feat(browser): add agentic browser control (#4187)

* feat(browser): add agentic browser control

* fix(frontend): format browser view changes

* fix(browser): keep browser optional and isolate sidecar layout

* fix(browser): address PR review security and IME findings

- Nginx: add a browser-stream WebSocket location before the generic
  /api/threads regex so Live upgrades instead of downgrading to HTTP
  (both nginx.conf and nginx.local.conf).
- Ownership: require an existing owned thread for the WS stream and REST
  navigate, and tear down the browser session on thread deletion so a
  later caller cannot reuse a retained page/cookies by guessing the id.
- SSRF: enforce the URL policy at the browser request boundary via a
  context-level route guard covering redirects, popups, iframes, and
  subresources (skipped for CDP-attached Chrome).
- IME: skip key forwarding while a composition is active so confirming a
  CJK candidate with Enter no longer submits the remote page form.

Adds regression tests for the request guard, session teardown on delete,
and the composing-Enter key decision.

* fix(frontend): smooth streaming in long tool threads

* Revert "fix(frontend): smooth streaming in long tool threads"

This reverts commit f0462516eabe77f138d4027ea1c714fb226683cf.

* fix(browser): address review security and lifecycle findings

- Reject cross-origin WebSocket upgrades on the live browser stream
  (Origin allow-list reuse of CORS/same-origin helpers) to close a
  WS-CSRF hole, and fail closed when the ownership store is absent.
- Warn when a CDP-attached session runs with the SSRF request guard
  off, and drop the unreachable CDP screencast teardown dead code.
- Read browser session launch config from a single canonical source
  (browser_navigate) so it is deterministic regardless of call order.
- Bound per-thread Chromium accumulation with idle-timeout eviction
  and an LRU max-sessions cap.
- Reset the Live reconnect counter on a successful open so the stream
  can't permanently stall after the cumulative attempt cap.

* fix(frontend): reduce long tool thread render stalls

Reuse stable historical message groups during streaming, defer heavy Markdown and browser previews, and lazy-decode message images.

* fix(browser): keep live control responsive during continuous input

Why: Manual browser control felt laggy — a physical click ran the remote
Playwright click three times and each non-move input synchronously awaited a
JPEG screenshot, so events queued behind capture (queue wait up to ~237ms).
The first async attempt used a trailing-edge debounce, which froze the visible
page until a wheel/keyboard gesture stopped ("scroll finishes, then it jumps").

What:
- Frontend forwards one `click` per physical click instead of also emitting
  `down`/`up`, so the remote page is not clicked twice per gesture.
- Backend detaches live-frame capture from input dispatch: non-move actions
  start a rate-limited background refresh loop (leading frame + bounded cadence)
  that keeps emitting frames while input continues and never blocks dispatch.
- Add regression tests: input dispatch no longer awaits the screenshot, rapid
  inputs coalesce, and continuous input keeps refreshing before it stops.

Scenarios: Verified in the live Browser panel — a single click completes in
~57ms (was blocked behind a 171ms capture), and a 1.14s sustained wheel gesture
renders ~7 frames throughout the scroll instead of one frame after it ends.

* fix(browser): harden worker and session lifecycle

* fix(browser): address latest review feedback

* fix(frontend): preserve optimistic new-chat message

* test(e2e): preserve mocked message run ids

* fix(browser): address capability review feedback

---------

Co-authored-by: Willem Jiang <willem.jiang@gmail.com>

README.md

@@ -778,6 +778,20 @@ This is the difference between a chatbot with tool access and an agent with an a
 └── outputs/          ← final deliverables
 ```
 
+### Agentic Browser Control
+
+Reading a page is not the same as *using* one. Alongside the read-only `web_fetch` and `web_capture` tools, DeerFlow ships an optional agentic browser tool group that keeps a live, per-conversation browser session so the agent can actually operate a page — navigate, read the interactive elements, click, type, submit forms, and follow multi-step flows on JavaScript-heavy sites.
+
+Each action returns a fresh snapshot of the page's interactive elements, each addressed by a stable `[ref]` number, so the agent acts on what it just observed instead of guessing selectors. Outbound URLs are SSRF-screened by default. It is powered by Playwright and shipped as an optional extra so the core install stays lean:
+
+```bash
+cd backend
+uv sync --extra browser
+uv run playwright install chromium
+```
+
+Then uncomment the `group: browser` tool entries in `config.yaml` (`browser_navigate`, `browser_snapshot`, `browser_click`, `browser_type`, `browser_get_text`, `browser_back`, `browser_screenshot`, `browser_close`). `make dev` / Docker startup detects an enabled `browser_navigate` tool and preserves the `browser` extra on dependency syncs. The Gateway fails startup if browser control is configured but Playwright is missing, and `/api/features` hides the Browser UI unless the backend can actually serve it. Keep `headless: true` and `allow_private_addresses: false` for anything but local, trusted debugging. Attaching to an existing Chrome with `cdp_url` cannot enforce DeerFlow's subresource/redirect SSRF guard and therefore fails closed unless `allow_unguarded_cdp: true` explicitly acknowledges that risk; use it only with a trusted local browser. Browser sessions are process-local; keep `GATEWAY_WORKERS=1` while this tool group is enabled because ordinary uvicorn worker dispatch does not provide thread affinity.
+
 ### Context Engineering
 
 **Isolated Sub-Agent Context**: Each sub-agent runs in its own isolated context. This means that the sub-agent will not be able to see the context of the main agent or other sub-agents. This is important to ensure that the sub-agent is able to focus on the task at hand and not be distracted by the context of the main agent or other sub-agents.

backend/AGENTS.md

@@ -163,6 +163,11 @@ Boundary check (harness → app import firewall):
 
 CI runs these regression tests for every pull request via [.github/workflows/backend-unit-tests.yml](../.github/workflows/backend-unit-tests.yml).
 
+Agentic browser sessions are process-local. The Gateway startup safety gate rejects
+`GATEWAY_WORKERS > 1` when `browser_navigate` is configured, because ordinary
+uvicorn worker dispatch does not provide thread affinity for browser tools, REST
+navigation, and the Live WebSocket.
+
 ## Architecture
 
 ### Harness / App Split
@@ -329,7 +334,7 @@ Localhost persistence deliberately reads the direct request `Host` and ignores `
 | Router | Endpoints |
 |--------|-----------|
 | **Models** (`/api/models`) | `GET /` - list models; `GET /{name}` - model details |
-| **Features** (`/api/features`) | `GET /` - report config-gated feature availability (currently `agents_api.enabled`) for frontend UI gating |
+| **Features** (`/api/features`) | `GET /` - report config-gated feature availability (`agents_api.enabled`, `browser_control.enabled`) for frontend UI gating |
 | **Console** (`/api/console`) | Read-only cross-thread observability for the current user (the data layer for an operations dashboard or external monitoring): `GET /stats` - headline counters (runs/threads/agents/tokens/cost); `GET /runs` - paginated run history joined with thread titles (per-run cost); `GET /usage` - zero-filled daily token series + per-model breakdown with spend. Queries `runs`/`threads_meta` directly as a reporting layer (no new `RunStore` methods); requires a SQL database backend — returns 503 on `database.backend: memory`. Real-cost estimation reads optional `models[*].pricing` (`currency`, `input_per_million`, `output_per_million`, `input_cache_hit_per_million`; `ModelConfig` is `extra="allow"`, so no schema change) and prices each run from its `token_usage_by_model` input/output split. Pricing is **cache-aware**: `RunJournal` accumulates prompt-cache hits from `usage_metadata.input_token_details.cache_read` into a sparse `cache_read_tokens` bucket key (also threaded through `SubagentTokenCollector` → `record_external_llm_usage_records`), and cache-hit input tokens are billed at `input_cache_hit_per_million` (omitted → billed at the miss price, a conservative upper bound). Legacy rows fall back to run-level totals at `model_name`; unpriced models yield `cost: null` and cost fields are null when no pricing is configured |
 | **MCP** (`/api/mcp`) | `GET /config` - get config; `PUT /config` - update config (saves to extensions_config.json) |
 | **Skills** (`/api/skills`) | `GET /` - list skills; `GET /{name}` - details; `PUT /{name}` - update enabled; `POST /install` - install from .skill archive (accepts standard optional frontmatter like `version`, `author`, `compatibility`); `POST /reload` - admin-only process-local prompt-cache invalidation after trusted external filesystem changes |
@@ -454,6 +459,8 @@ Scheduled-task runtime note:
 - `firecrawl/` - Web scraping via Firecrawl API
 - `image_search/` - Image search via DuckDuckGo
 - `aio_sandbox/` - Docker-based isolation (`AioSandboxProvider`)
+- `browser_automation/` - Agentic browser control (stateful `navigate → observe → click/type` loop) via Playwright, distinct from the read-only `web_fetch`/`web_capture` tools. Tools: `browser_navigate`, `browser_snapshot`, `browser_click`, `browser_type`, `browser_get_text`, `browser_back`, `browser_screenshot`, `browser_close` (config `group: browser`). A process-local `BrowserSessionManager` owns one private, loop-affine Playwright event-loop thread (same pattern as the BoxLite provider) so a per-thread browser session survives across turns regardless of the caller's loop (Gateway / TUI / test). Each action returns a fresh page snapshot whose interactive elements are addressed by a stable numeric `[ref]` index (stamped as `data-df-ref` during snapshot), so the model acts on what it just observed instead of holding stale handles or guessing selectors. URLs are SSRF-screened via the shared `validate_public_http_url` (opt-out `allow_private_addresses` only for intentional internal targets). CDP attachment cannot install the request guard on an existing Chrome context, so `cdp_url` fails closed unless the operator explicitly sets `allow_unguarded_cdp: true` for a trusted local browser. Browser REST/Live access also requires an exact non-NULL thread owner, rather than the general legacy shared-thread policy, because retained pages may contain authenticated state. Session admission is a hard `max_sessions` cap: pinned Live/operation sessions are never evicted, and a new thread is rejected when no unpinned session can be closed; one Live viewer owns a session at a time. Optional dependency: `cd backend && uv sync --extra browser && uv run playwright install chromium`; `scripts/detect_uv_extras.py` preserves the extra when `config.yaml` enables `browser_navigate`, and Gateway startup fails fast if configured browser control cannot import Playwright. Tests: `tests/test_browser_automation.py` (mocked tools + a real-Chromium integration test guarded by `importorskip`); `tests/manual_browser_live_check.py` is a manual DeepSeek-driven end-to-end check (not collected by pytest).
+  Live UI input dispatch is kept independent from JPEG capture: non-move actions start a rate-limited background refresh loop, so pointer, wheel, or keyboard input stays responsive while continuous gestures still produce frames throughout the interaction.
 
 Additional providers also live here (`boxlite`, `brave`, `browserless`, `crawl4ai`, `ddg_search`, `e2b_sandbox`, `exa`, `fastcrw`, `groundroute`, `infoquest`, `searxng`, `serper`); see each subpackage for specifics.
 

backend/app/gateway/app.py

@@ -8,6 +8,7 @@ from fastapi.middleware.cors import CORSMiddleware
 
 from app.gateway.auth_disabled import warn_if_auth_disabled_enabled
 from app.gateway.auth_middleware import AuthMiddleware
+from app.gateway.browser_capability import ensure_browser_runtime_available
 from app.gateway.config import get_gateway_config
 from app.gateway.csrf_middleware import CSRFMiddleware, get_configured_cors_origins
 from app.gateway.deps import langgraph_runtime
@@ -16,6 +17,7 @@ from app.gateway.routers import (
     artifacts,
     assistants_compat,
     auth,
+    browser,
     channel_connections,
     channels,
     console,
@@ -181,6 +183,7 @@ async def lifespan(app: FastAPI) -> AsyncGenerator[None, None]:
     try:
         startup_config = get_app_config()
         configure_logging(startup_config)
+        ensure_browser_runtime_available(startup_config)
         logger.info("Configuration loaded successfully")
         warn_if_auth_disabled_enabled()
     except Exception as e:
@@ -301,6 +304,23 @@ async def lifespan(app: FastAPI) -> AsyncGenerator[None, None]:
             except Exception:
                 logger.exception("Failed to stop scheduled task service")
 
+        try:
+            from deerflow.community.browser_automation import get_browser_session_manager
+
+            closed = await asyncio.wait_for(
+                get_browser_session_manager().close_all_sessions(),
+                timeout=_SHUTDOWN_HOOK_TIMEOUT_SECONDS,
+            )
+            if closed:
+                logger.info("Closed %d browser session(s)", closed)
+        except TimeoutError:
+            logger.warning(
+                "Browser session shutdown exceeded %.1fs; proceeding with worker exit.",
+                _SHUTDOWN_HOOK_TIMEOUT_SECONDS,
+            )
+        except Exception:
+            logger.exception("Failed to close browser sessions")
+
         # Drain the memory backend's pending-update buffer before the worker
         # exits (best-effort, bounded). IM channels and the scheduler are
         # already stopped above, so no new IM/scheduler updates arrive during
@@ -317,9 +337,10 @@ async def lifespan(app: FastAPI) -> AsyncGenerator[None, None]:
         # would (review #6 on the original PR).
         #
         # K8s caveat: ``shutdown_flush_timeout_seconds`` must fit inside the
-        # pod's ``terminationGracePeriodSeconds`` (channel stop + this drain +
+        # pod's ``terminationGracePeriodSeconds`` (channel stop + browser
-        # buffer), set on the gateway Helm deployment -- or K8s SIGKILLs the
+        # session close + this drain + buffer), set on the gateway Helm
-        # drain mid-flight and the loss this is fixing is silently re-introduced.
+        # deployment -- or K8s SIGKILLs the drain mid-flight and the loss this
+        # is fixing is silently re-introduced.
         try:
             app_cfg = get_app_config()
             if app_cfg.memory.enabled:
@@ -487,6 +508,9 @@ This gateway provides runtime endpoints for agent runs plus custom endpoints for
     # Artifacts API is mounted at /api/threads/{thread_id}/artifacts
     app.include_router(artifacts.router)
 
+    # Browser API is mounted at /api/threads/{thread_id}/browser
+    app.include_router(browser.router)
+
     # Uploads API is mounted at /api/threads/{thread_id}/uploads
     app.include_router(uploads.router)
 

backend/app/gateway/browser_capability.py

@@ -0,0 +1,72 @@
+"""Shared browser-control capability checks for Gateway surfaces."""
+
+from __future__ import annotations
+
+import importlib.util
+from dataclasses import dataclass
+from typing import Any
+
+from deerflow.community.browser_automation.session import browser_multi_worker_error
+from deerflow.config.app_config import AppConfig
+
+
+@dataclass(frozen=True)
+class BrowserCapability:
+    """Frontend/API availability for the agentic browser control surface."""
+
+    configured: bool
+    available: bool
+    reason: str | None = None
+
+
+def _tool_config(config: AppConfig) -> Any | None:
+    get_tool_config = getattr(config, "get_tool_config", None)
+    if callable(get_tool_config) and callable(getattr(type(config), "get_tool_config", None)):
+        return get_tool_config("browser_navigate")
+    return next(
+        (tool for tool in (getattr(config, "tools", None) or []) if getattr(tool, "name", None) == "browser_navigate"),
+        None,
+    )
+
+
+def _tool_extra(tool_cfg: Any) -> dict[str, Any]:
+    extra = getattr(tool_cfg, "model_extra", None)
+    return extra if isinstance(extra, dict) else {}
+
+
+def browser_capability(config: AppConfig) -> BrowserCapability:
+    """Return whether browser control can actually serve frontend requests."""
+
+    tool_cfg = _tool_config(config)
+    if tool_cfg is None:
+        return BrowserCapability(configured=False, available=False, reason="browser_navigate is not configured")
+
+    worker_error = browser_multi_worker_error()
+    if worker_error is not None:
+        return BrowserCapability(configured=True, available=False, reason=worker_error)
+
+    extra = _tool_extra(tool_cfg)
+    cdp_url = extra.get("cdp_url")
+    if isinstance(cdp_url, str) and cdp_url.strip() and extra.get("allow_unguarded_cdp") is not True:
+        return BrowserCapability(
+            configured=True,
+            available=False,
+            reason="cdp_url requires allow_unguarded_cdp: true because DeerFlow cannot enforce the SSRF request guard on a CDP-attached browser",
+        )
+
+    if importlib.util.find_spec("playwright") is None or importlib.util.find_spec("playwright.async_api") is None:
+        return BrowserCapability(
+            configured=True,
+            available=False,
+            reason="Playwright is not installed; install the backend browser extra and run `playwright install chromium`",
+        )
+
+    return BrowserCapability(configured=True, available=True)
+
+
+def ensure_browser_runtime_available(config: AppConfig) -> None:
+    """Fail startup when browser control is configured but cannot run."""
+
+    capability = browser_capability(config)
+    if capability.configured and not capability.available:
+        raise RuntimeError(capability.reason or "Browser automation is not available")

backend/app/gateway/deps.py

@@ -27,6 +27,7 @@ from typing import TYPE_CHECKING, TypeVar, cast
 from fastapi import FastAPI, HTTPException, Request
 from langgraph.types import Checkpointer
 
+from deerflow.community.browser_automation.session import browser_multi_worker_error
 from deerflow.config.app_config import AppConfig, get_app_config
 from deerflow.persistence.feedback import FeedbackRepository
 from deerflow.runtime import RunContext, RunManager, StreamBridge
@@ -46,14 +47,25 @@ logger = logging.getLogger(__name__)
 _RUN_DRAIN_TIMEOUT_SECONDS = 5.0
 
 
+def _browser_tools_enabled_in_config(config: AppConfig) -> bool:
+    """Return whether process-local agentic browser sessions are configured."""
+    get_tool_config = getattr(config, "get_tool_config", None)
+    if callable(get_tool_config):
+        return get_tool_config("browser_navigate") is not None
+    return any(getattr(tool, "name", None) == "browser_navigate" for tool in (getattr(config, "tools", None) or []))
+
+
 def _enforce_postgres_for_multi_worker(config: AppConfig) -> None:
-    """Refuse to start when GATEWAY_WORKERS > 1 and safety preconditions are not met.
+    """Refuse unsafe multi-worker configurations before persistence starts.
 
-    Two checks (both must pass for multi-worker):
+    Three checks (all must pass for multi-worker):
 
-    1. The DB backend must be Postgres — SQLite write-locks cannot support
+    1. Process-local browser sessions must be disabled. Browser tools keep
+       Chromium and Playwright objects in one worker's memory, while ordinary
+       uvicorn dispatch provides no thread-id affinity.
+    2. The DB backend must be Postgres — SQLite write-locks cannot support
        concurrent multi-process access.
-    2. ``run_ownership.heartbeat_enabled`` must be True — without heartbeat,
+    3. ``run_ownership.heartbeat_enabled`` must be True — without heartbeat,
        every run has a NULL lease, so reconciliation treats all inflight
        runs as orphans and Worker B would kill Worker A's live runs on
        every rolling update or scale-up.
@@ -70,6 +82,9 @@ def _enforce_postgres_for_multi_worker(config: AppConfig) -> None:
     if workers <= 1:
         return
 
+    if _browser_tools_enabled_in_config(config):
+        raise SystemExit(browser_multi_worker_error(workers))
+
     backend = getattr(config.database, "backend", None)
     if backend != "postgres":
         raise SystemExit(f"GATEWAY_WORKERS={workers} requires database.backend='postgres', but database.backend is '{backend}'. SQLite cannot support concurrent multi-process access. Set GATEWAY_WORKERS=1 or switch to Postgres.")

backend/app/gateway/routers/__init__.py

@@ -1,6 +1,7 @@
 from . import (
     artifacts,
     assistants_compat,
+    browser,
     input_polish,
     mcp,
     models,
@@ -15,6 +16,7 @@ from . import (
 __all__ = [
     "artifacts",
     "assistants_compat",
+    "browser",
     "input_polish",
     "mcp",
     "models",

backend/app/gateway/routers/browser.py

@@ -0,0 +1,467 @@
+import asyncio
+import contextlib
+import json
+import logging
+
+from fastapi import APIRouter, HTTPException, Request, WebSocket, WebSocketDisconnect
+from pydantic import BaseModel, Field
+
+from app.gateway.authz import require_permission
+from app.gateway.browser_capability import browser_capability
+from deerflow.config.paths import get_paths
+from deerflow.runtime.user_context import get_effective_user_id, reset_current_user, set_current_user
+
+logger = logging.getLogger(__name__)
+
+router = APIRouter(prefix="/api", tags=["browser"])
+
+
+class BrowserNavigateRequest(BaseModel):
+    url: str = Field(..., description="The http(s) URL to open in the thread's live browser session")
+
+
+class BrowserNavigateResponse(BaseModel):
+    screenshot: str | None = Field(default=None, description="Virtual artifact path of the captured screenshot")
+    url: str = Field(..., description="The resolved URL after navigation")
+    title: str = Field(default="", description="The page title after navigation")
+
+
+def _normalize_browser_seed_url(url: str | None) -> str:
+    return (url or "").split("#", 1)[0].rstrip("/")
+
+
+def _should_apply_browser_seed(current: str | None, seed: str | None) -> bool:
+    if not seed:
+        return False
+    if not current or current == "about:blank":
+        return True
+    return _normalize_browser_seed_url(current) != _normalize_browser_seed_url(seed)
+
+
+def _browser_tools_enabled() -> bool:
+    """Whether the browser tools are turned on in config.
+
+    The live browser HTTP/WS endpoints are an opt-in surface: they must only be
+    reachable when the operator has enabled the ``browser_navigate`` tool in
+    ``config.yaml``. Merely having Playwright importable (it may be preinstalled
+    in a base image) is not sufficient — otherwise the endpoints would expose
+    server-side browser control the operator never turned on.
+    """
+    from deerflow.config import get_app_config
+
+    with contextlib.suppress(Exception):
+        capability = browser_capability(get_app_config())
+        if capability.configured and not capability.available and capability.reason:
+            logger.error("browser automation disabled: %s", capability.reason)
+        return capability.available
+    return False
+
+
+async def _browser_thread_owned_by(thread_store, thread_id: str, user_id: str) -> bool:
+    """Require an explicit owner for browser control's high-trust surface.
+
+    General thread reads retain compatibility with legacy ``user_id=NULL``
+    rows. A retained browser may contain authenticated cookies and page data,
+    so REST navigation and the Live socket deliberately use a stricter policy:
+    only an existing row whose owner exactly matches may drive it.
+    """
+    record = await thread_store.get(thread_id, user_id=user_id)
+    return record is not None and record.get("user_id") == user_id
+
+
+@router.post(
+    "/threads/{thread_id}/browser/navigate",
+    response_model=BrowserNavigateResponse,
+    summary="Navigate The Live Browser Session",
+    description="Steer the thread's live browser session to a URL from the UI and capture a screenshot.",
+)
+@require_permission("threads", "write", owner_check=True, require_existing=True)
+async def navigate_browser(thread_id: str, body: BrowserNavigateRequest, request: Request) -> BrowserNavigateResponse:
+    user_id = str(request.state.auth.user.id)
+    thread_store = getattr(request.app.state, "thread_store", None)
+    if thread_store is None or not await _browser_thread_owned_by(thread_store, thread_id, user_id):
+        raise HTTPException(status_code=404, detail=f"Thread {thread_id} not found")
+
+    if not _browser_tools_enabled():
+        raise HTTPException(status_code=404, detail="Browser automation is not enabled")
+
+    try:
+        from deerflow.community.browser_automation import navigate_and_capture, redact_browser_url
+    except ImportError as exc:  # Playwright is an optional dependency.
+        raise HTTPException(status_code=501, detail="Browser automation is not available") from exc
+
+    url = body.url.strip()
+    if not url:
+        raise HTTPException(status_code=400, detail="URL is required")
+
+    outputs_path = get_paths().sandbox_outputs_dir(thread_id, user_id=get_effective_user_id())
+    try:
+        result = await navigate_and_capture(thread_id=thread_id, url=url, outputs_path=outputs_path)
+    except ValueError as exc:
+        # SSRF / URL validation failure.
+        raise HTTPException(status_code=400, detail=str(exc)) from exc
+    except Exception as exc:
+        logger.error(
+            "Browser navigate failed: thread_id=%s url=%s err_type=%s",
+            thread_id,
+            redact_browser_url(url),
+            type(exc).__name__,
+        )
+        raise HTTPException(status_code=502, detail="Browser navigation failed") from exc
+
+    return BrowserNavigateResponse(**result)
+
+
+async def _authenticate_ws(websocket: WebSocket):
+    """Resolve the user for a WebSocket, honoring auth-disabled mode.
+
+    WebSocket upgrades bypass ``AuthMiddleware`` (a BaseHTTPMiddleware), so we
+    replicate the minimal cookie → user resolution here. Returns the user or
+    ``None`` when unauthenticated.
+    """
+    from app.gateway.auth import decode_token
+    from app.gateway.auth.errors import TokenError
+    from app.gateway.auth_disabled import get_auth_disabled_user, is_auth_disabled
+    from app.gateway.deps import get_local_provider
+
+    access_token = websocket.cookies.get("access_token")
+    if access_token:
+        payload = decode_token(access_token)
+        if not isinstance(payload, TokenError):
+            provider = get_local_provider()
+            user = await provider.get_user(payload.sub)
+            if user is not None and user.token_version == payload.ver:
+                return user
+    if is_auth_disabled():
+        return get_auth_disabled_user()
+    return None
+
+
+def _ws_origin_allowed(websocket: WebSocket) -> bool:
+    """Reject cross-origin WebSocket upgrades (WS-CSRF defense).
+
+    WS upgrades bypass ``CSRFMiddleware`` (also a BaseHTTPMiddleware), so a
+    cross-origin page could otherwise open this socket riding the victim's
+    cookie and both observe frames and drive their authenticated browser. The
+    ``Origin`` header is browser-controlled but always sent on cross-origin
+    upgrades, so validating it is a standard, cheap mitigation.
+
+    Allow when: no ``Origin`` (non-browser clients such as native ws/tests do
+    not send it), the origin is an explicitly configured CORS origin, or it is
+    same-origin with the upgrade target's host. The WS scheme (ws/wss) differs
+    from the page scheme (http/https), so same-origin compares host[:port].
+    """
+    from app.gateway.csrf_middleware import (
+        _first_header_value,
+        _normalize_origin,
+        get_configured_cors_origins,
+    )
+
+    origin = websocket.headers.get("origin")
+    if not origin:
+        return True
+
+    normalized = _normalize_origin(origin)
+    if normalized is None:
+        return False
+    if normalized in get_configured_cors_origins():
+        return True
+
+    target_host = _first_header_value(websocket.headers.get("x-forwarded-host")) or websocket.headers.get("host")
+    if target_host:
+        normalized_host = normalized.split("://", 1)[-1]
+        if normalized_host == target_host.strip().lower():
+            return True
+    return False
+
+
+@router.websocket("/threads/{thread_id}/browser/stream")
+async def browser_stream(websocket: WebSocket, thread_id: str) -> None:
+    """Bidirectional live browser stream.
+
+    Server → client: JSON ``{"type":"frame","data":"<base64 jpeg>"}`` frames
+    captured via CDP screencast. Client → server: input events (click, move,
+    down, up, wheel, key, text, navigate) that drive the live page.
+    """
+    user = await _authenticate_ws(websocket)
+    if user is None:
+        await websocket.close(code=4401)
+        return
+
+    if not _ws_origin_allowed(websocket):
+        # Cross-origin upgrade — reject before touching any session (WS-CSRF).
+        await websocket.close(code=4403)
+        return
+
+    thread_store = getattr(websocket.app.state, "thread_store", None)
+    if thread_store is None:
+        # Fail closed: the live stream drives a real browser (cookies,
+        # logged-in pages), so if the ownership store can't be resolved we must
+        # deny rather than let any authenticated caller attach to any thread's
+        # retained session.
+        await websocket.close(code=4404)
+        return
+    # Browser control is stricter than ordinary legacy-thread access: NULL-owner
+    # rows are not shared because a retained page may expose cookies or account
+    # data from a previous authenticated browser session.
+    if not await _browser_thread_owned_by(thread_store, thread_id, str(user.id)):
+        await websocket.close(code=4404)
+        return
+
+    if not _browser_tools_enabled():
+        await websocket.close(code=4404)
+        return
+
+    try:
+        from deerflow.community.browser_automation import (
+            BrowserLiveViewerError,
+            BrowserSessionCapacityError,
+            get_browser_session_manager,
+            validate_browser_url,
+        )
+    except ImportError:
+        await websocket.close(code=4501)
+        return
+
+    await websocket.accept()
+
+    token = set_current_user(user)
+    loop = asyncio.get_running_loop()
+    frame_queue: asyncio.Queue[str] = asyncio.Queue(maxsize=4)
+    send_lock = asyncio.Lock()
+    input_event = asyncio.Event()
+    input_queue: asyncio.Queue[dict] = asyncio.Queue(maxsize=64)
+    pending_move: dict | None = None
+    pending_wheel: dict | None = None
+
+    async def _send_payload(payload: dict) -> None:
+        async with send_lock:
+            await websocket.send_text(json.dumps(payload))
+
+    def _on_frame(data: str) -> None:
+        # Invoked on the private Playwright loop; hop to this loop and drop the
+        # oldest frame when the client can't keep up (screencast is lossy).
+        def _enqueue() -> None:
+            if frame_queue.full():
+                with contextlib.suppress(asyncio.QueueEmpty):
+                    frame_queue.get_nowait()
+            with contextlib.suppress(asyncio.QueueFull):
+                frame_queue.put_nowait(data)
+
+        loop.call_soon_threadsafe(_enqueue)
+
+    # Match the tool's session config (headless/viewport/cdp_url) so the live
+    # stream reuses the same session the agent drives — including CDP-attach to
+    # the user's real Chrome when configured.
+    from deerflow.config import get_app_config
+
+    tool_cfg = get_app_config().get_tool_config("browser_navigate")
+    extra = (tool_cfg.model_extra or {}) if tool_cfg is not None else {}
+
+    def _cfg_int(key: str, default: int) -> int:
+        value = extra.get(key)
+        return value if isinstance(value, int) and not isinstance(value, bool) else default
+
+    def _cfg_bool(key: str, default: bool) -> bool:
+        value = extra.get(key)
+        return value if isinstance(value, bool) else default
+
+    def _cfg_str(key: str) -> str | None:
+        value = extra.get(key)
+        return value.strip() or None if isinstance(value, str) else None
+
+    manager = get_browser_session_manager()
+    try:
+        session_lease = manager.acquire_session(
+            thread_id,
+            headless=_cfg_bool("headless", True),
+            timeout_ms=_cfg_int("timeout_ms", 30000),
+            viewport={"width": _cfg_int("viewport_width", 1280), "height": _cfg_int("viewport_height", 720)},
+            cdp_url=_cfg_str("cdp_url"),
+            allow_unguarded_cdp=_cfg_bool("allow_unguarded_cdp", False),
+            url_guard=validate_browser_url,
+        )
+        session = session_lease.__enter__()
+    except BrowserSessionCapacityError:
+        await websocket.close(code=4429)
+        reset_current_user(token)
+        return
+    except Exception:
+        await websocket.close(code=4501)
+        reset_current_user(token)
+        return
+
+    async def _pump_frames() -> None:
+        while True:
+            data = await frame_queue.get()
+            await _send_payload({"type": "frame", "data": data})
+
+    async def _send_url() -> None:
+        # Report the page's real URL so the client's address bar reflects the
+        # actual location after navigations, redirects, and history moves — not
+        # the optimistic value the user typed.
+        with contextlib.suppress(Exception):
+            url = await session.current_url()
+            if url:
+                await _send_payload({"type": "url", "url": url})
+
+    async def _send_tabs() -> None:
+        with contextlib.suppress(Exception):
+            tabs = await session.tabs()
+            await _send_payload(
+                {
+                    "type": "tabs",
+                    "tabs": [
+                        {
+                            "index": tab.index,
+                            "title": tab.title,
+                            "url": tab.url,
+                            "active": tab.active,
+                        }
+                        for tab in tabs
+                    ],
+                },
+            )
+
+    async def _poll_location() -> None:
+        # The agent drives the same session through its tools (browser_navigate /
+        # click / type), which do not flow through this socket's input handler, so
+        # those location changes would otherwise never reach the address bar/tabs.
+        # Tool actions push their own settled Live frame after the inline
+        # screenshot is captured; this poll only keeps URL/tabs metadata in sync.
+        # Avoid screenshotting here, because opening a Live panel already primes a
+        # frame and GitHub-style SPAs can generate many URL/render transitions.
+        last_url: str | None = None
+        while True:
+            await asyncio.sleep(1.0)
+            with contextlib.suppress(Exception):
+                url = await session.current_url()
+                if url and url != last_url:
+                    last_url = url
+                    await _send_payload({"type": "url", "url": url})
+                    await _send_tabs()
+
+    def _queue_input(event: dict) -> None:
+        nonlocal pending_move, pending_wheel
+        etype = event.get("type")
+        if etype == "move":
+            pending_move = event
+        elif etype == "wheel":
+            if pending_wheel is None:
+                pending_wheel = event
+            else:
+                pending_wheel = {
+                    **event,
+                    "dx": float(pending_wheel.get("dx", 0)) + float(event.get("dx", 0)),
+                    "dy": float(pending_wheel.get("dy", 0)) + float(event.get("dy", 0)),
+                }
+        else:
+            pending_move = None
+            pending_wheel = None
+            if input_queue.full():
+                with contextlib.suppress(asyncio.QueueEmpty):
+                    input_queue.get_nowait()
+            with contextlib.suppress(asyncio.QueueFull):
+                input_queue.put_nowait(event)
+        input_event.set()
+
+    def _has_pending_input() -> bool:
+        return pending_move is not None or pending_wheel is not None or not input_queue.empty()
+
+    def _take_input() -> dict | None:
+        nonlocal pending_move, pending_wheel
+        if not input_queue.empty():
+            return input_queue.get_nowait()
+        if pending_wheel is not None:
+            event = pending_wheel
+            pending_wheel = None
+            return event
+        if pending_move is not None:
+            event = pending_move
+            pending_move = None
+            return event
+        return None
+
+    async def _read_inputs() -> None:
+        while True:
+            try:
+                raw = await websocket.receive_text()
+            except (WebSocketDisconnect, RuntimeError):
+                return
+            try:
+                event = json.loads(raw)
+            except (json.JSONDecodeError, TypeError):
+                continue
+            if isinstance(event, dict):
+                _queue_input(event)
+
+    async def _process_inputs() -> None:
+        while True:
+            await input_event.wait()
+            while True:
+                event = _take_input()
+                if event is None:
+                    input_event.clear()
+                    if _has_pending_input():
+                        input_event.set()
+                    break
+                if event.get("type") == "navigate":
+                    # SSRF-screen client-driven navigations with the same policy
+                    # the agent tools enforce; reject rather than dispatch.
+                    url = event.get("url")
+                    reason = validate_browser_url(url) if isinstance(url, str) else "Error: invalid navigation URL"
+                    if reason is not None:
+                        await _send_payload({"type": "nav_rejected", "url": url, "message": reason})
+                        continue
+                try:
+                    await session.dispatch_input(event)
+                except Exception as exc:
+                    logger.warning("browser stream input failed: %s", exc)
+                else:
+                    # A location may have changed — resync the client's URL bar.
+                    if event.get("type") in {"navigate", "back", "forward", "click", "activate_tab"}:
+                        await _send_url()
+                        await _send_tabs()
+
+    pump_task = asyncio.create_task(_pump_frames())
+    input_task: asyncio.Task | None = None
+    reader_task: asyncio.Task | None = None
+    poll_task: asyncio.Task | None = None
+    try:
+        # Seed the live page from the latest browser_view URL. A thread can have
+        # a stale browser session from an earlier panel/live attempt; if that
+        # page differs from the latest visible browser artifact, align Live with
+        # what the user expects instead of requiring an off/on reconnect.
+        seed = websocket.query_params.get("seed")
+        if seed and validate_browser_url(seed) is None:
+            with contextlib.suppress(Exception):
+                current = await session.current_url()
+                if _should_apply_browser_seed(current, seed):
+                    await session.navigate(seed)
+        try:
+            await session.start_screencast(_on_frame)
+        except BrowserLiveViewerError:
+            await websocket.close(code=4409)
+            return
+        await _send_url()
+        await _send_tabs()
+        input_task = asyncio.create_task(_process_inputs())
+        reader_task = asyncio.create_task(_read_inputs())
+        poll_task = asyncio.create_task(_poll_location())
+        await reader_task
+    except WebSocketDisconnect:
+        pass
+    except Exception as exc:
+        logger.exception("browser stream error: thread_id=%s err=%s", thread_id, exc)
+    finally:
+        pump_task.cancel()
+        if input_task is not None:
+            input_task.cancel()
+        if reader_task is not None:
+            reader_task.cancel()
+        if poll_task is not None:
+            poll_task.cancel()
+        with contextlib.suppress(Exception):
+            await session.stop_screencast(_on_frame)
+        session_lease.__exit__(None, None, None)
+        reset_current_user(token)

backend/app/gateway/routers/features.py

@@ -9,6 +9,7 @@ on the next request without a restart (config hot-reload boundary).
 from fastapi import APIRouter, Depends
 from pydantic import BaseModel, Field
 
+from app.gateway.browser_capability import browser_capability
 from app.gateway.deps import get_config
 from deerflow.config.app_config import AppConfig
 
@@ -21,10 +22,17 @@ class AgentsApiFeature(BaseModel):
     enabled: bool = Field(..., description="Whether the agents_api routes are exposed over HTTP")
 
 
+class BrowserControlFeature(BaseModel):
+    """Availability of live agentic browser control."""
+
+    enabled: bool = Field(..., description="Whether the live browser routes and UI are available")
+
+
 class FeaturesResponse(BaseModel):
     """Frontend-facing feature availability flags."""
 
     agents_api: AgentsApiFeature
+    browser_control: BrowserControlFeature
 
 
 @router.get(
@@ -35,6 +43,8 @@ class FeaturesResponse(BaseModel):
 )
 async def list_features(config: AppConfig = Depends(get_config)) -> FeaturesResponse:
     """Return availability of optional, config-gated frontend features."""
+    browser = browser_capability(config)
     return FeaturesResponse(
         agents_api=AgentsApiFeature(enabled=config.agents_api.enabled),
+        browser_control=BrowserControlFeature(enabled=browser.available),
     )

backend/app/gateway/routers/threads.py

@@ -501,6 +501,18 @@ async def delete_thread_data(thread_id: str, request: Request) -> ThreadDeleteRe
     except Exception:
         logger.debug("Could not delete thread_meta for %s (not critical)", sanitize_log_param(thread_id))
 
+    # Tear down any live browser session (best-effort). Sessions are keyed only
+    # by thread_id, so leaving one alive after the owner deletes the thread lets
+    # a later caller who guesses the id reuse the retained page/cookies.
+    try:
+        from deerflow.community.browser_automation import get_browser_session_manager
+
+        await get_browser_session_manager().close_session(thread_id)
+    except ImportError:
+        pass  # Playwright is an optional dependency.
+    except Exception:
+        logger.debug("Could not close browser session for %s (not critical)", sanitize_log_param(thread_id))
+
     return response
 
 

backend/docs/CONFIGURATION.md

@@ -248,7 +248,7 @@ Notes:
 - `enabled: false` keeps background polling off by default.
 - `max_concurrent_runs` is a global cap on active scheduled runs (queued/running run rows); each poll cycle claims only into the remaining budget, so long runs accumulating across cycles cannot exceed it.
 - All scheduler fields are restart-required; edits need a Gateway restart.
-- Multi-worker deployments (`GATEWAY_WORKERS > 1`) must use the Postgres database backend. SQLite silently ignores row-level locks, so multiple workers can double-fire the same task.
+- Multi-worker deployments (`GATEWAY_WORKERS > 1`) must use the Postgres database backend. SQLite silently ignores row-level locks, so multiple workers can double-fire the same task. The process-local agentic browser tool group is incompatible with multiple Gateway workers; keep `GATEWAY_WORKERS=1` while `browser_navigate` is enabled. Browser control also requires the backend `browser` extra (`cd backend && uv sync --extra browser && uv run playwright install chromium`); startup detects enabled browser config and fails fast when Playwright is missing, and `/api/features` reports `browser_control.enabled=false` until the runtime is available.
 - The MVP supports thread reuse and fresh-thread-per-run execution modes.
 - The MVP supports only `once` and `cron`.
 - Manual trigger uses the same scheduled-task resource and run lifecycle.

backend/packages/harness/deerflow/community/browser_automation/__init__.py

@@ -0,0 +1,49 @@
+from .session import (
+    BrowserLiveViewerError,
+    BrowserSession,
+    BrowserSessionCapacityError,
+    BrowserSessionManager,
+    BrowserTab,
+    PageSnapshot,
+    SnapshotElement,
+    browser_multi_worker_error,
+    get_browser_session_manager,
+    redact_browser_url,
+    reset_browser_session_manager,
+)
+from .tools import (
+    browser_back_tool,
+    browser_click_tool,
+    browser_close_tool,
+    browser_get_text_tool,
+    browser_navigate_tool,
+    browser_screenshot_tool,
+    browser_snapshot_tool,
+    browser_type_tool,
+    navigate_and_capture,
+    validate_browser_url,
+)
+
+__all__ = [
+    "BrowserSession",
+    "BrowserSessionCapacityError",
+    "BrowserSessionManager",
+    "BrowserTab",
+    "BrowserLiveViewerError",
+    "PageSnapshot",
+    "SnapshotElement",
+    "browser_multi_worker_error",
+    "browser_back_tool",
+    "browser_click_tool",
+    "browser_close_tool",
+    "browser_get_text_tool",
+    "browser_navigate_tool",
+    "browser_screenshot_tool",
+    "browser_snapshot_tool",
+    "browser_type_tool",
+    "get_browser_session_manager",
+    "navigate_and_capture",
+    "redact_browser_url",
+    "reset_browser_session_manager",
+    "validate_browser_url",
+]

backend/packages/harness/deerflow/community/browser_automation/tools.py

@@ -0,0 +1,448 @@
+"""Agentic browser tools — a stateful navigate → observe → act loop.
+
+Unlike the read-only ``web_fetch`` / ``web_capture`` tools, these keep a live
+per-thread browser session (see :mod:`.session`) so the agent can click, type,
+submit forms, and follow multi-step flows on JavaScript-heavy or authenticated
+pages. Every action returns a fresh page snapshot whose interactive elements are
+addressed by a stable ``[ref]`` index, so the model acts on what it just
+observed instead of guessing selectors.
+
+All URLs are SSRF-screened with the shared :func:`validate_public_http_url`
+helper (opt-out only for intentional internal targets).
+"""
+
+from __future__ import annotations
+
+import asyncio
+import contextlib
+import logging
+import re
+from datetime import UTC, datetime
+from pathlib import Path
+from typing import Annotated
+
+from langchain.tools import InjectedToolCallId, tool
+from langchain_core.messages import ToolMessage
+from langgraph.types import Command
+
+from deerflow.community.url_safety import resolve_host_addresses as _resolve_host_addresses
+from deerflow.community.url_safety import validate_public_http_url
+from deerflow.config import get_app_config
+from deerflow.config.paths import VIRTUAL_PATH_PREFIX
+from deerflow.constants import BROWSER_FRAMES_DIRNAME
+from deerflow.tools.types import Runtime
+
+from .session import BrowserSession, BrowserSessionManager, PageSnapshot, get_browser_session_manager
+
+logger = logging.getLogger(__name__)
+
+_OUTPUTS_VIRTUAL_PREFIX = f"{VIRTUAL_PATH_PREFIX}/outputs"
+# Auto-captured per-step screenshots are live progress feedback (shown in the
+# browser panel + inline thumbnails), not deliverables. Keep them in a hidden
+# subdir so the workspace-changes review does not list them as file changes.
+# The dir name is a shared constant so the scanner's ignore list cannot drift.
+_BROWSER_FRAMES_DIRNAME = BROWSER_FRAMES_DIRNAME
+_FRAMES_VIRTUAL_PREFIX = f"{_OUTPUTS_VIRTUAL_PREFIX}/{_BROWSER_FRAMES_DIRNAME}"
+_SAFE_FILENAME_RE = re.compile(r"[^A-Za-z0-9._-]+")
+
+
+def _get_tool_config(tool_name: str) -> dict:
+    config = get_app_config().get_tool_config(tool_name)
+    if config is None:
+        return {}
+    return config.model_extra or {}
+
+
+def _as_bool(value: object, default: bool) -> bool:
+    if isinstance(value, bool):
+        return value
+    if isinstance(value, str):
+        lowered = value.strip().lower()
+        if lowered in {"1", "true", "yes", "on"}:
+            return True
+        if lowered in {"0", "false", "no", "off"}:
+            return False
+    return default
+
+
+def _as_int(value: object, default: int) -> int:
+    if isinstance(value, int) and not isinstance(value, bool):
+        return value
+    if isinstance(value, str):
+        try:
+            return int(value.strip())
+        except ValueError:
+            return default
+    return default
+
+
+def _thread_id(runtime: Runtime) -> str | None:
+    return runtime.context.get("thread_id") if runtime.context else None
+
+
+def _as_str(value: object) -> str | None:
+    if isinstance(value, str):
+        trimmed = value.strip()
+        return trimmed or None
+    return None
+
+
+class _SessionLease:
+    """Context manager that keeps a process-local browser session pinned."""
+
+    def __init__(self, manager: BrowserSessionManager, thread_id: str | None, session: BrowserSession) -> None:
+        self._manager = manager
+        self._thread_id = thread_id
+        self.session = session
+
+    def __enter__(self) -> BrowserSession:
+        return self.session
+
+    def __exit__(self, exc_type, exc_value, traceback) -> None:
+        self._manager.release_session(self._thread_id, self.session)
+
+
+def _resolve_session(runtime: Runtime, tool_name: str) -> _SessionLease:
+    # Launch config (headless/viewport/timeout/cdp_url) is read from a single
+    # canonical source — always ``browser_navigate`` — regardless of which tool
+    # first creates the session. ``get_session`` caches per thread and ignores
+    # these params for later callers, so keying launch config off the calling
+    # tool made it "first tool to run wins": a ``headless: false`` set only on
+    # ``browser_navigate`` was silently dropped if another tool (or the live WS)
+    # initialized the session first. ``tool_name`` is retained for callers that
+    # read their own non-launch config (e.g. ``browser_get_text``'s max_chars).
+    del tool_name
+    cfg = _get_tool_config("browser_navigate")
+    headless = _as_bool(cfg.get("headless"), True)
+    timeout_ms = _as_int(cfg.get("timeout_ms"), 30000)
+    width = _as_int(cfg.get("viewport_width"), 1280)
+    height = _as_int(cfg.get("viewport_height"), 720)
+    cdp_url = _as_str(cfg.get("cdp_url"))
+    manager = get_browser_session_manager()
+    thread_id = _thread_id(runtime)
+    session = manager.get_session(
+        thread_id,
+        headless=headless,
+        timeout_ms=timeout_ms,
+        viewport={"width": width, "height": height},
+        cdp_url=cdp_url,
+        allow_unguarded_cdp=_as_bool(cfg.get("allow_unguarded_cdp"), False),
+        url_guard=validate_browser_url,
+        pin=True,
+    )
+    return _SessionLease(manager, thread_id, session)
+
+
+def validate_browser_url(url: str, *, tool_name: str = "browser_navigate") -> str | None:
+    """SSRF-screen a browser navigation URL using the tool's config policy.
+
+    Returns an ``"Error: ..."`` string when the URL must be rejected, or ``None``
+    when navigation may proceed. Shared by the agent tools and the Gateway live
+    stream so every path that can steer the browser enforces the same allow/deny
+    policy (``allow_private_addresses`` from the ``browser_navigate`` tool config).
+    """
+    cfg = _get_tool_config(tool_name)
+    allow_private = _as_bool(cfg.get("allow_private_addresses"), False)
+    return validate_public_http_url(
+        url,
+        allow_private_addresses=allow_private,
+        action="browse",
+        resolver=_resolve_host_addresses,
+    )
+
+
+def _validate_url(tool_name: str, url: str) -> str | None:
+    return validate_browser_url(url, tool_name=tool_name)
+
+
+def _snapshot_message(snapshot: PageSnapshot, prefix: str = "") -> str:
+    body = snapshot.render()
+    return f"{prefix}\n\n{body}" if prefix else body
+
+
+def _tool_message(content: str, tool_call_id: str) -> Command:
+    return Command(update={"messages": [ToolMessage(content, tool_call_id=tool_call_id)]})
+
+
+def _step_screenshot_name(action: str) -> str:
+    stamp = datetime.now(UTC).strftime("%Y%m%d-%H%M%S-%f")
+    safe = _SAFE_FILENAME_RE.sub("_", action).strip("._-") or "step"
+    return f"browser-{safe}-{stamp}.png"
+
+
+async def _capture_step_screenshot(runtime: Runtime, session: BrowserSession, action: str) -> str | None:
+    """Best-effort per-action screenshot saved as hidden live-progress feedback.
+
+    Returns the artifact virtual path (under the hidden ``.browser-frames`` dir so
+    it stays out of the workspace-changes review), or ``None`` when outputs are
+    unavailable or capture fails — a failed capture must never break the action.
+    """
+    outputs_path = _thread_outputs_path(runtime)
+    if isinstance(outputs_path, str):
+        return None
+    try:
+        content = await session.screenshot_bytes(full_page=False)
+        name = _step_screenshot_name(action)
+        frames_dir = outputs_path / _BROWSER_FRAMES_DIRNAME
+        final_name = await asyncio.to_thread(_write_screenshot, frames_dir, name, content)
+        with contextlib.suppress(Exception):
+            session.schedule_live_frames()
+        return f"{_FRAMES_VIRTUAL_PREFIX}/{final_name}"
+    except Exception as e:
+        logger.warning(f"browser step screenshot failed: {e}")
+        return None
+
+
+def _snapshot_command(
+    runtime: Runtime,
+    session: BrowserSession,
+    snapshot: PageSnapshot,
+    tool_call_id: str,
+    prefix: str,
+    screenshot_path: str | None,
+) -> Command:
+    """Build a ToolMessage carrying the text snapshot plus an inline screenshot.
+
+    The screenshot rides both as a thread ``artifacts`` entry (so it opens in the
+    artifacts side panel) and on ``ToolMessage.additional_kwargs.browser_view``
+    (so the chat can render an inline thumbnail per browser step).
+    """
+    text = _snapshot_message(snapshot, prefix)
+    additional_kwargs: dict = {}
+    update: dict = {}
+    if screenshot_path:
+        additional_kwargs["browser_view"] = {"screenshot": screenshot_path, "url": snapshot.url, "title": snapshot.title}
+        update["artifacts"] = [screenshot_path]
+    update["messages"] = [ToolMessage(text, tool_call_id=tool_call_id, additional_kwargs=additional_kwargs)]
+    return Command(update=update)
+
+
+async def navigate_and_capture(*, thread_id: str | None, url: str, outputs_path: Path) -> dict:
+    """Drive the per-thread browser session to *url* and capture a screenshot.
+
+    Used by the Gateway browser router so a user can steer the live session from
+    the UI URL bar. Shares the same per-thread session, SSRF policy, and
+    screenshot pipeline as :func:`browser_navigate_tool`.
+
+    Returns ``{"screenshot": virtual_path|None, "url": str, "title": str}``.
+    Raises :class:`ValueError` when the URL fails SSRF validation.
+    """
+    url_error = _validate_url("browser_navigate", url)
+    if url_error:
+        raise ValueError(url_error)
+    cfg = _get_tool_config("browser_navigate")
+    manager = get_browser_session_manager()
+    with manager.acquire_session(
+        thread_id,
+        headless=_as_bool(cfg.get("headless"), True),
+        timeout_ms=_as_int(cfg.get("timeout_ms"), 30000),
+        viewport={"width": _as_int(cfg.get("viewport_width"), 1280), "height": _as_int(cfg.get("viewport_height"), 720)},
+        cdp_url=_as_str(cfg.get("cdp_url")),
+        allow_unguarded_cdp=_as_bool(cfg.get("allow_unguarded_cdp"), False),
+        url_guard=validate_browser_url,
+    ) as session:
+        snapshot = await session.navigate(url)
+        screenshot_path: str | None = None
+        try:
+            content = await session.screenshot_bytes(full_page=False)
+            name = _step_screenshot_name("navigate")
+            frames_dir = outputs_path / _BROWSER_FRAMES_DIRNAME
+            final_name = await asyncio.to_thread(_write_screenshot, frames_dir, name, content)
+            screenshot_path = f"{_FRAMES_VIRTUAL_PREFIX}/{final_name}"
+        except Exception as e:
+            logger.warning(f"browser gateway navigate screenshot failed: {e}")
+    return {"screenshot": screenshot_path, "url": snapshot.url, "title": snapshot.title}
+
+
+@tool("browser_navigate", parse_docstring=True)
+async def browser_navigate_tool(runtime: Runtime, url: str, tool_call_id: Annotated[str, InjectedToolCallId]) -> Command:
+    """Open a URL in a live browser session and return the page's interactive elements.
+
+    Use this to START a browsing flow. Unlike web_fetch (read-only), this keeps a
+    stateful browser so you can then click and type on the page. The result lists
+    interactive elements as ``[ref] role: name`` — use those ``[ref]`` numbers with
+    browser_click and browser_type. The session persists across tool calls for this
+    conversation until browser_close. Every navigate/click/type step is
+    auto-captured as a screenshot the user can see, so you do not need to call
+    browser_screenshot just to show progress.
+    URLs must include the scheme, e.g. https://example.com.
+
+    Args:
+        url: The http(s) URL to open.
+    """
+    try:
+        url_error = _validate_url("browser_navigate", url)
+        if url_error:
+            return _tool_message(url_error, tool_call_id)
+        with _resolve_session(runtime, "browser_navigate") as session:
+            snapshot = await session.navigate(url)
+            screenshot = await _capture_step_screenshot(runtime, session, "navigate")
+            return _snapshot_command(runtime, session, snapshot, tool_call_id, f"Navigated to {url}.", screenshot)
+    except Exception as e:
+        logger.error(f"browser_navigate failed: {e}")
+        return _tool_message(f"Error: browser navigation failed: {e}", tool_call_id)
+
+
+@tool("browser_snapshot")
+async def browser_snapshot_tool(runtime: Runtime, tool_call_id: Annotated[str, InjectedToolCallId]) -> Command:
+    """Re-read the current page's interactive elements without acting. Use this to refresh the [ref] element list after the page changed on its own (e.g. async content loaded) or when you are unsure of the current state."""
+    try:
+        with _resolve_session(runtime, "browser_snapshot") as session:
+            snapshot = await session.snapshot()
+            screenshot = await _capture_step_screenshot(runtime, session, "snapshot")
+            return _snapshot_command(runtime, session, snapshot, tool_call_id, "", screenshot)
+    except Exception as e:
+        logger.error(f"browser_snapshot failed: {e}")
+        return _tool_message(f"Error: browser snapshot failed: {e}", tool_call_id)
+
+
+@tool("browser_click", parse_docstring=True)
+async def browser_click_tool(runtime: Runtime, ref: int, tool_call_id: Annotated[str, InjectedToolCallId]) -> Command:
+    """Click an interactive element by its ``[ref]`` number from the latest snapshot.
+
+    The ref comes from the numbered element list returned by browser_navigate,
+    browser_snapshot, browser_click, or browser_type. Returns the updated page
+    snapshot after the click (with new ``[ref]`` numbers).
+
+    Args:
+        ref: The element reference number to click.
+    """
+    try:
+        with _resolve_session(runtime, "browser_click") as session:
+            snapshot = await session.click(ref)
+            screenshot = await _capture_step_screenshot(runtime, session, "click")
+            return _snapshot_command(runtime, session, snapshot, tool_call_id, f"Clicked element [{ref}].", screenshot)
+    except Exception as e:
+        logger.error(f"browser_click failed: {e}")
+        return _tool_message(f"Error: could not click element [{ref}]: {e}", tool_call_id)
+
+
+@tool("browser_type", parse_docstring=True)
+async def browser_type_tool(
+    runtime: Runtime,
+    ref: int,
+    text: str,
+    tool_call_id: Annotated[str, InjectedToolCallId],
+    submit: bool = False,
+) -> Command:
+    """Type text into an input/textarea element by its ``[ref]`` number.
+
+    Fills the field identified by ref. Set submit=true to press Enter afterward
+    (e.g. to run a search or submit a form). Returns the updated page snapshot.
+
+    Args:
+        ref: The element reference number of the input to type into.
+        text: The text to enter.
+        submit: When true, press Enter after typing to submit.
+    """
+    try:
+        with _resolve_session(runtime, "browser_type") as session:
+            snapshot = await session.type_text(ref, text, submit=submit)
+            action = f"Typed into element [{ref}] and submitted." if submit else f"Typed into element [{ref}]."
+            screenshot = await _capture_step_screenshot(runtime, session, "type")
+            return _snapshot_command(runtime, session, snapshot, tool_call_id, action, screenshot)
+    except Exception as e:
+        logger.error(f"browser_type failed: {e}")
+        return _tool_message(f"Error: could not type into element [{ref}]: {e}", tool_call_id)
+
+
+@tool("browser_get_text")
+async def browser_get_text_tool(runtime: Runtime, tool_call_id: Annotated[str, InjectedToolCallId]) -> Command:
+    """Read the visible text content of the current page. Use this to extract readable text after navigating/interacting, e.g. to quote results or summarize content. Output is truncated for large pages."""
+    try:
+        with _resolve_session(runtime, "browser_get_text") as session:
+            cfg = _get_tool_config("browser_get_text")
+            max_chars = _as_int(cfg.get("max_chars"), 8000)
+            text = await session.get_text(max_chars=max_chars)
+        return _tool_message(text or "(page has no visible text)", tool_call_id)
+    except Exception as e:
+        logger.error(f"browser_get_text failed: {e}")
+        return _tool_message(f"Error: could not read page text: {e}", tool_call_id)
+
+
+@tool("browser_back")
+async def browser_back_tool(runtime: Runtime, tool_call_id: Annotated[str, InjectedToolCallId]) -> Command:
+    """Go back to the previous page in the browser session's history."""
+    try:
+        with _resolve_session(runtime, "browser_back") as session:
+            snapshot = await session.back()
+            screenshot = await _capture_step_screenshot(runtime, session, "back")
+            return _snapshot_command(runtime, session, snapshot, tool_call_id, "Went back.", screenshot)
+    except Exception as e:
+        logger.error(f"browser_back failed: {e}")
+        return _tool_message(f"Error: could not go back: {e}", tool_call_id)
+
+
+def _safe_screenshot_name(filename: str | None) -> str:
+    if filename:
+        stem = Path(filename).stem or "browser-capture"
+    else:
+        stem = f"browser-capture-{datetime.now(UTC).strftime('%Y%m%d-%H%M%S')}"
+    safe = _SAFE_FILENAME_RE.sub("_", stem).strip("._-") or "browser-capture"
+    return f"{safe[:100]}.png"
+
+
+def _thread_outputs_path(runtime: Runtime) -> Path | str:
+    if runtime.state is None:
+        return "Error: Thread runtime state is not available"
+    thread_data = runtime.state.get("thread_data") or {}
+    outputs_path = thread_data.get("outputs_path")
+    if not outputs_path:
+        return "Error: Thread outputs path is not available"
+    return Path(outputs_path)
+
+
+def _write_screenshot(outputs_path: Path, name: str, content: bytes) -> str:
+    outputs_path.mkdir(parents=True, exist_ok=True)
+    (outputs_path / name).write_bytes(content)
+    return name
+
+
+@tool("browser_screenshot", parse_docstring=True)
+async def browser_screenshot_tool(
+    runtime: Runtime,
+    tool_call_id: Annotated[str, InjectedToolCallId],
+    filename: str | None = None,
+    full_page: bool = False,
+) -> Command:
+    """Capture a screenshot of the current browser page and save it as an artifact.
+
+    Use this for visual evidence of the current interactive session state (after
+    clicking/typing), which web_capture cannot provide because it renders a fresh,
+    stateless page load.
+
+    Args:
+        filename: Optional output filename (extension is forced to .png).
+        full_page: Capture the full scrollable page instead of just the viewport.
+    """
+    try:
+        outputs_path = _thread_outputs_path(runtime)
+        if isinstance(outputs_path, str):
+            return _tool_message(outputs_path, tool_call_id)
+        with _resolve_session(runtime, "browser_screenshot") as session:
+            content = await session.screenshot_bytes(full_page=full_page)
+            name = _safe_screenshot_name(filename)
+            final_name = await asyncio.to_thread(_write_screenshot, outputs_path, name, content)
+        virtual_path = f"{_OUTPUTS_VIRTUAL_PREFIX}/{final_name}"
+        return Command(
+            update={
+                "artifacts": [virtual_path],
+                "messages": [ToolMessage(f"Saved browser screenshot: {virtual_path}", tool_call_id=tool_call_id)],
+            }
+        )
+    except Exception as e:
+        logger.error(f"browser_screenshot failed: {e}")
+        return _tool_message(f"Error: could not capture screenshot: {e}", tool_call_id)
+
+
+@tool("browser_close")
+async def browser_close_tool(runtime: Runtime, tool_call_id: Annotated[str, InjectedToolCallId]) -> Command:
+    """Close the current browser session and free its resources. Call this when done with the browsing flow; a later browser_navigate starts a fresh session."""
+    try:
+        manager = get_browser_session_manager()
+        closed = await manager.close_session(_thread_id(runtime))
+        msg = "Browser session closed." if closed else "No active browser session to close."
+        return _tool_message(msg, tool_call_id)
+    except Exception as e:
+        logger.error(f"browser_close failed: {e}")
+        return _tool_message(f"Error: could not close browser session: {e}", tool_call_id)

backend/packages/harness/deerflow/constants.py

@@ -1,3 +1,10 @@
 """Shared runtime protocol constants."""
 
 DEFAULT_SKILLS_CONTAINER_PATH = "/mnt/skills"
+
+# Hidden subdirectory (under a thread's outputs dir) that holds the browser
+# tools' per-step screenshots. These are transient live-progress frames, not
+# deliverables, so the workspace-changes scanner excludes this directory. Both
+# the browser tools (which write here) and the scanner (which ignores it) import
+# this single source of truth so the name cannot drift between them.
+BROWSER_FRAMES_DIRNAME = ".browser-frames"

backend/packages/harness/deerflow/workspace_changes/scanner.py

@@ -6,6 +6,8 @@ import os
 from codecs import BOM_UTF16_BE, BOM_UTF16_LE, getincrementaldecoder
 from pathlib import Path
 
+from deerflow.constants import BROWSER_FRAMES_DIRNAME
+
 from .types import (
     DiffUnavailableReason,
     FileSnapshot,
@@ -21,6 +23,10 @@ EXCLUDED_DIR_NAMES = {
     ".cache",
     ".next",
     ".venv",
+    # Transient per-step browser screenshots: live progress feedback surfaced in
+    # the browser panel + inline thumbnails, not workspace deliverables. Shared
+    # constant with the browser tools so the name cannot drift out of sync.
+    BROWSER_FRAMES_DIRNAME,
     "__pycache__",
     "build",
     "dist",

backend/packages/harness/pyproject.toml

@@ -68,6 +68,10 @@ boxlite = ["boxlite>=0.9.7"]
 # Agent observability (Monocle). Optional so a default install stays free of the
 # OpenTelemetry stack; only pulled in when MONOCLE_TRACING is used.
 monocle = ["monocle_apptrace>=0.8.8"]
+# Agentic browser control (browser_navigate/click/type/... tool group). Optional
+# so the core harness install stays lean; import is lazy inside the private
+# Playwright loop. After install, run `playwright install chromium` once.
+browser = ["playwright>=1.40"]
 
 [build-system]
 requires = ["hatchling"]

backend/pyproject.toml

@@ -29,6 +29,7 @@ postgres = ["deerflow-harness[postgres]"]
 redis = ["deerflow-harness[redis]"]
 discord = ["discord.py>=2.7.0"]
 monocle = ["deerflow-harness[monocle]"]
+browser = ["deerflow-harness[browser]"]
 
 [dependency-groups]
 dev = [

backend/tests/manual_browser_live_check.py

@@ -0,0 +1,171 @@
+"""Live end-to-end verification for the agentic browser tools.
+
+NOT a unit test. Run manually with DEEPSEEK_API_KEY in the environment:
+
+    DEEPSEEK_API_KEY=sk-... PYTHONPATH=. uv run python tests/manual_browser_live_check.py
+
+It:
+  1. serves a tiny local HTML form,
+  2. builds an isolated DeerFlow config (DeepSeek model + browser tool group),
+  3. runs a real agent turn that must navigate, type, submit, and read the result,
+  4. asserts the agent-visible tool trace shows the browser loop actually ran.
+"""
+
+from __future__ import annotations
+
+import os
+import tempfile
+import threading
+from http.server import BaseHTTPRequestHandler, HTTPServer
+from pathlib import Path
+
+FORM_PAGE = """<!doctype html><html><head><title>DeerFlow Browser Test</title></head>
+<body>
+<h1>Sign-in demo</h1>
+<form method="GET" action="/welcome">
+  <input type="text" name="username" placeholder="Username">
+  <button type="submit">Sign in</button>
+</form>
+</body></html>"""
+
+
+def _welcome_page(username: str) -> str:
+    return f"""<!doctype html><html><head><title>Welcome</title></head>
+<body><h1>Welcome, {username}!</h1><p>SECRET-TOKEN-4917</p></body></html>"""
+
+
+class _Handler(BaseHTTPRequestHandler):
+    def log_message(self, *args):  # silence server logs
+        pass
+
+    def do_GET(self):
+        from urllib.parse import parse_qs, urlparse
+
+        parsed = urlparse(self.path)
+        if parsed.path.startswith("/welcome"):
+            qs = parse_qs(parsed.query)
+            username = (qs.get("username") or ["friend"])[0]
+            body = _welcome_page(username)
+        else:
+            body = FORM_PAGE
+        data = body.encode()
+        self.send_response(200)
+        self.send_header("Content-Type", "text/html; charset=utf-8")
+        self.send_header("Content-Length", str(len(data)))
+        self.end_headers()
+        self.wfile.write(data)
+
+
+def _start_server() -> tuple[HTTPServer, int]:
+    server = HTTPServer(("127.0.0.1", 0), _Handler)
+    port = server.server_address[1]
+    thread = threading.Thread(target=server.serve_forever, daemon=True)
+    thread.start()
+    return server, port
+
+
+def _write_config(tmp: Path) -> Path:
+    config = f"""
+config_version: 1
+data_dir: {tmp / "data"}
+
+models:
+  - name: deepseek-chat
+    display_name: DeepSeek Chat
+    use: deerflow.models.patched_deepseek:PatchedChatDeepSeek
+    model: deepseek-chat
+    api_key: $DEEPSEEK_API_KEY
+    timeout: 120.0
+    max_retries: 2
+    max_tokens: 4096
+
+sandbox:
+  use: deerflow.sandbox.local:LocalSandboxProvider
+
+tool_groups:
+  - name: browser
+
+tools:
+  - name: browser_navigate
+    group: browser
+    use: deerflow.community.browser_automation.tools:browser_navigate_tool
+    headless: true
+    allow_private_addresses: true
+  - name: browser_snapshot
+    group: browser
+    use: deerflow.community.browser_automation.tools:browser_snapshot_tool
+  - name: browser_click
+    group: browser
+    use: deerflow.community.browser_automation.tools:browser_click_tool
+  - name: browser_type
+    group: browser
+    use: deerflow.community.browser_automation.tools:browser_type_tool
+  - name: browser_get_text
+    group: browser
+    use: deerflow.community.browser_automation.tools:browser_get_text_tool
+  - name: browser_close
+    group: browser
+    use: deerflow.community.browser_automation.tools:browser_close_tool
+
+memory:
+  enabled: false
+
+title:
+  enabled: false
+"""
+    path = tmp / "config.yaml"
+    path.write_text(config)
+    return path
+
+
+def main() -> int:
+    if not os.environ.get("DEEPSEEK_API_KEY"):
+        print("SKIP: DEEPSEEK_API_KEY not set")
+        return 0
+
+    server, port = _start_server()
+    base = f"http://127.0.0.1:{port}/"
+    tmpdir = Path(tempfile.mkdtemp(prefix="deerflow-browser-live-"))
+    try:
+        from deerflow.client import DeerFlowClient
+
+        config_path = _write_config(tmpdir)
+        # Make config resolution deterministic: get_available_tools() re-resolves
+        # via get_app_config(), which would otherwise pick up a project-root
+        # config.yaml. DEER_FLOW_CONFIG_PATH is resolution priority #2.
+        os.environ["DEER_FLOW_CONFIG_PATH"] = str(config_path)
+        client = DeerFlowClient(config_path=str(config_path))
+
+        prompt = (
+            f"Use the browser tools to complete this task. "
+            f"1) Navigate to {base} 2) type the username 'deerbot' into the username field "
+            f"3) click the Sign in button 4) read the resulting page's text. "
+            f"Then tell me the exact SECRET token shown on the welcome page."
+        )
+
+        tool_calls: list[str] = []
+        final_text = ""
+        for event in client.stream(prompt, thread_id="browser-live-check"):
+            if event.type == "messages-tuple":
+                data = event.data or {}
+                if data.get("type") == "tool":
+                    name = data.get("name") or ""
+                    if name.startswith("browser_"):
+                        tool_calls.append(name)
+                elif data.get("type") == "ai":
+                    final_text += data.get("content") or ""
+
+        print("Browser tool calls observed:", tool_calls)
+        print("Final answer:\n", final_text.strip()[:1000])
+
+        assert "browser_navigate" in tool_calls, "agent never navigated"
+        assert any(t in tool_calls for t in ("browser_type", "browser_click")), "agent never interacted"
+        assert "SECRET-TOKEN-4917" in final_text, "agent did not read the post-submit page content"
+        print("\nLIVE CHECK PASSED")
+        return 0
+    finally:
+        server.shutdown()
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

backend/tests/test_browser_router.py

@@ -0,0 +1,271 @@
+import logging
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+from _router_auth_helpers import make_authed_test_app
+from fastapi import FastAPI
+from fastapi.testclient import TestClient
+from starlette.websockets import WebSocketDisconnect
+
+from app.gateway.auth.models import User
+from app.gateway.routers import browser as browser_router
+from app.gateway.routers.browser import _should_apply_browser_seed, _ws_origin_allowed
+
+
+class _FakeWebSocket:
+    """Minimal stand-in exposing only the headers ``_ws_origin_allowed`` reads."""
+
+    def __init__(self, headers: dict[str, str]):
+        self.headers = {k.lower(): v for k, v in headers.items()}
+
+
+def _user(user_id: str = "browser-user") -> SimpleNamespace:
+    return SimpleNamespace(id=user_id)
+
+
+def _browser_ws_app(thread_store=...):
+    app = FastAPI()
+    app.include_router(browser_router.router)
+    if thread_store is not ...:
+        app.state.thread_store = thread_store
+    return app
+
+
+def _expect_ws_close(app: FastAPI, code: int, *, headers: dict[str, str] | None = None) -> None:
+    with TestClient(app) as client:
+        with pytest.raises(WebSocketDisconnect) as exc_info:
+            with client.websocket_connect("/api/threads/thread-1/browser/stream", headers=headers or {}):
+                pass
+    assert exc_info.value.code == code
+
+
+def test_browser_stream_closes_4401_when_unauthenticated():
+    app = _browser_ws_app()
+    with patch.object(browser_router, "_authenticate_ws", AsyncMock(return_value=None)):
+        _expect_ws_close(app, 4401)
+
+
+def test_browser_stream_closes_4403_for_cross_origin_upgrade():
+    app = _browser_ws_app()
+    with patch.object(browser_router, "_authenticate_ws", AsyncMock(return_value=_user())):
+        _expect_ws_close(app, 4403, headers={"origin": "https://evil.example.com"})
+
+
+def test_browser_stream_closes_4404_when_thread_store_missing():
+    app = _browser_ws_app()
+    with patch.object(browser_router, "_authenticate_ws", AsyncMock(return_value=_user())):
+        _expect_ws_close(app, 4404)
+
+
+def test_browser_stream_rejects_legacy_null_owner_thread():
+    store = MagicMock()
+    store.check_access = AsyncMock(return_value=True)
+    store.get = AsyncMock(return_value={"thread_id": "thread-1", "user_id": None})
+    app = _browser_ws_app(store)
+    with (
+        patch.object(browser_router, "_authenticate_ws", AsyncMock(return_value=_user())),
+        patch.object(browser_router, "_browser_tools_enabled", return_value=False),
+    ):
+        _expect_ws_close(app, 4404)
+    store.get.assert_awaited_once_with("thread-1", user_id="browser-user")
+
+
+def test_browser_stream_closes_4404_when_tools_disabled_for_owned_thread():
+    store = MagicMock()
+    store.check_access = AsyncMock(return_value=True)
+    store.get = AsyncMock(return_value={"thread_id": "thread-1", "user_id": "browser-user"})
+    app = _browser_ws_app(store)
+    with (
+        patch.object(browser_router, "_authenticate_ws", AsyncMock(return_value=_user())),
+        patch.object(browser_router, "_browser_tools_enabled", return_value=False),
+    ):
+        _expect_ws_close(app, 4404)
+
+
+def test_browser_stream_closes_4501_when_browser_runtime_unavailable():
+    store = MagicMock()
+    store.check_access = AsyncMock(return_value=True)
+    store.get = AsyncMock(return_value={"thread_id": "thread-1", "user_id": "browser-user"})
+    app = _browser_ws_app(store)
+    real_import = __import__
+
+    def fail_browser_runtime_import(name, globals=None, locals=None, fromlist=(), level=0):
+        if name == "deerflow.community.browser_automation" and "get_browser_session_manager" in fromlist:
+            raise ImportError("browser runtime unavailable")
+        return real_import(name, globals, locals, fromlist, level)
+
+    with (
+        patch.object(browser_router, "_authenticate_ws", AsyncMock(return_value=_user())),
+        patch.object(browser_router, "_browser_tools_enabled", return_value=True),
+        patch("builtins.__import__", side_effect=fail_browser_runtime_import),
+    ):
+        _expect_ws_close(app, 4501)
+
+
+def test_browser_navigate_rejects_legacy_null_owner_thread():
+    user = User(email="browser@example.com", password_hash="x", system_role="user")
+    app = make_authed_test_app(user_factory=lambda: user)
+    app.include_router(browser_router.router)
+    app.state.thread_store.get = AsyncMock(return_value={"thread_id": "thread-1", "user_id": None})
+
+    with (
+        patch.object(browser_router, "_browser_tools_enabled", return_value=True),
+        patch("deerflow.community.browser_automation.navigate_and_capture", new=AsyncMock()),
+    ):
+        response = TestClient(app).post(
+            "/api/threads/thread-1/browser/navigate",
+            json={"url": "https://example.com"},
+        )
+
+    assert response.status_code == 404
+    app.state.thread_store.get.assert_awaited_once_with("thread-1", user_id=str(user.id))
+
+
+def test_browser_tools_disabled_when_cdp_risk_not_explicitly_accepted():
+    tool_cfg = SimpleNamespace(name="browser_navigate", model_extra={"cdp_url": "http://127.0.0.1:9222"})
+    app_config = SimpleNamespace(tools=[tool_cfg])
+
+    with (
+        patch("deerflow.config.get_app_config", return_value=app_config),
+        patch("app.gateway.browser_capability.browser_multi_worker_error", return_value=None),
+        patch("app.gateway.browser_capability.importlib.util.find_spec", return_value=object()),
+    ):
+        assert browser_router._browser_tools_enabled() is False
+
+
+def test_browser_tools_enabled_when_cdp_risk_explicitly_accepted():
+    tool_cfg = SimpleNamespace(
+        name="browser_navigate",
+        model_extra={"cdp_url": "http://127.0.0.1:9222", "allow_unguarded_cdp": True},
+    )
+    app_config = SimpleNamespace(tools=[tool_cfg])
+
+    with (
+        patch("deerflow.config.get_app_config", return_value=app_config),
+        patch("app.gateway.browser_capability.browser_multi_worker_error", return_value=None),
+        patch("app.gateway.browser_capability.importlib.util.find_spec", return_value=object()),
+    ):
+        assert browser_router._browser_tools_enabled() is True
+
+
+def test_browser_navigate_redacts_failure_url_from_logs_and_response(caplog):
+    user = User(email="browser@example.com", password_hash="x", system_role="user")
+    app = make_authed_test_app(user_factory=lambda: user)
+    app.include_router(browser_router.router)
+    app.state.thread_store.get = AsyncMock(return_value={"thread_id": "thread-1", "user_id": str(user.id)})
+    failing_url = "https://example.com/callback?code=secret#fragment"
+
+    caplog.set_level(logging.ERROR, logger="app.gateway.routers.browser")
+    with (
+        patch.object(browser_router, "_browser_tools_enabled", return_value=True),
+        patch(
+            "deerflow.community.browser_automation.navigate_and_capture",
+            new=AsyncMock(side_effect=RuntimeError(f"timed out opening {failing_url}")),
+        ),
+    ):
+        response = TestClient(app).post(
+            "/api/threads/thread-1/browser/navigate",
+            json={"url": failing_url},
+        )
+
+    assert response.status_code == 502
+    assert response.json() == {"detail": "Browser navigation failed"}
+    assert "https://example.com/callback" in caplog.text
+    assert "code=secret" not in caplog.text
+    assert "fragment" not in caplog.text
+    assert "secret" not in response.text
+
+
+def test_browser_stream_seed_applies_to_blank_page():
+    assert _should_apply_browser_seed("about:blank", "https://github.com/bytedance/deer-flow")
+
+
+def test_browser_stream_seed_applies_when_current_url_differs():
+    assert _should_apply_browser_seed(
+        "https://docs.byteplus.com/en/docs/InfoQuest/What_is_Info_Quest",
+        "https://github.com/bytedance/deer-flow",
+    )
+
+
+def test_browser_stream_seed_ignores_hash_and_trailing_slash_for_same_page():
+    assert not _should_apply_browser_seed(
+        "https://github.com/bytedance/deer-flow/#readme",
+        "https://github.com/bytedance/deer-flow/",
+    )
+
+
+def test_ws_origin_allowed_without_origin_header():
+    # Native ws clients / tests do not send Origin — allow them.
+    assert _ws_origin_allowed(_FakeWebSocket({"host": "app.example.com"})) is True
+
+
+def test_ws_origin_allowed_same_origin_host():
+    # Browser page scheme (https) differs from the ws scheme, so same-origin
+    # compares host[:port] against the upgrade target Host.
+    ws = _FakeWebSocket({"origin": "https://app.example.com", "host": "app.example.com"})
+    assert _ws_origin_allowed(ws) is True
+
+
+def test_ws_origin_allowed_rejects_cross_origin():
+    ws = _FakeWebSocket({"origin": "https://evil.example.com", "host": "app.example.com"})
+    assert _ws_origin_allowed(ws) is False
+
+
+def test_ws_origin_allowed_rejects_malformed_origin():
+    ws = _FakeWebSocket({"origin": "not-a-url", "host": "app.example.com"})
+    assert _ws_origin_allowed(ws) is False
+
+
+def test_ws_origin_allowed_honors_configured_cors_origin(monkeypatch):
+    monkeypatch.setenv("GATEWAY_CORS_ORIGINS", "https://console.example.com")
+    ws = _FakeWebSocket({"origin": "https://console.example.com", "host": "gateway.internal"})
+    assert _ws_origin_allowed(ws) is True
+
+
+def test_ws_origin_allowed_honors_forwarded_host():
+    # Behind a proxy the real upgrade target is X-Forwarded-Host, not Host.
+    ws = _FakeWebSocket(
+        {
+            "origin": "https://app.example.com",
+            "host": "gateway.internal",
+            "x-forwarded-host": "app.example.com",
+        },
+    )
+    assert _ws_origin_allowed(ws) is True
+
+
+def test_browser_frames_dirname_shared_between_tools_and_scanner():
+    """The screenshots dir name must stay identical in the writer and the scanner.
+
+    Both sides import the single ``BROWSER_FRAMES_DIRNAME`` constant; this locks
+    that they resolve to the same value so the workspace-changes ignore cannot
+    silently drift away from where the browser tools write frames.
+    """
+    from deerflow.community.browser_automation import tools as browser_tools
+    from deerflow.constants import BROWSER_FRAMES_DIRNAME
+    from deerflow.workspace_changes.scanner import EXCLUDED_DIR_NAMES
+
+    assert browser_tools._BROWSER_FRAMES_DIRNAME == BROWSER_FRAMES_DIRNAME
+    assert BROWSER_FRAMES_DIRNAME in EXCLUDED_DIR_NAMES
+
+
+def test_validate_browser_url_rejects_private_and_non_http(monkeypatch):
+    """WS seed / navigate events reuse the same SSRF policy as the agent tools.
+
+    With no ``allow_private_addresses`` override the shared validator must reject
+    loopback / metadata / non-http targets, so the live stream cannot be steered
+    at internal infrastructure.
+    """
+    from deerflow.community.browser_automation import tools as browser_tools
+    from deerflow.community.browser_automation import validate_browser_url
+
+    # Isolate from any local config.yaml that may set allow_private_addresses.
+    monkeypatch.setattr(browser_tools, "_get_tool_config", lambda _tool_name: {})
+
+    assert validate_browser_url("http://169.254.169.254/latest/meta-data/") is not None
+    assert validate_browser_url("http://127.0.0.1:8001/") is not None
+    assert validate_browser_url("file:///etc/passwd") is not None
+    assert validate_browser_url("ftp://example.com") is not None
+    # A normal public URL passes (returns None = allowed).
+    assert validate_browser_url("https://github.com/bytedance/deer-flow") is None

backend/tests/test_detect_uv_extras.py

@@ -149,6 +149,22 @@ def test_detect_from_config_redis_via_stream_bridge(tmp_path):
     assert detect.detect_from_config(cfg) == ["redis"]
 
 
+def test_detect_from_config_browser_via_browser_navigate_tool(tmp_path):
+    cfg = tmp_path / "config.yaml"
+    cfg.write_text(
+        "tools:\n  - name: browser_navigate\n    group: browser\n    use: deerflow.community.browser_automation.tools:browser_navigate_tool\n",
+    )
+    assert detect.detect_from_config(cfg) == ["browser"]
+
+
+def test_detect_from_config_ignores_commented_browser_tool(tmp_path):
+    cfg = tmp_path / "config.yaml"
+    cfg.write_text(
+        "tools:\n  # - name: browser_navigate\n  #   group: browser\n  - name: web_fetch\n    group: web\n",
+    )
+    assert detect.detect_from_config(cfg) == []
+
+
 def test_detect_from_config_memory_stream_bridge_returns_no_extras(tmp_path):
     cfg = tmp_path / "config.yaml"
     cfg.write_text("stream_bridge:\n  type: memory\n  queue_maxsize: 256\n")

backend/tests/test_features_router.py

@@ -1,4 +1,5 @@
 from types import SimpleNamespace
+from unittest.mock import patch
 
 from fastapi import FastAPI
 from fastapi.testclient import TestClient
@@ -7,10 +8,17 @@ from app.gateway.deps import get_config
 from app.gateway.routers import features
 
 
-def _app_with_config(*, agents_api_enabled: bool) -> FastAPI:
+def _app_with_config(*, agents_api_enabled: bool, browser_enabled: bool = False, browser_extra: dict | None = None) -> FastAPI:
     app = FastAPI()
     app.include_router(features.router)
-    fake_config = SimpleNamespace(agents_api=SimpleNamespace(enabled=agents_api_enabled))
+    tools = (
+        [
+            SimpleNamespace(name="browser_navigate", model_extra=browser_extra or {}),
+        ]
+        if browser_enabled
+        else []
+    )
+    fake_config = SimpleNamespace(agents_api=SimpleNamespace(enabled=agents_api_enabled), tools=tools)
     app.dependency_overrides[get_config] = lambda: fake_config
     return app
 
@@ -19,11 +27,47 @@ def test_features_reports_agents_api_enabled() -> None:
     with TestClient(_app_with_config(agents_api_enabled=True)) as client:
         response = client.get("/api/features")
     assert response.status_code == 200
-    assert response.json() == {"agents_api": {"enabled": True}}
+    assert response.json() == {"agents_api": {"enabled": True}, "browser_control": {"enabled": False}}
 
 
 def test_features_reports_agents_api_disabled() -> None:
     with TestClient(_app_with_config(agents_api_enabled=False)) as client:
         response = client.get("/api/features")
     assert response.status_code == 200
-    assert response.json() == {"agents_api": {"enabled": False}}
+    assert response.json() == {"agents_api": {"enabled": False}, "browser_control": {"enabled": False}}
+
+
+def test_features_reports_browser_control_enabled_when_configured_and_runtime_available() -> None:
+    with (
+        patch("app.gateway.browser_capability.importlib.util.find_spec", return_value=object()),
+        TestClient(_app_with_config(agents_api_enabled=True, browser_enabled=True)) as client,
+    ):
+        response = client.get("/api/features")
+    assert response.status_code == 200
+    assert response.json()["browser_control"] == {"enabled": True}
+
+
+def test_features_reports_browser_control_disabled_when_runtime_missing() -> None:
+    with (
+        patch("app.gateway.browser_capability.importlib.util.find_spec", return_value=None),
+        TestClient(_app_with_config(agents_api_enabled=True, browser_enabled=True)) as client,
+    ):
+        response = client.get("/api/features")
+    assert response.status_code == 200
+    assert response.json()["browser_control"] == {"enabled": False}
+
+
+def test_features_reports_browser_control_disabled_for_unguarded_cdp() -> None:
+    with (
+        patch("app.gateway.browser_capability.importlib.util.find_spec", return_value=object()),
+        TestClient(
+            _app_with_config(
+                agents_api_enabled=True,
+                browser_enabled=True,
+                browser_extra={"cdp_url": "http://127.0.0.1:9222"},
+            ),
+        ) as client,
+    ):
+        response = client.get("/api/features")
+    assert response.status_code == 200
+    assert response.json()["browser_control"] == {"enabled": False}

backend/tests/test_lead_agent_skills.py

@@ -429,6 +429,47 @@ def test_make_lead_agent_passive_empty_skill_policy_preserves_mcp_and_other_tool
     assert captured_deferred_setups[0].deferred_names == frozenset({"lightrag_query"})
 
 
+def test_default_lead_agent_does_not_apply_installed_skill_allowlists(monkeypatch):
+    """Installed skills are discoverable but not active for ordinary default chat.
+
+    A public skill with ``allowed-tools`` must not globally hide configured
+    tools like ``browser_navigate`` before the user has selected a specific
+    skill-owned workflow.
+    """
+    from unittest.mock import MagicMock
+
+    from deerflow.agents.lead_agent import agent as lead_agent_module
+
+    monkeypatch.setattr(lead_agent_module, "_resolve_model_name", lambda x=None, **kwargs: "default-model")
+    monkeypatch.setattr(lead_agent_module, "create_chat_model", lambda **kwargs: "model")
+    monkeypatch.setattr(lead_agent_module, "build_middlewares", lambda *args, **kwargs: [])
+    monkeypatch.setattr(lead_agent_module, "apply_prompt_template", lambda **kwargs: "mock_prompt")
+    monkeypatch.setattr(lead_agent_module, "create_agent", lambda **kwargs: kwargs)
+    monkeypatch.setattr(
+        lead_agent_module,
+        "_load_enabled_available_skills",
+        lambda available_skills, *, app_config, user_id=None: [_make_skill("skill-reviewer", ["review_skill_package"])],
+    )
+    monkeypatch.setattr(
+        "deerflow.tools.get_available_tools",
+        lambda **kwargs: [NamedTool("bash"), NamedTool("browser_navigate"), NamedTool("review_skill_package")],
+    )
+
+    mock_app_config = MagicMock()
+    mock_app_config.get_model_config.return_value = SimpleNamespace(supports_thinking=False, supports_vision=False)
+    mock_app_config.tool_search.enabled = True
+    mock_app_config.skills.container_path = "/mnt/skills"
+    mock_app_config.skills.deferred_discovery = True
+    monkeypatch.setattr(lead_agent_module, "get_app_config", lambda: mock_app_config)
+
+    agent_kwargs = lead_agent_module.make_lead_agent({"configurable": {}})
+
+    tool_names = [tool.name for tool in agent_kwargs["tools"]]
+    assert "browser_navigate" in tool_names
+    assert "bash" in tool_names
+    assert "describe_skill" in tool_names
+
+
 def test_make_lead_agent_fails_closed_when_skill_policy_load_fails(monkeypatch):
     from unittest.mock import MagicMock
 

backend/tests/test_multi_worker_postgres_gate.py

@@ -19,13 +19,15 @@ import pytest
 from fastapi import FastAPI
 
 from app.gateway.deps import _enforce_postgres_for_multi_worker, langgraph_runtime
+from app.gateway.routers.browser import _browser_tools_enabled
 from deerflow.config.database_config import DatabaseConfig
 from deerflow.config.run_ownership_config import RunOwnershipConfig
 
 
-def _config_with_backend(backend: str, *, heartbeat_enabled: bool | None = None) -> SimpleNamespace:
+def _config_with_backend(backend: str, *, heartbeat_enabled: bool | None = None, browser_enabled: bool = False) -> SimpleNamespace:
     run_ownership = RunOwnershipConfig(heartbeat_enabled=heartbeat_enabled) if heartbeat_enabled is not None else None
-    return SimpleNamespace(database=DatabaseConfig(backend=backend), run_ownership=run_ownership)
+    tools = [SimpleNamespace(name="browser_navigate")] if browser_enabled else []
+    return SimpleNamespace(database=DatabaseConfig(backend=backend), run_ownership=run_ownership, tools=tools)
 
 
 # ---------------------------------------------------------------------------
@@ -52,6 +54,25 @@ def test_gate_allows_multi_worker_with_postgres_and_heartbeat(monkeypatch):
     _enforce_postgres_for_multi_worker(_config_with_backend("postgres", heartbeat_enabled=True))
 
 
+def test_gate_rejects_process_local_browser_with_multi_worker(monkeypatch):
+    monkeypatch.setenv("GATEWAY_WORKERS", "2")
+    with pytest.raises(SystemExit) as exc_info:
+        _enforce_postgres_for_multi_worker(
+            _config_with_backend("postgres", heartbeat_enabled=True, browser_enabled=True),
+        )
+    msg = str(exc_info.value)
+    assert "process-local" in msg
+    assert "GATEWAY_WORKERS=1" in msg
+
+
+def test_runtime_browser_surface_stays_disabled_after_incompatible_hot_reload(monkeypatch):
+    monkeypatch.setenv("GATEWAY_WORKERS", "2")
+    live_config = SimpleNamespace(tools=[SimpleNamespace(name="browser_navigate", model_extra={})])
+
+    with patch("deerflow.config.get_app_config", return_value=live_config):
+        assert _browser_tools_enabled() is False
+
+
 def test_gate_rejects_multi_worker_with_sqlite(monkeypatch):
     monkeypatch.setenv("GATEWAY_WORKERS", "2")
     with pytest.raises(SystemExit) as exc_info:

backend/tests/test_threads_router.py

@@ -1,7 +1,7 @@
 import asyncio
 import re
 from types import SimpleNamespace
-from unittest.mock import patch
+from unittest.mock import AsyncMock, patch
 
 import pytest
 from _router_auth_helpers import make_authed_test_app
@@ -158,6 +158,29 @@ def test_delete_thread_route_cleans_thread_directory(tmp_path):
     assert not thread_dir.exists()
 
 
+def test_delete_thread_route_closes_browser_session(tmp_path):
+    """Deleting a thread tears down its live browser session so a later caller
+    who reuses the id cannot inherit the retained page/cookies."""
+    paths = Paths(tmp_path)
+
+    app = make_authed_test_app()
+    app.include_router(threads.router)
+
+    manager = SimpleNamespace(close_session=AsyncMock(return_value=True))
+    with (
+        patch("app.gateway.routers.threads.get_paths", return_value=paths),
+        patch(
+            "deerflow.community.browser_automation.get_browser_session_manager",
+            return_value=manager,
+        ),
+    ):
+        with TestClient(app) as client:
+            response = client.delete("/api/threads/thread-browser")
+
+    assert response.status_code == 200
+    manager.close_session.assert_awaited_once_with("thread-browser")
+
+
 def test_delete_thread_route_rejects_invalid_thread_id(tmp_path):
     paths = Paths(tmp_path)
 

backend/tests/test_workspace_changes.py

@@ -198,6 +198,20 @@ def test_scan_workspace_roots_skips_excluded_directories(tmp_path):
     assert "/mnt/user-data/workspace/node_modules/ignored.js" not in snapshot.files
 
 
+def test_scan_workspace_roots_skips_browser_frames(tmp_path):
+    roots = _roots(tmp_path)
+    outputs = roots[1].host_path
+    (outputs / "report.md").write_text("keep", encoding="utf-8")
+    frames = outputs / ".browser-frames"
+    frames.mkdir()
+    (frames / "browser-navigate-1.png").write_bytes(b"\x89PNG\r\n\x1a\nshot")
+
+    snapshot = scan_workspace_roots(roots)
+
+    assert "/mnt/user-data/outputs/report.md" in snapshot.files
+    assert "/mnt/user-data/outputs/.browser-frames/browser-navigate-1.png" not in snapshot.files
+
+
 def test_scan_workspace_roots_can_skip_text_loading(tmp_path):
     roots = _roots(tmp_path)
     workspace = roots[0].host_path

backend/uv.lock

@@ -796,6 +796,9 @@ dependencies = [
 ]
 
 [package.optional-dependencies]
+browser = [
+    { name = "deerflow-harness", extra = ["browser"] },
+]
 discord = [
     { name = "discord-py" },
 ]
@@ -826,6 +829,7 @@ dev = [
 requires-dist = [
     { name = "bcrypt", specifier = ">=4.0.0" },
     { name = "deerflow-harness", editable = "packages/harness" },
+    { name = "deerflow-harness", extras = ["browser"], marker = "extra == 'browser'", editable = "packages/harness" },
     { name = "deerflow-harness", extras = ["monocle"], marker = "extra == 'monocle'", editable = "packages/harness" },
     { name = "deerflow-harness", extras = ["postgres"], marker = "extra == 'postgres'", editable = "packages/harness" },
     { name = "deerflow-harness", extras = ["redis"], marker = "extra == 'redis'", editable = "packages/harness" },
@@ -846,7 +850,7 @@ requires-dist = [
     { name = "uvicorn", extras = ["standard"], specifier = ">=0.34.0" },
     { name = "wecom-aibot-python-sdk", specifier = ">=0.1.6" },
 ]
-provides-extras = ["postgres", "redis", "discord", "monocle"]
+provides-extras = ["postgres", "redis", "discord", "monocle", "browser"]
 
 [package.metadata.requires-dev]
 dev = [
@@ -907,6 +911,9 @@ dependencies = [
 boxlite = [
     { name = "boxlite" },
 ]
+browser = [
+    { name = "playwright" },
+]
 monocle = [
     { name = "monocle-apptrace" },
 ]
@@ -965,6 +972,7 @@ requires-dist = [
     { name = "markdownify", specifier = ">=1.2.2" },
     { name = "markitdown", extras = ["all", "xlsx"], specifier = ">=0.0.1a2" },
     { name = "monocle-apptrace", marker = "extra == 'monocle'", specifier = ">=0.8.8" },
+    { name = "playwright", marker = "extra == 'browser'", specifier = ">=1.40" },
     { name = "psycopg", extras = ["binary"], marker = "extra == 'postgres'", specifier = ">=3.3.3" },
     { name = "psycopg-pool", marker = "extra == 'postgres'", specifier = ">=3.3.0" },
     { name = "pydantic", specifier = ">=2.12.5" },
@@ -977,7 +985,7 @@ requires-dist = [
     { name = "textual", marker = "extra == 'tui'", specifier = ">=0.80" },
     { name = "tiktoken", specifier = ">=0.8.0" },
 ]
-provides-extras = ["tui", "groundroute", "ollama", "postgres", "redis", "pymupdf", "boxlite", "monocle"]
+provides-extras = ["tui", "groundroute", "ollama", "postgres", "redis", "pymupdf", "boxlite", "monocle", "browser"]
 
 [[package]]
 name = "defusedxml"
@@ -3220,6 +3228,25 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/81/e6/cd9575ac904136b3cbf7aa7ee819ef86eedb7274e46f230e94ea4342e729/platformdirs-4.10.0-py3-none-any.whl", hash = "sha256:fb516cdb12eb0d857d0cd85a7c57cea4d060bee4578d6cf5a14dfdf8cbf8784a", size = 22743, upload-time = "2026-05-28T03:32:52.175Z" },
 ]
 
+[[package]]
+name = "playwright"
+version = "1.61.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "greenlet" },
+    { name = "pyee" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/44/ee/31e4e0db36588b817a10b299a0285082545fde7d36543c2abe498bb3d61a/playwright-1.61.0-py3-none-macosx_10_13_x86_64.whl", hash = "sha256:ff138c3a604f69911e9d42fd036e55c2a171e5616edf04c1e7f60a2a285540b0", size = 43421877, upload-time = "2026-06-29T10:32:48.428Z" },
+    { url = "https://files.pythonhosted.org/packages/42/35/71395dd3ecc798965be4a3ef8c443217d4abca168e7cb34536304f9489e6/playwright-1.61.0-py3-none-macosx_11_0_arm64.whl", hash = "sha256:009588c2a7e499bc5a8b425b61fa65490968bbda9cd69e0cf2cff10f8304659a", size = 42205016, upload-time = "2026-06-29T10:32:52.104Z" },
+    { url = "https://files.pythonhosted.org/packages/f4/44/323164cf5cd1647bdefce76ffce27651aadb959d089b48f53ea40918276e/playwright-1.61.0-py3-none-macosx_11_0_universal2.whl", hash = "sha256:9f7de4536088d12037c13a52b7ea34b59270b78926bb56935070597ffac6b1af", size = 43421884, upload-time = "2026-06-29T10:32:55.773Z" },
+    { url = "https://files.pythonhosted.org/packages/ab/f8/a35bf179e4ba2522c1893635094a64e407572547bd61528820fc0abc87fe/playwright-1.61.0-py3-none-manylinux1_x86_64.whl", hash = "sha256:54f3b39f6eab832e33458c1dd7da0b5682aedab3b09ae731b5c59fa12fd2024e", size = 47421381, upload-time = "2026-06-29T10:32:59.903Z" },
+    { url = "https://files.pythonhosted.org/packages/b7/eb/e3f922348ec17c315f98c463f72faa1181a1c3de0bfe31a8d2edf6561723/playwright-1.61.0-py3-none-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:93454322ade8c11d5d6c211bfd91bdfb9ffb4810e3e026371bcbc4bec1b7ee4c", size = 47120545, upload-time = "2026-06-29T10:33:03.574Z" },
+    { url = "https://files.pythonhosted.org/packages/c2/a6/5be4e52b40a9c0c8a073e7c5b0785c05cf5a9ea8f8a7b5b260e32d970342/playwright-1.61.0-py3-none-win32.whl", hash = "sha256:372d55a6f1248fa1dd47599686980cb8fb5bbe6fcda59eab793eb657c11d8a9b", size = 37844841, upload-time = "2026-06-29T10:33:07.361Z" },
+    { url = "https://files.pythonhosted.org/packages/6c/fd/2b78036e5fbe9d5f5645bbe08a1eac7160c51243c0093963edbcf67c35d9/playwright-1.61.0-py3-none-win_amd64.whl", hash = "sha256:35c6cc4589a5d00964a59d7b3e59641e0aac0c02f15479a7af77d20f6bc79597", size = 37844846, upload-time = "2026-06-29T10:33:10.637Z" },
+    { url = "https://files.pythonhosted.org/packages/27/0d/1b0f3c4ee4eb0514bc805b5c2f9a223e5b6de4f11a926f5235d51d0fc81b/playwright-1.61.0-py3-none-win_arm64.whl", hash = "sha256:e9fcbffcf557a8620fdedd92491eb59a32d18e23d6f3b4f6214b952be324fe51", size = 33955127, upload-time = "2026-06-29T10:33:14.008Z" },
+]
+
 [[package]]
 name = "pluggy"
 version = "1.6.0"

config.example.yaml

@@ -648,6 +648,7 @@ tool_groups:
   - name: file:read
   - name: file:write
   - name: bash
+  - name: browser
 
 # ============================================================================
 # Tools Configuration
@@ -801,6 +802,65 @@ tools:
   #   # allow_private_addresses: false      # SSRF guard: keep false in production. Set true ONLY to
   #   #                                     # capture internal/private targets (loopback, RFC1918, etc.)
 
+  # Agentic browser control (stateful navigate → observe → click/type loop).
+  # Unlike web_fetch/web_capture (read-only, stateless), these tools keep a live
+  # per-thread Playwright browser so the agent can interact with JavaScript-heavy
+  # pages, fill forms, and follow multi-step flows. Elements are addressed by the
+  # numeric [ref] index returned in each page snapshot.
+  #
+  # Requires the optional dependency, installed once per environment:
+  #   cd backend && uv sync --extra browser && uv run playwright install chromium
+  # make dev / Docker startup auto-detect an active browser_navigate entry and
+  # preserve the browser extra during dependency sync.
+  #
+  # Uncomment the whole block to enable. Launch and URL-policy settings are read
+  # canonically from browser_navigate and shared by every browser tool + Live.
+  # - name: browser_navigate
+  #   group: browser
+  #   use: deerflow.community.browser_automation.tools:browser_navigate_tool
+  #   headless: true                   # Set false only for local, trusted debugging
+  #   timeout_ms: 30000                # Per-action navigation/interaction timeout
+  #   viewport_width: 1280
+  #   viewport_height: 720
+  #   # allow_private_addresses: false # SSRF guard: keep false in production
+  #   # cdp_url: http://127.0.0.1:9222 # Attach to YOUR running Chrome (Codex-style)
+  #   # allow_unguarded_cdp: false      # REQUIRED true with cdp_url: attaching to an
+  #   #                                # existing Chrome prevents DeerFlow from
+  #   #                                # enforcing its subresource/redirect SSRF guard.
+  #   #                                # instead of launching a private headless one.
+  #   #                                # Start Chrome with --remote-debugging-port=9222
+  #   #                                # so you watch the agent drive your real browser
+  #   #                                # with your real login sessions. Local/trusted only;
+  #   #                                # never enable for an untrusted CDP endpoint.
+  #   #                                # Chrome must be version-matched to the bundled
+  #   #                                # Playwright; a much newer Chrome can reject the
+  #   #                                # CDP handshake. If so, use headless: false to launch
+  #   #                                # Playwright's own visible browser instead.
+  #   # Multi-worker note: browser sessions live in one Gateway worker's memory.
+  #   # Keep GATEWAY_WORKERS=1 while this tool group is enabled.
+  # - name: browser_snapshot
+  #   group: browser
+  #   use: deerflow.community.browser_automation.tools:browser_snapshot_tool
+  # - name: browser_click
+  #   group: browser
+  #   use: deerflow.community.browser_automation.tools:browser_click_tool
+  # - name: browser_type
+  #   group: browser
+  #   use: deerflow.community.browser_automation.tools:browser_type_tool
+  # - name: browser_get_text
+  #   group: browser
+  #   use: deerflow.community.browser_automation.tools:browser_get_text_tool
+  #   max_chars: 8000                  # Truncation cap for browser_get_text output
+  # - name: browser_back
+  #   group: browser
+  #   use: deerflow.community.browser_automation.tools:browser_back_tool
+  # - name: browser_screenshot
+  #   group: browser
+  #   use: deerflow.community.browser_automation.tools:browser_screenshot_tool
+  # - name: browser_close
+  #   group: browser
+  #   use: deerflow.community.browser_automation.tools:browser_close_tool
+
   # Web fetch tool (uses Exa)
   # NOTE: Only one web_fetch provider can be active at a time.
   # Comment out the Jina AI web_fetch entry below before enabling this one.

docker/nginx/nginx.conf

@@ -168,6 +168,26 @@ http {
             # Disable response buffering to avoid permission errors
         }
 
+        # Live browser stream is a WebSocket upgrade. It must be matched before
+        # the generic /api/threads regex below (which omits Upgrade/Connection
+        # forwarding and would downgrade it to plain HTTP).
+        location ~ ^/api/threads/[^/]+/browser/stream {
+            proxy_pass http://$gateway_upstream;
+            proxy_http_version 1.1;
+            proxy_set_header Host $http_host;
+            proxy_set_header X-Real-IP $remote_addr;
+            proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+            proxy_set_header X-Forwarded-Proto $forwarded_proto;
+            proxy_set_header Upgrade $http_upgrade;
+            proxy_set_header Connection 'upgrade';
+            proxy_cache_bypass $http_upgrade;
+
+            # Long-lived Live stream: keep the upgraded connection open.
+            proxy_connect_timeout 600s;
+            proxy_send_timeout 600s;
+            proxy_read_timeout 600s;
+        }
+
         # Custom API: Other endpoints under /api/threads
         location ~ ^/api/threads {
             proxy_pass http://$gateway_upstream;

docker/nginx/nginx.local.conf

@@ -171,6 +171,28 @@ http {
             proxy_cache off;
         }
 
+        # Live browser stream is a WebSocket upgrade. It must be matched before
+        # the generic /api/threads regex below (which omits Upgrade/Connection
+        # forwarding and would downgrade it to plain HTTP).
+        location ~ ^/api/threads/[^/]+/browser/stream {
+            proxy_pass http://gateway;
+            proxy_http_version 1.1;
+            proxy_set_header Host $http_host;
+            proxy_set_header X-Real-IP $remote_addr;
+            proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+            proxy_set_header X-Forwarded-Proto $scheme;
+            proxy_set_header Upgrade $http_upgrade;
+            proxy_set_header Connection 'upgrade';
+            proxy_cache_bypass $http_upgrade;
+
+            # Long-lived Live stream: keep the upgraded connection open.
+            proxy_buffering off;
+            proxy_cache off;
+            proxy_connect_timeout 600s;
+            proxy_send_timeout 600s;
+            proxy_read_timeout 600s;
+        }
+
         # Custom API: Other endpoints under /api/threads
         location ~ ^/api/threads {
             proxy_pass http://gateway;

docs/plans/2026-07-10-pluggable-authorization-rfc.md

@@ -529,5 +529,3 @@ A search of `bytedance/deer-flow` confirms **no RBAC implementation exists**; RB
 - **[#3929](https://github.com/bytedance/deer-flow/issues/3929)** (open) - sandbox NodePort->ClusterIP (same author family of security hardening RFCs).
 
 **Net takeaway:** the upstream has spent real effort plumbing identity into the guardrail execution point (#3665, #3839) and has explicitly deferred the RBAC policy brain. The two-layer design in this RFC is the natural next step the prior work points at - not a competing or redundant proposal.
-
-

frontend/AGENTS.md

@@ -93,8 +93,10 @@ Tool-calling AI messages can contain user-visible text as well as `tool_calls`.
 
 - `src/app/workspace/chats/[thread_id]/page.tsx` owns composer busy-state wiring.
 - `src/app/workspace/chats/[thread_id]/page.tsx` owns branch-from-turn submission and navigation; sidecar `MessageList` instances do not receive the branch action.
+- `src/app/workspace/chats/[thread_id]/page.tsx` gates the Workspace Browser trigger and browser right panel on `/api/features -> browser_control.enabled`; default/failed feature discovery hides the browser control so optional backend installs do not show a dead Live socket.
 - `src/app/workspace/chats/[thread_id]/page.tsx` and `src/app/workspace/agents/[agent_name]/chats/[thread_id]/page.tsx` own active-goal display state for their composer overlays.
 - `src/components/workspace/messages/message-list.tsx` owns human-input card answered/latest/pending gating; entry pages only translate a submitted card response into `sendMessage` calls.
+- `src/components/workspace/browser-view/browser-view-panel.tsx` forwards each physical pointer click as one `click` input; do not also emit `down`/`up` for the same gesture because the remote Playwright click would run twice.
 - `src/core/threads/hooks.ts` owns pre-submit upload state and thread submission.
 
 ## Code Style

frontend/src/app/workspace/chats/[thread_id]/page.tsx

@@ -7,6 +7,7 @@ import { toast } from "sonner";
 import { type PromptInputMessage } from "@/components/ai-elements/prompt-input";
 import { SidebarTrigger } from "@/components/ui/sidebar";
 import { ArtifactTrigger } from "@/components/workspace/artifacts";
+import { BrowserTrigger } from "@/components/workspace/browser-view";
 import {
   ChatBox,
   useSpecificChatMode,
@@ -33,6 +34,7 @@ import { TodoList } from "@/components/workspace/todo-list";
 import { TokenUsageIndicator } from "@/components/workspace/token-usage-indicator";
 import { useActiveGoal } from "@/components/workspace/use-active-goal";
 import { Welcome } from "@/components/workspace/welcome";
+import { useBrowserControlEnabled } from "@/core/features";
 import { useI18n } from "@/core/i18n/hooks";
 import {
   buildHumanInputResponseText,
@@ -68,6 +70,7 @@ export default function ChatPage() {
   const [isWelcomeMode, setIsWelcomeMode] = useState(isNewThread);
   const [settings, setSettings] = useThreadSettings(threadId);
   const [localSettings, setLocalSettings] = useLocalSettings();
+  const { enabled: browserControlEnabled } = useBrowserControlEnabled();
   const { tokenUsageEnabled } = useModels();
   const threadTokenUsage = useThreadTokenUsage(
     isNewThread || isMock ? undefined : threadId,
@@ -240,6 +243,7 @@ export default function ChatPage() {
     ? localSettings.tokenUsage.inlineMode
     : "off";
   const hasTodos = (thread.values.todos?.length ?? 0) > 0;
+  const browserEnabled = !isNewThread && browserControlEnabled;
   const { activeGoal, hasGoal, setLocalGoal } = useActiveGoal(
     threadId,
     thread.values.goal,
@@ -260,7 +264,7 @@ export default function ChatPage() {
         context={settings.context}
         isMock={isMock}
       >
-        <ChatBox threadId={threadId}>
+        <ChatBox threadId={threadId} browserEnabled={browserEnabled}>
           <div className="relative flex size-full min-h-0 justify-between">
             <header
               className={cn(
@@ -290,6 +294,7 @@ export default function ChatPage() {
                   }
                 />
                 <SidecarTrigger />
+                {browserEnabled && <BrowserTrigger />}
                 <ExportTrigger threadId={threadId} />
                 <ArtifactTrigger />
               </div>

frontend/src/app/workspace/chats/[thread_id]/providers.tsx

@@ -2,13 +2,16 @@
 
 import { PromptInputProvider } from "@/components/ai-elements/prompt-input";
 import { ArtifactsProvider } from "@/components/workspace/artifacts";
+import { BrowserViewProvider } from "@/components/workspace/browser-view";
 import { SubtasksProvider } from "@/core/tasks/context";
 
 export function ChatProviders({ children }: { children: React.ReactNode }) {
   return (
     <SubtasksProvider>
       <ArtifactsProvider>
-        <PromptInputProvider>{children}</PromptInputProvider>
+        <BrowserViewProvider>
+          <PromptInputProvider>{children}</PromptInputProvider>
+        </BrowserViewProvider>
       </ArtifactsProvider>
     </SubtasksProvider>
   );

frontend/src/components/workspace/browser-view/api.ts

@@ -0,0 +1,55 @@
+import { throwGatewayApiError } from "@/core/api/errors";
+import { fetch } from "@/core/api/fetcher";
+import { getBackendBaseURL } from "@/core/config";
+
+export interface BrowserNavigateResult {
+  screenshot: string | null;
+  url: string;
+  title: string;
+}
+
+export async function navigateBrowser(
+  threadId: string,
+  url: string,
+): Promise<BrowserNavigateResult> {
+  const response = await fetch(
+    `${getBackendBaseURL()}/api/threads/${encodeURIComponent(threadId)}/browser/navigate`,
+    {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({ url }),
+    },
+  );
+  if (!response.ok) {
+    await throwGatewayApiError(
+      response,
+      `Failed to navigate browser: ${response.statusText}`,
+    );
+  }
+  return response.json();
+}
+
+/**
+ * Build the WebSocket URL for the live browser stream.
+ *
+ * Uses the configured backend base URL when present (split-origin dev/prod),
+ * otherwise falls back to the current same-origin host (nginx proxies the
+ * upgrade in the unified deployment).
+ *
+ * When ``seedUrl`` is provided, the server aligns the live page to it when the
+ * current page is blank or points at a different URL — so reconnecting to a
+ * stale session lands on the page the user expects instead of a white
+ * about:blank screen or a leftover page. The seed is SSRF-screened server-side.
+ */
+export function browserStreamURL(threadId: string, seedUrl?: string): string {
+  const base = getBackendBaseURL();
+  const origin =
+    base && base.length > 0
+      ? base
+      : typeof window !== "undefined"
+        ? window.location.origin
+        : "";
+  const wsOrigin = origin.replace(/^http/i, "ws");
+  const query = seedUrl ? `?seed=${encodeURIComponent(seedUrl)}` : "";
+  return `${wsOrigin}/api/threads/${encodeURIComponent(threadId)}/browser/stream${query}`;
+}

frontend/src/components/workspace/browser-view/browser-trigger.tsx

@@ -0,0 +1,44 @@
+import { MonitorIcon } from "lucide-react";
+
+import { Button } from "@/components/ui/button";
+import { Tooltip } from "@/components/workspace/tooltip";
+import { useI18n } from "@/core/i18n/hooks";
+
+import { useMaybeSidecar } from "../sidecar/context";
+
+import { useMaybeBrowserView } from "./context";
+
+export const BrowserTrigger = () => {
+  const { t } = useI18n();
+  const browserView = useMaybeBrowserView();
+  const sidecar = useMaybeSidecar();
+
+  if (!browserView) {
+    return null;
+  }
+  const browserVisible = browserView.open && !sidecar?.open;
+  const label = browserVisible ? t.common.close : t.common.showBrowser;
+
+  return (
+    <Tooltip content={label}>
+      <Button
+        aria-label={label}
+        className="text-muted-foreground hover:text-foreground"
+        data-testid="browser-trigger"
+        size="icon"
+        type="button"
+        variant={browserVisible ? "secondary" : "ghost"}
+        onClick={() => {
+          if (browserVisible) {
+            browserView.close();
+            return;
+          }
+          sidecar?.close();
+          browserView.openPanel();
+        }}
+      >
+        <MonitorIcon />
+      </Button>
+    </Tooltip>
+  );
+};

frontend/src/components/workspace/browser-view/browser-view-panel.tsx

@@ -0,0 +1,462 @@
+"use client";
+
+import {
+  ArrowLeftIcon,
+  ArrowRightIcon,
+  GlobeIcon,
+  Loader2Icon,
+  MonitorIcon,
+  RadioIcon,
+  XIcon,
+} from "lucide-react";
+import { useCallback, useEffect, useRef, useState } from "react";
+import { toast } from "sonner";
+
+import { ConversationEmptyState } from "@/components/ai-elements/conversation";
+import { Button } from "@/components/ui/button";
+import { Input } from "@/components/ui/input";
+import { resolveArtifactURL } from "@/core/artifacts/utils";
+import { isIMEComposing } from "@/lib/ime";
+import { cn } from "@/lib/utils";
+
+import { navigateBrowser } from "./api";
+import { useMaybeBrowserView } from "./context";
+import { decideBrowserKeyInput } from "./keyboard";
+import { type BrowserInputEvent, useBrowserStream } from "./use-browser-stream";
+
+export function BrowserViewPanel({
+  threadId,
+  className,
+}: {
+  threadId: string;
+  className?: string;
+}) {
+  const browserView = useMaybeBrowserView();
+  const frame = browserView?.latestFrame ?? null;
+  const imageUrl = frame
+    ? resolveArtifactURL(frame.screenshot, threadId)
+    : null;
+
+  const [urlInput, setUrlInput] = useState("");
+  const [navigating, setNavigating] = useState(false);
+  const [live, setLive] = useState(true);
+  const [lastLiveUrl, setLastLiveUrl] = useState<string | null>(null);
+  const [liveFallback, setLiveFallback] = useState<{
+    frameUrl: string;
+    url?: string;
+  } | null>(null);
+
+  const streamSeedUrl = lastLiveUrl ?? liveFallback?.url ?? frame?.url;
+  const handleNavRejected = useCallback(
+    (url: string | undefined, message: string | undefined) => {
+      setNavigating(false);
+      toast.error(
+        message?.replace(/^Error:\s*/i, "") ??
+          `Cannot open ${url ?? "that URL"}`,
+      );
+    },
+    [],
+  );
+  const { status, frameUrl, liveUrl, sendInput } = useBrowserStream(
+    threadId,
+    live,
+    streamSeedUrl,
+    handleNavRejected,
+  );
+  const panelRef = useRef<HTMLDivElement | null>(null);
+  const surfaceRef = useRef<HTMLImageElement | null>(null);
+  const stageRef = useRef<HTMLDivElement | null>(null);
+  const staticScreenshotRef = useRef<string | null>(null);
+  // True while the user is actively editing the URL bar, so background live-URL
+  // reports do not clobber a half-typed address the user has not submitted yet.
+  const urlEditingRef = useRef(false);
+  // Handle for the live-navigate spinner timeout so it can be cleared on
+  // unmount / re-navigation instead of firing setState after teardown.
+  const navSpinnerTimerRef = useRef<number | null>(null);
+
+  useEffect(() => {
+    return () => {
+      if (navSpinnerTimerRef.current !== null) {
+        window.clearTimeout(navSpinnerTimerRef.current);
+      }
+    };
+  }, []);
+
+  useEffect(() => {
+    setLastLiveUrl(null);
+  }, [threadId]);
+
+  useEffect(() => {
+    if (frame?.url && !urlInput && !liveUrl) {
+      setUrlInput(frame.url);
+    }
+  }, [frame?.url, liveUrl, urlInput]);
+
+  // Keep the URL bar in sync with the latest visible browser state while not
+  // live. If the user leaves Live mode, the last streamed frame becomes the
+  // static baseline instead of falling back to an older artifact screenshot.
+  useEffect(() => {
+    if (!live) {
+      const url = liveFallback?.url ?? frame?.url;
+      if (url) {
+        setUrlInput(url);
+      }
+    }
+  }, [frame?.url, live, liveFallback?.url]);
+
+  useEffect(() => {
+    const screenshot = frame?.screenshot ?? null;
+    if (screenshot && screenshot !== staticScreenshotRef.current) {
+      staticScreenshotRef.current = screenshot;
+      setLiveFallback(null);
+    }
+  }, [frame?.screenshot]);
+
+  useEffect(() => {
+    if (frameUrl) {
+      setLiveFallback((prev) => ({
+        frameUrl,
+        url: liveUrl ?? prev?.url ?? (urlInput || frame?.url),
+      }));
+    }
+  }, [frame?.url, frameUrl, liveUrl, urlInput]);
+
+  // In live mode the server reports the page's real URL (after redirects and
+  // history moves). Reconcile the address bar + device-shell label + persisted
+  // frame so the top URL never goes stale after navigation.
+  useEffect(() => {
+    if (live && liveUrl) {
+      setLastLiveUrl(liveUrl);
+      if (!urlEditingRef.current) {
+        setUrlInput(liveUrl === "about:blank" ? "" : liveUrl);
+      }
+      setNavigating(false);
+      setLiveFallback((prev) => (prev ? { ...prev, url: liveUrl } : prev));
+    }
+  }, [liveUrl, live]);
+
+  const handleNavigate = async () => {
+    const target = urlInput.trim();
+    if (!target) {
+      return;
+    }
+    const normalized = /^https?:\/\//i.test(target)
+      ? target
+      : `https://${target}`;
+
+    // In live mode, steer the streamed page directly over the socket. The
+    // screencast streams continuously, so there is no single "done" frame to
+    // key off — show the reload spinner for a brief window so the user gets
+    // clear feedback that the navigation actually took. Also sync the URL
+    // everywhere immediately (URL bar, device-shell label, persisted frame).
+    if (live) {
+      sendInput({ type: "navigate", url: normalized });
+      setUrlInput(normalized);
+      urlEditingRef.current = false;
+      setLiveFallback((prev) => (prev ? { ...prev, url: normalized } : prev));
+      setNavigating(true);
+      if (navSpinnerTimerRef.current !== null) {
+        window.clearTimeout(navSpinnerTimerRef.current);
+      }
+      navSpinnerTimerRef.current = window.setTimeout(() => {
+        navSpinnerTimerRef.current = null;
+        setNavigating(false);
+      }, 1200);
+      return;
+    }
+
+    if (navigating) {
+      return;
+    }
+    setNavigating(true);
+    try {
+      const result = await navigateBrowser(threadId, normalized);
+      if (result.screenshot) {
+        setLiveFallback(null);
+        browserView?.pushFrame({
+          screenshot: result.screenshot,
+          url: result.url,
+          title: result.title,
+        });
+      } else {
+        setUrlInput(result.url);
+        toast.warning("Navigated, but no screenshot could be captured.");
+      }
+      browserView?.openPanel();
+    } catch (error) {
+      toast.error(error instanceof Error ? error.message : String(error));
+    } finally {
+      setNavigating(false);
+    }
+  };
+
+  const normalizedPoint = useCallback(
+    (
+      clientX: number,
+      clientY: number,
+      options?: { fallbackToCenter?: boolean },
+    ): { nx: number; ny: number } | null => {
+      const el = surfaceRef.current;
+      if (!el) {
+        return null;
+      }
+      const rect = el.getBoundingClientRect();
+      const natW = el.naturalWidth;
+      const natH = el.naturalHeight;
+      if (!rect.width || !rect.height || !natW || !natH) {
+        return null;
+      }
+      // The frame is drawn with object-contain: scaled to fit and centered, so
+      // there are letterbox bars whenever the panel and remote viewport differ
+      // in aspect. Map the pointer against the actual content box.
+      const scale = Math.min(rect.width / natW, rect.height / natH);
+      const contentW = natW * scale;
+      const contentH = natH * scale;
+      const offX = (rect.width - contentW) / 2;
+      const offY = (rect.height - contentH) / 2;
+      const px = clientX - rect.left - offX;
+      const py = clientY - rect.top - offY;
+      if (px < 0 || py < 0 || px > contentW || py > contentH) {
+        if (options?.fallbackToCenter) {
+          return { nx: 0.5, ny: 0.5 };
+        }
+        return null;
+      }
+      return { nx: px / contentW, ny: py / contentH };
+    },
+    [],
+  );
+
+  const forwardMouse = (
+    type: Extract<BrowserInputEvent, { nx: number }>["type"],
+    event: React.MouseEvent<HTMLImageElement>,
+  ) => {
+    if (!live) {
+      return;
+    }
+    const point = normalizedPoint(event.clientX, event.clientY);
+    if (point) {
+      sendInput({ type, ...point });
+    }
+  };
+
+  const liveActive = live && status === "open";
+  const liveConnecting = live && status === "connecting";
+  const displayUrl = live
+    ? (frameUrl ?? liveFallback?.frameUrl ?? imageUrl)
+    : (liveFallback?.frameUrl ?? imageUrl);
+
+  // React registers onWheel as a passive listener, so preventDefault() there is
+  // ignored and the wheel scrolls the host chat page. Bind a native, non-passive
+  // listener on the always-mounted stage (the <img> may not exist yet when live
+  // first opens) so scrolling stays captured inside the remote page.
+  useEffect(() => {
+    const el = stageRef.current;
+    if (!el || !liveActive) {
+      return;
+    }
+    let wheelFrame: number | null = null;
+    let pendingDx = 0;
+    let pendingDy = 0;
+    let pendingPoint: { nx: number; ny: number } | null = null;
+
+    const browserWheelDelta = (pixels: number) => {
+      return Math.abs(pixels) < 0.25 ? 0 : pixels * 2;
+    };
+
+    const flushWheel = () => {
+      wheelFrame = null;
+      const dx = browserWheelDelta(pendingDx);
+      const dy = browserWheelDelta(pendingDy);
+      const point = pendingPoint;
+      pendingDx = 0;
+      pendingDy = 0;
+      pendingPoint = null;
+      if (dx || dy) {
+        sendInput({ type: "wheel", dx, dy, ...(point ?? {}) });
+      }
+    };
+
+    const onWheel = (event: WheelEvent) => {
+      event.preventDefault();
+      event.stopPropagation();
+      // Normalize deltaMode (line/page → pixels), then batch one animation
+      // frame of deltas. A small 2x gain keeps touchpad gestures responsive
+      // without returning to the jumpy large-step behavior.
+      const unit = event.deltaMode === 1 ? 16 : event.deltaMode === 2 ? 800 : 1;
+      pendingDx += event.deltaX * unit;
+      pendingDy += event.deltaY * unit;
+      const point = normalizedPoint(event.clientX, event.clientY, {
+        fallbackToCenter: true,
+      });
+      if (point) {
+        pendingPoint = point;
+      }
+      wheelFrame ??= window.requestAnimationFrame(flushWheel);
+    };
+    el.addEventListener("wheel", onWheel, {
+      capture: true,
+      passive: false,
+    });
+    return () => {
+      el.removeEventListener("wheel", onWheel, { capture: true });
+      if (wheelFrame !== null) {
+        window.cancelAnimationFrame(wheelFrame);
+      }
+    };
+  }, [liveActive, normalizedPoint, sendInput]);
+
+  return (
+    <div
+      ref={panelRef}
+      className={cn(
+        "bg-background relative flex flex-col",
+        "size-full",
+        className,
+      )}
+      tabIndex={live ? 0 : undefined}
+      onKeyDown={(event) => {
+        const target = event.target;
+        const editableTarget =
+          target instanceof HTMLInputElement ||
+          target instanceof HTMLTextAreaElement ||
+          (target instanceof HTMLElement && target.isContentEditable);
+        const input = decideBrowserKeyInput({
+          live,
+          editableTarget,
+          composing: isIMEComposing(event),
+          key: event.key,
+          metaKey: event.metaKey,
+          ctrlKey: event.ctrlKey,
+        });
+        if (!input) {
+          return;
+        }
+        sendInput(input);
+        event.preventDefault();
+        event.stopPropagation();
+      }}
+    >
+      <header className="flex shrink-0 items-center gap-2 border-b px-3 py-2">
+        <MonitorIcon className="size-4 shrink-0" />
+        <span className="shrink-0 text-sm font-medium">Browser</span>
+        <div className="flex shrink-0 items-center">
+          <Button
+            size="icon-sm"
+            variant="ghost"
+            className="shrink-0"
+            disabled={!live}
+            onClick={() => sendInput({ type: "back" })}
+            title="Back"
+          >
+            <ArrowLeftIcon />
+          </Button>
+          <Button
+            size="icon-sm"
+            variant="ghost"
+            className="shrink-0"
+            disabled={!live}
+            onClick={() => sendInput({ type: "forward" })}
+            title="Forward"
+          >
+            <ArrowRightIcon />
+          </Button>
+        </div>
+        <form
+          className="relative flex min-w-0 flex-1 items-center"
+          onSubmit={(event) => {
+            event.preventDefault();
+            void handleNavigate();
+          }}
+        >
+          <GlobeIcon className="text-muted-foreground pointer-events-none absolute left-2 size-3.5" />
+          <Input
+            value={urlInput}
+            onChange={(event) => {
+              urlEditingRef.current = true;
+              setUrlInput(event.target.value);
+            }}
+            onFocus={(event) => {
+              urlEditingRef.current = true;
+              event.currentTarget.select();
+            }}
+            onBlur={() => {
+              urlEditingRef.current = false;
+            }}
+            onKeyDown={(event) => {
+              // Keep browser select-all/copy/cut/paste working inside the URL
+              // bar; stop the panel + global shortcut handlers from swallowing.
+              event.stopPropagation();
+            }}
+            placeholder="Enter a URL and press Enter"
+            spellCheck={false}
+            autoComplete="off"
+            className="h-8 pl-7 text-xs"
+          />
+          {navigating && (
+            <Loader2Icon className="text-muted-foreground absolute right-2 size-3.5 animate-spin" />
+          )}
+        </form>
+        <Button
+          size="sm"
+          variant={live ? "default" : "ghost"}
+          className="shrink-0 gap-1"
+          onClick={() => setLive((prev) => !prev)}
+          title={live ? "Stop live control" : "Take live control"}
+        >
+          <RadioIcon className="size-3.5" />
+          {live ? (status === "open" ? "Live" : "…") : "Live"}
+        </Button>
+        <Button
+          size="icon-sm"
+          variant="ghost"
+          className="shrink-0"
+          onClick={() => {
+            setLive(false);
+            browserView?.close();
+          }}
+        >
+          <XIcon />
+        </Button>
+      </header>
+      <main className="relative flex min-h-0 grow flex-col overflow-hidden bg-neutral-950">
+        <div
+          ref={stageRef}
+          className="relative min-h-0 grow bg-neutral-900"
+          onMouseDown={() => {
+            panelRef.current?.focus({ preventScroll: true });
+          }}
+        >
+          {displayUrl ? (
+            <img
+              ref={surfaceRef}
+              className="absolute inset-0 h-full w-full cursor-default object-contain object-center"
+              src={displayUrl}
+              alt={frame?.title ?? "Browser view"}
+              draggable={false}
+              onClick={(event) => forwardMouse("click", event)}
+            />
+          ) : (
+            <ConversationEmptyState
+              className="absolute inset-0 m-auto h-fit"
+              icon={<MonitorIcon />}
+              title={
+                live ? "Connecting to live browser…" : "No browser activity yet"
+              }
+              description={
+                live
+                  ? "Waiting for the first live frame."
+                  : "Enter a URL above or let the agent browse — the live view will appear here."
+              }
+            />
+          )}
+          {(navigating || liveConnecting) && displayUrl && (
+            <div className="absolute inset-0 flex items-center justify-center bg-white/40 backdrop-blur-[1px]">
+              <Loader2Icon className="text-muted-foreground size-8 animate-spin" />
+            </div>
+          )}
+        </div>
+      </main>
+    </div>
+  );
+}

frontend/src/components/workspace/browser-view/context.tsx

@@ -0,0 +1,76 @@
+"use client";
+
+import {
+  createContext,
+  useCallback,
+  useContext,
+  useMemo,
+  useState,
+  type ReactNode,
+} from "react";
+
+export interface BrowserViewFrame {
+  screenshot: string;
+  url?: string;
+  title?: string;
+  action?: string;
+}
+
+interface BrowserViewContextValue {
+  open: boolean;
+  latestFrame: BrowserViewFrame | null;
+  pushFrame: (frame: BrowserViewFrame) => void;
+  openPanel: () => void;
+  close: () => void;
+}
+
+const BrowserViewContext = createContext<BrowserViewContextValue | null>(null);
+
+export function BrowserViewProvider({ children }: { children: ReactNode }) {
+  const [open, setOpen] = useState(false);
+  const [latestFrame, setLatestFrame] = useState<BrowserViewFrame | null>(null);
+
+  const pushFrame = useCallback((frame: BrowserViewFrame) => {
+    setLatestFrame((prev) => {
+      const sameFrame =
+        prev?.screenshot === frame.screenshot &&
+        prev?.url === frame.url &&
+        prev?.title === frame.title &&
+        prev?.action === frame.action;
+      if (sameFrame) {
+        return prev;
+      }
+      if (prev?.screenshot !== frame.screenshot) {
+        // A new browser frame arrived — surface the panel automatically.
+        setOpen(true);
+      }
+      return frame;
+    });
+  }, []);
+
+  const openPanel = useCallback(() => setOpen(true), []);
+  const close = useCallback(() => setOpen(false), []);
+
+  const value = useMemo(
+    () => ({ open, latestFrame, pushFrame, openPanel, close }),
+    [open, latestFrame, pushFrame, openPanel, close],
+  );
+
+  return (
+    <BrowserViewContext.Provider value={value}>
+      {children}
+    </BrowserViewContext.Provider>
+  );
+}
+
+export function useMaybeBrowserView(): BrowserViewContextValue | null {
+  return useContext(BrowserViewContext);
+}
+
+export function useBrowserView(): BrowserViewContextValue {
+  const ctx = useContext(BrowserViewContext);
+  if (!ctx) {
+    throw new Error("useBrowserView must be used within a BrowserViewProvider");
+  }
+  return ctx;
+}

frontend/src/components/workspace/browser-view/index.ts

@@ -0,0 +1,19 @@
+export { BrowserViewPanel } from "./browser-view-panel";
+export { BrowserTrigger } from "./browser-trigger";
+export {
+  navigateBrowser,
+  browserStreamURL,
+  type BrowserNavigateResult,
+} from "./api";
+export {
+  useBrowserStream,
+  type BrowserTab,
+  type BrowserInputEvent,
+  type BrowserStreamStatus,
+} from "./use-browser-stream";
+export {
+  BrowserViewProvider,
+  useBrowserView,
+  useMaybeBrowserView,
+  type BrowserViewFrame,
+} from "./context";

frontend/src/components/workspace/browser-view/keyboard.ts

@@ -0,0 +1,55 @@
+import type { BrowserInputEvent } from "./use-browser-stream";
+
+// Named keys we forward to the remote page as key presses (everything else that
+// is a single printable char is forwarded as text). Chosen to cover editing and
+// navigation without leaking browser-level shortcuts.
+export const FORWARDED_NAMED_KEYS = [
+  "Enter",
+  "Backspace",
+  "Tab",
+  "ArrowUp",
+  "ArrowDown",
+  "ArrowLeft",
+  "ArrowRight",
+  "Escape",
+  "Delete",
+] as const;
+
+export interface BrowserKeyContext {
+  /** Panel is in Live (interactive) mode. */
+  live: boolean;
+  /** Focus is on an editable element (URL bar etc.) — keep it local. */
+  editableTarget: boolean;
+  /** An IME composition is active — the keystroke belongs to the composer. */
+  composing: boolean;
+  key: string;
+  metaKey: boolean;
+  ctrlKey: boolean;
+}
+
+/**
+ * Decide how a keydown maps to a remote browser input, or ``null`` to ignore it.
+ *
+ * Pure so the forwarding policy (including the IME-composition guard that must
+ * swallow a composing Enter) can be unit-tested without a DOM.
+ */
+export function decideBrowserKeyInput(
+  ctx: BrowserKeyContext,
+): BrowserInputEvent | null {
+  if (!ctx.live || ctx.editableTarget || ctx.composing) {
+    return null;
+  }
+  if ((ctx.ctrlKey || ctx.metaKey) && ctx.key.length === 1) {
+    return {
+      type: "key",
+      key: `${ctx.metaKey ? "Meta" : "Control"}+${ctx.key.toUpperCase()}`,
+    };
+  }
+  if (ctx.key.length === 1 && !ctx.metaKey && !ctx.ctrlKey) {
+    return { type: "text", text: ctx.key };
+  }
+  if ((FORWARDED_NAMED_KEYS as readonly string[]).includes(ctx.key)) {
+    return { type: "key", key: ctx.key };
+  }
+  return null;
+}

frontend/src/components/workspace/browser-view/use-browser-stream.ts

@@ -0,0 +1,237 @@
+"use client";
+
+import { useCallback, useEffect, useRef, useState } from "react";
+
+import { browserStreamURL } from "./api";
+
+export interface BrowserTab {
+  index: number;
+  title: string;
+  url: string;
+  active: boolean;
+}
+
+export type BrowserInputEvent =
+  | { type: "click"; nx: number; ny: number }
+  | { type: "move"; nx: number; ny: number }
+  | { type: "down"; nx: number; ny: number }
+  | { type: "up"; nx: number; ny: number }
+  | { type: "wheel"; dx: number; dy: number; nx?: number; ny?: number }
+  | { type: "key"; key: string }
+  | { type: "text"; text: string }
+  | { type: "navigate"; url: string }
+  | { type: "back" }
+  | { type: "forward" }
+  | { type: "activate_tab"; index: number };
+
+export type BrowserStreamStatus = "idle" | "connecting" | "open" | "closed";
+
+const RECONNECT_BASE_DELAY_MS = 800;
+const RECONNECT_MAX_DELAY_MS = 10_000;
+const RECONNECT_MAX_ATTEMPTS = 6;
+
+function normalizeSeedUrl(url: string | null | undefined): string {
+  return (url ?? "").split("#", 1)[0]?.replace(/\/+$/, "") ?? "";
+}
+
+/**
+ * Manage a live browser screencast WebSocket.
+ *
+ * When ``enabled`` is true, opens the stream, exposes the latest JPEG frame as
+ * a data URL, and returns a ``sendInput`` callback that forwards user input to
+ * the live page. Closes and cleans up when disabled or unmounted.
+ *
+ * ``seedUrl`` is only read when a connection is first established (via a ref, so
+ * it is NOT a reconnect trigger). A separate effect steers an already-open live
+ * page toward a changed seed with an in-band ``navigate`` event, so ordinary
+ * navigations no longer tear down and rebuild the socket.
+ */
+export function useBrowserStream(
+  threadId: string,
+  enabled: boolean,
+  seedUrl?: string,
+  onNavRejected?: (
+    url: string | undefined,
+    message: string | undefined,
+  ) => void,
+) {
+  const [status, setStatus] = useState<BrowserStreamStatus>("idle");
+  const [frameUrl, setFrameUrl] = useState<string | null>(null);
+  const [liveUrl, setLiveUrl] = useState<string | null>(null);
+  const [tabs, setTabs] = useState<BrowserTab[]>([]);
+  const [connectionAttempt, setConnectionAttempt] = useState(0);
+  const socketRef = useRef<WebSocket | null>(null);
+  const pendingNavigateRef = useRef<Extract<
+    BrowserInputEvent,
+    { type: "navigate" }
+  > | null>(null);
+  // Read the seed at connect time only; it must not be a reconnect dependency.
+  const seedRef = useRef(seedUrl);
+  seedRef.current = seedUrl;
+  // Latest live page URL reported by the server, used to decide whether an
+  // open stream already shows the seed target (avoids redundant navigations).
+  const liveUrlRef = useRef<string | null>(null);
+  const onNavRejectedRef = useRef(onNavRejected);
+  onNavRejectedRef.current = onNavRejected;
+
+  const sendInput = useCallback((event: BrowserInputEvent) => {
+    const socket = socketRef.current;
+    if (socket?.readyState === WebSocket.OPEN) {
+      socket.send(JSON.stringify(event));
+      return true;
+    }
+    // URL bar submissions are user intent and must not be lost during the
+    // short Live connection window right after opening the panel.
+    if (event.type === "navigate") {
+      pendingNavigateRef.current = event;
+    }
+    return false;
+  }, []);
+
+  useEffect(() => {
+    pendingNavigateRef.current = null;
+  }, [threadId]);
+
+  useEffect(() => {
+    if (enabled) {
+      return;
+    }
+    setConnectionAttempt(0);
+    setFrameUrl(null);
+    setLiveUrl(null);
+    setTabs([]);
+    liveUrlRef.current = null;
+  }, [enabled, threadId]);
+
+  useEffect(() => {
+    if (!enabled) {
+      setStatus("idle");
+      liveUrlRef.current = null;
+      return;
+    }
+
+    let closedByEffect = false;
+    let reconnectTimer: number | null = null;
+    setStatus("connecting");
+    // browserStreamURL treats empty/undefined seed identically (no seed param),
+    // so the raw ref value is fine here. Record the seed optimistically so the
+    // "steer to seed" effect below does not fire a duplicate navigate right
+    // after open (the server already aligns the page to the connect-time seed).
+    liveUrlRef.current = seedRef.current ?? null;
+    const socket = new WebSocket(browserStreamURL(threadId, seedRef.current));
+    socketRef.current = socket;
+
+    const scheduleReconnect = () => {
+      if (closedByEffect || !enabled) {
+        return;
+      }
+      if (reconnectTimer !== null) {
+        return;
+      }
+      // Exponential backoff with a ceiling + attempt cap so a server that keeps
+      // rejecting the upgrade cannot pin the client in a tight reconnect loop.
+      if (connectionAttempt >= RECONNECT_MAX_ATTEMPTS) {
+        return;
+      }
+      const delay = Math.min(
+        RECONNECT_BASE_DELAY_MS * 2 ** connectionAttempt,
+        RECONNECT_MAX_DELAY_MS,
+      );
+      reconnectTimer = window.setTimeout(() => {
+        setConnectionAttempt((attempt) => attempt + 1);
+      }, delay);
+    };
+
+    socket.onopen = () => {
+      const pendingNavigate = pendingNavigateRef.current;
+      if (pendingNavigate) {
+        socket.send(JSON.stringify(pendingNavigate));
+        pendingNavigateRef.current = null;
+      }
+      // Reset the reconnect budget on a successful open. Without this the
+      // cumulative attempt counter never returns to 0 while the panel stays
+      // mounted, so after RECONNECT_MAX_ATTEMPTS total reconnects — even across
+      // many healthy connections — scheduleReconnect would bail forever and
+      // Live would go permanently dead until the panel is toggled off/on.
+      setConnectionAttempt(0);
+      setStatus("open");
+    };
+    socket.onmessage = async (message) => {
+      try {
+        const raw =
+          typeof message.data === "string"
+            ? message.data
+            : message.data instanceof Blob
+              ? await message.data.text()
+              : message.data instanceof ArrayBuffer
+                ? new TextDecoder().decode(message.data)
+                : String(message.data);
+        // The message may resolve after cleanup (async Blob/ArrayBuffer decode);
+        // do not write state for a socket the effect already tore down.
+        if (closedByEffect) {
+          return;
+        }
+        const payload = JSON.parse(raw) as {
+          type?: string;
+          data?: string;
+          url?: string;
+          message?: string;
+          tabs?: BrowserTab[];
+        };
+        if (payload.type === "frame" && payload.data) {
+          setFrameUrl(`data:image/jpeg;base64,${payload.data}`);
+        } else if (payload.type === "url" && payload.url) {
+          liveUrlRef.current = payload.url;
+          setLiveUrl(payload.url);
+        } else if (payload.type === "tabs" && Array.isArray(payload.tabs)) {
+          setTabs(payload.tabs);
+        } else if (payload.type === "nav_rejected") {
+          onNavRejectedRef.current?.(payload.url, payload.message);
+        }
+      } catch (error) {
+        console.warn("Ignoring malformed browser stream message", error);
+      }
+    };
+    socket.onclose = () => {
+      if (!closedByEffect) {
+        setStatus("closed");
+        scheduleReconnect();
+      }
+    };
+    socket.onerror = () => {
+      if (!closedByEffect) {
+        setStatus("closed");
+        scheduleReconnect();
+      }
+    };
+
+    return () => {
+      closedByEffect = true;
+      if (reconnectTimer !== null) {
+        window.clearTimeout(reconnectTimer);
+      }
+      socketRef.current = null;
+      socket.close();
+    };
+  }, [connectionAttempt, enabled, threadId]);
+
+  // Steer an already-open stream toward a changed seed in-band instead of
+  // rebuilding the socket. Only navigates when the live page differs from the
+  // seed target, so redirects/history moves the server already reflects do not
+  // cause a redundant navigation loop.
+  useEffect(() => {
+    if (!enabled || status !== "open") {
+      return;
+    }
+    const target = seedUrl?.trim();
+    if (!target) {
+      return;
+    }
+    if (normalizeSeedUrl(target) === normalizeSeedUrl(liveUrlRef.current)) {
+      return;
+    }
+    sendInput({ type: "navigate", url: target });
+  }, [enabled, seedUrl, sendInput, status]);
+
+  return { status, frameUrl, liveUrl, tabs, sendInput };
+}

frontend/src/components/workspace/chats/chat-box.tsx

@@ -4,6 +4,11 @@ import { useEffect, useMemo, useRef, useState } from "react";
 
 import { ConversationEmptyState } from "@/components/ai-elements/conversation";
 import { Button } from "@/components/ui/button";
+import {
+  ResizableHandle,
+  ResizablePanel,
+  ResizablePanelGroup,
+} from "@/components/ui/resizable";
 import {
   Sheet,
   SheetContent,
@@ -20,17 +25,19 @@ import {
   ArtifactFileList,
   useArtifacts,
 } from "../artifacts";
+import { BrowserViewPanel, useMaybeBrowserView } from "../browser-view";
 import { useThread } from "../messages/context";
 import { SidecarPanel, useMaybeSidecar } from "../sidecar";
 
 const RIGHT_PANEL_ANIMATION_MS = 280;
 
-type RightPanelKind = "sidecar" | "artifacts";
+type RightPanelKind = "sidecar" | "artifacts" | "browser";
 
-const ChatBox: React.FC<{ children: React.ReactNode; threadId: string }> = ({
+const ChatBox: React.FC<{
-  children,
+  children: React.ReactNode;
-  threadId,
+  threadId: string;
-}) => {
+  browserEnabled?: boolean;
+}> = ({ children, threadId, browserEnabled = true }) => {
   const { thread } = useThread();
   const isMobile = useIsMobile();
   const pathname = usePathname();
@@ -47,6 +54,8 @@ const ChatBox: React.FC<{ children: React.ReactNode; threadId: string }> = ({
   } = useArtifacts();
   const sidecar = useMaybeSidecar();
   const sidecarOpen = sidecar?.open ?? false;
+  const browserView = useMaybeBrowserView();
+  const browserViewOpen = browserEnabled && (browserView?.open ?? false);
 
   const [autoSelectFirstArtifact, setAutoSelectFirstArtifact] = useState(true);
   useEffect(() => {
@@ -104,9 +113,11 @@ const ChatBox: React.FC<{ children: React.ReactNode; threadId: string }> = ({
 
   const activeRightPanel: RightPanelKind | null = sidecarOpen
     ? "sidecar"
-    : artifactPanelOpen
+    : browserViewOpen
-      ? "artifacts"
+      ? "browser"
-      : null;
+      : artifactPanelOpen
+        ? "artifacts"
+        : null;
   const rightPanelOpen = activeRightPanel !== null;
   const [renderedRightPanel, setRenderedRightPanel] =
     useState<RightPanelKind | null>(activeRightPanel);
@@ -136,7 +147,16 @@ const ChatBox: React.FC<{ children: React.ReactNode; threadId: string }> = ({
     }
   }, [artifactsOpen, setArtifactsOpen, sidecarOpen]);
 
+  useEffect(() => {
+    if (!browserEnabled && browserView?.open) {
+      browserView.close();
+    }
+  }, [browserEnabled, browserView]);
+
   const rightPanelContent = useMemo(() => {
+    if (renderedRightPanel === "browser") {
+      return <BrowserViewPanel threadId={threadId} className="size-full" />;
+    }
     if (renderedRightPanel === "sidecar") {
       return <SidecarPanel />;
     }
@@ -208,6 +228,9 @@ const ChatBox: React.FC<{ children: React.ReactNode; threadId: string }> = ({
             if (sidecarOpen) {
               sidecar?.close();
             }
+            if (browserViewOpen) {
+              browserView?.close();
+            }
             if (artifactsOpen) {
               setArtifactsOpen(false);
             }
@@ -219,7 +242,11 @@ const ChatBox: React.FC<{ children: React.ReactNode; threadId: string }> = ({
           >
             <SheetHeader className="sr-only">
               <SheetTitle>
-                {renderedRightPanel === "sidecar" ? "Sidecar" : "Artifacts"}
+                {renderedRightPanel === "sidecar"
+                  ? "Sidecar"
+                  : renderedRightPanel === "browser"
+                    ? "Browser"
+                    : "Artifacts"}
               </SheetTitle>
               <SheetDescription>
                 Browse the side panel for this conversation.
@@ -236,41 +263,79 @@ const ChatBox: React.FC<{ children: React.ReactNode; threadId: string }> = ({
     <div
       id={`${resizableIdBase}-panels`}
       className={cn(
-        "[container-type:inline-size] grid size-full min-h-0 transition-[grid-template-columns] duration-[280ms] ease-out motion-reduce:transition-none",
+        "[container-type:inline-size] size-full min-h-0",
-        rightPanelOpen
+        activeRightPanel !== "browser" &&
-          ? "grid-cols-[minmax(0,1fr)_1px_minmax(0,40%)]"
+          "grid transition-[grid-template-columns] duration-[280ms] ease-out motion-reduce:transition-none",
-          : "grid-cols-[minmax(0,1fr)_0px_0px]",
+        activeRightPanel !== "browser" &&
+          (rightPanelOpen
+            ? "grid-cols-[minmax(0,1fr)_1px_minmax(0,40%)]"
+            : "grid-cols-[minmax(0,1fr)_0px_0px]"),
       )}
     >
-      <div className="relative min-h-0 min-w-0" id="chat">
+      {activeRightPanel === "browser" ? (
-        {children}
+        <ResizablePanelGroup
-      </div>
+          id={`${resizableIdBase}-group`}
-      <div
+          orientation="horizontal"
-        id={`${resizableIdBase}-separator`}
+          className="size-full min-h-0"
-        aria-hidden="true"
-        className={cn(
-          "bg-border opacity-33 transition-opacity duration-200 ease-out motion-reduce:transition-none",
-          !rightPanelOpen && "pointer-events-none opacity-0",
-        )}
-      />
-      <aside
-        aria-hidden={!rightPanelOpen}
-        className={cn(
-          "min-h-0 min-w-0 overflow-hidden transition-opacity duration-[280ms] ease-out motion-reduce:transition-none",
-          !rightPanelOpen && "pointer-events-none opacity-0",
-        )}
-        id="artifacts"
-      >
-        <div
-          className={cn(
-            "ml-auto h-full w-[40cqw] transition-opacity duration-[280ms] ease-out motion-reduce:transition-none",
-            renderedRightPanel === "sidecar" ? "p-0" : "p-4",
-            rightPanelOpen ? "opacity-100" : "opacity-0",
-          )}
         >
-          {rightPanelContent}
+          <ResizablePanel
-        </div>
+            id={`${resizableIdBase}-chat`}
-      </aside>
+            minSize="30%"
+            className="relative min-h-0 min-w-0"
+          >
+            <div className="relative size-full min-h-0 min-w-0" id="chat">
+              {children}
+            </div>
+          </ResizablePanel>
+          <ResizableHandle withHandle />
+          <ResizablePanel
+            id={`${resizableIdBase}-side`}
+            defaultSize="40%"
+            minSize="20%"
+            maxSize="75%"
+            className="min-h-0 min-w-0"
+          >
+            <aside
+              className="size-full min-h-0 min-w-0 overflow-hidden p-0"
+              id="artifacts"
+            >
+              {rightPanelContent}
+            </aside>
+          </ResizablePanel>
+        </ResizablePanelGroup>
+      ) : (
+        <>
+          <div className="relative min-h-0 min-w-0" id="chat">
+            {children}
+          </div>
+          <div
+            id={`${resizableIdBase}-separator`}
+            aria-hidden="true"
+            className={cn(
+              "bg-border opacity-33 transition-opacity duration-200 ease-out motion-reduce:transition-none",
+              !rightPanelOpen && "pointer-events-none opacity-0",
+            )}
+          />
+          <aside
+            aria-hidden={!rightPanelOpen}
+            className={cn(
+              "min-h-0 min-w-0 overflow-hidden transition-opacity duration-[280ms] ease-out motion-reduce:transition-none",
+              !rightPanelOpen && "pointer-events-none opacity-0",
+            )}
+            id="artifacts"
+          >
+            <div
+              className={cn(
+                "ml-auto h-full w-[40cqw] transition-opacity duration-[280ms] ease-out motion-reduce:transition-none",
+                renderedRightPanel === "sidecar" ? "p-0" : "p-4",
+                rightPanelOpen ? "opacity-100" : "opacity-0",
+              )}
+            >
+              {rightPanelContent}
+            </div>
+          </aside>
+        </>
+      )}
     </div>
   );
 };

frontend/src/components/workspace/messages/markdown-content.tsx

@@ -6,7 +6,11 @@ import {
   isValidElement,
   type ReactNode,
   useContext,
+  useDeferredValue,
+  useEffect,
   useMemo,
+  useRef,
+  useState,
 } from "react";
 
 import { type ClipboardSafeStreamdownProps } from "@/components/ai-elements/streamdown";
@@ -33,8 +37,125 @@ type StreamingCodeProps = ComponentProps<"code"> & {
   children?: ReactNode;
 };
 
+const SMOOTH_REVEAL_MIN_DELTA = 80;
+const SMOOTH_REVEAL_MIN_CHARS_PER_FRAME = 8;
+const SMOOTH_REVEAL_DURATION_MS = 300;
+
 const StreamingCodeBlockContext = createContext(false);
 
+function useSmoothStreamingContent(content: string, isLoading: boolean) {
+  const initialContent =
+    isLoading && content.length >= SMOOTH_REVEAL_MIN_DELTA ? "" : content;
+  const [displayContent, setDisplayContent] = useState(initialContent);
+  const displayContentRef = useRef(initialContent);
+  const targetContentRef = useRef(content);
+  const sawLoadingRef = useRef(isLoading);
+
+  useEffect(() => {
+    if (isLoading) {
+      sawLoadingRef.current = true;
+    }
+  }, [isLoading]);
+
+  useEffect(() => {
+    targetContentRef.current = content;
+
+    const current = displayContentRef.current;
+    const delta = content.length - current.length;
+    const shouldSmoothReveal =
+      delta >= SMOOTH_REVEAL_MIN_DELTA &&
+      content.startsWith(current) &&
+      (isLoading || sawLoadingRef.current);
+
+    if (!shouldSmoothReveal) {
+      if (current !== content) {
+        displayContentRef.current = content;
+        setDisplayContent(content);
+      }
+      if (!isLoading) {
+        sawLoadingRef.current = false;
+      }
+      return;
+    }
+
+    let cancelled = false;
+    let frame: number | null = null;
+    let timer: ReturnType<typeof setTimeout> | null = null;
+    let generation = 0;
+    const startedAt = performance.now();
+    const startLength = current.length;
+
+    const tick = (now: number, scheduledGeneration: number) => {
+      if (cancelled || scheduledGeneration !== generation) {
+        return;
+      }
+      generation += 1;
+      if (frame !== null) {
+        cancelAnimationFrame(frame);
+        frame = null;
+      }
+      if (timer !== null) {
+        clearTimeout(timer);
+        timer = null;
+      }
+
+      const target = targetContentRef.current;
+      const latest = displayContentRef.current;
+      if (!target.startsWith(latest) || latest.length >= target.length) {
+        if (!isLoading) {
+          sawLoadingRef.current = false;
+        }
+        return;
+      }
+
+      const progress = Math.min(
+        1,
+        (now - startedAt) / SMOOTH_REVEAL_DURATION_MS,
+      );
+      const elapsedLength = startLength + Math.ceil(delta * progress);
+      const nextLength = Math.max(
+        latest.length + SMOOTH_REVEAL_MIN_CHARS_PER_FRAME,
+        elapsedLength,
+      );
+      const next = target.slice(0, nextLength);
+      displayContentRef.current = next;
+      setDisplayContent(next);
+
+      if (next.length < target.length) {
+        scheduleTick();
+      } else if (!isLoading) {
+        sawLoadingRef.current = false;
+      }
+    };
+
+    const scheduleTick = () => {
+      const scheduledGeneration = ++generation;
+      frame = requestAnimationFrame((now) => tick(now, scheduledGeneration));
+      timer = setTimeout(
+        () => tick(performance.now(), scheduledGeneration),
+        50,
+      );
+    };
+
+    scheduleTick();
+    return () => {
+      cancelled = true;
+      generation += 1;
+      if (frame !== null) {
+        cancelAnimationFrame(frame);
+      }
+      if (timer !== null) {
+        clearTimeout(timer);
+      }
+    };
+  }, [content, isLoading]);
+
+  return {
+    content: displayContent,
+    isRevealing: displayContent !== content,
+  };
+}
+
 function StreamingPre({ children }: ComponentProps<"pre">) {
   const childClassName = isValidElement<{ className?: string }>(children)
     ? children.props.className
@@ -101,9 +222,16 @@ export function MarkdownContent({
   remarkPlugins = streamdownPluginsWithoutRawHtml.remarkPlugins,
   components: componentsFromProps,
 }: MarkdownContentProps) {
+  const deferredContent = useDeferredValue(content);
+  const targetContent = isLoading ? deferredContent : content;
+  const { content: displayContent, isRevealing } = useSmoothStreamingContent(
+    targetContent,
+    isLoading,
+  );
+  const isStreamingRender = isLoading || isRevealing;
   const normalizedContent = useMemo(
-    () => preprocessStreamdownMarkdown(content),
+    () => preprocessStreamdownMarkdown(displayContent),
-    [content],
+    [displayContent],
   );
   const effectiveRehypePlugins = useMemo(() => {
     const base = streamdownPluginsWithoutRawHtml.rehypePlugins ?? [];
@@ -115,7 +243,7 @@ export function MarkdownContent({
       a: createMarkdownLinkComponent(),
       ...componentsFromProps,
     };
-    if (!isLoading) {
+    if (!isStreamingRender) {
       return baseComponents;
     }
     return {
@@ -123,9 +251,9 @@ export function MarkdownContent({
       code: componentsFromProps?.code ?? StreamingCode,
       pre: componentsFromProps?.pre ?? StreamingPre,
     };
-  }, [componentsFromProps, isLoading]);
+  }, [componentsFromProps, isStreamingRender]);
 
-  if (!content) return null;
+  if (!displayContent) return null;
 
   return (
     <SafeMessageResponse
@@ -133,7 +261,7 @@ export function MarkdownContent({
       remarkPlugins={remarkPlugins}
       rehypePlugins={effectiveRehypePlugins}
       components={components}
-      parseIncompleteMarkdown={isLoading}
+      parseIncompleteMarkdown={isStreamingRender}
     >
       {normalizedContent}
     </SafeMessageResponse>

frontend/src/components/workspace/messages/message-group.tsx

@@ -9,12 +9,13 @@ import {
   ListTodoIcon,
   MessageCircleQuestionMarkIcon,
   MessageSquareTextIcon,
+  MonitorIcon,
   NotebookPenIcon,
   SearchIcon,
   SquareTerminalIcon,
   WrenchIcon,
 } from "lucide-react";
-import { useMemo, useState } from "react";
+import { memo, useMemo, useState } from "react";
 
 import {
   ChainOfThought,
@@ -25,6 +26,7 @@ import {
 } from "@/components/ai-elements/chain-of-thought";
 import { CodeBlock } from "@/components/ai-elements/code-block";
 import { Button } from "@/components/ui/button";
+import { resolveArtifactURL } from "@/core/artifacts/utils";
 import { useI18n } from "@/core/i18n/hooks";
 import { formatTokenCount } from "@/core/messages/usage";
 import type { TokenDebugStep } from "@/core/messages/usage-model";
@@ -39,24 +41,31 @@ import { env } from "@/env";
 import { cn } from "@/lib/utils";
 
 import { useArtifacts } from "../artifacts";
+import { useMaybeBrowserView } from "../browser-view";
 import { FlipDisplay } from "../flip-display";
 import { Tooltip } from "../tooltip";
 
 import { MarkdownContent } from "./markdown-content";
 
-export function MessageGroup({
+interface MessageGroupProps {
-  className,
-  messages,
-  isLoading = false,
-  tokenDebugSteps = [],
-  showTokenDebugSummaries = false,
-}: {
   className?: string;
   messages: Message[];
   isLoading?: boolean;
+  deferBrowserPreviews?: boolean;
   tokenDebugSteps?: TokenDebugStep[];
   showTokenDebugSummaries?: boolean;
-}) {
+  threadId?: string;
+}
+
+function MessageGroupComponent({
+  className,
+  messages,
+  isLoading = false,
+  deferBrowserPreviews = false,
+  tokenDebugSteps = [],
+  showTokenDebugSummaries = false,
+  threadId,
+}: MessageGroupProps) {
   const { t } = useI18n();
   const [showAbove, setShowAbove] = useState(
     env.NEXT_PUBLIC_STATIC_WEBSITE_ONLY === "true",
@@ -218,8 +227,10 @@ export function MessageGroup({
       <ToolCall
         key={step.id}
         {...step}
+        threadId={threadId}
         isLast={options?.isLast}
         isLoading={isLoading}
+        deferBrowserPreview={deferBrowserPreviews}
         tokenDebugStep={
           debugStep && !debugStep.sharedAttribution ? debugStep : undefined
         }
@@ -391,6 +402,47 @@ export function MessageGroup({
   );
 }
 
+export const MessageGroup = memo(
+  MessageGroupComponent,
+  areMessageGroupPropsEqual,
+);
+MessageGroup.displayName = "MessageGroup";
+
+function areMessageGroupPropsEqual(
+  previous: MessageGroupProps,
+  next: MessageGroupProps,
+): boolean {
+  if (next.isLoading) {
+    return false;
+  }
+  return (
+    previous.className === next.className &&
+    Boolean(previous.isLoading) === Boolean(next.isLoading) &&
+    Boolean(previous.deferBrowserPreviews) ===
+      Boolean(next.deferBrowserPreviews) &&
+    Boolean(previous.showTokenDebugSummaries) ===
+      Boolean(next.showTokenDebugSummaries) &&
+    previous.threadId === next.threadId &&
+    sameReferences(previous.messages, next.messages) &&
+    sameReferences(previous.tokenDebugSteps, next.tokenDebugSteps)
+  );
+}
+
+function sameReferences<T>(
+  previous: readonly T[] | undefined,
+  next: readonly T[] | undefined,
+): boolean {
+  if (previous === next) {
+    return true;
+  }
+  const previousItems = previous ?? [];
+  const nextItems = next ?? [];
+  return (
+    previousItems.length === nextItems.length &&
+    previousItems.every((item, index) => item === nextItems[index])
+  );
+}
+
 function formatDebugToken(
   debugStep: TokenDebugStep,
   t: ReturnType<typeof useI18n>["t"],
@@ -445,6 +497,35 @@ function DebugStepLabel({
   );
 }
 
+function browserToolLabel(
+  name: string,
+  args: Record<string, unknown>,
+  t: ReturnType<typeof useI18n>["t"],
+): string {
+  switch (name) {
+    case "browser_navigate":
+      return typeof args.url === "string"
+        ? t.toolCalls.browserNavigate(args.url)
+        : t.toolCalls.browserNavigateGeneric;
+    case "browser_click":
+      return t.toolCalls.browserClick;
+    case "browser_type":
+      return t.toolCalls.browserType;
+    case "browser_snapshot":
+      return t.toolCalls.browserSnapshot;
+    case "browser_get_text":
+      return t.toolCalls.browserGetText;
+    case "browser_back":
+      return t.toolCalls.browserBack;
+    case "browser_screenshot":
+      return t.toolCalls.browserScreenshot;
+    case "browser_close":
+      return t.toolCalls.browserClose;
+    default:
+      return t.toolCalls.useTool(name);
+  }
+}
+
 function ToolCall({
   id,
   messageId,
@@ -453,7 +534,10 @@ function ToolCall({
   result,
   isLast = false,
   isLoading = false,
+  deferBrowserPreview = false,
   tokenDebugStep,
+  browserView,
+  threadId,
 }: {
   id?: string;
   messageId?: string;
@@ -462,11 +546,15 @@ function ToolCall({
   result?: string | Record<string, unknown>;
   isLast?: boolean;
   isLoading?: boolean;
+  deferBrowserPreview?: boolean;
   tokenDebugStep?: TokenDebugStep;
+  browserView?: BrowserViewMeta;
+  threadId?: string;
 }) {
   const { t } = useI18n();
   const { setOpen, autoOpen, autoSelect, selectedArtifact, select } =
     useArtifacts();
+  const browserViewPanel = useMaybeBrowserView();
   const tokenLabel = tokenDebugStep
     ? formatDebugToken(tokenDebugStep, t)
     : null;
@@ -477,7 +565,54 @@ function ToolCall({
       fallback
     );
 
-  if (name === "web_search") {
+  if (name.startsWith("browser_")) {
+    const shot = browserView?.screenshot;
+    const previewUrl =
+      shot && threadId ? resolveArtifactURL(shot, threadId) : undefined;
+    return (
+      <ChainOfThoughtStep
+        key={id}
+        label={resolveLabel(browserToolLabel(name, args, t))}
+        icon={MonitorIcon}
+      >
+        {previewUrl && !deferBrowserPreview && (
+          <button
+            type="button"
+            className="border-border mt-1 block w-full max-w-md cursor-pointer overflow-hidden rounded-lg border"
+            onClick={() => {
+              if (!shot) {
+                return;
+              }
+              if (browserViewPanel) {
+                browserViewPanel.pushFrame({
+                  screenshot: shot,
+                  url: browserView?.url,
+                  title: browserView?.title,
+                });
+                browserViewPanel.openPanel();
+              } else {
+                select(shot);
+                setOpen(true);
+              }
+            }}
+          >
+            <img
+              className="w-full object-contain"
+              src={previewUrl}
+              alt={browserView?.title ?? "browser view"}
+              loading="lazy"
+              decoding="async"
+            />
+            {browserView?.url && (
+              <div className="text-muted-foreground bg-muted/40 truncate px-2 py-1 text-left text-[11px]">
+                {browserView.url}
+              </div>
+            )}
+          </button>
+        )}
+      </ChainOfThoughtStep>
+    );
+  } else if (name === "web_search") {
     let label: React.ReactNode = t.toolCalls.searchForRelatedInfo;
     if (typeof args.query === "string") {
       label = t.toolCalls.searchOnWebFor(args.query);
@@ -732,6 +867,7 @@ interface CoTToolCallStep extends GenericCoTStep<"toolCall"> {
   name: string;
   args: Record<string, unknown>;
   result?: string;
+  browserView?: BrowserViewMeta;
 }
 
 interface CoTAssistantTextStep extends GenericCoTStep<"assistantText"> {
@@ -740,6 +876,31 @@ interface CoTAssistantTextStep extends GenericCoTStep<"assistantText"> {
 
 type CoTStep = CoTAssistantTextStep | CoTReasoningStep | CoTToolCallStep;
 
+interface BrowserViewMeta {
+  screenshot: string;
+  url?: string;
+  title?: string;
+}
+
+function findBrowserViewMeta(
+  toolCallId: string,
+  messages: Message[],
+): BrowserViewMeta | undefined {
+  for (const message of messages) {
+    if (message.type === "tool" && message.tool_call_id === toolCallId) {
+      const meta = (
+        message.additional_kwargs as
+          | { browser_view?: BrowserViewMeta }
+          | undefined
+      )?.browser_view;
+      if (meta && typeof meta.screenshot === "string") {
+        return meta;
+      }
+    }
+  }
+  return undefined;
+}
+
 function convertToSteps(messages: Message[]): CoTStep[] {
   const steps: CoTStep[] = [];
   for (const [messageIndex, message] of messages.entries()) {
@@ -785,6 +946,7 @@ function convertToSteps(messages: Message[]): CoTStep[] {
               step.result = toolCallResult;
             }
           }
+          step.browserView = findBrowserViewMeta(toolCallId, messages);
         }
         steps.push(step);
       }

frontend/src/components/workspace/messages/message-list-item.tsx

@@ -212,14 +212,30 @@ function MessageImage({
   const imgClassName = cn("overflow-hidden rounded-lg", `max-w-[${maxWidth}]`);
 
   if (typeof src !== "string") {
-    return <img className={imgClassName} src={src} alt={alt} {...props} />;
+    return (
+      <img
+        className={imgClassName}
+        src={src}
+        alt={alt}
+        loading="lazy"
+        decoding="async"
+        {...props}
+      />
+    );
   }
 
   const url = resolveMessageImageURL(src, threadId, artifactPaths);
 
   return (
     <a href={url} target="_blank" rel="noopener noreferrer">
-      <img className={imgClassName} src={url} alt={alt} {...props} />
+      <img
+        className={imgClassName}
+        src={url}
+        alt={alt}
+        loading="lazy"
+        decoding="async"
+        {...props}
+      />
     </a>
   );
 }
@@ -611,6 +627,8 @@ function RichFileCard({
         <img
           src={fileUrl}
           alt={file.filename}
+          loading="lazy"
+          decoding="async"
           className="h-32 w-auto max-w-60 object-cover transition-transform group-hover:scale-105"
         />
       </a>

frontend/src/components/workspace/messages/message-list.tsx

@@ -39,6 +39,7 @@ import {
 } from "@/core/messages/human-input";
 import {
   buildTokenDebugSteps,
+  type TokenDebugStep,
   type TokenUsageInlineMode,
 } from "@/core/messages/usage-model";
 import {
@@ -55,6 +56,7 @@ import {
   hasReasoning,
   isAssistantMessageGroupStreaming,
   isHiddenFromUIMessage,
+  type MessageGroup as ThreadMessageGroup,
 } from "@/core/messages/utils";
 import { useRehypeSplitWordsIntoSpans } from "@/core/rehype";
 import {
@@ -71,6 +73,7 @@ import type { AgentThreadState } from "@/core/threads";
 import { cn } from "@/lib/utils";
 
 import { ArtifactFileList } from "../artifacts/artifact-file-list";
+import { useMaybeBrowserView } from "../browser-view";
 import { CopyButton } from "../copy-button";
 import { useMaybeSidecar } from "../sidecar/context";
 import { Tooltip } from "../tooltip";
@@ -89,6 +92,101 @@ import {
 import { MessageListSkeleton } from "./skeleton";
 import { SubtaskCard } from "./subtask-card";
 
+const EMPTY_TOKEN_DEBUG_STEPS: TokenDebugStep[] = [];
+const EMPTY_ARTIFACT_PATHS: readonly string[] = [];
+
+function messageStableKey(message: Message) {
+  if (
+    message.type === "tool" &&
+    typeof message.tool_call_id === "string" &&
+    message.tool_call_id.length > 0
+  ) {
+    return `tool:${message.tool_call_id}`;
+  }
+  if (typeof message.id === "string" && message.id.length > 0) {
+    return `message:${message.id}`;
+  }
+  return null;
+}
+
+function sameMessageIdentity(previous: Message, next: Message) {
+  if (previous === next) {
+    return true;
+  }
+  if (previous.type !== next.type) {
+    return false;
+  }
+  const previousKey = messageStableKey(previous);
+  const nextKey = messageStableKey(next);
+  return previousKey !== null && previousKey === nextKey;
+}
+
+function canReuseMessageGroup(
+  previous: ThreadMessageGroup | undefined,
+  next: ThreadMessageGroup,
+): previous is ThreadMessageGroup {
+  if (
+    !previous ||
+    previous.id !== next.id ||
+    previous.type !== next.type ||
+    previous.messages.length !== next.messages.length
+  ) {
+    return false;
+  }
+  return previous.messages.every(
+    (message, index) =>
+      next.messages[index] !== undefined &&
+      sameMessageIdentity(message, next.messages[index]),
+  );
+}
+
+function sameStrings(previous: readonly string[], next: readonly string[]) {
+  return (
+    previous.length === next.length &&
+    previous.every((value, index) => value === next[index])
+  );
+}
+
+function useStableArtifactPaths(paths: readonly string[] | undefined) {
+  const previousPathsRef = useRef<readonly string[]>(EMPTY_ARTIFACT_PATHS);
+  return useMemo(() => {
+    const nextPaths = paths ?? EMPTY_ARTIFACT_PATHS;
+    const previousPaths = previousPathsRef.current;
+    if (sameStrings(previousPaths, nextPaths)) {
+      return previousPaths;
+    }
+    previousPathsRef.current = nextPaths;
+    return nextPaths;
+  }, [paths]);
+}
+
+function useStableMessageGroups(
+  messages: Message[],
+  isLoading: boolean,
+): ThreadMessageGroup[] {
+  const previousGroupsRef = useRef<ThreadMessageGroup[]>([]);
+  const previousIsLoadingRef = useRef(false);
+  return useMemo(() => {
+    const nextGroups = getMessageGroups(messages);
+    const previousGroups = previousGroupsRef.current;
+    const activeGroupIndex =
+      isLoading || previousIsLoadingRef.current ? nextGroups.length - 1 : -1;
+    const stableGroups = nextGroups.map((group, index) => {
+      // Keep the actively streaming group fresh even if the SDK mutates a
+      // message object in place while appending token content.
+      if (index === activeGroupIndex) {
+        return group;
+      }
+      return canReuseMessageGroup(previousGroups[index], group)
+        ? previousGroups[index]
+        : group;
+    });
+    previousGroupsRef.current = stableGroups;
+    previousIsLoadingRef.current = isLoading;
+    return stableGroups;
+  }, [isLoading, messages]);
+}
+
 export const MESSAGE_LIST_DEFAULT_PADDING_BOTTOM = 24;
 
 const LOAD_MORE_HISTORY_THROTTLE_MS = 1200;
@@ -273,7 +371,47 @@ export function MessageList({
     prevIsLoading.current = thread.isLoading;
   }, [thread.isLoading]);
   const messages = thread.messages;
-  const groupedMessages = getMessageGroups(messages);
+  const browserView = useMaybeBrowserView();
+  const pushBrowserFrame = browserView?.pushFrame;
+  const messageCount = messages.length;
+  useEffect(() => {
+    // Only the primary chat surface drives the shared browser panel. The
+    // sidecar renders a different thread's messages against the same
+    // BrowserViewProvider; pushing its frames would make the panel resolve
+    // another thread's screenshot with the primary threadId (404 / wrong page).
+    if (sidecarSurface || !pushBrowserFrame) {
+      return;
+    }
+    for (let i = messages.length - 1; i >= 0; i--) {
+      const message = messages[i];
+      if (message?.type !== "tool") {
+        continue;
+      }
+      const meta = (
+        message.additional_kwargs as
+          | {
+              browser_view?: {
+                screenshot?: string;
+                url?: string;
+                title?: string;
+              };
+            }
+          | undefined
+      )?.browser_view;
+      if (meta && typeof meta.screenshot === "string") {
+        pushBrowserFrame({
+          screenshot: meta.screenshot,
+          url: meta.url,
+          title: meta.title,
+        });
+        return;
+      }
+    }
+    // messages is intentionally read (not a dep) so token updates do not
+    // repeatedly scan long history looking for the last browser frame.
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [messageCount, pushBrowserFrame, sidecarSurface]);
+  const groupedMessages = useStableMessageGroups(messages, thread.isLoading);
   const [regeneratingMessageId, setRegeneratingMessageId] = useState<
     string | null
   >(null);
@@ -302,8 +440,47 @@ export function MessageList({
   const turnUsageMessagesByGroupIndex =
     getAssistantTurnUsageMessages(groupedMessages);
   const tokenDebugSteps = useMemo(
-    () => buildTokenDebugSteps(messages, t),
+    () =>
-    [messages, t],
+      tokenUsageInlineMode === "step_debug"
+        ? buildTokenDebugSteps(messages, t)
+        : EMPTY_TOKEN_DEBUG_STEPS,
+    [messages, t, tokenUsageInlineMode],
+  );
+  const showTokenDebugSummaries = tokenUsageInlineMode === "step_debug";
+  const tokenDebugStepsByMessageId = useMemo(() => {
+    const stepsByMessageId = new Map<string, TokenDebugStep[]>();
+    for (const step of tokenDebugSteps) {
+      const messageId = step.messageId;
+      if (!messageId) {
+        continue;
+      }
+      const steps = stepsByMessageId.get(messageId);
+      if (steps) {
+        steps.push(step);
+      } else {
+        stepsByMessageId.set(messageId, [step]);
+      }
+    }
+    return stepsByMessageId;
+  }, [tokenDebugSteps]);
+  const getTokenDebugStepsForMessages = useCallback(
+    (groupMessages: Message[]) => {
+      if (!showTokenDebugSummaries) {
+        return EMPTY_TOKEN_DEBUG_STEPS;
+      }
+      const steps: TokenDebugStep[] = [];
+      for (const message of groupMessages) {
+        if (!message.id) {
+          continue;
+        }
+        const matched = tokenDebugStepsByMessageId.get(message.id);
+        if (matched) {
+          steps.push(...matched);
+        }
+      }
+      return steps;
+    },
+    [showTokenDebugSummaries, tokenDebugStepsByMessageId],
   );
   const streamingMessages = useMemo(
     () =>
@@ -674,7 +851,7 @@ export function MessageList({
         );
       }
 
-      if (tokenUsageInlineMode === "step_debug" && inlineDebug) {
+      if (showTokenDebugSummaries && inlineDebug) {
         const messageIds = new Set(
           debugMessageIds ??
             messages
@@ -695,15 +872,22 @@ export function MessageList({
 
       return null;
     },
-    [thread.isLoading, tokenDebugSteps, tokenUsageInlineMode],
+    [
+      showTokenDebugSummaries,
+      thread.isLoading,
+      tokenDebugSteps,
+      tokenUsageInlineMode,
+    ],
+  );
+
+  const artifactPaths = useStableArtifactPaths(
+    extractArtifactsFromThread(thread),
   );
 
   if (thread.isThreadLoading && messages.length === 0) {
     return <MessageListSkeleton />;
   }
 
-  const artifactPaths = extractArtifactsFromThread(thread);
-
   return (
     <>
       <Conversation
@@ -949,12 +1133,9 @@ export function MessageList({
                       key={"thinking-group-" + message.id}
                       messages={[message]}
                       isLoading={groupIsLoading}
-                      tokenDebugSteps={tokenDebugSteps.filter(
+                      deferBrowserPreviews={thread.isLoading}
-                        (step) => step.messageId === message.id,
+                      tokenDebugSteps={getTokenDebugStepsForMessages([message])}
-                      )}
+                      showTokenDebugSummaries={showTokenDebugSummaries}
-                      showTokenDebugSummaries={
-                        tokenUsageInlineMode === "step_debug"
-                      }
                     />,
                   );
                 } else if (message.id) {
@@ -993,15 +1174,13 @@ export function MessageList({
               <div key={"group-" + group.id} className="w-full">
                 <MessageGroup
                   messages={group.messages}
-                  isLoading={thread.isLoading}
+                  isLoading={groupIsLoading}
-                  tokenDebugSteps={tokenDebugSteps.filter((step) =>
+                  deferBrowserPreviews={thread.isLoading}
-                    group.messages.some(
+                  threadId={threadId}
-                      (message) => message.id === step.messageId,
+                  tokenDebugSteps={getTokenDebugStepsForMessages(
-                    ),
+                    group.messages,
                   )}
-                  showTokenDebugSummaries={
+                  showTokenDebugSummaries={showTokenDebugSummaries}
-                    tokenUsageInlineMode === "step_debug"
-                  }
                 />
                 {renderTokenUsage({
                   messages: group.messages,

frontend/src/components/workspace/recent-chat-list.tsx

@@ -81,10 +81,16 @@ export function RecentChatList() {
     hasNextPage,
     isFetchingNextPage,
   } = useInfiniteThreads();
-  const threads = useMemo(
+  const threads = useMemo(() => {
-    () => infiniteThreads?.pages.flat() ?? [],
+    const seen = new Set<string>();
-    [infiniteThreads],
+    return (infiniteThreads?.pages.flat() ?? []).filter((thread) => {
-  );
+      if (seen.has(thread.thread_id)) {
+        return false;
+      }
+      seen.add(thread.thread_id);
+      return true;
+    });
+  }, [infiniteThreads]);
 
   const sentinelRef = useRef<HTMLDivElement | null>(null);
   useEffect(() => {

frontend/src/components/workspace/sidecar/sidecar-panel.tsx

@@ -550,7 +550,7 @@ export function SidecarPanel({ className }: { className?: string }) {
                 : t.sidecar.noContext}
           </div>
         </div>
-        {hasSidecarThread ? (
+        {hasSidecarThread && (
           <Tooltip content={t.sidecar.delete}>
             <Button
               aria-label={t.sidecar.delete}
@@ -563,22 +563,21 @@ export function SidecarPanel({ className }: { className?: string }) {
               <Trash2Icon />
             </Button>
           </Tooltip>
-        ) : (
-          // No conversation yet — nothing to delete, so this just discards the
-          // draft and closes the panel. A plain X (no confirm) keeps it light.
-          <Tooltip content={t.common.close}>
-            <Button
-              aria-label={t.common.close}
-              className="text-muted-foreground hover:text-foreground"
-              data-testid="sidecar-close-button"
-              size="icon-sm"
-              variant="ghost"
-              onClick={() => discardDraftAndClose()}
-            >
-              <XIcon />
-            </Button>
-          </Tooltip>
         )}
+        <Tooltip content={t.common.close}>
+          <Button
+            aria-label={t.common.close}
+            className="text-muted-foreground hover:text-foreground"
+            data-testid="sidecar-close-button"
+            size="icon-sm"
+            variant="ghost"
+            onClick={() =>
+              hasSidecarThread ? sidecar.close() : discardDraftAndClose()
+            }
+          >
+            <XIcon />
+          </Button>
+        </Tooltip>
       </header>
 
       <div className="min-h-0 flex-1">

frontend/src/core/agents/api.ts

@@ -1,5 +1,6 @@
 import { fetch } from "@/core/api/fetcher";
 import { getBackendBaseURL } from "@/core/config";
+export { fetchAgentsApiEnabled } from "@/core/features/api";
 
 import type { Agent, CreateAgentRequest, UpdateAgentRequest } from "./types";
 
@@ -87,19 +88,6 @@ export async function deleteAgent(name: string): Promise<void> {
   if (!res.ok) throw new Error(`Failed to delete agent: ${res.statusText}`);
 }
 
-interface FeaturesResponse {
-  agents_api: { enabled: boolean };
-}
-
-export async function fetchAgentsApiEnabled(): Promise<boolean> {
-  const res = await fetch(`${getBackendBaseURL()}/api/features`);
-  if (!res.ok) {
-    throw new Error(`Failed to load features: ${res.statusText}`);
-  }
-  const data = (await res.json()) as FeaturesResponse;
-  return data.agents_api.enabled;
-}
-
 export async function checkAgentName(
   name: string,
 ): Promise<{ available: boolean; name: string }> {

frontend/src/core/features/api.ts

@@ -0,0 +1,23 @@
+import { fetch } from "@/core/api/fetcher";
+import { getBackendBaseURL } from "@/core/config";
+
+export interface FeaturesResponse {
+  agents_api: { enabled: boolean };
+  browser_control?: { enabled: boolean };
+}
+
+export async function fetchFeatures(): Promise<FeaturesResponse> {
+  const res = await fetch(`${getBackendBaseURL()}/api/features`);
+  if (!res.ok) {
+    throw new Error(`Failed to load features: ${res.statusText}`);
+  }
+  return (await res.json()) as FeaturesResponse;
+}
+
+export async function fetchAgentsApiEnabled(): Promise<boolean> {
+  return (await fetchFeatures()).agents_api.enabled;
+}
+
+export async function fetchBrowserControlEnabled(): Promise<boolean> {
+  return (await fetchFeatures()).browser_control?.enabled ?? false;
+}

frontend/src/core/features/hooks.ts

@@ -0,0 +1,18 @@
+import { useQuery } from "@tanstack/react-query";
+
+import { fetchBrowserControlEnabled } from "./api";
+
+export function useBrowserControlEnabled() {
+  const { data, isPending } = useQuery({
+    queryKey: ["features", "browser_control"],
+    queryFn: () => fetchBrowserControlEnabled(),
+    staleTime: 0,
+    refetchOnMount: true,
+    retry: false,
+  });
+
+  return {
+    enabled: data ?? false,
+    isLoading: isPending,
+  };
+}

frontend/src/core/features/index.ts

@@ -0,0 +1,2 @@
+export * from "./api";
+export * from "./hooks";

frontend/src/core/i18n/locales/en-US.ts

@@ -53,6 +53,8 @@ export const enUS: Translations = {
     regenerate: "Regenerate",
     branch: "Branch conversation",
     showArtifacts: "Show artifacts of this conversation",
+    browser: "Browser",
+    showBrowser: "Open browser panel",
   },
 
   // Home
@@ -567,6 +569,15 @@ export const enUS: Translations = {
     clickToViewContent: "Click to view file content",
     writeTodos: "Update to-do list",
     skillInstallTooltip: "Install skill and make it available to DeerFlow",
+    browserNavigate: (url: string) => `Open ${url} in browser`,
+    browserNavigateGeneric: "Open page in browser",
+    browserClick: "Click element in browser",
+    browserType: "Type into browser field",
+    browserSnapshot: "Read page in browser",
+    browserGetText: "Read page text in browser",
+    browserBack: "Go back in browser",
+    browserScreenshot: "Capture browser screenshot",
+    browserClose: "Close browser",
   },
 
   humanInput: {

frontend/src/core/i18n/locales/types.ts

@@ -42,6 +42,8 @@ export interface Translations {
     regenerate: string;
     branch: string;
     showArtifacts: string;
+    browser: string;
+    showBrowser: string;
   };
 
   home: {
@@ -457,6 +459,15 @@ export interface Translations {
     clickToViewContent: string;
     writeTodos: string;
     skillInstallTooltip: string;
+    browserNavigate: (url: string) => string;
+    browserNavigateGeneric: string;
+    browserClick: string;
+    browserType: string;
+    browserSnapshot: string;
+    browserGetText: string;
+    browserBack: string;
+    browserScreenshot: string;
+    browserClose: string;
   };
 
   humanInput: {

frontend/src/core/i18n/locales/zh-CN.ts

@@ -53,6 +53,8 @@ export const zhCN: Translations = {
     regenerate: "重新生成",
     branch: "分叉",
     showArtifacts: "查看此对话的文件",
+    browser: "浏览器",
+    showBrowser: "打开浏览器面板",
   },
 
   // Home
@@ -545,6 +547,15 @@ export const zhCN: Translations = {
     clickToViewContent: "点击查看文件内容",
     writeTodos: "更新 To-do 列表",
     skillInstallTooltip: "安装技能并使其可在 DeerFlow 中使用",
+    browserNavigate: (url: string) => `在浏览器中打开 ${url}`,
+    browserNavigateGeneric: "在浏览器中打开页面",
+    browserClick: "在浏览器中点击元素",
+    browserType: "在浏览器中输入内容",
+    browserSnapshot: "在浏览器中读取页面",
+    browserGetText: "在浏览器中读取页面文本",
+    browserBack: "在浏览器中返回上一页",
+    browserScreenshot: "浏览器截图",
+    browserClose: "关闭浏览器",
   },
 
   humanInput: {

frontend/src/styles/globals.css

@@ -85,6 +85,16 @@
     }
   }
 
+  --animate-loading-bar: loading-bar 1s ease-in-out infinite;
+  @keyframes loading-bar {
+    0% {
+      transform: translateX(-100%);
+    }
+    100% {
+      transform: translateX(400%);
+    }
+  }
+
   --animate-fade-in-up: fade-in-up 0.15s ease-in-out forwards;
   @keyframes fade-in-up {
     0% {

frontend/tests/e2e/browser-feature.spec.ts

@@ -0,0 +1,39 @@
+import { expect, test } from "@playwright/test";
+
+import { MOCK_THREAD_ID, mockLangGraphAPI } from "./utils/mock-api";
+
+test.describe("Browser feature flag", () => {
+  test("shows browser trigger only when browser_control is enabled", async ({
+    page,
+  }) => {
+    mockLangGraphAPI(page, {
+      threads: [{ thread_id: MOCK_THREAD_ID, title: "Browser Enabled" }],
+    });
+
+    await page.goto(`/workspace/chats/${MOCK_THREAD_ID}`);
+
+    await expect(page.getByTestId("browser-trigger")).toBeVisible({
+      timeout: 15_000,
+    });
+  });
+
+  test("hides browser trigger when browser_control is disabled", async ({
+    page,
+  }) => {
+    mockLangGraphAPI(page, {
+      threads: [{ thread_id: MOCK_THREAD_ID, title: "Browser Disabled" }],
+      features: { browserControlEnabled: false },
+    });
+
+    const featuresResponse = page.waitForResponse((response) =>
+      response.url().includes("/api/features"),
+    );
+    await page.goto(`/workspace/chats/${MOCK_THREAD_ID}`);
+    const features = (await (await featuresResponse).json()) as {
+      browser_control?: { enabled?: boolean };
+    };
+    expect(features.browser_control?.enabled).toBe(false);
+
+    await expect(page.getByTestId("browser-trigger")).toHaveCount(0);
+  });
+});

frontend/tests/e2e/utils/mock-api.ts

@@ -87,6 +87,10 @@ export type MockAPIOptions = {
     max_file_size: number;
     max_total_size: number;
   };
+  features?: {
+    agentsApiEnabled?: boolean;
+    browserControlEnabled?: boolean;
+  };
 };
 
 const DEFAULT_SKILLS: MockSkill[] = [
@@ -126,6 +130,16 @@ function visibleInputMessages(messages: unknown[]) {
   return messages.filter((message) => !isHiddenInputMessage(message));
 }
 
+function mockMessageRunId(message: unknown, fallback: string) {
+  if (typeof message === "object" && message !== null) {
+    const runId = Reflect.get(message, "run_id");
+    if (typeof runId === "string" && runId.length > 0) {
+      return runId;
+    }
+  }
+  return fallback;
+}
+
 function visibleRunInputMessages(route: Route) {
   try {
     const body = route.request().postDataJSON() as {
@@ -244,6 +258,10 @@ export function mockLangGraphAPI(page: Page, options?: MockAPIOptions) {
     max_file_size: 50 * 1024 * 1024,
     max_total_size: 100 * 1024 * 1024,
   };
+  const featureFlags = {
+    agentsApiEnabled: options?.features?.agentsApiEnabled ?? true,
+    browserControlEnabled: options?.features?.browserControlEnabled ?? true,
+  };
 
   const upsertThread = (thread: MockThread) => {
     threads = [
@@ -962,7 +980,10 @@ export function mockLangGraphAPI(page: Page, options?: MockAPIOptions) {
         contentType: "application/json",
         body: JSON.stringify({
           data: (matchingThread?.messages ?? []).map((message, index) => ({
-            run_id: `run-${matchingThread?.thread_id ?? "unknown"}`,
+            run_id: mockMessageRunId(
+              message,
+              `run-${matchingThread?.thread_id ?? "unknown"}`,
+            ),
             seq: index + 1,
             content: message,
             metadata: { caller: "lead_agent" },
@@ -1016,7 +1037,7 @@ export function mockLangGraphAPI(page: Page, options?: MockAPIOptions) {
     return route.fallback();
   });
 
-  // Feature flags — frontend gates UI (e.g. agents) on these. Default to
+  // Feature flags — frontend gates UI (e.g. agents/browser) on these. Default to
   // enabled so existing tests exercise the normal path; tests that need the
   // disabled state override this route after calling mockLangGraphAPI.
   void page.route("**/api/features", (route) => {
@@ -1024,7 +1045,10 @@ export function mockLangGraphAPI(page: Page, options?: MockAPIOptions) {
       return route.fulfill({
         status: 200,
         contentType: "application/json",
-        body: JSON.stringify({ agents_api: { enabled: true } }),
+        body: JSON.stringify({
+          agents_api: { enabled: featureFlags.agentsApiEnabled },
+          browser_control: { enabled: featureFlags.browserControlEnabled },
+        }),
       });
     }
     return route.fallback();

frontend/tests/unit/components/workspace/browser-view/keyboard.test.ts

@@ -0,0 +1,71 @@
+import { describe, expect, it } from "@rstest/core";
+
+import { decideBrowserKeyInput } from "@/components/workspace/browser-view/keyboard";
+
+function ctx(
+  overrides: Partial<Parameters<typeof decideBrowserKeyInput>[0]> = {},
+) {
+  return {
+    live: true,
+    editableTarget: false,
+    composing: false,
+    key: "a",
+    metaKey: false,
+    ctrlKey: false,
+    ...overrides,
+  };
+}
+
+describe("decideBrowserKeyInput", () => {
+  it("forwards printable chars as text", () => {
+    expect(decideBrowserKeyInput(ctx({ key: "a" }))).toEqual({
+      type: "text",
+      text: "a",
+    });
+  });
+
+  it("forwards named keys as key presses", () => {
+    expect(decideBrowserKeyInput(ctx({ key: "Enter" }))).toEqual({
+      type: "key",
+      key: "Enter",
+    });
+    expect(decideBrowserKeyInput(ctx({ key: "ArrowLeft" }))).toEqual({
+      type: "key",
+      key: "ArrowLeft",
+    });
+  });
+
+  it("forwards modifier combos as a normalized key chord", () => {
+    expect(decideBrowserKeyInput(ctx({ key: "c", metaKey: true }))).toEqual({
+      type: "key",
+      key: "Meta+C",
+    });
+    expect(decideBrowserKeyInput(ctx({ key: "a", ctrlKey: true }))).toEqual({
+      type: "key",
+      key: "Control+A",
+    });
+  });
+
+  it("does not forward while an IME composition is active", () => {
+    // A CJK candidate confirmed with Enter must not submit the remote page.
+    expect(
+      decideBrowserKeyInput(ctx({ key: "Enter", composing: true })),
+    ).toBeNull();
+    expect(
+      decideBrowserKeyInput(ctx({ key: "a", composing: true })),
+    ).toBeNull();
+  });
+
+  it("ignores keys when not live or focus is on an editable target", () => {
+    expect(
+      decideBrowserKeyInput(ctx({ key: "Enter", live: false })),
+    ).toBeNull();
+    expect(
+      decideBrowserKeyInput(ctx({ key: "Enter", editableTarget: true })),
+    ).toBeNull();
+  });
+
+  it("ignores unmapped named keys", () => {
+    expect(decideBrowserKeyInput(ctx({ key: "F5" }))).toBeNull();
+  });
+});

frontend/tests/unit/components/workspace/messages/markdown-content.test.ts

@@ -82,6 +82,13 @@ describe("MarkdownContent streaming code blocks", () => {
     expect(html).toContain('data-custom-code="true"');
     expect(html).toContain("data-streaming-code-block");
   });
+
+  it("does not paint an initial large streaming chunk all at once", () => {
+    const content = "x".repeat(120);
+
+    expect(renderMarkdown(content, true)).not.toContain(content);
+    expect(renderMarkdown(content, false)).toContain(content);
+  });
 });
 
 describe("MarkdownContent strikethrough", () => {

frontend/tests/unit/components/workspace/messages/message-group.test.ts

@@ -1,6 +1,6 @@
 import type { Message } from "@langchain/langgraph-sdk";
 import { describe, expect, it, rs } from "@rstest/core";
-import { createElement } from "react";
+import { createElement, type ComponentProps } from "react";
 import { renderToStaticMarkup } from "react-dom/server";
 
 import { MessageGroup } from "@/components/workspace/messages/message-group";
@@ -121,9 +121,55 @@ describe("MessageGroup", () => {
     expect(html).toContain("1 more step");
     expect(html).not.toContain("Check how processing groups convert messages.");
   });
+
+  it("defers browser screenshot previews while the thread is loading", () => {
+    const messages = [
+      {
+        id: "ai-1",
+        type: "ai",
+        content: "",
+        tool_calls: [
+          {
+            id: "call-1",
+            name: "browser_navigate",
+            args: { url: "https://github.com/bytedance/deer-flow" },
+          },
+        ],
+      } as Message,
+      {
+        id: "tool-1",
+        type: "tool",
+        name: "browser_navigate",
+        tool_call_id: "call-1",
+        content: "Opened",
+        additional_kwargs: {
+          browser_view: {
+            screenshot: "/mnt/user-data/outputs/browser.png",
+            url: "https://github.com/bytedance/deer-flow",
+          },
+        },
+      } as Message,
+    ];
+
+    const visibleHtml = renderGroup(messages, {
+      threadId: "thread-1",
+      deferBrowserPreviews: false,
+    });
+    const deferredHtml = renderGroup(messages, {
+      threadId: "thread-1",
+      deferBrowserPreviews: true,
+    });
+
+    expect(visibleHtml).toContain("<img");
+    expect(visibleHtml).toContain('decoding="async"');
+    expect(deferredHtml).not.toContain("<img");
+  });
 });
 
-function renderGroup(messages: Message[]) {
+function renderGroup(
+  messages: Message[],
+  props: Omit<ComponentProps<typeof MessageGroup>, "messages"> = {},
+) {
   return renderToStaticMarkup(
     createElement(
       I18nContext.Provider,
@@ -133,7 +179,7 @@ function renderGroup(messages: Message[]) {
           setLocale: () => undefined,
         },
       },
-      createElement(MessageGroup, { messages }),
+      createElement(MessageGroup, { ...props, messages }),
     ),
   );
 }

frontend/tests/unit/core/agents/features.test.ts

@@ -10,6 +10,7 @@ rs.mock("@/core/config", () => ({
 
 import { fetchAgentsApiEnabled } from "@/core/agents/api";
 import { fetch as fetcher } from "@/core/api/fetcher";
+import { fetchBrowserControlEnabled } from "@/core/features/api";
 
 const mockedFetch = rs.mocked(fetcher);
 
@@ -45,3 +46,36 @@ describe("fetchAgentsApiEnabled", () => {
     await expect(fetchAgentsApiEnabled()).rejects.toThrow();
   });
 });
+
+describe("fetchBrowserControlEnabled", () => {
+  test("returns true when backend reports browser_control enabled", async () => {
+    mockedFetch.mockResolvedValueOnce(
+      jsonResponse(200, {
+        agents_api: { enabled: true },
+        browser_control: { enabled: true },
+      }),
+    );
+    await expect(fetchBrowserControlEnabled()).resolves.toBe(true);
+    expect(mockedFetch).toHaveBeenCalledWith("/api/features");
+  });
+
+  test("returns false when browser_control is disabled or omitted", async () => {
+    mockedFetch.mockResolvedValueOnce(
+      jsonResponse(200, {
+        agents_api: { enabled: true },
+        browser_control: { enabled: false },
+      }),
+    );
+    await expect(fetchBrowserControlEnabled()).resolves.toBe(false);
+
+    mockedFetch.mockResolvedValueOnce(
+      jsonResponse(200, { agents_api: { enabled: true } }),
+    );
+    await expect(fetchBrowserControlEnabled()).resolves.toBe(false);
+  });
+
+  test("throws when the features request fails", async () => {
+    mockedFetch.mockResolvedValueOnce(jsonResponse(500, {}));
+    await expect(fetchBrowserControlEnabled()).rejects.toThrow();
+  });
+});

scripts/detect_uv_extras.py

@@ -11,6 +11,7 @@ Order of resolution:
    - database.backend == postgres        -> postgres
    - checkpointer.type == postgres       -> postgres
    - stream_bridge.type == redis         -> redis
+   - tools[].name == browser_navigate    -> browser
    - sandbox.ownership.type == redis     -> redis
 3. Runtime environment toggles that enable optional backends:
    - DEER_FLOW_STREAM_BRIDGE_REDIS_URL   -> redis
@@ -76,6 +77,7 @@ def find_config_file() -> Path | None:
 _SECTION_RE = re.compile(r"^([A-Za-z_][\w-]*)\s*:\s*$")
 _INDENTED_SECTION_RE = re.compile(r"^\s+([A-Za-z_][\w-]*)\s*:\s*$")
 _KEY_RE = re.compile(r"^\s+([A-Za-z_][\w-]*)\s*:\s*(\S.*?)\s*$")
+_LIST_ITEM_NAME_RE = re.compile(r"^\s*-\s+name\s*:\s*(\S.*?)\s*$")
 
 
 def _strip_comment(line: str) -> str:
@@ -222,6 +224,30 @@ def nested_section_value(lines: list[str], section_path: str, key: str) -> str |
     return None
 
 
+def tools_include_name(lines: list[str], tool_name: str) -> bool:
+    """Return True when the top-level tools list has an active item name."""
+    inside = False
+    for raw in lines:
+        line = _strip_comment(raw)
+        if not line.strip():
+            continue
+        sect_match = _SECTION_RE.match(line)
+        if sect_match:
+            inside = sect_match.group(1) == "tools"
+            continue
+        if not inside:
+            continue
+        stripped = line.lstrip()
+        indent = len(line) - len(stripped)
+        if indent == 0:
+            inside = False
+            continue
+        name_match = _LIST_ITEM_NAME_RE.match(line)
+        if name_match and _unquote(name_match.group(1).strip()) == tool_name:
+            return True
+    return False
+
+
 def detect_from_config(path: Path) -> list[str]:
     try:
         text = path.read_text(encoding="utf-8", errors="replace")
@@ -239,6 +265,8 @@ def detect_from_config(path: Path) -> list[str]:
         extras.add("redis")
     if (nested_section_value(lines, "channels.discord", "enabled") or "").lower() == "true":
         extras.add("discord")
+    if tools_include_name(lines, "browser_navigate"):
+        extras.add("browser")
     return sorted(extras)