agent-zero/plugins/_browser/tools/browser.py
Alessandro 430c48d1a5 Make browser screenshots ephemeral and context scoped
Route no-path browser screenshots through an in-process ephemeral image registry that vision_load consumes into the existing data-url model boundary. Stop materializing host-browser artifacts into tmp/browser/host-screenshots, keep explicit path screenshots durable, and make browser log metadata point at the active chat/task context while preserving browser-context detail.
2026-05-22 09:50:47 +02:00

566 lines
21 KiB
Python

from __future__ import annotations
import json
import re
import time
import uuid
from pathlib import Path
from typing import Any
from helpers import files
from helpers.print_style import PrintStyle
from helpers.tool import Response, Tool
from plugins._browser.helpers.selector import get_tool_runtime
HISTORY_SCREENSHOT_QUALITY = 62
HISTORY_SCREENSHOT_ACTION_DENYLIST = {"close", "close_all"}
async def get_runtime(context_id: str, create: bool = True, agent: Any | None = None):
if agent is not None:
return await get_tool_runtime(agent)
from plugins._browser.helpers.runtime import get_runtime as get_container_runtime
return await get_container_runtime(context_id, create=create)
class Browser(Tool):
async def execute(
self,
action: str = "",
browser_id: int | str | None = None,
url: str = "",
ref: int | str | None = None,
target_ref: int | str | None = None,
text: str = "",
selector: str = "",
selectors: list[str] | None = None,
script: str = "",
modifiers: list[str] | str | None = None,
keys: list[str] | None = None,
key: str = "",
include_content: bool = False,
focus_popup: bool | None = None,
event_type: str = "",
x: float = 0.0,
y: float = 0.0,
to_x: float = 0.0,
to_y: float = 0.0,
offset_x: float = 0.0,
offset_y: float = 0.0,
target_offset_x: float = 0.0,
target_offset_y: float = 0.0,
delta_x: float = 0.0,
delta_y: float = 0.0,
button: str = "left",
quality: int = 80,
full_page: bool = False,
path: str = "",
paths: list[str] | None = None,
value: str = "",
values: list[str] | None = None,
checked: bool | None = None,
width: int = 0,
height: int = 0,
calls: list[dict[str, Any]] | None = None,
**kwargs: Any,
) -> Response:
method_action = str(self.method or "").strip().lower().replace("-", "_")
requested_action = str(action or "").strip().lower().replace("-", "_")
clipboard_action = ""
if method_action == "clipboard" and requested_action in {"copy", "cut", "paste"}:
clipboard_action = requested_action
action = "clipboard"
else:
action = str(action or self.method or "state").strip().lower().replace("-", "_")
try:
runtime = await get_runtime(self.agent.context.id, agent=self.agent)
except Exception as exc:
return Response(message=f"Browser runtime unavailable: {exc}", break_loop=False)
if isinstance(modifiers, str):
modifiers = [modifiers] if modifiers else None
elif isinstance(modifiers, list) and not modifiers:
modifiers = None
keys = self._normalize_keys(keys)
try:
if action == "open":
result = await runtime.call("open", url or "")
elif action == "screenshot":
result = await runtime.call(
"screenshot_file",
browser_id,
quality=quality,
full_page=full_page,
path=path,
)
elif action == "list":
result = await runtime.call("list", include_content=bool(include_content))
elif action == "state":
result = await runtime.call("state", browser_id)
elif action in {"set_active", "setactive", "activate", "focus"}:
result = await runtime.call("set_active", browser_id)
elif action == "navigate":
result = await runtime.call("navigate", browser_id, url)
elif action == "back":
result = await runtime.call("back", browser_id)
elif action == "forward":
result = await runtime.call("forward", browser_id)
elif action == "reload":
result = await runtime.call("reload", browser_id)
elif action == "content":
payload = self._selector_payload(selector, selectors)
result = await runtime.call("content", browser_id, payload)
elif action == "detail":
result = await runtime.call(
"detail",
browser_id,
await self._resolve_ref(runtime, browser_id, ref, selector, action),
)
elif action == "click":
resolved_ref = await self._resolve_ref(
runtime,
browser_id,
ref,
selector,
action,
required=not self._has_coordinates(x, y),
)
if resolved_ref is None and self._has_coordinates(x, y):
result = await runtime.call(
"mouse", browser_id, "click", x, y,
button=button or "left", modifiers=modifiers,
)
elif modifiers:
result = await runtime.call(
"click", browser_id, resolved_ref,
modifiers=modifiers, focus_popup=focus_popup,
)
else:
result = await runtime.call("click", browser_id, resolved_ref)
elif action == "type":
resolved_ref = await self._resolve_ref(
runtime,
browser_id,
ref,
selector,
action,
required=False,
)
if resolved_ref is None:
result = await runtime.call("keyboard", browser_id, key="", text=text)
else:
result = await runtime.call("type", browser_id, resolved_ref, text)
elif action == "submit":
result = await runtime.call(
"submit",
browser_id,
await self._resolve_ref(runtime, browser_id, ref, selector, action),
)
elif action in {"type_submit", "typesubmit"}:
result = await runtime.call(
"type_submit",
browser_id,
await self._resolve_ref(runtime, browser_id, ref, selector, action),
text,
)
elif action == "scroll":
result = await runtime.call(
"scroll",
browser_id,
await self._resolve_ref(runtime, browser_id, ref, selector, action),
)
elif action == "evaluate":
result = await runtime.call("evaluate", browser_id, script)
elif action in {"key_chord", "keychord"}:
if not keys:
raise ValueError("key_chord requires non-empty 'keys' list")
result = await runtime.call("key_chord", browser_id, keys)
elif action == "hover":
result = await runtime.call(
"hover",
browser_id,
ref=ref,
x=x,
y=y,
offset_x=offset_x,
offset_y=offset_y,
)
elif action == "double_click":
result = await runtime.call(
"double_click",
browser_id,
ref=ref,
x=x,
y=y,
button=button or "left",
modifiers=modifiers,
offset_x=offset_x,
offset_y=offset_y,
)
elif action == "right_click":
result = await runtime.call(
"right_click",
browser_id,
ref=ref,
x=x,
y=y,
modifiers=modifiers,
offset_x=offset_x,
offset_y=offset_y,
)
elif action == "drag":
result = await runtime.call(
"drag",
browser_id,
ref=ref,
target_ref=target_ref,
x=x,
y=y,
to_x=to_x,
to_y=to_y,
offset_x=offset_x,
offset_y=offset_y,
target_offset_x=target_offset_x,
target_offset_y=target_offset_y,
)
elif action == "wheel":
result = await runtime.call(
"wheel",
browser_id,
x,
y,
delta_x,
delta_y,
)
elif action == "keyboard":
result = await runtime.call(
"keyboard",
browser_id,
key=key,
text=text,
)
elif action == "clipboard":
normalized_clipboard_action = clipboard_action or str(
kwargs.get("clipboard_action")
or kwargs.get("operation")
or event_type
or ""
).strip().lower()
result = await runtime.call(
"clipboard",
browser_id,
action=normalized_clipboard_action,
text=text,
)
elif action in {"copy", "cut", "paste"}:
result = await runtime.call(
"clipboard",
browser_id,
action=action,
text=text,
)
elif action == "set_viewport":
result = await runtime.call("set_viewport", browser_id, width, height)
elif action == "select_option":
result = await runtime.call(
"select_option",
browser_id,
await self._resolve_ref(runtime, browser_id, ref, selector, action),
value=value,
values=values,
)
elif action == "set_checked":
result = await runtime.call(
"set_checked",
browser_id,
await self._resolve_ref(runtime, browser_id, ref, selector, action),
checked=True if checked is None else bool(checked),
)
elif action == "upload_file":
result = await runtime.call(
"upload_file",
browser_id,
await self._resolve_ref(runtime, browser_id, ref, selector, action),
path=path,
paths=paths,
)
elif action == "mouse":
result = await runtime.call(
"mouse", browser_id, event_type or "click", x, y,
button=button or "left", modifiers=modifiers,
)
elif action == "multi":
if not calls:
raise ValueError("multi requires non-empty 'calls' list")
result = await runtime.call("multi", list(calls))
elif action == "close":
result = await runtime.call("close_browser", browser_id)
elif action == "close_all":
result = await runtime.call("close_all_browsers")
else:
return Response(
message=f"Unknown browser action: {action}",
break_loop=False,
)
await self._record_history_screenshot(runtime, action, result, browser_id)
except Exception as exc:
return Response(message=f"Browser {action} failed: {exc}", break_loop=False)
return Response(message=self._format_result(action, result), break_loop=False)
def get_log_object(self):
return self.agent.context.log.log(
type="tool",
heading=f"icon://captive_portal {self.agent.agent_name}: Using browser",
content="",
kvps=self.args,
_tool_name=self.name,
)
@staticmethod
def _require_ref(ref: int | str | None) -> int | str:
if ref is None or str(ref).strip() == "":
raise ValueError("ref is required for this browser action")
return ref
@staticmethod
def _has_ref(ref: int | str | None) -> bool:
return ref is not None and str(ref).strip() != ""
@staticmethod
def _has_coordinates(x: float, y: float) -> bool:
return bool(float(x or 0) or float(y or 0))
@classmethod
async def _resolve_ref(
cls,
runtime: Any,
browser_id: int | str | None,
ref: int | str | None,
selector: str = "",
action: str = "action",
*,
required: bool = True,
) -> int | str | None:
if cls._has_ref(ref):
return ref
selector = str(selector or "").strip()
if selector:
content = await runtime.call("content", browser_id, {"selector": selector})
resolved = cls._first_ref_from_content(content, selector)
if resolved is not None:
return resolved
raise ValueError(
f"{action} could not resolve selector {selector!r} to a browser ref"
)
if required:
return cls._require_ref(ref)
return None
@staticmethod
def _first_ref_from_content(content: Any, selector: str = "") -> str | None:
if isinstance(content, dict):
values: list[Any] = []
if selector and selector in content:
values.append(content.get(selector))
values.extend(value for key, value in content.items() if key != selector)
text = "\n".join(str(value or "") for value in values)
else:
text = str(content or "")
match = re.search(r"\[[^\]\n]*?\b(\d+)\]", text)
return match.group(1) if match else None
@staticmethod
def _normalize_keys(keys: list[str] | str | None) -> list[str]:
if keys is None:
return []
if isinstance(keys, str):
raw = re.split(r"\s*\+\s*|\s*,\s*", keys.strip())
elif isinstance(keys, list):
raw = keys
else:
raw = [str(keys)]
aliases = {
"cmd": "Meta",
"command": "Meta",
"control": "Control",
"ctrl": "Control",
"escape": "Escape",
"esc": "Escape",
"meta": "Meta",
"option": "Alt",
"return": "Enter",
"space": "Space",
}
normalized: list[str] = []
for key in raw:
value = str(key or "").strip()
if not value:
continue
normalized.append(aliases.get(value.lower(), value.upper() if len(value) == 1 and value.isalpha() else value))
return normalized
@staticmethod
def _selector_payload(selector: str = "", selectors: list[str] | None = None) -> dict | None:
if selectors:
return {"selectors": selectors}
if selector:
return {"selector": selector}
return None
async def _record_history_screenshot(
self,
runtime: Any,
action: str,
result: Any,
requested_browser_id: int | str | None = None,
) -> None:
if not getattr(self, "log", None):
return
if action in HISTORY_SCREENSHOT_ACTION_DENYLIST:
return
screenshot = result if action == "screenshot" and isinstance(result, dict) else None
if not self._screenshot_has_reference(screenshot):
target_browser_id = self._browser_id_from_result(result) or requested_browser_id
try:
screenshot = await runtime.call(
"screenshot_file",
target_browser_id,
quality=HISTORY_SCREENSHOT_QUALITY,
full_page=False,
path="",
)
except Exception as exc:
PrintStyle.debug(
"Browser history screenshot capture failed:",
f"browser_id={target_browser_id}",
f"quality={HISTORY_SCREENSHOT_QUALITY}",
f"error={exc}",
)
return
if not self._screenshot_has_reference(screenshot):
return
a0_path = str(screenshot.get("a0_path") or "").strip()
local_path = str(screenshot.get("path") or (files.fix_dev_path(a0_path) if a0_path else ""))
state = screenshot.get("state") if isinstance(screenshot.get("state"), dict) else {}
chat_context_id = self._agent_context_id()
browser_context_id = str(screenshot.get("context_id") or state.get("context_id") or "").strip()
snapshot = {
"mime": screenshot.get("mime") or "image/jpeg",
"browser_id": screenshot.get("browser_id") or state.get("id") or requested_browser_id,
"context_id": chat_context_id or browser_context_id,
"browser_context_id": browser_context_id,
}
update_payload: dict[str, Any] = {"browser_snapshot": snapshot}
if local_path:
uri = f"img://{local_path}&t={time.time()}"
snapshot.update(
{
"uri": uri,
"path": local_path,
"a0_path": screenshot.get("a0_path") or files.normalize_a0_path(local_path),
"ephemeral": False,
}
)
update_payload["Screenshot"] = uri
else:
ephemeral_ref = self._screenshot_ephemeral_ref(screenshot)
snapshot.update(
{
"ephemeral": bool(ephemeral_ref),
"ephemeral_ref": ephemeral_ref,
}
)
self.log.update(**update_payload)
def _history_screenshot_path(self, action: str) -> str:
if not getattr(self, "agent", None) or not getattr(self.agent, "context", None):
return ""
context_id = self._agent_context_id()
if not context_id:
return ""
from helpers import persist_chat
token = str(getattr(getattr(self, "log", None), "id", "") or uuid.uuid4())
safe_action = files.safe_file_name(str(action or "browser"))
safe_token = files.safe_file_name(token)
timestamp = time.strftime("%Y%m%d-%H%M%S")
return str(
Path(persist_chat.get_chat_folder_path(context_id))
/ "browser"
/ "screenshots"
/ f"{timestamp}-{safe_action}-{safe_token}.jpg"
)
@staticmethod
def _browser_id_from_result(result: Any) -> Any:
if not isinstance(result, dict):
return None
browsers = result.get("browsers") if isinstance(result.get("browsers"), list) else []
last_interacted_id = result.get("last_interacted_browser_id")
listed_browser = None
if last_interacted_id is not None:
listed_browser = next(
(
browser
for browser in browsers
if isinstance(browser, dict) and str(browser.get("id")) == str(last_interacted_id)
),
None,
)
if listed_browser is None and browsers:
listed_browser = next((browser for browser in browsers if isinstance(browser, dict)), None)
state = result.get("state") if isinstance(result.get("state"), dict) else {}
return (
result.get("id")
or result.get("browser_id")
or state.get("id")
or last_interacted_id
or (listed_browser or {}).get("id")
)
@staticmethod
def _screenshot_has_path(screenshot: Any) -> bool:
return isinstance(screenshot, dict) and bool(screenshot.get("path") or screenshot.get("a0_path"))
@classmethod
def _screenshot_has_reference(cls, screenshot: Any) -> bool:
return cls._screenshot_has_path(screenshot) or bool(cls._screenshot_ephemeral_ref(screenshot))
@staticmethod
def _screenshot_ephemeral_ref(screenshot: Any) -> str:
if not isinstance(screenshot, dict):
return ""
ref = str(screenshot.get("ephemeral_ref") or "").strip()
if ref:
return ref
vision_load = screenshot.get("vision_load")
if isinstance(vision_load, dict):
tool_args = vision_load.get("tool_args")
if isinstance(tool_args, dict):
paths = tool_args.get("paths")
if isinstance(paths, list) and paths:
first = str(paths[0] or "").strip()
if first.startswith("a0-ephemeral-image://"):
return first
return ""
def _agent_context_id(self) -> str:
return str(getattr(getattr(self.agent, "context", None), "id", "") or "").strip()
@staticmethod
def _format_result(action: str, result: Any) -> str:
if action == "content" and isinstance(result, dict):
if set(result.keys()) == {"document"}:
return str(result.get("document") or "")
return json.dumps(result, indent=2, ensure_ascii=False)
return json.dumps(result, indent=2, ensure_ascii=False, default=str)