' : 'Create agent'}</span>
+      <Plus className="size-5" />
+      <span>Create agent</span>
     </button>
   )
 }
 
+/**
+ * 3-column grid of HomeAgentCards plus a trailing "Create agent"
+ * tile. The previous `compact` mode (rendered a horizontal pill rail)
+ * had no callers and was dropped along with the legacy AgentCard.
+ */
 export const AgentCardDock: FC<AgentCardDockProps> = ({
   agents,
+  adapters,
   activeAgentId,
   onSelectAgent,
   onCreateAgent,
-  compact,
 }) => {
   if (agents.length === 0 && !onCreateAgent) return null
 
-  const Card = compact ? AgentCardCompact : AgentCardExpanded
+  const adapterHealth = new Map<HarnessAgentAdapter, HarnessAdapterHealth>()
+  for (const descriptor of adapters) {
+    if (descriptor.health) adapterHealth.set(descriptor.id, descriptor.health)
+  }
 
   return (
-    <div
-      className={cn(
-        compact
-          ? 'flex items-center gap-2 overflow-x-auto pb-1'
-          : 'grid gap-4 md:grid-cols-3',
-      )}
-    >
+    <div className="grid gap-4 md:grid-cols-3">
       {agents.map((agent) => (
-        <Card
-          key={agent.agentId}
+        <HomeAgentCard
+          key={agent.id}
           agent={agent}
-          active={agent.agentId === activeAgentId}
-          onClick={() => onSelectAgent(agent.agentId)}
+          adapter={agent.adapter}
+          adapterHealth={adapterHealth.get(agent.adapter) ?? null}
+          active={agent.id === activeAgentId}
+          onClick={() => onSelectAgent(agent.id)}
         />
       ))}
       {onCreateAgent ? (
-        <CreateAgentButton compact={compact} onCreateAgent={onCreateAgent} />
+        <CreateAgentButton onCreateAgent={onCreateAgent} />
       ) : null}
     </div>
   )

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/AgentCommandConversation.tsx

@@ -2,6 +2,12 @@ import { ArrowLeft, Bot, Home } from 'lucide-react'
 import { type FC, useEffect, useMemo, useRef } from 'react'
 import { Navigate, useNavigate, useParams, useSearchParams } from 'react-router'
 import { Button } from '@/components/ui/button'
+import {
+  cancelHarnessTurn,
+  useEnqueueHarnessMessage,
+  useHarnessAgents,
+  useRemoveHarnessQueuedMessage,
+} from '@/entrypoints/app/agents/useAgents'
 import {
   type AgentEntry,
   getModelDisplayName,
@@ -15,6 +21,7 @@ import {
   filterTurnsPersistedInHistory,
   flattenHistoryPages,
 } from './claw-chat-types'
+import { QueuePanel } from './QueuePanel'
 import { useAgentConversation } from './useAgentConversation'
 import { useHarnessChatHistory } from './useHarnessChatHistory'
 
@@ -212,15 +219,33 @@ function AgentConversationController({
     [historyMessages],
   )
 
+  // Listing query feeds queue + active-turn state for this agent. We
+  // already poll it every 5s for the rail; reusing the same cache
+  // keeps cross-tab queue state in sync without a second poll.
+  const { harnessAgents } = useHarnessAgents()
+  const harnessAgent = harnessAgents.find((entry) => entry.id === agentId)
+  const queue = harnessAgent?.queue ?? []
+  const activeTurnId = harnessAgent?.activeTurnId ?? null
+
   const { turns, streaming, send } = useAgentConversation(agentId, {
     runtime: 'agent-harness',
     sessionKey: null,
     history: chatHistory,
+    activeTurnId,
     onComplete: () => {
       void harnessHistoryQuery.refetch()
     },
     onSessionKeyChange: () => {},
   })
+  const enqueueMessage = useEnqueueHarnessMessage()
+  const removeQueuedMessage = useRemoveHarnessQueuedMessage()
+
+  const handleStop = () => {
+    void cancelHarnessTurn(agentId, {
+      turnId: activeTurnId ?? undefined,
+      reason: 'user pressed stop',
+    })
+  }
   const visibleTurns = useMemo(
     () => filterTurnsPersistedInHistory(turns, historyMessages),
     [historyMessages, turns],
@@ -281,7 +306,15 @@ function AgentConversationController({
       />
 
       <div className="border-border/50 border-t bg-background/88 px-4 py-3 backdrop-blur-md">
-        <div className="mx-auto max-w-3xl">
+        <div className="mx-auto max-w-3xl space-y-3">
+          {queue.length > 0 ? (
+            <QueuePanel
+              queue={queue}
+              onRemove={(messageId) =>
+                removeQueuedMessage.mutate({ agentId, messageId })
+              }
+            />
+          ) : null}
           <ConversationInput
             variant="conversation"
             agents={agents}
@@ -296,14 +329,31 @@ function AgentConversationController({
                 name: a.name,
                 dataUrl: a.dataUrl,
               }))
+              // When the agent already has an in-flight turn, route
+              // the new message into the durable queue instead of
+              // starting a parallel turn. Drains automatically as
+              // soon as the active turn ends.
+              if (streaming || activeTurnId) {
+                enqueueMessage.mutate({
+                  agentId,
+                  message: input.text,
+                  attachments,
+                })
+                return
+              }
               void send({ text: input.text, attachments, attachmentPreviews })
             }}
             onCreateAgent={() => navigate(createAgentPath)}
+            onStop={handleStop}
             streaming={streaming}
             disabled={disabled}
             status="running"
             attachmentsEnabled={true}
-            placeholder={`Message ${agentName}...`}
+            placeholder={
+              streaming
+                ? `Type to queue another message for ${agentName}...`
+                : `Message ${agentName}...`
+            }
           />
         </div>
       </div>

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/AgentCommandHome.tsx

@@ -1,18 +1,25 @@
 import { Plus } from 'lucide-react'
-import { type FC, useEffect, useState } from 'react'
+import { type FC, useEffect, useMemo, useState } from 'react'
 import { useNavigate } from 'react-router'
 import { Button } from '@/components/ui/button'
 import { Card, CardContent } from '@/components/ui/card'
 import { Separator } from '@/components/ui/separator'
+import type {
+  HarnessAdapterDescriptor,
+  HarnessAgent,
+} from '@/entrypoints/app/agents/agent-harness-types'
+import {
+  useAgentAdapters,
+  useHarnessAgents,
+} from '@/entrypoints/app/agents/useAgents'
 import type { AgentEntry } from '@/entrypoints/app/agents/useOpenClaw'
 import { ImportDataHint } from '@/entrypoints/newtab/index/ImportDataHint'
 import { SignInHint } from '@/entrypoints/newtab/index/SignInHint'
 import { useActiveHint } from '@/entrypoints/newtab/index/useActiveHint'
-import type { AgentCardData } from '@/lib/agent-conversations/types'
 import { AgentCardDock } from './AgentCardDock'
 import { useAgentCommandData } from './agent-command-layout'
 import { ConversationInput } from './ConversationInput'
-import { buildAgentCardData } from './useAgentCardData'
+import { orderHomeAgents } from './home-agent-card.helpers'
 
 function EmptyAgentsState({ onOpenAgents }: { onOpenAgents: () => void }) {
   return (
@@ -38,11 +45,13 @@ function EmptyAgentsState({ onOpenAgents }: { onOpenAgents: () => void }) {
 function RecentThreads({
   activeAgentId,
   agents,
+  adapters,
   onOpenAgents,
   onSelectAgent,
 }: {
   activeAgentId?: string | null
-  agents: AgentCardData[]
+  agents: HarnessAgent[]
+  adapters: HarnessAdapterDescriptor[]
   onOpenAgents: () => void
   onSelectAgent: (agentId: string) => void
 }) {
@@ -68,6 +77,7 @@ function RecentThreads({
       </div>
       <AgentCardDock
         agents={agents}
+        adapters={adapters}
         activeAgentId={activeAgentId ?? undefined}
         onSelectAgent={onSelectAgent}
         onCreateAgent={onOpenAgents}
@@ -79,25 +89,32 @@ function RecentThreads({
 export const AgentCommandHome: FC = () => {
   const navigate = useNavigate()
   const activeHint = useActiveHint()
-  const { agents, status } = useAgentCommandData()
+  // The conversation input still consumes the merged AgentEntry list
+  // from the layout context (handles legacy /claw/agents entries that
+  // haven't yet been backfilled into the harness store). The Recent
+  // Agents grid below reads the richer harness payload directly.
+  const { agents: legacyAgents, status } = useAgentCommandData()
+  const { harnessAgents } = useHarnessAgents()
+  const { adapters } = useAgentAdapters()
   const [selectedAgentId, setSelectedAgentId] = useState<string | null>(null)
-  const cardData = buildAgentCardData(agents, status?.status, undefined)
+
+  const orderedAgents = useMemo(
+    () => orderHomeAgents(harnessAgents),
+    [harnessAgents],
+  )
 
   useEffect(() => {
-    if (agents.length === 0) {
-      if (selectedAgentId) {
-        setSelectedAgentId(null)
-      }
+    if (legacyAgents.length === 0) {
+      if (selectedAgentId) setSelectedAgentId(null)
       return
     }
-
     if (
       !selectedAgentId ||
-      !agents.some((agent) => agent.agentId === selectedAgentId)
+      !legacyAgents.some((agent) => agent.agentId === selectedAgentId)
     ) {
-      setSelectedAgentId(agents[0].agentId)
+      setSelectedAgentId(legacyAgents[0].agentId)
     }
-  }, [agents, selectedAgentId])
+  }, [legacyAgents, selectedAgentId])
 
   const handleSend = (input: { text: string }) => {
     if (!selectedAgentId) return
@@ -110,7 +127,7 @@ export const AgentCommandHome: FC = () => {
     setSelectedAgentId(agent.agentId)
   }
 
-  const selectedAgent = agents.find(
+  const selectedAgent = legacyAgents.find(
     (agent) => agent.agentId === selectedAgentId,
   )
   const selectedAgentReady = selectedAgent
@@ -118,13 +135,15 @@ export const AgentCommandHome: FC = () => {
     : false
   const selectedAgentStatus =
     selectedAgent?.source === 'agent-harness' ? 'running' : status?.status
-  const selectedCard =
-    cardData.find((agent) => agent.agentId === selectedAgentId) ?? cardData[0]
+  const selectedAgentName =
+    selectedAgent?.name ?? orderedAgents[0]?.name ?? 'your agent'
+
+  const hasAgents = legacyAgents.length > 0
 
   return (
     <div className="min-h-full px-4 py-6">
       <div className="mx-auto flex w-full max-w-5xl flex-col gap-8">
-        {cardData.length > 0 ? (
+        {hasAgents ? (
           <>
             <div className="flex flex-col items-center gap-5 pt-[max(10vh,24px)] text-center">
               <div className="space-y-3">
@@ -140,7 +159,7 @@ export const AgentCommandHome: FC = () => {
               <div className="w-full max-w-3xl">
                 <ConversationInput
                   variant="home"
-                  agents={agents}
+                  agents={legacyAgents}
                   selectedAgentId={selectedAgentId}
                   onSelectAgent={handleSelectAgent}
                   onSend={handleSend}
@@ -151,7 +170,7 @@ export const AgentCommandHome: FC = () => {
                   attachmentsEnabled={false}
                   placeholder={
                     selectedAgentReady
-                      ? `Ask ${selectedCard?.name ?? 'your agent'} to handle a task...`
+                      ? `Ask ${selectedAgentName} to handle a task...`
                       : 'Agent runtime is not running...'
                   }
                 />
@@ -162,7 +181,8 @@ export const AgentCommandHome: FC = () => {
 
             <RecentThreads
               activeAgentId={selectedAgentId}
-              agents={cardData}
+              agents={orderedAgents}
+              adapters={adapters}
               onOpenAgents={() => navigate('/agents')}
               onSelectAgent={(agentId) => navigate(`/home/agents/${agentId}`)}
             />

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/ConversationInput.tsx

@@ -54,25 +54,40 @@ interface ConversationInputProps {
   placeholder?: string
   attachmentsEnabled?: boolean
   variant?: 'home' | 'conversation'
+  /**
+   * When set, a Stop button surfaces to the left of the voice mic
+   * while `streaming === true`. Click cancels the active turn
+   * server-side via the chat-cancel endpoint. Absent → no Stop
+   * button (legacy behaviour for the home composer).
+   */
+  onStop?: () => void
 }
 
 function InputActionButton({
   disabled,
   onClick,
   streaming,
+  hasContent,
 }: {
   disabled: boolean
   onClick: () => void
   streaming: boolean
+  hasContent: boolean
 }) {
+  // Show the spinner while streaming only when there's nothing to
+  // send — once the user types something, the icon flips back to the
+  // paper-plane so it reads as "queue this message" instead of
+  // "still working".
+  const showSpinner = streaming && !hasContent
   return (
     <Button
       onClick={onClick}
       size="icon"
       disabled={disabled}
+      title={streaming && hasContent ? 'Queue message' : undefined}
       className="h-10 w-10 flex-shrink-0 rounded-xl bg-primary text-primary-foreground hover:bg-primary/90"
     >
-      {streaming ? (
+      {showSpinner ? (
         <Loader2 className="h-5 w-5 animate-spin" />
       ) : (
         <ArrowRight className="h-5 w-5" />
@@ -81,6 +96,22 @@ function InputActionButton({
   )
 }
 
+function StopButton({ onStop }: { onStop: () => void }) {
+  return (
+    <Button
+      type="button"
+      size="icon"
+      variant="ghost"
+      onClick={onStop}
+      title="Stop current turn — queued messages will start next."
+      aria-label="Stop current turn"
+      className="h-8 w-8 flex-shrink-0 rounded-lg bg-destructive/10 text-destructive transition-colors hover:bg-destructive/15 hover:text-destructive"
+    >
+      <Square className="h-3.5 w-3.5 fill-current" />
+    </Button>
+  )
+}
+
 function VoiceButton({
   isRecording,
   isTranscribing,
@@ -299,6 +330,7 @@ export const ConversationInput: FC<ConversationInputProps> = ({
   placeholder,
   attachmentsEnabled = true,
   variant = 'conversation',
+  onStop,
 }) => {
   const [input, setInput] = useState('')
   const [selectedTabs, setSelectedTabs] = useState<chrome.tabs.Tab[]>([])
@@ -379,10 +411,17 @@ export const ConversationInput: FC<ConversationInputProps> = ({
   }
 
   const hasContent = input.trim().length > 0 || attachments.length > 0
+  // Queue-aware composers (the conversation panel passes `onStop`)
+  // accept input while streaming — the parent decides whether the
+  // submission opens a new turn or enqueues onto the active one.
+  // Surfaces without a Stop hook (home) keep the legacy behaviour
+  // and block input until the current turn finishes.
+  const queueAware = Boolean(onStop)
 
   const handleSend = () => {
     const text = input.trim()
-    if (disabled || isStaging || streaming) return
+    if (disabled || isStaging) return
+    if (streaming && !queueAware) return
     if (!text && attachments.length === 0) return
     onSend({ text, attachments })
     setInput('')
@@ -512,6 +551,7 @@ export const ConversationInput: FC<ConversationInputProps> = ({
               )}
             />
           </div>
+          {streaming && onStop ? <StopButton onStop={onStop} /> : null}
           <VoiceButton
             isRecording={voice.isRecording}
             isTranscribing={voice.isTranscribing}
@@ -529,12 +569,13 @@ export const ConversationInput: FC<ConversationInputProps> = ({
               !!disabled ||
               voice.isRecording ||
               voice.isTranscribing ||
-              streaming
+              (streaming && !queueAware)
             }
             onClick={handleSend}
             // Spinner stays the user-facing "agent is busy" hint; with the
             // queue active we still spin while a turn is in flight.
             streaming={streaming}
+            hasContent={hasContent}
           />
         </div>
         {voice.error ? (

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/HomeAgentCard.tsx

@@ -0,0 +1,243 @@
+import { Quote, TriangleAlert } from 'lucide-react'
+import type { FC } from 'react'
+import { Badge } from '@/components/ui/badge'
+import {
+  HoverCard,
+  HoverCardContent,
+  HoverCardTrigger,
+} from '@/components/ui/hover-card'
+import { adapterLabel } from '@/entrypoints/app/agents/AdapterIcon'
+import { formatRelativeTime } from '@/entrypoints/app/agents/agent-display.helpers'
+import type {
+  HarnessAdapterHealth,
+  HarnessAgent,
+  HarnessAgentAdapter,
+} from '@/entrypoints/app/agents/agent-harness-types'
+import { AgentTile } from '@/entrypoints/app/agents/agent-row/AgentTile'
+import {
+  firstNonBlankLine,
+  truncate,
+} from '@/entrypoints/app/agents/agent-row/agent-row.helpers'
+import type { AgentLiveness } from '@/entrypoints/app/agents/LivenessDot'
+import { cn } from '@/lib/utils'
+
+interface HomeAgentCardProps {
+  agent: HarnessAgent
+  adapter: HarnessAgentAdapter | 'unknown'
+  /** Per-adapter health snapshot, shared across cards rendering the
+   *  same adapter. `null` when the /adapters response hasn't surfaced
+   *  health yet (we treat that as healthy until proven otherwise). */
+  adapterHealth: HarnessAdapterHealth | null
+  /** Highlights the card with an accent ring; tells the user which
+   *  agent the conversation input is bound to. */
+  active?: boolean
+  onClick: () => void
+}
+
+const PREVIEW_CHARS = 100
+
+/**
+ * Grid-shaped card for the /home Recent agents section. Composition
+ * mirrors the rail's `AgentRowCard` but the layout is a vertical
+ * column sized for a 1/3-width tile rather than a full-width row.
+ *
+ * Reuses `<AgentTile>`, `<LivenessDot>`, `livenessDetail`,
+ * `formatRelativeTime`, `firstNonBlankLine`, `truncate`, and the
+ * inline `Unavailable` chip pattern so the visual language is
+ * continuous between rail and grid.
+ */
+export const HomeAgentCard: FC<HomeAgentCardProps> = ({
+  agent,
+  adapter,
+  adapterHealth,
+  active,
+  onClick,
+}) => {
+  const status = agent.status ?? 'unknown'
+  const lastUsedAt = agent.lastUsedAt ?? null
+  const isWorking = status === 'working'
+  const isAsleep = status === 'asleep'
+  const isError = status === 'error'
+  const hasActiveTurn = Boolean(agent.activeTurnId)
+
+  return (
+    <button
+      type="button"
+      onClick={onClick}
+      className={cn(
+        'group flex min-h-32 w-full min-w-0 flex-col rounded-2xl border bg-card p-4 text-left shadow-sm transition-colors',
+        active && 'ring-1 ring-[var(--accent-orange)]/30',
+        isWorking
+          ? 'border-[var(--accent-orange)]/40'
+          : isError
+            ? 'border-destructive/30'
+            : 'border-border/60 hover:border-[var(--accent-orange)]/30',
+      )}
+    >
+      <div className="flex items-start gap-3">
+        <AgentTile adapter={adapter} status={status} lastUsedAt={lastUsedAt} />
+        <div className="min-w-0 flex-1">
+          <div className="flex items-center gap-1.5">
+            <span className="truncate font-semibold text-sm">
+              {displayName(agent)}
+            </span>
+            {isWorking && (
+              <Badge
+                variant="secondary"
+                className="ml-auto bg-amber-50 text-amber-900 hover:bg-amber-50"
+              >
+                Working
+              </Badge>
+            )}
+          </div>
+          <SummaryLine
+            adapter={adapter}
+            modelId={agent.modelId ?? null}
+            reasoningEffort={agent.reasoningEffort ?? null}
+            adapterHealth={adapterHealth}
+          />
+        </div>
+      </div>
+
+      <LastMessage message={agent.lastUserMessage ?? null} />
+
+      <div className="mt-3 flex items-center justify-between gap-2 text-muted-foreground text-xs">
+        <span>{statusFootnote(status, lastUsedAt)}</span>
+        {hasActiveTurn ? (
+          <ResumeChip />
+        ) : isAsleep ? (
+          <Badge variant="outline" className="text-muted-foreground">
+            Asleep
+          </Badge>
+        ) : isError ? (
+          <ErrorChip lastError={agent.lastError ?? null} />
+        ) : null}
+      </div>
+    </button>
+  )
+}
+
+const SummaryLine: FC<{
+  adapter: HarnessAgentAdapter | 'unknown'
+  modelId: string | null
+  reasoningEffort: string | null
+  adapterHealth: HarnessAdapterHealth | null
+}> = ({ adapter, modelId, reasoningEffort, adapterHealth }) => {
+  const parts = [adapterLabel(adapter)]
+  if (modelId) parts.push(modelId)
+  if (reasoningEffort) parts.push(reasoningEffort)
+  const unhealthy = adapterHealth?.healthy === false
+  return (
+    <div
+      className={cn(
+        'mt-0.5 flex items-center gap-1.5 text-muted-foreground text-xs',
+        unhealthy && 'text-muted-foreground/70',
+      )}
+    >
+      <span className="truncate">{parts.join(' · ')}</span>
+      {unhealthy && (
+        <HoverCard openDelay={200}>
+          <HoverCardTrigger asChild>
+            <Badge
+              variant="outline"
+              className="h-5 cursor-default gap-1 border-amber-500/40 bg-amber-50 px-1.5 text-amber-900 hover:bg-amber-50"
+            >
+              <TriangleAlert className="size-2.5" />
+              <span className="font-normal">Unavailable</span>
+            </Badge>
+          </HoverCardTrigger>
+          <HoverCardContent side="right" className="w-72 text-sm">
+            <div className="font-medium">
+              {adapterLabel(adapter)} CLI not available
+            </div>
+            <div className="mt-1 text-muted-foreground text-xs">
+              {adapterHealth?.reason ??
+                'Adapter binary missing on $PATH. Install it from the adapter docs to use this agent.'}
+            </div>
+          </HoverCardContent>
+        </HoverCard>
+      )}
+    </div>
+  )
+}
+
+const LastMessage: FC<{ message: string | null }> = ({ message }) => {
+  if (!message) {
+    return (
+      <p className="mt-3 flex-1 text-muted-foreground/70 text-xs italic">
+        No messages yet — start a chat
+      </p>
+    )
+  }
+  return (
+    <p className="mt-3 line-clamp-2 flex flex-1 items-start gap-1.5 text-foreground/85 text-sm italic leading-snug">
+      <Quote
+        className="mt-1 size-3 shrink-0 text-muted-foreground/60"
+        aria-hidden
+      />
+      <span className="line-clamp-2">
+        {truncate(firstNonBlankLine(message), PREVIEW_CHARS)}
+      </span>
+    </p>
+  )
+}
+
+const ResumeChip: FC = () => (
+  <span className="inline-flex items-center gap-1.5 rounded-full bg-[var(--accent-orange)] px-2.5 py-0.5 font-medium text-[11px] text-white shadow-sm">
+    <span className="relative flex size-1.5">
+      <span className="absolute inline-flex h-full w-full animate-ping rounded-full bg-white/70 opacity-75" />
+      <span className="relative inline-flex size-1.5 rounded-full bg-white" />
+    </span>
+    Resume
+  </span>
+)
+
+const ErrorChip: FC<{ lastError: string | null }> = ({ lastError }) => {
+  if (!lastError) {
+    return <Badge variant="destructive">Attention</Badge>
+  }
+  return (
+    <HoverCard openDelay={200}>
+      <HoverCardTrigger asChild>
+        <Badge variant="destructive" className="cursor-default">
+          Attention
+        </Badge>
+      </HoverCardTrigger>
+      <HoverCardContent
+        side="left"
+        className="max-w-xs whitespace-pre-wrap font-mono text-xs"
+      >
+        {lastError}
+      </HoverCardContent>
+    </HoverCard>
+  )
+}
+
+/**
+ * Footer left side: relative time on every state EXCEPT working,
+ * which shows `now` (the dot is already pulsing — restating it as
+ * "Working" would duplicate the pill in the title row).
+ */
+function statusFootnote(
+  status: AgentLiveness,
+  lastUsedAt: number | null,
+): string {
+  if (status === 'working') return 'now'
+  return formatRelativeTime(lastUsedAt)
+}
+
+const UUID_PATTERN =
+  /^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/i
+const OC_UUID_PATTERN =
+  /^oc-[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/i
+
+function displayName(agent: HarnessAgent): string {
+  const name = agent.name?.trim()
+  const id = agent.id
+  if (!name || name === id) {
+    if (OC_UUID_PATTERN.test(id)) return id.slice(0, 11)
+    if (UUID_PATTERN.test(id)) return id.slice(0, 8)
+    return id
+  }
+  return name
+}

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/QueuePanel.tsx

@@ -0,0 +1,94 @@
+import { ListPlus, X } from 'lucide-react'
+import type { FC } from 'react'
+import {
+  Queue,
+  QueueItem,
+  QueueItemAction,
+  QueueItemActions,
+  QueueItemAttachment,
+  QueueItemContent,
+  QueueItemFile,
+  QueueItemImage,
+  QueueList,
+  QueueSection,
+  QueueSectionContent,
+  QueueSectionLabel,
+  QueueSectionTrigger,
+} from '@/components/ai-elements/queue'
+import type {
+  HarnessQueuedMessage,
+  HarnessQueuedMessageAttachment,
+} from '@/entrypoints/app/agents/agent-harness-types'
+import { firstNonBlankLine } from '@/entrypoints/app/agents/agent-row/agent-row.helpers'
+
+interface QueuePanelProps {
+  queue: HarnessQueuedMessage[]
+  onRemove: (messageId: string) => void
+}
+
+/**
+ * Renders the agent's pending message queue using the shared AI
+ * Elements `Queue` primitives. Caller is expected to gate render on
+ * `queue.length > 0` — when empty, this returns null so the panel
+ * disappears cleanly between turns.
+ */
+export const QueuePanel: FC<QueuePanelProps> = ({ queue, onRemove }) => {
+  if (queue.length === 0) return null
+  return (
+    <Queue>
+      <QueueSection>
+        <QueueSectionTrigger>
+          <QueueSectionLabel
+            count={queue.length}
+            label={queue.length === 1 ? 'queued message' : 'queued messages'}
+            icon={<ListPlus className="size-3.5" />}
+          />
+        </QueueSectionTrigger>
+        <QueueSectionContent>
+          <QueueList>
+            {queue.map((entry) => (
+              <QueueItem key={entry.id}>
+                <div className="flex items-center gap-2">
+                  <QueueItemContent>
+                    {firstNonBlankLine(entry.message)}
+                  </QueueItemContent>
+                  <QueueItemActions>
+                    <QueueItemAction
+                      aria-label="Remove from queue"
+                      onClick={() => onRemove(entry.id)}
+                    >
+                      <X className="size-3" />
+                    </QueueItemAction>
+                  </QueueItemActions>
+                </div>
+                {entry.attachments && entry.attachments.length > 0 ? (
+                  <QueueItemAttachment>
+                    {entry.attachments.map((attachment, idx) =>
+                      renderAttachment(entry.id, attachment, idx),
+                    )}
+                  </QueueItemAttachment>
+                ) : null}
+              </QueueItem>
+            ))}
+          </QueueList>
+        </QueueSectionContent>
+      </QueueSection>
+    </Queue>
+  )
+}
+
+function renderAttachment(
+  messageId: string,
+  attachment: HarnessQueuedMessageAttachment,
+  idx: number,
+) {
+  if (attachment.mediaType.startsWith('image/')) {
+    const src = `data:${attachment.mediaType};base64,${attachment.data}`
+    return <QueueItemImage key={`${messageId}-${idx}`} src={src} />
+  }
+  return (
+    <QueueItemFile key={`${messageId}-${idx}`}>
+      {attachment.mediaType}
+    </QueueItemFile>
+  )
+}

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/home-agent-card.helpers.test.ts

@@ -0,0 +1,69 @@
+import { describe, expect, it } from 'bun:test'
+import type { HarnessAgent } from '@/entrypoints/app/agents/agent-harness-types'
+import { orderHomeAgents } from './home-agent-card.helpers'
+
+function agent(overrides: Partial<HarnessAgent>): HarnessAgent {
+  return {
+    id: overrides.id ?? 'agent-x',
+    name: overrides.name ?? overrides.id ?? 'agent-x',
+    adapter: overrides.adapter ?? 'codex',
+    permissionMode: 'approve-all',
+    sessionKey: `agent:${overrides.id ?? 'agent-x'}:main`,
+    createdAt: 1000,
+    updatedAt: 1000,
+    ...overrides,
+  }
+}
+
+describe('orderHomeAgents', () => {
+  it('places active-turn agents before everyone else', () => {
+    const sorted = orderHomeAgents([
+      agent({ id: 'a', lastUsedAt: 5000 }),
+      agent({ id: 'b', lastUsedAt: 9000, activeTurnId: 'turn-1' }),
+      agent({ id: 'c', lastUsedAt: 7000 }),
+    ])
+    expect(sorted.map((a) => a.id)).toEqual(['b', 'c', 'a'])
+  })
+
+  it('orders non-active agents by lastUsedAt desc', () => {
+    const sorted = orderHomeAgents([
+      agent({ id: 'old', lastUsedAt: 1000 }),
+      agent({ id: 'new', lastUsedAt: 9000 }),
+      agent({ id: 'mid', lastUsedAt: 5000 }),
+    ])
+    expect(sorted.map((a) => a.id)).toEqual(['new', 'mid', 'old'])
+  })
+
+  it('puts the gateway `main` seed agent above other never-used agents', () => {
+    const sorted = orderHomeAgents([
+      agent({ id: 'oc-aaaaaa', lastUsedAt: null }),
+      agent({ id: 'main', lastUsedAt: null }),
+      agent({ id: 'oc-bbbbbb', lastUsedAt: null }),
+    ])
+    expect(sorted.map((a) => a.id)).toEqual(['main', 'oc-aaaaaa', 'oc-bbbbbb'])
+  })
+
+  it('sends never-used agents to the bottom even when `main` is among them', () => {
+    const sorted = orderHomeAgents([
+      agent({ id: 'main', lastUsedAt: null }),
+      agent({ id: 'used', lastUsedAt: 5000 }),
+    ])
+    expect(sorted.map((a) => a.id)).toEqual(['used', 'main'])
+  })
+
+  it('does NOT sort by pinned — pinned agents are treated like any other', () => {
+    const sorted = orderHomeAgents([
+      agent({ id: 'unpinned-recent', lastUsedAt: 9000, pinned: false }),
+      agent({ id: 'pinned-old', lastUsedAt: 1000, pinned: true }),
+    ])
+    expect(sorted.map((a) => a.id)).toEqual(['unpinned-recent', 'pinned-old'])
+  })
+
+  it('falls back to id-stable ordering when lastUsedAt ties', () => {
+    const sorted = orderHomeAgents([
+      agent({ id: 'b', lastUsedAt: 5000 }),
+      agent({ id: 'a', lastUsedAt: 5000 }),
+    ])
+    expect(sorted.map((a) => a.id)).toEqual(['a', 'b'])
+  })
+})

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/home-agent-card.helpers.ts

@@ -0,0 +1,42 @@
+import type { HarnessAgent } from '@/entrypoints/app/agents/agent-harness-types'
+
+/**
+ * Order for the /home Recent agents grid.
+ *
+ * 1. Active turn first — agents mid-turn float to the top so the
+ *    Resume affordance is the first thing the user sees on /home.
+ * 2. The protected gateway-side `main` agent stays pinned-to-top in
+ *    the never-used group on a fresh install (mirrors the rail).
+ * 3. Recency (`lastUsedAt` desc).
+ * 4. `id` tiebreaker for stability so the grid doesn't reshuffle on
+ *    every 5-second poll.
+ *
+ * Pin is NOT a sort key. The home grid is action-oriented and trusts
+ * recency + active-turn to surface the right agent; pinning is an
+ * organisation tool that lives on the rail at /agents.
+ */
+export function orderHomeAgents(agents: HarnessAgent[]): HarnessAgent[] {
+  return [...agents].sort((a, b) => {
+    const aActive = a.activeTurnId != null
+    const bActive = b.activeTurnId != null
+    if (aActive !== bActive) return aActive ? -1 : 1
+
+    // Recency wins outright. Never-used agents (`lastUsedAt == null`)
+    // both fall to the same `-Infinity` bucket and the seed/id rules
+    // below decide their order — but a used agent always beats any
+    // never-used agent regardless of id.
+    const aValue = a.lastUsedAt ?? Number.NEGATIVE_INFINITY
+    const bValue = b.lastUsedAt ?? Number.NEGATIVE_INFINITY
+    if (aValue !== bValue) return bValue - aValue
+
+    // Inside the never-used (or exact-tie) group: pin the gateway
+    // `main` seed to the top of the group on a fresh install, then
+    // fall back to id-stable order so the grid doesn't reshuffle on
+    // every poll.
+    const aSeed = a.id === 'main' && a.lastUsedAt == null
+    const bSeed = b.id === 'main' && b.lastUsedAt == null
+    if (aSeed !== bSeed) return aSeed ? -1 : 1
+
+    return a.id.localeCompare(b.id)
+  })
+}

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/useAgentCardData.ts

@@ -1,53 +0,0 @@
-import {
-  type AgentEntry,
-  getModelDisplayName,
-  type OpenClawStatus,
-} from '@/entrypoints/app/agents/useOpenClaw'
-import type { AgentCardData } from '@/lib/agent-conversations/types'
-import type { AgentOverview } from './useAgentDashboard'
-
-function resolveAgentStatus(
-  gatewayStatus: OpenClawStatus['status'] | undefined,
-  liveStatus: AgentOverview['status'] | undefined,
-): AgentCardData['status'] {
-  // Gateway-level errors take precedence
-  if (gatewayStatus === 'error') return 'error'
-  if (gatewayStatus === 'starting') return 'working'
-
-  // Per-agent live status from the WS observer
-  if (liveStatus === 'working') return 'working'
-  if (liveStatus === 'error') return 'error'
-
-  return 'idle'
-}
-
-/**
- * Build agent card display data by merging the raw agent entries from
- * the gateway with enriched overview data from the dashboard API.
- *
- * Pure function — no hooks, no IndexedDB, no async.
- */
-export function buildAgentCardData(
-  agents: AgentEntry[],
-  status: OpenClawStatus['status'] | undefined,
-  dashboard: AgentOverview[] | undefined,
-): AgentCardData[] {
-  return agents.map((agent) => {
-    const overview = dashboard?.find((d) => d.agentId === agent.agentId)
-
-    return {
-      agentId: agent.agentId,
-      name: agent.name,
-      model: getModelDisplayName(agent.model),
-      status:
-        agent.source === 'agent-harness'
-          ? 'idle'
-          : resolveAgentStatus(status, overview?.status),
-      lastMessage: overview?.latestMessage?.slice(0, 200) ?? undefined,
-      lastMessageTimestamp: overview?.latestMessageAt ?? undefined,
-      activitySummary: overview?.activitySummary ?? undefined,
-      currentTool: overview?.currentTool ?? undefined,
-      costUsd: overview?.totalCostUsd ?? undefined,
-    }
-  })
-}

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/useAgentConversation.ts

@@ -36,6 +36,15 @@ interface UseAgentConversationOptions {
   history?: OpenClawChatHistoryMessage[]
   onComplete?: () => void
   onSessionKeyChange?: (sessionKey: string) => void
+  /**
+   * Server-side active turn id, surfaced via the listing query. When
+   * this changes from null/<id> to a different non-null id while we
+   * aren't already streaming (e.g. the server just popped a queued
+   * message and started a new turn), the hook reattaches via
+   * /chat/active so the chat panel picks up the live stream without
+   * waiting for a remount.
+   */
+  activeTurnId?: string | null
 }
 
 export function useAgentConversation(
@@ -211,31 +220,46 @@ export function useAgentConversation(
   }
   processEventRef.current = processAgentHarnessStreamEvent
 
-  // On mount (and whenever the agent changes), check whether the
-  // server has an in-flight turn for this agent and reattach to it.
-  // This is what makes the chat resilient across tab close/reopen,
-  // refresh, and navigation: the runtime call kept running on the
-  // server while we were away. Effect only depends on `agentId` —
-  // the event handler is read off a ref so this doesn't re-subscribe
-  // every render.
+  const activeTurnIdDep = options.activeTurnId ?? null
+
+  // On mount, on agent change, and whenever the listing reports a
+  // *new* active turn id, check whether the server has an in-flight
+  // turn for this agent and reattach to it. This catches three
+  // cases at once: the chat resilience flow (tab close/reopen),
+  // navigation between agents, AND queue drain (the server starts a
+  // new turn from a queued message → activeTurnId flips → attach).
   useEffect(() => {
     let cancelled = false
     const abortController = new AbortController()
+    // Reference the dep inside the body so biome's exhaustive-deps
+    // rule sees it consumed; the value is just an "any non-null
+    // active turn id" trigger — the actual id we attach to comes
+    // from the fresh fetchActiveHarnessTurn call below.
+    void activeTurnIdDep
 
     const attemptResume = async () => {
+      // Track whether *we* started a stream in this run. When the
+      // early-return paths fire (no active turn, or a `send()` /
+      // earlier resume already owns `streamAbortRef`), the finally
+      // block must NOT touch streaming/turnIdRef/lastSeqRef —
+      // otherwise we clobber the in-flight stream's state and the
+      // Stop button drops out mid-turn while events keep arriving.
+      let weStartedStream = false
       try {
         const active = await fetchActiveHarnessTurn(agentId)
         if (cancelled || !active || active.status !== 'running') return
-        if (streamAbortRef.current) return // a fresh send already in flight
+        if (streamAbortRef.current) return // someone else already owns the stream
 
         // Stage a placeholder turn so the streamed events have a row
-        // to render into. We don't have the user message text on
-        // resume; the assistant turn is what we're catching up on.
+        // to render into. The server now persists the kicking-off
+        // prompt on the active turn, so we render it as the user
+        // bubble immediately — no empty-bubble flicker when a queued
+        // message starts running.
         setTurns((prev) => [
           ...prev,
           {
             id: crypto.randomUUID(),
-            userText: '',
+            userText: active.prompt ?? '',
             parts: [],
             done: false,
             timestamp: active.startedAt,
@@ -247,6 +271,7 @@ export function useAgentConversation(
         lastSeqRef.current = null
         streamAbortRef.current = abortController
         setStreaming(true)
+        weStartedStream = true
 
         const response = await attachToHarnessTurn(agentId, {
           turnId: active.turnId,
@@ -265,10 +290,20 @@ export function useAgentConversation(
         // Resume is best-effort; transient errors fall back to the
         // user starting a new turn manually.
       } finally {
-        if (!cancelled) {
-          if (streamAbortRef.current === abortController) {
-            streamAbortRef.current = null
-          }
+        // Always release `streamAbortRef` if we owned it — even when
+        // the effect was cancelled mid-stream (a listing poll
+        // captured the next queue-drain turn id, for example). If we
+        // don't, the next effect run hits `if (streamAbortRef.current)
+        // return` against our now-aborted controller and never
+        // reattaches, leaving `streaming === true` with no live stream.
+        if (weStartedStream && streamAbortRef.current === abortController) {
+          streamAbortRef.current = null
+        }
+        // The other state (streaming flag, turn id, lastSeq) is the
+        // *current run's* lifecycle: only reset it on a clean exit.
+        // When `cancelled` is true the next run will set these
+        // itself, so resetting here would only cause a brief flicker.
+        if (!cancelled && weStartedStream) {
           turnIdRef.current = null
           lastSeqRef.current = null
           setStreaming(false)
@@ -281,7 +316,7 @@ export function useAgentConversation(
       cancelled = true
       abortController.abort()
     }
-  }, [agentId])
+  }, [agentId, activeTurnIdDep])
 
   const send = async (input: string | SendInput) => {
     const normalized: SendInput =

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/useAgentDashboard.ts

@@ -1,95 +0,0 @@
-import { useQuery, useQueryClient } from '@tanstack/react-query'
-import { useEffect } from 'react'
-import { useAgentServerUrl } from '@/lib/browseros/useBrowserOSProviders'
-
-export interface AgentOverview {
-  agentId: string
-  status: 'working' | 'idle' | 'error' | 'unknown'
-  latestMessage: string | null
-  latestMessageAt: number | null
-  activitySummary: string | null
-  currentTool: string | null
-  totalCostUsd: number
-  sessionCount: number
-}
-
-export interface DashboardResponse {
-  agents: AgentOverview[]
-  summary: {
-    totalAgents: number
-    totalCostUsd: number
-  }
-}
-
-interface StatusEvent {
-  agentId: string
-  status: AgentOverview['status']
-  currentTool: string | null
-  error: string | null
-  timestamp: number
-}
-
-const DASHBOARD_QUERY_KEY = ['claw', 'dashboard']
-
-export function useAgentDashboard(enabled: boolean) {
-  const { baseUrl, isLoading: urlLoading } = useAgentServerUrl()
-  const queryClient = useQueryClient()
-  const ready = enabled && Boolean(baseUrl) && !urlLoading
-
-  // Initial data load + periodic refresh as fallback
-  const query = useQuery<DashboardResponse>({
-    queryKey: [...DASHBOARD_QUERY_KEY, baseUrl],
-    queryFn: async () => {
-      const url = new URL('/claw/dashboard', baseUrl as string)
-      const response = await fetch(url.toString())
-      if (!response.ok) throw new Error('Failed to fetch dashboard')
-      return response.json()
-    },
-    enabled: ready,
-  })
-
-  // SSE subscription for real-time status patches
-  useEffect(() => {
-    if (!ready || !baseUrl) return
-
-    const streamUrl = new URL('/claw/dashboard/stream', baseUrl)
-    const eventSource = new EventSource(streamUrl.toString())
-
-    eventSource.addEventListener('snapshot', (event) => {
-      try {
-        const dashboard = JSON.parse(event.data) as DashboardResponse
-        queryClient.setQueryData([...DASHBOARD_QUERY_KEY, baseUrl], dashboard)
-      } catch {}
-    })
-
-    eventSource.addEventListener('status', (event) => {
-      try {
-        const status = JSON.parse(event.data) as StatusEvent
-        queryClient.setQueryData<DashboardResponse>(
-          [...DASHBOARD_QUERY_KEY, baseUrl],
-          (prev) => {
-            if (!prev) return prev
-            return {
-              ...prev,
-              agents: prev.agents.map((agent) =>
-                agent.agentId === status.agentId
-                  ? {
-                      ...agent,
-                      status: status.status,
-                      currentTool: status.currentTool,
-                    }
-                  : agent,
-              ),
-            }
-          },
-        )
-      } catch {}
-    })
-
-    return () => {
-      eventSource.close()
-    }
-  }, [ready, baseUrl, queryClient])
-
-  return query
-}

packages/browseros-agent/apps/agent/entrypoints/app/agents/AgentList.tsx

@@ -2,67 +2,87 @@ import { Loader2 } from 'lucide-react'
 import { type FC, useMemo } from 'react'
 import { AgentRowCard } from './AgentRowCard'
 import { AgentsEmptyState } from './AgentsEmptyState'
-import type { HarnessAgent, HarnessAgentAdapter } from './agent-harness-types'
+import type {
+  HarnessAdapterDescriptor,
+  HarnessAgent,
+  HarnessAgentAdapter,
+} from './agent-harness-types'
+import type {
+  AgentAdapterHealth,
+  AgentRowData,
+} from './agent-row/agent-row.types'
 import type { AgentListItem } from './agents-page-types'
 import type { AgentLiveness } from './LivenessDot'
 
 interface AgentListProps {
   agents: AgentListItem[]
-  /**
-   * Optional per-agent activity metadata. Keyed by `agentId`. Missing
-   * entries fall back to status='unknown' / lastUsedAt=null and the
-   * row renders an "unknown" dot. The server will populate this once
-   * the activity tracker ships; the page works without it.
-   */
+  /** Optional per-agent activity metadata, keyed by `agentId`. */
   activity?: Record<
     string,
     { status: AgentLiveness; lastUsedAt: number | null }
   >
-  /**
-   * Lookup table from harness agent id → adapter + reasoning effort,
-   * sourced from `useHarnessAgents`. Lets the row card render the
-   * correct adapter icon and chips for harness agents (legacy
-   * /claw/agents entries fall back to inferring from `runtimeLabel`).
-   */
+  /** Lookup table from harness id → enriched agent record. */
   harnessAgentLookup?: Map<string, HarnessAgent>
+  /** Adapter catalog (carries per-adapter health). */
+  adapters: HarnessAdapterDescriptor[]
   loading: boolean
   deletingAgentKey: string | null
   onCreateAgent: () => void
   onDeleteAgent: (agent: AgentListItem) => void
+  onPinToggle: (agent: AgentListItem, next: boolean) => void
 }
 
 export const AgentList: FC<AgentListProps> = ({
   agents,
   activity,
   harnessAgentLookup,
+  adapters,
   loading,
   deletingAgentKey,
   onCreateAgent,
   onDeleteAgent,
+  onPinToggle,
 }) => {
-  // Sort by recency: most recently used first; never-used agents drop
-  // to the bottom in id-stable order so the list doesn't reshuffle on
-  // every refresh. The pinned exception is the gateway's `main` agent
-  // when it's never been touched — keep it at the top so a fresh
-  // install has an obvious starting point.
+  const adapterHealth = useMemo(() => {
+    const map = new Map<HarnessAgentAdapter, AgentAdapterHealth>()
+    for (const adapter of adapters) {
+      if (adapter.health) {
+        map.set(adapter.id, {
+          healthy: adapter.health.healthy,
+          reason: adapter.health.reason,
+        })
+      }
+    }
+    return map
+  }, [adapters])
+
+  // Sort: pinned rows first, then most recently used, then never-used
+  // agents in id-stable order. The gateway's `main` agent stays
+  // pinned-to-top when never touched so a fresh install has an
+  // obvious starting point.
   const ordered = useMemo(() => {
-    const withScore = agents.map((agent) => {
-      const lastUsedAt = activity?.[agent.agentId]?.lastUsedAt ?? null
-      return { agent, lastUsedAt }
+    const withMeta = agents.map((agent) => {
+      const harness = harnessAgentLookup?.get(agent.agentId)
+      return {
+        agent,
+        pinned: harness?.pinned ?? false,
+        lastUsedAt: activity?.[agent.agentId]?.lastUsedAt ?? null,
+      }
     })
-    return withScore
+    return withMeta
       .sort((a, b) => {
-        const aPinned = a.agent.agentId === 'main' && a.lastUsedAt === null
-        const bPinned = b.agent.agentId === 'main' && b.lastUsedAt === null
-        if (aPinned && !bPinned) return -1
-        if (!aPinned && bPinned) return 1
+        if (a.pinned !== b.pinned) return a.pinned ? -1 : 1
+        const aSeed = a.agent.agentId === 'main' && a.lastUsedAt === null
+        const bSeed = b.agent.agentId === 'main' && b.lastUsedAt === null
+        if (aSeed && !bSeed) return -1
+        if (!aSeed && bSeed) return 1
         const aValue = a.lastUsedAt ?? -Infinity
         const bValue = b.lastUsedAt ?? -Infinity
         if (aValue !== bValue) return bValue - aValue
         return a.agent.agentId.localeCompare(b.agent.agentId)
       })
       .map((entry) => entry.agent)
-  }, [activity, agents])
+  }, [activity, agents, harnessAgentLookup])
 
   if (loading && agents.length === 0) {
     return (
@@ -80,18 +100,23 @@ export const AgentList: FC<AgentListProps> = ({
     <div className="grid gap-3">
       {ordered.map((agent) => {
         const harness = harnessAgentLookup?.get(agent.agentId)
-        const adapter: HarnessAgentAdapter | undefined =
+        const adapter: HarnessAgentAdapter | 'unknown' =
           harness?.adapter ?? inferAdapterFromLabel(agent.runtimeLabel)
+        const data = buildRowData({
+          agent,
+          adapter,
+          harness,
+          activity: activity?.[agent.agentId],
+          adapterHealth:
+            adapterHealth.get(adapter as HarnessAgentAdapter) ?? null,
+        })
         return (
           <AgentRowCard
             key={agent.key}
-            agent={agent}
-            status={activity?.[agent.agentId]?.status}
-            lastUsedAt={activity?.[agent.agentId]?.lastUsedAt}
-            adapter={adapter}
-            reasoningEffort={harness?.reasoningEffort ?? null}
-            onDelete={onDeleteAgent}
+            data={data}
             deleting={deletingAgentKey === agent.key}
+            onDelete={onDeleteAgent}
+            onPinToggle={onPinToggle}
           />
         )
       })}
@@ -99,10 +124,53 @@ export const AgentList: FC<AgentListProps> = ({
   )
 }
 
-function inferAdapterFromLabel(label: string): HarnessAgentAdapter | undefined {
+function inferAdapterFromLabel(label: string): HarnessAgentAdapter | 'unknown' {
   const lower = label?.toLowerCase()
   if (lower === 'claude code') return 'claude'
   if (lower === 'codex') return 'codex'
   if (lower === 'openclaw') return 'openclaw'
-  return undefined
+  return 'unknown'
+}
+
+const ZERO_BUCKETS = (): number[] => Array.from({ length: 14 }, () => 0)
+
+function buildRowData(input: {
+  agent: AgentListItem
+  adapter: HarnessAgentAdapter | 'unknown'
+  harness: HarnessAgent | undefined
+  activity: { status: AgentLiveness; lastUsedAt: number | null } | undefined
+  adapterHealth: AgentAdapterHealth | null
+}): AgentRowData {
+  const { agent, adapter, harness, activity, adapterHealth } = input
+  return {
+    agent,
+    adapter,
+    modelLabel: deriveModelLabel(agent, harness),
+    reasoningEffort: harness?.reasoningEffort ?? null,
+    status: activity?.status ?? 'unknown',
+    lastUsedAt: activity?.lastUsedAt ?? harness?.lastUsedAt ?? null,
+    pinned: harness?.pinned ?? false,
+    cwd: harness?.cwd ?? null,
+    lastUserMessage: harness?.lastUserMessage ?? null,
+    tokens: harness?.tokens ?? null,
+    turnsByDay: harness?.turnsByDay ?? ZERO_BUCKETS(),
+    failedByDay: harness?.failedByDay ?? ZERO_BUCKETS(),
+    lastError: harness?.lastError ?? null,
+    lastErrorAt: harness?.lastErrorAt ?? null,
+    activeTurnId: harness?.activeTurnId ?? null,
+    adapterHealth,
+  }
+}
+
+function deriveModelLabel(
+  agent: AgentListItem,
+  harness: HarnessAgent | undefined,
+): string | null {
+  // Prefer the agent rail's modelLabel when meaningful; harness's
+  // modelId is a stable identifier but the rail's `modelLabel`
+  // already maps to a friendly display string.
+  if (agent.modelLabel && agent.modelLabel !== 'default') {
+    return agent.modelLabel
+  }
+  return harness?.modelId ?? null
 }

packages/browseros-agent/apps/agent/entrypoints/app/agents/AgentRowCard.tsx

@@ -1,270 +1,99 @@
-import {
-  Copy,
-  Loader2,
-  MessageSquare,
-  MoreHorizontal,
-  Pencil,
-  RotateCcw,
-  Trash2,
-} from 'lucide-react'
 import type { FC } from 'react'
-import { useNavigate } from 'react-router'
-import { toast } from 'sonner'
-import { Badge } from '@/components/ui/badge'
-import { Button } from '@/components/ui/button'
-import {
-  DropdownMenu,
-  DropdownMenuContent,
-  DropdownMenuItem,
-  DropdownMenuSeparator,
-  DropdownMenuTrigger,
-} from '@/components/ui/dropdown-menu'
-import {
-  Tooltip,
-  TooltipContent,
-  TooltipProvider,
-  TooltipTrigger,
-} from '@/components/ui/tooltip'
 import { cn } from '@/lib/utils'
-import { AdapterIcon, adapterLabel } from './AdapterIcon'
-import {
-  canDelete as canDeleteAgent,
-  canRename as canRenameAgent,
-  displayName,
-  formatRelativeTime,
-  workspaceLabel,
-} from './agent-display.helpers'
-import type { HarnessAgentAdapter } from './agent-harness-types'
-import type { AgentListItem } from './agents-page-types'
-import { type AgentLiveness, LivenessDot } from './LivenessDot'
+import { AgentActions } from './agent-row/AgentActions'
+import { AgentErrorPanel } from './agent-row/AgentErrorPanel'
+import { AgentLastMessage } from './agent-row/AgentLastMessage'
+import { AgentMetaRow } from './agent-row/AgentMetaRow'
+import { AgentSummaryChips } from './agent-row/AgentSummaryChips'
+import { AgentTile } from './agent-row/AgentTile'
+import { AgentTitleRow } from './agent-row/AgentTitleRow'
+import type {
+  AgentRowCallbacks,
+  AgentRowData,
+} from './agent-row/agent-row.types'
 
-interface AgentRowCardProps {
-  agent: AgentListItem
-  /**
-   * Per-agent extras the listing surface provides on top of the
-   * minimal `AgentListItem` shape. `lastUsedAt` survives server
-   * restart (sourced from acpx session record); `status` is in-memory
-   * server-side.
-   */
-  status?: AgentLiveness
-  lastUsedAt?: number | null
-  /** Adapter the agent belongs to. Drives icon + label. */
-  adapter?: HarnessAgentAdapter
-  /** Reasoning effort chip (claude/codex/openclaw catalog). */
-  reasoningEffort?: string | null
-  /** Modeled directly off the inbound delete handler so the parent owns the dialog. */
-  onDelete: (agent: AgentListItem) => void
-  /** Whether THIS agent is mid-delete; renders a spinner in place of the trash icon. */
+interface AgentRowCardProps extends AgentRowCallbacks {
+  data: AgentRowData
+  /** Whether THIS agent is mid-delete; renders a spinner in the menu. */
   deleting?: boolean
 }
 
+/**
+ * Composition shell for the agent rail. Owns no state; sub-components
+ * each handle their own micro-state (error-panel collapse, etc.) and
+ * emit callbacks (delete, pin/unpin) for the page to act on.
+ *
+ * The whole card carries state — not just the tile — so the row's
+ * border subtly tells the user what's going on at a glance:
+ *   working → accent-orange border with a soft glow
+ *   error   → destructive border
+ *   idle    → muted border, lifts on hover
+ */
 export const AgentRowCard: FC<AgentRowCardProps> = ({
-  agent,
-  status = 'unknown',
-  lastUsedAt,
-  adapter,
-  reasoningEffort,
-  onDelete,
+  data,
   deleting,
+  onDelete,
+  onPinToggle,
 }) => {
-  const navigate = useNavigate()
-  const adapterId = adapter ?? inferAdapterFromListItem(agent)
-  const workspace = workspaceLabel(agent)
-  const lastUsedLabel = formatRelativeTime(lastUsedAt ?? null)
-  const allowDelete = canDeleteAgent(agent)
-  const allowRename = canRenameAgent(agent)
-
-  const handleChat = () => navigate(`/agents/${agent.agentId}`)
-  const handleCopyId = async () => {
-    try {
-      await navigator.clipboard.writeText(agent.agentId)
-      toast.success('Agent id copied')
-    } catch {
-      toast.error('Could not copy agent id')
-    }
-  }
-
   return (
     <div
       className={cn(
-        'group rounded-xl border border-border bg-card p-4 shadow-sm transition-all',
-        'hover:border-[var(--accent-orange)]/50 hover:shadow-sm',
+        // Layout-stable hover. No translate, no shadow change — both
+        // visibly perturb neighbouring rows. Only the border tint
+        // shifts on hover, and the rail's vertical rhythm stays
+        // exactly the same in every state.
+        'group rounded-xl border bg-card p-4 shadow-sm transition-colors',
+        data.status === 'working'
+          ? 'border-[var(--accent-orange)]/40'
+          : data.status === 'error'
+            ? 'border-destructive/40'
+            : 'border-border hover:border-[var(--accent-orange)]/30',
       )}
     >
       <div className="flex items-start gap-4">
-        {/* Adapter tile + liveness dot in the corner. */}
-        <div className="relative shrink-0">
-          <div className="flex h-12 w-12 items-center justify-center rounded-xl bg-muted text-muted-foreground">
-            <AdapterIcon adapter={adapterId} className="h-6 w-6" />
-          </div>
-          <LivenessDot
-            status={status}
-            detail={livenessDetail(status, lastUsedAt)}
-            className="absolute -right-0.5 -bottom-0.5"
-          />
-        </div>
+        <AgentTile
+          adapter={data.adapter}
+          status={data.status}
+          lastUsedAt={data.lastUsedAt}
+        />
 
         <div className="min-w-0 flex-1">
-          <div className="mb-1 flex items-center gap-2">
-            <span className="truncate font-semibold">{displayName(agent)}</span>
-            {status === 'working' && (
-              <Badge
-                variant="secondary"
-                className="bg-amber-50 text-amber-900 hover:bg-amber-50"
-              >
-                Working
-              </Badge>
-            )}
-            {status === 'asleep' && (
-              <Badge variant="outline" className="text-muted-foreground">
-                Asleep
-              </Badge>
-            )}
-            {status === 'error' && (
-              <Badge variant="destructive">Attention</Badge>
-            )}
-          </div>
+          <AgentTitleRow
+            agent={data.agent}
+            status={data.status}
+            pinned={data.pinned}
+            turnsByDay={data.turnsByDay}
+            failedByDay={data.failedByDay}
+            onPinToggle={(next) => onPinToggle(data.agent, next)}
+          />
 
-          <div className="mb-2 flex flex-wrap items-center gap-1.5 text-xs">
-            <Badge variant="secondary" className="font-normal">
-              {adapterLabel(adapterId)}
-            </Badge>
-            {agent.modelLabel && agent.modelLabel !== 'default' && (
-              <Badge variant="outline" className="font-normal">
-                {agent.modelLabel}
-              </Badge>
-            )}
-            {reasoningEffort && reasoningEffort !== 'medium' && (
-              <Badge variant="outline" className="font-normal">
-                {reasoningEffort}
-              </Badge>
-            )}
-          </div>
+          <AgentSummaryChips
+            adapter={data.adapter}
+            modelLabel={data.modelLabel}
+            reasoningEffort={data.reasoningEffort}
+            adapterHealth={data.adapterHealth}
+          />
 
-          <div className="flex flex-wrap items-center gap-2 text-muted-foreground text-xs">
-            <span>Last used {lastUsedLabel}</span>
-            {workspace && (
-              <>
-                <span aria-hidden>•</span>
-                <span className="truncate font-mono" title={workspace}>
-                  {workspace}
-                </span>
-              </>
-            )}
-          </div>
+          <AgentLastMessage message={data.lastUserMessage} />
+
+          <AgentMetaRow lastUsedAt={data.lastUsedAt} tokens={data.tokens} />
+
+          {data.status === 'error' && data.lastError && (
+            <AgentErrorPanel
+              agentId={data.agent.agentId}
+              message={data.lastError}
+              errorAt={data.lastErrorAt}
+            />
+          )}
         </div>
 
-        <div className="flex shrink-0 items-center gap-2">
-          <Button variant="outline" size="sm" onClick={handleChat}>
-            <MessageSquare className="mr-1.5 h-3 w-3" />
-            Chat
-          </Button>
-          <DropdownMenu>
-            <DropdownMenuTrigger asChild>
-              <Button
-                variant="ghost"
-                size="icon"
-                aria-label={`More actions for ${displayName(agent)}`}
-                className="h-8 w-8"
-              >
-                <MoreHorizontal className="h-4 w-4" />
-              </Button>
-            </DropdownMenuTrigger>
-            <DropdownMenuContent align="end" className="w-44">
-              <DropdownMenuItem onSelect={() => void handleCopyId()}>
-                <Copy className="mr-2 h-3.5 w-3.5" />
-                Copy id
-              </DropdownMenuItem>
-              <RenameMenuItem disabled={!allowRename} />
-              <ResetHistoryMenuItem />
-              <DropdownMenuSeparator />
-              <DropdownMenuItem
-                onSelect={() => onDelete(agent)}
-                disabled={!allowDelete || deleting}
-                className="text-destructive focus:text-destructive"
-              >
-                {deleting ? (
-                  <Loader2 className="mr-2 h-3.5 w-3.5 animate-spin" />
-                ) : (
-                  <Trash2 className="mr-2 h-3.5 w-3.5" />
-                )}
-                Delete
-              </DropdownMenuItem>
-            </DropdownMenuContent>
-          </DropdownMenu>
-        </div>
+        <AgentActions
+          agent={data.agent}
+          activeTurnId={data.activeTurnId}
+          deleting={deleting}
+          onDelete={onDelete}
+        />
       </div>
     </div>
   )
 }
-
-const RenameMenuItem: FC<{ disabled: boolean }> = ({ disabled }) => {
-  const item = (
-    <DropdownMenuItem disabled className="text-muted-foreground">
-      <Pencil className="mr-2 h-3.5 w-3.5" />
-      Rename
-    </DropdownMenuItem>
-  )
-  if (!disabled) return item
-  // Disabled but with a hint so users know it's coming, not broken.
-  return (
-    <TooltipProvider delayDuration={300}>
-      <Tooltip>
-        <TooltipTrigger asChild>
-          <span className="block w-full">{item}</span>
-        </TooltipTrigger>
-        <TooltipContent side="left" className="text-xs">
-          Rename coming soon
-        </TooltipContent>
-      </Tooltip>
-    </TooltipProvider>
-  )
-}
-
-const ResetHistoryMenuItem: FC = () => {
-  const item = (
-    <DropdownMenuItem disabled className="text-muted-foreground">
-      <RotateCcw className="mr-2 h-3.5 w-3.5" />
-      Reset history
-    </DropdownMenuItem>
-  )
-  return (
-    <TooltipProvider delayDuration={300}>
-      <Tooltip>
-        <TooltipTrigger asChild>
-          <span className="block w-full">{item}</span>
-        </TooltipTrigger>
-        <TooltipContent side="left" className="text-xs">
-          Reset history coming soon
-        </TooltipContent>
-      </Tooltip>
-    </TooltipProvider>
-  )
-}
-
-function inferAdapterFromListItem(
-  agent: AgentListItem,
-): HarnessAgentAdapter | 'unknown' {
-  const label = agent.runtimeLabel?.toLowerCase()
-  if (label?.includes('claude')) return 'claude'
-  if (label?.includes('codex')) return 'codex'
-  if (label?.includes('openclaw')) return 'openclaw'
-  return 'unknown'
-}
-
-function livenessDetail(
-  status: AgentLiveness,
-  lastUsedAt: number | null | undefined,
-): string | undefined {
-  if (lastUsedAt == null) return undefined
-  const diffMin = Math.floor((Date.now() - lastUsedAt) / 60_000)
-  if (status === 'idle') return `Idle for ${Math.max(0, diffMin)} min`
-  if (status === 'asleep') {
-    if (diffMin < 60) return `Asleep — quiet for ${diffMin} min`
-    const hr = Math.floor(diffMin / 60)
-    return `Asleep — quiet for ${hr} hr`
-  }
-  if (status === 'working') return 'Working on a turn'
-  if (status === 'error') return 'Attention — last turn failed'
-  return undefined
-}

packages/browseros-agent/apps/agent/entrypoints/app/agents/AgentsPage.tsx

@@ -44,6 +44,7 @@ import {
   useCreateHarnessAgent,
   useDeleteHarnessAgent,
   useHarnessAgents,
+  useUpdateHarnessAgent,
 } from './useAgents'
 import { useOpenClawAgents, useOpenClawMutations } from './useOpenClaw'
 
@@ -76,6 +77,7 @@ export const AgentsPage: FC = () => {
   } = useOpenClawAgents(openClawAgentsEnabled)
   const createHarnessAgent = useCreateHarnessAgent()
   const deleteHarnessAgent = useDeleteHarnessAgent()
+  const updateHarnessAgent = useUpdateHarnessAgent()
   const {
     setupOpenClaw,
     createAgent: createOpenClawAgent,
@@ -342,12 +344,24 @@ export const AgentsPage: FC = () => {
           agents={agentListItems}
           activity={agentActivity}
           harnessAgentLookup={harnessAgentLookup}
+          adapters={adapters}
           loading={agentsLoading}
           deletingAgentKey={deletingAgent ? deletingAgentKey : null}
           onCreateAgent={() => setCreateOpen(true)}
           onDeleteAgent={(agent) => {
             void handleDelete(agent)
           }}
+          onPinToggle={(agent, next) => {
+            // Optimistic mutation; harness-only — gateway-original
+            // OpenClaw entries are gated server-side via the harness
+            // backfill, so we only fire when the row maps to a
+            // harness agent record.
+            if (!harnessAgentLookup.has(agent.agentId)) return
+            updateHarnessAgent.mutate({
+              agentId: agent.agentId,
+              patch: { pinned: next },
+            })
+          }}
         />
 
         <SetupOpenClawDialog

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-display.helpers.ts

@@ -1,4 +1,5 @@
 import type { AgentListItem } from './agents-page-types'
+import type { AgentLiveness } from './LivenessDot'
 
 /**
  * Display rules for the redesigned agent rows. Pure helpers — no React,
@@ -82,3 +83,25 @@ export function formatRelativeTime(epochMs: number | null): string {
   const d = Math.floor(diff / ONE_DAY)
   return d === 1 ? '1 day ago' : `${d} days ago`
 }
+
+/**
+ * Tooltip-friendly description of a row's current liveness state.
+ * Returns `undefined` when the state has nothing extra to add (e.g.
+ * `unknown` with no timestamp).
+ */
+export function livenessDetail(
+  status: AgentLiveness,
+  lastUsedAt: number | null | undefined,
+): string | undefined {
+  if (lastUsedAt == null) return undefined
+  const diffMin = Math.floor((Date.now() - lastUsedAt) / 60_000)
+  if (status === 'idle') return `Idle for ${Math.max(0, diffMin)} min`
+  if (status === 'asleep') {
+    if (diffMin < 60) return `Asleep — quiet for ${diffMin} min`
+    const hr = Math.floor(diffMin / 60)
+    return `Asleep — quiet for ${hr} hr`
+  }
+  if (status === 'working') return 'Working on a turn'
+  if (status === 'error') return 'Attention — last turn failed'
+  return undefined
+}

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-harness-types.ts

@@ -56,6 +56,43 @@ export interface HarnessAgent {
    * agents. Drives the recency sort and the "Last used X min ago" copy.
    */
   lastUsedAt?: number | null
+  /** Pinned agents float to the top of the list. Defaults to `false`. */
+  pinned?: boolean
+  /** First non-blank line of the most recent user message; null if none. */
+  lastUserMessage?: string | null
+  /** Working directory the agent runs in; null when no session record yet. */
+  cwd?: string | null
+  /** Cumulative + 7-day rolling token usage; null when no record. */
+  tokens?: {
+    last7d: { input: number; output: number; requestCount: number }
+    cumulative: { input: number; output: number }
+  } | null
+  turnsByDay?: number[]
+  failedByDay?: number[]
+  lastError?: string | null
+  lastErrorAt?: number | null
+  /** When non-null, an in-flight turn this row can be resumed from. */
+  activeTurnId?: string | null
+  /** Persistent FIFO queue of messages waiting for this agent. */
+  queue?: HarnessQueuedMessage[]
+}
+
+export interface HarnessQueuedMessageAttachment {
+  mediaType: string
+  data: string
+}
+
+export interface HarnessQueuedMessage {
+  id: string
+  createdAt: number
+  message: string
+  attachments?: ReadonlyArray<HarnessQueuedMessageAttachment>
+}
+
+export interface HarnessAdapterHealth {
+  healthy: boolean
+  reason?: string
+  checkedAt: number
 }
 
 export interface HarnessAdapterDescriptor {
@@ -66,6 +103,7 @@ export interface HarnessAdapterDescriptor {
   modelControl: 'runtime-supported' | 'best-effort'
   models: Array<{ id: string; label: string; recommended?: boolean }>
   reasoningEfforts: Array<{ id: string; label: string; recommended?: boolean }>
+  health?: HarnessAdapterHealth
 }
 
 export interface CreateHarnessAgentInput {

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/AgentActions.tsx

@@ -0,0 +1,160 @@
+import {
+  Copy,
+  Loader2,
+  MessageSquare,
+  MoreHorizontal,
+  Pencil,
+  RotateCcw,
+  Trash2,
+} from 'lucide-react'
+import type { FC } from 'react'
+import { useNavigate } from 'react-router'
+import { toast } from 'sonner'
+import { Button } from '@/components/ui/button'
+import {
+  DropdownMenu,
+  DropdownMenuContent,
+  DropdownMenuItem,
+  DropdownMenuSeparator,
+  DropdownMenuTrigger,
+} from '@/components/ui/dropdown-menu'
+import {
+  Tooltip,
+  TooltipContent,
+  TooltipProvider,
+  TooltipTrigger,
+} from '@/components/ui/tooltip'
+import {
+  canDelete as canDeleteAgent,
+  canRename as canRenameAgent,
+  displayName,
+} from '../agent-display.helpers'
+import type { AgentListItem } from '../agents-page-types'
+
+interface AgentActionsProps {
+  agent: AgentListItem
+  activeTurnId: string | null
+  deleting?: boolean
+  onDelete: (agent: AgentListItem) => void
+}
+
+/**
+ * Single primary CTA per row: `Resume` (filled, accent-orange, with a
+ * pulsing dot) when an active turn exists; otherwise `Chat` (outline).
+ * Both navigate to the same place — the chat hook auto-attaches via
+ * `/chat/active` when there's a live turn — but the row signals which
+ * action the user is actually taking.
+ */
+export const AgentActions: FC<AgentActionsProps> = ({
+  agent,
+  activeTurnId,
+  deleting,
+  onDelete,
+}) => {
+  const navigate = useNavigate()
+  const allowDelete = canDeleteAgent(agent)
+  const allowRename = canRenameAgent(agent)
+
+  const handleChat = () => navigate(`/agents/${agent.agentId}`)
+  const handleCopyId = async () => {
+    try {
+      await navigator.clipboard.writeText(agent.agentId)
+      toast.success('Agent id copied')
+    } catch {
+      toast.error('Could not copy agent id')
+    }
+  }
+
+  return (
+    <div className="flex shrink-0 items-center gap-1.5">
+      {activeTurnId ? (
+        <Button
+          variant="default"
+          size="sm"
+          onClick={handleChat}
+          className="gap-2 bg-[var(--accent-orange)] text-white shadow-sm hover:bg-[var(--accent-orange)]/90"
+        >
+          <span className="relative flex size-2">
+            <span className="absolute inline-flex h-full w-full animate-ping rounded-full bg-white/70 opacity-75" />
+            <span className="relative inline-flex size-2 rounded-full bg-white" />
+          </span>
+          Resume
+        </Button>
+      ) : (
+        <Button variant="outline" size="sm" onClick={handleChat}>
+          <MessageSquare className="mr-1.5 size-3" />
+          Chat
+        </Button>
+      )}
+      <DropdownMenu>
+        <DropdownMenuTrigger asChild>
+          <Button
+            variant="ghost"
+            size="icon"
+            aria-label={`More actions for ${displayName(agent)}`}
+            className="size-8 text-muted-foreground hover:text-foreground"
+          >
+            <MoreHorizontal className="size-4" />
+          </Button>
+        </DropdownMenuTrigger>
+        <DropdownMenuContent align="end" className="w-44">
+          <DropdownMenuItem onSelect={() => void handleCopyId()}>
+            <Copy className="mr-2 size-3.5" />
+            Copy id
+          </DropdownMenuItem>
+          <ComingSoonItem
+            icon={Pencil}
+            label="Rename"
+            disabled={!allowRename}
+          />
+          <ComingSoonItem icon={RotateCcw} label="Reset history" disabled />
+          <DropdownMenuSeparator />
+          <DropdownMenuItem
+            onSelect={() => onDelete(agent)}
+            disabled={!allowDelete || deleting}
+            className="text-destructive focus:text-destructive"
+          >
+            {deleting ? (
+              <Loader2 className="mr-2 size-3.5 animate-spin" />
+            ) : (
+              <Trash2 className="mr-2 size-3.5" />
+            )}
+            Delete
+          </DropdownMenuItem>
+        </DropdownMenuContent>
+      </DropdownMenu>
+    </div>
+  )
+}
+
+interface ComingSoonItemProps {
+  icon: typeof Pencil
+  label: string
+  disabled: boolean
+}
+
+const ComingSoonItem: FC<ComingSoonItemProps> = ({
+  icon: Icon,
+  label,
+  disabled,
+}) => {
+  const item = (
+    <DropdownMenuItem disabled className="text-muted-foreground">
+      <Icon className="mr-2 size-3.5" />
+      {label}
+    </DropdownMenuItem>
+  )
+  if (!disabled) return item
+  return (
+    <TooltipProvider delayDuration={300}>
+      <Tooltip>
+        <TooltipTrigger asChild>
+          <span className="block w-full">{item}</span>
+        </TooltipTrigger>
+        <TooltipContent side="left" className="text-xs">
+          {label} coming soon
+        </TooltipContent>
+      </Tooltip>
+    </TooltipProvider>
+  )
+}

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/AgentErrorPanel.tsx

@@ -0,0 +1,96 @@
+import { AlertTriangle, ChevronDown } from 'lucide-react'
+import { type FC, useEffect, useState } from 'react'
+import { Button } from '@/components/ui/button'
+import {
+  Collapsible,
+  CollapsibleContent,
+  CollapsibleTrigger,
+} from '@/components/ui/collapsible'
+import {
+  HoverCard,
+  HoverCardContent,
+  HoverCardTrigger,
+} from '@/components/ui/hover-card'
+import { cn } from '@/lib/utils'
+import { truncate } from './agent-row.helpers'
+
+interface AgentErrorPanelProps {
+  agentId: string
+  message: string
+  errorAt: number | null
+}
+
+const STORAGE_PREFIX = 'agent-row:lastErrorSeenAt:'
+const PREVIEW_CHARS = 200
+
+export const AgentErrorPanel: FC<AgentErrorPanelProps> = ({
+  agentId,
+  message,
+  errorAt,
+}) => {
+  const storageKey = `${STORAGE_PREFIX}${agentId}`
+  // Open if we've never seen this `errorAt` for this agent. Once the
+  // user collapses the panel (or refreshes after seeing it), we mark
+  // it seen so it doesn't re-pop on every poll.
+  const [open, setOpen] = useState<boolean>(() => {
+    if (typeof window === 'undefined' || !errorAt) return true
+    const seen = Number(window.localStorage.getItem(storageKey) ?? 0)
+    return !Number.isFinite(seen) || errorAt > seen
+  })
+
+  useEffect(() => {
+    if (!open && errorAt && typeof window !== 'undefined') {
+      window.localStorage.setItem(storageKey, String(errorAt))
+    }
+  }, [open, errorAt, storageKey])
+
+  const preview = truncate(message, PREVIEW_CHARS)
+  const truncated = preview.length < message.length
+
+  return (
+    <Collapsible open={open} onOpenChange={setOpen} className="mt-3">
+      <div className="flex items-center justify-between rounded-md border border-destructive/30 bg-destructive/5 px-3 py-2">
+        <div className="flex items-center gap-2 font-medium text-destructive text-xs">
+          <AlertTriangle className="size-3.5" />
+          Last error
+        </div>
+        <CollapsibleTrigger asChild>
+          <Button
+            variant="ghost"
+            size="sm"
+            className="h-6 px-2 text-muted-foreground"
+          >
+            <span className="text-xs">{open ? 'hide' : 'show'}</span>
+            <ChevronDown
+              className={cn(
+                'ml-1 size-3 transition-transform',
+                open && 'rotate-180',
+              )}
+            />
+          </Button>
+        </CollapsibleTrigger>
+      </div>
+      <CollapsibleContent>
+        <div className="mt-1 rounded-md border-destructive/30 border-x border-b bg-destructive/5 px-3 pb-2 text-xs">
+          {truncated ? (
+            <HoverCard openDelay={300}>
+              <HoverCardTrigger asChild>
+                <span className="cursor-default font-mono text-foreground/80">
+                  {preview}…
+                </span>
+              </HoverCardTrigger>
+              <HoverCardContent
+                side="bottom"
+                className="max-w-md whitespace-pre-wrap font-mono text-xs"
+              >
+                {message}
+              </HoverCardContent>
+            </HoverCard>
+          ) : (
+            <span className="font-mono text-foreground/80">{message}</span>
+          )}
+        </div>
+      </CollapsibleContent>
+    </Collapsible>
+  )
+}

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/AgentLastMessage.tsx

@@ -0,0 +1,35 @@
+import { Quote } from 'lucide-react'
+import type { FC } from 'react'
+import { firstNonBlankLine, truncate } from './agent-row.helpers'
+
+interface AgentLastMessageProps {
+  message: string | null
+}
+
+const PREVIEW_CHARS = 110
+
+/**
+ * Inline preview of the most recent user message. Renders as a quoted,
+ * italic line so the row reads like a conversation snippet rather than
+ * a label-and-value pair. No hover-card — opening the agent's chat is
+ * the canonical way to read the full message.
+ */
+export const AgentLastMessage: FC<AgentLastMessageProps> = ({ message }) => {
+  if (!message) {
+    return (
+      <p className="mt-1 text-muted-foreground/70 text-xs italic">
+        No messages yet — start a chat
+      </p>
+    )
+  }
+  const preview = truncate(firstNonBlankLine(message), PREVIEW_CHARS)
+  return (
+    <p className="mt-1.5 flex items-start gap-1.5 text-foreground/85 text-sm italic leading-snug">
+      <Quote
+        className="mt-1 size-3 shrink-0 text-muted-foreground/60"
+        aria-hidden
+      />
+      <span className="truncate">{preview}</span>
+    </p>
+  )
+}

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/AgentMetaRow.tsx

@@ -0,0 +1,37 @@
+import type { FC } from 'react'
+import { formatRelativeTime } from '../agent-display.helpers'
+import { AgentTokenSummary } from './AgentTokenSummary'
+import type { AgentTokenUsage } from './agent-row.types'
+
+interface AgentMetaRowProps {
+  lastUsedAt: number | null
+  tokens: AgentTokenUsage | null
+}
+
+/**
+ * Bottom-of-row meta line. Intentionally sparse — last activity time
+ * and lifetime tokens. CWD is no longer surfaced here because the path
+ * the server happens to be running from isn't actionable; if a future
+ * surface needs the cwd (chat panel, debug view) it reads from the
+ * listing payload directly.
+ */
+export const AgentMetaRow: FC<AgentMetaRowProps> = ({ lastUsedAt, tokens }) => {
+  const lastUsedLabel = formatRelativeTime(lastUsedAt)
+  const tokensTotal =
+    (tokens?.cumulative.input ?? 0) + (tokens?.cumulative.output ?? 0)
+  const showTokens = tokensTotal > 0
+
+  return (
+    <div className="mt-2 flex flex-wrap items-center gap-x-2 text-muted-foreground text-xs">
+      <span>{lastUsedLabel}</span>
+      {showTokens && (
+        <>
+          <span aria-hidden className="text-muted-foreground/50">
+            ·
+          </span>
+          <AgentTokenSummary tokens={tokens} />
+        </>
+      )}
+    </div>
+  )
+}

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/AgentSparkline.tsx

@@ -0,0 +1,92 @@
+import type { FC } from 'react'
+import {
+  HoverCard,
+  HoverCardContent,
+  HoverCardTrigger,
+} from '@/components/ui/hover-card'
+import { cn } from '@/lib/utils'
+import { formatLocalDate, ROW_BAR_COUNT } from './agent-row.helpers'
+
+interface AgentSparklineProps {
+  /** 14 entries, oldest → newest. Today's bucket is the last index. */
+  turnsByDay: number[]
+  /** Same length, same order. Failed turns counted separately. */
+  failedByDay: number[]
+  className?: string
+}
+
+const MIN_BAR_HEIGHT_PX = 2
+const MAX_BAR_HEIGHT_PX = 18
+
+export const AgentSparkline: FC<AgentSparklineProps> = ({
+  turnsByDay,
+  failedByDay,
+  className,
+}) => {
+  if (turnsByDay.length === 0 || turnsByDay.every((n) => n === 0)) return null
+  const max = Math.max(1, ...turnsByDay)
+
+  return (
+    <HoverCard openDelay={250}>
+      <HoverCardTrigger asChild>
+        <div
+          role="img"
+          aria-label={`Last ${ROW_BAR_COUNT} days of activity`}
+          className={cn('flex h-5 items-end gap-px', className)}
+        >
+          {turnsByDay.map((count, idx) => {
+            const ratio = count / max
+            const height = Math.max(
+              MIN_BAR_HEIGHT_PX,
+              Math.round(ratio * MAX_BAR_HEIGHT_PX),
+            )
+            const isToday = idx === ROW_BAR_COUNT - 1
+            const failed = failedByDay[idx] ?? 0
+            return (
+              <div
+                // biome-ignore lint/suspicious/noArrayIndexKey: fixed-length sparkline buckets keyed by day position
+                key={`bar-${idx}`}
+                className={cn(
+                  'w-1.5 rounded-sm',
+                  count === 0
+                    ? 'bg-muted-foreground/15'
+                    : failed > 0
+                      ? 'bg-destructive/50'
+                      : 'bg-[var(--accent-orange)]/50',
+                  isToday && 'ring-1 ring-foreground/30',
+                )}
+                style={{ height }}
+              />
+            )
+          })}
+        </div>
+      </HoverCardTrigger>
+      <HoverCardContent side="left" className="w-56 text-xs">
+        <div className="mb-2 font-medium text-sm">Last 14 days</div>
+        <ul className="space-y-0.5">
+          {turnsByDay.map((count, idx) => {
+            const failed = failedByDay[idx] ?? 0
+            const dayLabel = formatLocalDate(idx)
+            return (
+              <li
+                // biome-ignore lint/suspicious/noArrayIndexKey: fixed-length list keyed by day position
+                key={`day-${idx}`}
+                className="flex items-center justify-between text-muted-foreground"
+              >
+                <span>{dayLabel}</span>
+                <span>
+                  {count}
+                  {failed > 0 && (
+                    <span className="ml-1 text-destructive">
+                      ({failed} failed)
+                    </span>
+                  )}
+                </span>
+              </li>
+            )
+          })}
+        </ul>
+      </HoverCardContent>
+    </HoverCard>
+  )
+}

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/AgentSummaryChips.tsx

@@ -0,0 +1,71 @@
+import { TriangleAlert } from 'lucide-react'
+import type { FC } from 'react'
+import { Badge } from '@/components/ui/badge'
+import {
+  HoverCard,
+  HoverCardContent,
+  HoverCardTrigger,
+} from '@/components/ui/hover-card'
+import { cn } from '@/lib/utils'
+import { adapterLabel } from '../AdapterIcon'
+import type { HarnessAgentAdapter } from '../agent-harness-types'
+import type { AgentAdapterHealth } from './agent-row.types'
+
+interface AgentSummaryChipsProps {
+  adapter: HarnessAgentAdapter | 'unknown'
+  modelLabel: string | null
+  reasoningEffort: string | null
+  /** When unhealthy, the adapter label dims and a warning chip appears. */
+  adapterHealth: AgentAdapterHealth | null
+}
+
+/**
+ * Adapter / model / reasoning summary line. Always rendered (so OpenClaw
+ * rows that fall back to defaults still expose what they're set up to do)
+ * and surfaces adapter-health *only when unhealthy* — keeping the calm
+ * default state silent and reserving visual noise for things the user
+ * needs to act on.
+ */
+export const AgentSummaryChips: FC<AgentSummaryChipsProps> = ({
+  adapter,
+  modelLabel,
+  reasoningEffort,
+  adapterHealth,
+}) => {
+  const parts = [adapterLabel(adapter)]
+  if (modelLabel) parts.push(modelLabel)
+  if (reasoningEffort) parts.push(reasoningEffort)
+  const unhealthy = adapterHealth?.healthy === false
+  return (
+    <div
+      className={cn(
+        'flex items-center gap-1.5 text-muted-foreground text-xs',
+        unhealthy && 'text-muted-foreground/70',
+      )}
+    >
+      <span className="truncate">{parts.join(' · ')}</span>
+      {unhealthy && adapterHealth && (
+        <HoverCard openDelay={200}>
+          <HoverCardTrigger asChild>
+            <Badge
+              variant="outline"
+              className="h-5 cursor-default gap-1 border-amber-500/40 bg-amber-50 px-1.5 text-amber-900 hover:bg-amber-50"
+            >
+              <TriangleAlert className="size-2.5" />
+              <span className="font-normal">Unavailable</span>
+            </Badge>
+          </HoverCardTrigger>
+          <HoverCardContent side="right" className="w-72 text-sm">
+            <div className="font-medium">
+              {adapterLabel(adapter)} CLI not available
+            </div>
+            <div className="mt-1 text-muted-foreground text-xs">
+              {adapterHealth.reason ??
+                'Adapter binary missing on $PATH. Install it from the adapter docs to use this agent.'}
+            </div>
+          </HoverCardContent>
+        </HoverCard>
+      )}
+    </div>
+  )
+}

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/AgentTile.tsx

@@ -0,0 +1,37 @@
+import type { FC } from 'react'
+import { cn } from '@/lib/utils'
+import { AdapterIcon } from '../AdapterIcon'
+import { livenessDetail } from '../agent-display.helpers'
+import type { HarnessAgentAdapter } from '../agent-harness-types'
+import { type AgentLiveness, LivenessDot } from '../LivenessDot'
+
+export interface AgentTileProps {
+  adapter: HarnessAgentAdapter | 'unknown'
+  status: AgentLiveness
+  lastUsedAt: number | null
+}
+
+/**
+ * Adapter glyph + a single liveness dot. Adapter health is no longer
+ * surfaced here — it lives as an inline pill inside `AgentSummaryChips`
+ * so the user isn't asked to disambiguate two dots on the same tile.
+ */
+export const AgentTile: FC<AgentTileProps> = ({
+  adapter,
+  status,
+  lastUsedAt,
+}) => (
+  <div className="relative shrink-0">
+    <div className="flex h-12 w-12 items-center justify-center rounded-xl bg-muted text-muted-foreground">
+      <AdapterIcon adapter={adapter} className="h-6 w-6" />
+    </div>
+    <LivenessDot
+      status={status}
+      detail={livenessDetail(status, lastUsedAt)}
+      className={cn(
+        'absolute -right-0.5 -bottom-0.5',
+        status === 'working' && 'animate-pulse',
+      )}
+    />
+  </div>
+)

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/AgentTitleRow.tsx

@@ -0,0 +1,55 @@
+import type { FC } from 'react'
+import { Badge } from '@/components/ui/badge'
+import { displayName } from '../agent-display.helpers'
+import type { AgentListItem } from '../agents-page-types'
+import type { AgentLiveness } from '../LivenessDot'
+import { AgentSparkline } from './AgentSparkline'
+import { PinToggle } from './PinToggle'
+
+interface AgentTitleRowProps {
+  agent: AgentListItem
+  status: AgentLiveness
+  pinned: boolean
+  turnsByDay: number[]
+  failedByDay: number[]
+  onPinToggle: (next: boolean) => void
+}
+
+/**
+ * Title strip: name + status badge + (right-aligned) sparkline. The
+ * pin toggle sits trailing the title so the title always flushes left
+ * regardless of pin state — moving the star left of the title indents
+ * the row's first line off-axis from the model/preview/meta lines
+ * below it. When unpinned and not hovered, the toggle is removed from
+ * layout entirely so it reserves no space at all.
+ */
+export const AgentTitleRow: FC<AgentTitleRowProps> = ({
+  agent,
+  status,
+  pinned,
+  turnsByDay,
+  failedByDay,
+  onPinToggle,
+}) => (
+  <div className="mb-1 flex items-center gap-2">
+    <span className="truncate font-semibold">{displayName(agent)}</span>
+    {status === 'working' && (
+      <Badge
+        variant="secondary"
+        className="bg-amber-50 text-amber-900 hover:bg-amber-50"
+      >
+        Working
+      </Badge>
+    )}
+    {status === 'asleep' && (
+      <Badge variant="outline" className="text-muted-foreground">
+        Asleep
+      </Badge>
+    )}
+    {status === 'error' && <Badge variant="destructive">Attention</Badge>}
+    <PinToggle pinned={pinned} onToggle={onPinToggle} />
+    <div className="ml-auto">
+      <AgentSparkline turnsByDay={turnsByDay} failedByDay={failedByDay} />
+    </div>
+  </div>
+)

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/AgentTokenSummary.tsx

@@ -0,0 +1,63 @@
+import type { FC } from 'react'
+import {
+  HoverCard,
+  HoverCardContent,
+  HoverCardTrigger,
+} from '@/components/ui/hover-card'
+import { Progress } from '@/components/ui/progress'
+import { formatTokens } from './agent-row.helpers'
+import type { AgentTokenUsage } from './agent-row.types'
+
+interface AgentTokenSummaryProps {
+  tokens: AgentTokenUsage | null
+}
+
+/**
+ * Inline token total + a HoverCard breakdown. Surfaces lifetime tokens
+ * (the only window we can compute reliably from the session record).
+ * Per-window stats land in a follow-up once the activity ledger ships.
+ */
+export const AgentTokenSummary: FC<AgentTokenSummaryProps> = ({ tokens }) => {
+  if (!tokens) return null
+  const { input, output } = tokens.cumulative
+  const total = input + output
+  if (total === 0) return null
+  const inputPct = (input / total) * 100
+
+  return (
+    <HoverCard openDelay={200}>
+      <HoverCardTrigger asChild>
+        <span className="cursor-default text-muted-foreground tabular-nums transition-colors hover:text-foreground">
+          {formatTokens(total)} tokens
+        </span>
+      </HoverCardTrigger>
+      <HoverCardContent side="top" align="end" className="w-72 text-sm">
+        <div className="mb-3 flex items-center justify-between">
+          <span className="font-medium">Lifetime tokens</span>
+          <span className="text-muted-foreground text-xs tabular-nums">
+            {formatTokens(total)} total
+          </span>
+        </div>
+
+        <div className="space-y-2">
+          <div className="flex items-center justify-between text-xs">
+            <span className="text-muted-foreground">Input</span>
+            <span className="tabular-nums">{formatTokens(input)}</span>
+          </div>
+          <Progress value={inputPct} className="h-1.5" />
+
+          <div className="mt-2 flex items-center justify-between text-xs">
+            <span className="text-muted-foreground">Output</span>
+            <span className="tabular-nums">{formatTokens(output)}</span>
+          </div>
+          <Progress value={100 - inputPct} className="h-1.5" />
+        </div>
+
+        <p className="mt-3 border-t pt-2 text-muted-foreground text-xs leading-snug">
+          Cumulative across every turn this agent has run. Per-window stats
+          arrive in a future release.
+        </p>
+      </HoverCardContent>
+    </HoverCard>
+  )
+}

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/PinToggle.tsx

@@ -0,0 +1,60 @@
+import { Star } from 'lucide-react'
+import type { FC } from 'react'
+import { Button } from '@/components/ui/button'
+import {
+  Tooltip,
+  TooltipContent,
+  TooltipProvider,
+  TooltipTrigger,
+} from '@/components/ui/tooltip'
+import { cn } from '@/lib/utils'
+
+interface PinToggleProps {
+  pinned: boolean
+  onToggle: (next: boolean) => void
+}
+
+/**
+ * Trailing star toggle. The button is *always rendered* — only its
+ * opacity changes between pinned/unpinned/hover states — so the title
+ * row's height is constant. Hiding the slot via `display: none` would
+ * collapse the row's vertical metrics on hover and shift every card
+ * below in the rail.
+ *
+ * Placement is trailing the title (after the status badge) so the
+ * title itself flushes left regardless of pin state — leading the
+ * row with the star would indent the title relative to the model /
+ * preview / meta lines beneath it.
+ */
+export const PinToggle: FC<PinToggleProps> = ({ pinned, onToggle }) => (
+  <TooltipProvider delayDuration={300}>
+    <Tooltip>
+      <TooltipTrigger asChild>
+        <Button
+          variant="ghost"
+          size="icon"
+          className={cn(
+            'size-6 text-muted-foreground transition-opacity hover:text-foreground',
+            pinned ? 'opacity-100' : 'opacity-0 group-hover:opacity-100',
+          )}
+          aria-pressed={pinned}
+          aria-label={pinned ? 'Unpin agent' : 'Pin agent'}
+          onClick={(event) => {
+            event.stopPropagation()
+            onToggle(!pinned)
+          }}
+        >
+          <Star
+            className={cn(
+              'size-3.5',
+              pinned && 'fill-amber-400 text-amber-500',
+            )}
+          />
+        </Button>
+      </TooltipTrigger>
+      <TooltipContent side="top" className="text-xs">
+        {pinned ? 'Unpin' : 'Pin to top'}
+      </TooltipContent>
+    </Tooltip>
+  </TooltipProvider>
+)

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/agent-row.helpers.test.ts

@@ -0,0 +1,73 @@
+import { describe, expect, it } from 'bun:test'
+import {
+  firstNonBlankLine,
+  formatLocalDate,
+  formatTokens,
+  ROW_BAR_COUNT,
+  truncate,
+} from './agent-row.helpers'
+
+describe('formatTokens', () => {
+  it('renders zero / NaN as "0"', () => {
+    expect(formatTokens(0)).toBe('0')
+    expect(formatTokens(Number.NaN)).toBe('0')
+  })
+
+  it('renders sub-1K as integer', () => {
+    expect(formatTokens(142)).toBe('142')
+  })
+
+  it('renders K with one decimal under 10', () => {
+    expect(formatTokens(8_400)).toBe('8.4K')
+  })
+
+  it('drops the decimal at >=10K', () => {
+    expect(formatTokens(120_000)).toBe('120K')
+  })
+
+  it('renders M with one decimal under 10', () => {
+    expect(formatTokens(1_200_000)).toBe('1.2M')
+  })
+})
+
+describe('firstNonBlankLine', () => {
+  it('returns the first non-blank line', () => {
+    expect(firstNonBlankLine('\n\nhello\nworld')).toBe('hello')
+  })
+
+  it('skips USER_QUERY envelope tags', () => {
+    expect(firstNonBlankLine('<USER_QUERY>\nfix tests\n</USER_QUERY>')).toBe(
+      'fix tests',
+    )
+  })
+
+  it('falls back to the trimmed input when nothing matches', () => {
+    expect(firstNonBlankLine('   single   ')).toBe('single')
+  })
+})
+
+describe('truncate', () => {
+  it('returns input unchanged when within limit', () => {
+    expect(truncate('hello', 10)).toBe('hello')
+  })
+
+  it('appends an ellipsis when over limit', () => {
+    expect(truncate('hello world', 6)).toBe('hello…')
+  })
+})
+
+describe('formatLocalDate', () => {
+  const today = new Date('2026-04-30T12:00:00Z')
+
+  it('labels today and yesterday explicitly', () => {
+    expect(formatLocalDate(ROW_BAR_COUNT - 1, today)).toBe('today')
+    expect(formatLocalDate(ROW_BAR_COUNT - 2, today)).toBe('yesterday')
+  })
+
+  it('returns a "Mon D" format for older days', () => {
+    const label = formatLocalDate(0, today)
+    // "Apr 17" or "Apr 17," depending on locale; just assert it
+    // contains a month abbreviation and a day number.
+    expect(label).toMatch(/[A-Za-z]+ \d+/)
+  })
+})

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/agent-row.helpers.ts

@@ -0,0 +1,64 @@
+/**
+ * Pure formatters consumed by row sub-components. Kept distinct from
+ * `agent-display.helpers.ts` (page-level helpers) so the row internals
+ * have an obvious single home.
+ */
+
+const TOKEN_THRESHOLDS: Array<[number, string]> = [
+  [1_000_000, 'M'],
+  [1_000, 'K'],
+]
+
+/** `1.2M`, `820K`, `8.4K`, `142`, `0`. */
+export function formatTokens(n: number): string {
+  if (!Number.isFinite(n) || n <= 0) return '0'
+  for (const [threshold, suffix] of TOKEN_THRESHOLDS) {
+    if (n >= threshold) {
+      const value = n / threshold
+      const decimal = value < 10 ? value.toFixed(1) : value.toFixed(0)
+      return `${decimal}${suffix}`
+    }
+  }
+  return String(Math.round(n))
+}
+
+const USER_QUERY_OPEN = /^<USER_QUERY>$/i
+const USER_QUERY_CLOSE = /^<\/USER_QUERY>$/i
+
+/**
+ * First non-blank line, with the BrowserOS user-system-prompt
+ * `<USER_QUERY>` envelope tags stripped so previews don't show
+ * structural noise.
+ */
+export function firstNonBlankLine(text: string): string {
+  const lines = text.split('\n').map((line) => line.trim())
+  for (const line of lines) {
+    if (!line) continue
+    if (USER_QUERY_OPEN.test(line) || USER_QUERY_CLOSE.test(line)) continue
+    return line
+  }
+  return text.trim()
+}
+
+export function truncate(text: string, max: number): string {
+  if (text.length <= max) return text
+  return `${text.slice(0, max - 1).trimEnd()}…`
+}
+
+const SPARKLINE_DAYS = 14
+
+/**
+ * "today" / "yesterday" / "Apr 17" — given an index 0..13 from
+ * oldest → newest. `today` defaults to `new Date()` so callers don't
+ * have to thread a clock through.
+ */
+export function formatLocalDate(idx: number, today: Date = new Date()): string {
+  if (idx === SPARKLINE_DAYS - 1) return 'today'
+  if (idx === SPARKLINE_DAYS - 2) return 'yesterday'
+  const offset = SPARKLINE_DAYS - 1 - idx
+  const date = new Date(today)
+  date.setDate(date.getDate() - offset)
+  return date.toLocaleDateString(undefined, { month: 'short', day: 'numeric' })
+}
+
+export const ROW_BAR_COUNT = SPARKLINE_DAYS

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/agent-row.types.ts

@@ -0,0 +1,51 @@
+import type { HarnessAgentAdapter } from '../agent-harness-types'
+import type { AgentListItem } from '../agents-page-types'
+import type { AgentLiveness } from '../LivenessDot'
+
+/**
+ * Window-bounded token usage. Server returns `null` when no session
+ * record exists yet for the agent.
+ */
+export interface AgentTokenUsage {
+  last7d: { input: number; output: number; requestCount: number }
+  cumulative: { input: number; output: number }
+}
+
+export interface AgentAdapterHealth {
+  healthy: boolean
+  reason?: string
+}
+
+/**
+ * Everything an `AgentRowCard` needs to render. Mirrors the shape
+ * `useHarnessAgents` exposes; the page assembles one entry per row in
+ * `AgentList` and passes it down. Sub-components only see slices of
+ * this object — no prop drilling beyond two levels.
+ */
+export interface AgentRowData {
+  agent: AgentListItem
+  adapter: HarnessAgentAdapter | 'unknown'
+  modelLabel: string | null
+  reasoningEffort: string | null
+  status: AgentLiveness
+  lastUsedAt: number | null
+  pinned: boolean
+  cwd: string | null
+  lastUserMessage: string | null
+  tokens: AgentTokenUsage | null
+  /** 14 entries, oldest → newest. Today is the last index. */
+  turnsByDay: number[]
+  /** Same length and ordering as `turnsByDay`. */
+  failedByDay: number[]
+  lastError: string | null
+  lastErrorAt: number | null
+  /** When non-null, an in-flight turn this row can be resumed from. */
+  activeTurnId: string | null
+  /** Adapter-level health, shared across rows for the same adapter. */
+  adapterHealth: AgentAdapterHealth | null
+}
+
+export interface AgentRowCallbacks {
+  onDelete: (agent: AgentListItem) => void
+  onPinToggle: (agent: AgentListItem, next: boolean) => void
+}

packages/browseros-agent/apps/agent/entrypoints/app/agents/useAgents.ts

@@ -8,6 +8,7 @@ import {
   type HarnessAdapterDescriptor,
   type HarnessAgent,
   type HarnessAgentHistoryPage,
+  type HarnessQueuedMessage,
   mapHarnessAgentToEntry,
 } from './agent-harness-types'
 import type { OpenClawStatus } from './useOpenClaw'
@@ -135,6 +136,63 @@ export function useCreateHarnessAgent() {
   })
 }
 
+/**
+ * Apply a partial update to a harness agent. Used by the pin-toggle
+ * star and (eventually) the inline rename UI. Optimistically writes
+ * the patch into the listing query cache so the row updates instantly,
+ * then rolls back if the server rejects the change.
+ */
+export function useUpdateHarnessAgent() {
+  const { baseUrl, isLoading: urlLoading } = useAgentServerUrl()
+  const queryClient = useQueryClient()
+
+  return useMutation({
+    mutationFn: async (input: {
+      agentId: string
+      patch: { name?: string; pinned?: boolean }
+    }) => {
+      if (!baseUrl || urlLoading) {
+        throw new Error('BrowserOS agent server URL is not ready')
+      }
+      const data = await agentsFetch<{ agent: HarnessAgent }>(
+        baseUrl,
+        `/${encodeURIComponent(input.agentId)}`,
+        {
+          method: 'PATCH',
+          headers: { 'Content-Type': 'application/json' },
+          body: JSON.stringify(input.patch),
+        },
+      )
+      return data.agent
+    },
+    onMutate: async ({ agentId, patch }) => {
+      const queryKey = [AGENT_QUERY_KEYS.agents, baseUrl]
+      await queryClient.cancelQueries({ queryKey })
+      const previous = queryClient.getQueryData<HarnessAgentsResponse>(queryKey)
+      if (!previous) return { previous: undefined }
+      queryClient.setQueryData<HarnessAgentsResponse>(queryKey, {
+        ...previous,
+        agents: previous.agents.map((agent) =>
+          agent.id === agentId ? { ...agent, ...patch } : agent,
+        ),
+      })
+      return { previous }
+    },
+    onError: (_err, _vars, context) => {
+      if (!context?.previous) return
+      queryClient.setQueryData(
+        [AGENT_QUERY_KEYS.agents, baseUrl],
+        context.previous,
+      )
+    },
+    onSettled: async () => {
+      await queryClient.invalidateQueries({
+        queryKey: [AGENT_QUERY_KEYS.agents],
+      })
+    },
+  })
+}
+
 export function useDeleteHarnessAgent() {
   const { baseUrl, isLoading: urlLoading } = useAgentServerUrl()
   const queryClient = useQueryClient()
@@ -206,6 +264,8 @@ export interface HarnessActiveTurnInfo {
   lastSeq: number
   startedAt: number
   endedAt?: number
+  /** User message that kicked off the turn; null when not captured. */
+  prompt: string | null
 }
 
 /**
@@ -260,3 +320,145 @@ export async function fetchHarnessAgentHistory(
     `/${encodeURIComponent(agentId)}/sessions/main/history`,
   )
 }
+
+export interface EnqueueMessageInput {
+  message: string
+  attachments?: ReadonlyArray<unknown>
+}
+
+export async function enqueueHarnessMessage(
+  agentId: string,
+  input: EnqueueMessageInput,
+): Promise<HarnessQueuedMessage> {
+  const baseUrl = await getAgentServerUrl()
+  const response = await fetch(
+    `${baseUrl}/agents/${encodeURIComponent(agentId)}/queue`,
+    {
+      method: 'POST',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify({
+        message: input.message,
+        ...(input.attachments && input.attachments.length > 0
+          ? { attachments: input.attachments }
+          : {}),
+      }),
+    },
+  )
+  if (!response.ok) {
+    let message = `Request failed with status ${response.status}`
+    try {
+      const body = (await response.json()) as { error?: string }
+      if (body.error) message = body.error
+    } catch {}
+    throw new Error(message)
+  }
+  const body = (await response.json()) as { queued: HarnessQueuedMessage }
+  return body.queued
+}
+
+export async function removeHarnessQueuedMessage(
+  agentId: string,
+  messageId: string,
+): Promise<{ removed: boolean }> {
+  const baseUrl = await getAgentServerUrl()
+  const response = await fetch(
+    `${baseUrl}/agents/${encodeURIComponent(agentId)}/queue/${encodeURIComponent(
+      messageId,
+    )}`,
+    { method: 'DELETE' },
+  )
+  if (!response.ok) return { removed: false }
+  return (await response.json()) as { removed: boolean }
+}
+
+/**
+ * Optimistic enqueue: writes the new queued message into the listing
+ * cache immediately so the queue panel reflects the change without
+ * waiting for the next poll. Rolls back if the server rejects.
+ */
+export function useEnqueueHarnessMessage() {
+  const { baseUrl } = useAgentServerUrl()
+  const queryClient = useQueryClient()
+
+  return useMutation({
+    mutationFn: async (input: { agentId: string } & EnqueueMessageInput) =>
+      enqueueHarnessMessage(input.agentId, input),
+    onMutate: async (input) => {
+      const queryKey = [AGENT_QUERY_KEYS.agents, baseUrl]
+      await queryClient.cancelQueries({ queryKey })
+      const previous = queryClient.getQueryData<HarnessAgentsResponse>(queryKey)
+      if (!previous) return { previous: undefined }
+      const optimistic: HarnessQueuedMessage = {
+        id: `optimistic-${Math.random().toString(36).slice(2, 10)}`,
+        createdAt: Date.now(),
+        message: input.message,
+      }
+      queryClient.setQueryData<HarnessAgentsResponse>(queryKey, {
+        ...previous,
+        agents: previous.agents.map((agent) =>
+          agent.id === input.agentId
+            ? { ...agent, queue: [...(agent.queue ?? []), optimistic] }
+            : agent,
+        ),
+      })
+      return { previous }
+    },
+    onError: (_err, _vars, context) => {
+      if (!context?.previous) return
+      queryClient.setQueryData(
+        [AGENT_QUERY_KEYS.agents, baseUrl],
+        context.previous,
+      )
+    },
+    onSettled: async () => {
+      await queryClient.invalidateQueries({
+        queryKey: [AGENT_QUERY_KEYS.agents],
+      })
+    },
+  })
+}
+
+/**
+ * Optimistic queue removal mirror of `useEnqueueHarnessMessage`.
+ */
+export function useRemoveHarnessQueuedMessage() {
+  const { baseUrl } = useAgentServerUrl()
+  const queryClient = useQueryClient()
+
+  return useMutation({
+    mutationFn: async (input: { agentId: string; messageId: string }) =>
+      removeHarnessQueuedMessage(input.agentId, input.messageId),
+    onMutate: async (input) => {
+      const queryKey = [AGENT_QUERY_KEYS.agents, baseUrl]
+      await queryClient.cancelQueries({ queryKey })
+      const previous = queryClient.getQueryData<HarnessAgentsResponse>(queryKey)
+      if (!previous) return { previous: undefined }
+      queryClient.setQueryData<HarnessAgentsResponse>(queryKey, {
+        ...previous,
+        agents: previous.agents.map((agent) =>
+          agent.id === input.agentId
+            ? {
+                ...agent,
+                queue: (agent.queue ?? []).filter(
+                  (entry) => entry.id !== input.messageId,
+                ),
+              }
+            : agent,
+        ),
+      })
+      return { previous }
+    },
+    onError: (_err, _vars, context) => {
+      if (!context?.previous) return
+      queryClient.setQueryData(
+        [AGENT_QUERY_KEYS.agents, baseUrl],
+        context.previous,
+      )
+    },
+    onSettled: async () => {
+      await queryClient.invalidateQueries({
+        queryKey: [AGENT_QUERY_KEYS.agents],
+      })
+    },
+  })
+}

packages/browseros-agent/apps/agent/lib/agent-conversations/types.ts

@@ -59,15 +59,3 @@ export interface AgentConversation {
   createdAt: number
   updatedAt: number
 }
-
-export interface AgentCardData {
-  agentId: string
-  name: string
-  model?: string
-  status: 'idle' | 'working' | 'error'
-  lastMessage?: string
-  lastMessageTimestamp?: number
-  activitySummary?: string
-  currentTool?: string
-  costUsd?: number
-}

packages/browseros-agent/apps/agent/package.json

@@ -9,6 +9,7 @@
     "build": "bun run codegen && wxt build",
     "build:dev": "bun --env-file=.env.development wxt build --mode development",
     "zip": "wxt zip",
+    "test": "bun run ../../scripts/run-bun-test.ts ./apps/agent",
     "compile": "bun --env-file=.env.development wxt prepare && tsgo --noEmit",
     "lint": "bunx biome check",
     "typecheck": "bun --env-file=.env.development wxt prepare && tsgo --noEmit",

packages/browseros-agent/apps/eval/.env.example

@@ -0,0 +1,51 @@
+# Copy to .env.development for local eval runs.
+
+# Provider keys used by existing config files.
+OPENROUTER_API_KEY=
+FIREWORKS_API_KEY=
+ANTHROPIC_API_KEY=
+OPENAI_API_KEY=
+GOOGLE_GENERATIVE_AI_API_KEY=
+
+# Claude Agent SDK token used by performance_grader.
+CLAUDE_CODE_OAUTH_TOKEN=
+
+# Suite-mode model selection.
+EVAL_VARIANT=local
+EVAL_AGENT_PROVIDER=openai-compatible
+EVAL_AGENT_MODEL=
+EVAL_AGENT_API_KEY=
+EVAL_AGENT_BASE_URL=
+EVAL_AGENT_SUPPORTS_IMAGES=true
+
+# Optional suite-mode executor override for orchestrator suites.
+EVAL_EXECUTOR_MODEL=
+EVAL_EXECUTOR_API_KEY=
+EVAL_EXECUTOR_BASE_URL=
+
+# Clado visual action executor.
+CLADO_ACTION_MODEL=
+CLADO_ACTION_API_KEY=
+CLADO_ACTION_BASE_URL=
+# Backward-compatible alias used by older local scripts.
+CLADO_ACTION_URL=
+
+# BrowserOS runner.
+BROWSEROS_BINARY=/Applications/BrowserOS.app/Contents/MacOS/BrowserOS
+BROWSEROS_SERVER_URL=http://127.0.0.1:9110
+BROWSEROS_SERVER_LOG_DIR=/tmp/browseros-server-logs
+BROWSEROS_CONFIG_URL=
+
+# Captcha solver extension.
+NOPECHA_API_KEY=
+
+# WebArena-Infinity.
+WEBARENA_INFINITY_DIR=
+INFINITY_APP_URL=
+
+# R2 publishing and weekly report.
+EVAL_R2_ACCOUNT_ID=
+EVAL_R2_ACCESS_KEY_ID=
+EVAL_R2_SECRET_ACCESS_KEY=
+EVAL_R2_BUCKET=browseros-eval
+EVAL_R2_CDN_BASE_URL=https://eval.browseros.com

packages/browseros-agent/apps/eval/README.md

@@ -14,6 +14,7 @@ Evaluation framework for BrowserOS browser automation agents. Runs tasks from st
 
 ```bash
 cd apps/eval
+cp .env.example .env.development
 # Edit .env.development with your keys, then:
 bun run eval
 ```
@@ -23,11 +24,55 @@ Opens the eval dashboard at `http://localhost:9900` in config mode. From there:
 ### CLI mode
 
 ```bash
-bun run eval -c configs/browseros-agent-weekly.json
+bun run eval -c configs/legacy/browseros-agent-weekly.json
+bun run eval suite --config configs/legacy/browseros-agent-weekly.json --publish r2
 ```
 
 Runs immediately. Dashboard still available at `http://localhost:9900` for live progress.
 
+The `suite` command is the workflow-compatible full loop: execute tasks, run graders, write artifacts, and optionally publish to R2. The old `-c` form remains supported during migration.
+
+```bash
+bun run eval run --config configs/legacy/browseros-agent-weekly.json
+bun run eval suite --suite configs/suites/agisdk-daily-10.json --variant kimi-fireworks --publish r2
+bun run eval grade --run results/browseros-agent-weekly/2026-04-29-1430
+bun run eval publish --run results/browseros-agent-weekly/2026-04-29-1430 --target r2
+```
+
+Config files live in two groups:
+
+```txt
+configs/legacy/  # Complete EvalConfig files used by older workflows and the dashboard
+configs/suites/  # Suite definitions; model/provider comes from CLI flags or env
+```
+
+Suite mode takes model settings from CLI flags first, then env:
+
+```bash
+EVAL_VARIANT=kimi-fireworks \
+EVAL_AGENT_PROVIDER=openai-compatible \
+EVAL_AGENT_MODEL=accounts/fireworks/models/kimi-k2p5 \
+EVAL_AGENT_API_KEY=$FIREWORKS_API_KEY \
+EVAL_AGENT_BASE_URL=https://api.fireworks.ai/inference/v1 \
+bun run eval suite --suite configs/suites/agisdk-daily-10.json --publish r2
+```
+
+### Suites and variants
+
+A **suite** is what we run: the task dataset, graders, worker count, timeout, and browser settings. For example, `agisdk-daily-10` means "run these 10 AGI SDK tasks and grade them with `agisdk_state_diff`."
+
+A **variant** is the model setup we are testing on that suite. `EVAL_VARIANT` is just the human-readable name for that setup. The actual model connection still comes from `EVAL_AGENT_PROVIDER`, `EVAL_AGENT_MODEL`, `EVAL_AGENT_API_KEY`, and `EVAL_AGENT_BASE_URL`.
+
+This lets us run the same suite against multiple model setups without copying the benchmark config:
+
+```txt
+agisdk-daily-10 + kimi-fireworks
+agisdk-daily-10 + claude-sonnet
+agisdk-daily-10 + clado-action-000159
+```
+
+For `orchestrator-executor` suites, there can also be an executor model/backend. The `EVAL_AGENT_*` vars describe the main agent or orchestrator. The optional `EVAL_EXECUTOR_*` or `CLADO_ACTION_*` vars describe the delegated executor.
+
 ## Agent types
 
 | Type | Description |
@@ -66,9 +111,9 @@ The orchestrator works with any LLM provider. The executor can be another LLM, o
     },
     "executor": {
       "provider": "clado-action",
-      "model": "qwen3-vl-30b-a3b-instruct",
+      "model": "Qwen3.5-35B-A3B-action-000159-merged",
       "apiKey": "",
-      "baseUrl": "https://clado-ai--clado-browseros-action-actionmodel-generate.modal.run"
+      "baseUrl": "https://clado-ai--clado-browseros-action-000159-merged-actionmod-f4a6ef.modal.run"
     }
   }
 }
@@ -96,6 +141,20 @@ The `apiKey` field supports two formats:
 - **Env var name**: `"OPENAI_API_KEY"` — resolved from `.env.development` at runtime
 - **Direct value**: `"sk-xxxxx"` — used as-is (not recommended)
 
+### Environment variables
+
+| Variable | Used for |
+|----------|----------|
+| `EVAL_AGENT_PROVIDER`, `EVAL_AGENT_MODEL`, `EVAL_AGENT_API_KEY`, `EVAL_AGENT_BASE_URL`, `EVAL_AGENT_SUPPORTS_IMAGES` | Suite variant model selection |
+| `FIREWORKS_API_KEY`, `OPENROUTER_API_KEY`, `ANTHROPIC_API_KEY`, provider-specific keys | Config-file or provider-backed model calls |
+| `EVAL_EXECUTOR_MODEL`, `EVAL_EXECUTOR_API_KEY`, `EVAL_EXECUTOR_BASE_URL` | Suite-mode orchestrator executor override |
+| `CLADO_ACTION_MODEL`, `CLADO_ACTION_API_KEY`, `CLADO_ACTION_BASE_URL` | Clado executor defaults |
+| `BROWSEROS_BINARY` | BrowserOS binary path in CI/local smoke runs |
+| `BROWSEROS_SERVER_URL` | Optional grader MCP URL override |
+| `WEBARENA_INFINITY_DIR` | Local WebArena-Infinity checkout for Infinity tasks |
+| `NOPECHA_API_KEY` | CAPTCHA solver extension |
+| `EVAL_R2_ACCOUNT_ID`, `EVAL_R2_ACCESS_KEY_ID`, `EVAL_R2_SECRET_ACCESS_KEY`, `EVAL_R2_BUCKET`, `EVAL_R2_CDN_BASE_URL` | R2 upload and viewer URL |
+
 ### Supported providers
 
 | Provider | `provider` value | Requires `baseUrl` |
@@ -110,6 +169,22 @@ The `apiKey` field supports two formats:
 | Ollama | `ollama` | No |
 | Clado Action (executor only) | `clado-action` | Yes |
 
+### R2 publishing
+
+`suite --config ... --publish r2` and `publish --target r2` upload the run artifacts plus `viewer.html` to the viewer-compatible R2 layout:
+
+```bash
+export EVAL_R2_ACCOUNT_ID=...
+export EVAL_R2_ACCESS_KEY_ID=...
+export EVAL_R2_SECRET_ACCESS_KEY=...
+export EVAL_R2_BUCKET=browseros-eval
+export EVAL_R2_CDN_BASE_URL=https://eval.browseros.com
+```
+
+`EVAL_R2_CDN_BASE_URL` must be a public R2 custom domain, `r2.dev` URL, or Worker URL. Do not set it to the private `*.r2.cloudflarestorage.com` S3 API endpoint.
+
+Published runs are available at `EVAL_R2_CDN_BASE_URL/viewer.html?run=<run-id>`.
+
 ### BrowserOS infrastructure
 
 ```json
@@ -137,10 +212,12 @@ Each worker gets its own Chrome instance. Worker N uses `base_port + N` for CDP
 
 | File | Tasks | Description |
 |------|-------|-------------|
+| `agisdk-daily-10.jsonl` | 10 | Daily AGI SDK / REAL Bench subset |
 | `webvoyager.jsonl` | 643 | Full WebVoyager benchmark |
 | `mind2web.jsonl` | 300 | Online-Mind2Web |
 | `webbench-{0,1,2}of4-50.jsonl` | 50 each | WebBench shards (50-task subsets) |
-| `agisdk-real.jsonl` | 40 | AGI SDK / REAL Bench (action-only tasks) |
+| `agisdk-real-smoke.jsonl` | 1 | AGI SDK / REAL Bench smoke task |
+| `agisdk-real.jsonl` | 36 | AGI SDK / REAL Bench (action-only tasks) |
 | `webarena-infinity-hard-50.jsonl` | 50 | WebArena-Infinity hard set |
 | `browsecomp-medium-hard-50.jsonl` | 50 | BrowseComp medium-hard |
 | `browsecomp-very-hard-50.jsonl` | 50 | BrowseComp very-hard |
@@ -167,14 +244,47 @@ results/
   browseros-agent-weekly/
     2026-04-29-1430/
       Amazon--0/
+        attempt.json          # Stable attempt summary for viewer/reporting
         metadata.json         # Task result, timing, grader scores
+        grades.json           # Compact grader results
         messages.jsonl         # Full message log
+        grader-artifacts/      # Grader-specific inputs/outputs/stderr
         screenshots/
           001.png              # Step-by-step screenshots
           002.png
       summary.json             # Aggregate pass rates
 ```
 
+R2 publishing preserves the task files under `runs/<run-id>/...`, writes `runs/<run-id>/manifest.json`, and uploads `viewer.html` at the bucket root. The viewer URL is `EVAL_R2_CDN_BASE_URL/viewer.html?run=<run-id>`.
+
+### R2 viewer manifest
+
+`runs/<run-id>/manifest.json` is the source of truth for the public viewer. New manifests include `schemaVersion: 2` and each task includes explicit artifact paths:
+
+```json
+{
+  "schemaVersion": 2,
+  "runId": "agisdk-real-smoke-2026-04-30-0000",
+  "tasks": [
+    {
+      "queryId": "agisdk-dashdish-10",
+      "paths": {
+        "metadata": "tasks/agisdk-dashdish-10/metadata.json",
+        "messages": "tasks/agisdk-dashdish-10/messages.jsonl",
+        "grades": "tasks/agisdk-dashdish-10/grades.json",
+        "trace": "tasks/agisdk-dashdish-10/trace.jsonl",
+        "screenshots": "tasks/agisdk-dashdish-10/screenshots",
+        "graderArtifacts": "tasks/agisdk-dashdish-10/grader-artifacts"
+      }
+    }
+  ]
+}
+```
+
+The static viewer uses `task.paths` when present. Older uploaded runs without `schemaVersion` or `task.paths` still work through the legacy inferred layout: `runs/<run-id>/<task-id>/metadata.json`, `messages.jsonl`, and `screenshots/<n>.png`.
+
+Manifest paths are stable artifact locations, not a guarantee that every optional artifact exists for every task. For example, `attempt.json`, `trace.jsonl`, or grader artifact directories may be absent when that artifact was not produced by the run.
+
 ## Troubleshooting
 
 **BrowserOS not found**: Expects `/Applications/BrowserOS.app/Contents/MacOS/BrowserOS`. Set `BROWSEROS_BINARY` to override.

packages/browseros-agent/apps/eval/configs/legacy/agisdk-real-smoke.json

@@ -7,8 +7,8 @@
     "baseUrl": "https://openrouter.ai/api/v1",
     "supportsImages": true
   },
-  "dataset": "../data/agisdk-real.jsonl",
-  "num_workers": 10,
+  "dataset": "../../data/agisdk-real-smoke.jsonl",
+  "num_workers": 1,
   "restart_server_per_task": true,
   "browseros": {
     "server_url": "http://127.0.0.1:9110",

packages/browseros-agent/apps/eval/configs/legacy/agisdk-real.json

@@ -0,0 +1,26 @@
+{
+  "agent": {
+    "type": "single",
+    "provider": "openai-compatible",
+    "model": "accounts/fireworks/models/kimi-k2p5",
+    "apiKey": "FIREWORKS_API_KEY",
+    "baseUrl": "https://api.fireworks.ai/inference/v1",
+    "supportsImages": true
+  },
+  "dataset": "../../data/agisdk-real.jsonl",
+  "num_workers": 4,
+  "restart_server_per_task": true,
+  "browseros": {
+    "server_url": "http://127.0.0.1:9110",
+    "base_cdp_port": 9010,
+    "base_server_port": 9110,
+    "base_extension_port": 9310,
+    "load_extensions": false,
+    "headless": false
+  },
+  "captcha": {
+    "api_key_env": "NOPECHA_API_KEY"
+  },
+  "graders": ["agisdk_state_diff"],
+  "timeout_ms": 1800000
+}

packages/browseros-agent/apps/eval/configs/legacy/browseros-agent-weekly.json

@@ -7,7 +7,7 @@
     "baseUrl": "https://openrouter.ai/api/v1",
     "supportsImages": true
   },
-  "dataset": "../data/webbench-2of4-50.jsonl",
+  "dataset": "../../data/webbench-2of4-50.jsonl",
   "num_workers": 10,
   "restart_server_per_task": true,
   "browseros": {

packages/browseros-agent/apps/eval/configs/legacy/browseros-oe-agent-weekly.json

@@ -14,7 +14,7 @@
       "baseUrl": "https://api.fireworks.ai/inference/v1"
     }
   },
-  "dataset": "../data/webbench-2of4-50.jsonl",
+  "dataset": "../../data/webbench-2of4-50.jsonl",
   "num_workers": 10,
   "restart_server_per_task": true,
   "browseros": {

packages/browseros-agent/apps/eval/configs/legacy/browseros-oe-clado-weekly.json

@@ -9,12 +9,12 @@
     },
     "executor": {
       "provider": "clado-action",
-      "model": "qwen3-vl-30b-a3b-instruct",
+      "model": "Qwen3.5-35B-A3B-action-000159-merged",
       "apiKey": "",
-      "baseUrl": "https://clado-ai--clado-browseros-action-actionmodel-generate.modal.run"
+      "baseUrl": "https://clado-ai--clado-browseros-action-000159-merged-actionmod-f4a6ef.modal.run"
     }
   },
-  "dataset": "../data/webbench-2of4-50.jsonl",
+  "dataset": "../../data/agisdk-real.jsonl",
   "num_workers": 10,
   "restart_server_per_task": true,
   "browseros": {
@@ -23,11 +23,11 @@
     "base_server_port": 9110,
     "base_extension_port": 9310,
     "load_extensions": false,
-    "headless": false
+    "headless": true
   },
   "captcha": {
     "api_key_env": "NOPECHA_API_KEY"
   },
-  "graders": ["performance_grader"],
+  "graders": ["agisdk_state_diff"],
   "timeout_ms": 1800000
 }

packages/browseros-agent/apps/eval/configs/legacy/infinity-hard-50.json

@@ -7,7 +7,7 @@
     "baseUrl": "https://openrouter.ai/api/v1",
     "supportsImages": true
   },
-  "dataset": "../data/webarena-infinity-hard-50.jsonl",
+  "dataset": "../../data/webarena-infinity-hard-50.jsonl",
   "num_workers": 10,
   "restart_server_per_task": true,
   "browseros": {

packages/browseros-agent/apps/eval/configs/legacy/test-mind2web.json

@@ -5,7 +5,7 @@
     "model": "openai/gpt-4.1",
     "apiKey": "OPENROUTER_API_KEY"
   },
-  "dataset": "../data/mind2web.jsonl",
+  "dataset": "../../data/mind2web.jsonl",
   "num_workers": 5,
   "restart_server_per_task": true,
   "browseros": {

packages/browseros-agent/apps/eval/configs/legacy/test-webvoyager.json

@@ -7,7 +7,7 @@
     "baseUrl": "https://api.fireworks.ai/inference/v1",
     "supportsImages": true
   },
-  "dataset": "../data/webvoyager.jsonl",
+  "dataset": "../../data/webvoyager.jsonl",
   "num_workers": 3,
   "restart_server_per_task": true,
   "browseros": {

packages/browseros-agent/apps/eval/configs/suites/agisdk-daily-10.json

@@ -0,0 +1,22 @@
+{
+  "id": "agisdk-daily-10",
+  "dataset": "../../data/agisdk-daily-10.jsonl",
+  "agent": {
+    "type": "single"
+  },
+  "graders": ["agisdk_state_diff"],
+  "workers": 1,
+  "restartBrowserPerTask": true,
+  "timeoutMs": 1800000,
+  "browseros": {
+    "server_url": "http://127.0.0.1:9110",
+    "base_cdp_port": 9010,
+    "base_server_port": 9110,
+    "base_extension_port": 9310,
+    "load_extensions": false,
+    "headless": true
+  },
+  "captcha": {
+    "api_key_env": "NOPECHA_API_KEY"
+  }
+}

packages/browseros-agent/apps/eval/configs/suites/agisdk-real-smoke.json

@@ -0,0 +1,22 @@
+{
+  "id": "agisdk-real-smoke",
+  "dataset": "../../data/agisdk-real-smoke.jsonl",
+  "agent": {
+    "type": "single"
+  },
+  "graders": ["agisdk_state_diff"],
+  "workers": 1,
+  "restartBrowserPerTask": true,
+  "timeoutMs": 1800000,
+  "browseros": {
+    "server_url": "http://127.0.0.1:9110",
+    "base_cdp_port": 9010,
+    "base_server_port": 9110,
+    "base_extension_port": 9310,
+    "load_extensions": false,
+    "headless": false
+  },
+  "captcha": {
+    "api_key_env": "NOPECHA_API_KEY"
+  }
+}

packages/browseros-agent/apps/eval/configs/suites/agisdk-real.json

@@ -0,0 +1,22 @@
+{
+  "id": "agisdk-real",
+  "dataset": "../../data/agisdk-real.jsonl",
+  "agent": {
+    "type": "single"
+  },
+  "graders": ["agisdk_state_diff"],
+  "workers": 1,
+  "restartBrowserPerTask": true,
+  "timeoutMs": 1800000,
+  "browseros": {
+    "server_url": "http://127.0.0.1:9110",
+    "base_cdp_port": 9010,
+    "base_server_port": 9110,
+    "base_extension_port": 9310,
+    "load_extensions": false,
+    "headless": false
+  },
+  "captcha": {
+    "api_key_env": "NOPECHA_API_KEY"
+  }
+}

packages/browseros-agent/apps/eval/data/agisdk-daily-10.jsonl

@@ -0,0 +1,10 @@
+{"query_id": "agisdk-dashdish-10", "dataset": "agisdk-real", "query": "Place an order from \"Souvla\" for a \"Medium Classic Cheeseburger\" and a \"Small Bacon Double Cheeseburger\" with \"Standard Delivery\" as the method with the default charged options.", "graders": ["agisdk_state_diff"], "start_url": "https://evals-dashdish.vercel.app", "metadata": {"original_task_id": "dashdish-10", "website": "DashDish", "category": "agisdk-real", "additional": {"agisdk_task_id": "dashdish-10", "challenge_type": "action", "difficulty": "hard", "similar_to": "Doordash"}}}
+{"query_id": "agisdk-fly-unified-5", "dataset": "agisdk-real", "query": "Find me the cheapest fare for a flight from Orlando to Milwaukee on December 5th, 2024 and book it.\nPassenger: John Doe\nDate of Birth: 01/01/1990\nSex: Male\nSeat Selection: No\nPayment: Credit Card (378342143523967), Exp: 12/30, Security Code: 420 Address: 123 Main St, San Francisco, CA, 94105, USA, Phone: 555-123-4567, Email: johndoe@example.com.", "graders": ["agisdk_state_diff"], "start_url": "https://evals-fly-unified.vercel.app", "metadata": {"original_task_id": "fly-unified-5", "website": "Fly Unified", "category": "agisdk-real", "additional": {"agisdk_task_id": "fly-unified-5", "challenge_type": "retrieval-action", "difficulty": "medium", "similar_to": "United Airlines"}}}
+{"query_id": "agisdk-udriver-10", "dataset": "agisdk-real", "query": "Order me a ride for 4pm, I'll be at the de Young muesum headed to the Waterbar, fanciest option possible please.", "graders": ["agisdk_state_diff"], "start_url": "https://evals-udriver.vercel.app", "metadata": {"original_task_id": "udriver-10", "website": "UDriver", "category": "agisdk-real", "additional": {"agisdk_task_id": "udriver-10", "challenge_type": "action", "difficulty": "hard", "similar_to": "Uber"}}}
+{"query_id": "agisdk-udriver-9", "dataset": "agisdk-real", "query": "Book me a ride from the thai restaurant I last took a ride to for later today at 2pm, I'll be at 333 Apartments on Fremont", "graders": ["agisdk_state_diff"], "start_url": "https://evals-udriver.vercel.app", "metadata": {"original_task_id": "udriver-9", "website": "UDriver", "category": "agisdk-real", "additional": {"agisdk_task_id": "udriver-9", "challenge_type": "retrieval-action", "difficulty": "hard", "similar_to": "Uber"}}}
+{"query_id": "agisdk-topwork-4", "dataset": "agisdk-real", "query": "Create a job post for a UI/UX Designer with expertise in Figma, Sketch, and Adobe Creative Suite, including project details, timeline, and required skills (Wireframing, Prototyping, Responsive Design).", "graders": ["agisdk_state_diff"], "start_url": "https://evals-topwork.vercel.app", "metadata": {"original_task_id": "topwork-4", "website": "TopWork", "category": "agisdk-real", "additional": {"agisdk_task_id": "topwork-4", "challenge_type": "action", "difficulty": "medium", "similar_to": "Upwork"}}}
+{"query_id": "agisdk-gocalendar-4", "dataset": "agisdk-real", "query": "Change the \"Team Check-In\" event on July 18, 2024, name to \"Project Kickoff\" and update the location to \"Zoom\"", "graders": ["agisdk_state_diff"], "start_url": "https://evals-gocalendar.vercel.app", "metadata": {"original_task_id": "gocalendar-4", "website": "GoCalendar", "category": "agisdk-real", "additional": {"agisdk_task_id": "gocalendar-4", "challenge_type": "action", "difficulty": "medium", "similar_to": "Google Calendar"}}}
+{"query_id": "agisdk-staynb-6", "dataset": "agisdk-real", "query": "Find and book the stay with the best value for money (cheapest stay with the best reviews) for 1 day. For fields you don't know the answer for, just fill them in with anything of your choice.", "graders": ["agisdk_state_diff"], "start_url": "https://evals-staynb.vercel.app", "metadata": {"original_task_id": "staynb-6", "website": "StayNB", "category": "agisdk-real", "additional": {"agisdk_task_id": "staynb-6", "challenge_type": "retrieval-action", "difficulty": "medium", "similar_to": "Airbnb"}}}
+{"query_id": "agisdk-udriver-11", "dataset": "agisdk-real", "query": "I need to go from Pacific Catch on Chestnut back home to 333 Fremont now. If the fancy version is within ten dollars of the regular one, book that.", "graders": ["agisdk_state_diff"], "start_url": "https://evals-udriver.vercel.app", "metadata": {"original_task_id": "udriver-11", "website": "UDriver", "category": "agisdk-real", "additional": {"agisdk_task_id": "udriver-11", "challenge_type": "action", "difficulty": "hard", "similar_to": "Uber"}}}
+{"query_id": "agisdk-networkin-5", "dataset": "agisdk-real", "query": "Send a connection request to John Smith.", "graders": ["agisdk_state_diff"], "start_url": "https://evals-networkin.vercel.app", "metadata": {"original_task_id": "networkin-5", "website": "Networkin", "category": "agisdk-real", "additional": {"agisdk_task_id": "networkin-5", "challenge_type": "action", "difficulty": "easy", "similar_to": "LinkedIn"}}}
+{"query_id": "agisdk-zilloft-6", "dataset": "agisdk-real", "query": "Select a property listed in San Francisco as \"Condos\" within a price range under $300,000 and request a tour for tomorrow at 4:00 PM. Use these contact details: Name: Sarah Brown, Email: sarahbrown@example.com, Phone: 555-987-6543.", "graders": ["agisdk_state_diff"], "start_url": "https://evals-zilloft.vercel.app", "metadata": {"original_task_id": "zilloft-6", "website": "Zilloft", "category": "agisdk-real", "additional": {"agisdk_task_id": "zilloft-6", "challenge_type": "action", "difficulty": "medium", "similar_to": "Zillow"}}}

packages/browseros-agent/apps/eval/data/agisdk-real-smoke.jsonl

@@ -0,0 +1 @@
+{"query_id": "agisdk-dashdish-10", "dataset": "agisdk-real", "query": "Place an order from \"Souvla\" for a \"Medium Classic Cheeseburger\" and a \"Small Bacon Double Cheeseburger\" with \"Standard Delivery\" as the method with the default charged options.", "graders": ["agisdk_state_diff"], "start_url": "https://evals-dashdish.vercel.app", "metadata": {"original_task_id": "dashdish-10", "website": "DashDish", "category": "agisdk-real", "additional": {"agisdk_task_id": "dashdish-10", "challenge_type": "action", "difficulty": "hard", "similar_to": "Doordash"}}}

packages/browseros-agent/apps/eval/package.json

@@ -5,6 +5,7 @@
   "type": "module",
   "scripts": {
     "eval": "bun --env-file=.env.development run src/index.ts",
+    "test": "bun run ../../scripts/run-bun-test.ts ./apps/eval/tests",
     "typecheck": "tsc --noEmit"
   },
   "dependencies": {

packages/browseros-agent/apps/eval/scripts/test-clado-api.ts

@@ -1,34 +1,73 @@
 /**
- * Test script for Clado API endpoints (grounding + action models)
+ * Smoke-test for the Clado BrowserOS Action endpoint.
+ *
+ * Health-checks the model, then runs a generate call and prints every
+ * field the new contract documents (action, coordinates, text, key,
+ * direction, scroll/drag fields, wait, end+final_answer, thinking,
+ * parse_error, raw_response).
  *
  * Usage:
  *   bun apps/eval/scripts/test-clado-api.ts [screenshot-path]
  *
- * If no screenshot provided, captures one from a running BrowserOS server.
+ * If no screenshot path is given, captures one over MCP from a
+ * running BrowserOS server (default http://127.0.0.1:9110, override
+ * with BROWSEROS_URL).
+ *
+ * Cold start can take ~5 minutes; the script waits up to 6.
  */
 
 import { readFile } from 'node:fs/promises'
 import { resolve } from 'node:path'
 
 const ACTION_URL =
-  'https://clado-ai--clado-browseros-action-actionmodel-generate.modal.run'
+  'https://clado-ai--clado-browseros-action-000159-merged-actionmod-f4a6ef.modal.run'
 const ACTION_HEALTH_URL =
-  'https://clado-ai--clado-browseros-action-actionmodel-health.modal.run'
-const GROUNDING_URL =
-  'https://clado-ai--clado-browseros-grounding-groundingmodel-generate.modal.run'
-const GROUNDING_HEALTH_URL =
-  'https://clado-ai--clado-browseros-grounding-groundingmodel-health.modal.run'
+  'https://clado-ai--clado-browseros-action-000159-merged-actionmod-5e5033.modal.run'
 
-async function checkHealth(name: string, url: string): Promise<boolean> {
-  console.log(`\n--- ${name} health check ---`)
-  console.log(`  URL: ${url}`)
+const COLD_START_BUDGET_MS = 360_000 // 6 min — Clado cold start is ~5 min
+const COLD_START_WARN_MS = 30_000
+
+interface CladoResponse {
+  action?: string | null
+  thinking?: string | null
+  raw_response?: string
+  parse_error?: string | null
+  inference_time_seconds?: number
+  x?: number
+  y?: number
+  text?: string
+  key?: string
+  direction?: string
+  amount?: number
+  startX?: number
+  startY?: number
+  endX?: number
+  endY?: number
+  time?: number
+  final_answer?: string | null
+}
+
+async function checkHealth(): Promise<boolean> {
+  console.log(`\n--- Action model health ---`)
+  console.log(`  URL:   ${ACTION_HEALTH_URL}`)
+  console.log(
+    `  Note:  cold start can take ~5 min; waiting up to ${COLD_START_BUDGET_MS / 1000}s.`,
+  )
   const start = performance.now()
+  const warn = setTimeout(() => {
+    console.log(
+      `  ...still waiting (${COLD_START_WARN_MS / 1000}s in) — model is likely cold-starting on Modal.`,
+    )
+  }, COLD_START_WARN_MS)
+
   try {
-    const resp = await fetch(url, { signal: AbortSignal.timeout(30_000) })
+    const resp = await fetch(ACTION_HEALTH_URL, {
+      signal: AbortSignal.timeout(COLD_START_BUDGET_MS),
+    })
     const elapsed = ((performance.now() - start) / 1000).toFixed(2)
     const body = await resp.text()
     console.log(`  Status: ${resp.status} (${elapsed}s)`)
-    console.log(`  Body: ${body.slice(0, 200)}`)
+    console.log(`  Body:   ${body.slice(0, 400)}`)
     return resp.ok
   } catch (err) {
     const elapsed = ((performance.now() - start) / 1000).toFixed(2)
@@ -36,63 +75,34 @@ async function checkHealth(name: string, url: string): Promise<boolean> {
       `  FAILED (${elapsed}s): ${err instanceof Error ? err.message : err}`,
     )
     return false
+  } finally {
+    clearTimeout(warn)
   }
 }
 
-async function testGenerate(
-  name: string,
-  url: string,
+async function generate(
+  label: string,
   payload: Record<string, unknown>,
-): Promise<Record<string, unknown> | null> {
-  console.log(`\n--- ${name} generate ---`)
-  console.log(`  URL: ${url}`)
+): Promise<CladoResponse | null> {
+  console.log(`\n--- ${label} ---`)
+  console.log(`  URL:         ${ACTION_URL}`)
   console.log(`  Instruction: ${payload.instruction}`)
   console.log(
-    `  Image size: ${((payload.image_base64 as string).length / 1024).toFixed(0)} KB (base64)`,
+    `  Image size:  ${((payload.image_base64 as string).length / 1024).toFixed(0)} KB (base64)`,
   )
-  if (payload.history) console.log(`  History: ${payload.history}`)
+  if (payload.history && payload.history !== 'None') {
+    console.log(`  History:     ${payload.history}`)
+  }
 
   const start = performance.now()
+  let resp: Response
   try {
-    const resp = await fetch(url, {
+    resp = await fetch(ACTION_URL, {
       method: 'POST',
       headers: { 'Content-Type': 'application/json' },
       body: JSON.stringify(payload),
-      signal: AbortSignal.timeout(120_000),
+      signal: AbortSignal.timeout(COLD_START_BUDGET_MS),
     })
-    const elapsed = ((performance.now() - start) / 1000).toFixed(2)
-
-    if (!resp.ok) {
-      const body = await resp.text()
-      console.log(`  FAILED: HTTP ${resp.status} (${elapsed}s)`)
-      console.log(`  Body: ${body.slice(0, 400)}`)
-      return null
-    }
-
-    const result = (await resp.json()) as Record<string, unknown>
-    console.log(`  Status: ${resp.status} (${elapsed}s)`)
-    console.log(`  Action: ${result.action}`)
-    if (result.x !== null && result.x !== undefined)
-      console.log(`  Coordinates: (${result.x}, ${result.y})`)
-    if (result.text)
-      console.log(`  Text: ${(result.text as string).slice(0, 100)}`)
-    if (result.key) console.log(`  Key: ${result.key}`)
-    if (result.inference_time_seconds)
-      console.log(`  Inference: ${result.inference_time_seconds}s`)
-
-    // Show thinking if present
-    const raw = result.raw_response as string | undefined
-    if (raw) {
-      const thinkMatch = raw.match(/<thinking>([\s\S]*?)<\/thinking>/)
-      if (thinkMatch) {
-        const thinking = thinkMatch[1].trim()
-        console.log(
-          `  Thinking: ${thinking.slice(0, 200)}${thinking.length > 200 ? '...' : ''}`,
-        )
-      }
-    }
-
-    return result
   } catch (err) {
     const elapsed = ((performance.now() - start) / 1000).toFixed(2)
     console.log(
@@ -100,6 +110,50 @@ async function testGenerate(
     )
     return null
   }
+  const elapsed = ((performance.now() - start) / 1000).toFixed(2)
+
+  if (!resp.ok) {
+    const body = await resp.text()
+    console.log(`  HTTP ${resp.status} ${resp.statusText} (${elapsed}s)`)
+    console.log(`  Body: ${body.slice(0, 400)}`)
+    return null
+  }
+
+  const result = (await resp.json()) as CladoResponse
+  console.log(`  HTTP ${resp.status} (${elapsed}s)`)
+  console.log(`  action:                ${result.action ?? 'null'}`)
+  if (result.parse_error) {
+    console.log(`  parse_error:           ${result.parse_error}`)
+  }
+  if (result.thinking) {
+    const trimmed = result.thinking.replace(/\s+/g, ' ').trim()
+    console.log(
+      `  thinking:              ${trimmed.slice(0, 240)}${trimmed.length > 240 ? '…' : ''}`,
+    )
+  }
+  if (typeof result.x === 'number' || typeof result.y === 'number') {
+    console.log(`  x, y:                  ${result.x}, ${result.y}`)
+  }
+  if (typeof result.text === 'string')
+    console.log(`  text:                  ${result.text.slice(0, 120)}`)
+  if (typeof result.key === 'string')
+    console.log(`  key:                   ${result.key}`)
+  if (typeof result.direction === 'string')
+    console.log(`  direction:             ${result.direction}`)
+  if (typeof result.amount === 'number')
+    console.log(`  amount:                ${result.amount}`)
+  if (typeof result.startX === 'number' || typeof result.endX === 'number') {
+    console.log(
+      `  drag:                  (${result.startX}, ${result.startY}) → (${result.endX}, ${result.endY})`,
+    )
+  }
+  if (typeof result.time === 'number')
+    console.log(`  time:                  ${result.time}s`)
+  if (result.final_answer)
+    console.log(`  final_answer:          ${result.final_answer.slice(0, 240)}`)
+  if (typeof result.inference_time_seconds === 'number')
+    console.log(`  inference_time_seconds: ${result.inference_time_seconds}`)
+  return result
 }
 
 async function loadScreenshot(path?: string): Promise<string> {
@@ -110,10 +164,9 @@ async function loadScreenshot(path?: string): Promise<string> {
     return data.toString('base64')
   }
 
-  // Try to capture from a running BrowserOS server
   const serverUrl = process.env.BROWSEROS_URL || 'http://127.0.0.1:9110'
   console.log(
-    `No screenshot path provided. Trying to capture from ${serverUrl}...`,
+    `No screenshot path provided. Capturing from ${serverUrl} via MCP...`,
   )
 
   const { Client } = await import('@modelcontextprotocol/sdk/client/index.js')
@@ -134,82 +187,101 @@ async function loadScreenshot(path?: string): Promise<string> {
       arguments: { format: 'png', page: 1 },
     })) as { content: Array<{ type: string; data?: string }> }
 
-    const imageContent = result.content?.find((c) => c.type === 'image')
-    if (!imageContent?.data)
-      throw new Error('No image data in screenshot response')
+    const image = result.content?.find((c) => c.type === 'image')
+    if (!image?.data)
+      throw new Error('No image data in take_screenshot response')
 
     console.log(
-      `Captured screenshot (${(imageContent.data.length / 1024).toFixed(0)} KB base64)`,
+      `Captured screenshot (${(image.data.length / 1024).toFixed(0)} KB base64)`,
     )
-    return imageContent.data
+    return image.data
   } finally {
     try {
       await transport.close()
-    } catch {}
+    } catch {
+      /* ignore */
+    }
   }
 }
 
+function summarize(history: CladoResponse[]): string {
+  if (history.length === 0) return 'None'
+  return history
+    .map((h) => {
+      switch (h.action) {
+        case 'click':
+        case 'double_click':
+        case 'right_click':
+        case 'hover':
+          return `${h.action}(${h.x}, ${h.y})`
+        case 'type':
+          return `type(${JSON.stringify(h.text ?? '')})`
+        case 'press_key':
+          return `press_key(${JSON.stringify(h.key ?? '')})`
+        case 'scroll':
+          return `scroll(${h.direction ?? 'down'})`
+        case 'drag':
+          return `drag(${h.startX},${h.startY} -> ${h.endX},${h.endY})`
+        case 'wait':
+          return `wait(${h.time ?? 1}s)`
+        case 'end':
+          return 'end()'
+        default:
+          return h.action ?? 'invalid'
+      }
+    })
+    .join(' -> ')
+}
+
 async function main() {
-  const screenshotPath = process.argv[2]
+  console.log('=== Clado action endpoint smoke test ===')
 
-  console.log('=== Clado API Test ===\n')
-
-  // Health checks (parallel)
-  const [actionHealthy, groundingHealthy] = await Promise.all([
-    checkHealth('Action Model', ACTION_HEALTH_URL),
-    checkHealth('Grounding Model', GROUNDING_HEALTH_URL),
-  ])
-
-  if (!actionHealthy && !groundingHealthy) {
-    console.log('\nBoth endpoints are down. Exiting.')
+  const healthy = await checkHealth()
+  if (!healthy) {
+    console.log('\nHealth check failed. Exiting.')
     process.exit(1)
   }
 
-  // Load screenshot
   let imageBase64: string
   try {
-    imageBase64 = await loadScreenshot(screenshotPath)
+    imageBase64 = await loadScreenshot(process.argv[2])
   } catch (err) {
     console.log(
       `\nFailed to load screenshot: ${err instanceof Error ? err.message : err}`,
     )
     console.log(
-      'Provide a screenshot path: bun apps/eval/scripts/test-clado-api.ts path/to/screenshot.png',
+      'Pass a path: bun apps/eval/scripts/test-clado-api.ts path/to/screenshot.png',
     )
     process.exit(1)
   }
 
-  const instruction = 'Click on the search button or search bar'
+  const history: CladoResponse[] = []
 
-  // Test grounding model
-  if (groundingHealthy) {
-    await testGenerate('Grounding Model', GROUNDING_URL, {
-      instruction,
+  // Step 1: open task — let the model decide what to do.
+  const step1 = await generate('Step 1: cold task', {
+    instruction: 'Find the search bar and click it',
+    image_base64: imageBase64,
+    history: 'None',
+  })
+  if (step1?.action) history.push(step1)
+
+  // Step 2: continuation with history, asks for typing.
+  if (step1?.action) {
+    const step2 = await generate('Step 2: with history', {
+      instruction: 'Type "hello world" into the search bar',
       image_base64: imageBase64,
+      history: summarize(history),
     })
-  } else {
-    console.log('\nSkipping grounding model (unhealthy)')
+    if (step2?.action) history.push(step2)
   }
 
-  // Test action model (no history)
-  if (actionHealthy) {
-    const result = await testGenerate('Action Model (step 1)', ACTION_URL, {
-      instruction,
-      image_base64: imageBase64,
-      history: 'None',
-    })
-
-    // Test action model with history (simulate multi-turn)
-    if (result && result.action === 'click') {
-      await testGenerate('Action Model (step 2, with history)', ACTION_URL, {
-        instruction: 'Type "hello world" in the search bar',
-        image_base64: imageBase64,
-        history: `click(${result.x}, ${result.y})`,
-      })
-    }
-  } else {
-    console.log('\nSkipping action model (unhealthy)')
-  }
+  // Step 3: ask for end with a final answer to exercise that field.
+  await generate('Step 3: ask for end+final_answer', {
+    instruction:
+      'You have completed the task. Reply with end() and final_answer="done".',
+    image_base64: imageBase64,
+    history: summarize(history),
+  })
 
   console.log('\n=== Done ===')
 }

packages/browseros-agent/apps/eval/scripts/upload-run.ts

@@ -1,349 +1,43 @@
+#!/usr/bin/env bun
+
 /**
  * Upload eval runs to R2.
  *
  * Two modes:
  *   bun scripts/upload-run.ts results/browseros-agent-weekly/2026-03-21-1730
- *       → uploads that specific run
- *
  *   bun scripts/upload-run.ts results/browseros-agent-weekly
- *       → finds all timestamped subfolders, uploads any not yet in R2
- *
- * Env vars: EVAL_R2_ACCOUNT_ID, EVAL_R2_ACCESS_KEY_ID, EVAL_R2_SECRET_ACCESS_KEY
- *           EVAL_R2_BUCKET (default: browseros-eval)
- *           EVAL_R2_CDN_BASE_URL (default: https://eval.browseros.com)
  */
 
-import { readdir, readFile, stat } from 'node:fs/promises'
-import { basename, dirname, extname, join } from 'node:path'
 import {
-  GetObjectCommand,
-  PutObjectCommand,
-  S3Client,
-} from '@aws-sdk/client-s3'
+  loadR2ConfigFromEnv,
+  R2Publisher,
+} from '../src/publishing/r2-publisher'
 
-const CONCURRENCY = 20
-
-const CONTENT_TYPES: Record<string, string> = {
-  '.json': 'application/json',
-  '.jsonl': 'application/x-ndjson',
-  '.png': 'image/png',
-}
-
-interface R2Config {
-  accountId: string
-  accessKeyId: string
-  secretAccessKey: string
-  bucket: string
-  cdnBaseUrl: string
-}
-
-function loadConfig(): R2Config {
-  const accountId = process.env.EVAL_R2_ACCOUNT_ID
-  const accessKeyId = process.env.EVAL_R2_ACCESS_KEY_ID
-  const secretAccessKey = process.env.EVAL_R2_SECRET_ACCESS_KEY
-
-  if (!accountId || !accessKeyId || !secretAccessKey) {
-    console.error(
-      'Missing required env vars: EVAL_R2_ACCOUNT_ID, EVAL_R2_ACCESS_KEY_ID, EVAL_R2_SECRET_ACCESS_KEY',
-    )
-    process.exit(1)
-  }
-
-  return {
-    accountId,
-    accessKeyId,
-    secretAccessKey,
-    bucket: process.env.EVAL_R2_BUCKET || 'browseros-eval',
-    cdnBaseUrl: (
-      process.env.EVAL_R2_CDN_BASE_URL || 'https://eval.browseros.com'
-    ).replace(/\/+$/, ''),
-  }
-}
-
-function createClient(config: R2Config): S3Client {
-  return new S3Client({
-    region: 'auto',
-    endpoint: `https://${config.accountId}.r2.cloudflarestorage.com`,
-    credentials: {
-      accessKeyId: config.accessKeyId,
-      secretAccessKey: config.secretAccessKey,
-    },
-  })
-}
-
-async function upload(
-  client: S3Client,
-  bucket: string,
-  key: string,
-  body: Buffer,
-  contentType: string,
-) {
-  await client.send(
-    new PutObjectCommand({
-      Bucket: bucket,
-      Key: key,
-      Body: body,
-      ContentType: contentType,
-    }),
-  )
-}
-
-async function collectFiles(dir: string): Promise<string[]> {
-  const files: string[] = []
-  const entries = await readdir(dir, { withFileTypes: true })
-  for (const entry of entries) {
-    const full = join(dir, entry.name)
-    if (entry.isDirectory()) {
-      files.push(...(await collectFiles(full)))
-    } else {
-      files.push(full)
-    }
-  }
-  return files
-}
-
-async function runPool<T>(
-  items: T[],
-  concurrency: number,
-  fn: (item: T) => Promise<void>,
-) {
-  let i = 0
-  const workers = Array.from({ length: concurrency }, async () => {
-    while (i < items.length) {
-      const idx = i++
-      await fn(items[idx])
-    }
-  })
-  await Promise.all(workers)
-}
-
-// Check if a run has already been uploaded to R2
-async function isUploaded(
-  client: S3Client,
-  bucket: string,
-  runId: string,
-): Promise<boolean> {
-  try {
-    await client.send(
-      new GetObjectCommand({
-        Bucket: bucket,
-        Key: `runs/${runId}/manifest.json`,
-      }),
-    )
-    return true
-  } catch {
-    return false
-  }
-}
-
-// Detect if a directory is a run dir (has task subdirs with metadata.json)
-// vs a config dir (has timestamped subdirs like 2026-03-21-1730/)
-async function isRunDir(dir: string): Promise<boolean> {
-  const entries = await readdir(dir, { withFileTypes: true })
-  const subdirs = entries.filter((e) => e.isDirectory())
-  for (const subdir of subdirs) {
-    const metaPath = join(dir, subdir.name, 'metadata.json')
-    const metaStat = await stat(metaPath).catch(() => null)
-    if (metaStat?.isFile()) return true
-  }
-  return false
-}
-
-async function uploadSingleRun(
-  runDir: string,
-  runId: string,
-  r2Config: R2Config,
-  client: S3Client,
-): Promise<void> {
-  const taskDirs = await readdir(runDir, { withFileTypes: true })
-  const taskEntries = taskDirs.filter((d) => d.isDirectory())
-
-  if (taskEntries.length === 0) {
-    console.warn(`  No task subdirectories in ${runId}, skipping`)
-    return
-  }
-
-  const manifestTasks: Record<string, unknown>[] = []
-  const jobs: { key: string; filePath: string; contentType: string }[] = []
-
-  // Extract agent config from first task
-  let agentConfig: Record<string, unknown> | undefined
-  let dataset: string | undefined
-
-  for (const taskDir of taskEntries) {
-    const taskId = taskDir.name
-    const taskPath = join(runDir, taskId)
-    const metaPath = join(taskPath, 'metadata.json')
-
-    let meta: Record<string, unknown> = {}
-    try {
-      meta = JSON.parse(await readFile(metaPath, 'utf-8'))
-    } catch {
-      continue
-    }
-
-    if (!agentConfig && meta.agent_config)
-      agentConfig = meta.agent_config as Record<string, unknown>
-    if (!dataset && meta.dataset) dataset = meta.dataset as string
-
-    const files = await collectFiles(taskPath)
-    let screenshotCount = 0
-
-    for (const file of files) {
-      const relative = file.slice(taskPath.length + 1)
-      const ext = extname(file)
-      if (relative.startsWith('screenshots/') && ext === '.png')
-        screenshotCount++
-
-      jobs.push({
-        key: `runs/${runId}/${taskId}/${relative}`,
-        filePath: file,
-        contentType: CONTENT_TYPES[ext] || 'application/octet-stream',
-      })
-    }
-
-    manifestTasks.push({
-      queryId: meta.query_id || taskId,
-      query: meta.query || '',
-      startUrl: meta.start_url || '',
-      status:
-        meta.termination_reason === 'completed'
-          ? 'completed'
-          : meta.termination_reason || 'unknown',
-      durationMs: meta.total_duration_ms || 0,
-      screenshotCount: (meta.screenshot_count as number) || screenshotCount,
-      graderResults: meta.grader_results || {},
-    })
-  }
-
-  if (manifestTasks.length === 0) {
-    console.warn(`  No completed tasks in ${runId}, skipping`)
-    return
-  }
-
-  console.log(
-    `  Uploading ${jobs.length} files across ${manifestTasks.length} tasks...`,
-  )
-
-  let uploaded = 0
-  await runPool(jobs, CONCURRENCY, async (job) => {
-    const body = await readFile(job.filePath)
-    await upload(client, r2Config.bucket, job.key, body, job.contentType)
-    uploaded++
-    if (uploaded % 50 === 0 || uploaded === jobs.length) {
-      console.log(`    ${uploaded}/${jobs.length}`)
-    }
-  })
-
-  // Read summary.json if it exists
-  let summaryData: Record<string, unknown> | undefined
-  try {
-    summaryData = JSON.parse(
-      await readFile(join(runDir, 'summary.json'), 'utf-8'),
-    )
-  } catch {}
-
-  // Upload manifest
-  const manifest = {
-    runId,
-    uploadedAt: new Date().toISOString(),
-    agentConfig,
-    dataset,
-    summary: summaryData
-      ? {
-          passRate: summaryData.passRate,
-          avgDurationMs: summaryData.avgDurationMs,
-        }
-      : undefined,
-    tasks: manifestTasks,
-  }
-  const manifestBody = Buffer.from(JSON.stringify(manifest, null, 2))
-  await upload(
-    client,
-    r2Config.bucket,
-    `runs/${runId}/manifest.json`,
-    manifestBody,
-    'application/json',
-  )
-
-  // Upload viewer.html to bucket root
-  const viewerPath = join(
-    import.meta.dir,
-    '..',
-    'src',
-    'dashboard',
-    'viewer.html',
-  )
-  const viewerBody = await readFile(viewerPath)
-  await upload(client, r2Config.bucket, 'viewer.html', viewerBody, 'text/html')
-
-  console.log(`  Uploaded ${uploaded + 2} files`)
-  console.log(`  ${r2Config.cdnBaseUrl}/viewer.html?run=${runId}`)
-}
-
-async function main() {
+async function main(): Promise<void> {
   const inputDir = process.argv[2]
   if (!inputDir) {
-    console.error(
+    throw new Error(
       'Usage:\n' +
-        '  bun scripts/upload-run.ts results/config-name/2026-03-21-1730  (specific run)\n' +
-        '  bun scripts/upload-run.ts results/config-name                   (all un-uploaded runs)',
-    )
-    process.exit(1)
-  }
-
-  const dirStat = await stat(inputDir).catch(() => null)
-  if (!dirStat?.isDirectory()) {
-    console.error(`Not a directory: ${inputDir}`)
-    process.exit(1)
-  }
-
-  const r2Config = loadConfig()
-  const client = createClient(r2Config)
-
-  if (await isRunDir(inputDir)) {
-    // Single run: results/config-name/2026-03-21-1730
-    const timestamp = basename(inputDir)
-    const configName = basename(dirname(inputDir))
-    const runId = `${configName}-${timestamp}`
-    console.log(`Uploading run: ${runId}`)
-    await uploadSingleRun(inputDir, runId, r2Config, client)
-  } else {
-    // Config dir: results/config-name/ — upload all un-uploaded runs
-    const configName = basename(inputDir)
-    const entries = await readdir(inputDir, { withFileTypes: true })
-    const runDirs = entries
-      .filter((e) => e.isDirectory())
-      .map((e) => e.name)
-      .sort()
-
-    if (runDirs.length === 0) {
-      console.error('No run subdirectories found')
-      process.exit(1)
-    }
-
-    console.log(
-      `Found ${runDirs.length} runs for config "${configName}", checking R2...`,
-    )
-
-    let uploadedCount = 0
-    for (const dir of runDirs) {
-      const runId = `${configName}-${dir}`
-      const alreadyUploaded = await isUploaded(client, r2Config.bucket, runId)
-      if (alreadyUploaded) {
-        console.log(`  ${runId}: already uploaded, skipping`)
-        continue
-      }
-
-      console.log(`  ${runId}: uploading...`)
-      await uploadSingleRun(join(inputDir, dir), runId, r2Config, client)
-      uploadedCount++
-    }
-
-    console.log(
-      `\nDone. Uploaded ${uploadedCount} new run(s), ${runDirs.length - uploadedCount} already in R2.`,
+        '  bun scripts/upload-run.ts results/config-name/2026-03-21-1730\n' +
+        '  bun scripts/upload-run.ts results/config-name',
     )
   }
+
+  const publisher = new R2Publisher({ config: loadR2ConfigFromEnv() })
+  const result = await publisher.publishPath(inputDir)
+  for (const run of result.uploadedRuns) {
+    console.log(`Uploaded ${run.uploadedFiles} files for ${run.runId}`)
+    console.log(run.viewerUrl)
+  }
+  for (const runId of result.skippedRuns) {
+    console.log(`${runId}: already uploaded, skipping`)
+  }
+  console.log(
+    `Done. Uploaded ${result.uploadedRuns.length} run(s), skipped ${result.skippedRuns.length}.`,
+  )
 }
 
-main()
+main().catch((error) => {
+  console.error(error instanceof Error ? error.message : String(error))
+  process.exit(1)
+})

packages/browseros-agent/apps/eval/scripts/weekly-report.ts

@@ -24,45 +24,11 @@ import {
   PutObjectCommand,
   S3Client,
 } from '@aws-sdk/client-s3'
-
-interface ManifestTask {
-  queryId: string
-  query: string
-  status: string
-  durationMs: number
-  screenshotCount: number
-  graderResults: Record<string, { pass: boolean; score: number }>
-}
-
-interface Manifest {
-  runId: string
-  uploadedAt: string
-  agentConfig?: { type?: string; model?: string }
-  dataset?: string
-  summary?: { passRate?: number; avgDurationMs?: number }
-  tasks: ManifestTask[]
-}
-
-interface RunSummary {
-  runId: string
-  configName: string
-  date: string
-  avgScore: number
-  total: number
-  completed: number
-  failed: number
-  timeout: number
-  avgDurationMs: number
-  model: string
-  dataset: string
-  agentType: string
-}
-
-const PASS_FAIL_GRADER_ORDER = [
-  'agisdk_state_diff',
-  'infinity_state',
-  'performance_grader',
-]
+import {
+  buildRunSummaries,
+  type ReportManifest,
+  type RunSummary,
+} from '../src/reporting/run-summary'
 
 function requireEnv(name: string): string {
   const value = process.env[name]
@@ -87,7 +53,7 @@ const client = new S3Client({
 // Step 1: List all manifest.json files in runs/
 console.log('Scanning R2 for eval runs...')
 
-const manifests: Manifest[] = []
+const manifests: ReportManifest[] = []
 let continuationToken: string | undefined
 
 do {
@@ -127,64 +93,9 @@ if (manifests.length === 0) {
 }
 
 // Step 2: Build run summaries
-const runs: RunSummary[] = manifests
-  .map((m) => {
-    const total = m.tasks.length
-    const completed = m.tasks.filter((t) => t.status === 'completed').length
-    const failed = m.tasks.filter((t) => t.status === 'failed').length
-    const timeout = m.tasks.filter((t) => t.status === 'timeout').length
-
-    let scoredCount = 0
-    let scoreSum = 0
-    for (const task of m.tasks) {
-      if (!task.graderResults) continue
-      for (const name of PASS_FAIL_GRADER_ORDER) {
-        if (task.graderResults[name]) {
-          scoredCount++
-          scoreSum += task.graderResults[name].score ?? 0
-          break
-        }
-      }
-    }
-
-    const avgScore = scoredCount > 0 ? (scoreSum / scoredCount) * 100 : 0
-    const durations = m.tasks
-      .filter((t) => t.durationMs > 0)
-      .map((t) => t.durationMs)
-    const avgDurationMs =
-      durations.length > 0
-        ? durations.reduce((a, b) => a + b, 0) / durations.length
-        : 0
-
-    const date = m.uploadedAt
-      ? `${m.uploadedAt.split('T')[0]} ${m.uploadedAt.split('T')[1]?.slice(0, 5) || ''}`
-      : m.runId.slice(0, 15)
-
-    const model = m.agentConfig?.model || 'unknown'
-    const dataset = m.dataset || m.runId
-    const agentType = m.agentConfig?.type || 'unknown'
-
-    const configName = extractConfigName(m.runId)
-    return {
-      runId: m.runId,
-      configName,
-      date,
-      avgScore,
-      total,
-      completed,
-      failed,
-      timeout,
-      avgDurationMs,
-      model,
-      dataset,
-      agentType,
-    }
-  })
-  .sort((a, b) => a.date.localeCompare(b.date))
+const runs: RunSummary[] = buildRunSummaries(manifests)
 
 // Step 3: Identify unique config groups
-// runId can be "ci-weekly" (old) or "ci-weekly-2026-03-21-1730" (timestamped)
-// Extract config name by stripping the date-time suffix pattern
 function escHtml(s: string): string {
   return s
     .replace(/&/g, '&amp;')
@@ -193,12 +104,6 @@ function escHtml(s: string): string {
     .replace(/"/g, '&quot;')
 }
 
-function extractConfigName(runId: string): string {
-  // "browseros-agent-weekly-2026-03-21-1730" → "browseros-agent-weekly"
-  // "ci-weekly" → "ci-weekly" (no timestamp, old format)
-  return runId.replace(/-\d{4}-\d{2}-\d{2}-\d{4}$/, '')
-}
-
 const configGroups = [...new Set(runs.map((r) => r.configName))]
 const defaultConfig = configGroups.includes('ci-weekly')
   ? 'ci-weekly'

packages/browseros-agent/apps/eval/src/agents/orchestrated/backends/clado/clado-action-executor.ts

@@ -1,113 +1,67 @@
 import { randomUUID } from 'node:crypto'
+import { MAX_ACTIONS_PER_DELEGATION } from '../../../../constants'
+import { McpClient, type McpToolResult } from '../../../../utils/mcp-client'
+import { sleep } from '../../../../utils/sleep'
+import type {
+  ExecutorConfig,
+  ExecutorResult,
+} from '../../../orchestrator-executor/types'
+import type { ExecutorCallbacks } from '../../executor-backend'
 import {
-  CLADO_REQUEST_TIMEOUT_MS,
-  MAX_ACTIONS_PER_DELEGATION,
-} from '../../constants'
-import { McpClient, type McpToolResult } from '../../utils/mcp-client'
-import { sleep } from '../../utils/sleep'
-import type { ExecutorCallbacks } from './executor'
-import type { ExecutorConfig, ExecutorResult } from './types'
+  extractCladoThinking,
+  formatCladoHistory,
+  getCladoActionSignature,
+  parseCladoActions,
+  summarizeCladoPrediction,
+} from './clado-actions'
+import {
+  normalizeCladoDirection,
+  normalizeCladoPressKey,
+  normalizeCladoScrollAmount,
+  prepareCladoToolArgs,
+  resolveCladoPoint,
+} from './clado-browser-driver'
+import { CladoActionClient } from './clado-client'
+import {
+  CLADO_ACTION_PROVIDER,
+  type CladoAction,
+  type CladoActionPoint,
+  type CladoActionResponse,
+  type CladoViewport,
+  isCladoActionProvider,
+} from './types'
 
-const CLADO_ACTION_PROVIDER = 'clado-action'
-const PAGE_SCOPED_TOOLS = new Set<string>([
-  'take_screenshot',
-  'evaluate_script',
-  'click',
-  'click_at',
-  'hover',
-  'hover_at',
-  'clear',
-  'fill',
-  'press_key',
-  'type_at',
-  'drag',
-  'drag_at',
-  'scroll',
-  'handle_dialog',
-  'select_option',
-  'navigate_page',
-  'close_page',
-  'wait_for',
-])
-
-interface CladoActionResponse {
-  action?: string
-  x?: number
-  y?: number
-  text?: string
-  key?: string
-  direction?: string
-  startX?: number
-  startY?: number
-  endX?: number
-  endY?: number
-  amount?: number
-  time?: number
-  inference_time_seconds?: number
-  raw_response?: string
-}
-
-interface Viewport {
-  width: number
-  height: number
-}
-
-interface CladoAction {
-  action: string
-  x?: number
-  y?: number
-  text?: string
-  key?: string
-  direction?: string
-  startX?: number
-  startY?: number
-  endX?: number
-  endY?: number
-  amount?: number
-  time?: number
-}
-
-type RawActionPayload = Partial<CladoAction>
-
-interface ActionPoint {
-  x: number
-  y: number
-}
+const MAX_CONSECUTIVE_PARSE_FAILURES = 3
 
 function asErrorMessage(error: unknown): string {
   return error instanceof Error ? error.message : String(error)
 }
 
-function clampNormalized(value: number): number {
-  return Math.min(999, Math.max(0, Math.round(value)))
-}
-
-function isCladoProvider(provider: string): boolean {
-  return provider === CLADO_ACTION_PROVIDER
-}
-
 export class CladoActionExecutor {
   private readonly mcpClient: McpClient
+  private readonly cladoClient: CladoActionClient
   private readonly pageId: number
   private callbacks: ExecutorCallbacks = {}
   private stepsUsed = 0
-  private viewport: Viewport | null = null
-  private lastPoint: ActionPoint | null = null
+  private viewport: CladoViewport | null = null
+  private lastPoint: CladoActionPoint | null = null
   private currentUrl = ''
 
   constructor(
-    private readonly config: ExecutorConfig,
+    config: ExecutorConfig,
     serverUrl: string,
-    readonly _windowId?: number,
-    readonly _tabId?: number,
     initialPageId?: number,
   ) {
-    if (!isCladoProvider(config.provider)) {
+    if (!isCladoActionProvider(config.provider)) {
       throw new Error(
         `CladoActionExecutor requires provider="${CLADO_ACTION_PROVIDER}"`,
       )
     }
     this.mcpClient = new McpClient(`${serverUrl}/mcp`)
+    this.cladoClient = new CladoActionClient({
+      baseUrl: config.baseUrl,
+      apiKey: config.apiKey,
+    })
     this.pageId = initialPageId ?? 1
   }
 
@@ -135,6 +89,8 @@ export class CladoActionExecutor {
     const actionHistory: CladoAction[] = []
     let predictionCalls = 0
     const thinkingTrace: string[] = []
+    let consecutiveParseFailures = 0
+    let finalAnswer: string | undefined
 
     let status: ExecutorResult['status'] = 'done'
     let reason = 'Goal executed.'
@@ -155,7 +111,7 @@ export class CladoActionExecutor {
         break
       }
 
-      const historyForPrediction = this.formatHistory(actionHistory)
+      const historyForPrediction = formatCladoHistory(actionHistory)
       const actionToolCallId = randomUUID()
       const predictionInput = {
         instruction,
@@ -177,7 +133,7 @@ export class CladoActionExecutor {
           signal,
         )
         predictionCalls++
-        const thinking = this.extractThinking(prediction.raw_response)
+        const thinking = extractCladoThinking(prediction.raw_response)
         if (thinking) {
           const previous = thinkingTrace[thinkingTrace.length - 1]
           if (previous !== thinking) {
@@ -207,8 +163,19 @@ export class CladoActionExecutor {
         break
       }
 
-      const predictedActions = this.parseActions(prediction)
+      const predictedActions = parseCladoActions(prediction)
       if (predictedActions.length === 0) {
+        // Per Clado contract: HTTP 200 with action=null on parse failure.
+        // Count as an invalid step so the model can self-correct on the
+        // next call instead of dropping the trajectory.
+        consecutiveParseFailures++
+        const parseError =
+          prediction.parse_error ?? 'no parsable <answer> in raw_response'
+        actionHistory.push({
+          action: 'invalid',
+          text: `parse_error: ${parseError}`,
+        })
+        this.stepsUsed++
         await this.callbacks.onStepFinish?.({
           toolCalls: [
             {
@@ -222,16 +189,23 @@ export class CladoActionExecutor {
               toolCallId: actionToolCallId,
               toolName: 'clado_action_predict',
               output: {
-                prediction: this.summarizePrediction(prediction),
+                prediction: summarizeCladoPrediction(prediction),
                 parsedActions: [],
+                parseError,
+                consecutiveParseFailures,
               },
             },
           ],
         })
-        status = 'blocked'
-        reason = 'Clado action response did not contain a valid action.'
-        break
+
+        if (consecutiveParseFailures >= MAX_CONSECUTIVE_PARSE_FAILURES) {
+          status = 'blocked'
+          reason = `Clado returned ${consecutiveParseFailures} consecutive unparseable responses.`
+          break
+        }
+        continue
       }
+      consecutiveParseFailures = 0
 
       let requestedStop = false
       const executionNotes: string[] = []
@@ -257,7 +231,7 @@ export class CladoActionExecutor {
                 toolCallId: actionToolCallId,
                 toolName: 'clado_action_predict',
                 output: {
-                  prediction: this.summarizePrediction(prediction),
+                  prediction: summarizeCladoPrediction(prediction),
                   parsedActions: predictedActions,
                   executed: executionNotes,
                 },
@@ -272,7 +246,12 @@ export class CladoActionExecutor {
 
         actionHistory.push(predictedAction)
         if (predictedAction.action === 'end') {
-          reason = 'Model requested end() and marked task complete.'
+          if (predictedAction.final_answer) {
+            finalAnswer = predictedAction.final_answer
+            reason = `Model requested end() with final_answer: ${predictedAction.final_answer.slice(0, 240)}`
+          } else {
+            reason = 'Model requested end() and marked task complete.'
+          }
           requestedStop = true
           break
         }
@@ -293,7 +272,7 @@ export class CladoActionExecutor {
               toolCallId: actionToolCallId,
               toolName: 'clado_action_predict',
               output: {
-                prediction: this.summarizePrediction(prediction),
+                prediction: summarizeCladoPrediction(prediction),
                 parsedActions: predictedActions,
                 executed: executionNotes,
               },
@@ -327,6 +306,7 @@ export class CladoActionExecutor {
       actions: actionHistory,
       url: this.currentUrl,
       thinkingTrace,
+      finalAnswer,
     })
 
     return {
@@ -344,121 +324,12 @@ export class CladoActionExecutor {
     actionHistory: CladoAction[],
     signal?: AbortSignal,
   ): Promise<CladoActionResponse> {
-    if (!this.config.baseUrl) {
-      throw new Error('executor.baseUrl must be set for clado-action provider')
-    }
-
-    const requestController = new AbortController()
-    const onAbort = () => requestController.abort()
-    signal?.addEventListener('abort', onAbort, { once: true })
-
-    const timeoutHandle = setTimeout(() => {
-      requestController.abort()
-    }, CLADO_REQUEST_TIMEOUT_MS)
-
-    try {
-      const headers: Record<string, string> = {
-        'Content-Type': 'application/json',
-      }
-      if (this.config.apiKey) {
-        headers.Authorization = `Bearer ${this.config.apiKey}`
-      }
-
-      const response = await fetch(this.config.baseUrl, {
-        method: 'POST',
-        headers,
-        body: JSON.stringify({
-          instruction,
-          image_base64: imageBase64,
-          history: this.formatHistory(actionHistory),
-        }),
-        signal: requestController.signal,
-      })
-
-      if (!response.ok) {
-        const body = await response.text()
-        throw new Error(
-          `HTTP ${response.status} ${response.statusText}: ${body.slice(0, 400)}`,
-        )
-      }
-
-      return (await response.json()) as CladoActionResponse
-    } finally {
-      clearTimeout(timeoutHandle)
-      signal?.removeEventListener('abort', onAbort)
-    }
-  }
-
-  private parseActions(prediction: CladoActionResponse): CladoAction[] {
-    const actionFromField =
-      typeof prediction.action === 'string' ? prediction.action : null
-
-    const rawActions = this.parseActionsFromRawResponse(prediction.raw_response)
-    const primaryFromRaw = rawActions[0] ?? null
-    const mergedPrimary = {
-      ...primaryFromRaw,
-      ...prediction,
-      action: actionFromField ?? primaryFromRaw?.action,
-    }
-
-    const normalized: CladoAction[] = []
-    const primary = this.normalizeActionPayload(mergedPrimary)
-    if (primary) normalized.push(primary)
-
-    for (const candidate of rawActions.slice(1)) {
-      const parsed = this.normalizeActionPayload(candidate)
-      if (!parsed) continue
-      const prev = normalized[normalized.length - 1]
-      if (
-        !prev ||
-        this.getActionSignature(prev) !== this.getActionSignature(parsed)
-      ) {
-        normalized.push(parsed)
-      }
-    }
-
-    return normalized
-  }
-
-  private normalizeActionPayload(
-    payload: RawActionPayload,
-  ): CladoAction | null {
-    if (!payload.action || typeof payload.action !== 'string') {
-      return null
-    }
-    return {
-      action: payload.action,
-      x: typeof payload.x === 'number' ? payload.x : undefined,
-      y: typeof payload.y === 'number' ? payload.y : undefined,
-      text: typeof payload.text === 'string' ? payload.text : undefined,
-      key: typeof payload.key === 'string' ? payload.key : undefined,
-      direction:
-        typeof payload.direction === 'string' ? payload.direction : undefined,
-      startX: typeof payload.startX === 'number' ? payload.startX : undefined,
-      startY: typeof payload.startY === 'number' ? payload.startY : undefined,
-      endX: typeof payload.endX === 'number' ? payload.endX : undefined,
-      endY: typeof payload.endY === 'number' ? payload.endY : undefined,
-      amount: typeof payload.amount === 'number' ? payload.amount : undefined,
-      time: typeof payload.time === 'number' ? payload.time : undefined,
-    }
-  }
-
-  private parseActionsFromRawResponse(
-    rawResponse: string | undefined,
-  ): RawActionPayload[] {
-    if (!rawResponse) return []
-    const matches = [
-      ...rawResponse.matchAll(/<answer>\s*([\s\S]*?)\s*<\/answer>/gi),
-    ]
-    const parsed: RawActionPayload[] = []
-    for (const match of matches) {
-      try {
-        parsed.push(JSON.parse(match[1]) as RawActionPayload)
-      } catch {
-        // ignore malformed answer blocks
-      }
-    }
-    return parsed
+    return this.cladoClient.requestActionPrediction({
+      instruction,
+      imageBase64,
+      actionHistory,
+      signal,
+    })
   }
 
   private async executeAction(
@@ -529,14 +400,14 @@ export class CladoActionExecutor {
       }
 
       case 'press_key': {
-        const key = this.normalizePressKey(action.key)
+        const key = normalizeCladoPressKey(action.key)
         await this.runTool('press_key', { key }, signal)
         return `Pressed key "${key}".`
       }
 
       case 'scroll': {
-        const direction = this.normalizeDirection(action.direction)
-        const amountPx = this.normalizeScrollAmount(action.amount)
+        const direction = normalizeCladoDirection(action.direction)
+        const amountPx = normalizeCladoScrollAmount(action.amount)
         const ticks = Math.max(1, Math.round(amountPx / 120))
 
         await this.runTool('scroll', { direction, amount: ticks }, signal)
@@ -578,7 +449,9 @@ export class CladoActionExecutor {
       }
 
       case 'end': {
-        return 'Model requested end().'
+        return action.final_answer
+          ? `Model requested end() with final_answer: ${action.final_answer.slice(0, 240)}`
+          : 'Model requested end().'
       }
 
       default: {
@@ -588,9 +461,10 @@ export class CladoActionExecutor {
   }
 
   private async captureScreenshotBase64(signal?: AbortSignal): Promise<string> {
+    // Clado contract is PNG or JPEG; use PNG for lossless input.
     const result = await this.runTool(
       'take_screenshot',
-      { format: 'webp', quality: 80 },
+      { format: 'png' },
       signal,
     )
 
@@ -604,7 +478,7 @@ export class CladoActionExecutor {
     return image.data
   }
 
-  private async getViewport(signal?: AbortSignal): Promise<Viewport> {
+  private async getViewport(signal?: AbortSignal): Promise<CladoViewport> {
     if (this.viewport) return this.viewport
 
     try {
@@ -635,15 +509,9 @@ export class CladoActionExecutor {
     normalizedX: number | undefined,
     normalizedY: number | undefined,
     signal?: AbortSignal,
-  ): Promise<ActionPoint> {
+  ): Promise<CladoActionPoint> {
     const viewport = await this.getViewport(signal)
-    const nx = clampNormalized(normalizedX ?? 500)
-    const ny = clampNormalized(normalizedY ?? 500)
-
-    return {
-      x: Math.round((nx / 1000) * viewport.width),
-      y: Math.round((ny / 1000) * viewport.height),
-    }
+    return resolveCladoPoint(viewport, normalizedX, normalizedY)
   }
 
   private async getCurrentUrl(signal?: AbortSignal): Promise<string> {
@@ -670,7 +538,7 @@ export class CladoActionExecutor {
       throw new Error('aborted')
     }
 
-    const toolArgs = this.prepareToolArgs(toolName, args)
+    const toolArgs = prepareCladoToolArgs(toolName, args, this.pageId)
 
     try {
       const raw = await this.mcpClient.callTool(toolName, toolArgs)
@@ -689,211 +557,22 @@ export class CladoActionExecutor {
     }
   }
 
-  private prepareToolArgs(
-    toolName: string,
-    args: Record<string, unknown>,
-  ): Record<string, unknown> {
-    const prepared: Record<string, unknown> = { ...args }
-
-    if (
-      toolName === 'evaluate_script' &&
-      typeof prepared.function === 'string' &&
-      prepared.expression === undefined
-    ) {
-      prepared.expression = this.toEvaluateExpression(prepared.function)
-      delete prepared.function
-    }
-
-    if (
-      toolName === 'click_at' &&
-      typeof prepared.dblClick === 'boolean' &&
-      prepared.clickCount === undefined
-    ) {
-      prepared.clickCount = prepared.dblClick ? 2 : 1
-      delete prepared.dblClick
-    }
-
-    // Use fixed page ID for all page-scoped tools (single-page operation)
-    if (PAGE_SCOPED_TOOLS.has(toolName) && typeof prepared.page !== 'number') {
-      prepared.page = this.pageId
-    }
-
-    return prepared
-  }
-
-  private toEvaluateExpression(rawFunction: unknown): string {
-    const source = String(rawFunction).trim()
-    if (source.startsWith('() =>') || source.startsWith('async () =>')) {
-      return `(${source})()`
-    }
-    if (source.startsWith('function')) {
-      return `(${source})()`
-    }
-    return source
-  }
-
-  private normalizePressKey(key: string | undefined): string {
-    const raw = (key ?? '').trim()
-    if (!raw) throw new Error('press_key action missing key field')
-
-    const map: Record<string, string> = {
-      'C-a': 'Control+A',
-      'C-c': 'Control+C',
-      'C-v': 'Control+V',
-      'C-x': 'Control+X',
-      'C-z': 'Control+Z',
-      'C-y': 'Control+Y',
-      'C-s': 'Control+S',
-      'C-t': 'Control+T',
-      'C-w': 'Control+W',
-      'C-h': 'Control+H',
-      'C-f': 'Control+F',
-      'C-+': 'Control++',
-      'C--': 'Control+-',
-      'C-tab': 'Control+Tab',
-      'C-S-tab': 'Control+Shift+Tab',
-      'C-S-n': 'Control+Shift+N',
-      'C-down': 'Control+ArrowDown',
-      'M-f4': 'Alt+F4',
-    }
-    return map[raw] ?? raw
-  }
-
-  private normalizeDirection(
-    direction: string | undefined,
-  ): 'up' | 'down' | 'left' | 'right' {
-    if (
-      direction === 'up' ||
-      direction === 'down' ||
-      direction === 'left' ||
-      direction === 'right'
-    ) {
-      return direction
-    }
-    return 'down'
-  }
-
-  private normalizeScrollAmount(amount: number | undefined): number {
-    if (typeof amount !== 'number') return 500
-    if (amount <= 0) return 100
-    const clamped = Math.min(amount, 1000)
-    return Math.max(100, Math.round((clamped / 1000) * 900))
-  }
-
-  private summarizePrediction(
-    prediction: CladoActionResponse,
-  ): Record<string, unknown> {
-    const preview =
-      typeof prediction.raw_response === 'string' &&
-      prediction.raw_response.length > 0
-        ? prediction.raw_response.slice(0, 240)
-        : undefined
-
-    return {
-      action: prediction.action,
-      x: prediction.x,
-      y: prediction.y,
-      text: prediction.text,
-      key: prediction.key,
-      direction: prediction.direction,
-      startX: prediction.startX,
-      startY: prediction.startY,
-      endX: prediction.endX,
-      endY: prediction.endY,
-      amount: prediction.amount,
-      time: prediction.time,
-      inference_time_seconds: prediction.inference_time_seconds,
-      raw_response_preview: preview,
-    }
-  }
-
-  private extractThinking(rawResponse: string | undefined): string | undefined {
-    if (!rawResponse) return undefined
-    const matches = [
-      ...rawResponse.matchAll(/<thinking>\s*([\s\S]*?)\s*<\/thinking>/gi),
-    ]
-    if (matches.length === 0) return undefined
-
-    const merged = matches
-      .map((match) => match[1]?.replace(/\s+/g, ' ').trim() ?? '')
-      .filter((value) => value.length > 0)
-      .join(' ')
-
-    if (!merged) return undefined
-    return merged
-  }
-
-  private getActionSignature(action: CladoAction): string {
-    switch (action.action) {
-      case 'click':
-      case 'double_click':
-      case 'right_click':
-      case 'hover':
-        return `${action.action}:${action.x ?? 'x'}:${action.y ?? 'y'}`
-      case 'type':
-        return `${action.action}:${(action.text ?? '').slice(0, 16)}`
-      case 'press_key':
-        return `${action.action}:${action.key ?? 'key'}`
-      case 'scroll':
-        return `${action.action}:${action.direction ?? 'down'}:${action.amount ?? 500}`
-      case 'drag':
-        return `${action.action}:${action.startX}:${action.startY}:${action.endX}:${action.endY}`
-      case 'wait':
-        return `${action.action}:${action.time ?? 1}`
-      case 'end':
-        return 'end()'
-      default:
-        return action.action
-    }
-  }
-
-  private formatHistory(actions: CladoAction[]): string {
-    if (actions.length === 0) return 'None'
-
-    const parts = actions.map((action) => {
-      switch (action.action) {
-        case 'click':
-        case 'double_click':
-        case 'right_click':
-        case 'hover':
-          return `${action.action}(${Math.round(action.x ?? 500)}, ${Math.round(action.y ?? 500)})`
-        case 'type': {
-          const text = (action.text ?? '').replace(/'/g, "\\'")
-          return `type('${text}')`
-        }
-        case 'press_key':
-          return `press_key('${action.key ?? 'Enter'}')`
-        case 'scroll':
-          return `scroll(${action.direction ?? 'down'})`
-        case 'drag':
-          return `drag(${Math.round(action.startX ?? 500)},${Math.round(action.startY ?? 500)} -> ${Math.round(action.endX ?? 500)},${Math.round(action.endY ?? 500)})`
-        case 'wait':
-          return `wait(${Math.round(action.time ?? 1)}s)`
-        case 'end':
-          return 'end()'
-        default:
-          return action.action
-      }
-    })
-
-    return parts.join(' -> ')
-  }
-
   private buildObservation(params: {
     status: ExecutorResult['status']
     reason: string
     actions: CladoAction[]
     url: string
     thinkingTrace: string[]
+    finalAnswer?: string
   }): string {
-    const { status, reason, actions, url, thinkingTrace } = params
+    const { status, reason, actions, url, thinkingTrace, finalAnswer } = params
     const actionSummary =
       actions.length === 0
         ? 'No actions were executed.'
         : actions
             .slice(-5)
             .map(
-              (action, idx) => `${idx + 1}. ${this.getActionSignature(action)}`,
+              (action, idx) => `${idx + 1}. ${getCladoActionSignature(action)}`,
             )
             .join('\n')
     const thinkingSummary =
@@ -907,6 +586,7 @@ export class CladoActionExecutor {
       `Status: ${status}`,
       `Reason: ${reason}`,
       `URL: ${url || 'unknown'}`,
+      finalAnswer ? `Final answer: ${finalAnswer}` : '',
       '',
       'Recent actions:',
       actionSummary,

packages/browseros-agent/apps/eval/src/agents/orchestrated/backends/clado/clado-actions.ts

@@ -0,0 +1,191 @@
+import type {
+  CladoAction,
+  CladoActionResponse,
+  RawCladoActionPayload,
+} from './types'
+
+/** Parses Clado's structured response plus any raw `<answer>` blocks into executable actions. */
+export function parseCladoActions(
+  prediction: CladoActionResponse,
+): CladoAction[] {
+  const actionFromField =
+    typeof prediction.action === 'string' ? prediction.action : null
+
+  const rawActions = parseCladoActionsFromRawResponse(prediction.raw_response)
+  const primaryFromRaw = rawActions[0] ?? null
+  const mergedPrimary = {
+    ...primaryFromRaw,
+    ...prediction,
+    action: actionFromField ?? primaryFromRaw?.action,
+  }
+
+  const normalized: CladoAction[] = []
+  const primary = normalizeCladoActionPayload(mergedPrimary)
+  if (primary) normalized.push(primary)
+
+  for (const candidate of rawActions.slice(1)) {
+    const parsed = normalizeCladoActionPayload(candidate)
+    if (!parsed) continue
+    const prev = normalized[normalized.length - 1]
+    if (
+      !prev ||
+      getCladoActionSignature(prev) !== getCladoActionSignature(parsed)
+    ) {
+      normalized.push(parsed)
+    }
+  }
+
+  return normalized
+}
+
+export function normalizeCladoActionPayload(
+  payload: RawCladoActionPayload,
+): CladoAction | null {
+  if (!payload.action || typeof payload.action !== 'string') {
+    return null
+  }
+  return {
+    action: payload.action,
+    x: typeof payload.x === 'number' ? payload.x : undefined,
+    y: typeof payload.y === 'number' ? payload.y : undefined,
+    text: typeof payload.text === 'string' ? payload.text : undefined,
+    key: typeof payload.key === 'string' ? payload.key : undefined,
+    direction:
+      typeof payload.direction === 'string' ? payload.direction : undefined,
+    startX: typeof payload.startX === 'number' ? payload.startX : undefined,
+    startY: typeof payload.startY === 'number' ? payload.startY : undefined,
+    endX: typeof payload.endX === 'number' ? payload.endX : undefined,
+    endY: typeof payload.endY === 'number' ? payload.endY : undefined,
+    amount: typeof payload.amount === 'number' ? payload.amount : undefined,
+    time: typeof payload.time === 'number' ? payload.time : undefined,
+    final_answer:
+      typeof payload.final_answer === 'string'
+        ? payload.final_answer
+        : undefined,
+  }
+}
+
+export function parseCladoActionsFromRawResponse(
+  rawResponse: string | undefined,
+): RawCladoActionPayload[] {
+  if (!rawResponse) return []
+  const matches = [
+    ...rawResponse.matchAll(/<answer>\s*([\s\S]*?)\s*<\/answer>/gi),
+  ]
+  const parsed: RawCladoActionPayload[] = []
+  for (const match of matches) {
+    try {
+      parsed.push(JSON.parse(match[1]) as RawCladoActionPayload)
+    } catch {
+      // Ignore malformed answer blocks so one bad block does not drop the whole prediction.
+    }
+  }
+  return parsed
+}
+
+export function extractCladoThinking(
+  rawResponse: string | undefined,
+): string | undefined {
+  if (!rawResponse) return undefined
+  const matches = [
+    ...rawResponse.matchAll(/<thinking>\s*([\s\S]*?)\s*<\/thinking>/gi),
+  ]
+  if (matches.length === 0) return undefined
+
+  const merged = matches
+    .map((match) => match[1]?.replace(/\s+/g, ' ').trim() ?? '')
+    .filter((value) => value.length > 0)
+    .join(' ')
+
+  if (!merged) return undefined
+  return merged
+}
+
+export function summarizeCladoPrediction(
+  prediction: CladoActionResponse,
+): Record<string, unknown> {
+  const preview =
+    typeof prediction.raw_response === 'string' &&
+    prediction.raw_response.length > 0
+      ? prediction.raw_response.slice(0, 240)
+      : undefined
+
+  return {
+    action: prediction.action,
+    x: prediction.x,
+    y: prediction.y,
+    text: prediction.text,
+    key: prediction.key,
+    direction: prediction.direction,
+    startX: prediction.startX,
+    startY: prediction.startY,
+    endX: prediction.endX,
+    endY: prediction.endY,
+    amount: prediction.amount,
+    time: prediction.time,
+    inference_time_seconds: prediction.inference_time_seconds,
+    raw_response_preview: preview,
+  }
+}
+
+export function getCladoActionSignature(action: CladoAction): string {
+  switch (action.action) {
+    case 'click':
+    case 'double_click':
+    case 'right_click':
+    case 'hover':
+      return `${action.action}:${action.x ?? 'x'}:${action.y ?? 'y'}`
+    case 'type':
+      return `${action.action}:${(action.text ?? '').slice(0, 16)}`
+    case 'press_key':
+      return `${action.action}:${action.key ?? 'key'}`
+    case 'scroll':
+      return `${action.action}:${action.direction ?? 'down'}:${action.amount ?? 500}`
+    case 'drag':
+      return `${action.action}:${action.startX}:${action.startY}:${action.endX}:${action.endY}`
+    case 'wait':
+      return `${action.action}:${action.time ?? 1}`
+    case 'end':
+      return action.final_answer
+        ? `end(${action.final_answer.slice(0, 32)})`
+        : 'end()'
+    case 'invalid':
+      return `invalid(${(action.text ?? '').slice(0, 40)})`
+    default:
+      return action.action
+  }
+}
+
+export function formatCladoHistory(actions: CladoAction[]): string {
+  if (actions.length === 0) return 'None'
+
+  const parts = actions.map((action) => {
+    switch (action.action) {
+      case 'click':
+      case 'double_click':
+      case 'right_click':
+      case 'hover':
+        return `${action.action}(${Math.round(action.x ?? 500)}, ${Math.round(action.y ?? 500)})`
+      case 'type': {
+        const text = (action.text ?? '').replace(/'/g, "\\'")
+        return `type('${text}')`
+      }
+      case 'press_key':
+        return `press_key('${action.key ?? 'Enter'}')`
+      case 'scroll':
+        return `scroll(${action.direction ?? 'down'})`
+      case 'drag':
+        return `drag(${Math.round(action.startX ?? 500)},${Math.round(action.startY ?? 500)} -> ${Math.round(action.endX ?? 500)},${Math.round(action.endY ?? 500)})`
+      case 'wait':
+        return `wait(${Math.round(action.time ?? 1)}s)`
+      case 'end':
+        return 'end()'
+      case 'invalid':
+        return 'invalid()'
+      default:
+        return action.action
+    }
+  })
+
+  return parts.join(' -> ')
+}

packages/browseros-agent/apps/eval/src/agents/orchestrated/backends/clado/clado-browser-driver.ts

@@ -0,0 +1,123 @@
+import {
+  CLADO_PAGE_SCOPED_TOOLS,
+  type CladoActionPoint,
+  type CladoViewport,
+} from './types'
+
+export function clampCladoNormalizedCoordinate(value: number): number {
+  return Math.min(999, Math.max(0, Math.round(value)))
+}
+
+/** Converts Clado's 0-1000 normalized coordinate space into BrowserOS viewport pixels. */
+export function resolveCladoPoint(
+  viewport: CladoViewport,
+  normalizedX: number | undefined,
+  normalizedY: number | undefined,
+): CladoActionPoint {
+  const nx = clampCladoNormalizedCoordinate(normalizedX ?? 500)
+  const ny = clampCladoNormalizedCoordinate(normalizedY ?? 500)
+
+  return {
+    x: Math.round((nx / 1000) * viewport.width),
+    y: Math.round((ny / 1000) * viewport.height),
+  }
+}
+
+/** Adapts Clado action tool arguments to the BrowserOS MCP tool argument contract. */
+export function prepareCladoToolArgs(
+  toolName: string,
+  args: Record<string, unknown>,
+  pageId: number,
+): Record<string, unknown> {
+  const prepared: Record<string, unknown> = { ...args }
+
+  if (
+    toolName === 'evaluate_script' &&
+    typeof prepared.function === 'string' &&
+    prepared.expression === undefined
+  ) {
+    prepared.expression = toCladoEvaluateExpression(prepared.function)
+    delete prepared.function
+  }
+
+  if (
+    toolName === 'click_at' &&
+    typeof prepared.dblClick === 'boolean' &&
+    prepared.clickCount === undefined
+  ) {
+    prepared.clickCount = prepared.dblClick ? 2 : 1
+    delete prepared.dblClick
+  }
+
+  if (
+    CLADO_PAGE_SCOPED_TOOLS.has(toolName) &&
+    typeof prepared.page !== 'number'
+  ) {
+    prepared.page = pageId
+  }
+
+  return prepared
+}
+
+export function toCladoEvaluateExpression(rawFunction: unknown): string {
+  const source = String(rawFunction).trim()
+  if (source.startsWith('() =>') || source.startsWith('async () =>')) {
+    return `(${source})()`
+  }
+  if (source.startsWith('function')) {
+    return `(${source})()`
+  }
+  return source
+}
+
+export function normalizeCladoPressKey(key: string | undefined): string {
+  const raw = (key ?? '').trim()
+  if (!raw) throw new Error('press_key action missing key field')
+
+  const map: Record<string, string> = {
+    'C-a': 'Control+A',
+    'C-c': 'Control+C',
+    'C-v': 'Control+V',
+    'C-x': 'Control+X',
+    'C-z': 'Control+Z',
+    'C-y': 'Control+Y',
+    'C-s': 'Control+S',
+    'C-t': 'Control+T',
+    'C-w': 'Control+W',
+    'C-h': 'Control+H',
+    'C-f': 'Control+F',
+    'C-+': 'Control++',
+    'C--': 'Control+-',
+    'C-tab': 'Control+Tab',
+    'C-S-tab': 'Control+Shift+Tab',
+    'C-S-n': 'Control+Shift+N',
+    'C-down': 'Control+ArrowDown',
+    'M-a': 'Meta+A',
+    'M-c': 'Meta+C',
+    'M-v': 'Meta+V',
+    'M-x': 'Meta+X',
+    'M-f4': 'Alt+F4',
+  }
+  return map[raw] ?? raw
+}
+
+export function normalizeCladoDirection(
+  direction: string | undefined,
+): 'up' | 'down' | 'left' | 'right' {
+  if (
+    direction === 'up' ||
+    direction === 'down' ||
+    direction === 'left' ||
+    direction === 'right'
+  ) {
+    return direction
+  }
+  return 'down'
+}
+
+export function normalizeCladoScrollAmount(amount: number | undefined): number {
+  if (typeof amount !== 'number') return 500
+  if (amount <= 0) return 100
+  const clamped = Math.min(amount, 1000)
+  return Math.max(100, Math.round((clamped / 1000) * 900))
+}

packages/browseros-agent/apps/eval/src/agents/orchestrated/backends/clado/clado-client.ts

@@ -0,0 +1,68 @@
+import { CLADO_REQUEST_TIMEOUT_MS } from '../../../../constants'
+import { formatCladoHistory } from './clado-actions'
+import type { CladoAction, CladoActionResponse } from './types'
+
+export interface CladoActionClientOptions {
+  baseUrl?: string
+  apiKey?: string
+}
+
+export interface CladoActionPredictionInput {
+  instruction: string
+  imageBase64: string
+  actionHistory: CladoAction[]
+  signal?: AbortSignal
+}
+
+/** Calls the Clado action model without exposing credentials in process arguments or artifacts. */
+export class CladoActionClient {
+  constructor(private readonly options: CladoActionClientOptions) {}
+
+  async requestActionPrediction(
+    input: CladoActionPredictionInput,
+  ): Promise<CladoActionResponse> {
+    if (!this.options.baseUrl) {
+      throw new Error('executor.baseUrl must be set for clado-action provider')
+    }
+
+    const requestController = new AbortController()
+    const onAbort = () => requestController.abort()
+    input.signal?.addEventListener('abort', onAbort, { once: true })
+
+    const timeoutHandle = setTimeout(() => {
+      requestController.abort()
+    }, CLADO_REQUEST_TIMEOUT_MS)
+
+    try {
+      const headers: Record<string, string> = {
+        'Content-Type': 'application/json',
+      }
+      if (this.options.apiKey) {
+        headers.Authorization = `Bearer ${this.options.apiKey}`
+      }
+
+      const response = await fetch(this.options.baseUrl, {
+        method: 'POST',
+        headers,
+        body: JSON.stringify({
+          instruction: input.instruction,
+          image_base64: input.imageBase64,
+          history: formatCladoHistory(input.actionHistory),
+        }),
+        signal: requestController.signal,
+      })
+
+      if (!response.ok) {
+        const body = await response.text()
+        throw new Error(
+          `HTTP ${response.status} ${response.statusText}: ${body.slice(0, 400)}`,
+        )
+      }
+
+      return (await response.json()) as CladoActionResponse
+    } finally {
+      clearTimeout(timeoutHandle)
+      input.signal?.removeEventListener('abort', onAbort)
+    }
+  }
+}

packages/browseros-agent/apps/eval/src/agents/orchestrated/backends/clado/clado-executor-backend.ts

@@ -0,0 +1,56 @@
+import type { ResolvedAgentConfig } from '@browseros/server/agent/types'
+import type {
+  DelegationResult,
+  ExecutorBackend,
+  ExecutorCallbacks,
+} from '../../executor-backend'
+import { CladoActionExecutor } from './clado-action-executor'
+
+export interface CladoExecutorBackendOptions {
+  configTemplate: ResolvedAgentConfig
+  serverUrl: string
+  initialPageId?: number
+  callbacks?: ExecutorCallbacks
+}
+
+/** Executes delegated goals through the Clado visual action model. */
+export class CladoExecutorBackend implements ExecutorBackend {
+  readonly kind = 'clado'
+  private executor: CladoActionExecutor | null = null
+
+  constructor(private readonly options: CladoExecutorBackendOptions) {}
+
+  async execute(
+    instruction: string,
+    signal?: AbortSignal,
+  ): Promise<DelegationResult> {
+    const executor = this.getExecutor()
+    const result = await executor.execute(instruction, signal)
+    return result
+  }
+
+  async close(): Promise<void> {
+    await this.executor?.close()
+  }
+
+  getTotalSteps(): number {
+    return this.executor?.getTotalSteps() ?? 0
+  }
+
+  private getExecutor(): CladoActionExecutor {
+    if (this.executor) return this.executor
+
+    this.executor = new CladoActionExecutor(
+      {
+        provider: this.options.configTemplate.provider,
+        model: this.options.configTemplate.model,
+        apiKey: this.options.configTemplate.apiKey ?? '',
+        baseUrl: this.options.configTemplate.baseUrl,
+      },
+      this.options.serverUrl,
+      this.options.initialPageId,
+    )
+    this.executor.setCallbacks(this.options.callbacks ?? {})
+    return this.executor
+  }
+}

packages/browseros-agent/apps/eval/src/agents/orchestrated/backends/clado/types.ts

@@ -0,0 +1,78 @@
+export const CLADO_ACTION_PROVIDER = 'clado-action'
+
+export const CLADO_PAGE_SCOPED_TOOLS = new Set<string>([
+  'take_screenshot',
+  'evaluate_script',
+  'click',
+  'click_at',
+  'hover',
+  'hover_at',
+  'clear',
+  'fill',
+  'press_key',
+  'type_at',
+  'drag',
+  'drag_at',
+  'scroll',
+  'handle_dialog',
+  'select_option',
+  'navigate_page',
+  'close_page',
+  'wait_for',
+])
+
+export interface CladoActionResponse {
+  action?: string | null
+  x?: number
+  y?: number
+  text?: string
+  key?: string
+  direction?: string
+  startX?: number
+  startY?: number
+  endX?: number
+  endY?: number
+  amount?: number
+  time?: number
+  final_answer?: string | null
+  inference_time_seconds?: number
+  raw_response?: string
+  thinking?: string | null
+  parse_error?: string | null
+}
+
+export interface CladoViewport {
+  width: number
+  height: number
+}
+
+export interface CladoAction {
+  action: string
+  x?: number
+  y?: number
+  text?: string
+  key?: string
+  direction?: string
+  startX?: number
+  startY?: number
+  endX?: number
+  endY?: number
+  amount?: number
+  time?: number
+  final_answer?: string
+}
+
+export type RawCladoActionPayload = Partial<
+  Omit<CladoAction, 'final_answer'>
+> & {
+  final_answer?: string | null
+}
+
+export interface CladoActionPoint {
+  x: number
+  y: number
+}
+
+export function isCladoActionProvider(provider: string): boolean {
+  return provider === CLADO_ACTION_PROVIDER
+}

packages/browseros-agent/apps/eval/src/agents/orchestrated/backends/create-executor-backend.ts

@@ -0,0 +1,60 @@
+import type { ResolvedAgentConfig } from '@browseros/server/agent/types'
+import type { Browser } from '@browseros/server/browser'
+import type {
+  ExecutorBackend,
+  ExecutorBackendKind,
+  ExecutorCallbacks,
+} from '../executor-backend'
+import { CladoExecutorBackend } from './clado/clado-executor-backend'
+import { isCladoActionProvider } from './clado/types'
+import { ToolLoopExecutorBackend } from './tool-loop/tool-loop-executor-backend'
+
+export interface CreateExecutorBackendOptions {
+  backendKind?: ExecutorBackendKind
+  provider?: string
+  configTemplate?: ResolvedAgentConfig
+  browser?: Browser | null
+  serverUrl?: string
+  windowId?: number
+  tabId?: number
+  initialPageId?: number
+  callbacks?: ExecutorCallbacks
+  executor?: ExecutorBackend
+}
+
+export function backendKindForProvider(provider: string): ExecutorBackendKind {
+  return isCladoActionProvider(provider) ? 'clado' : 'tool-loop'
+}
+
+/** Creates the backend used for one orchestrator delegation. */
+export function createExecutorBackend(
+  options: CreateExecutorBackendOptions,
+): ExecutorBackend {
+  if (options.executor) return options.executor
+
+  const kind =
+    options.backendKind ??
+    backendKindForProvider(
+      options.provider ?? options.configTemplate?.provider ?? '',
+    )
+
+  if (kind === 'clado') {
+    return new CladoExecutorBackend({
+      configTemplate: required(options.configTemplate, 'configTemplate'),
+      serverUrl: required(options.serverUrl, 'serverUrl'),
+      initialPageId: options.initialPageId,
+      callbacks: options.callbacks,
+    })
+  }
+
+  return new ToolLoopExecutorBackend({
+    configTemplate: required(options.configTemplate, 'configTemplate'),
+    browser: options.browser ?? null,
+    callbacks: options.callbacks,
+  })
+}
+
+function required<T>(value: T | undefined, name: string): T {
+  if (value === undefined) throw new Error(`${name} is required`)
+  return value
+}

packages/browseros-agent/apps/eval/src/agents/orchestrated/backends/tool-loop/tool-loop-executor-backend.ts

@@ -0,0 +1,144 @@
+import { randomUUID } from 'node:crypto'
+import { AiSdkAgent } from '@browseros/server/agent/tool-loop'
+import type { ResolvedAgentConfig } from '@browseros/server/agent/types'
+import type { Browser } from '@browseros/server/browser'
+import { registry } from '@browseros/server/tools/registry'
+import type { BrowserContext } from '@browseros/shared/schemas/browser-context'
+import type {
+  DelegationResult,
+  ExecutorBackend,
+  ExecutorCallbacks,
+} from '../../executor-backend'
+import { TOOL_LOOP_EXECUTOR_SYSTEM_PROMPT } from './tool-loop-executor-prompt'
+
+export interface ToolLoopExecutorBackendOptions {
+  configTemplate: ResolvedAgentConfig
+  browser: Browser | null
+  callbacks?: ExecutorCallbacks
+}
+
+/** Executes delegated goals through the BrowserOS ToolLoopAgent. */
+export class ToolLoopExecutorBackend implements ExecutorBackend {
+  readonly kind = 'tool-loop'
+  private stepsUsed = 0
+  private currentUrl = ''
+
+  constructor(private readonly options: ToolLoopExecutorBackendOptions) {}
+
+  async execute(
+    instruction: string,
+    signal?: AbortSignal,
+  ): Promise<DelegationResult> {
+    const browser = this.options.browser
+    if (!browser) {
+      throw new Error('Browser instance is required for tool-loop executor')
+    }
+
+    const stepsAtStart = this.stepsUsed
+    const toolsUsed: string[] = []
+    let status: DelegationResult['status'] = 'done'
+    let resultText = ''
+
+    const conversationId = randomUUID()
+    const agentConfig: ResolvedAgentConfig = {
+      ...this.options.configTemplate,
+      conversationId,
+      userSystemPrompt: TOOL_LOOP_EXECUTOR_SYSTEM_PROMPT,
+      evalMode: true,
+      workingDir: `/tmp/browseros-eval-executor-${conversationId}`,
+    }
+
+    const browserContext = await this.browserContext(browser)
+    let agent: AiSdkAgent | null = null
+
+    try {
+      agent = await AiSdkAgent.create({
+        resolvedConfig: agentConfig,
+        browser,
+        registry,
+        browserContext,
+      })
+
+      await agent.toolLoopAgent.generate({
+        prompt: instruction,
+        abortSignal: signal,
+
+        experimental_onToolCallStart: ({ toolCall }) => {
+          const input = toolCall.input as Record<string, unknown> | undefined
+          if (input && typeof input.url === 'string' && input.url.length > 0) {
+            this.currentUrl = input.url
+          }
+          this.options.callbacks?.onToolCallStart?.({
+            toolCallId: toolCall.toolCallId,
+            toolName: toolCall.toolName,
+            input: toolCall.input,
+          })
+        },
+
+        experimental_onToolCallFinish: async () => {
+          this.stepsUsed++
+          await this.options.callbacks?.onToolCallFinish?.()
+        },
+
+        onStepFinish: async ({ toolCalls, toolResults, text }) => {
+          if (toolCalls) {
+            for (const toolCall of toolCalls) {
+              if (!toolsUsed.includes(toolCall.toolName)) {
+                toolsUsed.push(toolCall.toolName)
+              }
+            }
+          }
+
+          if (text) resultText = text
+
+          await this.options.callbacks?.onStepFinish?.({
+            toolCalls,
+            toolResults,
+            text,
+          })
+        },
+      })
+    } catch {
+      status = signal?.aborted ? 'timeout' : 'blocked'
+    } finally {
+      if (agent) await agent.dispose().catch(() => {})
+    }
+
+    if (status === 'done' && signal?.aborted) {
+      status = 'timeout'
+    }
+
+    return {
+      observation: resultText || 'Execution completed with no actions taken.',
+      status,
+      url: this.currentUrl,
+      actionsPerformed: this.stepsUsed - stepsAtStart,
+      toolsUsed,
+    }
+  }
+
+  async close(): Promise<void> {
+    // No persistent resources; AiSdkAgent is disposed at the end of each execute() call.
+  }
+
+  getTotalSteps(): number {
+    return this.stepsUsed
+  }
+
+  private async browserContext(
+    browser: Browser,
+  ): Promise<BrowserContext | undefined> {
+    const pages = await browser.listPages()
+    const activePage = pages[0]
+    if (!activePage) return undefined
+
+    return {
+      activeTab: {
+        id: activePage.tabId,
+        pageId: activePage.pageId,
+        url: activePage.url,
+        title: activePage.title,
+      },
+    }
+  }
+}

packages/browseros-agent/apps/eval/src/agents/orchestrated/backends/tool-loop/tool-loop-executor-prompt.ts

@@ -0,0 +1,21 @@
+export const TOOL_LOOP_EXECUTOR_SYSTEM_PROMPT = `You are a browser executor. You receive a single goal-level instruction and execute it using browser tools.
+
+## Your Job
+1. Execute browser actions to achieve the given goal
+2. Stop as soon as the goal is accomplished -- do NOT perform extra actions
+3. Write a final observation describing the result
+
+## Final Response Format
+When done, your response MUST include:
+- What you accomplished (or what went wrong)
+- What the page currently shows: key headings, links, data, or content visible
+- The current URL from the address bar
+- If you got stuck, what is blocking progress
+
+## Rules
+- Only do what was asked. Do not navigate away, open extra tabs, or reorganize the browser.
+- If the goal is to navigate somewhere, confirm you arrived by describing what you see.
+- If the goal is to click something, confirm the result of the click.
+- If you cannot find what was asked for, say so clearly -- do not guess or improvise.
+- Prefer browser_navigate over browser_open_tab for going to URLs.
+- Do NOT call browser_group_tabs or other organizational tools.`

packages/browseros-agent/apps/eval/src/agents/orchestrated/executor-backend.ts

@@ -0,0 +1,33 @@
+import type { ExecutorResult } from '../orchestrator-executor/types'
+
+export type ExecutorBackendKind = 'tool-loop' | 'clado'
+export type DelegationResult = ExecutorResult
+
+export interface ToolCallInfo {
+  toolCallId: string
+  toolName: string
+  input: unknown
+}
+
+export interface ToolResultInfo {
+  toolCallId: string
+  toolName: string
+  output: unknown
+}
+
+export interface ExecutorCallbacks {
+  onToolCallStart?: (toolCall: ToolCallInfo) => void
+  onToolCallFinish?: () => Promise<void>
+  onStepFinish?: (step: {
+    toolCalls?: ReadonlyArray<ToolCallInfo>
+    toolResults?: ReadonlyArray<ToolResultInfo>
+    text?: string
+  }) => Promise<void>
+}
+
+export interface ExecutorBackend {
+  readonly kind: ExecutorBackendKind
+  execute(instruction: string, signal?: AbortSignal): Promise<DelegationResult>
+  close(): Promise<void>
+  getTotalSteps(): number
+}

packages/browseros-agent/apps/eval/src/agents/orchestrator-executor/executor.ts

@@ -1,243 +0,0 @@
-/**
- * Executor - Wraps AiSdkAgent for page-level browser actions (direct CDP)
- *
- * The executor:
- * - Receives goal-level instructions from orchestrator
- * - Executes browser actions until the goal is accomplished
- * - Returns observation to orchestrator (not full history)
- */
-
-import { randomUUID } from 'node:crypto'
-import { AiSdkAgent } from '@browseros/server/agent/tool-loop'
-import type { ResolvedAgentConfig } from '@browseros/server/agent/types'
-import type { Browser } from '@browseros/server/browser'
-import { registry } from '@browseros/server/tools/registry'
-import type { BrowserContext } from '@browseros/shared/schemas/browser-context'
-import { CladoActionExecutor } from './clado-action-executor'
-import type { ExecutorResult } from './types'
-
-const EXECUTOR_SYSTEM_PROMPT = `You are a browser executor. You receive a single goal-level instruction and execute it using browser tools.
-
-## Your Job
-1. Execute browser actions to achieve the given goal
-2. Stop as soon as the goal is accomplished — do NOT perform extra actions
-3. Write a final observation describing the result
-
-## Final Response Format
-When done, your response MUST include:
-- What you accomplished (or what went wrong)
-- What the page currently shows: key headings, links, data, or content visible
-- The current URL from the address bar
-- If you got stuck, what is blocking progress
-
-## Rules
-- Only do what was asked. Do not navigate away, open extra tabs, or reorganize the browser.
-- If the goal is to navigate somewhere, confirm you arrived by describing what you see.
-- If the goal is to click something, confirm the result of the click.
-- If you cannot find what was asked for, say so clearly — do not guess or improvise.
-- Prefer browser_navigate over browser_open_tab for going to URLs.
-- Do NOT call browser_group_tabs or other organizational tools.`
-
-export interface ToolCallInfo {
-  toolCallId: string
-  toolName: string
-  input: unknown
-}
-
-export interface ToolResultInfo {
-  toolCallId: string
-  toolName: string
-  output: unknown
-}
-
-export interface ExecutorCallbacks {
-  onToolCallStart?: (toolCall: ToolCallInfo) => void
-  onToolCallFinish?: () => Promise<void>
-  onStepFinish?: (step: {
-    toolCalls?: ReadonlyArray<ToolCallInfo>
-    toolResults?: ReadonlyArray<ToolResultInfo>
-    text?: string
-  }) => Promise<void>
-}
-
-export class Executor {
-  private cladoExecutor: CladoActionExecutor | null = null
-  private stepsUsed = 0
-  private currentUrl = ''
-  private configTemplate: ResolvedAgentConfig
-  private isCladoAction: boolean
-  private browser: Browser | null
-  private serverUrl: string
-  private windowId?: number
-  private tabId?: number
-  private initialPageId?: number
-  private callbacks: ExecutorCallbacks
-
-  constructor(
-    configTemplate: ResolvedAgentConfig,
-    browser: Browser | null,
-    serverUrl: string,
-    options?: {
-      isCladoAction?: boolean
-      windowId?: number
-      tabId?: number
-      initialPageId?: number
-      callbacks?: ExecutorCallbacks
-    },
-  ) {
-    this.configTemplate = configTemplate
-    this.isCladoAction = options?.isCladoAction ?? false
-    this.browser = browser
-    this.serverUrl = serverUrl
-    this.windowId = options?.windowId
-    this.tabId = options?.tabId
-    this.initialPageId = options?.initialPageId
-    this.callbacks = options?.callbacks ?? {}
-  }
-
-  async execute(
-    instruction: string,
-    signal?: AbortSignal,
-  ): Promise<ExecutorResult> {
-    if (this.isCladoAction) {
-      if (!this.cladoExecutor) {
-        this.cladoExecutor = new CladoActionExecutor(
-          {
-            provider: this.configTemplate.provider,
-            model: this.configTemplate.model,
-            apiKey: this.configTemplate.apiKey ?? '',
-            baseUrl: this.configTemplate.baseUrl,
-          },
-          this.serverUrl,
-          this.windowId,
-          this.tabId,
-          this.initialPageId,
-        )
-        this.cladoExecutor.setCallbacks(this.callbacks)
-      }
-
-      const result = await this.cladoExecutor.execute(instruction, signal)
-      this.stepsUsed = this.cladoExecutor.getTotalSteps()
-      this.currentUrl = result.url || this.currentUrl
-      return result
-    }
-
-    if (!this.browser) {
-      throw new Error('Browser instance is required for standard executor path')
-    }
-
-    const stepsAtStart = this.stepsUsed
-    const toolsUsed: string[] = []
-    let status: 'done' | 'blocked' | 'timeout' = 'done'
-    let resultText = ''
-
-    const conversationId = randomUUID()
-    const agentConfig: ResolvedAgentConfig = {
-      ...this.configTemplate,
-      conversationId,
-      userSystemPrompt: EXECUTOR_SYSTEM_PROMPT,
-      evalMode: true,
-      workingDir: `/tmp/browseros-eval-executor-${conversationId}`,
-    }
-
-    // Build browser context so executor agent knows the correct page ID
-    let browserContext: BrowserContext | undefined
-    if (this.browser) {
-      const pages = await this.browser.listPages()
-      const activePage = pages[0]
-      if (activePage) {
-        browserContext = {
-          activeTab: {
-            id: activePage.tabId,
-            pageId: activePage.pageId,
-            url: activePage.url,
-            title: activePage.title,
-          },
-        }
-      }
-    }
-
-    let agent: AiSdkAgent | null = null
-
-    try {
-      agent = await AiSdkAgent.create({
-        resolvedConfig: agentConfig,
-        browser: this.browser,
-        registry,
-        browserContext,
-      })
-
-      await agent.toolLoopAgent.generate({
-        prompt: instruction,
-        abortSignal: signal,
-
-        experimental_onToolCallStart: ({ toolCall }) => {
-          const input = toolCall.input as Record<string, unknown> | undefined
-          if (input && typeof input.url === 'string' && input.url.length > 0) {
-            this.currentUrl = input.url
-          }
-          this.callbacks.onToolCallStart?.({
-            toolCallId: toolCall.toolCallId,
-            toolName: toolCall.toolName,
-            input: toolCall.input,
-          })
-        },
-
-        experimental_onToolCallFinish: async () => {
-          this.stepsUsed++
-          await this.callbacks.onToolCallFinish?.()
-        },
-
-        onStepFinish: async ({ toolCalls, toolResults, text }) => {
-          if (toolCalls) {
-            for (const tc of toolCalls) {
-              if (!toolsUsed.includes(tc.toolName)) {
-                toolsUsed.push(tc.toolName)
-              }
-            }
-          }
-
-          if (text) {
-            resultText = text
-          }
-
-          await this.callbacks.onStepFinish?.({ toolCalls, toolResults, text })
-        },
-      })
-    } catch {
-      if (signal?.aborted) {
-        status = 'timeout'
-      } else {
-        status = 'blocked'
-      }
-    } finally {
-      if (agent) await agent.dispose().catch(() => {})
-    }
-
-    if (status === 'done' && signal?.aborted) {
-      status = 'timeout'
-    }
-
-    const observation =
-      resultText || 'Execution completed with no actions taken.'
-
-    return {
-      observation,
-      status,
-      url: this.currentUrl,
-      actionsPerformed: this.stepsUsed - stepsAtStart,
-      toolsUsed,
-    }
-  }
-
-  async close(): Promise<void> {
-    await this.cladoExecutor?.close()
-  }
-
-  getTotalSteps(): number {
-    if (this.isCladoAction) {
-      return this.cladoExecutor?.getTotalSteps() ?? 0
-    }
-    return this.stepsUsed
-  }
-}

packages/browseros-agent/apps/eval/src/agents/orchestrator-executor/index.ts

@@ -24,15 +24,16 @@ import {
   resolveProviderConfig,
 } from '../../utils/resolve-provider-config'
 import { withEvalTimeout } from '../../utils/with-eval-timeout'
+import { isCladoActionProvider } from '../orchestrated/backends/clado/types'
+import { createExecutorBackend } from '../orchestrated/backends/create-executor-backend'
+import type { ExecutorCallbacks } from '../orchestrated/executor-backend'
 import type { AgentContext, AgentEvaluator, AgentResult } from '../types'
-import { Executor, type ExecutorCallbacks } from './executor'
 import { OrchestratorAgent } from './orchestrator-agent'
 import type { ExecutorFactory, ExecutorResult } from './types'
 
 interface ResolvedConfigs {
   orchestratorConfig: ResolvedAgentConfig & { maxTurns?: number }
   executorConfig: ResolvedAgentConfig
-  isCladoAction: boolean
 }
 
 function toResolvedAgentConfig(
@@ -67,7 +68,10 @@ async function resolveAgentConfig(
   if (!executorModel) {
     throw new Error('executor.model is required in config')
   }
-  if (config.executor.provider === 'clado-action' && !config.executor.baseUrl) {
+  if (
+    isCladoActionProvider(config.executor.provider) &&
+    !config.executor.baseUrl
+  ) {
     throw new Error(
       'executor.baseUrl is required in config for clado-action provider',
     )
@@ -75,10 +79,8 @@ async function resolveAgentConfig(
 
   const resolvedOrchestrator = await resolveProviderConfig(config.orchestrator)
 
-  const isCladoAction = config.executor.provider === 'clado-action'
-
   let executorConfig: ResolvedAgentConfig
-  if (isCladoAction) {
+  if (isCladoActionProvider(config.executor.provider)) {
     executorConfig = {
       conversationId: crypto.randomUUID(),
       provider: config.executor.provider as ResolvedAgentConfig['provider'],
@@ -107,7 +109,7 @@ async function resolveAgentConfig(
     maxTurns: config.orchestrator.maxTurns,
   }
 
-  return { orchestratorConfig, executorConfig, isCladoAction }
+  return { orchestratorConfig, executorConfig }
 }
 
 export class OrchestratorExecutorEvaluator implements AgentEvaluator {
@@ -127,7 +129,7 @@ export class OrchestratorExecutorEvaluator implements AgentEvaluator {
     }
 
     const agentConfig = config.agent as OrchestratorExecutorConfig
-    const { orchestratorConfig, executorConfig, isCladoAction } =
+    const { orchestratorConfig, executorConfig } =
       await resolveAgentConfig(agentConfig)
 
     // Connect to Chrome via CDP — same per-worker offset used by app-manager.
@@ -235,12 +237,12 @@ export class OrchestratorExecutorEvaluator implements AgentEvaluator {
         await capture.messageLogger.logStreamEvent(delegateInputEvent)
         capture.emitEvent(task.query_id, delegateInputEvent)
 
-        const executor = new Executor(
-          executorConfig,
+        const executor = createExecutorBackend({
+          configTemplate: executorConfig,
           browser,
-          config.browseros.server_url,
-          { isCladoAction, callbacks },
-        )
+          serverUrl: config.browseros.server_url,
+          callbacks,
+        })
         let result: ExecutorResult
         try {
           result = await executor.execute(instruction, signal)
@@ -329,6 +331,5 @@ export class OrchestratorExecutorEvaluator implements AgentEvaluator {
   }
 }
 
-export { Executor } from './executor'
 export { OrchestratorAgent } from './orchestrator-agent'
 export * from './types'

packages/browseros-agent/apps/eval/src/capture/trajectory-saver.ts

@@ -57,6 +57,20 @@ export class TrajectorySaver {
     )
   }
 
+  async saveAttempt(attempt: Record<string, unknown>): Promise<void> {
+    await writeFile(
+      join(this.outputDir, 'attempt.json'),
+      JSON.stringify(attempt, null, 2),
+    )
+  }
+
+  async saveGrades(graderResults: Record<string, GraderResult>): Promise<void> {
+    await writeFile(
+      join(this.outputDir, 'grades.json'),
+      JSON.stringify(graderResults, null, 2),
+    )
+  }
+
   async loadMetadata(): Promise<TaskMetadata> {
     const content = await readFile(
       join(this.outputDir, 'metadata.json'),
@@ -70,6 +84,7 @@ export class TrajectorySaver {
   ): Promise<void> {
     const metadata = await this.loadMetadata()
     metadata.grader_results = graderResults
+    await this.saveGrades(graderResults)
     await this.saveMetadata(metadata)
   }
 

packages/browseros-agent/apps/eval/src/cli/args.ts

@@ -0,0 +1,170 @@
+import { parseArgs } from 'node:util'
+
+export type PublishTarget = 'r2'
+
+export interface LegacyCliArgs {
+  command: 'legacy'
+  configPath?: string
+  help?: boolean
+}
+
+export interface SuiteCliArgs {
+  command: 'suite'
+  configPath?: string
+  suitePath?: string
+  variantId?: string
+  provider?: string
+  model?: string
+  apiKey?: string
+  baseUrl?: string
+  publishTarget?: PublishTarget
+}
+
+export interface RunCliArgs
+  extends Omit<SuiteCliArgs, 'command' | 'publishTarget'> {
+  command: 'run'
+}
+
+export interface GradeCliArgs {
+  command: 'grade'
+  runDir: string
+}
+
+export interface PublishCliArgs {
+  command: 'publish'
+  runDir: string
+  target: PublishTarget
+}
+
+export type EvalCliArgs =
+  | LegacyCliArgs
+  | SuiteCliArgs
+  | RunCliArgs
+  | GradeCliArgs
+  | PublishCliArgs
+
+const COMMANDS = new Set(['suite', 'run', 'grade', 'publish'])
+
+function stringValue(value: string | boolean | undefined): string | undefined {
+  return typeof value === 'string' && value.length > 0 ? value : undefined
+}
+
+function publishTarget(value: string | undefined): PublishTarget | undefined {
+  if (value === undefined) return undefined
+  if (value === 'r2') return 'r2'
+  throw new Error(`Unsupported publish target: ${value}`)
+}
+
+function requireOne(
+  command: string,
+  configPath: string | undefined,
+  suitePath: string | undefined,
+): void {
+  if (!configPath && !suitePath) {
+    throw new Error(`${command} requires --config or --suite`)
+  }
+  if (configPath && suitePath) {
+    throw new Error(`${command} accepts either --config or --suite, not both`)
+  }
+}
+
+function parseSuiteLikeArgs(
+  command: 'suite' | 'run',
+  argv: string[],
+): SuiteCliArgs | RunCliArgs {
+  const { values } = parseArgs({
+    args: argv,
+    options: {
+      config: { type: 'string' },
+      suite: { type: 'string' },
+      variant: { type: 'string' },
+      provider: { type: 'string' },
+      model: { type: 'string' },
+      'api-key': { type: 'string' },
+      'base-url': { type: 'string' },
+      publish: { type: 'string' },
+    },
+  })
+
+  const configPath = stringValue(values.config)
+  const suitePath = stringValue(values.suite)
+  requireOne(command, configPath, suitePath)
+
+  const parsed: SuiteCliArgs | RunCliArgs =
+    command === 'suite' ? { command: 'suite' } : { command: 'run' }
+  if (configPath) parsed.configPath = configPath
+  if (suitePath) parsed.suitePath = suitePath
+  const variantId = stringValue(values.variant)
+  if (variantId) parsed.variantId = variantId
+  const provider = stringValue(values.provider)
+  if (provider) parsed.provider = provider
+  const model = stringValue(values.model)
+  if (model) parsed.model = model
+  const apiKey = stringValue(values['api-key'])
+  if (apiKey) parsed.apiKey = apiKey
+  const baseUrl = stringValue(values['base-url'])
+  if (baseUrl) parsed.baseUrl = baseUrl
+
+  if (command === 'suite') {
+    const target = publishTarget(stringValue(values.publish))
+    if (target) {
+      const suiteArgs = parsed as SuiteCliArgs
+      suiteArgs.publishTarget = target
+    }
+  }
+
+  return parsed
+}
+
+function parseLegacyArgs(argv: string[]): LegacyCliArgs {
+  const { values } = parseArgs({
+    args: argv,
+    options: {
+      config: { type: 'string', short: 'c' },
+      help: { type: 'boolean', short: 'h' },
+    },
+  })
+
+  const parsed: LegacyCliArgs = { command: 'legacy' }
+  const configPath = stringValue(values.config)
+  if (configPath) parsed.configPath = configPath
+  if (values.help) parsed.help = true
+  return parsed
+}
+
+/** Parses the eval CLI command without running browser or publishing side effects. */
+export function parseEvalCliArgs(argv: string[]): EvalCliArgs {
+  const [command, ...rest] = argv
+  if (!COMMANDS.has(command ?? '')) {
+    return parseLegacyArgs(argv)
+  }
+
+  switch (command) {
+    case 'suite':
+      return parseSuiteLikeArgs('suite', rest)
+    case 'run':
+      return parseSuiteLikeArgs('run', rest)
+    case 'grade': {
+      const { values } = parseArgs({
+        args: rest,
+        options: { run: { type: 'string' } },
+      })
+      const runDir = stringValue(values.run)
+      if (!runDir) throw new Error('grade requires --run')
+      return { command: 'grade', runDir }
+    }
+    case 'publish': {
+      const { values } = parseArgs({
+        args: rest,
+        options: { run: { type: 'string' }, target: { type: 'string' } },
+      })
+      const runDir = stringValue(values.run)
+      if (!runDir) throw new Error('publish requires --run')
+      const target = publishTarget(stringValue(values.target))
+      if (!target) throw new Error('publish requires --target')
+      return { command: 'publish', runDir, target }
+    }
+    default:
+      return parseLegacyArgs(argv)
+  }
+}

packages/browseros-agent/apps/eval/src/cli/commands/grade.ts

@@ -0,0 +1,84 @@
+import { readdir, readFile, stat } from 'node:fs/promises'
+import { join } from 'node:path'
+import { TrajectorySaver } from '../../capture/trajectory-saver'
+import { runGraders } from '../../grading/grader-runner'
+import { type Message, MessageSchema, TaskMetadataSchema } from '../../types'
+import type { GradeCliArgs } from '../args'
+
+async function loadMessages(taskDir: string): Promise<Message[]> {
+  const content = await readFile(
+    join(taskDir, 'messages.jsonl'),
+    'utf-8',
+  ).catch(() => '')
+  return content
+    .split('\n')
+    .filter((line) => line.trim().length > 0)
+    .map((line) => MessageSchema.parse(JSON.parse(line)))
+}
+
+async function findTaskDirs(runDir: string): Promise<string[]> {
+  const entries = await readdir(runDir, { withFileTypes: true })
+  const taskDirs: string[] = []
+  for (const entry of entries) {
+    if (!entry.isDirectory()) continue
+    const taskDir = join(runDir, entry.name)
+    const metadata = await stat(join(taskDir, 'metadata.json')).catch(
+      () => null,
+    )
+    if (metadata?.isFile()) taskDirs.push(taskDir)
+  }
+  return taskDirs
+}
+
+/** Re-runs graders for task artifacts that already contain metadata and messages. */
+export async function runGradeCommand(args: GradeCliArgs): Promise<void> {
+  const runStat = await stat(args.runDir).catch(() => null)
+  if (!runStat?.isDirectory()) {
+    throw new Error(`Not a run directory: ${args.runDir}`)
+  }
+
+  const taskDirs = await findTaskDirs(args.runDir)
+  if (taskDirs.length === 0) {
+    throw new Error(`No task metadata found under ${args.runDir}`)
+  }
+
+  let graded = 0
+  for (const taskDir of taskDirs) {
+    const metadata = TaskMetadataSchema.parse(
+      JSON.parse(await readFile(join(taskDir, 'metadata.json'), 'utf-8')),
+    )
+    const graderNames = Object.keys(metadata.grader_results ?? {})
+    if (graderNames.length === 0) {
+      console.warn(`Skipping ${metadata.query_id}: no existing grader names`)
+      continue
+    }
+
+    const messages = await loadMessages(taskDir)
+    const graderResults = await runGraders(graderNames, {
+      task: {
+        query_id: metadata.query_id,
+        query: metadata.query,
+        dataset: metadata.dataset,
+      },
+      messages,
+      screenshotCount: metadata.screenshot_count ?? metadata.total_steps,
+      finalAnswer: metadata.final_answer,
+      taskArtifactDir: taskDir,
+      outputDir: taskDir,
+      mcpUrl: `${process.env.BROWSEROS_SERVER_URL || 'http://127.0.0.1:9110'}/mcp`,
+    })
+
+    await new TrajectorySaver(
+      args.runDir,
+      metadata.query_id,
+    ).updateGraderResults(graderResults)
+    graded++
+  }
+
+  if (graded === 0) {
+    throw new Error(
+      `No tasks with existing grader names found under ${args.runDir}`,
+    )
+  }
+  console.log(`Re-graded ${graded} task(s) in ${args.runDir}`)
+}

packages/browseros-agent/apps/eval/src/cli/commands/publish.ts

@@ -0,0 +1,25 @@
+import { publishPathToR2 } from '../../publishing/r2-publisher'
+import type { PublishCliArgs, PublishTarget } from '../args'
+
+export interface PublishRunOptions {
+  runDir: string
+  target: PublishTarget
+}
+
+/** Publishes run artifacts through the R2 viewer upload path. */
+export async function publishRun(options: PublishRunOptions): Promise<void> {
+  if (options.target !== 'r2') {
+    throw new Error(`Unsupported publish target: ${options.target}`)
+  }
+  const result = await publishPathToR2(options.runDir)
+  for (const run of result.uploadedRuns) {
+    console.log(run.viewerUrl)
+  }
+  for (const runId of result.skippedRuns) {
+    console.log(`${runId}: already uploaded, skipping`)
+  }
+}
+
+export async function runPublishCommand(args: PublishCliArgs): Promise<void> {
+  await publishRun({ runDir: args.runDir, target: args.target })
+}

packages/browseros-agent/apps/eval/src/cli/commands/run.ts

@@ -0,0 +1,21 @@
+import type { RunCliArgs } from '../args'
+import { runSuiteCommand, type SuiteCommandDeps } from './suite'
+
+/** Executes tasks from a config or suite without publishing artifacts. */
+export async function runRunCommand(
+  args: RunCliArgs,
+  deps: SuiteCommandDeps = {},
+): Promise<void> {
+  await runSuiteCommand(
+    {
+      configPath: args.configPath,
+      suitePath: args.suitePath,
+      variantId: args.variantId,
+      provider: args.provider,
+      model: args.model,
+      apiKey: args.apiKey,
+      baseUrl: args.baseUrl,
+    },
+    deps,
+  )
+}

packages/browseros-agent/apps/eval/src/cli/commands/suite.ts

@@ -0,0 +1,187 @@
+import type { RunEvalOptions, RunEvalResult } from '../../runner/types'
+import { runEval as defaultRunEval } from '../../runs/eval-runner'
+import {
+  type AdaptedEvalConfig,
+  adaptEvalConfigFile,
+} from '../../suites/config-adapter'
+import { loadSuite } from '../../suites/load-suite'
+import { type EvalVariant, resolveVariant } from '../../suites/resolve-variant'
+import type { EvalSuite } from '../../suites/schema'
+import { type EvalConfig, EvalConfigSchema } from '../../types'
+import type { PublishTarget } from '../args'
+
+type Env = Record<string, string | undefined>
+
+export interface SuiteCommandOptions {
+  configPath?: string
+  suitePath?: string
+  variantId?: string
+  provider?: string
+  model?: string
+  apiKey?: string
+  baseUrl?: string
+  publishTarget?: PublishTarget
+  env?: Env
+}
+
+export type ResolvedSuiteCommand =
+  | (AdaptedEvalConfig & { kind: 'config'; datasetPath?: undefined })
+  | {
+      kind: 'suite'
+      suitePath: string
+      suite: EvalSuite
+      variant: EvalVariant
+      datasetPath: string
+      evalConfig: EvalConfig
+    }
+
+export interface SuiteCommandDeps {
+  runEval?: (options: RunEvalOptions) => Promise<RunEvalResult | undefined>
+  publishRun?: (options: {
+    runDir: string
+    target: PublishTarget
+  }) => Promise<void>
+}
+
+function ensureRunnableSuite(suite: EvalSuite): void {
+  if (!suite.browseros) {
+    throw new Error('suite browseros config is required to run suite commands')
+  }
+}
+
+function suiteToEvalConfig(
+  suite: EvalSuite,
+  datasetPath: string,
+  variant: EvalVariant,
+  env: Env,
+): EvalConfig {
+  ensureRunnableSuite(suite)
+
+  const base = {
+    dataset: datasetPath,
+    num_workers: suite.workers,
+    restart_server_per_task: suite.restartBrowserPerTask,
+    browseros: suite.browseros,
+    graders: suite.graders,
+    timeout_ms: suite.timeoutMs,
+    captcha: suite.captcha,
+  }
+
+  if (suite.agent.type === 'single' || suite.agent.type === 'tool-loop') {
+    // The legacy runner names the BrowserOS tool-loop agent "single".
+    return EvalConfigSchema.parse({
+      ...base,
+      agent: {
+        type: 'single',
+        provider: variant.agent.provider,
+        model: variant.agent.model,
+        apiKey: variant.agent.apiKey,
+        baseUrl: variant.agent.baseUrl,
+        supportsImages: variant.agent.supportsImages,
+      },
+    })
+  }
+
+  const executorBackend = suite.agent.executorBackend ?? 'tool-loop'
+  const executor =
+    executorBackend === 'clado'
+      ? {
+          provider: 'clado-action' as const,
+          model:
+            env.EVAL_EXECUTOR_MODEL ?? env.CLADO_ACTION_MODEL ?? 'clado-action',
+          apiKey: env.EVAL_EXECUTOR_API_KEY ?? env.CLADO_ACTION_API_KEY ?? '',
+          baseUrl:
+            env.EVAL_EXECUTOR_BASE_URL ??
+            env.CLADO_ACTION_BASE_URL ??
+            env.CLADO_ACTION_URL,
+        }
+      : {
+          provider: variant.agent.provider,
+          model: variant.agent.model,
+          apiKey: variant.agent.apiKey,
+          baseUrl: variant.agent.baseUrl,
+        }
+
+  return EvalConfigSchema.parse({
+    ...base,
+    agent: {
+      type: 'orchestrator-executor',
+      orchestrator: {
+        provider: variant.agent.provider,
+        model: variant.agent.model,
+        apiKey: variant.agent.apiKey,
+        baseUrl: variant.agent.baseUrl,
+      },
+      executor,
+    },
+  })
+}
+
+/** Resolves config-backed or suite-backed CLI input into the run shape used by the runner. */
+export async function resolveSuiteCommand(
+  options: SuiteCommandOptions,
+): Promise<ResolvedSuiteCommand> {
+  const env = options.env ?? process.env
+  if (options.configPath) {
+    return {
+      kind: 'config',
+      ...(await adaptEvalConfigFile(options.configPath, { env })),
+    }
+  }
+  if (!options.suitePath) {
+    throw new Error('suite requires --config or --suite')
+  }
+
+  const loaded = await loadSuite(options.suitePath)
+  const variant = resolveVariant({
+    variantId: options.variantId,
+    provider: options.provider,
+    model: options.model,
+    apiKey: options.apiKey,
+    baseUrl: options.baseUrl,
+    env,
+  })
+
+  return {
+    kind: 'suite',
+    suitePath: loaded.suitePath,
+    suite: loaded.suite,
+    variant,
+    datasetPath: loaded.datasetPath,
+    evalConfig: suiteToEvalConfig(
+      loaded.suite,
+      loaded.datasetPath,
+      variant,
+      env,
+    ),
+  }
+}
+
+/** Runs the full suite loop: resolve input, execute tasks, then optionally publish the run. */
+export async function runSuiteCommand(
+  options: SuiteCommandOptions,
+  deps: SuiteCommandDeps = {},
+): Promise<void> {
+  const runEval = deps.runEval ?? defaultRunEval
+  const resolved = await resolveSuiteCommand(options)
+  const runOptions: RunEvalOptions =
+    resolved.kind === 'config'
+      ? { configPath: resolved.configPath }
+      : {
+          configPath: resolved.suitePath,
+          dataPath: resolved.datasetPath,
+          config: resolved.evalConfig,
+        }
+
+  const result = await runEval(runOptions)
+  if (!options.publishTarget) return
+
+  const outputDir = result?.outputDir
+  if (!outputDir) {
+    throw new Error('publish requested but runner did not return an outputDir')
+  }
+  if (!deps.publishRun) {
+    throw new Error('publish requested before the publisher is configured')
+  }
+  await deps.publishRun({ runDir: outputDir, target: options.publishTarget })
+}

packages/browseros-agent/apps/eval/src/cli/index.ts

@@ -0,0 +1,70 @@
+import { startDashboard } from '../dashboard/server'
+import { runEval } from '../runs/eval-runner'
+import { type EvalCliArgs, parseEvalCliArgs } from './args'
+import { runGradeCommand } from './commands/grade'
+import { publishRun, runPublishCommand } from './commands/publish'
+import { runRunCommand } from './commands/run'
+import { runSuiteCommand } from './commands/suite'
+
+export function usage(): string {
+  return `
+BrowserOS Eval
+
+Usage:
+  bun run eval suite --config <config.json> [--publish r2]
+  bun run eval suite --suite <suite.json> --variant <id> [--publish r2]
+  bun run eval run --config <config.json>
+  bun run eval run --suite <suite.json> --variant <id>
+  bun run eval grade --run <results/run-dir>
+  bun run eval publish --run <results/run-dir> --target r2
+  bun run eval -c <config.json>
+`
+}
+
+async function runLegacyCommand(args: EvalCliArgs): Promise<void> {
+  if (args.command !== 'legacy') return
+  if (args.help) {
+    console.log(usage())
+    return
+  }
+  if (args.configPath) {
+    await runEval({ configPath: args.configPath })
+    return
+  }
+
+  startDashboard({
+    tasks: [],
+    configName: '',
+    agentType: '',
+    outputDir: '',
+    configMode: true,
+  })
+  console.log(
+    'Dashboard running at http://localhost:9900 — configure and run from the UI',
+  )
+  await new Promise(() => {})
+}
+
+/** Dispatches the eval CLI while preserving the old config/dashboard entry points. */
+export async function runCli(
+  argv: string[] = Bun.argv.slice(2),
+): Promise<void> {
+  const args = parseEvalCliArgs(argv)
+  switch (args.command) {
+    case 'legacy':
+      await runLegacyCommand(args)
+      break
+    case 'suite':
+      await runSuiteCommand(args, { publishRun })
+      break
+    case 'run':
+      await runRunCommand(args)
+      break
+    case 'grade':
+      await runGradeCommand(args)
+      break
+    case 'publish':
+      await runPublishCommand(args)
+      break
+  }
+}

packages/browseros-agent/apps/eval/src/constants.ts

@@ -5,4 +5,5 @@
 export const DEFAULT_TIMEOUT_MS = 30 * 60 * 1000 // 30 minutes
 export const SCREENSHOT_TIMEOUT_MS = 65_000 // 65s — ensures we get extension's error (60s)
 export const MAX_ACTIONS_PER_DELEGATION = 15
-export const CLADO_REQUEST_TIMEOUT_MS = 120_000
+// Cold start can take ~5 minutes per Clado; 6 minutes leaves headroom.
+export const CLADO_REQUEST_TIMEOUT_MS = 360_000

packages/browseros-agent/apps/eval/src/dashboard/server.ts

@@ -1,5 +1,5 @@
 import { mkdir, readdir, readFile, stat } from 'node:fs/promises'
-import { join, resolve } from 'node:path'
+import { dirname, join, resolve, sep } from 'node:path'
 import { Hono } from 'hono'
 import { streamSSE } from 'hono/streaming'
 import { ParallelExecutor } from '../runner/parallel-executor'
@@ -128,6 +128,35 @@ let dashboardConfigMode = false
 const configsDir = join(import.meta.dir, '..', '..', 'configs')
 const projectRoot = resolve(import.meta.dir, '..', '..', '..', '..')
 
+async function listConfigFiles(dir: string, prefix = ''): Promise<string[]> {
+  const entries = await readdir(join(dir, prefix), { withFileTypes: true })
+  const files: string[] = []
+  for (const entry of entries) {
+    const relativePath = prefix ? join(prefix, entry.name) : entry.name
+    if (entry.isDirectory()) {
+      files.push(...(await listConfigFiles(dir, relativePath)))
+    } else if (entry.isFile() && entry.name.endsWith('.json')) {
+      files.push(relativePath.split(sep).join('/'))
+    }
+  }
+  return files.sort()
+}
+
+function resolveConfigPath(name: string): string | null {
+  if (!name.endsWith('.json')) return null
+  if (name.split('/').some((part) => !part || part === '.' || part === '..')) {
+    return null
+  }
+
+  const resolvedPath = resolve(configsDir, name)
+  const resolvedConfigsDir = resolve(configsDir)
+  const configRootPrefix = resolvedConfigsDir.endsWith(sep)
+    ? resolvedConfigsDir
+    : `${resolvedConfigsDir}${sep}`
+  if (!resolvedPath.startsWith(configRootPrefix)) return null
+  return resolvedPath
+}
+
 // ============================================================================
 // Hono App
 // ============================================================================
@@ -339,21 +368,21 @@ app.get('/api/mode', (c) => {
 // List saved config files
 app.get('/api/configs', async (c) => {
   try {
-    const files = await readdir(configsDir)
-    return c.json(files.filter((f) => f.endsWith('.json')))
+    return c.json(await listConfigFiles(configsDir))
   } catch {
     return c.json([])
   }
 })
 
 // Read a specific config file
-app.get('/api/config/:name', async (c) => {
-  const name = c.req.param('name')
-  if (name.includes('/') || name.includes('..')) {
+app.get('/api/config/*', async (c) => {
+  const name = decodeURIComponent(c.req.path.slice('/api/config/'.length))
+  const configPath = resolveConfigPath(name)
+  if (!configPath) {
     return c.json({ error: 'Invalid config name' }, 400)
   }
   try {
-    const content = await readFile(join(configsDir, name), 'utf-8')
+    const content = await readFile(configPath, 'utf-8')
     return c.json(JSON.parse(content))
   } catch {
     return c.notFound()
@@ -382,8 +411,17 @@ app.post('/api/run', async (c) => {
 
   const config = parseResult.data
 
-  // Resolve relative paths from configs/ dir (dataset dropdown values are relative to it)
-  const baseDir = configsDir
+  let baseDir = configsDir
+  if (body.configName) {
+    const configPath = resolveConfigPath(body.configName)
+    if (!configPath) {
+      return c.json({ error: 'Invalid config name' }, 400)
+    }
+    baseDir = dirname(configPath)
+  }
+
+  // Resolve relative paths from the loaded config location. Unsaved dashboard
+  // configs keep using apps/eval/configs as their base for dropdown values.
   const datasetPath = resolve(
     config.dataset.startsWith('/')
       ? config.dataset

packages/browseros-agent/apps/eval/src/dashboard/viewer.html

@@ -685,6 +685,59 @@
     });
   }
 
+  // Test harness note: these ASCII section markers are used by r2-viewer-compat.test.ts.
+  // -- Artifact path resolution
+  function taskKey(task) {
+    return task.queryId || task.id || 'unknown-task';
+  }
+
+  function legacyArtifactPath(task, artifact) {
+    const id = taskKey(task);
+    switch (artifact) {
+      case 'attempt':
+        return `${id}/attempt.json`;
+      case 'metadata':
+        return `${id}/metadata.json`;
+      case 'messages':
+        return `${id}/messages.jsonl`;
+      case 'trace':
+        return `${id}/trace.jsonl`;
+      case 'grades':
+        return `${id}/grades.json`;
+      case 'screenshots':
+        return `${id}/screenshots`;
+      case 'graderArtifacts':
+        return `${id}/grader-artifacts`;
+      default:
+        return `${id}/${artifact}`;
+    }
+  }
+
+  function artifactPath(task, artifact) {
+    const manifestPath = task.paths && task.paths[artifact];
+    if (typeof manifestPath === 'string' && manifestPath.length > 0) {
+      return manifestPath.replace(/^\/+/, '');
+    }
+    return legacyArtifactPath(task, artifact);
+  }
+
+  function artifactUrl(task, artifact) {
+    return `${basePath}/${artifactPath(task, artifact)}`;
+  }
+
+  function metadataUrl(task) {
+    return artifactUrl(task, 'metadata');
+  }
+
+  function messagesUrl(task) {
+    return artifactUrl(task, 'messages');
+  }
+
+  function screenshotUrl(task, n) {
+    return `${artifactUrl(task, 'screenshots')}/${n}.png`;
+  }
+
+  // -- Task selection
   // ── Task selection ─────────────────────────────────────────────
   function selectTask(task) {
     stopAutoplay();
@@ -716,6 +769,7 @@
     }
   }
 
+  // -- Center panel
   // ── Center panel: screenshot viewer ────────────────────────────
   function renderCenterPanel(task) {
     const panel = document.getElementById('center-panel');
@@ -763,10 +817,6 @@
     updateControls();
   }
 
-  function screenshotUrl(task, n) {
-    return `${basePath}/${task.queryId || task.id}/screenshots/${n}.png`;
-  }
-
   function goToStep(n) {
     if (!selectedTask || n < 1 || n > totalSteps) return;
     currentStep = n;
@@ -914,7 +964,7 @@
     body.innerHTML = '<div class="placeholder"><div class="ph-text" style="color: #6e7681;">Loading messages...</div></div>';
     countEl.textContent = '';
 
-    const msgUrl = `${basePath}/${task.queryId || task.id}/messages.jsonl`;
+    const msgUrl = messagesUrl(task);
 
     fetch(msgUrl)
       .then((res) => {
@@ -1075,7 +1125,7 @@
 
   // ── Load task metadata for rich grader details ──────────────────
   function loadTaskMetadata(task) {
-    const metaUrl = `${basePath}/${task.queryId || task.id}/metadata.json`;
+    const metaUrl = metadataUrl(task);
     fetch(metaUrl)
       .then((res) => res.ok ? res.json() : null)
       .then((meta) => {

packages/browseros-agent/apps/eval/src/graders/benchmark/agisdk-state-diff.ts

@@ -1,5 +1,12 @@
-import { spawn } from 'node:child_process'
 import { join } from 'node:path'
+import {
+  writeGraderJsonArtifact,
+  writeGraderTextArtifact,
+} from '../../grading/artifacts'
+import {
+  type PythonEvaluatorResult,
+  runPythonJsonEvaluator,
+} from '../../grading/python-evaluator'
 import type { GraderResult } from '../../types'
 import { callMcpTool } from '../../utils/mcp-client'
 import type { Grader, GraderInput } from '../types'
@@ -7,12 +14,23 @@ import type { Grader, GraderInput } from '../types'
 const EVAL_SCRIPT = join(
   import.meta.dirname,
   '..',
-  '..',
-  '..',
-  'scripts',
+  'python',
   'agisdk-evaluate.py',
 )
 
+interface AgisdkEvaluatorInput {
+  task_id: string
+  env_state: Record<string, unknown>
+  model_response: string
+}
+
+interface AgisdkEvaluatorOutput {
+  reward: number
+  pass: boolean
+  message: string
+  per_criterion: unknown[]
+}
+
 export class AgisdkStateDiffGrader implements Grader {
   name = 'agisdk_state_diff'
 
@@ -36,6 +54,16 @@ export class AgisdkStateDiffGrader implements Grader {
     let envState: Record<string, unknown>
     try {
       envState = await this.fetchFinishState(origin, mcpEndpoint)
+      await writeGraderJsonArtifact(
+        input,
+        this.name,
+        'finish-state.json',
+        envState,
+      )
+      await writeGraderJsonArtifact(input, this.name, 'context.json', {
+        origin,
+        agisdk_task_id: taskId,
+      })
     } catch (error) {
       return {
         score: 0,
@@ -46,10 +74,30 @@ export class AgisdkStateDiffGrader implements Grader {
     }
 
     try {
-      const result = await this.runPythonEvaluator(
-        taskId,
-        envState,
-        input.finalAnswer || '',
+      const evaluatorInput: AgisdkEvaluatorInput = {
+        task_id: taskId,
+        env_state: envState,
+        model_response: input.finalAnswer || '',
+      }
+      await writeGraderJsonArtifact(
+        input,
+        this.name,
+        'evaluator-input.json',
+        evaluatorInput,
+      )
+      const evaluation = await this.runPythonEvaluator(evaluatorInput)
+      const result = evaluation.output
+      await writeGraderJsonArtifact(
+        input,
+        this.name,
+        'evaluator-output.json',
+        result,
+      )
+      await writeGraderTextArtifact(
+        input,
+        this.name,
+        'stderr.txt',
+        evaluation.stderr,
       )
       return {
         score: result.reward,
@@ -144,59 +192,12 @@ export class AgisdkStateDiffGrader implements Grader {
   }
 
   private runPythonEvaluator(
-    taskId: string,
-    envState: Record<string, unknown>,
-    modelResponse: string,
-  ): Promise<{
-    reward: number
-    pass: boolean
-    message: string
-    per_criterion: unknown[]
-  }> {
-    return new Promise((resolve, reject) => {
-      const proc = spawn('python3', [EVAL_SCRIPT], {
-        stdio: ['pipe', 'pipe', 'pipe'],
-      })
-
-      const inputData = JSON.stringify({
-        task_id: taskId,
-        env_state: envState,
-        model_response: modelResponse,
-      })
-
-      let stdout = ''
-      let stderr = ''
-
-      proc.stdout.on('data', (data: Buffer) => {
-        stdout += data.toString()
-      })
-
-      proc.stderr.on('data', (data: Buffer) => {
-        stderr += data.toString()
-      })
-
-      proc.on('close', (code) => {
-        if (code !== 0) {
-          reject(
-            new Error(`Python evaluator exited with code ${code}: ${stderr}`),
-          )
-          return
-        }
-
-        try {
-          const result = JSON.parse(stdout.trim())
-          resolve(result)
-        } catch {
-          reject(new Error(`Failed to parse evaluator output: ${stdout}`))
-        }
-      })
-
-      proc.on('error', (err) => {
-        reject(new Error(`Failed to spawn Python evaluator: ${err.message}`))
-      })
-
-      proc.stdin.write(inputData)
-      proc.stdin.end()
+    evalInput: AgisdkEvaluatorInput,
+  ): Promise<PythonEvaluatorResult<AgisdkEvaluatorOutput>> {
+    return runPythonJsonEvaluator<AgisdkEvaluatorOutput>({
+      scriptPath: EVAL_SCRIPT,
+      input: evalInput,
+      timeoutMs: 300_000,
     })
   }
 }

packages/browseros-agent/apps/eval/src/graders/benchmark/infinity-state.ts

@@ -1,4 +1,12 @@
 import { join, resolve } from 'node:path'
+import {
+  writeGraderJsonArtifact,
+  writeGraderTextArtifact,
+} from '../../grading/artifacts'
+import {
+  type PythonEvaluatorResult,
+  runPythonJsonEvaluator,
+} from '../../grading/python-evaluator'
 import type { GraderResult } from '../../types'
 import type { Grader, GraderInput } from '../types'
 
@@ -14,10 +22,7 @@ interface InfinityEvalOutput {
   message: string
 }
 
-const EVAL_SCRIPT = resolve(
-  import.meta.dir,
-  '../../../scripts/infinity-evaluate.py',
-)
+const EVAL_SCRIPT = resolve(import.meta.dir, '../python/infinity-evaluate.py')
 
 export class InfinityStateGrader implements Grader {
   name = 'infinity_state'
@@ -66,7 +71,32 @@ export class InfinityStateGrader implements Grader {
     }
 
     try {
-      const result = await this.runPythonEvaluator(evalInput)
+      await writeGraderJsonArtifact(input, this.name, 'verifier.json', {
+        appName: parsed.appName,
+        taskId: parsed.taskId,
+        verifierPath,
+        appServerUrl,
+      })
+      await writeGraderJsonArtifact(
+        input,
+        this.name,
+        'evaluator-input.json',
+        evalInput,
+      )
+      const evaluation = await this.runPythonEvaluator(evalInput)
+      const result = evaluation.output
+      await writeGraderJsonArtifact(
+        input,
+        this.name,
+        'evaluator-output.json',
+        result,
+      )
+      await writeGraderTextArtifact(
+        input,
+        this.name,
+        'stderr.txt',
+        evaluation.stderr,
+      )
       return {
         score: result.pass ? 1 : 0,
         pass: result.pass,
@@ -108,27 +138,11 @@ export class InfinityStateGrader implements Grader {
 
   private async runPythonEvaluator(
     evalInput: InfinityEvalInput,
-  ): Promise<InfinityEvalOutput> {
-    const proc = Bun.spawn(['python3', EVAL_SCRIPT], {
-      stdin: 'pipe',
-      stdout: 'pipe',
-      stderr: 'pipe',
+  ): Promise<PythonEvaluatorResult<InfinityEvalOutput>> {
+    return runPythonJsonEvaluator<InfinityEvalOutput>({
+      scriptPath: EVAL_SCRIPT,
+      input: evalInput,
+      timeoutMs: 300_000,
     })
-
-    const inputJson = JSON.stringify(evalInput)
-    proc.stdin.write(inputJson)
-    proc.stdin.end()
-
-    const stdout = await new Response(proc.stdout).text()
-    const stderr = await new Response(proc.stderr).text()
-    const exitCode = await proc.exited
-
-    if (exitCode !== 0) {
-      throw new Error(
-        `Python evaluator exited with code ${exitCode}: ${stderr || stdout}`,
-      )
-    }
-
-    return JSON.parse(stdout.trim()) as InfinityEvalOutput
   }
 }

packages/browseros-agent/apps/eval/src/graders/performance/performance-grader.ts

@@ -1,6 +1,7 @@
 import { readFile } from 'node:fs/promises'
 import { join } from 'node:path'
 import { query } from '@anthropic-ai/claude-agent-sdk'
+import { writeGraderJsonArtifact } from '../../grading/artifacts'
 import type { GraderResult } from '../../types'
 import type { Grader, GraderInput } from '../types'
 import {
@@ -63,6 +64,7 @@ export class PerformanceGrader implements Grader {
         input.screenshotCount,
         terminationReason,
       )
+      await writeGraderJsonArtifact(input, this.name, 'metrics.json', metrics)
 
       const systemPrompt = PERFORMANCE_SYSTEM_PROMPT.replace(
         /\{screenshot_count\}/g,
@@ -82,6 +84,14 @@ export class PerformanceGrader implements Grader {
         userPrompt,
         input.outputDir,
       )
+      if (response) {
+        await writeGraderJsonArtifact(
+          input,
+          this.name,
+          'agent-output.json',
+          response,
+        )
+      }
 
       if (!response) {
         return {
@@ -140,6 +150,7 @@ export class PerformanceGrader implements Grader {
           `Perf grader: LLM returned ${returnedAxes.size}/${expectedAxes.size} axes, missing: ${missingAxes.join(', ')}`,
         )
       }
+      await writeGraderJsonArtifact(input, this.name, 'axes.json', axisResults)
 
       return {
         score: compositeScore / 100,

packages/browseros-agent/apps/eval/src/graders/registry.ts

@@ -1,51 +1,2 @@
-import type { GraderResult } from '../types'
-import { AgisdkStateDiffGrader } from './benchmark/agisdk-state-diff'
-import { InfinityStateGrader } from './benchmark/infinity-state'
-import { PerformanceGrader } from './performance/performance-grader'
-import type { Grader, GraderInput } from './types'
-
-export const PASS_FAIL_GRADER_ORDER = [
-  'agisdk_state_diff',
-  'infinity_state',
-  'performance_grader',
-] as const
-
-export function createGrader(name: string): Grader | null {
-  switch (name) {
-    case 'agisdk_state_diff':
-      return new AgisdkStateDiffGrader()
-    case 'infinity_state':
-      return new InfinityStateGrader()
-    case 'performance_grader':
-      return new PerformanceGrader()
-    default:
-      console.warn(`Unknown grader: ${name}`)
-      return null
-  }
-}
-
-export async function runGraders(
-  graderNames: string[],
-  input: GraderInput,
-): Promise<Record<string, GraderResult>> {
-  const results: Record<string, GraderResult> = {}
-
-  for (const name of graderNames) {
-    const grader = createGrader(name)
-    if (!grader) continue
-    try {
-      console.log(`  Running grader: ${name}`)
-      results[name] = await grader.grade(input)
-    } catch (error) {
-      results[name] = {
-        score: 0,
-        pass: false,
-        reasoning: `Error running grader: ${error}`,
-      }
-    }
-  }
-
-  return results
-}
-
-export { AgisdkStateDiffGrader, InfinityStateGrader, PerformanceGrader }
+export * from '../grading/grader-registry'
+export { runConfiguredGraders, runGraders } from '../grading/grader-runner'

packages/browseros-agent/apps/eval/src/graders/types.ts

@@ -1,21 +1 @@
-import type { GraderResult, Message } from '../types'
-
-export interface GraderInput {
-  task: {
-    query_id: string
-    query: string
-    dataset: string
-  }
-  messages: Message[]
-  screenshotCount: number
-  finalAnswer: string | null
-  expectedAnswer?: string | null
-  outputDir: string
-  mcpUrl?: string
-  infinityAppUrl?: string
-}
-
-export interface Grader {
-  name: string
-  grade(input: GraderInput): Promise<GraderResult>
-}
+export type { Grader, GraderInput } from '../grading/types'

packages/browseros-agent/apps/eval/src/grading/artifacts.ts

@@ -0,0 +1,34 @@
+import { mkdir, writeFile } from 'node:fs/promises'
+import { join } from 'node:path'
+import type { GraderInput } from './types'
+
+function artifactDir(input: GraderInput, graderName: string): string {
+  return join(
+    input.taskArtifactDir || input.outputDir,
+    'grader-artifacts',
+    graderName,
+  )
+}
+
+/** Writes a JSON artifact for a grader under the task artifact directory. */
+export async function writeGraderJsonArtifact(
+  input: GraderInput,
+  graderName: string,
+  filename: string,
+  value: unknown,
+): Promise<void> {
+  const dir = artifactDir(input, graderName)
+  await mkdir(dir, { recursive: true })
+  await writeFile(join(dir, filename), JSON.stringify(value, null, 2))
+}
+
+export async function writeGraderTextArtifact(
+  input: GraderInput,
+  graderName: string,
+  filename: string,
+  value: string,
+): Promise<void> {
+  const dir = artifactDir(input, graderName)
+  await mkdir(dir, { recursive: true })
+  await writeFile(join(dir, filename), value)
+}

packages/browseros-agent/apps/eval/src/grading/grader-registry.ts

@@ -0,0 +1,26 @@
+import { AgisdkStateDiffGrader } from '../graders/benchmark/agisdk-state-diff'
+import { InfinityStateGrader } from '../graders/benchmark/infinity-state'
+import { PerformanceGrader } from '../graders/performance/performance-grader'
+import type { Grader } from './types'
+
+export const PASS_FAIL_GRADER_ORDER = [
+  'agisdk_state_diff',
+  'infinity_state',
+  'performance_grader',
+] as const
+
+export function createGrader(name: string): Grader | null {
+  switch (name) {
+    case 'agisdk_state_diff':
+      return new AgisdkStateDiffGrader()
+    case 'infinity_state':
+      return new InfinityStateGrader()
+    case 'performance_grader':
+      return new PerformanceGrader()
+    default:
+      console.warn(`Unknown grader: ${name}`)
+      return null
+  }
+}
+
+export { AgisdkStateDiffGrader, InfinityStateGrader, PerformanceGrader }

packages/browseros-agent/apps/eval/src/grading/grader-runner.ts

@@ -0,0 +1,36 @@
+import type { GraderResult } from '../types'
+import { createGrader as defaultCreateGrader } from './grader-registry'
+import type { Grader, GraderInput } from './types'
+
+export interface GraderRunnerDeps {
+  createGrader?: (name: string) => Grader | null
+}
+
+/** Runs configured graders independently so one failure does not hide others. */
+export async function runConfiguredGraders(
+  graderNames: string[],
+  input: GraderInput,
+  deps: GraderRunnerDeps = {},
+): Promise<Record<string, GraderResult>> {
+  const create = deps.createGrader ?? defaultCreateGrader
+  const results: Record<string, GraderResult> = {}
+
+  for (const name of graderNames) {
+    const grader = create(name)
+    if (!grader) continue
+    try {
+      console.log(`  Running grader: ${name}`)
+      results[name] = await grader.grade(input)
+    } catch (error) {
+      results[name] = {
+        score: 0,
+        pass: false,
+        reasoning: `Error running grader: ${error instanceof Error ? error.message : String(error)}`,
+      }
+    }
+  }
+
+  return results
+}
+
+export const runGraders = runConfiguredGraders

packages/browseros-agent/apps/eval/src/grading/python-evaluator.ts

@@ -0,0 +1,65 @@
+export interface PythonEvaluatorOptions {
+  scriptPath: string
+  input: unknown
+  timeoutMs: number
+}
+
+export interface PythonEvaluatorResult<T> {
+  output: T
+  stdout: string
+  stderr: string
+  exitCode: number
+}
+
+/** Runs a Python evaluator that accepts stdin JSON and emits stdout JSON. */
+export async function runPythonJsonEvaluator<T>(
+  options: PythonEvaluatorOptions,
+): Promise<PythonEvaluatorResult<T>> {
+  const proc = Bun.spawn(['python3', options.scriptPath], {
+    stdin: 'pipe',
+    stdout: 'pipe',
+    stderr: 'pipe',
+  })
+
+  proc.stdin.write(JSON.stringify(options.input))
+  proc.stdin.end()
+
+  let timeoutHandle: ReturnType<typeof setTimeout> | undefined
+  const timeout = new Promise<never>((_, reject) => {
+    timeoutHandle = setTimeout(() => {
+      proc.kill('SIGKILL')
+      reject(
+        new Error(`Python evaluator timed out after ${options.timeoutMs}ms`),
+      )
+    }, options.timeoutMs)
+  })
+
+  const completed = (async (): Promise<PythonEvaluatorResult<T>> => {
+    const stdout = await new Response(proc.stdout).text()
+    const stderr = await new Response(proc.stderr).text()
+    const exitCode = await proc.exited
+
+    if (exitCode !== 0) {
+      throw new Error(
+        `Python evaluator exited with code ${exitCode}: ${stderr || stdout}`,
+      )
+    }
+
+    try {
+      return {
+        output: JSON.parse(stdout.trim()) as T,
+        stdout,
+        stderr,
+        exitCode,
+      }
+    } catch {
+      throw new Error(`Failed to parse Python evaluator output: ${stdout}`)
+    }
+  })()
+
+  try {
+    return await Promise.race([completed, timeout])
+  } finally {
+    clearTimeout(timeoutHandle)
+  }
+}

packages/browseros-agent/apps/eval/src/grading/types.ts

@@ -0,0 +1,22 @@
+import type { GraderResult, Message } from '../types'
+
+export interface GraderInput {
+  task: {
+    query_id: string
+    query: string
+    dataset: string
+  }
+  messages: Message[]
+  screenshotCount: number
+  finalAnswer: string | null
+  expectedAnswer?: string | null
+  taskArtifactDir: string
+  outputDir: string
+  mcpUrl?: string
+  infinityAppUrl?: string
+}
+
+export interface Grader {
+  name: string
+  grade(input: GraderInput): Promise<GraderResult>
+}

packages/browseros-agent/apps/eval/src/index.ts

@@ -1,72 +1,10 @@
 #!/usr/bin/env bun
 
-import { parseArgs } from 'node:util'
-import { runEval } from './runner/eval-runner'
+import { runCli } from './cli'
 
-const { values } = parseArgs({
-  args: Bun.argv.slice(2),
-  options: {
-    config: { type: 'string', short: 'c' },
-    help: { type: 'boolean', short: 'h' },
-  },
-})
-
-if (values.help) {
-  console.log(`
-BrowserOS Eval
-
-Usage:
-  bun run eval                          # Opens dashboard in config mode
-  bun run eval --config <config.json>   # Runs eval with config file
-
-Available agent types:
-  - single                  Single LLM agent driven by the BrowserOS tool loop
-  - orchestrator-executor   High-level planner + visual/text executor
-
-Available graders:
-  - performance_grader      Multi-axis grader using Claude Agent SDK
-  - agisdk_state_diff       AGI SDK / REAL Bench state-diff grader
-  - infinity_state          WebArena-Infinity verifier-script grader
-
-Preset configs in configs/:
-  - browseros-agent-weekly.json       Weekly eval (single agent)
-  - browseros-oe-agent-weekly.json    Weekly eval (orchestrator + LLM executor)
-  - browseros-oe-clado-weekly.json    Weekly eval (orchestrator + Clado executor)
-  - agisdk-real-smoke.json            AGI SDK smoke run
-  - infinity-hard-50.json             WebArena-Infinity hard-50 set
-  - test-webvoyager.json              WebVoyager test
-  - test-mind2web.json                Mind2Web test
-
-Examples:
-  bun run eval                                       # Dashboard config mode
-  bun run eval -c configs/browseros-agent-weekly.json
-  bun run eval -c configs/test-webvoyager.json
-`)
-  process.exit(0)
-}
-
-if (values.config) {
-  try {
-    await runEval({ configPath: values.config })
-  } catch (error) {
-    console.error(error instanceof Error ? error.message : String(error))
-    process.exit(1)
-  }
-  process.exit(0)
-} else {
-  // No config — start dashboard in config mode, wait for user to configure and run
-  const { startDashboard } = await import('./dashboard/server')
-  startDashboard({
-    tasks: [],
-    configName: '',
-    agentType: '',
-    outputDir: '',
-    configMode: true,
-  })
-  console.log(
-    'Dashboard running at http://localhost:9900 — configure and run from the UI',
-  )
-
-  // Keep process alive until SIGINT
-  await new Promise(() => {})
+try {
+  await runCli(Bun.argv.slice(2))
+} catch (error) {
+  console.error(error instanceof Error ? error.message : String(error))
+  process.exit(1)
 }

packages/browseros-agent/apps/eval/src/publishing/r2-manifest.ts

@@ -0,0 +1,28 @@
+import type {
+  ViewerManifest,
+  ViewerManifestTask,
+} from '../viewer/viewer-manifest'
+
+export interface R2UploadConfig {
+  accountId: string
+  accessKeyId: string
+  secretAccessKey: string
+  bucket: string
+  cdnBaseUrl: string
+}
+
+export type R2ManifestTask = ViewerManifestTask
+
+export type R2RunManifest = ViewerManifest
+
+export interface R2PublishRunResult {
+  runId: string
+  uploadedFiles: number
+  viewerUrl: string
+  manifest: R2RunManifest
+}
+
+export interface R2PublishPathResult {
+  uploadedRuns: R2PublishRunResult[]
+  skippedRuns: string[]
+}

packages/browseros-agent/apps/eval/src/publishing/r2-publisher.ts

@@ -0,0 +1,427 @@
+import { readdir, readFile, stat } from 'node:fs/promises'
+import { basename, dirname, extname, join } from 'node:path'
+import {
+  GetObjectCommand,
+  PutObjectCommand,
+  S3Client,
+} from '@aws-sdk/client-s3'
+import {
+  buildViewerManifest,
+  type ViewerManifestTaskInput,
+} from '../viewer/viewer-manifest'
+import type {
+  R2PublishPathResult,
+  R2PublishRunResult,
+  R2RunManifest,
+  R2UploadConfig,
+} from './r2-manifest'
+
+const DEFAULT_CONCURRENCY = 20
+
+const CONTENT_TYPES: Record<string, string> = {
+  '.json': 'application/json',
+  '.jsonl': 'application/x-ndjson',
+  '.png': 'image/png',
+  '.html': 'text/html',
+}
+
+export interface R2Client {
+  send(command: unknown): Promise<unknown>
+}
+
+export interface R2PublisherOptions {
+  config: R2UploadConfig
+  client?: R2Client
+  viewerPath?: string
+  concurrency?: number
+  now?: () => Date
+}
+
+interface UploadJob {
+  key: string
+  filePath: string
+  contentType: string
+}
+
+interface TaskDirEntry {
+  taskId: string
+  taskPath: string
+}
+
+export function contentTypeForPath(filePath: string): string {
+  return CONTENT_TYPES[extname(filePath)] || 'application/octet-stream'
+}
+
+export function loadR2ConfigFromEnv(
+  env: Record<string, string | undefined> = process.env,
+): R2UploadConfig {
+  const accountId = env.EVAL_R2_ACCOUNT_ID
+  const accessKeyId = env.EVAL_R2_ACCESS_KEY_ID
+  const secretAccessKey = env.EVAL_R2_SECRET_ACCESS_KEY
+
+  if (!accountId || !accessKeyId || !secretAccessKey) {
+    throw new Error(
+      'Missing required env vars: EVAL_R2_ACCOUNT_ID, EVAL_R2_ACCESS_KEY_ID, EVAL_R2_SECRET_ACCESS_KEY',
+    )
+  }
+
+  return {
+    accountId,
+    accessKeyId,
+    secretAccessKey,
+    bucket: env.EVAL_R2_BUCKET || 'browseros-eval',
+    cdnBaseUrl: (
+      env.EVAL_R2_CDN_BASE_URL || 'https://eval.browseros.com'
+    ).replace(/\/+$/, ''),
+  }
+}
+
+export function createR2Client(config: R2UploadConfig): S3Client {
+  return new S3Client({
+    region: 'auto',
+    endpoint: `https://${config.accountId}.r2.cloudflarestorage.com`,
+    credentials: {
+      accessKeyId: config.accessKeyId,
+      secretAccessKey: config.secretAccessKey,
+    },
+  })
+}
+
+async function collectFiles(dir: string): Promise<string[]> {
+  const files: string[] = []
+  const entries = await readdir(dir, { withFileTypes: true })
+  for (const entry of entries) {
+    const full = join(dir, entry.name)
+    if (entry.isDirectory()) {
+      files.push(...(await collectFiles(full)))
+    } else {
+      files.push(full)
+    }
+  }
+  return files
+}
+
+async function runPool<T>(
+  items: T[],
+  concurrency: number,
+  fn: (item: T) => Promise<void>,
+): Promise<void> {
+  let i = 0
+  const workers = Array.from({ length: concurrency }, async () => {
+    while (i < items.length) {
+      const idx = i++
+      await fn(items[idx])
+    }
+  })
+  await Promise.all(workers)
+}
+
+async function hasMetadata(dir: string): Promise<boolean> {
+  const metaStat = await stat(join(dir, 'metadata.json')).catch(() => null)
+  return !!metaStat?.isFile()
+}
+
+async function findTaskDirs(runDir: string): Promise<TaskDirEntry[]> {
+  const entries = await readdir(runDir, { withFileTypes: true })
+  const legacyTasks: TaskDirEntry[] = []
+  for (const entry of entries) {
+    if (!entry.isDirectory() || entry.name === 'tasks') continue
+    const taskPath = join(runDir, entry.name)
+    if (await hasMetadata(taskPath)) {
+      legacyTasks.push({
+        taskId: entry.name,
+        taskPath,
+      })
+    }
+  }
+
+  const tasksRoot = join(runDir, 'tasks')
+  const canonicalEntries = await readdir(tasksRoot, {
+    withFileTypes: true,
+  }).catch(() => [])
+  const canonicalTasks: TaskDirEntry[] = []
+  for (const entry of canonicalEntries) {
+    if (!entry.isDirectory()) continue
+    const taskPath = join(tasksRoot, entry.name)
+    if (await hasMetadata(taskPath)) {
+      canonicalTasks.push({
+        taskId: entry.name,
+        taskPath,
+      })
+    }
+  }
+
+  return legacyTasks.length > 0 ? legacyTasks : canonicalTasks
+}
+
+async function isRunDir(dir: string): Promise<boolean> {
+  return (await findTaskDirs(dir)).length > 0
+}
+
+async function collectRunRootFiles(runDir: string): Promise<UploadJob[]> {
+  const entries = await readdir(runDir, { withFileTypes: true })
+  return entries
+    .filter((entry) => entry.isFile())
+    .map((entry) => {
+      const filePath = join(runDir, entry.name)
+      return {
+        key: entry.name,
+        filePath,
+        contentType: contentTypeForPath(filePath),
+      }
+    })
+}
+
+function statusFromMetadata(meta: Record<string, unknown>): string {
+  return meta.termination_reason === 'completed'
+    ? 'completed'
+    : ((meta.termination_reason as string | undefined) ?? 'unknown')
+}
+
+function runIdForDir(runDir: string): string {
+  const timestamp = basename(runDir)
+  const configName = basename(dirname(runDir))
+  return `${configName}-${timestamp}`
+}
+
+/** Publishes eval artifacts in the viewer-compatible R2 layout. */
+export class R2Publisher {
+  private readonly client: R2Client
+  private readonly config: R2UploadConfig
+  private readonly viewerPath: string
+  private readonly concurrency: number
+  private readonly now: () => Date
+
+  constructor(options: R2PublisherOptions) {
+    this.config = options.config
+    this.client = options.client ?? createR2Client(options.config)
+    this.viewerPath =
+      options.viewerPath ??
+      join(import.meta.dirname, '..', 'dashboard', 'viewer.html')
+    this.concurrency = options.concurrency ?? DEFAULT_CONCURRENCY
+    this.now = options.now ?? (() => new Date())
+  }
+
+  async isUploaded(runId: string): Promise<boolean> {
+    try {
+      await this.client.send(
+        new GetObjectCommand({
+          Bucket: this.config.bucket,
+          Key: `runs/${runId}/manifest.json`,
+        }),
+      )
+      return true
+    } catch {
+      return false
+    }
+  }
+
+  async publishPath(inputDir: string): Promise<R2PublishPathResult> {
+    const dirStat = await stat(inputDir).catch(() => null)
+    if (!dirStat?.isDirectory()) {
+      throw new Error(`Not a directory: ${inputDir}`)
+    }
+
+    if (await isRunDir(inputDir)) {
+      const result = await this.publishRun(inputDir, runIdForDir(inputDir))
+      return { uploadedRuns: [result], skippedRuns: [] }
+    }
+
+    const configName = basename(inputDir)
+    const entries = await readdir(inputDir, { withFileTypes: true })
+    const runDirs = entries
+      .filter((entry) => entry.isDirectory())
+      .map((entry) => entry.name)
+      .sort()
+
+    if (runDirs.length === 0) {
+      throw new Error('No run subdirectories found')
+    }
+
+    const uploadedRuns: R2PublishRunResult[] = []
+    const skippedRuns: string[] = []
+    for (const dir of runDirs) {
+      const runId = `${configName}-${dir}`
+      if (await this.isUploaded(runId)) {
+        skippedRuns.push(runId)
+        continue
+      }
+      uploadedRuns.push(await this.publishRun(join(inputDir, dir), runId))
+    }
+
+    return { uploadedRuns, skippedRuns }
+  }
+
+  async publishRun(
+    runDir: string,
+    runId: string = runIdForDir(runDir),
+  ): Promise<R2PublishRunResult> {
+    const taskEntries = await findTaskDirs(runDir)
+
+    if (taskEntries.length === 0) {
+      throw new Error(`No task subdirectories in ${runId}`)
+    }
+
+    const manifestTasks: ViewerManifestTaskInput[] = []
+    const jobs: UploadJob[] = (await collectRunRootFiles(runDir)).map(
+      (job) => ({
+        ...job,
+        key: `runs/${runId}/${job.key}`,
+      }),
+    )
+    let agentConfig: Record<string, unknown> | undefined
+    let dataset: string | undefined
+
+    for (const taskDirEntry of taskEntries) {
+      const { taskId, taskPath } = taskDirEntry
+      const meta = await this.readMetadata(taskPath)
+      if (!meta) continue
+
+      if (!agentConfig && meta.agent_config) {
+        agentConfig = meta.agent_config as Record<string, unknown>
+      }
+      if (!dataset && meta.dataset) dataset = meta.dataset as string
+
+      const files = await collectFiles(taskPath)
+      let screenshotCount = 0
+      for (const file of files) {
+        const relative = file.slice(taskPath.length + 1)
+        if (relative.startsWith('screenshots/') && extname(file) === '.png') {
+          screenshotCount++
+        }
+        // Keep legacy keys during the manifest v2 rollout so cached viewers and
+        // old manifests can still resolve task artifacts.
+        jobs.push({
+          key: `runs/${runId}/${taskId}/${relative}`,
+          filePath: file,
+          contentType: contentTypeForPath(file),
+        })
+        jobs.push({
+          key: `runs/${runId}/tasks/${taskId}/${relative}`,
+          filePath: file,
+          contentType: contentTypeForPath(file),
+        })
+      }
+
+      manifestTasks.push({
+        queryId: (meta.query_id as string | undefined) || taskId,
+        artifactId: taskId,
+        query: (meta.query as string | undefined) || '',
+        startUrl: (meta.start_url as string | undefined) || '',
+        status: statusFromMetadata(meta),
+        durationMs: (meta.total_duration_ms as number | undefined) || 0,
+        screenshotCount:
+          (meta.screenshot_count as number | undefined) || screenshotCount,
+        graderResults:
+          (meta.grader_results as ViewerManifestTaskInput['graderResults']) ||
+          {},
+      })
+    }
+
+    if (manifestTasks.length === 0) {
+      throw new Error(`No completed tasks in ${runId}`)
+    }
+
+    let uploaded = 0
+    await runPool(jobs, this.concurrency, async (job) => {
+      await this.uploadFile(job)
+      uploaded++
+    })
+
+    const manifest = await this.buildManifest(
+      runDir,
+      runId,
+      agentConfig,
+      dataset,
+      manifestTasks,
+    )
+    await this.uploadBuffer(
+      `runs/${runId}/manifest.json`,
+      Buffer.from(JSON.stringify(manifest, null, 2)),
+      'application/json',
+    )
+    await this.uploadBuffer(
+      'viewer.html',
+      await readFile(this.viewerPath),
+      'text/html',
+    )
+
+    return {
+      runId,
+      uploadedFiles: uploaded + 2,
+      viewerUrl: `${this.config.cdnBaseUrl}/viewer.html?run=${encodeURIComponent(runId)}`,
+      manifest,
+    }
+  }
+
+  private async readMetadata(
+    taskPath: string,
+  ): Promise<Record<string, unknown> | null> {
+    try {
+      return JSON.parse(
+        await readFile(join(taskPath, 'metadata.json'), 'utf-8'),
+      ) as Record<string, unknown>
+    } catch {
+      return null
+    }
+  }
+
+  private async buildManifest(
+    runDir: string,
+    runId: string,
+    agentConfig: Record<string, unknown> | undefined,
+    dataset: string | undefined,
+    tasks: ViewerManifestTaskInput[],
+  ): Promise<R2RunManifest> {
+    let summaryData: Record<string, unknown> | undefined
+    try {
+      summaryData = JSON.parse(
+        await readFile(join(runDir, 'summary.json'), 'utf-8'),
+      ) as Record<string, unknown>
+    } catch {}
+
+    return buildViewerManifest({
+      runId,
+      uploadedAt: this.now().toISOString(),
+      agentConfig,
+      dataset,
+      summary: summaryData
+        ? {
+            passRate: summaryData.passRate,
+            avgDurationMs: summaryData.avgDurationMs,
+          }
+        : undefined,
+      tasks,
+    })
+  }
+
+  private async uploadFile(job: UploadJob): Promise<void> {
+    await this.uploadBuffer(
+      job.key,
+      await readFile(job.filePath),
+      job.contentType,
+    )
+  }
+
+  private async uploadBuffer(
+    key: string,
+    body: Buffer,
+    contentType: string,
+  ): Promise<void> {
+    await this.client.send(
+      new PutObjectCommand({
+        Bucket: this.config.bucket,
+        Key: key,
+        Body: body,
+        ContentType: contentType,
+      }),
+    )
+  }
+}
+
+export async function publishPathToR2(
+  inputDir: string,
+): Promise<R2PublishPathResult> {
+  const config = loadR2ConfigFromEnv()
+  return new R2Publisher({ config }).publishPath(inputDir)
+}

packages/browseros-agent/apps/eval/src/reporting/run-summary.ts

@@ -0,0 +1,104 @@
+export interface ReportManifestTask {
+  queryId: string
+  query?: string
+  status: string
+  durationMs: number
+  screenshotCount?: number
+  paths?: Record<string, string>
+  graderResults?: Record<string, { pass?: boolean; score?: number }>
+}
+
+export interface ReportManifest {
+  schemaVersion?: number
+  runId: string
+  uploadedAt?: string
+  agentConfig?: { type?: string; model?: string }
+  dataset?: string
+  summary?: { passRate?: number; avgDurationMs?: number }
+  tasks?: ReportManifestTask[]
+}
+
+export interface RunSummary {
+  runId: string
+  configName: string
+  date: string
+  avgScore: number
+  total: number
+  completed: number
+  failed: number
+  timeout: number
+  avgDurationMs: number
+  model: string
+  dataset: string
+  agentType: string
+}
+
+// Report score uses the primary pass/fail grader so mixed-grader runs keep
+// the same precedence as the eval summary.
+const PASS_FAIL_GRADER_ORDER = [
+  'agisdk_state_diff',
+  'infinity_state',
+  'performance_grader',
+]
+
+export function extractConfigName(runId: string): string {
+  return runId.replace(/-\d{4}-\d{2}-\d{2}-\d{4}$/, '')
+}
+
+function reportDate(manifest: ReportManifest): string {
+  if (!manifest.uploadedAt) return 'unknown'
+  const [date, time] = manifest.uploadedAt.split('T')
+  return `${date} ${time?.slice(0, 5) || ''}`
+}
+
+function primaryScore(task: ReportManifestTask): number | null {
+  if (!task.graderResults) return null
+  for (const name of PASS_FAIL_GRADER_ORDER) {
+    const result = task.graderResults[name]
+    if (result) return result.score ?? 0
+  }
+  return null
+}
+
+export function buildRunSummaries(manifests: ReportManifest[]): RunSummary[] {
+  return manifests
+    .map((manifest) => {
+      const tasks = Array.isArray(manifest.tasks) ? manifest.tasks : []
+      const total = tasks.length
+      const completed = tasks.filter((t) => t.status === 'completed').length
+      const failed = tasks.filter((t) => t.status === 'failed').length
+      const timeout = tasks.filter((t) => t.status === 'timeout').length
+
+      let scoredCount = 0
+      let scoreSum = 0
+      for (const task of tasks) {
+        const score = primaryScore(task)
+        if (score === null) continue
+        scoredCount++
+        scoreSum += score
+      }
+
+      const durations = tasks
+        .filter((t) => t.durationMs > 0)
+        .map((t) => t.durationMs)
+
+      return {
+        runId: manifest.runId,
+        configName: extractConfigName(manifest.runId),
+        date: reportDate(manifest),
+        avgScore: scoredCount > 0 ? (scoreSum / scoredCount) * 100 : 0,
+        total,
+        completed,
+        failed,
+        timeout,
+        avgDurationMs:
+          durations.length > 0
+            ? durations.reduce((a, b) => a + b, 0) / durations.length
+            : 0,
+        model: manifest.agentConfig?.model || 'unknown',
+        dataset: manifest.dataset || manifest.runId,
+        agentType: manifest.agentConfig?.type || 'unknown',
+      }
+    })
+    .sort((a, b) => a.date.localeCompare(b.date))
+}

packages/browseros-agent/apps/eval/src/runner/browseros-app-manager.ts

@@ -14,8 +14,11 @@
  */
 
 import {
+  closeSync,
   existsSync,
+  mkdirSync,
   mkdtempSync,
+  openSync,
   readFileSync,
   rmSync,
   writeFileSync,
@@ -33,7 +36,17 @@ export interface EvalPorts {
 
 const MAX_RESTART_ATTEMPTS = 3
 const CDP_WAIT_TIMEOUT_MS = 30_000
-const SERVER_HEALTH_TIMEOUT_MS = 30_000
+// Bumped from 30s → 90s while debugging dev-CI startup. Dev's server module
+// graph is ~108 files larger than main's; cold-cache module load on a CI
+// runner can take much longer than the original 30s budget allowed.
+const SERVER_HEALTH_TIMEOUT_MS = 90_000
+
+// Where per-worker server stderr is written. Captured (rather than ignored)
+// so eval-weekly.yml can upload these as workflow artifacts on failure for
+// post-mortem debugging. Path is also referenced in the workflow's artifact
+// upload step.
+const SERVER_LOG_DIR =
+  process.env.BROWSEROS_SERVER_LOG_DIR || '/tmp/browseros-server-logs'
 
 const MONOREPO_ROOT = join(
   dirname(fileURLToPath(import.meta.url)),
@@ -53,6 +66,7 @@ export class BrowserOSAppManager {
   private ports: EvalPorts
   private chromeProc: Subprocess | null = null
   private serverProc: Subprocess | null = null
+  private serverLogFd: number | null = null
   private tempDir: string | null = null
   private readonly workerIndex: number
   private readonly loadExtensions: boolean
@@ -183,15 +197,36 @@ export class BrowserOSAppManager {
       VITE_BROWSEROS_SERVER_PORT: String(server),
     }
 
+    // Capture both stdout and stderr to a per-worker file so we can
+    // post-mortem startup hangs. The server uses pino which writes logs to
+    // stdout by default — capturing stderr alone misses everything. The
+    // eval-weekly workflow uploads /tmp/browseros-server-logs/ as a workflow
+    // artifact on failure.
+    // Open the per-worker log file under SERVER_LOG_DIR. If the directory
+    // can't be created or the file can't be opened (e.g. unwritable custom
+    // BROWSEROS_SERVER_LOG_DIR), fall back to /dev/null so spawn still works.
+    const logPath = join(SERVER_LOG_DIR, `server-W${this.workerIndex}.log`)
+    let logFd: number
+    try {
+      mkdirSync(SERVER_LOG_DIR, { recursive: true })
+      logFd = openSync(logPath, 'a')
+    } catch {
+      logFd = openSync('/dev/null', 'w')
+    }
+    this.serverLogFd = logFd
+
+    // `start:ci` skips `--watch` (no file-watcher overhead in CI). Falls back
+    // to the regular `start` script outside CI for the dev-watch experience.
+    const startScript = process.env.CI ? 'start:ci' : 'start'
     this.serverProc = spawn({
-      cmd: ['bun', 'run', '--filter', '@browseros/server', 'start'],
+      cmd: ['bun', 'run', '--filter', '@browseros/server', startScript],
       cwd: MONOREPO_ROOT,
-      stdout: 'ignore',
-      stderr: 'ignore',
+      stdout: logFd,
+      stderr: logFd,
       env: serverEnv,
     })
     console.log(
-      `  [W${this.workerIndex}] Server started (PID: ${this.serverProc.pid})`,
+      `  [W${this.workerIndex}] Server started (PID: ${this.serverProc.pid}, logs → ${logPath})`,
     )
 
     // --- Wait for Server Health ---
@@ -244,6 +279,18 @@ export class BrowserOSAppManager {
     await this.killProcess(this.serverProc)
     this.serverProc = null
 
+    // Close the parent's copy of the server log fd. Child kept its own dup
+    // until it exited above, so closing here doesn't truncate any output.
+    // Without this we'd leak one fd per restart attempt across all workers.
+    if (this.serverLogFd !== null) {
+      try {
+        closeSync(this.serverLogFd)
+      } catch {
+        // already closed or invalid — ignore
+      }
+      this.serverLogFd = null
+    }
+
     // Kill Chrome (graceful → force)
     await this.killProcess(this.chromeProc)
     this.chromeProc = null

packages/browseros-agent/apps/eval/src/runner/eval-runner.ts

@@ -1,362 +1 @@
-import { mkdir, writeFile } from 'node:fs/promises'
-import { basename, dirname, join, resolve } from 'node:path'
-import {
-  dashboardState,
-  setActiveExecutor,
-  startDashboard,
-  stopDashboard,
-} from '../dashboard/server'
-import type { ErrorSource, EvalConfig, Task } from '../types'
-import {
-  printValidationResult,
-  validateConfig,
-} from '../utils/config-validator'
-import { ParallelExecutor } from './parallel-executor'
-import {
-  getTaskSourceDescription,
-  loadTasks,
-  TaskLoadError,
-} from './task-loader'
-import type {
-  BatchSummary,
-  RunEvalOptions,
-  TaskResult,
-  TaskResultSummary,
-  TaskSource,
-} from './types'
-import { getPrimaryGraderResult, isSuccessfulResult } from './types'
-
-// ============================================================================
-// Main Entry Point
-// ============================================================================
-
-export async function runEval(options: RunEvalOptions): Promise<void> {
-  // Step 1: Validate configuration
-  const config = await loadAndValidateConfig(options.configPath)
-
-  // Step 2: Resolve paths relative to config location
-  const configDir = dirname(resolve(options.configPath))
-  const resolvedPaths = resolvePaths(options, config, configDir)
-
-  // Log configuration
-  console.log('Eval Configuration:')
-  console.log(`  Config: ${options.configPath}`)
-  console.log(`  Dataset: ${resolvedPaths.dataPath}`)
-  console.log(`  Output: ${resolvedPaths.outputDir}`)
-  console.log(`  Workers: ${config.num_workers}`)
-  console.log(`  Agent: ${config.agent.type}`)
-  console.log()
-
-  // Step 3: Load tasks
-  const taskSource = resolveTaskSource(options, resolvedPaths.dataPath)
-  const { tasks } = await loadTasksWithLogging(taskSource)
-
-  // Step 4: Setup
-  await mkdir(resolvedPaths.outputDir, { recursive: true })
-
-  // Step 5: Start dashboard
-  startDashboard({
-    tasks,
-    configName: options.configPath,
-    agentType: config.agent.type,
-    outputDir: resolvedPaths.outputDir,
-  })
-
-  // Step 6: Execute tasks (parallel or sequential based on num_workers)
-  const results = await executeTasks(tasks, config, resolvedPaths.outputDir)
-
-  // Step 7: Summary
-  const summary = buildSummary(results)
-  await saveSummary(summary, resolvedPaths.outputDir)
-  printSummary(summary)
-  console.log(`\nResults saved to: ${resolvedPaths.outputDir}`)
-
-  stopDashboard()
-}
-
-// ============================================================================
-// Configuration
-// ============================================================================
-
-async function loadAndValidateConfig(configPath: string) {
-  console.log('Validating configuration...')
-  const validationResult = await validateConfig(configPath)
-  printValidationResult(validationResult)
-
-  if (!validationResult.valid || !validationResult.config) {
-    throw new Error(
-      'Configuration validation failed. Fix the above errors and try again.',
-    )
-  }
-
-  return validationResult.config
-}
-
-interface ResolvedPaths {
-  dataPath: string
-  outputDir: string
-}
-
-function resolvePaths(
-  options: RunEvalOptions,
-  config: EvalConfig,
-  configDir: string,
-): ResolvedPaths {
-  // Resolve dataset path: use options.dataPath if provided, otherwise resolve from config
-  const dataPath = options.dataPath
-    ? options.dataPath
-    : config.dataset.startsWith('/')
-      ? config.dataset
-      : resolve(configDir, config.dataset)
-
-  // Resolve output directory: results/{config-name}/{timestamp}/
-  // Config name derived from config filename (e.g., "browseros-agent-weekly.json" → "browseros-agent-weekly")
-  const configName = options.configPath
-    ? basename(resolve(options.configPath), '.json')
-    : 'eval'
-  const timestamp = formatTimestamp(new Date())
-  const resultsBase = config.output_dir
-    ? config.output_dir.startsWith('/')
-      ? config.output_dir
-      : resolve(configDir, config.output_dir)
-    : resolve(configDir, '..', 'results')
-  const outputDir = join(resultsBase, configName, timestamp)
-
-  return { dataPath, outputDir }
-}
-
-function formatTimestamp(date: Date): string {
-  const y = date.getFullYear()
-  const m = String(date.getMonth() + 1).padStart(2, '0')
-  const d = String(date.getDate()).padStart(2, '0')
-  const h = String(date.getHours()).padStart(2, '0')
-  const min = String(date.getMinutes()).padStart(2, '0')
-  return `${y}-${m}-${d}-${h}${min}`
-}
-
-// ============================================================================
-// Task Loading
-// ============================================================================
-
-function resolveTaskSource(
-  options: RunEvalOptions,
-  dataPath: string,
-): TaskSource {
-  // If query is provided, use single task mode
-  if (options.query) {
-    return { type: 'single', query: options.query, startUrl: options.startUrl }
-  }
-
-  // Otherwise use file mode with the resolved dataPath
-  return { type: 'file', path: dataPath }
-}
-
-async function loadTasksWithLogging(
-  source: TaskSource,
-): Promise<{ tasks: Awaited<ReturnType<typeof loadTasks>>['tasks'] }> {
-  console.log(`Loading tasks from ${getTaskSourceDescription(source)}...`)
-
-  try {
-    const result = await loadTasks(source)
-    console.log(`Loaded ${result.tasks.length} task(s)`)
-    return { tasks: result.tasks }
-  } catch (error) {
-    if (error instanceof TaskLoadError) {
-      throw new Error(`Failed to load tasks: ${error.message}`)
-    }
-    throw new Error(`Failed to load tasks: ${error}`)
-  }
-}
-
-// ============================================================================
-// Task Execution
-// ============================================================================
-
-async function executeTasks(
-  tasks: Task[],
-  config: EvalConfig,
-  outputDir: string,
-): Promise<TaskResult[]> {
-  console.log(`\n${'='.repeat(60)}`)
-  console.log('STARTING EVALUATION')
-  console.log(`${'='.repeat(60)}\n`)
-
-  const numWorkers = config.num_workers || 1
-  console.log(`Running with ${numWorkers} worker(s)`)
-  if (config.restart_server_per_task) {
-    console.log(`Server restart per task: enabled`)
-  }
-  console.log()
-
-  const executor = new ParallelExecutor({
-    numWorkers,
-    config,
-    outputDir,
-    restartServerPerTask: config.restart_server_per_task,
-    onEvent: (taskId, event) =>
-      dashboardState.broadcastStreamEvent(taskId, event),
-  })
-
-  // Register so dashboard stop button works for CLI runs too
-  setActiveExecutor(executor)
-  try {
-    return await executor.execute(tasks, (completed, total, task, result) => {
-      printTaskProgress(completed, total, task, result)
-    })
-  } finally {
-    setActiveExecutor(null)
-  }
-}
-
-function printTaskProgress(
-  completed: number,
-  total: number,
-  task: Task,
-  result: TaskResult,
-): void {
-  const status =
-    result.status === 'completed'
-      ? 'DONE'
-      : result.status === 'timeout'
-        ? 'TIMEOUT'
-        : 'FAILED'
-
-  const duration =
-    result.durationMs > 0 ? ` (${(result.durationMs / 1000).toFixed(1)}s)` : ''
-
-  console.log(`[${completed}/${total}] ${task.query_id}: ${status}${duration}`)
-
-  if (result.status === 'failed') {
-    console.log(`  ERROR: ${result.error.message}`)
-  } else if (isSuccessfulResult(result)) {
-    // Log agent errors (e.g., LLM API failures) even if task "completed"
-    if (result.agentResult.metadata.errors?.length) {
-      for (const err of result.agentResult.metadata.errors) {
-        console.log(`    ERROR [${err.source}]: ${err.message}`)
-      }
-    }
-    for (const [name, gr] of Object.entries(result.graderResults)) {
-      const icon = gr.pass ? 'PASS' : 'FAIL'
-      console.log(`    ${name}: ${icon}`)
-    }
-  }
-}
-
-// ============================================================================
-// Summary
-// ============================================================================
-
-function buildSummary(results: TaskResult[]): BatchSummary {
-  // Track errors by source
-  const errorsBySource: Partial<Record<ErrorSource, number>> = {}
-  let totalWarnings = 0
-
-  const taskSummaries: TaskResultSummary[] = results.map((r) => {
-    let errorCount = 0
-    let warningCount = 0
-    let errorSources: ErrorSource[] | undefined
-    let failureReason: string | undefined
-
-    if (isSuccessfulResult(r)) {
-      // Count errors and warnings from agent metadata
-      errorCount = r.agentResult.metadata.errors?.length ?? 0
-      warningCount = r.agentResult.metadata.warnings?.length ?? 0
-      totalWarnings += warningCount
-
-      // Track error sources
-      if (r.agentResult.metadata.errors?.length) {
-        errorSources = r.agentResult.metadata.errors.map((e) => e.source)
-        for (const err of r.agentResult.metadata.errors) {
-          errorsBySource[err.source] = (errorsBySource[err.source] ?? 0) + 1
-        }
-      }
-    } else {
-      // Failed task
-      errorCount = 1
-      errorSources = [r.errorSource]
-      failureReason = r.error.message
-      errorsBySource[r.errorSource] = (errorsBySource[r.errorSource] ?? 0) + 1
-    }
-
-    return {
-      queryId: r.task.query_id,
-      status: r.status,
-      durationMs: r.durationMs,
-      graderResults: isSuccessfulResult(r)
-        ? Object.fromEntries(
-            Object.entries(r.graderResults).map(([name, gr]) => [
-              name,
-              { pass: gr.pass, score: gr.score },
-            ]),
-          )
-        : undefined,
-      errorCount,
-      warningCount,
-      errorSources: errorSources?.length ? errorSources : undefined,
-      failureReason,
-    }
-  })
-
-  const completed = results.filter((r) => r.status === 'completed').length
-  const timeout = results.filter((r) => r.status === 'timeout').length
-  const failed = results.filter((r) => r.status === 'failed').length
-
-  // Calculate pass rate using primary grader (fallback order)
-  let totalGraded = 0
-  let totalPasses = 0
-
-  for (const result of results) {
-    if (isSuccessfulResult(result)) {
-      const primary = getPrimaryGraderResult(result.graderResults)
-      if (primary) {
-        totalGraded++
-        if (primary.pass) totalPasses++
-      }
-    }
-  }
-
-  const passRate = totalGraded > 0 ? totalPasses / totalGraded : 0
-
-  // Calculate average duration for non-failed tasks
-  const durations = results
-    .filter((r) => r.status !== 'failed')
-    .map((r) => r.durationMs)
-  const avgDurationMs =
-    durations.length > 0
-      ? durations.reduce((a, b) => a + b, 0) / durations.length
-      : 0
-
-  return {
-    total: results.length,
-    completed,
-    failed,
-    timeout,
-    passRate,
-    avgDurationMs,
-    errorsBySource,
-    totalWarnings,
-    results: taskSummaries,
-  }
-}
-
-async function saveSummary(
-  summary: BatchSummary,
-  outputDir: string,
-): Promise<void> {
-  await writeFile(
-    join(outputDir, 'summary.json'),
-    JSON.stringify(summary, null, 2),
-  )
-}
-
-function printSummary(summary: BatchSummary): void {
-  console.log('='.repeat(60))
-  console.log('EVALUATION COMPLETE')
-  console.log('='.repeat(60))
-  console.log(`Total: ${summary.total} tasks`)
-  console.log(`  Completed: ${summary.completed}`)
-  console.log(`  Timeout: ${summary.timeout}`)
-  console.log(`  Failed: ${summary.failed}`)
-  console.log(`  Pass Rate: ${(summary.passRate * 100).toFixed(1)}%`)
-  console.log(`  Avg Duration: ${(summary.avgDurationMs / 1000).toFixed(1)}s`)
-}
+export { runEval } from '../runs/eval-runner'

packages/browseros-agent/apps/eval/src/runner/parallel-executor.ts

@@ -1,266 +1,5 @@
-/**
- * Parallel Executor
- *
- * Each worker gets its own isolated BrowserOS stack:
- *   - BrowserOSAppManager (Chrome + Server on unique ports)
- *   - TaskExecutor (uses that worker's server URL)
- *
- * Port allocation: Worker N → CDP=base+N, Server=base+N, Extension=base+N
- */
-
-import type { EvalConfig, Task } from '../types'
-import { BrowserOSAppManager, type EvalPorts } from './browseros-app-manager'
-import { createTaskExecutor } from './task-executor'
-import type { TaskResult } from './types'
-
-// ============================================================================
-// Types
-// ============================================================================
-
-export interface ParallelExecutorConfig {
-  numWorkers: number
-  config: EvalConfig
-  outputDir: string
-  restartServerPerTask?: boolean
-  onEvent?: (taskId: string, event: Record<string, unknown>) => void
-}
-
-export type ProgressCallback = (
-  completed: number,
-  total: number,
-  task: Task,
-  result: TaskResult,
-) => void
-
-// ============================================================================
-// Task Queue (thread-safe for single-threaded async — index is atomic)
-// ============================================================================
-
-class TaskQueue {
-  private tasks: Task[]
-  private index: number = 0
-  private stopped: boolean = false
-
-  constructor(tasks: Task[]) {
-    this.tasks = [...tasks]
-  }
-
-  next(): Task | null {
-    if (this.stopped || this.index >= this.tasks.length) return null
-    return this.tasks[this.index++]
-  }
-
-  stop(): void {
-    this.stopped = true
-  }
-}
-
-// ============================================================================
-// Parallel Executor
-// ============================================================================
-
-export class ParallelExecutor {
-  private readonly numWorkers: number
-  private readonly appManagers = new Map<number, BrowserOSAppManager>()
-  private completedCount: number = 0
-  private readonly resultLock = new Map<string, TaskResult>()
-  private queue: TaskQueue | null = null
-
-  constructor(private readonly config: ParallelExecutorConfig) {
-    this.numWorkers = Math.max(1, config.numWorkers)
-  }
-
-  async stop(): Promise<void> {
-    console.log('\nStopping eval run...')
-    this.queue?.stop()
-    const kills = [...this.appManagers.values()].map((m) => m.killApp())
-    await Promise.allSettled(kills)
-  }
-
-  async execute(
-    tasks: Task[],
-    onProgress?: ProgressCallback,
-  ): Promise<TaskResult[]> {
-    if (tasks.length === 0) return []
-
-    const cleanup = this.setupSignalHandlers()
-
-    const loadExtensions = this.config.config.browseros.load_extensions ?? false
-
-    // Patch NopeCHA API key before launching any workers
-    const captchaConfig = this.config.config.captcha
-    if (captchaConfig) {
-      const apiKey = process.env[captchaConfig.api_key_env]
-      if (apiKey) {
-        BrowserOSAppManager.patchNopechaApiKey(apiKey)
-      }
-    }
-
-    this.queue = new TaskQueue(tasks)
-    const totalTasks = tasks.length
-
-    try {
-      const queue = this.queue
-      // Launch N workers in parallel — each gets its own Chrome + Server
-      const workers = Array.from({ length: this.numWorkers }, (_, i) =>
-        this.runWorker(i, queue, totalTasks, loadExtensions, onProgress),
-      )
-      await Promise.all(workers)
-
-      // Return results in original task order
-      return tasks.map((task) => {
-        const result = this.resultLock.get(task.query_id)
-        if (!result) {
-          return {
-            status: 'failed' as const,
-            task,
-            error: new Error('Task result not found'),
-            errorSource: 'unknown' as const,
-            durationMs: 0,
-          }
-        }
-        return result
-      })
-    } finally {
-      cleanup()
-    }
-  }
-
-  private async runWorker(
-    workerIndex: number,
-    queue: TaskQueue,
-    totalTasks: number,
-    loadExtensions: boolean,
-    onProgress?: ProgressCallback,
-  ): Promise<void> {
-    // Per-worker isolated ports
-    const basePorts: EvalPorts = {
-      cdp: this.config.config.browseros.base_cdp_port,
-      server: this.config.config.browseros.base_server_port,
-      extension: this.config.config.browseros.base_extension_port,
-    }
-    const headless = this.config.config.browseros.headless ?? false
-    const appManager = new BrowserOSAppManager(
-      workerIndex,
-      basePorts,
-      loadExtensions,
-      headless,
-    )
-    this.appManagers.set(workerIndex, appManager)
-
-    // Per-worker executor pointing to this worker's server
-    const workerConfig: typeof this.config.config = {
-      ...this.config.config,
-      browseros: {
-        ...this.config.config.browseros,
-        server_url: appManager.getServerUrl(),
-      },
-    }
-    const executor = createTaskExecutor(
-      workerConfig,
-      workerIndex,
-      this.config.outputDir,
-      this.config.onEvent,
-    )
-
-    try {
-      // Always start Chrome+Server once for this worker
-      console.log(`\n  Worker ${workerIndex}: Starting BrowserOS stack...`)
-      await appManager.restart()
-
-      while (true) {
-        const task = queue.next()
-        if (!task) break
-
-        const taskStartTime = Date.now()
-        let result: TaskResult
-
-        try {
-          // Restart between tasks if configured
-          if (this.config.restartServerPerTask) {
-            console.log(`\n${'─'.repeat(60)}`)
-            console.log(`  Worker ${workerIndex}: Task: ${task.query_id}`)
-            console.log(`${'─'.repeat(60)}`)
-            await appManager.restart()
-          }
-
-          this.config.onEvent?.(task.query_id, {
-            type: 'task-state',
-            taskId: task.query_id,
-            status: 'running',
-          })
-          result = await executor.execute(task)
-          console.log(
-            `  Worker ${workerIndex}: ${task.query_id}: ${result.status}`,
-          )
-        } catch (error) {
-          console.error(
-            `  Worker ${workerIndex}: ${task.query_id}: FAILED - ${error instanceof Error ? error.message : String(error)}`,
-          )
-          result = {
-            status: 'failed',
-            task,
-            error: error instanceof Error ? error : new Error(String(error)),
-            errorSource: 'unknown',
-            durationMs: Date.now() - taskStartTime,
-          }
-        }
-
-        this.resultLock.set(task.query_id, result)
-        this.completedCount++
-
-        // Emit task completion to dashboard
-        const stateEvent: Record<string, unknown> = {
-          type: 'task-state',
-          taskId: task.query_id,
-          status: result.status,
-          durationMs: result.durationMs,
-        }
-        if (result.status !== 'failed' && 'graderResults' in result) {
-          stateEvent.graderResults = Object.fromEntries(
-            Object.entries(result.graderResults).map(([name, gr]) => [
-              name,
-              {
-                pass: gr.pass,
-                score: gr.score,
-                reasoning: gr.reasoning,
-                details: gr.details,
-              },
-            ]),
-          )
-          stateEvent.screenshotCount =
-            result.agentResult?.metadata?.total_steps ?? 0
-        }
-        this.config.onEvent?.(task.query_id, stateEvent)
-
-        onProgress?.(this.completedCount, totalTasks, task, result)
-
-        if (this.config.restartServerPerTask) {
-          await new Promise((resolve) => setTimeout(resolve, 2000))
-        }
-      }
-    } finally {
-      await appManager.killApp()
-    }
-  }
-
-  /**
-   * SIGINT/SIGTERM kills all Chrome + Server instances across all workers.
-   * Returns a cleanup function that removes the listeners after execute() completes.
-   */
-  private setupSignalHandlers(): () => void {
-    const onSignal = async () => {
-      console.log('\nShutting down all workers...')
-      this.queue?.stop()
-      const kills = [...this.appManagers.values()].map((m) => m.killApp())
-      await Promise.allSettled(kills)
-      process.exit(0)
-    }
-    process.on('SIGINT', onSignal)
-    process.on('SIGTERM', onSignal)
-    return () => {
-      process.off('SIGINT', onSignal)
-      process.off('SIGTERM', onSignal)
-    }
-  }
-}
+export {
+  type ProgressCallback,
+  TaskWorkerPool as ParallelExecutor,
+  type TaskWorkerPoolConfig as ParallelExecutorConfig,
+} from '../runs/task-worker-pool'

packages/browseros-agent/apps/eval/src/runner/task-executor.ts

@@ -1,316 +1,6 @@
-import { join } from 'node:path'
-import { createAgent } from '../agents'
-import type { AgentContext, AgentResult } from '../agents/types'
-import { CaptureContext } from '../capture/context'
-import {
-  hasExistingGraderResults,
-  TrajectorySaver,
-} from '../capture/trajectory-saver'
-import { runGraders } from '../graders/registry'
-import type { ErrorSource, EvalConfig, GraderResult, Task } from '../types'
-import { callMcpTool } from '../utils/mcp-client'
-import { InfinityAppManager } from './infinity-app-manager'
-import type { TaskResult } from './types'
-
-// ============================================================================
-// Errors
-// ============================================================================
-
-export class TaskExecutionError extends Error {
-  public readonly errorSource: ErrorSource
-
-  constructor(
-    message: string,
-    public readonly task: Task,
-    public readonly phase:
-      | 'navigation'
-      | 'agent_execution'
-      | 'grading'
-      | 'cleanup',
-    public readonly cause?: Error,
-  ) {
-    super(message)
-    this.name = 'TaskExecutionError'
-    this.errorSource = phase as ErrorSource
-  }
-}
-
-// ============================================================================
-// Task Executor
-// ============================================================================
-
-export interface TaskExecutorDeps {
-  onEvent?: (taskId: string, event: Record<string, unknown>) => void
-}
-
-export class TaskExecutor {
-  constructor(
-    private readonly config: EvalConfig,
-    private readonly workerIndex: number,
-    private readonly outputDir: string,
-    private readonly deps: TaskExecutorDeps,
-  ) {}
-
-  /**
-   * Resolve the initial page ID via list_pages MCP call.
-   * Called once per task on a fresh browser — there's exactly one page.
-   */
-  private async resolveInitialPageId(mcpUrl: string): Promise<number> {
-    try {
-      const result = await callMcpTool(mcpUrl, 'list_pages', {})
-      if (!result.isError) {
-        const textContent = result.content?.find(
-          (c: { type: string }) => c.type === 'text',
-        )
-        const match = textContent?.text?.match(/^\s*(\d+)\./m)
-        if (match) return Number.parseInt(match[1], 10)
-      }
-    } catch {
-      // Fall through to default
-    }
-    // Fresh browser always has page 1
-    return 1
-  }
-
-  async execute(task: Task): Promise<TaskResult> {
-    const startTime = Date.now()
-    const mcpUrl = `${this.config.browseros.server_url}/mcp`
-
-    // Check if task already has grader results (resume capability)
-    const existing = await hasExistingGraderResults(
-      this.outputDir,
-      task.query_id,
-    )
-    if (existing.exists && existing.metadata) {
-      console.log(`  Skipping: already has grader results`)
-      return {
-        status:
-          existing.metadata.termination_reason === 'timeout'
-            ? 'timeout'
-            : 'completed',
-        task,
-        agentResult: {
-          metadata: existing.metadata,
-          messages: [],
-          finalAnswer: existing.metadata.final_answer,
-        },
-        graderResults: existing.metadata.grader_results,
-        durationMs: existing.metadata.total_duration_ms,
-      }
-    }
-
-    // Resolve page ID once — fresh browser has exactly one page
-    const pageId = await this.resolveInitialPageId(mcpUrl)
-
-    // For Infinity tasks, start a fresh app server per task
-    let infinityManager: InfinityAppManager | null = null
-    let actualStartUrl = task.start_url
-
-    if (task.dataset === 'webarena-infinity') {
-      const appName = (task.metadata?.additional as Record<string, unknown>)
-        ?.app_name as string
-      const appBasePort =
-        ((task.metadata?.additional as Record<string, unknown>)
-          ?.app_base_port as number) || 8000
-
-      if (appName && process.env.WEBARENA_INFINITY_DIR) {
-        infinityManager = new InfinityAppManager(this.workerIndex, appBasePort)
-        try {
-          actualStartUrl = await infinityManager.startApp(appName)
-          console.log(
-            `  Infinity app "${appName}" started on port ${infinityManager.getPort()}`,
-          )
-        } catch (error) {
-          throw new TaskExecutionError(
-            `Failed to start Infinity app: ${error instanceof Error ? error.message : String(error)}`,
-            task,
-            'navigation',
-            error instanceof Error ? error : undefined,
-          )
-        }
-      }
-    }
-
-    try {
-      // Phase 1: Set viewport + navigate to start URL
-      try {
-        await callMcpTool(mcpUrl, 'evaluate_script', {
-          page: pageId,
-          expression: 'window.resizeTo(1440, 900)',
-        })
-      } catch (vpError) {
-        console.warn(
-          `  Viewport resize failed: ${vpError instanceof Error ? vpError.message : String(vpError)}`,
-        )
-      }
-
-      if (actualStartUrl && actualStartUrl !== 'about:blank') {
-        try {
-          await callMcpTool(mcpUrl, 'navigate_page', {
-            url: actualStartUrl,
-            page: pageId,
-          })
-        } catch (error) {
-          throw new TaskExecutionError(
-            `Failed to navigate to start URL: ${error instanceof Error ? error.message : String(error)}`,
-            task,
-            'navigation',
-            error instanceof Error ? error : undefined,
-          )
-        }
-      }
-
-      // Phase 2: Execute agent
-      const agentResult = await this.executeAgent(task, pageId)
-
-      // Phase 3: Run graders
-      const graderResults = await this.runGraders(
-        task,
-        agentResult,
-        infinityManager?.getUrl(),
-      )
-
-      const status =
-        agentResult.metadata.termination_reason === 'timeout'
-          ? 'timeout'
-          : 'completed'
-
-      return {
-        status,
-        task,
-        agentResult,
-        graderResults,
-        durationMs: Date.now() - startTime,
-      }
-    } catch (error) {
-      const errorSource: ErrorSource =
-        error instanceof TaskExecutionError ? error.errorSource : 'unknown'
-
-      return {
-        status: 'failed',
-        task,
-        error: error instanceof Error ? error : new Error(String(error)),
-        errorSource,
-        durationMs: Date.now() - startTime,
-      }
-    } finally {
-      // Navigate to about:blank to clean up
-      try {
-        await callMcpTool(mcpUrl, 'navigate_page', {
-          url: 'about:blank',
-          page: pageId,
-        })
-      } catch {
-        // Ignore cleanup errors
-      }
-
-      // Stop Infinity app server if running
-      if (infinityManager) {
-        await infinityManager.stop().catch(() => {})
-      }
-    }
-  }
-
-  private async executeAgent(task: Task, pageId: number): Promise<AgentResult> {
-    try {
-      const { capture, taskOutputDir } = await CaptureContext.create({
-        serverUrl: this.config.browseros.server_url,
-        outputDir: this.outputDir,
-        taskId: task.query_id,
-        initialPageId: pageId,
-        onEvent: this.deps.onEvent,
-      })
-
-      const context: AgentContext = {
-        config: this.config,
-        task,
-        workerIndex: this.workerIndex,
-        initialPageId: pageId,
-        outputDir: this.outputDir,
-        taskOutputDir,
-        capture,
-      }
-
-      const agent = createAgent(context)
-      return await agent.execute()
-    } catch (error) {
-      if (error instanceof TaskExecutionError) {
-        throw error
-      }
-      throw new TaskExecutionError(
-        `Agent execution failed: ${error instanceof Error ? error.message : String(error)}`,
-        task,
-        'agent_execution',
-        error instanceof Error ? error : undefined,
-      )
-    }
-  }
-
-  private async runGraders(
-    task: Task,
-    agentResult: AgentResult,
-    infinityAppUrl?: string,
-  ): Promise<Record<string, GraderResult>> {
-    const configGraders = this.config.graders ?? []
-    const taskGraders = task.graders ?? []
-    const graderNames = configGraders.length > 0 ? configGraders : taskGraders
-    if (graderNames.length === 0) {
-      return {}
-    }
-
-    try {
-      const graderResults = await runGraders(graderNames, {
-        task: {
-          query_id: task.query_id,
-          query: task.query,
-          dataset: task.dataset,
-        },
-        messages: agentResult.messages,
-        screenshotCount:
-          agentResult.metadata.screenshot_count ??
-          agentResult.metadata.total_steps,
-        finalAnswer: agentResult.finalAnswer,
-        expectedAnswer: (task.metadata?.additional as Record<string, unknown>)
-          ?.answer as string | undefined,
-        outputDir: join(this.outputDir, task.query_id),
-        mcpUrl: `${this.config.browseros.server_url}/mcp`,
-        infinityAppUrl,
-      })
-
-      try {
-        const saver = new TrajectorySaver(this.outputDir, task.query_id)
-        await saver.updateGraderResults(graderResults)
-      } catch (saveError) {
-        console.warn(
-          `  Failed to persist grader results: ${saveError instanceof Error ? saveError.message : String(saveError)}`,
-        )
-      }
-
-      return graderResults
-    } catch (error) {
-      console.warn(
-        `  Grading failed: ${error instanceof Error ? error.message : String(error)}`,
-      )
-      return {
-        _error: {
-          score: 0,
-          pass: false,
-          reasoning: `Grading failed: ${error instanceof Error ? error.message : String(error)}`,
-        },
-      }
-    }
-  }
-}
-
-// ============================================================================
-// Factory
-// ============================================================================
-
-export function createTaskExecutor(
-  config: EvalConfig,
-  workerIndex: number,
-  outputDir: string,
-  onEvent?: (taskId: string, event: Record<string, unknown>) => void,
-): TaskExecutor {
-  return new TaskExecutor(config, workerIndex, outputDir, { onEvent })
-}
+export {
+  createTaskRunPipeline as createTaskExecutor,
+  TaskExecutionError,
+  TaskRunPipeline as TaskExecutor,
+  type TaskRunPipelineDeps as TaskExecutorDeps,
+} from '../runs/task-run-pipeline'

packages/browseros-agent/apps/eval/src/runner/types.ts

@@ -8,12 +8,18 @@ import type { ErrorSource, EvalConfig, GraderResult, Task } from '../types'
 
 export interface RunEvalOptions {
   configPath: string
+  config?: EvalConfig
   dataPath?: string
   query?: string
   startUrl?: string
   outputDir?: string
 }
 
+export interface RunEvalResult {
+  outputDir: string
+  summary: BatchSummary
+}
+
 // ============================================================================
 // Task Loading
 // ============================================================================