Add per-conversation multiprocess execution and browser interaction/eval improvements

## Summary

This branch started as a per-conversation multi-process isolation change, but it also includes a broader set of browser automation, prompt, and evaluation improvements that landed along the way.

At a high level, this PR:

- adds worker-process execution for agent messages and browser command handling
- improves screenshot and highlight stability for background-tab automation
- extends element interaction support with swipe/carousel handling
- adds a new BlueBook evaluation site and datasets
- refreshes prompt guidance, model/tool selection behavior, and dependency pins
- fixes evaluation cost extraction and updates the checked-in evaluation report

## Main Changes

### 1. Per-conversation multi-process execution

- Added optional multi-process mode in the agent manager, with one worker process per conversation.
- Introduced `ProcessManager` and `BrowserExecutorBundle` to encapsulate process lifecycle, worker queues, agent manager initialization, and command execution in workers.
- Updated agent message processing so SSE can stream from worker queues instead of only in-process threads.
- Added worker-side handling for:
  - agent message execution
  - browser command execution
  - pause control
  - clean shutdown
- Updated session/status handling around worker execution and disconnect paths.
- Added unit coverage for multi-process API flow, process lifecycle, bundle behavior, and related command/model cases.

Key files:
`server/agent/manager.py`
`server/agent/api.py`
`server/api/routes/agent.py`
`server/core/process_manager.py`
`server/core/browser_executor_bundle.py`

### 2. Highlight readiness and screenshot stability

- Reworked highlight readiness to use a snapshot-first approach instead of relying on page-side polling loops that can be throttled in background tabs.
- Added consistency/stability handling around highlight capture and screenshot timing.
- Improved screenshot capture by waiting for page settle before capture, including fonts, viewport mutations, media readiness, and quiet windows.
- Adjusted collision/highlight behavior so `element_type="any"` better surfaces scrollable regions and remains more stable visually.
- Documented the new highlight-readiness design in project docs.

Key files:
`extension/src/commands/highlight-detection.injected.js`
`extension/src/commands/highlight-detection.ts`
`extension/src/commands/screenshot.ts`
`extension/src/utils/layout-stability.ts`
`extension/src/utils/collision-detection.ts`
`AGENTS.md`

### 3. Element interaction improvements, including swipe support

- Added swipe/carousel interaction support for swipable regions.
- Updated element-interaction prompts and tool guidance so the agent can distinguish scrollable containers from swipable carousel/slider regions.
- Fixed async JavaScript result handling for click flows where dialogs can open during JS-driven interactions, avoiding false failures.
- Expanded command/type support so swipe is treated as a first-class interaction.

Key files:
`extension/src/commands/element-actions.ts`
`extension/src/background/index.ts`
`server/models/commands.py`
`server/agent/prompts/element_interaction_tool.j2`
`server/agent/prompts/highlight_tool.j2`
`server/agent/prompts/tab_tool.j2`

### 4. Prompt and model/tooling refinements

- Refined highlight prompt guidance, especially around when to use keywords vs pagination.
- Tightened JavaScript prompt guidance so JS is positioned as a narrow system-specific fallback instead of a primary interaction path.
- Updated prompt context/tool profile behavior to better match model tier and browser tool availability.
- Bumped `openhands-sdk` / `openhands-tools` dependency revisions and refreshed the lockfile.

Key files:
`server/agent/prompts/highlight_tool.j2`
`server/agent/prompts/javascript_tool.j2`
`server/agent/tools/base.py`
`server/agent/tools/browser_executor.py`
`server/agent/tools/prompt_context.py`
`pyproject.toml`
`uv.lock`

### 5. Evaluation expansion and maintenance

- Added the new BlueBook evaluation site, assets, and two datasets:
  - `bluebook_simple`
  - `bluebook_complex`
- Updated the eval server/docs to expose the new scenario.
- Refreshed the checked-in evaluation report.
- Fixed evaluation cost extraction to use the final `usage_metrics` SSE snapshot rather than the first one, which was underreporting costs.
- Recomputed `eval/evaluation_report.json` from the saved SSE output.

Key files:
`eval/bluebook/index.html`
`eval/bluebook/js/bluebook.js`
`eval/bluebook/css/bluebook.css`
`eval/dataset/bluebook_simple.yaml`
`eval/dataset/bluebook_complex.yaml`
`eval/server.py`
`eval/README.md`
`eval/evaluate_browser_agent.py`
`eval/evaluation_report.json`

## Testing

Added or updated tests for:

- multi-process agent API behavior
- process manager / worker lifecycle
- browser executor bundle behavior
- command model coverage
- prompt/profile behavior
- highlight detection and layout stability
- screenshot capture behavior
- element-action regression coverage
- eval client cost extraction behavior

Representative test files:
`server/tests/unit/test_agent_api_multiprocess.py`
`server/tests/unit/test_agent_manager_process.py`
`server/tests/unit/test_browser_executor_bundle.py`
`server/tests/unit/test_eval_client.py`
`extension/src/__tests__/highlight-detection.test.ts`
`extension/src/__tests__/highlight-layout-stability.test.ts`
`extension/src/__tests__/element-actions-regression.test.ts`

## Notes

This branch also contains a few non-core artifacts alongside product changes, including:

- checked-in eval output/report updates
- lock files under `eval/.locks/`
- `bug_report_highlight_any_scrollable.md`

Author: softpudding

AGENTS.md

@@ -205,6 +205,16 @@ Elements are paginated to ensure **no visual overlap** in each screenshot:
 - AI calls `page=1, page=2, page=3...` to see all elements of that type
 - No offset/limit - pages are determined by collision geometry
 
+### Highlight Readiness Behavior
+
+- `highlight_elements` now uses a **snapshot-first** readiness check instead of page-side polling loops.
+- Reason: OpenBrowser intentionally keeps automated tabs in the browser background, and Chrome may heavily throttle hidden-tab timers. A page-side `setTimeout` stability loop can therefore take far longer than its nominal budget and become the main cause of highlight timeouts.
+- The extension samples viewport readiness signals once per attempt: document readiness, viewport text/media density, pending images, and loading placeholders such as skeleton/shimmer/spinner indicators.
+- Readiness is graded as `ready`, `provisionally_ready`, or `not_ready`.
+- If readiness is `not_ready`, the extension performs only a couple of short **background-side** retries before proceeding or returning the latest result.
+- After screenshot capture, highlight still runs a **consistency check**. This is a drift detector, not a loading detector: it verifies whether sampled highlighted elements moved or disappeared between detection and screenshot.
+- Design rule: prefer snapshot classification plus bounded retries; avoid depending on repeated timers inside the target page for highlight stability.
+
 ```
 # Highlight clickable elements (default)
 highlight_elements()                  → Page 1 of clickable elements

bug_report_highlight_any_scrollable.md

@@ -0,0 +1,66 @@
+## Bug Report: `highlight_elements` with `type='any'` Never Returns Scrollable Elements
+
+**Date:** 2026-03-22
+**Component:** `extension/src/commands/highlight-detection.injected.js`
+**Severity:** Medium
+**Type:** Logic Error / Dead Code
+
+---
+
+### Summary
+
+When calling `highlight_elements(element_type='any')`, scrollable elements are effectively **never returned** due to short-circuit logic in `resolveElementCandidate()`. The scrollable check (lines 948-957) is dead code for `type='any'`.
+
+---
+
+### Root Cause
+
+In `resolveElementCandidate(el, 'any')` (lines 837-973):
+
+```javascript
+const clickableCandidate = resolveClickableCandidate(el);  // Traverses UP DOM tree
+
+if (clickableCandidate) {
+    return { type: 'clickable', ... };  // ← EARLY RETURN
+}
+
+// Lines 924-969 below are NEVER reached when clickableCandidate exists:
+if (isInputableCandidate(el))      // dead code
+if (isSelectableCandidate(el))     // dead code
+if (isScrollableCandidate(el))    // dead code ← SCROLLABLE NEVER RETURNED
+if (isHoverableCandidate(el))      // dead code
+```
+
+The function short-circuits on the first clickable ancestor found (by traversing UP the DOM tree), making all subsequent type checks dead code for `type='any'`.
+
+---
+
+### Expected Behavior
+
+Per collision-aware pagination design, `type='any'` should return elements across **all types** (clickable, inputable, selectable, scrollable, hoverable), prioritized by `HIGHLIGHT_TYPE_PRIORITY`.
+
+---
+
+### Actual Behavior
+
+`type='any'` effectively becomes `type='clickable'`. Elements are only returned if:
+- They themselves are clickable, OR
+- Any **ancestor** has clickable characteristics (pointer cursor + text content)
+
+This causes:
+- Scrollable divs to be misclassified as "clickable" (their ancestor)
+- Hoverable elements to be skipped entirely
+- False positives where a non-interactive ancestor shadows the actual target element
+
+---
+
+### Affected Code Paths
+
+- `resolveElementCandidate()` — lines 912-922 (early return on `clickableCandidate`)
+- `resolveClickableCandidate()` — lines 632-689 (DOM tree traversal with `isTightClickableWrapper` checks)
+
+---
+
+### Suggested Fix
+
+For `type='any'`, collect candidates across ALL types and use `compareCandidates()` with `HIGHLIGHT_TYPE_PRIORITY` for selection — matching the behavior of the specific-type paths (lines 840-910) rather than short-circuiting.

eval/.locks/evaluation_7a77a63b-ab2a-4e1b-9734-66a4dfe1d6fe.lock

@@ -0,0 +1 @@
+{"pid": 48460, "browser_uuid": "ff2b4397-b8f6-4346-9f66-bf1b4d9a9804", "started_at": "2026-03-22T11:28:57.972109"}

eval/.locks/evaluation_ff2b4397-b8f6-4346-9f66-bf1b4d9a9804.lock

@@ -1,6 +1,6 @@
 # Mock Websites for AI Agent Evaluation
 
-This directory contains 5 mocked frontend websites designed to test browser-operating AI agents' capabilities.
+This directory contains mocked frontend websites designed to test browser-operating AI agents' capabilities.
 
 ## Websites
 
@@ -61,6 +61,18 @@ This directory contains 5 mocked frontend websites designed to test browser-oper
   - Multiple view modes (Overview, Valuation, Financial, etc.)
   - Dark theme matching original finviz.com
 
+### 6. BlueBook Feed (Hard)
+- **URL**: `/bluebook/`
+- **Difficulty**: Hard
+- **Purpose**: Test Xiaohongshu-style visual browsing, search, dense card layouts, modal note reading, and comment interactions
+- **Features**:
+  - Dense masonry feed with 70+ mocked posts
+  - Search bar with separate clear/search icon buttons
+  - Floating "graphic only" and "reload" buttons
+  - Note detail modal with left media area and right comment panel
+  - Comment like / reply interactions with author-specific tracking
+  - Shared tracker integration plus site-specific events
+
 ## Event Tracking
 
 All websites include comprehensive event tracking that records: