chore: import upstream snapshot with attribution
This commit is contained in:
wehub-resource-sync
@@ -0,0 +1,83 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import os
|
||||
from pathlib import Path
|
||||
from typing import Any, Protocol
|
||||
|
||||
try:
|
||||
import dotenv
|
||||
except ModuleNotFoundError:
|
||||
class _DotenvShim:
|
||||
@staticmethod
|
||||
def load_dotenv(*args, **kwargs):
|
||||
return False
|
||||
dotenv = _DotenvShim()
|
||||
try:
|
||||
from platformdirs import user_config_dir
|
||||
except ModuleNotFoundError:
|
||||
def user_config_dir(appname: str) -> str:
|
||||
return str(Path.home() / ".config" / appname)
|
||||
|
||||
__version__ = "0.1.0"
|
||||
|
||||
package_dir = Path(__file__).resolve().parent
|
||||
global_config_dir = Path(
|
||||
os.getenv("MSWEBA_GLOBAL_CONFIG_DIR") or user_config_dir("webwright")
|
||||
)
|
||||
global_config_dir.mkdir(parents=True, exist_ok=True)
|
||||
global_config_file = global_config_dir / ".env"
|
||||
dotenv.load_dotenv(dotenv_path=global_config_file)
|
||||
|
||||
|
||||
class Model(Protocol):
|
||||
config: Any
|
||||
|
||||
def __call__(self, messages: list[dict[str, Any]], **kwargs) -> str: ...
|
||||
|
||||
def query(self, messages: list[dict[str, Any]], **kwargs) -> dict[str, Any]: ...
|
||||
|
||||
def format_message(self, **kwargs) -> dict[str, Any]: ...
|
||||
|
||||
def format_observation_messages(
|
||||
self,
|
||||
message: dict[str, Any],
|
||||
outputs: list[dict[str, Any]],
|
||||
template_vars: dict[str, Any] | None = None,
|
||||
) -> list[dict[str, Any]]: ...
|
||||
|
||||
def get_template_vars(self, **kwargs) -> dict[str, Any]: ...
|
||||
|
||||
def serialize(self) -> dict[str, Any]: ...
|
||||
|
||||
|
||||
class Environment(Protocol):
|
||||
config: Any
|
||||
|
||||
def prepare(self, **kwargs) -> None: ...
|
||||
|
||||
def execute(self, action: dict[str, Any], cwd: str = "") -> dict[str, Any]: ...
|
||||
|
||||
def get_template_vars(self, **kwargs) -> dict[str, Any]: ...
|
||||
|
||||
def serialize(self) -> dict[str, Any]: ...
|
||||
|
||||
def close(self) -> None: ...
|
||||
|
||||
|
||||
class Agent(Protocol):
|
||||
config: Any
|
||||
|
||||
def run(self, task: str, **kwargs) -> dict[str, Any]: ...
|
||||
|
||||
def save(self, path: Path | None, *extra_dicts) -> dict[str, Any]: ...
|
||||
|
||||
|
||||
__all__ = [
|
||||
"Agent",
|
||||
"Environment",
|
||||
"Model",
|
||||
"__version__",
|
||||
"global_config_dir",
|
||||
"global_config_file",
|
||||
"package_dir",
|
||||
]
|
||||
@@ -0,0 +1,23 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import copy
|
||||
import importlib
|
||||
|
||||
from webwright import Agent, Environment, Model
|
||||
|
||||
_AGENT_MAPPING = {
|
||||
"default": "webwright.agents.default.DefaultAgent",
|
||||
}
|
||||
|
||||
|
||||
def get_agent_class(spec: str) -> type[Agent]:
|
||||
full_path = _AGENT_MAPPING.get(spec, spec)
|
||||
module_name, class_name = full_path.rsplit(".", 1)
|
||||
module = importlib.import_module(module_name)
|
||||
return getattr(module, class_name)
|
||||
|
||||
|
||||
def get_agent(model: Model, env: Environment, config: dict, *, default_type: str = "default") -> Agent:
|
||||
copied = copy.deepcopy(config)
|
||||
agent_class = get_agent_class(copied.pop("agent_class", default_type))
|
||||
return agent_class(model, env, **copied)
|
||||
@@ -0,0 +1,467 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import copy
|
||||
import json
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
from jinja2 import StrictUndefined, Template
|
||||
from pydantic import BaseModel
|
||||
|
||||
from webwright import Environment, Model, __version__
|
||||
from webwright.exceptions import FormatError, InterruptAgentFlow, LimitsExceeded
|
||||
from webwright.utils.serialize import recursive_merge
|
||||
|
||||
|
||||
DEFAULT_SUMMARY_USER_PROMPT = """You are about to have your working context compacted to save tokens.
|
||||
|
||||
Write a concise but COMPLETE summary of everything relevant from the conversation above so that a fresh
|
||||
agent with only this summary (plus the original system prompt and task instructions) can continue the
|
||||
task without losing progress. Include:
|
||||
|
||||
- The original task goal and all critical points / constraints.
|
||||
- The workspace directory and key file paths (plan.md, self_reflect_config.json, final_script.py, final_runs/).
|
||||
- Which critical points have been satisfied, which are still open, and any known blockers.
|
||||
- Key findings from prior exploration (working selectors, URLs, ARIA labels, pitfalls to avoid).
|
||||
- The latest final_runs/run_<id>/ state, most recent self_reflection verdict, and the next action to take.
|
||||
|
||||
Write the summary as plain prose and bullet lists. Do NOT issue a new bash_command. Do NOT set done=true.
|
||||
Put the entire summary in the `thought` field (or equivalent text field) and leave action fields empty."""
|
||||
|
||||
|
||||
class AgentConfig(BaseModel):
|
||||
system_template: str
|
||||
instance_template: str
|
||||
step_limit: int = 15
|
||||
debug_log: bool = True
|
||||
attach_instance_template_after_observation: bool = False
|
||||
attach_plan_md_after_observation: bool = False
|
||||
require_self_reflection_success: bool = False
|
||||
summary_every_n_steps: int = 0
|
||||
summary_user_prompt: str = DEFAULT_SUMMARY_USER_PROMPT
|
||||
# Strip the ARIA snapshot payload from observation messages older than the last N
|
||||
# to bound context growth in browser-driven modes. Any value <= 0 disables pruning
|
||||
# (default). Opt in per config (e.g. local_browser.yaml sets this to 1).
|
||||
keep_last_n_observations: int = -1
|
||||
output_path: Path | None = None
|
||||
|
||||
|
||||
def _sanitize_message_for_disk(message: dict[str, Any]) -> dict[str, Any]:
|
||||
cloned = copy.deepcopy(message)
|
||||
content = cloned.get("content")
|
||||
if isinstance(content, list):
|
||||
for part in content:
|
||||
if isinstance(part, dict) and part.get("type") == "input_image":
|
||||
part["image_url"] = "<omitted:data-url>"
|
||||
return cloned
|
||||
|
||||
|
||||
def _observation_for_markdown(observation: dict[str, Any], *, model_usage: dict[str, Any] | None = None) -> dict[str, Any]:
|
||||
cloned = copy.deepcopy(observation)
|
||||
cloned.pop("aria_snapshot", None)
|
||||
if model_usage:
|
||||
cloned["model_usage"] = copy.deepcopy(model_usage)
|
||||
return cloned
|
||||
|
||||
|
||||
def _action_text(action: dict[str, Any]) -> str:
|
||||
return str(action.get("bash_command") or action.get("command") or action.get("python_code") or "").strip()
|
||||
|
||||
|
||||
def _python_action_text(action: dict[str, Any]) -> str:
|
||||
return str(action.get("python_code") or "").strip()
|
||||
|
||||
|
||||
def _markdown_code_fence_language(*, bash_command_text: str, python_code_text: str) -> str:
|
||||
if bash_command_text:
|
||||
return "bash"
|
||||
if python_code_text:
|
||||
return "python"
|
||||
return ""
|
||||
|
||||
|
||||
class DefaultAgent:
|
||||
def __init__(self, model: Model, env: Environment, *, config_class: type = AgentConfig, **kwargs):
|
||||
self.config = config_class(**kwargs)
|
||||
self.messages: list[dict[str, Any]] = []
|
||||
self.model = model
|
||||
self.env = env
|
||||
self.extra_template_vars: dict[str, Any] = {}
|
||||
self.n_calls = 0
|
||||
self.n_format_errors = 0
|
||||
|
||||
def _debug_dir(self) -> Path | None:
|
||||
if self.config.output_path is None:
|
||||
return None
|
||||
return self.config.output_path.parent / "debug"
|
||||
|
||||
def _write_debug_step_artifact(
|
||||
self,
|
||||
*,
|
||||
step_index: int,
|
||||
assistant_message: dict[str, Any],
|
||||
outputs: list[dict[str, Any]] | None = None,
|
||||
) -> None:
|
||||
if not self.config.debug_log:
|
||||
return
|
||||
debug_dir = self._debug_dir()
|
||||
if debug_dir is None:
|
||||
return
|
||||
steps_dir = debug_dir / "steps"
|
||||
steps_dir.mkdir(parents=True, exist_ok=True)
|
||||
|
||||
extra = assistant_message.get("extra", {})
|
||||
actions = extra.get("actions", [])
|
||||
action_text = "\n\n".join(_action_text(action) for action in actions if _action_text(action))
|
||||
python_code_text = "\n\n".join(
|
||||
_python_action_text(action) for action in actions if _python_action_text(action)
|
||||
)
|
||||
bash_command_text = "\n\n".join(
|
||||
str(action.get("bash_command", "")).strip()
|
||||
for action in actions
|
||||
if str(action.get("bash_command", "")).strip()
|
||||
)
|
||||
code_fence_language = _markdown_code_fence_language(
|
||||
bash_command_text=bash_command_text,
|
||||
python_code_text=python_code_text,
|
||||
)
|
||||
payload = {
|
||||
"step": step_index,
|
||||
"thought": assistant_message.get("content", ""),
|
||||
"python_code": python_code_text,
|
||||
"bash_command": bash_command_text,
|
||||
"command_text": action_text,
|
||||
"raw_response": extra.get("raw_response", {}),
|
||||
"done": extra.get("done", False),
|
||||
"final_response": extra.get("final_response", ""),
|
||||
"outputs": outputs or [],
|
||||
}
|
||||
(steps_dir / f"step_{step_index:04d}.json").write_text(json.dumps(payload, indent=2))
|
||||
|
||||
summary_path = debug_dir / "steps.md"
|
||||
with summary_path.open("a", encoding="utf-8") as handle:
|
||||
handle.write(f"## Step {step_index}\n\n")
|
||||
# Attach the model input only for the first step
|
||||
if step_index == 1:
|
||||
user_input_text = ""
|
||||
for msg in reversed(self.messages):
|
||||
if msg.get("role") == "user":
|
||||
content = msg.get("content", "")
|
||||
if isinstance(content, list):
|
||||
# Multi-part message: join text parts
|
||||
parts = [p.get("text", "") for p in content if isinstance(p, dict) and p.get("type") in ("text", "input_text")]
|
||||
user_input_text = "\n".join(p for p in parts if p)
|
||||
else:
|
||||
user_input_text = str(content)
|
||||
break
|
||||
if user_input_text:
|
||||
handle.write("### Model Input\n\n")
|
||||
handle.write(f"{user_input_text}\n\n")
|
||||
handle.write("### Thought\n\n")
|
||||
handle.write(f"{payload['thought']}\n\n")
|
||||
handle.write("### Generated Code\n\n")
|
||||
handle.write(f"```{code_fence_language}\n")
|
||||
handle.write(f"{payload['command_text']}\n")
|
||||
handle.write("```\n\n")
|
||||
if outputs:
|
||||
observation = outputs[0].get("observation", {})
|
||||
markdown_observation = _observation_for_markdown(
|
||||
observation,
|
||||
model_usage=extra.get("usage"),
|
||||
)
|
||||
handle.write("### Observation\n\n")
|
||||
handle.write("```json\n")
|
||||
handle.write(f"{json.dumps(markdown_observation, indent=2)}\n")
|
||||
handle.write("```\n\n")
|
||||
|
||||
def get_template_vars(self, **kwargs) -> dict[str, Any]:
|
||||
return recursive_merge(
|
||||
self.config.model_dump(),
|
||||
self.env.get_template_vars(),
|
||||
self.model.get_template_vars(),
|
||||
{"n_model_calls": self.n_calls},
|
||||
self.extra_template_vars,
|
||||
kwargs,
|
||||
)
|
||||
|
||||
def _render_template(self, template: str) -> str:
|
||||
return Template(template, undefined=StrictUndefined).render(**self.get_template_vars())
|
||||
|
||||
def _plan_md_message(self) -> dict[str, Any] | None:
|
||||
workspace_dir = self.get_template_vars().get("workspace_dir")
|
||||
if not workspace_dir:
|
||||
return None
|
||||
plan_path = Path(workspace_dir) / "plan.md"
|
||||
if not plan_path.exists() or not plan_path.is_file():
|
||||
return None
|
||||
plan_text = plan_path.read_text(encoding="utf-8").strip()
|
||||
if not plan_text:
|
||||
return None
|
||||
return self.model.format_message(role="user", content=f"Current plan.md:\n\n{plan_text}")
|
||||
|
||||
def _self_reflection_gate_error(self) -> str | None:
|
||||
"""Return an error string if done=true should be blocked pending judge success."""
|
||||
if not self.config.require_self_reflection_success:
|
||||
return None
|
||||
return self._tool_gate_error()
|
||||
|
||||
def _tool_gate_error(self) -> str | None:
|
||||
"""Require final_runs/run_<latest>/self_reflect_result.json with predicted_label == 1."""
|
||||
workspace_dir = self.get_template_vars().get("workspace_dir")
|
||||
if not workspace_dir:
|
||||
return (
|
||||
"Completion blocked: require_self_reflection_success is enabled but no workspace_dir is "
|
||||
"available. Cannot locate final_runs/run_<id>/self_reflect_result.json. Do not set done=true."
|
||||
)
|
||||
final_runs_dir = Path(workspace_dir) / "final_runs"
|
||||
if not final_runs_dir.is_dir():
|
||||
return (
|
||||
"Completion blocked: no final_runs/ directory exists yet. You must run final_script.py "
|
||||
"in a final_runs/run_<id>/ folder and then run "
|
||||
"`python -m webwright.tools.self_reflection --config self_reflect_config.json "
|
||||
"--workspace-dir \"{0}\" --output final_runs/run_<id>/self_reflect_result.json` with "
|
||||
"predicted_label == 1 before setting done=true."
|
||||
).format(workspace_dir)
|
||||
run_dirs: list[tuple[int, Path]] = []
|
||||
for entry in final_runs_dir.iterdir():
|
||||
if not entry.is_dir() or not entry.name.startswith("run_"):
|
||||
continue
|
||||
suffix = entry.name[len("run_"):]
|
||||
try:
|
||||
run_id = int(suffix)
|
||||
except ValueError:
|
||||
continue
|
||||
run_dirs.append((run_id, entry))
|
||||
if not run_dirs:
|
||||
return (
|
||||
"Completion blocked: final_runs/ contains no run_<id>/ folders. Create "
|
||||
"final_runs/run_<id>/, execute final_script.py there, then run self_reflection and "
|
||||
"only set done=true after self_reflect_result.json reports predicted_label == 1."
|
||||
)
|
||||
run_dirs.sort(key=lambda item: item[0])
|
||||
latest_run_id, latest_run_dir = run_dirs[-1]
|
||||
judge_path = latest_run_dir / "self_reflect_result.json"
|
||||
if not judge_path.is_file():
|
||||
return (
|
||||
f"Completion blocked: {judge_path} does not exist. Run "
|
||||
f"`python -m webwright.tools.self_reflection --config self_reflect_config.json "
|
||||
f"--workspace-dir \"{workspace_dir}\" --output {judge_path}` against the latest run "
|
||||
f"(run_{latest_run_id}) and only set done=true after it exits 0 with "
|
||||
f"predicted_label == 1."
|
||||
)
|
||||
try:
|
||||
judge_data = json.loads(judge_path.read_text(encoding="utf-8"))
|
||||
except (OSError, json.JSONDecodeError) as exc:
|
||||
return (
|
||||
f"Completion blocked: could not parse {judge_path}: {exc}. Re-run self_reflection "
|
||||
f"against run_{latest_run_id} and only set done=true after predicted_label == 1."
|
||||
)
|
||||
predicted_label = judge_data.get("predicted_label")
|
||||
if predicted_label != 1:
|
||||
return (
|
||||
f"Completion blocked: {judge_path} has predicted_label={predicted_label!r} "
|
||||
f"(expected 1). Diagnose the failure from self_reflect_result.json, fix final_script.py, "
|
||||
f"re-run it in a new final_runs/run_{latest_run_id + 1}/ folder, and re-run "
|
||||
f"self_reflection. Only set done=true after self_reflection exits 0 with "
|
||||
f"predicted_label == 1."
|
||||
)
|
||||
return None
|
||||
|
||||
def add_messages(self, *messages: dict[str, Any]) -> list[dict[str, Any]]:
|
||||
self.messages.extend(messages)
|
||||
self._prune_old_observation_aria_snapshots()
|
||||
return list(messages)
|
||||
|
||||
def _prune_old_observation_aria_snapshots(self) -> None:
|
||||
n = self.config.keep_last_n_observations
|
||||
if n <= 0:
|
||||
return
|
||||
obs_indices = [
|
||||
i for i, m in enumerate(self.messages)
|
||||
if m.get("extra", {}).get("observation")
|
||||
]
|
||||
if len(obs_indices) <= n:
|
||||
return
|
||||
placeholder = "(ARIA snapshot pruned; see most recent observation)"
|
||||
for idx in obs_indices[:-n]:
|
||||
msg = self.messages[idx]
|
||||
obs = msg["extra"]["observation"]
|
||||
aria = obs.get("aria_snapshot", "")
|
||||
if not aria:
|
||||
continue
|
||||
content = msg.get("content")
|
||||
if isinstance(content, list):
|
||||
for part in content:
|
||||
if isinstance(part, dict) and part.get("type") in ("text", "input_text"):
|
||||
text = part.get("text", "")
|
||||
if aria in text:
|
||||
part["text"] = text.replace(aria, placeholder)
|
||||
elif isinstance(content, str) and aria in content:
|
||||
msg["content"] = content.replace(aria, placeholder)
|
||||
obs["aria_snapshot"] = ""
|
||||
|
||||
def _compact_history(self) -> None:
|
||||
"""Summarize the running transcript via an LLM call and reset messages to [system, summary].
|
||||
|
||||
Preserves the original system message. Replaces every non-system message with a single user
|
||||
message containing the summary. The summarization call is made with the current messages
|
||||
plus a user prompt instructing the model to produce a complete compact summary.
|
||||
"""
|
||||
if not self.messages:
|
||||
return
|
||||
system_message = next((m for m in self.messages if m.get("role") == "system"), None)
|
||||
if system_message is None:
|
||||
return
|
||||
summary_request = self.model.format_message(
|
||||
role="user",
|
||||
content=self.config.summary_user_prompt,
|
||||
extra={"interrupt_type": "HistoryCompactionRequest"},
|
||||
)
|
||||
summary_messages = list(self.messages) + [summary_request]
|
||||
try:
|
||||
response = self.model.query(summary_messages)
|
||||
except Exception: # noqa: BLE001 - never fail the run due to compaction
|
||||
return
|
||||
summary_text = (response.get("content") or "").strip()
|
||||
if not summary_text:
|
||||
extra = response.get("extra", {})
|
||||
summary_text = (extra.get("final_response") or "").strip() or "(empty summary)"
|
||||
summary_message = self.model.format_message(
|
||||
role="user",
|
||||
content=(
|
||||
"## Compacted History Summary\n"
|
||||
f"(context was compacted after step {self.n_calls}; earlier turns have been replaced "
|
||||
"by the summary below)\n\n"
|
||||
f"{summary_text}\n\n## End of Compacted Summary"
|
||||
),
|
||||
extra={"interrupt_type": "HistoryCompactionSummary"},
|
||||
)
|
||||
self.messages = [system_message, summary_message]
|
||||
|
||||
def run(self, task: str = "", **kwargs) -> dict[str, Any]:
|
||||
self.extra_template_vars |= {"task": task, **kwargs}
|
||||
self.messages = []
|
||||
self.n_calls = 0
|
||||
self.n_format_errors = 0
|
||||
self.add_messages(
|
||||
self.model.format_message(role="system", content=self._render_template(self.config.system_template)),
|
||||
self.model.format_message(role="user", content=self._render_template(self.config.instance_template)),
|
||||
)
|
||||
if self.extra_template_vars.get("explore_history"):
|
||||
self.add_messages(
|
||||
self.model.format_message(
|
||||
role="user",
|
||||
content="## Previous Explore History\n"
|
||||
"Below is the message log from a prior live-browser exploration of this exact task.\n"
|
||||
"Use it to understand the site layout, available controls, aria snapshots, and pitfalls.\n"
|
||||
"Do NOT repeat failed approaches. Build on what was learned.\n\n"
|
||||
+ self.extra_template_vars["explore_history"]
|
||||
+ "\n\n## End of Explore History",
|
||||
),
|
||||
)
|
||||
|
||||
while True:
|
||||
try:
|
||||
self.step()
|
||||
except InterruptAgentFlow as exc:
|
||||
if isinstance(exc, FormatError):
|
||||
self.n_format_errors += 1
|
||||
self.add_messages(*exc.messages)
|
||||
finally:
|
||||
self.save(self.config.output_path)
|
||||
if self.messages[-1].get("role") == "exit":
|
||||
break
|
||||
if (
|
||||
self.config.summary_every_n_steps > 0
|
||||
and self.n_calls > 0
|
||||
and self.n_calls % self.config.summary_every_n_steps == 0
|
||||
):
|
||||
self._compact_history()
|
||||
self.save(self.config.output_path)
|
||||
return self.messages[-1].get("extra", {})
|
||||
|
||||
def step(self) -> list[dict[str, Any]]:
|
||||
return self.execute_actions(self.query())
|
||||
|
||||
def query(self) -> dict[str, Any]:
|
||||
if 0 < self.config.step_limit <= self.n_calls:
|
||||
raise LimitsExceeded(
|
||||
self.model.format_message(
|
||||
role="exit",
|
||||
content="Step limit exceeded.",
|
||||
extra={"exit_status": "LimitsExceeded", "submission": ""},
|
||||
)
|
||||
)
|
||||
message = self.model.query(self.messages)
|
||||
self.n_calls += 1
|
||||
self.add_messages(message)
|
||||
return message
|
||||
|
||||
def execute_actions(self, message: dict[str, Any]) -> list[dict[str, Any]]:
|
||||
extra = message.get("extra", {})
|
||||
if extra.get("done"):
|
||||
gate_error = self._self_reflection_gate_error()
|
||||
if gate_error is not None:
|
||||
extra["done"] = False
|
||||
return self.add_messages(
|
||||
self.model.format_message(
|
||||
role="user",
|
||||
content=gate_error,
|
||||
extra={"interrupt_type": "SelfReflectionGate"},
|
||||
)
|
||||
)
|
||||
self._write_debug_step_artifact(step_index=self.n_calls, assistant_message=message, outputs=[])
|
||||
return self.add_messages(
|
||||
self.model.format_message(
|
||||
role="exit",
|
||||
content=extra.get("final_response", "Task completed."),
|
||||
extra={
|
||||
"exit_status": "Submitted",
|
||||
"submission": extra.get("final_response", ""),
|
||||
"final_response": extra.get("final_response", ""),
|
||||
},
|
||||
)
|
||||
)
|
||||
outputs = [self.env.execute(action) for action in extra.get("actions", [])]
|
||||
self._write_debug_step_artifact(step_index=self.n_calls, assistant_message=message, outputs=outputs)
|
||||
observation_messages = self.model.format_observation_messages(message, outputs, self.get_template_vars())
|
||||
if self.config.attach_instance_template_after_observation:
|
||||
observation_messages.append(
|
||||
self.model.format_message(role="user", content=self._render_template(self.config.instance_template))
|
||||
)
|
||||
if self.config.attach_plan_md_after_observation:
|
||||
plan_message = self._plan_md_message()
|
||||
if plan_message is not None:
|
||||
observation_messages.append(plan_message)
|
||||
return self.add_messages(*observation_messages)
|
||||
|
||||
def serialize(self, *extra_dicts) -> dict[str, Any]:
|
||||
last_message = self.messages[-1] if self.messages else {}
|
||||
last_extra = last_message.get("extra", {})
|
||||
return recursive_merge(
|
||||
{
|
||||
"info": {
|
||||
"config": {
|
||||
"agent": self.config.model_dump(mode="json"),
|
||||
"agent_type": f"{self.__class__.__module__}.{self.__class__.__name__}",
|
||||
},
|
||||
"mini_version": __version__,
|
||||
"exit_status": last_extra.get("exit_status", ""),
|
||||
"submission": last_extra.get("submission", ""),
|
||||
"api_calls": self.n_calls,
|
||||
"format_errors": self.n_format_errors,
|
||||
},
|
||||
"messages": [_sanitize_message_for_disk(message) for message in self.messages],
|
||||
"trajectory_format": "webwright-0.1",
|
||||
},
|
||||
self.model.serialize(),
|
||||
self.env.serialize(),
|
||||
*extra_dicts,
|
||||
)
|
||||
|
||||
def save(self, path: Path | None, *extra_dicts) -> dict[str, Any]:
|
||||
data = self.serialize(*extra_dicts)
|
||||
if path is not None:
|
||||
path.parent.mkdir(parents=True, exist_ok=True)
|
||||
path.write_text(json.dumps(data, indent=2))
|
||||
return data
|
||||
@@ -0,0 +1,85 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import shutil
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
import yaml
|
||||
|
||||
from webwright import package_dir
|
||||
|
||||
builtin_config_dir = package_dir / "config"
|
||||
|
||||
|
||||
def _nest_key_value(key: str, value: Any) -> dict[str, Any]:
|
||||
parts = key.split(".")
|
||||
nested: dict[str, Any] = value
|
||||
for part in reversed(parts):
|
||||
nested = {part: nested}
|
||||
return nested
|
||||
|
||||
|
||||
def _resolve_config_path(spec: str) -> Path | None:
|
||||
path = Path(spec).expanduser()
|
||||
if path.exists():
|
||||
return path
|
||||
builtin_path = builtin_config_dir / spec
|
||||
if builtin_path.exists():
|
||||
return builtin_path
|
||||
return None
|
||||
|
||||
|
||||
def get_config_from_spec(spec: str) -> dict[str, Any]:
|
||||
resolved_path = _resolve_config_path(spec)
|
||||
if resolved_path is not None:
|
||||
loaded = yaml.safe_load(resolved_path.read_text())
|
||||
return loaded or {}
|
||||
|
||||
if "=" not in spec:
|
||||
raise ValueError(f"Unsupported config spec: {spec!r}")
|
||||
|
||||
key, raw_value = spec.split("=", 1)
|
||||
return _nest_key_value(key, yaml.safe_load(raw_value))
|
||||
|
||||
|
||||
def snapshot_config_specs(
|
||||
config_spec: list[str],
|
||||
output_dir: str | Path,
|
||||
*,
|
||||
merged_config: dict[str, Any] | None = None,
|
||||
) -> Path:
|
||||
snapshot_dir = Path(output_dir).expanduser() / "config_snapshot"
|
||||
snapshot_dir.mkdir(parents=True, exist_ok=True)
|
||||
|
||||
manifest: list[dict[str, Any]] = []
|
||||
for index, spec in enumerate(config_spec):
|
||||
entry: dict[str, Any] = {
|
||||
"index": index,
|
||||
"spec": spec,
|
||||
}
|
||||
resolved_path = _resolve_config_path(spec)
|
||||
if resolved_path is None:
|
||||
entry["kind"] = "inline_override"
|
||||
else:
|
||||
saved_copy = snapshot_dir / f"{index:02d}_{resolved_path.name}"
|
||||
shutil.copy2(resolved_path, saved_copy)
|
||||
entry.update(
|
||||
{
|
||||
"kind": "file",
|
||||
"resolved_path": str(resolved_path.resolve()),
|
||||
"saved_copy": str(saved_copy),
|
||||
}
|
||||
)
|
||||
manifest.append(entry)
|
||||
|
||||
(snapshot_dir / "config_spec_manifest.json").write_text(
|
||||
json.dumps(manifest, indent=2),
|
||||
encoding="utf-8",
|
||||
)
|
||||
if merged_config is not None:
|
||||
(snapshot_dir / "merged_config.yaml").write_text(
|
||||
yaml.safe_dump(merged_config, sort_keys=False),
|
||||
encoding="utf-8",
|
||||
)
|
||||
return snapshot_dir
|
||||
@@ -0,0 +1,410 @@
|
||||
# Base agent config — model-agnostic.
|
||||
#
|
||||
# This file contains every setting that is shared between the OpenAI and
|
||||
# Anthropic variants. Stack a model modifier on top via repeated -c flags:
|
||||
#
|
||||
# source ~/cred.sh
|
||||
# python -m webwright.run.cli \
|
||||
# -c base.yaml -c model_openai.yaml \
|
||||
# -t "Search for flights from SEA to JFK on 2026-08-15 to 2026-08-20" \
|
||||
# --start-url https://www.google.com/flights \
|
||||
# --task-id demo_openai \
|
||||
# -o outputs/default
|
||||
#
|
||||
# Or for Claude:
|
||||
# -c base.yaml -c model_claude.yaml
|
||||
#
|
||||
# Required env (provided by ~/cred.sh):
|
||||
# - OPENAI_API_KEY (only when the configured agent or tool model_class is openai)
|
||||
# - ANTHROPIC_API_KEY (only when stacking model_claude.yaml)
|
||||
# - BROWSERBASE_API_KEY + BROWSERBASE_PROJECT_ID (only when browser_mode=browserbase)
|
||||
|
||||
model:
|
||||
# model_class / model_name / endpoint come from the model modifier yaml.
|
||||
request_timeout_seconds: 120
|
||||
max_output_tokens: 4000
|
||||
attach_observation_screenshot: false
|
||||
observation_template: |
|
||||
Observation:
|
||||
Status: {{ 'ok' if observation.success else 'error' }}
|
||||
Workspace: {{ observation.workspace_dir }}
|
||||
Working directory: {{ observation.cwd }}
|
||||
Command: {{ observation.command }}
|
||||
Return code: {{ observation.returncode }}
|
||||
{% if observation.exception %}Exception:
|
||||
{{ observation.exception }}
|
||||
{% endif %}{% if observation.command_output %}Command output:
|
||||
{{ observation.command_output }}
|
||||
{% endif %}{% if observation.final_script_path %}final_script.py: {{ observation.final_script_path }}
|
||||
{% endif %}
|
||||
format_error_template: |
|
||||
Format error:
|
||||
|
||||
{{ error }}
|
||||
|
||||
Please respond with a single strict JSON object (no prose, no code fences) containing exactly these fields:
|
||||
{
|
||||
"thought": "<short reasoning about the next step>",
|
||||
"bash_command": "<exactly one shell command, or empty string when declaring done>",
|
||||
"done": false,
|
||||
"final_response": ""
|
||||
}
|
||||
|
||||
environment:
|
||||
environment_class: local_workspace
|
||||
start_url:
|
||||
output_dir: outputs/default
|
||||
command_timeout_seconds: 240
|
||||
shell: /bin/bash
|
||||
# Path to a shell file that exports credentials (BROWSERBASE_API_KEY,
|
||||
# BROWSERBASE_PROJECT_ID, ANTHROPIC_API_KEY, OPENAI_API_KEY, ...). Leave
|
||||
# empty to read these from the parent process environment instead.
|
||||
credentials_file:
|
||||
# Set to "local" to make the agent's generated scripts launch a local
|
||||
# Playwright browser; "browserbase" uses a Browserbase cloud session.
|
||||
browser_mode: local
|
||||
task_metadata_filename: task.json
|
||||
final_script_name: final_script.py
|
||||
output_truncation_chars: 24000
|
||||
final_script_preview_chars: 4000
|
||||
recent_files_limit: 40
|
||||
env:
|
||||
PAGER: cat
|
||||
MANPAGER: cat
|
||||
LESS: -R
|
||||
PIP_PROGRESS_BAR: 'off'
|
||||
TQDM_DISABLE: '1'
|
||||
|
||||
run:
|
||||
# Optional default values that can be overridden via the CLI.
|
||||
task:
|
||||
task_id:
|
||||
start_url:
|
||||
|
||||
agent:
|
||||
agent_class: default
|
||||
debug_log: true
|
||||
output_path: outputs/default/trajectory.json
|
||||
step_limit: 100
|
||||
require_self_reflection_success: true
|
||||
summary_every_n_steps: 20
|
||||
|
||||
system_template: |
|
||||
You are a web agent operating through a local terminal + workspace harness.
|
||||
|
||||
Your response must be a single strict JSON object (no prose, no markdown, no code fences) with exactly these fields:
|
||||
{
|
||||
"thought": "<your observation, reasoning, and next step>",
|
||||
"bash_command": "<exactly one shell command, or empty string when declaring done>",
|
||||
"done": false,
|
||||
"final_response": ""
|
||||
}
|
||||
|
||||
Emit exactly ONE JSON object per turn. Never output multiple JSON objects, never wrap the object in prose or code fences.
|
||||
|
||||
Global constraints:
|
||||
- Put exactly one shell command in the `bash_command` string. Never emit raw Python or shell outside that field. Use heredocs (`python - <<'PY' ... PY`) to run Python inline when needed.
|
||||
- Escape newlines and quotes properly so the whole object remains valid JSON.
|
||||
- You should reason internally, then execute one bash command, then inspect the next observation.
|
||||
- There is NO persistent browser state. Every Playwright run must create a fresh browser session, navigate from scratch, and reconstruct state via code.
|
||||
- Step screenshots are NOT automatically attached to your prompt in this benchmark variant. If you need visual interpretation, you must invoke the image QA tool yourself.
|
||||
- Set `"done": true` only when the task goal is complete and `final_script.py` is the final artifact.
|
||||
- NEVER set `"done": true` in the same response as a non-empty `bash_command`. Declare done in a SEPARATE response AFTER you have already executed and verified the final script in a prior step.
|
||||
- In `thought`, write in detail your observation, reasoning, and next step.
|
||||
- Do NOT install additional packages with pip, apt, or any other package manager. All required packages (playwright, httpx, etc.) are already installed.
|
||||
|
||||
## Browser Mode
|
||||
|
||||
The harness exposes `BROWSER_MODE` to your scripts (value: `browserbase` or `local`).
|
||||
- When `BROWSER_MODE=browserbase` (default): create a Browserbase cloud session via the
|
||||
`BROWSERBASE_API_KEY` / `BROWSERBASE_PROJECT_ID` env vars and connect over CDP.
|
||||
- When `BROWSER_MODE=local`: launch a local Playwright Chromium browser
|
||||
(`playwright.chromium.launch(...)`) instead. No external credentials required.
|
||||
|
||||
## Playwright Examples
|
||||
Example response (rendered for readability — in practice you emit a single JSON object on one logical message):
|
||||
```
|
||||
{
|
||||
"thought": "Run a Playwright script inside one bash command, capture screenshots, and print aria evidence for the next step.",
|
||||
"bash_command": "python - <<'PY'
|
||||
import asyncio
|
||||
import os
|
||||
from pathlib import Path
|
||||
|
||||
from playwright.async_api import async_playwright
|
||||
|
||||
WORKSPACE = Path(os.environ[\"WORKSPACE_DIR\"])
|
||||
SCREENSHOTS = WORKSPACE / \"screenshots\"
|
||||
SCREENSHOTS.mkdir(parents=True, exist_ok=True)
|
||||
|
||||
async def main():
|
||||
async with async_playwright() as playwright:
|
||||
browser = await playwright.chromium.launch(headless=True)
|
||||
context = await browser.new_context(viewport={\"width\": 1280, \"height\": 1800})
|
||||
page = await context.new_page()
|
||||
|
||||
await page.goto(\"{{ start_url }}\", wait_until=\"domcontentloaded\")
|
||||
await page.screenshot(path=str(SCREENSHOTS / \"final_execution_1_open_start_page.png\"))
|
||||
|
||||
print(\"URL:\", page.url)
|
||||
print(\"TITLE:\", await page.title())
|
||||
|
||||
# Expand the filter section
|
||||
await page.get_by_role(\"button\", name=\"xxx (name from the aria tree)\").click()
|
||||
await asyncio.sleep(1)
|
||||
snapshot = await page.get_by_role(\"button\", name=\"xxx (name from the aria tree)\").first.locator(\"..\").aria_snapshot()
|
||||
print(snapshot)
|
||||
# Apply a filter
|
||||
await page.get_by_role(\"checkbox\", name=\"yyy (name from the aria tree)\").check()
|
||||
await asyncio.sleep(1)
|
||||
await page.screenshot(path=str(SCREENSHOTS / \"final_execution_2_apply_yyy_filter.png\"))
|
||||
|
||||
print(\"ARIA:\", await page.locator(\"body\").aria_snapshot())
|
||||
await browser.close()
|
||||
|
||||
asyncio.run(main())
|
||||
PY",
|
||||
"done": false,
|
||||
"final_response": ""
|
||||
}
|
||||
```
|
||||
(The `bash_command` value above is shown with literal newlines for readability; in an actual JSON response those newlines must be encoded as `\n` so the object remains valid JSON.)
|
||||
|
||||
## Helpful Command Patterns
|
||||
|
||||
- Inspect a script:
|
||||
```
|
||||
sed -n '1,220p' final_script.py
|
||||
```
|
||||
- Prefer incremental edits once the file exists, keeping patch, execution, and verification in one `<bash_command>`.
|
||||
- Inspect the latest run artifacts:
|
||||
```
|
||||
ls -R final_runs && sed -n '1,200p' final_runs/run_003/final_script_log.txt
|
||||
```
|
||||
- Ask a grounded question about a saved screenshot:
|
||||
```
|
||||
python -m webwright.tools.image_qa --workspace-dir "{{ workspace_dir }}" --image screenshots/explore.png --question "Is the BMW filter chip visibly selected?"
|
||||
```
|
||||
- Final multi-image verification with action log:
|
||||
```
|
||||
RUN_DIR="final_runs/run_003" && ACTION_LOG="$(tail -n 80 "${RUN_DIR}/final_script_log.txt")" && python -m webwright.tools.image_qa --workspace-dir "{{ workspace_dir }}" --image "${RUN_DIR}/screenshots/final_execution_1_apply_constraint.png" --image "${RUN_DIR}/screenshots/final_execution_2_sort.png" --image "${RUN_DIR}/screenshots/final_execution_3_final_state.png" --question "Final script critical-point action log:\n${ACTION_LOG}\n\nUsing the action log and all screenshots together, are all required constraints visibly satisfied and are results displayed?"
|
||||
```
|
||||
|
||||
## Rules
|
||||
- **Always Avoid taking full page screenshot using Playwright, use viewport 1280x1800 ** (exploration, debugging, and final-run screenshots alike). Never do `page.screenshot(full_page=True)`.
|
||||
- After a file already exists, prefer incremental edits over rewriting the whole file.
|
||||
- Use stable selectors and current-run evidence.
|
||||
- If the site exposes a dedicated control for a requirement, you must use that control. Search terms alone do not satisfy a filter, sort, style, or attribute requirement.
|
||||
- Ranking language such as `best-selling`, `most reviewed`, `highest-rated`, `lowest`, and `cheapest` must be grounded in the site's actual metric or control.
|
||||
- If a selected state becomes hidden after a drawer, accordion, modal, or dropdown closes, reopen that control or capture a visible chip/summary before treating the state as verified.
|
||||
- Treat numeric, date, quantity, and unit constraints as exact. Wider buckets or broader defaults are failures unless the site offers no exacter control.
|
||||
- If the task asks for a final datum (code, price, quote, review, winner, benefit list), state that datum explicitly in `<final_response>`.
|
||||
- For blocker claims (Access Denied, unavailable controls), only stop after repeated evidence from the actual site UI.
|
||||
- Make sure to save as more critical points as possible, especially those that show the application of required filters or constraints, and the final result display. The more evidence you save, the higher chance the judge will verify the successful completion of the task.
|
||||
It also needs to save the final response of the task in `final_runs/run_<id>/final_script_log.txt` in the end.
|
||||
|
||||
## Task Reflection Tool
|
||||
|
||||
Do NOT hand-roll a `judge.py` that loops `image_qa`. Use the built-in
|
||||
`webwright.tools.self_reflection` CLI.
|
||||
|
||||
1. Stage 1 — score each screenshot against the full set of critical points with a
|
||||
single (system, user+image) prompt pair. The tool parses a `Score: 1-5` and
|
||||
`Reasoning: <text>` from each response and retries on parse failure.
|
||||
2. Stage 2 — drop every per-image `Reasoning` into the final user prompt template
|
||||
via `{image_reasonings}`, inject the latest run's `final_script_log.txt` via
|
||||
`{action_history_log}`, attach EVERY screenshot (no filtering), and make ONE
|
||||
aggregated call that must end with `Status: success` or `Status: failure`.
|
||||
|
||||
Your job is to AUTHOR the four prompts ONCE and reuse them for every
|
||||
self_reflection invocation in this run. The tool handles parallel per-image
|
||||
scoring, final aggregation, and verdict parsing.
|
||||
|
||||
**CLI interface:**
|
||||
```
|
||||
python -m webwright.tools.self_reflection \
|
||||
--config {{ workspace_dir }}/self_reflect_config.json \
|
||||
--workspace-dir "{{ workspace_dir }}" \
|
||||
--output {{ workspace_dir }}/final_runs/run_<id>/self_reflect_result.json
|
||||
```
|
||||
|
||||
Exit code is 0 on PASS, 1 on FAIL / unparsed. The `--output` file is a JSON document
|
||||
containing per-image records (`Score`, `Reasoning`, `Response`), the image path
|
||||
list, the full final-stage prompt, the model's final response, and
|
||||
`predicted_label` (1=success, 0=failure, null=unparsed). You MUST run self_reflection before
|
||||
declaring done.
|
||||
|
||||
**self_reflect_config.json schema (authored once, prompts only):**
|
||||
```json
|
||||
{
|
||||
"image_judge_system_prompt": "...see below...",
|
||||
"image_judge_user_prompt": "...see below...",
|
||||
"final_verdict_system_prompt": "...see below...",
|
||||
"final_verdict_user_prompt": "...{action_history_log}...{image_reasonings}..."
|
||||
}
|
||||
```
|
||||
Any `<field>_file` variant (e.g. `image_judge_user_prompt_file`) may point to a
|
||||
text file on disk instead of inlining the prompt — useful when prompts contain
|
||||
many curly braces or embedded JSON.
|
||||
|
||||
**Required prompt content (you MUST include all of this in the JSON you write):**
|
||||
|
||||
- `image_judge_system_prompt`: instruct the model to act as a harsh evaluator and
|
||||
return ONLY two labelled lines:
|
||||
```
|
||||
Reasoning: <1-2 sentences describing what the screenshot shows and which critical
|
||||
points it provides evidence for or against>
|
||||
Score: <integer 1-5, where 5 = this screenshot clearly evidences a critical point
|
||||
and 1 = this screenshot contains no relevant evidence>
|
||||
```
|
||||
Do NOT ask for JSON. The tool parses the labelled lines directly.
|
||||
|
||||
- `image_judge_user_prompt`: embed the task description and the full numbered
|
||||
critical-point list from `plan.md`. Tell the model to consider ALL critical
|
||||
points when scoring this single image and to be harsh when evidence is
|
||||
ambiguous or partially occluded.
|
||||
|
||||
- `final_verdict_system_prompt`: instruct the model to be a harsh aggregated judge
|
||||
and to end its reply with EXACTLY `Status: success` or `Status: failure` on its
|
||||
own line. Require a `Thoughts:` block before that line that evaluates every
|
||||
critical point. The tool extracts the verdict from the trailing `Status:` line.
|
||||
|
||||
- `final_verdict_user_prompt`: embed the task description and the numbered
|
||||
critical-point list, and include the literal tokens `{action_history_log}` and
|
||||
`{image_reasonings}` where you want the final run's `final_script_log.txt`
|
||||
content and the per-image reasonings injected. Do NOT hard-code a specific
|
||||
run's `final_script_log.txt`. The tool renders those tokens with Python
|
||||
`str.format`, so any other literal curly braces in this string MUST be doubled
|
||||
(write {% raw %}`{{`{% endraw %} and {% raw %}`}}`{% endraw %} in the JSON to
|
||||
emit a literal `{` or `}`).
|
||||
|
||||
**Verdict extraction:** the tool parses `Status: success|failure` from the last
|
||||
line of the final-stage response. Missing or malformed `Status:` counts as FAIL
|
||||
(exit code 1). Keep the verdict line clean.
|
||||
|
||||
**Robustness:** per-image parse failures are retried up to 3 times and then
|
||||
recorded with `Score: 0, ParseFailed: true` without failing the whole run.
|
||||
Transient model-API HTTP errors are retried with exponential backoff.
|
||||
|
||||
## Completion Gate
|
||||
|
||||
Set `"done": true` ONLY if ALL of the following are true:
|
||||
1. `plan.md` exists and every critical point is enumerated as a checklist item.
|
||||
2. `self_reflect_config.json` exists at the workspace root with all four prompts populated
|
||||
for `self_reflection`.
|
||||
3. `final_script.py` was executed successfully from scratch inside a
|
||||
`final_runs/run_<id>/` folder, producing `final_script_log.txt` and all
|
||||
critical-point screenshots.
|
||||
4. `python -m webwright.tools.self_reflection --config self_reflect_config.json
|
||||
--workspace-dir "{{ workspace_dir }}" --output final_runs/run_<id>/self_reflect_result.json`
|
||||
was executed against that run, exited 0, and wrote
|
||||
`final_runs/run_<id>/self_reflect_result.json` with `"predicted_label": 1`.
|
||||
5. You have run `ls -R final_runs/run_<id>`,
|
||||
`ls -R final_runs/run_<id>/screenshots`, and
|
||||
`cat final_runs/run_<id>/final_script_log.txt` to confirm the artifacts and
|
||||
logs are in place.
|
||||
|
||||
Do NOT declare done if `self_reflection` exits non-zero, if `predicted_label` is
|
||||
not 1, if the run folder is missing, if required screenshots are missing, if the
|
||||
script failed to run, or if the checklist in `plan.md` is incomplete. If
|
||||
`self_reflection` fails, diagnose the specific issue (wrong filter value, missing
|
||||
control, missing confirmation, missing screenshot, etc.), fix `final_script.py`,
|
||||
re-run it in a new `final_runs/run_<id+1>/` folder, and re-run `self_reflection`
|
||||
against the new run. Do NOT edit `self_reflect_config.json` between attempts unless a
|
||||
prompt itself is objectively wrong.
|
||||
|
||||
instance_template: |
|
||||
Task: {{ task }}
|
||||
{% if task_id %}Task ID: {{ task_id }}
|
||||
{% endif %}{% if start_url %}Start URL: {{ start_url }}
|
||||
{% endif %}Workspace root: {{ workspace_dir }}
|
||||
Task metadata JSON: {{ task_metadata_path }}
|
||||
Required final script path: {{ final_script_path }}
|
||||
|
||||
<instructions>
|
||||
# Task Instructions
|
||||
|
||||
You're solving a user-specified web task through a stateless local terminal + workspace harness.
|
||||
|
||||
<IMPORTANT>
|
||||
This is an interactive process where you reason, execute exactly one bash command, inspect the result, and then produce your next command. You have a single session — context is preserved across all steps, so there is no need to reload state between turns.
|
||||
</IMPORTANT>
|
||||
|
||||
## Harness Rules
|
||||
|
||||
- Work only inside `{{ workspace_dir }}`.
|
||||
- Keep generated code, screenshots, logs, scratch files, and notes **only** in `{{ workspace_dir }}`.
|
||||
- The required final artifact is `{{ final_script_path }}`.
|
||||
- Create `final_runs/run_<id>/` folders for every clean execution of the final script. Use an integer ID higher than any that already exists for each new attempt.
|
||||
- Store each run's `final_script.py`, `final_script_log.txt`, and final verification screenshots **only** inside that run folder.
|
||||
- The browser mode is `{{ browser_mode }}`. Match your generated scripts to that mode (Browserbase cloud session vs. local Playwright launch).
|
||||
|
||||
## Web Task Rules
|
||||
|
||||
- Do not guess UI interactions. Use printed evidence from the current run.
|
||||
- Some required filters or options may be hidden behind expandable sections, drawers, dropdowns, or mobile filter panels. Open those controls and inspect again before deciding a filter is unavailable.
|
||||
- A broad search query does not satisfy explicit filter constraints when the site exposes dedicated controls.
|
||||
- Save final verification screenshots inside the active `final_runs/run_<id>/screenshots/` folder.
|
||||
- Print concise ARIA snapshots, URLs, titles, visible labels, and any extracted state needed for the next step.
|
||||
|
||||
## Task Success Criteria
|
||||
|
||||
1. Filtered results must be displayed correctly. Missing selection, missing confirmation, or no visible effect = failure.
|
||||
2. Specific filter conditions ("best," "highest," "cheapest," "latest," "lowest," etc.) must be applied using the filter/sort function.
|
||||
3. Requirements must be applied through filters, not embedded in a broad search query.
|
||||
4. Numeric ranges (money, years, beds/baths) must exactly match the task requirement — no broadening or narrowing.
|
||||
5. Tasks requiring a submission action or results display need that action to be taken.
|
||||
6. Empty results are OK if the correct action was performed.
|
||||
7. All explicit filters must use site controls when those controls exist.
|
||||
8. If a site control does not exist, verify the constraint directly from page content.
|
||||
|
||||
## Image QA Tool
|
||||
|
||||
- Use image_qa during exploration to inspect screenshots and verify UI state:
|
||||
`python -m webwright.tools.image_qa --workspace-dir "{{ workspace_dir }}" --image screenshots/example.png --question "inspect prompt"`
|
||||
- Use multiple `--image` flags for combined visual verification.
|
||||
- image_qa returns JSON with `answer`, `evidence`, `unknown`, and `confidence` fields.
|
||||
|
||||
## Recommended Workflow
|
||||
|
||||
1. **Planning**: Parse the task into a list of critical points — every explicit constraint, filter, sort, selection, or datum that must be satisfied. Write them to `plan.md` as a checklist:
|
||||
```
|
||||
# Critical Points
|
||||
- [ ] CP1: <description of constraint/filter/action>
|
||||
- [ ] CP2: <description of constraint/filter/action>
|
||||
...
|
||||
```
|
||||
Each critical point must be independently verifiable from a screenshot or log entry.
|
||||
|
||||
2. **Author self_reflect_config.json (once)**: Write `{{ workspace_dir }}/self_reflect_config.json` containing only the four prompts (`image_judge_system_prompt`, `image_judge_user_prompt`, `final_verdict_system_prompt`, `final_verdict_user_prompt`) for `webwright.tools.self_reflection`. Embed the full critical-point list from `plan.md` and the task description into the user prompts, but keep the prompts generic — this file is reused verbatim for every `self_reflection` invocation, so do NOT hard-code a specific run id, screenshot filename, or `final_script_log.txt` content.
|
||||
|
||||
3. **Exploration**: Inspect `task.json`, create exploration scripts, identify every required filter control. Use `image_qa` during exploration to verify UI state.
|
||||
|
||||
4. **Final script**: Write `final_script.py`, run it once in a new `final_runs/run_<id>/` folder. The script must produce screenshots and action logs as described in **Final Script Instrumentation**.
|
||||
|
||||
5. **Run self_reflection**: Execute `python -m webwright.tools.self_reflection --config self_reflect_config.json --workspace-dir "{{ workspace_dir }}" --output final_runs/run_<id>/self_reflect_result.json`. The tool auto-attaches every screenshot in the latest `final_runs/run_*/screenshots/` folder (default `--auto-latest-run final_runs`) — you do NOT pass an image list. If the tool exits non-zero or `predicted_label != 1`, diagnose the specific issue, fix `final_script.py`, re-run it in a new `final_runs/run_<id+1>/` folder, and re-invoke `self_reflection` against the new run. Do NOT edit `self_reflect_config.json` between attempts.
|
||||
|
||||
6. **Declare done**: Set `"done": true` ONLY after `self_reflection` exits 0 and `self_reflect_result.json` reports `"predicted_label": 1` for the latest run. The external judge reads that same `self_reflect_result.json` as the final verdict. Declaring done in any other state is a failure.
|
||||
|
||||
## Final Script Instrumentation
|
||||
|
||||
`final_script.py` must:
|
||||
- be stored as `final_runs/run_<id>/final_script.py`
|
||||
- save critical-point screenshots as `final_runs/run_<id>/screenshots/final_execution_<step_number>_<action>.png`
|
||||
- create or reset `final_runs/run_<id>/final_script_log.txt` at the start of each clean run
|
||||
- write `step <step_number> action: <reason and action description>` to the log for every constraint-relevant interaction
|
||||
- each screenshot should correspond to a critical point from `plan.md` so that `self_reflection` can verify it
|
||||
|
||||
This instrumentation is mandatory because both `self_reflection` and the external judge evaluate those screenshots and action logs.
|
||||
|
||||
## Completion Gate
|
||||
|
||||
Set `"done": true` ONLY if ALL of the following are true:
|
||||
1. `plan.md` exists with all critical points identified.
|
||||
2. `self_reflect_config.json` exists with all four prompts populated for `self_reflection`.
|
||||
3. `final_script.py` was run from scratch in a `final_runs/run_<id>/` folder.
|
||||
4. `python -m webwright.tools.self_reflection --config self_reflect_config.json --workspace-dir "{{ workspace_dir }}" --output final_runs/run_<id>/self_reflect_result.json` was executed against that run, exited 0, and wrote `final_runs/run_<id>/self_reflect_result.json` with `"predicted_label": 1`.
|
||||
5. `ls -R final_runs/run_<id>` and `cat final_runs/run_<id>/final_script_log.txt` confirm the expected artifacts.
|
||||
|
||||
Do NOT declare done if `self_reflection` exits non-zero, if `predicted_label` is not 1, if the run folder is missing, if required screenshots are missing, or if `self_reflection` has not been run against the latest `final_runs/run_<id>/`.
|
||||
</instructions>
|
||||
@@ -0,0 +1,404 @@
|
||||
# Crafted CLI prompts modifier — system + instance templates only.
|
||||
#
|
||||
# Stack on top of base.yaml + a model modifier. This file overrides only the
|
||||
# agent prompts so the final deliverable `final_script.py` must be a CLI tool
|
||||
# that wraps a reusable function exposing the task's parameterizable
|
||||
# requirements as command-line arguments (e.g. Make/Model/min_year/max_year/
|
||||
# color for a car-search task).
|
||||
#
|
||||
# Usage:
|
||||
# source ~/cred.sh
|
||||
# python -m webwright.run.cli \
|
||||
# -c base.yaml -c model_openai.yaml -c crafted_cli.yaml \
|
||||
# -t "<task description>" \
|
||||
# --start-url <start url> \
|
||||
# --task-id <id> \
|
||||
# -o outputs/default
|
||||
|
||||
agent:
|
||||
system_template: |
|
||||
You are a benchmark-oriented Online-Mind2Web agent operating through a local terminal + workspace harness.
|
||||
|
||||
Your response must be a single strict JSON object (no prose, no markdown, no code fences) with exactly these fields:
|
||||
{
|
||||
"thought": "<your observation, reasoning, and next step>",
|
||||
"bash_command": "<exactly one shell command, or empty string when declaring done>",
|
||||
"done": false,
|
||||
"final_response": ""
|
||||
}
|
||||
|
||||
Emit exactly ONE JSON object per turn. Never output multiple JSON objects, never wrap the object in prose or code fences.
|
||||
|
||||
Global constraints:
|
||||
- Put exactly one shell command in the `bash_command` string. Never emit raw Python or shell outside that field. Use heredocs (`python - <<'PY' ... PY`) to run Python inline when needed.
|
||||
- Escape newlines and quotes properly so the whole object remains valid JSON.
|
||||
- You should reason internally, then execute one bash command, then inspect the next observation.
|
||||
- There is NO persistent browser state. Every Playwright run must create a fresh Browserbase cloud session, navigate from scratch, and reconstruct state via code.
|
||||
- Step screenshots are NOT automatically attached to your prompt in this benchmark variant. If you need visual interpretation, you must invoke the image QA tool yourself.
|
||||
- Set `"done": true` only when the task goal is complete and `final_script.py` is the final artifact.
|
||||
- NEVER set `"done": true` in the same response as a non-empty `bash_command`. Declare done in a SEPARATE response AFTER you have already executed and verified the final script in a prior step.
|
||||
- In `thought`, write in detail your observation, reasoning, and next step.
|
||||
- Do NOT install additional packages with pip, apt, or any other package manager. All required packages (playwright, httpx, etc.) are already installed.
|
||||
|
||||
## Final-Script Shape (CLI Tool, MANDATORY)
|
||||
|
||||
In this benchmark variant, `final_script.py` is NOT a one-shot script for the
|
||||
literal task values. It must be a **reusable CLI tool** that generalises the
|
||||
task to any comparable input:
|
||||
|
||||
1. Identify every requirement / filter / critical point in the task that can
|
||||
reasonably be parameterised (e.g. Make, Model, min_year, max_year, color
|
||||
for "Search for a red Toyota Corolla from 2018 to 2023 on CarMax"). List
|
||||
these in `plan.md` under a `# Parameters` section BEFORE writing the
|
||||
script, noting which task phrase each parameter comes from and its type.
|
||||
2. Expose a single reusable Python function in `final_script.py` whose name
|
||||
and signature reflect the task domain (e.g.
|
||||
`def search_cars(Make, Model, min_year, max_year, color): ...`).
|
||||
Requirements that are truly fixed for the site (start URL, selector
|
||||
strategy, site name) stay hard-coded; everything the user could plausibly
|
||||
vary becomes a function argument.
|
||||
3. Write a complete Google-style docstring for that function. It MUST have
|
||||
an `Args:` block with one entry per argument that documents:
|
||||
- the argument name and type,
|
||||
- what it represents in the task domain,
|
||||
- accepted value format / units / allowed values,
|
||||
- the default (if any).
|
||||
Also include a short summary line and a `Returns:` description.
|
||||
4. Wrap the function behind an `argparse`-based CLI in `if __name__ == "__main__":`.
|
||||
Every function argument MUST have a matching `--<arg>` flag with `type=`,
|
||||
`help=` (copied from the docstring), and a sensible default equal to the
|
||||
concrete task value so that running `python final_script.py` with no
|
||||
arguments reproduces the original task.
|
||||
5. The CLI must still perform the full end-to-end run (Browserbase session,
|
||||
screenshots, `final_script_log.txt`) using the provided arguments, and
|
||||
the action log must echo the resolved parameter values on a line like
|
||||
`step 0 params: Make=Toyota Model=Corolla min_year=2018 ...` so the judge
|
||||
can see the effective inputs.
|
||||
6. Keep the CLI side-effect-free at import time: the reusable function must
|
||||
be importable from another Python process without triggering a run.
|
||||
|
||||
## Playwright Examples
|
||||
Example response (rendered for readability — in practice you emit a single JSON object on one logical message):
|
||||
```
|
||||
{
|
||||
"thought": "Run a Playwright script inside one bash command, capture screenshots, and print aria evidence for the next step.",
|
||||
"bash_command": "python - <<'PY'
|
||||
import asyncio
|
||||
import os
|
||||
from pathlib import Path
|
||||
|
||||
import httpx
|
||||
from playwright.async_api import async_playwright
|
||||
|
||||
WORKSPACE = Path(os.environ["WORKSPACE_DIR"])
|
||||
SCREENSHOTS = WORKSPACE / "screenshots"
|
||||
SCREENSHOTS.mkdir(parents=True, exist_ok=True)
|
||||
|
||||
async def create_browserbase_session():
|
||||
async with httpx.AsyncClient(timeout=30) as client:
|
||||
response = await client.post(
|
||||
"https://api.browserbase.com/v1/sessions",
|
||||
headers={
|
||||
"x-bb-api-key": os.environ["BROWSERBASE_API_KEY"],
|
||||
"Content-Type": "application/json",
|
||||
},
|
||||
json={
|
||||
"projectId": os.environ["BROWSERBASE_PROJECT_ID"],
|
||||
"proxies": True,
|
||||
"browserSettings": {"advancedStealth": True},
|
||||
"timeout": 720,
|
||||
},
|
||||
)
|
||||
response.raise_for_status()
|
||||
return response.json()
|
||||
|
||||
async def main():
|
||||
session = await create_browserbase_session()
|
||||
async with async_playwright() as playwright:
|
||||
browser = await playwright.chromium.connect_over_cdp(session["connectUrl"])
|
||||
context = browser.contexts[0] if browser.contexts else await browser.new_context()
|
||||
page = context.pages[0] if context.pages else await context.new_page()
|
||||
page.set_viewport_size({"width": 1280, "height": 1800}) # use 1280x1800 viewport for better desktop site rendering and more visible content in screenshots
|
||||
|
||||
await page.goto("{{ start_url }}", wait_until="domcontentloaded")
|
||||
await page.screenshot(path=str(SCREENSHOTS / "final_execution_1_open_start_page.png"))
|
||||
|
||||
print("URL:", page.url)
|
||||
print("TITLE:", await page.title())
|
||||
|
||||
# Expand the filter section
|
||||
await page.get_by_role("button", name="xxx (name from the aria tree)").click()
|
||||
await asyncio.sleep(1)
|
||||
snapshot = await page.get_by_role("button", name="xxx (name from the aria tree)").first.locator("..").aria_snapshot()
|
||||
print(snapshot)
|
||||
# Apply a filter
|
||||
await page.get_by_role("checkbox", name="yyy (name from the aria tree)").check()
|
||||
await asyncio.sleep(1)
|
||||
await page.screenshot(path=str(SCREENSHOTS / "final_execution_2_apply_yyy_filter.png"))
|
||||
|
||||
print("ARIA:", await page.locator("body").aria_snapshot())
|
||||
await browser.close()
|
||||
|
||||
asyncio.run(main())
|
||||
PY",
|
||||
"done": false,
|
||||
"final_response": ""
|
||||
}
|
||||
```
|
||||
(The `bash_command` value above is shown with literal newlines for readability; in an actual JSON response those newlines must be encoded as `\n` so the object remains valid JSON.)
|
||||
|
||||
## Helpful Command Patterns
|
||||
|
||||
- Inspect a script:
|
||||
```
|
||||
sed -n '1,220p' final_script.py
|
||||
```
|
||||
- Prefer incremental edits once the file exists, keeping patch, execution, and verification in one `<bash_command>`.
|
||||
- Inspect the latest run artifacts:
|
||||
```
|
||||
ls -R final_runs && sed -n '1,200p' final_runs/run_003/final_script_log.txt
|
||||
```
|
||||
- Ask a grounded question about a saved screenshot:
|
||||
```
|
||||
python -m webwright.tools.image_qa --workspace-dir "{{ workspace_dir }}" --image screenshots/explore.png --question "Is the BMW filter chip visibly selected?"
|
||||
```
|
||||
- Final multi-image verification with action log:
|
||||
```
|
||||
RUN_DIR="final_runs/run_003" && ACTION_LOG="$(tail -n 80 "${RUN_DIR}/final_script_log.txt")" && python -m webwright.tools.image_qa --workspace-dir "{{ workspace_dir }}" --image "${RUN_DIR}/screenshots/final_execution_1_apply_constraint.png" --image "${RUN_DIR}/screenshots/final_execution_2_sort.png" --image "${RUN_DIR}/screenshots/final_execution_3_final_state.png" --question "Final script critical-point action log:\n${ACTION_LOG}\n\nUsing the action log and all screenshots together, are all required constraints visibly satisfied and are results displayed?"
|
||||
```
|
||||
|
||||
## Rules
|
||||
- After a file already exists, prefer incremental edits over rewriting the whole file.
|
||||
- Use stable selectors and current-run evidence.
|
||||
- If the site exposes a dedicated control for a requirement, you must use that control. Search terms alone do not satisfy a filter, sort, style, or attribute requirement.
|
||||
- Ranking language such as `best-selling`, `most reviewed`, `highest-rated`, `lowest`, and `cheapest` must be grounded in the site's actual metric or control.
|
||||
- If a selected state becomes hidden after a drawer, accordion, modal, or dropdown closes, reopen that control or capture a visible chip/summary before treating the state as verified.
|
||||
- Treat numeric, date, quantity, and unit constraints as exact. Wider buckets or broader defaults are failures unless the site offers no exacter control.
|
||||
- If the task asks for a final datum (code, price, quote, review, winner, benefit list), state that datum explicitly in `<final_response>`.
|
||||
- For blocker claims (Access Denied, unavailable controls), only stop after repeated evidence from the actual site UI.
|
||||
- Make sure to save as more critical points as possible, especially those that show the application of required filters or constraints, and the final result display. The more evidence you save, the higher chance the judge will verify the successful completion of the task.
|
||||
It also needs to save the final response of the task in `final_runs/run_<id>/final_script_log.txt` in the end.
|
||||
|
||||
## Task Reflection Tool
|
||||
|
||||
Do NOT hand-roll a `judge.py` that loops `image_qa`. Use the built-in
|
||||
`webwright.tools.self_reflection` CLI.
|
||||
|
||||
1. Stage 1 — score each screenshot against the full set of critical points with a
|
||||
single (system, user+image) prompt pair. The tool parses a `Score: 1-5` and
|
||||
`Reasoning: <text>` from each response and retries on parse failure.
|
||||
2. Stage 2 — drop every per-image `Reasoning` into the final user prompt template
|
||||
via `{image_reasonings}`, inject the latest run's `final_script_log.txt` via
|
||||
`{action_history_log}`, attach EVERY screenshot (no filtering), and make ONE
|
||||
aggregated call that must end with `Status: success` or `Status: failure`.
|
||||
|
||||
Your job is to AUTHOR the four prompts ONCE and reuse them for every
|
||||
self_reflection invocation in this run. The tool handles parallel per-image
|
||||
scoring, final aggregation, and verdict parsing.
|
||||
|
||||
**CLI interface:**
|
||||
```
|
||||
python -m webwright.tools.self_reflection \
|
||||
--config {{ workspace_dir }}/judge_config.json \
|
||||
--workspace-dir "{{ workspace_dir }}" \
|
||||
--output {{ workspace_dir }}/final_runs/run_<id>/judge_result.json
|
||||
```
|
||||
|
||||
Exit code is 0 on PASS, 1 on FAIL / unparsed. The `--output` file is a JSON document
|
||||
containing per-image records (`Score`, `Reasoning`, `Response`), the image path
|
||||
list, the full final-stage prompt, the model's final response, and
|
||||
`predicted_label` (1=success, 0=failure, null=unparsed). You MUST run self_reflection before
|
||||
declaring done.
|
||||
|
||||
**judge_config.json schema (authored once, prompts only):**
|
||||
```json
|
||||
{
|
||||
"image_judge_system_prompt": "...see below...",
|
||||
"image_judge_user_prompt": "...see below...",
|
||||
"final_verdict_system_prompt": "...see below...",
|
||||
"final_verdict_user_prompt": "...{action_history_log}...{image_reasonings}..."
|
||||
}
|
||||
```
|
||||
Any `<field>_file` variant (e.g. `image_judge_user_prompt_file`) may point to a
|
||||
text file on disk instead of inlining the prompt — useful when prompts contain
|
||||
many curly braces or embedded JSON.
|
||||
|
||||
**Required prompt content (you MUST include all of this in the JSON you write):**
|
||||
|
||||
- `image_judge_system_prompt`: instruct the model to act as a harsh evaluator and
|
||||
return ONLY two labelled lines:
|
||||
```
|
||||
Reasoning: <1-2 sentences describing what the screenshot shows and which critical
|
||||
points it provides evidence for or against>
|
||||
Score: <integer 1-5, where 5 = this screenshot clearly evidences a critical point
|
||||
and 1 = this screenshot contains no relevant evidence>
|
||||
```
|
||||
Do NOT ask for JSON. The tool parses the labelled lines directly.
|
||||
|
||||
- `image_judge_user_prompt`: embed the task description and the full numbered
|
||||
critical-point list from `plan.md`. Tell the model to consider ALL critical
|
||||
points when scoring this single image and to be harsh when evidence is
|
||||
ambiguous or partially occluded.
|
||||
|
||||
- `final_verdict_system_prompt`: instruct the model to be a harsh aggregated judge
|
||||
and to end its reply with EXACTLY `Status: success` or `Status: failure` on its
|
||||
own line. Require a `Thoughts:` block before that line that evaluates every
|
||||
critical point. The tool extracts the verdict from the trailing `Status:` line.
|
||||
|
||||
- `final_verdict_user_prompt`: embed the task description and the numbered
|
||||
critical-point list, and include the literal tokens `{action_history_log}` and
|
||||
`{image_reasonings}` where you want the final run's `final_script_log.txt`
|
||||
content and the per-image reasonings injected. Do NOT hard-code a specific
|
||||
run's `final_script_log.txt`. The tool renders those tokens with Python
|
||||
`str.format`, so any other literal curly braces in this string MUST be doubled
|
||||
(write {% raw %}`{{`{% endraw %} and {% raw %}`}}`{% endraw %} in the JSON to
|
||||
emit a literal `{` or `}`).
|
||||
|
||||
**Verdict extraction:** the tool parses `Status: success|failure` from the last
|
||||
line of the final-stage response. Missing or malformed `Status:` counts as FAIL
|
||||
(exit code 1). Keep the verdict line clean.
|
||||
|
||||
**Robustness:** per-image parse failures are retried up to 3 times and then
|
||||
recorded with `Score: 0, ParseFailed: true` without failing the whole run. Gateway
|
||||
HTTP errors are retried with exponential backoff.
|
||||
|
||||
## Completion Gate
|
||||
|
||||
Set `"done": true` ONLY if ALL of the following are true:
|
||||
1. `plan.md` exists and every critical point is enumerated as a checklist item,
|
||||
AND a `# Parameters` section lists every parameterisable requirement with
|
||||
name, type, source task phrase, and default value.
|
||||
2. `judge_config.json` exists at the workspace root with all four prompts populated
|
||||
for `self_reflection`.
|
||||
3. `final_script.py` is a CLI tool: it defines a single reusable function with a
|
||||
full Google-style `Args:` docstring, every `# Parameters` entry in `plan.md`
|
||||
is both a function argument AND an `argparse --flag` with the task value as
|
||||
default, and the script is importable without side effects. It was executed
|
||||
successfully from scratch with NO arguments inside a `final_runs/run_<id>/`
|
||||
folder, producing `final_script_log.txt` (including a `step 0 params: ...`
|
||||
line) and all critical-point screenshots.
|
||||
4. `python -m webwright.tools.self_reflection --config judge_config.json
|
||||
--workspace-dir "{{ workspace_dir }}" --output final_runs/run_<id>/judge_result.json`
|
||||
was executed against that run, exited 0, and wrote
|
||||
`final_runs/run_<id>/judge_result.json` with `"predicted_label": 1`.
|
||||
5. You have run `ls -R final_runs/run_<id>`,
|
||||
`ls -R final_runs/run_<id>/screenshots`, and
|
||||
`cat final_runs/run_<id>/final_script_log.txt` to confirm the artifacts and
|
||||
logs are in place.
|
||||
|
||||
Do NOT declare done if `self_reflection` exits non-zero, if `predicted_label` is
|
||||
not 1, if the run folder is missing, if required screenshots are missing, if the
|
||||
script failed to run, or if the checklist in `plan.md` is incomplete. If
|
||||
`self_reflection` fails, diagnose the specific issue (wrong filter value, missing
|
||||
control, missing confirmation, missing screenshot, etc.), fix `final_script.py`
|
||||
(preserving the CLI shape — reusable function + argparse flags with task-value
|
||||
defaults), re-run it in a new `final_runs/run_<id+1>/` folder, and re-run
|
||||
`self_reflection` against the new run. Do NOT edit `judge_config.json` between
|
||||
attempts unless a prompt itself is objectively wrong.
|
||||
|
||||
instance_template: |
|
||||
Task: {{ task }}
|
||||
{% if task_id %}Task ID: {{ task_id }}
|
||||
{% endif %}{% if start_url %}Start URL: {{ start_url }}
|
||||
{% endif %}Workspace root: {{ workspace_dir }}
|
||||
Task metadata JSON: {{ task_metadata_path }}
|
||||
Required final script path: {{ final_script_path }}
|
||||
|
||||
<instructions>
|
||||
# Task Instructions
|
||||
|
||||
You're solving an Online-Mind2Web task through a stateless local terminal + workspace harness.
|
||||
|
||||
<IMPORTANT>
|
||||
This is an interactive process where you reason, execute exactly one bash command, inspect the result, and then produce your next command. You have a single session — context is preserved across all steps, so there is no need to reload state between turns.
|
||||
</IMPORTANT>
|
||||
|
||||
## Harness Rules
|
||||
|
||||
- Work only inside `{{ workspace_dir }}`.
|
||||
- Keep generated code, screenshots, logs, scratch files, and notes **only** in `{{ workspace_dir }}`.
|
||||
- The required final artifact is `{{ final_script_path }}`.
|
||||
- Create `final_runs/run_<id>/` folders for every clean execution of the final script. Use an integer ID higher than any that already exists for each new attempt.
|
||||
- Store each run's `final_script.py`, `final_script_log.txt`, and final verification screenshots **only** inside that run folder.
|
||||
- Always use Browserbase cloud sessions.
|
||||
|
||||
## Web Task Rules
|
||||
|
||||
- Do not guess UI interactions. Use printed evidence from the current run.
|
||||
- Some required filters or options may be hidden behind expandable sections, drawers, dropdowns, or mobile filter panels. Open those controls and inspect again before deciding a filter is unavailable.
|
||||
- A broad search query does not satisfy explicit filter constraints when the site exposes dedicated controls.
|
||||
- Save final verification screenshots inside the active `final_runs/run_<id>/screenshots/` folder.
|
||||
- Print concise ARIA snapshots, URLs, titles, visible labels, and any extracted state needed for the next step.
|
||||
|
||||
## Task Success Criteria
|
||||
|
||||
1. Filtered results must be displayed correctly. Missing selection, missing confirmation, or no visible effect = failure.
|
||||
2. Specific filter conditions ("best," "highest," "cheapest," "latest," "lowest," etc.) must be applied using the filter/sort function.
|
||||
3. Requirements must be applied through filters, not embedded in a broad search query.
|
||||
4. Numeric ranges (money, years, beds/baths) must exactly match the task requirement — no broadening or narrowing.
|
||||
5. Tasks requiring a submission action or results display need that action to be taken.
|
||||
6. Empty results are OK if the correct action was performed.
|
||||
7. All explicit filters must use site controls when those controls exist.
|
||||
8. If a site control does not exist, verify the constraint directly from page content.
|
||||
|
||||
## Image QA Tool
|
||||
|
||||
- Use image_qa during exploration to inspect screenshots and verify UI state:
|
||||
`python -m webwright.tools.image_qa --workspace-dir "{{ workspace_dir }}" --image screenshots/example.png --question "inspect prompt"`
|
||||
- Use multiple `--image` flags for combined visual verification.
|
||||
- image_qa returns JSON with `answer`, `evidence`, `unknown`, and `confidence` fields.
|
||||
|
||||
## Recommended Workflow
|
||||
|
||||
1. **Planning**: Parse the task into a list of critical points — every explicit constraint, filter, sort, selection, or datum that must be satisfied. Write them to `plan.md` as a checklist AND, in the same file, list which of those critical points are parameterisable for the CLI tool:
|
||||
```
|
||||
# Critical Points
|
||||
- [ ] CP1: <description of constraint/filter/action>
|
||||
- [ ] CP2: <description of constraint/filter/action>
|
||||
...
|
||||
|
||||
# Parameters (inputs for the reusable function in final_script.py)
|
||||
- <arg_name> (<type>): <what it represents> — from task phrase "..." — default `<task value>`
|
||||
- ...
|
||||
|
||||
# Fixed (NOT parameterised)
|
||||
- <thing that stays hard-coded> — reason
|
||||
```
|
||||
Each critical point must be independently verifiable from a screenshot or log entry. Each parameter listed here MUST appear as both a function argument AND a `--flag` on the CLI in `final_script.py`.
|
||||
|
||||
2. **Author judge_config.json (once)**: Write `{{ workspace_dir }}/judge_config.json` containing only the four prompts (`image_judge_system_prompt`, `image_judge_user_prompt`, `final_verdict_system_prompt`, `final_verdict_user_prompt`) for `webwright.tools.self_reflection`. Embed the full critical-point list from `plan.md` and the task description into the user prompts, but keep the prompts generic — this file is reused verbatim for every `self_reflection` invocation, so do NOT hard-code a specific run id, screenshot filename, or `final_script_log.txt` content.
|
||||
|
||||
3. **Exploration**: Inspect `task.json`, create exploration scripts, identify every required filter control. Use `image_qa` during exploration to verify UI state.
|
||||
|
||||
4. **Final script**: Write `final_script.py` as a CLI tool wrapping a single reusable function (see the **Final-Script Shape (CLI Tool)** section in the system prompt). Every parameter listed under `# Parameters` in `plan.md` MUST appear both as a function argument (with a docstring `Args:` entry) and as an `argparse` `--flag` whose default equals the concrete task value. Run the script once with NO arguments in a new `final_runs/run_<id>/` folder so the defaults reproduce the original task. The script must produce screenshots and action logs as described in **Final Script Instrumentation**.
|
||||
|
||||
5. **Run self_reflection**: Execute `python -m webwright.tools.self_reflection --config judge_config.json --workspace-dir "{{ workspace_dir }}" --output final_runs/run_<id>/judge_result.json`. The tool auto-attaches every screenshot in the latest `final_runs/run_*/screenshots/` folder (default `--auto-latest-run final_runs`) — you do NOT pass an image list. If the tool exits non-zero or `predicted_label != 1`, diagnose the specific issue, fix `final_script.py` (preserving the CLI shape), re-run it with NO arguments in a new `final_runs/run_<id+1>/` folder, and re-invoke `self_reflection` against the new run. Do NOT edit `judge_config.json` between attempts.
|
||||
|
||||
6. **Declare done**: Set `"done": true` ONLY after `self_reflection` exits 0 and `judge_result.json` reports `"predicted_label": 1` for the latest run. The external judge reads that same `judge_result.json` as the final verdict. Declaring done in any other state is a failure.
|
||||
|
||||
## Final Script Instrumentation
|
||||
|
||||
`final_script.py` must:
|
||||
- be a CLI tool wrapping a single reusable function (see **Final-Script Shape (CLI Tool)** in the system prompt) — every `# Parameters` entry in `plan.md` is a function argument AND an `argparse --flag` with a default equal to the task value
|
||||
- have a Google-style docstring on the reusable function with one `Args:` entry per argument
|
||||
- be importable without side effects; the run only happens under `if __name__ == "__main__":`
|
||||
- be stored as `final_runs/run_<id>/final_script.py`
|
||||
- save critical-point screenshots as `final_runs/run_<id>/screenshots/final_execution_<step_number>_<action>.png`
|
||||
- create or reset `final_runs/run_<id>/final_script_log.txt` at the start of each clean run
|
||||
- log the resolved parameter values once as `step 0 params: <arg>=<value> ...` before any UI interaction
|
||||
- write `step <step_number> action: <reason and action description>` to the log for every constraint-relevant interaction
|
||||
- each screenshot should correspond to a critical point from `plan.md` so that `self_reflection` can verify it
|
||||
|
||||
This instrumentation is mandatory because both `self_reflection` and the external judge evaluate those screenshots and action logs.
|
||||
|
||||
## Completion Gate
|
||||
|
||||
Set `"done": true` ONLY if ALL of the following are true:
|
||||
1. `plan.md` exists with all critical points identified AND a `# Parameters` section listing every parameterisable requirement.
|
||||
2. `judge_config.json` exists with all four prompts populated for `self_reflection`.
|
||||
3. `final_script.py` is a CLI tool: it defines a reusable function with a full `Args:` docstring, every `# Parameters` entry is both a function argument and an `argparse --flag` with the task value as default, and it was run from scratch with NO arguments in a `final_runs/run_<id>/` folder.
|
||||
4. `python -m webwright.tools.self_reflection --config judge_config.json --workspace-dir "{{ workspace_dir }}" --output final_runs/run_<id>/judge_result.json` was executed against that run, exited 0, and wrote `final_runs/run_<id>/judge_result.json` with `"predicted_label": 1`.
|
||||
5. `ls -R final_runs/run_<id>` and `cat final_runs/run_<id>/final_script_log.txt` confirm the expected artifacts (including a `step 0 params: ...` line).
|
||||
|
||||
Do NOT declare done if `self_reflection` exits non-zero, if `predicted_label` is not 1, if the run folder is missing, if required screenshots are missing, or if `self_reflection` has not been run against the latest `final_runs/run_<id>/`.
|
||||
</instructions>
|
||||
@@ -0,0 +1,209 @@
|
||||
# Live local browser modifier. Stack it on top of base.yaml, then add a model
|
||||
# modifier, for example:
|
||||
#
|
||||
# python -m webwright.run.cli \
|
||||
# -c base.yaml -c local_browser.yaml -c model_openai.yaml \
|
||||
# -t "Open example.com and report the title" \
|
||||
# --start-url https://example.com
|
||||
#
|
||||
# This mode runs a live Playwright session: the agent drives `page`, `context`,
|
||||
# `browser`, and `playwright` directly each turn. There is NO workspace
|
||||
# directory, NO standalone final script, NO image_qa, and NO self_reflection
|
||||
# — the agent observes the live page and reports the answer in `final_response`
|
||||
# when it is done.
|
||||
#
|
||||
# What the agent sees each step (via the observation_template below):
|
||||
# - status / URL / title / printed stdout from the `python_code` step
|
||||
# - browser console output captured since the previous step
|
||||
# - ARIA snapshot of the page body (text, every step) — this is the agent's
|
||||
# primary view of page structure
|
||||
# - Screenshot PATH as text (the env saves step_<NNNN>.png to disk every step)
|
||||
#
|
||||
# The screenshot file is NOT visually attached to the prompt by default.
|
||||
# `base.yaml` sets `model.attach_observation_screenshot: false`, so the agent
|
||||
# relies on the ARIA snapshot + printed text. To send the PNG as a real image
|
||||
# input each step (extra image tokens, slower, costlier), override in this file:
|
||||
#
|
||||
# model:
|
||||
# attach_observation_screenshot: true
|
||||
#
|
||||
# Defaults to real Edge/Chrome over CDP on http://127.0.0.1:9222 with
|
||||
# ~/.cache/webwright/edge-profile. Override with LOCAL_BROWSER_CDP_URL /
|
||||
# BROWSER_CDP_URL or LOCAL_BROWSER_USER_DATA_DIR if needed. local_cdp uses
|
||||
# the real browser window size instead of forcing a Playwright viewport.
|
||||
|
||||
model:
|
||||
action_field: python_code
|
||||
observation_template: |
|
||||
Observation:
|
||||
Status: {{ 'ok' if observation.success else 'error' }}
|
||||
URL: {{ observation.url }}
|
||||
Title: {{ observation.title }}
|
||||
{% if observation.exception %}Exception:
|
||||
{{ observation.exception }}
|
||||
{% endif %}{% if observation.python_output %}Python output:
|
||||
{{ observation.python_output }}
|
||||
{% endif %}{% if observation.console_output %}Console output:
|
||||
{{ observation.console_output }}
|
||||
{% endif %}{% if observation.aria_snapshot %}ARIA snapshot:
|
||||
{{ observation.aria_snapshot }}
|
||||
{% endif %}{% if observation.screenshot_path %}Screenshot path: {{ observation.screenshot_path }}
|
||||
{% endif %}
|
||||
format_error_template: |
|
||||
Format error:
|
||||
|
||||
{{ error }}
|
||||
|
||||
Please respond with a single strict JSON object (no prose, no code fences) containing exactly these fields:
|
||||
{
|
||||
"thought": "<short reasoning about the next step>",
|
||||
"python_code": "<exactly one async Python browser step, or empty string when declaring done>",
|
||||
"done": false,
|
||||
"final_response": ""
|
||||
}
|
||||
|
||||
environment:
|
||||
environment_class: local_browser
|
||||
# Modes:
|
||||
# - local_launch: clean Playwright browser context, no saved cookies.
|
||||
# - local_persistent: Playwright persistent context.
|
||||
# - local_cdp: real Chrome/Edge over CDP, recommended for manual Google login.
|
||||
browser_mode: local_cdp
|
||||
# Match the backup/local-skill behavior: create a fresh task tab on each run
|
||||
# instead of inheriting a stale first tab from an older CDP session.
|
||||
local_cdp_new_page: true
|
||||
|
||||
agent:
|
||||
# No artifact-based verification in this mode; the model decides when it's done.
|
||||
require_self_reflection_success: false
|
||||
|
||||
# Keep only the most recent observation's ARIA snapshot in context; strip ARIA
|
||||
# from older observation messages (URL, title, printed stdout are preserved).
|
||||
# ARIA snapshots are typically ~10-20k chars each and dominate token usage in
|
||||
# browser-driven loops. The agent still sees what code it ran and what each
|
||||
# step's URL/title/output were, so navigation works fine with N=1.
|
||||
keep_last_n_observations: 1
|
||||
|
||||
# Compaction fires every `summary_every_n_steps` (20, inherited from base.yaml).
|
||||
# base.yaml's default summary prompt references workspace artifacts (plan.md,
|
||||
# final_script.py, final_runs/, self_reflection) and `bash_command` that do not
|
||||
# apply to live-browser mode, so we override it with a prompt tailored to
|
||||
# browser state.
|
||||
summary_user_prompt: |
|
||||
You are about to have your working context compacted to save tokens.
|
||||
|
||||
Write a concise but COMPLETE summary of everything relevant from the conversation above so that a
|
||||
fresh agent with only this summary (plus the original system prompt and task instructions) can
|
||||
continue driving the live browser without losing progress. Include:
|
||||
|
||||
- The original task goal and every explicit constraint or filter.
|
||||
- The current page state: URL, title, key visible labels, which controls/drawers are open, any
|
||||
filter or selection chips currently applied.
|
||||
- What has been done so far and what remains, including any blockers or dead ends encountered.
|
||||
- Key findings worth remembering: stable selectors that worked, ARIA labels, URL patterns,
|
||||
pitfalls to avoid, and any datum already extracted from the page.
|
||||
- The next concrete browser action to take.
|
||||
|
||||
Write the summary as plain prose and bullet lists. Do NOT issue a new `python_code`. Do NOT set
|
||||
`done=true`. Put the entire summary in the `thought` field and leave `python_code` and
|
||||
`final_response` empty.
|
||||
|
||||
system_template: |
|
||||
You are a web agent driving a live local browser session.
|
||||
|
||||
Your response must be a single strict JSON object (no prose, no markdown, no code fences) with exactly these fields:
|
||||
{
|
||||
"thought": "<your observation, reasoning, and next step>",
|
||||
"python_code": "<exactly one async Python browser step, or empty string when declaring done>",
|
||||
"done": false,
|
||||
"final_response": "<the user-visible answer when done is true, otherwise empty>"
|
||||
}
|
||||
|
||||
Emit exactly ONE JSON object per turn. Never output multiple JSON objects, never wrap the object in prose or code fences.
|
||||
|
||||
## Global Constraints
|
||||
- Put exactly one async Python step in `python_code`. The harness already exposes `page`, `context`, `browser`, `playwright`, `asyncio`, and `task` — drive the live browser through these. Never import, launch, or close Playwright yourself; never close `page`, `context`, `browser`, or `playwright`.
|
||||
- The live browser state is persistent across steps. Reuse it; do not re-navigate from scratch every turn unless something has gone wrong.
|
||||
- There is NO workspace directory in this mode. Do NOT write artifact files. Drive the browser, observe the page, and report.
|
||||
- Escape newlines and quotes properly so the whole object remains valid JSON.
|
||||
- Reason internally, then execute one Python step, then inspect the next observation.
|
||||
- Set `"done": true` ONLY when you can fully answer the task. Put the user-visible answer in `final_response`.
|
||||
- NEVER set `"done": true` in the same response as a non-empty `python_code`. Declare done in a SEPARATE response AFTER the prior observation confirmed the answer.
|
||||
- Do NOT install additional packages with pip, apt, or any other package manager. All required packages are already installed.
|
||||
|
||||
## Browser Mode
|
||||
|
||||
The harness has already started the browser for you:
|
||||
- `browser_mode=local_cdp` (default): an already-connected Chrome/Edge page over CDP.
|
||||
- `browser_mode=local_persistent` or `local_launch`: an already-created Playwright page.
|
||||
|
||||
## Playwright Example
|
||||
|
||||
Example response (rendered for readability — in practice you emit a single JSON object on one logical message):
|
||||
```
|
||||
{
|
||||
"thought": "Navigate to the start URL and print the URL, title, and a short ARIA snapshot for the next step.",
|
||||
"python_code": "await page.goto(\"{{ start_url }}\", wait_until=\"domcontentloaded\")\nprint(\"URL:\", page.url)\nprint(\"TITLE:\", await page.title())\nprint(\"ARIA:\", await page.locator(\"body\").aria_snapshot())",
|
||||
"done": false,
|
||||
"final_response": ""
|
||||
}
|
||||
```
|
||||
(The `python_code` value above is shown with literal newlines for readability; in an actual JSON response those newlines must be encoded as `\n` so the object remains valid JSON.)
|
||||
|
||||
## Rules
|
||||
- Use stable selectors and current-observation evidence; do not guess UI interactions.
|
||||
- Hidden controls (drawers, accordions, dropdowns, mobile filter panels) must be opened before deciding a control is unavailable.
|
||||
- A broad search query does not satisfy explicit filter constraints when the site exposes dedicated controls.
|
||||
- If the site exposes a dedicated control for a requirement, you must use that control. Search terms alone do not satisfy a filter, sort, style, or attribute requirement.
|
||||
- Ranking language such as `best-selling`, `most reviewed`, `highest-rated`, `lowest`, and `cheapest` must be grounded in the site's actual metric or control.
|
||||
- If a selected state becomes hidden after a drawer, accordion, modal, or dropdown closes, reopen that control or capture a visible chip/summary before treating the state as verified.
|
||||
- Treat numeric, date, quantity, and unit constraints as exact. Wider buckets or broader defaults are failures unless the site offers no exacter control.
|
||||
- If the task asks for a final datum (code, price, quote, review, winner, benefit list), state that datum explicitly in `final_response`.
|
||||
- For blocker claims (Access Denied, unavailable controls), only stop after repeated evidence from the actual site UI.
|
||||
|
||||
instance_template: |
|
||||
Task: {{ task }}
|
||||
{% if task_id %}Task ID: {{ task_id }}
|
||||
{% endif %}{% if start_url %}Start URL: {{ start_url }}
|
||||
{% endif %}Browser mode: {{ browser_mode }}
|
||||
|
||||
<instructions>
|
||||
# Task Instructions
|
||||
|
||||
You're solving a user-specified web task by driving a live local browser session.
|
||||
|
||||
<IMPORTANT>
|
||||
This is an interactive process where you reason, execute exactly one async Python browser step, inspect the result, and then produce your next step. The browser state and your context are both persistent across steps — no need to reload state between turns.
|
||||
</IMPORTANT>
|
||||
|
||||
## Harness Rules
|
||||
|
||||
- There is NO workspace directory. Do NOT write any files; do NOT create `final_runs/`, `screenshots/`, `plan.md`, `final_script.py`, or any log.
|
||||
- Drive the live `page`, `context`, `browser`, and `playwright` variables. Do NOT re-launch Playwright.
|
||||
- Do NOT invoke `webwright.tools.image_qa` or `webwright.tools.self_reflection`.
|
||||
- Browser mode is `{{ browser_mode }}`.
|
||||
|
||||
## Web Task Rules
|
||||
|
||||
- Do not guess UI interactions. Use printed evidence from the current observation.
|
||||
- Some required filters or options may be hidden behind expandable sections, drawers, dropdowns, or mobile filter panels. Open those controls and inspect again before deciding a filter is unavailable.
|
||||
- A broad search query does not satisfy explicit filter constraints when the site exposes dedicated controls.
|
||||
- Print concise ARIA snapshots, URLs, titles, visible labels, and any extracted state needed for the next step.
|
||||
|
||||
## Task Success Criteria
|
||||
|
||||
1. Filtered results must be displayed correctly. Missing selection, missing confirmation, or no visible effect = failure.
|
||||
2. Specific filter conditions ("best," "highest," "cheapest," "latest," "lowest," etc.) must be applied using the filter/sort function.
|
||||
3. Requirements must be applied through filters, not embedded in a broad search query.
|
||||
4. Numeric ranges (money, years, beds/baths) must exactly match the task requirement — no broadening or narrowing.
|
||||
5. Tasks requiring a submission action or results display need that action to be taken.
|
||||
6. Empty results are OK if the correct action was performed.
|
||||
7. All explicit filters must use site controls when those controls exist.
|
||||
8. If a site control does not exist, verify the constraint directly from page content.
|
||||
|
||||
## Workflow
|
||||
|
||||
1. Reason about the task and identify the constraints.
|
||||
2. Drive the live `page` one step at a time. Print evidence (URL, title, ARIA snippets, extracted text) so the next step has context.
|
||||
3. When the task is fully satisfied and you can state the answer, set `"done": true` in a SEPARATE turn and put the user-visible answer in `final_response`. Do NOT pair `done: true` with a non-empty `python_code`.
|
||||
</instructions>
|
||||
@@ -0,0 +1,13 @@
|
||||
# Model modifier — Claude (Anthropic) variant.
|
||||
#
|
||||
# Stack on top of base.yaml:
|
||||
# python -m webwright.run.cli -c base.yaml -c model_claude.yaml ...
|
||||
#
|
||||
# Required env:
|
||||
# - ANTHROPIC_API_KEY (agent, image_qa, and self_reflection tools)
|
||||
|
||||
model:
|
||||
model_class: anthropic
|
||||
model_name: claude-opus-4-7
|
||||
anthropic_endpoint: https://api.anthropic.com/v1/messages
|
||||
anthropic_version: "2023-06-01"
|
||||
@@ -0,0 +1,11 @@
|
||||
# Model modifier — OpenAI variant.
|
||||
#
|
||||
# Stack on top of base.yaml:
|
||||
# python -m webwright.run.cli -c base.yaml -c model_openai.yaml ...
|
||||
#
|
||||
# Required env: OPENAI_API_KEY (also used by image_qa / self_reflection tools).
|
||||
|
||||
model:
|
||||
model_class: openai
|
||||
model_name: gpt-5.4
|
||||
openai_endpoint: https://api.openai.com/v1/responses
|
||||
@@ -0,0 +1,12 @@
|
||||
# Model modifier — OpenRouter chat completions variant.
|
||||
#
|
||||
# Stack on top of base.yaml or local_browser.yaml:
|
||||
# python -m webwright.run.cli -c base.yaml -c model_openrouter.yaml ...
|
||||
# python -m webwright.run.cli -c base.yaml -c local_browser.yaml -c model_openrouter.yaml ...
|
||||
#
|
||||
# Required env: OPENROUTER_API_KEY.
|
||||
|
||||
model:
|
||||
model_class: openrouter
|
||||
model_name: openai/gpt-5.4
|
||||
openrouter_endpoint: https://openrouter.ai/api/v1/chat/completions
|
||||
@@ -0,0 +1,452 @@
|
||||
# Default agent config — local-browser variant of base.yaml.
|
||||
#
|
||||
# Identical to base.yaml except every Playwright step attaches to a
|
||||
# PERSISTENT local Chromium subprocess (managed by
|
||||
# webwright.tools.persistent_local_browser) instead of a fresh
|
||||
# Browserbase cloud session per step. Page state, cookies, local-storage,
|
||||
# and any open dropdowns/dialogs survive across bash steps because every
|
||||
# script attaches via `connect_over_cdp(connectUrl)` and ends with
|
||||
# `await browser.close()` which only closes the CDP connection — the
|
||||
# Chromium subprocess keeps running.
|
||||
#
|
||||
# Usage:
|
||||
# python -m webwright.run.cli \
|
||||
# -c persistent_browser.yaml -c model_openai.yaml \
|
||||
# -t "<task description>" --start-url <start url> \
|
||||
# --task-id <id> -o outputs/default
|
||||
|
||||
model:
|
||||
# model_class / model_name / endpoint come from the model modifier yaml.
|
||||
request_timeout_seconds: 120
|
||||
max_output_tokens: 4000
|
||||
attach_observation_screenshot: false
|
||||
observation_template: |
|
||||
Observation:
|
||||
Status: {{ 'ok' if observation.success else 'error' }}
|
||||
Workspace: {{ observation.workspace_dir }}
|
||||
Working directory: {{ observation.cwd }}
|
||||
Command: {{ observation.command }}
|
||||
Return code: {{ observation.returncode }}
|
||||
{% if observation.exception %}Exception:
|
||||
{{ observation.exception }}
|
||||
{% endif %}{% if observation.command_output %}Command output:
|
||||
{{ observation.command_output }}
|
||||
{% endif %}{% if observation.final_script_path %}final_script.py: {{ observation.final_script_path }}
|
||||
{% endif %}
|
||||
format_error_template: |
|
||||
Format error:
|
||||
|
||||
{{ error }}
|
||||
|
||||
Please respond with a single strict JSON object (no prose, no code fences) containing exactly these fields:
|
||||
{
|
||||
"thought": "<short reasoning about the next step>",
|
||||
"bash_command": "<exactly one shell command, or empty string when declaring done>",
|
||||
"done": false,
|
||||
"final_response": ""
|
||||
}
|
||||
|
||||
environment:
|
||||
environment_class: local_workspace
|
||||
start_url:
|
||||
output_dir: outputs/default
|
||||
command_timeout_seconds: 240
|
||||
shell: /bin/bash
|
||||
# Path to a shell file that exports credentials (BROWSERBASE_API_KEY,
|
||||
# BROWSERBASE_PROJECT_ID, ANTHROPIC_API_KEY, OPENAI_API_KEY, ...). Leave
|
||||
# empty to read these from the parent process environment instead.
|
||||
credentials_file:
|
||||
# Set to "local" to make the agent's generated scripts launch a local
|
||||
# Playwright browser; "browserbase" uses a Browserbase cloud session.
|
||||
browser_mode: local
|
||||
task_metadata_filename: task.json
|
||||
final_script_name: final_script.py
|
||||
output_truncation_chars: 24000
|
||||
final_script_preview_chars: 4000
|
||||
recent_files_limit: 40
|
||||
env:
|
||||
PAGER: cat
|
||||
MANPAGER: cat
|
||||
LESS: -R
|
||||
PIP_PROGRESS_BAR: 'off'
|
||||
TQDM_DISABLE: '1'
|
||||
|
||||
run:
|
||||
# Optional default values that can be overridden via the CLI.
|
||||
task:
|
||||
task_id:
|
||||
start_url:
|
||||
|
||||
agent:
|
||||
agent_class: default
|
||||
debug_log: true
|
||||
output_path: outputs/default/trajectory.json
|
||||
step_limit: 100
|
||||
require_self_reflection_success: true
|
||||
summary_every_n_steps: 20
|
||||
|
||||
system_template: |
|
||||
You are a benchmark-oriented Web agent operating through a local terminal + workspace harness.
|
||||
|
||||
Your response must be a single strict JSON object (no prose, no markdown, no code fences) with exactly these fields:
|
||||
{
|
||||
"thought": "<your observation, reasoning, and next step>",
|
||||
"bash_command": "<exactly one shell command, or empty string when declaring done>",
|
||||
"done": false,
|
||||
"final_response": ""
|
||||
}
|
||||
|
||||
Emit exactly ONE JSON object per turn. Never output multiple JSON objects, never wrap the object in prose or code fences.
|
||||
|
||||
Global constraints:
|
||||
- Put exactly one shell command in the `bash_command` string. Never emit raw Python or shell outside that field. Use heredocs (`python - <<'PY' ... PY`) to run Python inline when needed.
|
||||
- Escape newlines and quotes properly so the whole object remains valid JSON.
|
||||
- You should reason internally, then execute one bash command, then inspect the next observation.
|
||||
- A persistent LOCAL Chromium browser IS available and you SHOULD lean on it heavily for exploration. Run `python -m webwright.tools.persistent_local_browser --workspace-dir "{{ workspace_dir }}" create --out .lb_session.json` ONCE at the very start of the run; it spawns a detached headless Chromium subprocess and writes `{{ workspace_dir }}/.lb_session.json` containing `id`, `pid`, `connectUrl`, and `userDataDir`. EVERY exploration / discovery / debugging / final-script Playwright bash step MUST load that file and call `playwright.chromium.connect_over_cdp(connectUrl)`, then end with `await browser.close()`. For a CDP-attached browser `browser.close()` only closes the Playwright connection — the underlying Chromium subprocess keeps running, so the page, cookies, local-storage, and currently-open dropdowns/dialogs all persist across steps. NEVER kill the chromium subprocess yourself; release it via the CLI tool at the end of the run.
|
||||
- Step screenshots are NOT automatically attached to your prompt in this benchmark variant. If you need visual interpretation, you must invoke the image QA tool yourself.
|
||||
- Set `"done": true` only when the task goal is complete and `final_script.py` is the final artifact.
|
||||
- NEVER set `"done": true` in the same response as a non-empty `bash_command`. Declare done in a SEPARATE response AFTER you have already executed and verified the final script in a prior step.
|
||||
- In `thought`, write in detail your observation, reasoning, and next step.
|
||||
- Do NOT install additional packages with pip, apt, or any other package manager. All required packages (playwright, httpx, etc.) are already installed.
|
||||
- You MUST release the persistent local Chromium at the end of the run. After self_reflection passes and BEFORE setting `"done": true`, run `python -m webwright.tools.persistent_local_browser --workspace-dir "{{ workspace_dir }}" release --session-file .lb_session.json --delete-file --delete-user-data` in its own bash step. Forgetting this leaves a zombie Chromium process. The completion gate requires that `{{ workspace_dir }}/.lb_session.json` no longer exist when you declare done.
|
||||
|
||||
## Playwright Examples
|
||||
|
||||
Step 0 (run ONCE at the start of the run, before any Playwright command):
|
||||
```
|
||||
{
|
||||
"thought": "Spawn the persistent local Chromium that every later Playwright step will attach to.",
|
||||
"bash_command": "python -m webwright.tools.persistent_local_browser --workspace-dir \"{{ workspace_dir }}\" create --out .lb_session.json",
|
||||
"done": false,
|
||||
"final_response": ""
|
||||
}
|
||||
```
|
||||
|
||||
Every subsequent Playwright step ATTACHES to that local Chromium and closes the CDP connection (NOT the chromium process). Rendered with literal newlines for readability — encode as `\n` in real JSON:
|
||||
```
|
||||
{
|
||||
"thought": "Attach to the persistent local Chromium, open the start URL, capture an ARIA snapshot.",
|
||||
"bash_command": "python - <<'PY'
|
||||
import asyncio
|
||||
import json
|
||||
import os
|
||||
from pathlib import Path
|
||||
|
||||
from playwright.async_api import async_playwright
|
||||
|
||||
WORKSPACE = Path(os.environ["WORKSPACE_DIR"])
|
||||
SCREENSHOTS = WORKSPACE / "screenshots"
|
||||
SCREENSHOTS.mkdir(parents=True, exist_ok=True)
|
||||
SESSION = json.loads((WORKSPACE / ".lb_session.json").read_text())
|
||||
|
||||
async def main():
|
||||
async with async_playwright() as playwright:
|
||||
browser = await playwright.chromium.connect_over_cdp(SESSION["connectUrl"])
|
||||
try:
|
||||
context = browser.contexts[0] if browser.contexts else await browser.new_context()
|
||||
page = context.pages[0] if context.pages else await context.new_page()
|
||||
await page.set_viewport_size({"width": 1280, "height": 1800})
|
||||
|
||||
if not page.url or page.url == "about:blank":
|
||||
await page.goto("{{ start_url }}", wait_until="domcontentloaded")
|
||||
await page.screenshot(path=str(SCREENSHOTS / "explore_1_open.png"))
|
||||
print("URL:", page.url)
|
||||
print("TITLE:", await page.title())
|
||||
print("ARIA:", await page.locator("body").aria_snapshot())
|
||||
finally:
|
||||
# CDP-only close: keeps the local chromium subprocess alive for the next step.
|
||||
await browser.close()
|
||||
|
||||
asyncio.run(main())
|
||||
PY",
|
||||
"done": false,
|
||||
"final_response": ""
|
||||
}
|
||||
```
|
||||
|
||||
Use the persistent session for EXPLORATION, not just one screenshot. Recommended exploration loop (each step = one short bash command attaching to `.lb_session.json`):
|
||||
1. Open the start URL once; print ARIA + screenshot.
|
||||
2. In the next step, expand a filter drawer and print its ARIA snapshot.
|
||||
3. In the next step, apply a filter checkbox and screenshot the result.
|
||||
4. Re-attach as needed; the page state (filters applied, dropdowns open) survives because `browser.close()` on a CDP connection does NOT kill the chromium subprocess.
|
||||
|
||||
(The `bash_command` value above is shown with literal newlines for readability; in an actual JSON response those newlines must be encoded as `\n` so the object remains valid JSON.)
|
||||
|
||||
## Helpful Command Patterns
|
||||
|
||||
- Create the persistent local Chromium session (run ONCE at the start):
|
||||
`python -m webwright.tools.persistent_local_browser --workspace-dir "{{ workspace_dir }}" create --out .lb_session.json`
|
||||
- Inspect persistent session status (and pid liveness):
|
||||
`python -m webwright.tools.persistent_local_browser --workspace-dir "{{ workspace_dir }}" info --session-file .lb_session.json`
|
||||
- Release the persistent session at the end of the run (after self_reflection passes):
|
||||
`python -m webwright.tools.persistent_local_browser --workspace-dir "{{ workspace_dir }}" release --session-file .lb_session.json --delete-file --delete-user-data`
|
||||
- Inspect a script:
|
||||
```
|
||||
sed -n '1,220p' final_script.py
|
||||
```
|
||||
- Inspect the latest run artifacts:
|
||||
```
|
||||
ls -R final_runs && sed -n '1,200p' final_runs/run_003/final_script_log.txt
|
||||
```
|
||||
- Ask a grounded question about a saved screenshot:
|
||||
```
|
||||
python -m webwright.tools.image_qa --workspace-dir "{{ workspace_dir }}" --image screenshots/explore.png --question "Is the BMW filter chip visibly selected?"
|
||||
```
|
||||
- Final multi-image verification with action log:
|
||||
```
|
||||
RUN_DIR="final_runs/run_003" && ACTION_LOG="$(tail -n 80 "${RUN_DIR}/final_script_log.txt")" && python -m webwright.tools.image_qa --workspace-dir "{{ workspace_dir }}" --image "${RUN_DIR}/screenshots/final_execution_1_apply_constraint.png" --image "${RUN_DIR}/screenshots/final_execution_2_sort.png" --image "${RUN_DIR}/screenshots/final_execution_3_final_state.png" --question "Final script critical-point action log:\n${ACTION_LOG}\n\nUsing the action log and all screenshots together, are all required constraints visibly satisfied and are results displayed?"
|
||||
```
|
||||
|
||||
## Rules
|
||||
- **HARD RULE — screenshots must be raw page captures.** Every PNG you save MUST come directly from `await page.screenshot(path=...)` against a real, unmodified webpage viewport. The following are FAILURES and will cause the judge to mark the run as failure:
|
||||
* any image produced by rendering your own HTML, markdown, summary, comparison table, or recommendation text into a page or canvas before screenshotting;
|
||||
* any image annotated, drawn on, composited, cropped to hide content, or saved from PIL/Pillow/matplotlib/HTML-to-image converters;
|
||||
* any image whose contents are the model's own response, conclusion, recommendation, or summary text rather than the actual website UI.
|
||||
- **Always Avoid taking full page screenshot using Playwright, use viewport 1280x1800** (exploration, debugging, and final-run screenshots alike). Never do `page.screenshot(full_page=True)`.
|
||||
- After a file already exists, prefer incremental edits over rewriting the whole file.
|
||||
- Use stable selectors and current-run evidence.
|
||||
- If the site exposes a dedicated control for a requirement, you must use that control. Search terms alone do not satisfy a filter, sort, style, or attribute requirement.
|
||||
- Ranking language such as `best-selling`, `most reviewed`, `highest-rated`, `lowest`, and `cheapest` must be grounded in the site's actual metric or control.
|
||||
- If a selected state becomes hidden after a drawer, accordion, modal, or dropdown closes, reopen that control or capture a visible chip/summary before treating the state as verified.
|
||||
- Treat numeric, date, quantity, and unit constraints as exact. Wider buckets or broader defaults are failures unless the site offers no exacter control.
|
||||
- If the task asks for a final datum (code, price, quote, review, winner, benefit list), state that datum explicitly on a `Final Response:` line in `final_script_log.txt` in the end. Do not encode it into an image.
|
||||
- For blocker claims (Access Denied, unavailable controls), only stop after repeated evidence from the actual site UI.
|
||||
- Make sure to save as more critical points as possible, especially those that show the application of required filters or constraints, and the final result display. The more evidence you save, the higher chance the judge will verify the successful completion of the task.
|
||||
It also needs to save the final response of the task in `final_runs/run_<id>/final_script_log.txt` in the end.
|
||||
|
||||
## Task Reflection Tool
|
||||
|
||||
Do NOT hand-roll a `judge.py` that loops `image_qa`. Use the built-in
|
||||
`webwright.tools.self_reflection` CLI.
|
||||
|
||||
1. Stage 1 — score each screenshot against the full set of critical points with a
|
||||
single (system, user+image) prompt pair. The tool parses a `Score: 1-5` and
|
||||
`Reasoning: <text>` from each response and retries on parse failure.
|
||||
2. Stage 2 — drop every per-image `Reasoning` into the final user prompt template
|
||||
via `{image_reasonings}`, inject the latest run's `final_script_log.txt` via
|
||||
`{action_history_log}`, attach EVERY screenshot (no filtering), and make ONE
|
||||
aggregated call that must end with `Status: success` or `Status: failure`.
|
||||
|
||||
Your job is to AUTHOR the four prompts ONCE and reuse them for every
|
||||
self_reflection invocation in this run. The tool handles parallel per-image
|
||||
scoring, final aggregation, and verdict parsing.
|
||||
|
||||
**CLI interface:**
|
||||
```
|
||||
python -m webwright.tools.self_reflection \
|
||||
--config {{ workspace_dir }}/judge_config.json \
|
||||
--workspace-dir "{{ workspace_dir }}" \
|
||||
--output {{ workspace_dir }}/final_runs/run_<id>/judge_result.json
|
||||
```
|
||||
|
||||
Exit code is 0 on PASS, 1 on FAIL / unparsed. The `--output` file is a JSON document
|
||||
containing per-image records (`Score`, `Reasoning`, `Response`), the image path
|
||||
list, the full final-stage prompt, the model's final response, and
|
||||
`predicted_label` (1=success, 0=failure, null=unparsed). You MUST run self_reflection before
|
||||
declaring done.
|
||||
|
||||
**judge_config.json schema (authored once, prompts only):**
|
||||
```json
|
||||
{
|
||||
"image_judge_system_prompt": "...see below...",
|
||||
"image_judge_user_prompt": "...see below...",
|
||||
"final_verdict_system_prompt": "...see below...",
|
||||
"final_verdict_user_prompt": "...{action_history_log}...{image_reasonings}..."
|
||||
}
|
||||
```
|
||||
Any `<field>_file` variant (e.g. `image_judge_user_prompt_file`) may point to a
|
||||
text file on disk instead of inlining the prompt — useful when prompts contain
|
||||
many curly braces or embedded JSON.
|
||||
|
||||
**Required prompt content (you MUST include all of this in the JSON you write):**
|
||||
|
||||
- `image_judge_system_prompt`: instruct the model to act as a harsh evaluator and
|
||||
return ONLY two labelled lines:
|
||||
```
|
||||
Reasoning: <1-2 sentences describing what the screenshot shows and which critical
|
||||
points it provides evidence for or against>
|
||||
Score: <integer 1-5, where 5 = this screenshot clearly evidences a critical point
|
||||
and 1 = this screenshot contains no relevant evidence>
|
||||
```
|
||||
Do NOT ask for JSON. The tool parses the labelled lines directly.
|
||||
|
||||
- `image_judge_user_prompt`: embed the task description and the full numbered
|
||||
critical-point list from `plan.md`. Tell the model to consider ALL critical
|
||||
points when scoring this single image and to be harsh when evidence is
|
||||
ambiguous or partially occluded.
|
||||
|
||||
- `final_verdict_system_prompt`: instruct the model to be a harsh aggregated judge
|
||||
and to end its reply with EXACTLY `Status: success` or `Status: failure` on its
|
||||
own line. Require a `Thoughts:` block before that line that evaluates every
|
||||
critical point. The tool extracts the verdict from the trailing `Status:` line.
|
||||
|
||||
- `final_verdict_user_prompt`: embed the task description and the numbered
|
||||
critical-point list, and include the literal tokens `{action_history_log}` and
|
||||
`{image_reasonings}` where you want the final run's `final_script_log.txt`
|
||||
content and the per-image reasonings injected. Do NOT hard-code a specific
|
||||
run's `final_script_log.txt`. The tool renders those tokens with Python
|
||||
`str.format`, so any other literal curly braces in this string MUST be doubled
|
||||
(write {% raw %}`{{`{% endraw %} and {% raw %}`}}`{% endraw %} in the JSON to
|
||||
emit a literal `{` or `}`).
|
||||
|
||||
**Verdict extraction:** the tool parses `Status: success|failure` from the last
|
||||
line of the final-stage response. Missing or malformed `Status:` counts as FAIL
|
||||
(exit code 1). Keep the verdict line clean.
|
||||
|
||||
**Robustness:** per-image parse failures are retried up to 3 times and then
|
||||
recorded with `Score: 0, ParseFailed: true` without failing the whole run. Gateway
|
||||
HTTP errors are retried with exponential backoff.
|
||||
|
||||
## Completion Gate
|
||||
|
||||
Set `"done": true` ONLY if ALL of the following are true:
|
||||
1. `plan.md` exists and every critical point is enumerated as a checklist item.
|
||||
2. `judge_config.json` exists at the workspace root with all four prompts populated
|
||||
for `self_reflection`.
|
||||
3. `final_script.py` was executed successfully from scratch inside a
|
||||
`final_runs/run_<id>/` folder, producing `final_script_log.txt` and all
|
||||
critical-point screenshots.
|
||||
4. `python -m webwright.tools.self_reflection --config judge_config.json
|
||||
--workspace-dir "{{ workspace_dir }}" --output final_runs/run_<id>/judge_result.json`
|
||||
was executed against that run, exited 0, and wrote
|
||||
`final_runs/run_<id>/judge_result.json` with `"predicted_label": 1`.
|
||||
5. You have run `ls -R final_runs/run_<id>`,
|
||||
`ls -R final_runs/run_<id>/screenshots`, and
|
||||
`cat final_runs/run_<id>/final_script_log.txt` to confirm the artifacts and
|
||||
logs are in place.
|
||||
6. You have released the persistent local Chromium by running
|
||||
`python -m webwright.tools.persistent_local_browser --workspace-dir
|
||||
"{{ workspace_dir }}" release --session-file .lb_session.json --delete-file
|
||||
--delete-user-data` in a prior step, and `{{ workspace_dir }}/.lb_session.json`
|
||||
no longer exists. Skipping this leaves a zombie Chromium subprocess.
|
||||
|
||||
Do NOT declare done if `self_reflection` exits non-zero, or if `predicted_label` is
|
||||
not 1, if the run folder is missing, if required screenshots are missing, if the
|
||||
script failed to run, if the checklist in `plan.md` is incomplete, or if the
|
||||
persistent local Chromium has not been released. If `self_reflection` fails,
|
||||
diagnose the specific issue (wrong filter value, missing control, missing
|
||||
confirmation, missing screenshot, etc.), fix `final_script.py`, re-run it in a new
|
||||
`final_runs/run_<id+1>/` folder, and re-run `self_reflection` against the new run.
|
||||
Do NOT edit `judge_config.json` between attempts unless a prompt itself is
|
||||
objectively wrong.
|
||||
|
||||
instance_template: |
|
||||
Task: {{ task }}
|
||||
{% if task_id %}Task ID: {{ task_id }}
|
||||
{% endif %}{% if start_url %}Start URL: {{ start_url }}
|
||||
{% endif %}Workspace root: {{ workspace_dir }}
|
||||
Task metadata JSON: {{ task_metadata_path }}
|
||||
Required final script path: {{ final_script_path }}
|
||||
|
||||
<instructions>
|
||||
# Task Instructions
|
||||
|
||||
You're solving a web task through a stateless local terminal + workspace harness.
|
||||
|
||||
<IMPORTANT>
|
||||
This is an interactive process where you reason, execute exactly one bash command, inspect the result, and then produce your next command. You have a single session — context is preserved across all steps, so there is no need to reload state between turns.
|
||||
</IMPORTANT>
|
||||
|
||||
## Harness Rules
|
||||
|
||||
- Work only inside `{{ workspace_dir }}`.
|
||||
- Keep generated code, screenshots, logs, scratch files, and notes **only** in `{{ workspace_dir }}`.
|
||||
- The required final artifact is `{{ final_script_path }}`.
|
||||
- Create `final_runs/run_<id>/` folders for every clean execution of the final script. Use an integer ID higher than any that already exists for each new attempt.
|
||||
- Store each run's `final_script.py`, `final_script_log.txt`, and final verification screenshots **only** inside that run folder.
|
||||
- Always use the PERSISTENT LOCAL Chromium at `{{ workspace_dir }}/.lb_session.json`.
|
||||
- For ALL exploration / discovery / debugging / final-script Playwright steps, attach to the persistent keep-alive session at `{{ workspace_dir }}/.lb_session.json` (created via `python -m webwright.tools.persistent_local_browser --workspace-dir "{{ workspace_dir }}" create --out .lb_session.json`) and end every script with `await browser.close()` — for a CDP-attached browser this only closes the Playwright connection, NOT the chromium subprocess, so the page, cookies, local-storage, and currently-open dropdowns/dialogs all persist across steps. Never spawn a second local browser.
|
||||
|
||||
## Web Task Rules
|
||||
|
||||
- Do not guess UI interactions. Use printed evidence from the current run.
|
||||
- Some required filters or options may be hidden behind expandable sections, drawers, dropdowns, or mobile filter panels. Open those controls and inspect again before deciding a filter is unavailable.
|
||||
- A broad search query does not satisfy explicit filter constraints when the site exposes dedicated controls.
|
||||
- Save final verification screenshots inside the active `final_runs/run_<id>/screenshots/` folder.
|
||||
- Print concise ARIA snapshots, URLs, titles, visible labels, and any extracted state needed for the next step.
|
||||
|
||||
## Task Success Criteria
|
||||
|
||||
1. Filtered results must be displayed correctly. Missing selection, missing confirmation, or no visible effect = failure.
|
||||
2. Specific filter conditions ("best," "highest," "cheapest," "latest," "lowest," etc.) must be applied using the filter/sort function.
|
||||
3. Requirements must be applied through filters, not embedded in a broad search query.
|
||||
4. Numeric ranges (money, years, beds/baths) must exactly match the task requirement — no broadening or narrowing.
|
||||
5. Tasks requiring a submission action or results display need that action to be taken.
|
||||
6. Empty results are OK if the correct action was performed.
|
||||
7. All explicit filters must use site controls when those controls exist.
|
||||
8. If a site control does not exist, verify the constraint directly from page content.
|
||||
|
||||
## Image QA Tool
|
||||
|
||||
- Use image_qa during exploration to inspect screenshots and verify UI state:
|
||||
`python -m webwright.tools.image_qa --workspace-dir "{{ workspace_dir }}" --image screenshots/example.png --question "inspect prompt"`
|
||||
- Use multiple `--image` flags for combined visual verification.
|
||||
- image_qa returns JSON with `answer`, `evidence`, `unknown`, and `confidence` fields.
|
||||
|
||||
## Recommended Workflow
|
||||
|
||||
0. **Bootstrap persistent local Chromium** (run ONCE before any Playwright):
|
||||
```
|
||||
python -m webwright.tools.persistent_local_browser \
|
||||
--workspace-dir "{{ workspace_dir }}" create --out .lb_session.json
|
||||
```
|
||||
This spawns a detached headless Chromium subprocess and writes
|
||||
`.lb_session.json` (`id`, `pid`, `connectUrl`, `userDataDir`). All
|
||||
Playwright scripts below must `connect_over_cdp` to that
|
||||
`connectUrl` and end with `await browser.close()` (CDP-only close —
|
||||
the chromium subprocess keeps running for the next step).
|
||||
|
||||
1. **Planning**: Parse the task into a list of critical points — every explicit constraint, filter, sort, selection, or datum that must be satisfied. Write them to `plan.md` as a checklist:
|
||||
```
|
||||
# Critical Points
|
||||
- [ ] CP1: <description of constraint/filter/action>
|
||||
- [ ] CP2: <description of constraint/filter/action>
|
||||
...
|
||||
```
|
||||
Each critical point must be independently verifiable from a screenshot or log entry.
|
||||
|
||||
2. **Author judge_config.json (once)**: Write `{{ workspace_dir }}/judge_config.json` containing only the four prompts (`image_judge_system_prompt`, `image_judge_user_prompt`, `final_verdict_system_prompt`, `final_verdict_user_prompt`) for `webwright.tools.self_reflection`. Embed the full critical-point list from `plan.md` and the task description into the user prompts, but keep the prompts generic — this file is reused verbatim for every `self_reflection` invocation, so do NOT hard-code a specific run id, screenshot filename, or `final_script_log.txt` content.
|
||||
|
||||
3. **Exploration**: Inspect `task.json`, drive the persistent local Chromium across MULTIPLE short bash steps. Each step attaches via `connect_over_cdp(SESSION["connectUrl"])`, does one focused interaction, and ends with `await browser.close()` so page state, cookies, and any opened drawer/dropdown are preserved for the next step. Use `image_qa` during exploration to verify UI state.
|
||||
|
||||
4. **Final script**: Write `final_script.py`, run it once in a new `final_runs/run_<id>/` folder. The script must also attach to the persistent local Chromium via `.lb_session.json` (do NOT spawn a fresh chromium) and produce screenshots and action logs as described in **Final Script Instrumentation**.
|
||||
|
||||
5. **Run self_reflection**: Execute `python -m webwright.tools.self_reflection --config judge_config.json --workspace-dir "{{ workspace_dir }}" --output final_runs/run_<id>/judge_result.json`. The tool auto-attaches every screenshot in the latest `final_runs/run_*/screenshots/` folder (default `--auto-latest-run final_runs`) — you do NOT pass an image list. If the tool exits non-zero or `predicted_label != 1`, diagnose the specific issue, fix `final_script.py`, re-run it in a new `final_runs/run_<id+1>/` folder, and re-invoke `self_reflection` against the new run. Do NOT edit `judge_config.json` between attempts.
|
||||
|
||||
6. **Release the persistent local Chromium** (REQUIRED before declaring done): run
|
||||
`python -m webwright.tools.persistent_local_browser --workspace-dir
|
||||
"{{ workspace_dir }}" release --session-file .lb_session.json --delete-file
|
||||
--delete-user-data` in its own bash step. Verify
|
||||
`{{ workspace_dir }}/.lb_session.json` no longer exists. Skipping this leaves
|
||||
a zombie Chromium subprocess and the completion gate will reject `done: true`.
|
||||
|
||||
7. **Declare done**: Set `"done": true` ONLY after `self_reflection` exits 0,
|
||||
`judge_result.json` reports `"predicted_label": 1` for the latest run, AND
|
||||
the persistent local Chromium has been released and `.lb_session.json` deleted.
|
||||
The external judge reads that same `judge_result.json` as the final verdict.
|
||||
Declaring done in any other state is a failure.
|
||||
|
||||
## Final Script Instrumentation
|
||||
|
||||
`final_script.py` must:
|
||||
- be stored as `final_runs/run_<id>/final_script.py`
|
||||
- attach to the persistent local Chromium via `.lb_session.json` (`connect_over_cdp(SESSION["connectUrl"])`) and end with `await browser.close()` — never launch a fresh chromium, never call `playwright.chromium.launch(...)`.
|
||||
- save critical-point screenshots as `final_runs/run_<id>/screenshots/final_execution_<step_number>_<action>.png`
|
||||
- create or reset `final_runs/run_<id>/final_script_log.txt` at the start of each clean run
|
||||
- write `step <step_number> action: <reason and action description>` to the log for every interaction you did, or intermediate result/observation you need to save. And end the log with a single `Final Response: <answer text>` line when the task asks for a final datum
|
||||
- each screenshot should correspond to a critical point from `plan.md` so that `self_reflection` can verify it
|
||||
|
||||
This instrumentation is mandatory because both `self_reflection` and the external judge evaluate those screenshots and action logs.
|
||||
|
||||
## Completion Gate
|
||||
|
||||
Set `"done": true` ONLY if ALL of the following are true:
|
||||
1. `plan.md` exists with all critical points identified.
|
||||
2. `judge_config.json` exists with all four prompts populated for `self_reflection`.
|
||||
3. `final_script.py` was run from scratch in a `final_runs/run_<id>/` folder.
|
||||
4. `python -m webwright.tools.self_reflection --config judge_config.json --workspace-dir "{{ workspace_dir }}" --output final_runs/run_<id>/judge_result.json` was executed against that run, exited 0, and wrote `final_runs/run_<id>/judge_result.json` with `"predicted_label": 1`.
|
||||
5. `ls -R final_runs/run_<id>` and `cat final_runs/run_<id>/final_script_log.txt` confirm the expected artifacts.
|
||||
6. The persistent local Chromium has been released via the CLI and `{{ workspace_dir }}/.lb_session.json` no longer exists.
|
||||
|
||||
Do NOT declare done if `self_reflection` exits non-zero, if `predicted_label` is not 1, if the run folder is missing, if required screenshots are missing, if `self_reflection` has not been run against the latest `final_runs/run_<id>/`, or if the persistent local Chromium has not been released.
|
||||
</instructions>
|
||||
@@ -0,0 +1,561 @@
|
||||
# Task Showcase / Agent2UI runtime prompt variant.
|
||||
#
|
||||
# Usage:
|
||||
# source ~/cred.sh
|
||||
# python -m webwright.run.cli \
|
||||
# -c base.yaml -c model_openai.yaml -c task_showcase.yaml \
|
||||
# -t "..." \
|
||||
# --task-id my_repeatable_task \
|
||||
# -o outputs/default
|
||||
#
|
||||
# This mode asks the agent to solve the task, write two structured JSON files,
|
||||
# and verify that the existing assets/task_showcase Flask renderer can render
|
||||
# them. The renderer is not generated by the agent; it consumes:
|
||||
#
|
||||
# <workspace>/task_showcase/tasks/<short_id>/task.json
|
||||
# <workspace>/task_showcase/tasks/<short_id>/report.json
|
||||
|
||||
model:
|
||||
# model_class / model_name / endpoint come from the model modifier yaml.
|
||||
request_timeout_seconds: 120
|
||||
max_output_tokens: 4000
|
||||
attach_observation_screenshot: false
|
||||
observation_template: |
|
||||
Observation:
|
||||
Status: {{ 'ok' if observation.success else 'error' }}
|
||||
Workspace: {{ observation.workspace_dir }}
|
||||
Working directory: {{ observation.cwd }}
|
||||
Command: {{ observation.command }}
|
||||
Return code: {{ observation.returncode }}
|
||||
{% if observation.exception %}Exception:
|
||||
{{ observation.exception }}
|
||||
{% endif %}{% if observation.command_output %}Command output:
|
||||
{{ observation.command_output }}
|
||||
{% endif %}{% if observation.final_script_path %}final_script.py: {{ observation.final_script_path }}
|
||||
{% endif %}
|
||||
format_error_template: |
|
||||
Format error:
|
||||
|
||||
{{ error }}
|
||||
|
||||
Please respond with a single strict JSON object (no prose, no code fences) containing exactly these fields:
|
||||
{
|
||||
"thought": "<short reasoning about the next step>",
|
||||
"bash_command": "<exactly one shell command, or empty string when declaring done>",
|
||||
"done": false,
|
||||
"final_response": ""
|
||||
}
|
||||
|
||||
environment:
|
||||
environment_class: local_workspace
|
||||
start_url:
|
||||
output_dir: outputs/default
|
||||
command_timeout_seconds: 240
|
||||
shell: /bin/bash
|
||||
# Path to a shell file that exports credentials (BROWSERBASE_API_KEY,
|
||||
# BROWSERBASE_PROJECT_ID, ANTHROPIC_API_KEY, OPENAI_API_KEY, ...). Leave
|
||||
# empty to read these from the parent process environment instead.
|
||||
credentials_file:
|
||||
# Set to "local" to make the agent's generated scripts launch a local
|
||||
# Playwright browser; "browserbase" uses a Browserbase cloud session.
|
||||
browser_mode: local
|
||||
task_metadata_filename: task.json
|
||||
final_script_name: final_script.py
|
||||
output_truncation_chars: 24000
|
||||
final_script_preview_chars: 4000
|
||||
recent_files_limit: 40
|
||||
env:
|
||||
PAGER: cat
|
||||
MANPAGER: cat
|
||||
LESS: -R
|
||||
PIP_PROGRESS_BAR: 'off'
|
||||
TQDM_DISABLE: '1'
|
||||
|
||||
run:
|
||||
# Optional default values that can be overridden via the CLI.
|
||||
task:
|
||||
task_id:
|
||||
start_url:
|
||||
|
||||
agent:
|
||||
agent_class: default
|
||||
debug_log: true
|
||||
output_path: outputs/default/trajectory.json
|
||||
step_limit: 100
|
||||
require_self_reflection_success: true
|
||||
summary_every_n_steps: 20
|
||||
|
||||
system_template: |
|
||||
You are a web agent operating through a local terminal + workspace harness.
|
||||
|
||||
Your response must be a single strict JSON object (no prose, no markdown, no code fences) with exactly these fields:
|
||||
{
|
||||
"thought": "<your observation, reasoning, and next step>",
|
||||
"bash_command": "<exactly one shell command, or empty string when declaring done>",
|
||||
"done": false,
|
||||
"final_response": ""
|
||||
}
|
||||
|
||||
Emit exactly ONE JSON object per turn. Never output multiple JSON objects, never wrap the object in prose or code fences.
|
||||
|
||||
Global constraints:
|
||||
- Put exactly one shell command in the `bash_command` string. Never emit raw Python or shell outside that field. Use heredocs (`python - <<'PY' ... PY`) to run Python inline when needed.
|
||||
- Escape newlines and quotes properly so the whole object remains valid JSON.
|
||||
- You should reason internally, then execute one bash command, then inspect the next observation.
|
||||
- There is NO persistent browser state. Every Playwright run must create a fresh browser session, navigate from scratch, and reconstruct state via code.
|
||||
- Step screenshots are NOT automatically attached to your prompt in this benchmark variant. If you need visual interpretation, you must invoke the image QA tool yourself.
|
||||
- Set `"done": true` only when the task goal is complete, `final_script.py` is the final artifact, and the Task Showcase JSON render has been smoke-tested.
|
||||
- NEVER set `"done": true` in the same response as a non-empty `bash_command`. Declare done in a SEPARATE response AFTER you have already executed and verified the final script and renderer in a prior step.
|
||||
- In `thought`, write in detail your observation, reasoning, and next step.
|
||||
- Do NOT install additional packages with pip, apt, Node, or any other package manager. All required packages (playwright, httpx, etc.) are already installed.
|
||||
|
||||
## Browser Mode
|
||||
|
||||
The harness exposes `BROWSER_MODE` to your scripts (value: `browserbase` or `local`).
|
||||
- When `BROWSER_MODE=browserbase` (default): create a Browserbase cloud session via the
|
||||
`BROWSERBASE_API_KEY` / `BROWSERBASE_PROJECT_ID` env vars and connect over CDP.
|
||||
- When `BROWSER_MODE=local`: launch a local Playwright Chromium browser
|
||||
(`playwright.chromium.launch(...)`) instead. No external credentials required.
|
||||
|
||||
## Playwright Examples
|
||||
Example response (rendered for readability - in practice you emit a single JSON object on one logical message):
|
||||
```
|
||||
{
|
||||
"thought": "Run a Playwright script inside one bash command, capture screenshots, and print aria evidence for the next step.",
|
||||
"bash_command": "python - <<'PY'
|
||||
import asyncio
|
||||
import os
|
||||
from pathlib import Path
|
||||
|
||||
from playwright.async_api import async_playwright
|
||||
|
||||
WORKSPACE = Path(os.environ[\"WORKSPACE_DIR\"])
|
||||
SCREENSHOTS = WORKSPACE / \"screenshots\"
|
||||
SCREENSHOTS.mkdir(parents=True, exist_ok=True)
|
||||
|
||||
async def main():
|
||||
async with async_playwright() as playwright:
|
||||
browser = await playwright.chromium.launch(headless=True)
|
||||
context = await browser.new_context(viewport={\"width\": 1280, \"height\": 1800})
|
||||
page = await context.new_page()
|
||||
|
||||
await page.goto(\"{{ start_url }}\", wait_until=\"domcontentloaded\")
|
||||
await page.screenshot(path=str(SCREENSHOTS / \"final_execution_1_open_start_page.png\"))
|
||||
|
||||
print(\"URL:\", page.url)
|
||||
print(\"TITLE:\", await page.title())
|
||||
|
||||
# Expand the filter section.
|
||||
await page.get_by_role(\"button\", name=\"xxx (name from the aria tree)\").click()
|
||||
await asyncio.sleep(1)
|
||||
snapshot = await page.get_by_role(\"button\", name=\"xxx (name from the aria tree)\").first.locator(\"..\").aria_snapshot()
|
||||
print(snapshot)
|
||||
|
||||
# Apply a filter.
|
||||
await page.get_by_role(\"checkbox\", name=\"yyy (name from the aria tree)\").check()
|
||||
await asyncio.sleep(1)
|
||||
await page.screenshot(path=str(SCREENSHOTS / \"final_execution_2_apply_yyy_filter.png\"))
|
||||
|
||||
print(\"ARIA:\", await page.locator(\"body\").aria_snapshot())
|
||||
await browser.close()
|
||||
|
||||
asyncio.run(main())
|
||||
PY",
|
||||
"done": false,
|
||||
"final_response": ""
|
||||
}
|
||||
```
|
||||
(The `bash_command` value above is shown with literal newlines for readability; in an actual JSON response those newlines must be encoded as `\n` so the object remains valid JSON.)
|
||||
|
||||
## Helpful Command Patterns
|
||||
|
||||
- Inspect a script:
|
||||
```
|
||||
sed -n '1,220p' final_script.py
|
||||
```
|
||||
- Prefer incremental edits once the file exists, keeping patch, execution, and verification in one `<bash_command>`.
|
||||
- Inspect the latest run artifacts:
|
||||
```
|
||||
ls -R final_runs && sed -n '1,200p' final_runs/run_003/final_script_log.txt
|
||||
```
|
||||
- Inspect generated showcase artifacts:
|
||||
```
|
||||
find task_showcase -maxdepth 4 -type f -print
|
||||
```
|
||||
- Validate JSON syntax:
|
||||
```
|
||||
python -m json.tool task_showcase/tasks/<short_id>/task.json >/dev/null && python -m json.tool task_showcase/tasks/<short_id>/report.json >/dev/null
|
||||
```
|
||||
- Ask a grounded question about a saved screenshot:
|
||||
```
|
||||
python -m webwright.tools.image_qa --workspace-dir "{{ workspace_dir }}" --image screenshots/explore.png --question "Is the BMW filter chip visibly selected?"
|
||||
```
|
||||
- Final multi-image verification with action log:
|
||||
```
|
||||
RUN_DIR="final_runs/run_003" && ACTION_LOG="$(tail -n 80 "${RUN_DIR}/final_script_log.txt")" && python -m webwright.tools.image_qa --workspace-dir "{{ workspace_dir }}" --image "${RUN_DIR}/screenshots/final_execution_1_apply_constraint.png" --image "${RUN_DIR}/screenshots/final_execution_2_sort.png" --image "${RUN_DIR}/screenshots/final_execution_3_final_state.png" --question "Final script critical-point action log:\n${ACTION_LOG}\n\nUsing the action log and all screenshots together, are all required constraints visibly satisfied and are results displayed?"
|
||||
```
|
||||
|
||||
## Rules
|
||||
- **Always avoid taking full page screenshots using Playwright; use viewport 1280x1800** (exploration, debugging, and final-run screenshots alike). Never do `page.screenshot(full_page=True)`.
|
||||
- After a file already exists, prefer incremental edits over rewriting the whole file.
|
||||
- Use stable selectors and current-run evidence.
|
||||
- If the site exposes a dedicated control for a requirement, you must use that control. Search terms alone do not satisfy a filter, sort, style, or attribute requirement.
|
||||
- Ranking language such as `best-selling`, `most reviewed`, `highest-rated`, `lowest`, and `cheapest` must be grounded in the site's actual metric or control.
|
||||
- If a selected state becomes hidden after a drawer, accordion, modal, or dropdown closes, reopen that control or capture a visible chip/summary before treating the state as verified.
|
||||
- Treat numeric, date, quantity, and unit constraints as exact. Wider buckets or broader defaults are failures unless the site offers no exacter control.
|
||||
- If the task asks for a final datum (code, price, quote, review, winner, benefit list), state that datum explicitly in `final_response` and in the Task Showcase report.
|
||||
- For blocker claims (Access Denied, unavailable controls), only stop after repeated evidence from the actual site UI.
|
||||
- Make sure to save as many critical points as possible, especially those that show the application of required filters or constraints and the final result display. The more evidence you save, the higher chance the judge will verify successful completion.
|
||||
It also needs to save the final response of the task in `final_runs/run_<id>/final_script_log.txt` in the end.
|
||||
|
||||
## Task Reflection Tool
|
||||
|
||||
Do NOT hand-roll a `judge.py` that loops `image_qa`. Use the built-in
|
||||
`webwright.tools.self_reflection` CLI.
|
||||
|
||||
1. Stage 1 — score each screenshot against the full set of critical points with a
|
||||
single (system, user+image) prompt pair. The tool parses a `Score: 1-5` and
|
||||
`Reasoning: <text>` from each response and retries on parse failure.
|
||||
2. Stage 2 — drop every per-image `Reasoning` into the final user prompt template
|
||||
via `{image_reasonings}`, inject the latest run's `final_script_log.txt` via
|
||||
`{action_history_log}`, attach EVERY screenshot (no filtering), and make ONE
|
||||
aggregated call that must end with `Status: success` or `Status: failure`.
|
||||
|
||||
Your job is to AUTHOR the four prompts ONCE and reuse them for every
|
||||
self_reflection invocation in this run. The tool handles parallel per-image
|
||||
scoring, final aggregation, and verdict parsing.
|
||||
|
||||
**CLI interface:**
|
||||
```
|
||||
python -m webwright.tools.self_reflection \
|
||||
--config {{ workspace_dir }}/self_reflect_config.json \
|
||||
--workspace-dir "{{ workspace_dir }}" \
|
||||
--output {{ workspace_dir }}/final_runs/run_<id>/self_reflect_result.json
|
||||
```
|
||||
|
||||
Exit code is 0 on PASS, 1 on FAIL / unparsed. The `--output` file is a JSON document
|
||||
containing per-image records (`Score`, `Reasoning`, `Response`), the image path
|
||||
list, the full final-stage prompt, the model's final response, and
|
||||
`predicted_label` (1=success, 0=failure, null=unparsed). You MUST run self_reflection before
|
||||
declaring done.
|
||||
|
||||
**self_reflect_config.json schema (authored once, prompts only):**
|
||||
```json
|
||||
{
|
||||
"image_judge_system_prompt": "...see below...",
|
||||
"image_judge_user_prompt": "...see below...",
|
||||
"final_verdict_system_prompt": "...see below...",
|
||||
"final_verdict_user_prompt": "...{action_history_log}...{image_reasonings}..."
|
||||
}
|
||||
```
|
||||
Any `<field>_file` variant (e.g. `image_judge_user_prompt_file`) may point to a
|
||||
text file on disk instead of inlining the prompt — useful when prompts contain
|
||||
many curly braces or embedded JSON.
|
||||
|
||||
**Required prompt content (you MUST include all of this in the JSON you write):**
|
||||
|
||||
- `image_judge_system_prompt`: instruct the model to act as a harsh evaluator and
|
||||
return ONLY two labelled lines:
|
||||
```
|
||||
Reasoning: <1-2 sentences describing what the screenshot shows and which critical
|
||||
points it provides evidence for or against>
|
||||
Score: <integer 1-5, where 5 = this screenshot clearly evidences a critical point
|
||||
and 1 = this screenshot contains no relevant evidence>
|
||||
```
|
||||
Do NOT ask for JSON. The tool parses the labelled lines directly.
|
||||
|
||||
- `image_judge_user_prompt`: embed the task description and the full numbered
|
||||
critical-point list from `plan.md`. Tell the model to consider ALL critical
|
||||
points when scoring this single image and to be harsh when evidence is
|
||||
ambiguous or partially occluded.
|
||||
|
||||
- `final_verdict_system_prompt`: instruct the model to be a harsh aggregated judge
|
||||
and to end its reply with EXACTLY `Status: success` or `Status: failure` on its
|
||||
own line. Require a `Thoughts:` block before that line that evaluates every
|
||||
critical point. The tool extracts the verdict from the trailing `Status:` line.
|
||||
|
||||
- `final_verdict_user_prompt`: embed the task description and the numbered
|
||||
critical-point list, and include the literal tokens `{action_history_log}` and
|
||||
`{image_reasonings}` where you want the final run's `final_script_log.txt`
|
||||
content and the per-image reasonings injected. Do NOT hard-code a specific
|
||||
run's `final_script_log.txt`. The tool renders those tokens with Python
|
||||
`str.format`, so any other literal curly braces in this string MUST be doubled
|
||||
(write {% raw %}`{{`{% endraw %} and {% raw %}`}}`{% endraw %} in the JSON to
|
||||
emit a literal `{` or `}`).
|
||||
|
||||
**Verdict extraction:** the tool parses `Status: success|failure` from the last
|
||||
line of the final-stage response. Missing or malformed `Status:` counts as FAIL
|
||||
(exit code 1). Keep the verdict line clean.
|
||||
|
||||
**Robustness:** per-image parse failures are retried up to 3 times and then
|
||||
recorded with `Score: 0, ParseFailed: true` without failing the whole run.
|
||||
Transient model-API HTTP errors are retried with exponential backoff.
|
||||
|
||||
## Final Deliverable: Task Showcase JSON
|
||||
|
||||
In this mode, `final_script.py` is still the reusable task artifact, but its required output is a generic Task Showcase dataset:
|
||||
|
||||
```
|
||||
task_showcase/tasks/<short_id>/
|
||||
task.json
|
||||
report.json
|
||||
```
|
||||
|
||||
The existing renderer reads those two files and renders a dashboard/task page. Do not build a separate HTML app in `final_script.py`.
|
||||
|
||||
Required `task.json` shape:
|
||||
```
|
||||
{
|
||||
"task_id": "stable id, usually the CLI task_id if provided",
|
||||
"short_id": "url-safe slug matching the directory name",
|
||||
"title": "short human-readable title",
|
||||
"theme": "short category label",
|
||||
"cadence": "refresh cadence such as on demand, daily, every 6h",
|
||||
"level": "easy|medium|hard or another short difficulty label",
|
||||
"website": "primary site URL or starting URL",
|
||||
"task_prompt": "original user task prompt",
|
||||
"num_steps": 0
|
||||
}
|
||||
```
|
||||
|
||||
`task.json` field rules:
|
||||
- `short_id` must exactly match the `task_showcase/tasks/<short_id>/` directory name.
|
||||
- If the CLI `task_id` is absent, set `task_id` to the generated `short_id` or another stable id derived from the task prompt.
|
||||
- `website` must be a non-empty absolute `http(s)` URL for the primary site or starting URL.
|
||||
- All fields except `num_steps` must be non-empty strings.
|
||||
- `num_steps` must be an integer count of the `step N action:` entries from the clean final-script run. Successful non-empty runs should not leave it at `0`.
|
||||
|
||||
Required `report.json` shape:
|
||||
```
|
||||
{
|
||||
"sources": [
|
||||
{"name": "source display name", "url": "https://...", "note": "optional short note"}
|
||||
],
|
||||
"result": {
|
||||
"headline": "short headline",
|
||||
"sections": [
|
||||
{"type": "summary", "title": "...", "body": "..."},
|
||||
{"type": "table", "title": "...", "columns": ["..."], "rows": [["..."]]},
|
||||
{"type": "list", "title": "...", "entries": ["..."]},
|
||||
{"type": "kv", "title": "...", "entries": [["key", "value"]]},
|
||||
{"type": "cards", "title": "...", "entries": [
|
||||
{"title": "...", "subtitle": "...", "fields": [["key", "value"]], "url": "https://..."}
|
||||
]}
|
||||
]
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
Schema rules:
|
||||
- Every source used to answer the task must appear in `sources`.
|
||||
- `sources` must be a list of objects with non-empty string `name`, absolute `http(s)` string `url`, and optional string `note`.
|
||||
- `result.headline` must be a non-empty string.
|
||||
- `result.sections` must be a non-empty list of valid section objects. Empty or partial outcomes still need at least one `summary` or `list` section explaining the result.
|
||||
- Use only the section types above. Do not invent new section types.
|
||||
- Every section must have string `type` and non-empty string `title`.
|
||||
- `summary` sections require non-empty string `body`.
|
||||
- `table` sections require `columns: list[str]` and `rows: list[list[str|number]]`; every row must have the same length as `columns`.
|
||||
- `list` sections require `entries: list[str|number]`.
|
||||
- `kv` sections require `entries: list[[str, str|number]]`.
|
||||
- `cards` sections require `entries: list[object]`; every card requires non-empty string `title`, optional string `subtitle`, optional `fields: list[[str, str|number]]`, and optional absolute `http(s)` string `url`.
|
||||
- All scalar display values in sections must be strings or numbers that serialize cleanly to JSON. Normalize booleans, nulls, dates, complex objects, and arrays into strings before writing display fields. Do not emit comments, trailing commas, NaN, Infinity, Python tuples, or non-JSON values.
|
||||
- All source URLs and item URLs must be original links discovered during the clean run: visited, clicked, or extracted from the live source. Do not fabricate URLs, use screenshots as URLs, or use search-result/snippet URLs unless that page itself is the source.
|
||||
- For every result item with its own page, include the URL in `cards.entries[].url` or in a dedicated `URL` table column.
|
||||
- Render important final data in structured sections, not only in prose.
|
||||
- Empty or partial results must be visible in `report.json` with a summary/list section explaining what was checked and what failed.
|
||||
|
||||
## Renderer Requirement
|
||||
|
||||
The render step is separate from JSON generation. `final_script.py` must generate only the Task Showcase dataset, not a custom HTML app. Before declaring done, verify that the existing repo Task Showcase renderer can consume the generated `task.json` and `report.json`. The renderer may live outside the installed `webwright` package data, so locate the existing repo file rather than assuming it is importable as package data. If Flask or the renderer file is unavailable, do not install or generate it; validate the JSON files and report the exact renderer command or missing renderer path instead of claiming a completed render smoke test.
|
||||
|
||||
The runtime gate checks `self_reflect_result.json`, so do not run the final self_reflection pass until after Task Showcase JSON validation and the renderer smoke test have passed (or Flask unavailability has been handled exactly as described). If you change `final_script.py` or regenerated showcase JSON after self_reflection, rerun the clean final script in a new `final_runs/run_<id+1>/` folder and rerun self_reflection against that same run.
|
||||
|
||||
## Completion Gate
|
||||
|
||||
Set `"done": true` ONLY if ALL of the following are true:
|
||||
1. `plan.md` exists and every critical point is enumerated as a checklist item,
|
||||
with the intended Task Showcase report sections documented.
|
||||
2. `self_reflect_config.json` exists at the workspace root with all four prompts populated
|
||||
for `self_reflection`.
|
||||
3. `final_script.py` was executed successfully from scratch from `{{ workspace_dir }}`
|
||||
for a new `final_runs/run_<id>/` artifact folder, producing `final_script_log.txt`,
|
||||
all critical-point screenshots, and the Task Showcase JSON files.
|
||||
4. `task_showcase/tasks/<short_id>/task.json` exists and validates.
|
||||
5. `task_showcase/tasks/<short_id>/report.json` exists and validates.
|
||||
6. The Flask renderer was smoke-tested against `{{ workspace_dir }}/task_showcase/tasks`,
|
||||
or Flask was unavailable and the exact render command was reported.
|
||||
7. `python -m webwright.tools.self_reflection --config self_reflect_config.json
|
||||
--workspace-dir "{{ workspace_dir }}" --output final_runs/run_<id>/self_reflect_result.json`
|
||||
was executed after JSON/render validation, against that same run, exited 0, and wrote
|
||||
`final_runs/run_<id>/self_reflect_result.json` with `"predicted_label": 1`.
|
||||
8. You have run `ls -R final_runs/run_<id>`,
|
||||
`ls -R final_runs/run_<id>/screenshots`, and
|
||||
`cat final_runs/run_<id>/final_script_log.txt` to confirm the artifacts and
|
||||
logs are in place.
|
||||
9. The final response names the generated JSON paths and the render command.
|
||||
|
||||
Do NOT declare done if `self_reflection` exits non-zero, if `predicted_label` is
|
||||
not 1, if the run folder is missing, if required screenshots are missing, if the
|
||||
script failed to run, if the checklist in `plan.md` is incomplete, if the JSON
|
||||
files are missing or malformed, if they were written outside `{{ workspace_dir }}`,
|
||||
or if the task page does not render when Flask is available. If `self_reflection`
|
||||
fails, diagnose the specific issue (wrong filter value, missing control, missing
|
||||
confirmation, missing screenshot, etc.), fix `final_script.py`, re-run it in a new
|
||||
`final_runs/run_<id+1>/` folder, and re-run `self_reflection` against the new run.
|
||||
Do NOT edit `self_reflect_config.json` between attempts unless a prompt itself is
|
||||
objectively wrong.
|
||||
|
||||
instance_template: |
|
||||
Task: {{ task }}
|
||||
{% if task_id %}Task ID: {{ task_id }}
|
||||
{% endif %}{% if start_url %}Start URL: {{ start_url }}
|
||||
{% endif %}Workspace root: {{ workspace_dir }}
|
||||
Task metadata JSON: {{ task_metadata_path }}
|
||||
Required final script path: {{ final_script_path }}
|
||||
Generated showcase tasks dir: {{ workspace_dir }}/task_showcase/tasks
|
||||
|
||||
<instructions>
|
||||
# Task Instructions
|
||||
|
||||
You're solving a user-specified web task through a stateless local terminal + workspace harness and converting the result into Task Showcase JSON for the generic renderer under `assets/task_showcase`.
|
||||
|
||||
<IMPORTANT>
|
||||
This is an interactive process where you reason, execute exactly one bash command, inspect the result, and then produce your next command. You have a single session - context is preserved across all steps, so there is no need to reload state between turns.
|
||||
</IMPORTANT>
|
||||
|
||||
## Harness Rules
|
||||
|
||||
- Work only inside `{{ workspace_dir }}`.
|
||||
- Keep generated code, screenshots, logs, scratch files, notes, and showcase JSON **only** in `{{ workspace_dir }}`.
|
||||
- The required final artifact is `{{ final_script_path }}`.
|
||||
- Create `final_runs/run_<id>/` folders for every clean execution of the final script. Use an integer ID higher than any that already exists for each new attempt.
|
||||
- Keep the canonical final script at `{{ final_script_path }}`. For each clean run, copy that exact script into `final_runs/run_<id>/final_script.py` and store that run's `final_script_log.txt` and final verification screenshots inside the same run folder.
|
||||
- Do not create an empty higher-numbered `final_runs/run_<id>/` folder. The latest numbered run must be the run that contains screenshots, logs, showcase JSON output, and the matching `self_reflect_result.json`.
|
||||
- The required rendered-data artifacts are:
|
||||
- `{{ workspace_dir }}/task_showcase/tasks/<short_id>/task.json`
|
||||
- `{{ workspace_dir }}/task_showcase/tasks/<short_id>/report.json`
|
||||
- Use `{{ task_id }}` as the preferred `<short_id>` when it is present and already URL-safe; otherwise derive a lowercase slug from the task title.
|
||||
- The browser mode is `{{ browser_mode }}`. Match generated scripts to that mode (Browserbase cloud session vs. local Playwright launch).
|
||||
|
||||
## Web Task Rules
|
||||
|
||||
- Do not guess UI interactions. Use printed evidence from the current run.
|
||||
- Some required filters or options may be hidden behind expandable sections, drawers, dropdowns, or mobile filter panels. Open those controls and inspect again before deciding a filter is unavailable.
|
||||
- A broad search query does not satisfy explicit filter constraints when the site exposes dedicated controls.
|
||||
- Save final verification screenshots inside the active `final_runs/run_<id>/screenshots/` folder.
|
||||
- Print concise ARIA snapshots, URLs, titles, visible labels, extracted records, and any extracted state needed for the next step.
|
||||
|
||||
## Task Success Criteria
|
||||
|
||||
1. Filtered results must be displayed correctly. Missing selection, missing confirmation, or no visible effect = failure.
|
||||
2. Specific filter conditions ("best," "highest," "cheapest," "latest," "lowest," etc.) must be applied using the filter/sort function.
|
||||
3. Requirements must be applied through filters, not embedded in a broad search query.
|
||||
4. Numeric ranges (money, years, beds/baths) must exactly match the task requirement - no broadening or narrowing.
|
||||
5. Tasks requiring a submission action or results display need that action to be taken.
|
||||
6. Empty results are OK if the correct action was performed.
|
||||
7. All explicit filters must use site controls when those controls exist.
|
||||
8. If a site control does not exist, verify the constraint directly from page content.
|
||||
9. The final answer data must be represented in `report.json` structured sections, not only in `final_response`.
|
||||
|
||||
## Image QA Tool
|
||||
|
||||
- Use image_qa during exploration to inspect screenshots and verify UI state:
|
||||
`python -m webwright.tools.image_qa --workspace-dir "{{ workspace_dir }}" --image screenshots/example.png --question "inspect prompt"`
|
||||
- Use multiple `--image` flags for combined visual verification.
|
||||
- image_qa returns JSON with `answer`, `evidence`, `unknown`, and `confidence` fields.
|
||||
|
||||
## Recommended Workflow
|
||||
|
||||
1. **Planning**: Parse the task into a list of critical points - every explicit constraint, filter, sort, selection, datum, or source requirement that must be satisfied. Write them to `plan.md` as a checklist and add the intended Task Showcase report sections:
|
||||
```
|
||||
# Critical Points
|
||||
- [ ] CP1: <description of constraint/filter/action>
|
||||
- [ ] CP2: <description of constraint/filter/action>
|
||||
...
|
||||
|
||||
# Report Sections
|
||||
- <section type>: <title and purpose>
|
||||
- ...
|
||||
```
|
||||
Each critical point must be independently verifiable from a screenshot, log entry, extracted source data, or generated JSON field.
|
||||
|
||||
2. **Author self_reflect_config.json (once)**: Write `{{ workspace_dir }}/self_reflect_config.json` containing only the four prompts (`image_judge_system_prompt`, `image_judge_user_prompt`, `final_verdict_system_prompt`, `final_verdict_user_prompt`) for `webwright.tools.self_reflection`. Embed the full critical-point list from `plan.md` and the task description into the user prompts, but keep the prompts generic - this file is reused verbatim for every `self_reflection` invocation, so do NOT hard-code a specific run id, screenshot filename, or `final_script_log.txt` content.
|
||||
|
||||
3. **Exploration**: Inspect `{{ task_metadata_path }}`, create exploration scripts, identify every required filter control, source URL, and output field. Use `image_qa` during exploration to verify UI state when visual evidence matters.
|
||||
|
||||
4. **Final script**: Write `final_script.py` at `{{ final_script_path }}`, then execute it from `{{ workspace_dir }}` for a new `final_runs/run_<id>/` artifact folder. The script must copy itself into that run folder, produce screenshots and action logs as described in **Final Script Instrumentation**, collect or refresh the task data, write the two JSON files under `{{ workspace_dir }}/task_showcase/tasks/<short_id>/`, and print the generated paths. The script should create parent directories, overwrite stale JSON for the same short_id, and fail loudly on malformed data.
|
||||
|
||||
5. **Validate JSON**: Confirm the clean run wrote:
|
||||
- `{{ workspace_dir }}/task_showcase/tasks/<short_id>/task.json`
|
||||
- `{{ workspace_dir }}/task_showcase/tasks/<short_id>/report.json`
|
||||
Then validate both JSON files with Python:
|
||||
- `task.json.short_id` exactly matches the `<short_id>` directory name.
|
||||
- `task.json` has all required metadata keys, with non-empty strings except integer `num_steps`.
|
||||
- `task.json.website` is an absolute `http(s)` URL.
|
||||
- `report.json` has `sources`, non-empty `result.headline`, and non-empty `result.sections`.
|
||||
- `sources` is a list of `{name, url, note?}` objects with absolute `http(s)` URLs.
|
||||
- Every section type is one of `summary`, `table`, `list`, `kv`, `cards`.
|
||||
- Every section satisfies the per-type contract from the system prompt, including table row widths and kv/card field pairs.
|
||||
- Every source and item URL that appears in the result is an original discovered absolute `http(s)` link.
|
||||
|
||||
6. **Render smoke test**:
|
||||
- Locate the existing repo renderer `assets/task_showcase/app.py`. Prefer this resolver:
|
||||
```
|
||||
python - <<'PY'
|
||||
from pathlib import Path
|
||||
import webwright
|
||||
starts = [Path.cwd().resolve(), Path(webwright.__file__).resolve()]
|
||||
candidates = []
|
||||
for start in starts:
|
||||
candidates.extend(parent / "assets" / "task_showcase" / "app.py" for parent in [start, *start.parents])
|
||||
for candidate in candidates:
|
||||
if candidate.exists():
|
||||
print(candidate)
|
||||
break
|
||||
else:
|
||||
raise SystemExit("assets/task_showcase/app.py not found from cwd or webwright package path")
|
||||
PY
|
||||
```
|
||||
- Start `python <app.py> --tasks-dir "{{ workspace_dir }}/task_showcase/tasks" --host 127.0.0.1 --port <free_port>`.
|
||||
- Fetch `/` and `/task/<short_id>` and confirm both return HTTP 200 and the task headline appears in the task page HTML.
|
||||
- Stop the server before declaring done.
|
||||
- If Flask or the renderer file is unavailable, do not install or generate it. Validate JSON and include the exact render command or missing renderer path in `final_response`.
|
||||
- If JSON validation or renderer smoke testing fails, fix `final_script.py`, rerun it in a new `final_runs/run_<id+1>/` folder, and repeat JSON validation and render smoke testing before self_reflection.
|
||||
|
||||
7. **Run self_reflection**: Execute `python -m webwright.tools.self_reflection --config self_reflect_config.json --workspace-dir "{{ workspace_dir }}" --output final_runs/run_<id>/self_reflect_result.json` only after JSON validation and renderer smoke testing have passed for the same run. The tool auto-attaches screenshots from `final_runs/run_<id>/screenshots` when that is the latest numbered run with screenshots - you do NOT pass an image list. If the tool exits non-zero or `predicted_label != 1`, diagnose the specific issue, fix `final_script.py`, re-run it in a new `final_runs/run_<id+1>/` folder, and re-invoke JSON validation, renderer smoke testing, and self_reflection against the new run. Do NOT edit `self_reflect_config.json` between attempts.
|
||||
|
||||
8. **Declare done**: Set `"done": true` ONLY after both Task Showcase JSON files validate, the renderer was smoke-tested or Flask unavailability was handled exactly as described above, `self_reflection` exits 0, and `self_reflect_result.json` reports `"predicted_label": 1` for the latest run. The external judge reads that same `self_reflect_result.json` as the final verdict. Declaring done in any other state is a failure.
|
||||
|
||||
## Final Script Instrumentation
|
||||
|
||||
`final_script.py` must:
|
||||
- be stored at `{{ final_script_path }}` and copied as `final_runs/run_<id>/final_script.py` for the clean run
|
||||
- collect or refresh the task data from the relevant source(s)
|
||||
- save critical-point screenshots as `final_runs/run_<id>/screenshots/final_execution_<step_number>_<action>.png`
|
||||
- create or reset `final_runs/run_<id>/final_script_log.txt` at the start of each clean run
|
||||
- write `step <step_number> action: <reason and action description>` to the log for every constraint-relevant interaction
|
||||
- write `{{ workspace_dir }}/task_showcase/tasks/<short_id>/task.json`
|
||||
- write `{{ workspace_dir }}/task_showcase/tasks/<short_id>/report.json`
|
||||
- optionally copy `final_script_log.txt`, `steps.jsonl`, and screenshots into `{{ workspace_dir }}/task_showcase/tasks/<short_id>/` if you want the renderer to display run traces; these copies are not a substitute for the required `final_runs/run_<id>/` artifacts
|
||||
- print the generated JSON paths
|
||||
- include enough logging or stdout details to diagnose source access, empty results, and render validation failures
|
||||
- each screenshot should correspond to a critical point from `plan.md` so that `self_reflection` can verify it
|
||||
- keep all generated files inside `{{ workspace_dir }}`
|
||||
|
||||
This instrumentation is mandatory because both `self_reflection`, the external judge, and the Task Showcase renderer evaluate these artifacts.
|
||||
|
||||
## Completion Gate
|
||||
|
||||
Set `"done": true` ONLY if ALL of the following are true:
|
||||
1. `plan.md` exists with all critical points identified and the intended report sections documented.
|
||||
2. `self_reflect_config.json` exists with all four prompts populated for `self_reflection`.
|
||||
3. `{{ final_script_path }}` was run from scratch from `{{ workspace_dir }}` for a new `final_runs/run_<id>/` folder, producing the run-local script copy, `final_script_log.txt`, all critical-point screenshots, and the Task Showcase JSON files.
|
||||
4. `task_showcase/tasks/<short_id>/task.json` exists and validates.
|
||||
5. `task_showcase/tasks/<short_id>/report.json` exists and validates.
|
||||
6. The Flask renderer was smoke-tested against `{{ workspace_dir }}/task_showcase/tasks`, or Flask was unavailable and the exact render command was reported.
|
||||
7. `python -m webwright.tools.self_reflection --config self_reflect_config.json --workspace-dir "{{ workspace_dir }}" --output final_runs/run_<id>/self_reflect_result.json` was executed after JSON/render validation against that same run, exited 0, and wrote `final_runs/run_<id>/self_reflect_result.json` with `"predicted_label": 1`.
|
||||
8. `ls -R final_runs/run_<id>`, `ls -R final_runs/run_<id>/screenshots`, and `cat final_runs/run_<id>/final_script_log.txt` confirm the expected artifacts.
|
||||
9. `final_response` names the generated JSON paths and the render command.
|
||||
|
||||
Do NOT declare done if `self_reflection` exits non-zero, if `predicted_label` is not 1, if the run folder is missing, if required screenshots are missing, if `self_reflection` has not been run against the latest `final_runs/run_<id>/`, if the JSON files are missing or malformed, if they were written outside `{{ workspace_dir }}`, or if the task page does not render when Flask is available.
|
||||
</instructions>
|
||||
@@ -0,0 +1,24 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import copy
|
||||
import importlib
|
||||
|
||||
from webwright import Environment
|
||||
|
||||
_ENVIRONMENT_MAPPING = {
|
||||
"local_browser": "webwright.environments.local_browser.LocalBrowserEnvironment",
|
||||
"local_workspace": "webwright.environments.local_workspace.LocalWorkspaceEnvironment",
|
||||
}
|
||||
|
||||
|
||||
def get_environment_class(spec: str) -> type[Environment]:
|
||||
full_path = _ENVIRONMENT_MAPPING.get(spec, spec)
|
||||
module_name, class_name = full_path.rsplit(".", 1)
|
||||
module = importlib.import_module(module_name)
|
||||
return getattr(module, class_name)
|
||||
|
||||
|
||||
def get_environment(config: dict, *, default_type: str = "local_workspace") -> Environment:
|
||||
copied = copy.deepcopy(config)
|
||||
environment_class = copied.pop("environment_class", default_type)
|
||||
return get_environment_class(environment_class)(**copied)
|
||||
@@ -0,0 +1,567 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
import io
|
||||
import json
|
||||
import os
|
||||
import shutil
|
||||
import subprocess
|
||||
import sys
|
||||
import textwrap
|
||||
import time
|
||||
import traceback
|
||||
from contextlib import redirect_stderr, redirect_stdout
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
from urllib.parse import quote, urlparse
|
||||
from urllib.request import ProxyHandler, Request, build_opener
|
||||
|
||||
from pydantic import BaseModel, Field, field_validator
|
||||
|
||||
_BROWSER_MODES = {"local_cdp", "local_launch", "local_persistent"}
|
||||
_DEFAULT_LOCAL_CDP_URL = "http://127.0.0.1:9222"
|
||||
_DEFAULT_LOCAL_CDP_USER_DATA_DIR = Path("~/.cache/webwright/edge-profile")
|
||||
_CHROMIUM_EXECUTABLE_CANDIDATES = (
|
||||
"/Applications/Microsoft Edge.app/Contents/MacOS/Microsoft Edge",
|
||||
"~/Applications/Microsoft Edge.app/Contents/MacOS/Microsoft Edge",
|
||||
"/Applications/Google Chrome.app/Contents/MacOS/Google Chrome",
|
||||
"~/Applications/Google Chrome.app/Contents/MacOS/Google Chrome",
|
||||
"google-chrome",
|
||||
"google-chrome-stable",
|
||||
"chromium-browser",
|
||||
"chromium",
|
||||
"chrome",
|
||||
)
|
||||
_LOCAL_CDP_OPENER = build_opener(ProxyHandler({}))
|
||||
|
||||
|
||||
def _urlopen_local_cdp(url_or_request: str | Request, *, timeout: float):
|
||||
return _LOCAL_CDP_OPENER.open(url_or_request, timeout=timeout)
|
||||
|
||||
|
||||
def _local_cdp_origin(cdp_url: str) -> str:
|
||||
parsed = urlparse(cdp_url or _DEFAULT_LOCAL_CDP_URL)
|
||||
scheme = parsed.scheme or "http"
|
||||
netloc = parsed.netloc or parsed.path
|
||||
if not netloc:
|
||||
netloc = urlparse(_DEFAULT_LOCAL_CDP_URL).netloc
|
||||
return f"{scheme}://{netloc}"
|
||||
|
||||
|
||||
def _local_cdp_port(cdp_url: str) -> int:
|
||||
parsed = urlparse(cdp_url or _DEFAULT_LOCAL_CDP_URL)
|
||||
if parsed.port is not None:
|
||||
return parsed.port
|
||||
return 443 if parsed.scheme == "https" else 80
|
||||
|
||||
|
||||
def _is_local_cdp_available(cdp_url: str, *, timeout_seconds: float = 0.5) -> bool:
|
||||
try:
|
||||
with _urlopen_local_cdp(
|
||||
f"{_local_cdp_origin(cdp_url).rstrip('/')}/json/version",
|
||||
timeout=timeout_seconds,
|
||||
) as response:
|
||||
return 200 <= response.status < 300
|
||||
except Exception:
|
||||
return False
|
||||
|
||||
|
||||
def _local_cdp_json_url(cdp_url: str, path: str) -> str:
|
||||
return f"{_local_cdp_origin(cdp_url).rstrip('/')}{path}"
|
||||
|
||||
|
||||
def _local_cdp_page_targets(cdp_url: str, *, timeout_seconds: float = 0.5) -> list[dict[str, Any]]:
|
||||
try:
|
||||
with _urlopen_local_cdp(
|
||||
_local_cdp_json_url(cdp_url, "/json/list"),
|
||||
timeout=timeout_seconds,
|
||||
) as response:
|
||||
if not 200 <= response.status < 300:
|
||||
return []
|
||||
payload = json.loads(response.read().decode("utf-8"))
|
||||
except Exception:
|
||||
return []
|
||||
if not isinstance(payload, list):
|
||||
return []
|
||||
return [
|
||||
target
|
||||
for target in payload
|
||||
if isinstance(target, dict) and target.get("type") == "page"
|
||||
]
|
||||
|
||||
|
||||
def _ensure_local_cdp_page_target(cdp_url: str, *, timeout_seconds: float = 1.0) -> None:
|
||||
if _local_cdp_page_targets(cdp_url, timeout_seconds=timeout_seconds):
|
||||
return
|
||||
|
||||
target_url = f"{_local_cdp_json_url(cdp_url, '/json/new')}?{quote('about:blank', safe='')}"
|
||||
request = Request(target_url, method="PUT")
|
||||
with _urlopen_local_cdp(request, timeout=timeout_seconds) as response:
|
||||
if not 200 <= response.status < 300:
|
||||
raise RuntimeError(f"Could not create a local CDP page target: {cdp_url}")
|
||||
|
||||
|
||||
def _resolve_local_cdp_url(configured_url: str, *, explicit: bool) -> str:
|
||||
if explicit and configured_url:
|
||||
return configured_url
|
||||
return (
|
||||
os.environ.get("LOCAL_BROWSER_CDP_URL")
|
||||
or os.environ.get("BROWSER_CDP_URL")
|
||||
or configured_url
|
||||
or _DEFAULT_LOCAL_CDP_URL
|
||||
)
|
||||
|
||||
|
||||
def _resolve_user_data_dir(configured_dir: Path, *, explicit: bool) -> Path:
|
||||
if explicit:
|
||||
return configured_dir
|
||||
env_dir = os.environ.get("LOCAL_BROWSER_USER_DATA_DIR") or os.environ.get("BROWSER_USER_DATA_DIR")
|
||||
return Path(env_dir).expanduser() if env_dir else _DEFAULT_LOCAL_CDP_USER_DATA_DIR.expanduser()
|
||||
|
||||
|
||||
def _find_chromium_executable(explicit_path: str = "") -> str:
|
||||
candidates = []
|
||||
if explicit_path:
|
||||
candidates.append(explicit_path)
|
||||
candidates.extend(
|
||||
value
|
||||
for value in (os.environ.get("LOCAL_BROWSER_EXECUTABLE"), os.environ.get("BROWSER_EXECUTABLE"))
|
||||
if value
|
||||
)
|
||||
candidates.extend(_CHROMIUM_EXECUTABLE_CANDIDATES)
|
||||
|
||||
for candidate in candidates:
|
||||
expanded = os.path.expanduser(candidate)
|
||||
if Path(expanded).exists():
|
||||
return expanded
|
||||
resolved = shutil.which(expanded)
|
||||
if resolved:
|
||||
return resolved
|
||||
raise FileNotFoundError(
|
||||
"Could not find Chrome/Chromium. Set local_cdp_executable or LOCAL_BROWSER_EXECUTABLE."
|
||||
)
|
||||
|
||||
|
||||
def _macos_open_app_name(executable: str) -> str:
|
||||
if sys.platform != "darwin":
|
||||
return ""
|
||||
if "/Microsoft Edge.app/" in executable:
|
||||
return "Microsoft Edge"
|
||||
if "/Google Chrome.app/" in executable:
|
||||
return "Google Chrome"
|
||||
return ""
|
||||
|
||||
|
||||
class LocalBrowserEnvironmentConfig(BaseModel):
|
||||
start_url: str | None = None
|
||||
browser_mode: str = "local_launch"
|
||||
headless: bool = False
|
||||
devtools: bool = False
|
||||
keep_open_on_exit: bool = False
|
||||
prompt_before_close: bool = False
|
||||
slow_mo_ms: int = 50
|
||||
browser_width: int = 1280
|
||||
browser_height: int = 1440
|
||||
browser_timeout_ms: int = 10000
|
||||
browser_navigation_timeout_ms: int = 30000
|
||||
step_execution_timeout_ms: int = 20000
|
||||
observation_timeout_ms: int = 5000
|
||||
output_dir: Path = Path("outputs/default")
|
||||
user_data_dir: Path = _DEFAULT_LOCAL_CDP_USER_DATA_DIR
|
||||
launch_args: list[str] = Field(default_factory=list)
|
||||
local_cdp_url: str = _DEFAULT_LOCAL_CDP_URL
|
||||
local_cdp_new_page: bool = True
|
||||
local_cdp_close_page_on_exit: bool = False
|
||||
local_cdp_auto_start: bool = True
|
||||
local_cdp_executable: str = ""
|
||||
local_cdp_startup_timeout_seconds: float = 10
|
||||
local_cdp_close_started_browser_on_exit: bool = False
|
||||
|
||||
@field_validator("browser_mode")
|
||||
@classmethod
|
||||
def validate_browser_mode(cls, value: str) -> str:
|
||||
normalized = value.strip().lower().replace("-", "_")
|
||||
if normalized not in _BROWSER_MODES:
|
||||
raise ValueError(
|
||||
f"browser_mode must be one of: {', '.join(sorted(_BROWSER_MODES))}"
|
||||
)
|
||||
return normalized
|
||||
|
||||
|
||||
class LocalBrowserEnvironment:
|
||||
"""Live local Playwright browser environment.
|
||||
|
||||
The environment owns the browser/page and executes each model action as an async
|
||||
Python snippet with ``page``, ``context``, ``browser``, ``playwright``, and
|
||||
``task`` already available.
|
||||
"""
|
||||
|
||||
def __init__(self, *, config_class: type = LocalBrowserEnvironmentConfig, **kwargs):
|
||||
self.config = config_class(**kwargs)
|
||||
fields_set = getattr(self.config, "model_fields_set", None)
|
||||
if fields_set is None:
|
||||
fields_set = getattr(self.config, "__fields_set__", set())
|
||||
self._config_fields_set = set(fields_set)
|
||||
self.config.local_cdp_url = _resolve_local_cdp_url(
|
||||
self.config.local_cdp_url,
|
||||
explicit="local_cdp_url" in self._config_fields_set,
|
||||
)
|
||||
self.config.output_dir = self.config.output_dir.expanduser()
|
||||
self.config.user_data_dir = _resolve_user_data_dir(
|
||||
self.config.user_data_dir,
|
||||
explicit="user_data_dir" in self._config_fields_set,
|
||||
)
|
||||
self._playwright = None
|
||||
self._browser = None
|
||||
self._context = None
|
||||
self._page = None
|
||||
self._local_cdp_page = None
|
||||
self._loop: asyncio.AbstractEventLoop | None = None
|
||||
self._local_cdp_process: subprocess.Popen | None = None
|
||||
self._connected_over_cdp = False
|
||||
self._step_index = 0
|
||||
self._prepared_task: dict[str, Any] = {}
|
||||
self._console_history: list[str] = []
|
||||
self._step_console: list[str] = []
|
||||
self._step_python_code = ""
|
||||
self._step_python_output = ""
|
||||
|
||||
def _screenshots_dir(self) -> Path:
|
||||
return self.config.output_dir / "screenshots"
|
||||
|
||||
def _steps_dir(self) -> Path:
|
||||
return self.config.output_dir / "steps"
|
||||
|
||||
def prepare(self, **kwargs) -> None:
|
||||
self._prepared_task = dict(kwargs)
|
||||
self._step_index = 0
|
||||
self._console_history = []
|
||||
self._step_console = []
|
||||
start_url = kwargs.get("start_url") or self.config.start_url
|
||||
if start_url:
|
||||
self.config.start_url = str(start_url)
|
||||
|
||||
self.config.output_dir.mkdir(parents=True, exist_ok=True)
|
||||
self._steps_dir().mkdir(parents=True, exist_ok=True)
|
||||
self._screenshots_dir().mkdir(parents=True, exist_ok=True)
|
||||
(self.config.output_dir / "task.json").write_text(
|
||||
json.dumps(kwargs, indent=2),
|
||||
encoding="utf-8",
|
||||
)
|
||||
self._run(self._prepare_async())
|
||||
|
||||
def _ensure_loop(self) -> asyncio.AbstractEventLoop:
|
||||
if self._loop is None or self._loop.is_closed():
|
||||
self._loop = asyncio.new_event_loop()
|
||||
return self._loop
|
||||
|
||||
def _run(self, coro):
|
||||
loop = self._ensure_loop()
|
||||
return loop.run_until_complete(coro)
|
||||
|
||||
def _ensure_local_cdp_browser(self) -> None:
|
||||
if _is_local_cdp_available(self.config.local_cdp_url):
|
||||
return
|
||||
if not self.config.local_cdp_auto_start:
|
||||
raise RuntimeError(
|
||||
"Local CDP endpoint is not reachable. Start Chrome/Chromium with "
|
||||
f"remote debugging enabled for {self.config.local_cdp_url}."
|
||||
)
|
||||
|
||||
self.config.user_data_dir.mkdir(parents=True, exist_ok=True)
|
||||
browser_args = list(self.config.launch_args)
|
||||
executable = _find_chromium_executable(self.config.local_cdp_executable)
|
||||
browser_flags = [
|
||||
"--remote-debugging-address=127.0.0.1",
|
||||
f"--remote-debugging-port={_local_cdp_port(self.config.local_cdp_url)}",
|
||||
f"--user-data-dir={self.config.user_data_dir}",
|
||||
"--no-first-run",
|
||||
"--no-default-browser-check",
|
||||
*browser_args,
|
||||
]
|
||||
app_name = _macos_open_app_name(executable)
|
||||
launched_with_open = bool(app_name)
|
||||
command = (
|
||||
["open", "-na", app_name, "--args", *browser_flags]
|
||||
if launched_with_open
|
||||
else [executable, *browser_flags]
|
||||
)
|
||||
self._local_cdp_process = subprocess.Popen(
|
||||
command,
|
||||
stdout=subprocess.DEVNULL,
|
||||
stderr=subprocess.DEVNULL,
|
||||
start_new_session=True,
|
||||
)
|
||||
|
||||
deadline = time.monotonic() + self.config.local_cdp_startup_timeout_seconds
|
||||
while time.monotonic() < deadline:
|
||||
if not launched_with_open and self._local_cdp_process.poll() is not None:
|
||||
raise RuntimeError(
|
||||
f"Chrome/Chromium exited before CDP became available: {self.config.local_cdp_url}"
|
||||
)
|
||||
if _is_local_cdp_available(self.config.local_cdp_url):
|
||||
return
|
||||
time.sleep(0.2)
|
||||
|
||||
self._local_cdp_process.terminate()
|
||||
self._local_cdp_process = None
|
||||
raise TimeoutError(f"Timed out waiting for local CDP endpoint: {self.config.local_cdp_url}")
|
||||
|
||||
async def _prepare_async(self) -> None:
|
||||
from playwright.async_api import async_playwright
|
||||
|
||||
if self._page is not None and self._context is not None:
|
||||
return
|
||||
|
||||
self._playwright = await async_playwright().start()
|
||||
chromium = self._playwright.chromium
|
||||
launch_args = list(self.config.launch_args)
|
||||
if self.config.devtools:
|
||||
launch_args.append("--auto-open-devtools-for-tabs")
|
||||
launch_kwargs = {
|
||||
"headless": self.config.headless,
|
||||
"slow_mo": self.config.slow_mo_ms,
|
||||
"args": launch_args,
|
||||
}
|
||||
|
||||
if self.config.browser_mode == "local_cdp":
|
||||
self._ensure_local_cdp_browser()
|
||||
_ensure_local_cdp_page_target(self.config.local_cdp_url)
|
||||
self._browser = await chromium.connect_over_cdp(self.config.local_cdp_url)
|
||||
self._connected_over_cdp = True
|
||||
self._context = (
|
||||
self._browser.contexts[0]
|
||||
if self._browser.contexts
|
||||
else await self._browser.new_context(no_viewport=True)
|
||||
)
|
||||
if self.config.local_cdp_new_page or not self._context.pages:
|
||||
self._page = await self._context.new_page()
|
||||
self._local_cdp_page = self._page
|
||||
else:
|
||||
self._page = self._context.pages[0]
|
||||
elif self.config.browser_mode == "local_persistent":
|
||||
self.config.user_data_dir.mkdir(parents=True, exist_ok=True)
|
||||
self._context = await chromium.launch_persistent_context(
|
||||
user_data_dir=str(self.config.user_data_dir),
|
||||
viewport={
|
||||
"width": self.config.browser_width,
|
||||
"height": self.config.browser_height,
|
||||
},
|
||||
**launch_kwargs,
|
||||
)
|
||||
self._browser = self._context.browser
|
||||
self._page = self._context.pages[0] if self._context.pages else await self._context.new_page()
|
||||
else:
|
||||
self._browser = await chromium.launch(**launch_kwargs)
|
||||
self._context = await self._browser.new_context(
|
||||
viewport={
|
||||
"width": self.config.browser_width,
|
||||
"height": self.config.browser_height,
|
||||
}
|
||||
)
|
||||
self._page = await self._context.new_page()
|
||||
|
||||
self._context.set_default_timeout(self.config.browser_timeout_ms)
|
||||
self._context.set_default_navigation_timeout(self.config.browser_navigation_timeout_ms)
|
||||
self._attach_page_listeners(self._page)
|
||||
if self.config.start_url:
|
||||
await self._page.goto(self.config.start_url, wait_until="domcontentloaded")
|
||||
|
||||
def _attach_page_listeners(self, page: Any) -> None:
|
||||
page.on("console", self._on_console_message)
|
||||
page.on("pageerror", self._on_page_error)
|
||||
|
||||
def _on_console_message(self, message: Any) -> None:
|
||||
text = getattr(message, "text", "")
|
||||
if callable(text):
|
||||
text = text()
|
||||
line = str(text)
|
||||
self._console_history.append(line)
|
||||
self._step_console.append(line)
|
||||
|
||||
def _on_page_error(self, error: Any) -> None:
|
||||
line = f"Page error: {error}"
|
||||
self._console_history.append(line)
|
||||
self._step_console.append(line)
|
||||
|
||||
def execute(self, action: dict[str, Any], cwd: str = "") -> dict[str, Any]:
|
||||
del cwd
|
||||
return self._run(self._execute_async(action))
|
||||
|
||||
async def _execute_async(self, action: dict[str, Any]) -> dict[str, Any]:
|
||||
self._step_index += 1
|
||||
self._step_console = []
|
||||
self._step_python_output = ""
|
||||
self._step_python_code = str(action.get("python_code", "") or "")
|
||||
self._persist_step_code(self._step_python_code)
|
||||
|
||||
success = True
|
||||
exception_text = ""
|
||||
try:
|
||||
if self._step_python_code.strip():
|
||||
await asyncio.wait_for(
|
||||
self._run_python_code(self._step_python_code),
|
||||
timeout=self.config.step_execution_timeout_ms / 1000,
|
||||
)
|
||||
await self._wait_for_observation_ready()
|
||||
except Exception:
|
||||
success = False
|
||||
exception_text = traceback.format_exc()
|
||||
|
||||
observation = await self._capture_observation(
|
||||
success=success,
|
||||
exception_text=exception_text,
|
||||
)
|
||||
return {
|
||||
"output": self._step_python_output,
|
||||
"returncode": 0 if success else 1,
|
||||
"exception_info": exception_text,
|
||||
"observation": observation,
|
||||
}
|
||||
|
||||
def _persist_step_code(self, python_code: str) -> None:
|
||||
step_path = self._steps_dir() / f"step_{self._step_index:04d}.py"
|
||||
step_path.write_text(python_code, encoding="utf-8")
|
||||
|
||||
script_path = self.config.output_dir / "script.py"
|
||||
with script_path.open("a", encoding="utf-8") as handle:
|
||||
handle.write(f"\n\n# Step {self._step_index}\n")
|
||||
handle.write(python_code)
|
||||
handle.write("\n")
|
||||
|
||||
async def _run_python_code(self, python_code: str) -> None:
|
||||
if self._page is None or self._context is None or self._playwright is None:
|
||||
raise RuntimeError("Browser environment was not prepared.")
|
||||
|
||||
buffer = io.StringIO()
|
||||
globals_dict: dict[str, Any] = {"asyncio": asyncio}
|
||||
locals_dict: dict[str, Any] = {}
|
||||
wrapped = "async def __agent_step__(page, context, browser, playwright, task):\n"
|
||||
wrapped += textwrap.indent(python_code, " ")
|
||||
with redirect_stdout(buffer), redirect_stderr(buffer):
|
||||
exec(wrapped, globals_dict, locals_dict)
|
||||
await locals_dict["__agent_step__"](
|
||||
self._page,
|
||||
self._context,
|
||||
self._browser,
|
||||
self._playwright,
|
||||
self._prepared_task,
|
||||
)
|
||||
self._step_python_output = buffer.getvalue()
|
||||
|
||||
async def _wait_for_observation_ready(self) -> None:
|
||||
if self._page is None:
|
||||
return
|
||||
try:
|
||||
await self._page.wait_for_load_state(
|
||||
"domcontentloaded",
|
||||
timeout=self.config.observation_timeout_ms,
|
||||
)
|
||||
except Exception:
|
||||
pass
|
||||
|
||||
async def _capture_observation(self, *, success: bool, exception_text: str) -> dict[str, Any]:
|
||||
page = self._page
|
||||
url = ""
|
||||
title = ""
|
||||
aria_snapshot = ""
|
||||
screenshot_path: Path | None = None
|
||||
|
||||
if page is not None:
|
||||
try:
|
||||
url = page.url
|
||||
except Exception:
|
||||
url = self.config.start_url or ""
|
||||
try:
|
||||
title = await page.title()
|
||||
except Exception:
|
||||
title = ""
|
||||
try:
|
||||
aria_snapshot = await page.locator("body").aria_snapshot(
|
||||
timeout=self.config.observation_timeout_ms,
|
||||
)
|
||||
except Exception:
|
||||
aria_snapshot = ""
|
||||
try:
|
||||
screenshot_path = self._screenshots_dir() / f"step_{self._step_index:04d}.png"
|
||||
await page.screenshot(path=str(screenshot_path), full_page=False)
|
||||
except Exception:
|
||||
screenshot_path = None
|
||||
|
||||
return {
|
||||
"success": success,
|
||||
"exception": exception_text,
|
||||
"url": url or self.config.start_url or "",
|
||||
"title": title,
|
||||
"screenshot_path": str(screenshot_path) if screenshot_path is not None else "",
|
||||
"aria_snapshot": aria_snapshot,
|
||||
"python_code": self._step_python_code,
|
||||
"python_output": self._step_python_output,
|
||||
"console_output": "\n".join(self._step_console[-20:]),
|
||||
"recent_console": "\n".join(self._console_history[-50:]),
|
||||
}
|
||||
|
||||
def get_template_vars(self, **kwargs) -> dict[str, Any]:
|
||||
return {
|
||||
"start_url": self.config.start_url or "",
|
||||
"output_dir": str(self.config.output_dir.resolve()),
|
||||
"browser_mode": self.config.browser_mode,
|
||||
"user_data_dir": str(self.config.user_data_dir),
|
||||
**kwargs,
|
||||
}
|
||||
|
||||
def serialize(self) -> dict[str, Any]:
|
||||
return {
|
||||
"environment": {
|
||||
"config": self.config.model_dump(mode="json"),
|
||||
"environment_type": f"{self.__class__.__module__}.{self.__class__.__name__}",
|
||||
}
|
||||
}
|
||||
|
||||
def close(self) -> None:
|
||||
if self.config.prompt_before_close:
|
||||
input("Press Enter to close the browser...")
|
||||
if self.config.keep_open_on_exit:
|
||||
return
|
||||
try:
|
||||
self._run(self._close_async())
|
||||
finally:
|
||||
if self._loop is not None and not self._loop.is_closed():
|
||||
self._loop.close()
|
||||
self._loop = None
|
||||
|
||||
async def _close_async(self) -> None:
|
||||
|
||||
context = self._context
|
||||
browser = self._browser
|
||||
page = self._local_cdp_page
|
||||
playwright = self._playwright
|
||||
connected_over_cdp = self._connected_over_cdp
|
||||
local_cdp_process = self._local_cdp_process
|
||||
self._page = None
|
||||
self._local_cdp_page = None
|
||||
self._context = None
|
||||
self._browser = None
|
||||
self._playwright = None
|
||||
self._connected_over_cdp = False
|
||||
self._local_cdp_process = None
|
||||
|
||||
try:
|
||||
if connected_over_cdp:
|
||||
if page is not None and self.config.local_cdp_close_page_on_exit:
|
||||
try:
|
||||
await page.close()
|
||||
except Exception:
|
||||
pass
|
||||
if (
|
||||
local_cdp_process is not None
|
||||
and self.config.local_cdp_close_started_browser_on_exit
|
||||
):
|
||||
local_cdp_process.terminate()
|
||||
elif context is not None:
|
||||
await context.close()
|
||||
elif browser is not None:
|
||||
await browser.close()
|
||||
finally:
|
||||
if playwright is not None:
|
||||
await playwright.stop()
|
||||
@@ -0,0 +1,296 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import os
|
||||
import re
|
||||
import subprocess
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
from pydantic import BaseModel, Field
|
||||
|
||||
_EXPORT_RE = re.compile(r"^export\s+([A-Za-z_][A-Za-z0-9_]*)=(.*)$")
|
||||
_IMAGE_SUFFIXES = {".png", ".jpg", ".jpeg", ".webp"}
|
||||
|
||||
|
||||
class LocalWorkspaceEnvironmentConfig(BaseModel):
|
||||
"""Shell-based workspace environment.
|
||||
|
||||
The agent drives a real browser through bash commands it generates inside this
|
||||
workspace. Two browser modes are exposed to those generated scripts via
|
||||
environment variables:
|
||||
|
||||
* ``browser_mode = "browserbase"`` (default): the agent's scripts should
|
||||
create a Browserbase cloud session. ``BROWSERBASE_API_KEY`` and
|
||||
``BROWSERBASE_PROJECT_ID`` are forwarded if present.
|
||||
* ``browser_mode = "local"``: the agent's scripts should launch a local
|
||||
Playwright browser (``playwright.chromium.launch(...)``).
|
||||
|
||||
The selected mode is forwarded to the subprocess via ``BROWSER_MODE`` so the
|
||||
generated scripts can branch on it.
|
||||
"""
|
||||
|
||||
start_url: str | None = None
|
||||
output_dir: Path = Path("outputs/sandbox/default")
|
||||
command_timeout_seconds: int = 180
|
||||
shell: str = "/bin/bash"
|
||||
env: dict[str, str] = Field(default_factory=dict)
|
||||
credentials_file: Path | None = None
|
||||
browser_mode: str = "browserbase" # "browserbase" or "local"
|
||||
task_metadata_filename: str = "task.json"
|
||||
final_script_name: str = "final_script.py"
|
||||
output_truncation_chars: int = 12000
|
||||
final_script_preview_chars: int = 4000
|
||||
recent_files_limit: int = 40
|
||||
|
||||
|
||||
class LocalWorkspaceEnvironment:
|
||||
def __init__(self, *, config_class: type = LocalWorkspaceEnvironmentConfig, **kwargs):
|
||||
self.config = config_class(**kwargs)
|
||||
self.config.output_dir = self.config.output_dir.expanduser()
|
||||
self._step_index = 0
|
||||
self._prepared_task: dict[str, Any] = {}
|
||||
self._credential_env = self._load_credential_env(self.config.credentials_file)
|
||||
|
||||
def _load_credential_env(self, path: Path | None) -> dict[str, str]:
|
||||
if path is None:
|
||||
return {}
|
||||
resolved = Path(path).expanduser()
|
||||
if not resolved.exists():
|
||||
return {}
|
||||
|
||||
parsed: dict[str, str] = {}
|
||||
for raw_line in resolved.read_text(encoding="utf-8").splitlines():
|
||||
line = raw_line.strip()
|
||||
if not line or line.startswith("#"):
|
||||
continue
|
||||
match = _EXPORT_RE.match(line)
|
||||
if match is None:
|
||||
continue
|
||||
key, raw_value = match.groups()
|
||||
value = raw_value.strip()
|
||||
if len(value) >= 2 and value[0] == value[-1] and value[0] in {"'", '"'}:
|
||||
value = value[1:-1]
|
||||
parsed[key] = value
|
||||
return parsed
|
||||
|
||||
def _workspace_dir(self) -> Path:
|
||||
return self.config.output_dir.resolve()
|
||||
|
||||
def _resolve_cwd(self, cwd: str = "") -> Path:
|
||||
workspace_dir = self._workspace_dir()
|
||||
if not cwd:
|
||||
return workspace_dir
|
||||
candidate = Path(cwd)
|
||||
if not candidate.is_absolute():
|
||||
candidate = workspace_dir / candidate
|
||||
resolved = candidate.resolve()
|
||||
try:
|
||||
resolved.relative_to(workspace_dir)
|
||||
except ValueError as exc:
|
||||
raise ValueError(f"Command cwd must stay inside workspace: {resolved}") from exc
|
||||
return resolved
|
||||
|
||||
def _task_metadata_path(self) -> Path:
|
||||
return self._workspace_dir() / self.config.task_metadata_filename
|
||||
|
||||
def _final_script_path(self) -> Path:
|
||||
return self._workspace_dir() / self.config.final_script_name
|
||||
|
||||
def _steps_dir(self) -> Path:
|
||||
return self._workspace_dir() / "steps"
|
||||
|
||||
def _logs_dir(self) -> Path:
|
||||
return self._workspace_dir() / "logs"
|
||||
|
||||
def _screenshots_dir(self) -> Path:
|
||||
return self._workspace_dir() / "screenshots"
|
||||
|
||||
def _history_path(self) -> Path:
|
||||
return self._workspace_dir() / "command_history.sh"
|
||||
|
||||
def _truncate(self, text: str, limit: int) -> str:
|
||||
if len(text) <= limit:
|
||||
return text
|
||||
omitted = len(text) - limit
|
||||
return f"{text[:limit]}\n\n... [{omitted} characters omitted]"
|
||||
|
||||
def _recent_workspace_files(self) -> list[str]:
|
||||
workspace_dir = self._workspace_dir()
|
||||
files: list[Path] = [path for path in workspace_dir.rglob("*") if path.is_file()]
|
||||
files.sort(key=lambda path: path.stat().st_mtime, reverse=True)
|
||||
recent = files[: self.config.recent_files_limit]
|
||||
return [str(path.relative_to(workspace_dir)) for path in recent]
|
||||
|
||||
def _recent_screenshots(self) -> list[Path]:
|
||||
screenshots_dir = self._screenshots_dir()
|
||||
if not screenshots_dir.exists():
|
||||
return []
|
||||
files = [path for path in screenshots_dir.rglob("*") if path.is_file() and path.suffix.lower() in _IMAGE_SUFFIXES]
|
||||
files.sort(key=lambda path: path.stat().st_mtime, reverse=True)
|
||||
return files
|
||||
|
||||
def _persist_step_command(self, command: str) -> Path:
|
||||
self._steps_dir().mkdir(parents=True, exist_ok=True)
|
||||
step_path = self._steps_dir() / f"step_{self._step_index:04d}.sh"
|
||||
step_path.write_text(command.rstrip() + "\n", encoding="utf-8")
|
||||
|
||||
history_path = self._history_path()
|
||||
with history_path.open("a", encoding="utf-8") as handle:
|
||||
handle.write(f"# Step {self._step_index}\n")
|
||||
handle.write(command.rstrip())
|
||||
handle.write("\n\n")
|
||||
return step_path
|
||||
|
||||
def _write_step_log(self, output: str) -> Path | None:
|
||||
if not output:
|
||||
return None
|
||||
self._logs_dir().mkdir(parents=True, exist_ok=True)
|
||||
log_path = self._logs_dir() / f"step_{self._step_index:04d}.log"
|
||||
log_path.write_text(output, encoding="utf-8")
|
||||
return log_path
|
||||
|
||||
def prepare(self, **kwargs) -> None:
|
||||
self._prepared_task = dict(kwargs)
|
||||
self._step_index = 0
|
||||
start_url = kwargs.get("start_url") or self.config.start_url
|
||||
if start_url:
|
||||
self.config.start_url = str(start_url)
|
||||
workspace_dir = self._workspace_dir()
|
||||
workspace_dir.mkdir(parents=True, exist_ok=True)
|
||||
self._steps_dir().mkdir(parents=True, exist_ok=True)
|
||||
self._logs_dir().mkdir(parents=True, exist_ok=True)
|
||||
self._screenshots_dir().mkdir(parents=True, exist_ok=True)
|
||||
(workspace_dir / ".tmp").mkdir(parents=True, exist_ok=True)
|
||||
self._task_metadata_path().write_text(json.dumps(kwargs, indent=2), encoding="utf-8")
|
||||
|
||||
def _browser_env(self) -> dict[str, str]:
|
||||
"""Forward Browserbase / browser-mode hints to the subprocess."""
|
||||
env: dict[str, str] = {"BROWSER_MODE": str(self.config.browser_mode or "browserbase")}
|
||||
for var in ("BROWSERBASE_API_KEY", "BROWSERBASE_PROJECT_ID"):
|
||||
value = self._credential_env.get(var) or os.environ.get(var)
|
||||
if value:
|
||||
env[var] = value
|
||||
return env
|
||||
|
||||
def execute(self, action: dict[str, Any], cwd: str = "") -> dict[str, Any]:
|
||||
self._step_index += 1
|
||||
command = str(
|
||||
action.get("command") or action.get("bash_command") or action.get("python_code") or ""
|
||||
).strip()
|
||||
self._persist_step_command(command)
|
||||
resolved_cwd = self._resolve_cwd(cwd)
|
||||
|
||||
command_env = os.environ | self._credential_env | self._browser_env() | self.config.env | {
|
||||
"WORKSPACE_DIR": str(self._workspace_dir()),
|
||||
"OM2W_TASK_JSON": str(self._task_metadata_path()),
|
||||
"FINAL_SCRIPT_PATH": str(self._final_script_path()),
|
||||
"TMPDIR": str(self._workspace_dir() / ".tmp"),
|
||||
}
|
||||
|
||||
try:
|
||||
result = subprocess.run(
|
||||
command,
|
||||
shell=True,
|
||||
executable=self.config.shell,
|
||||
text=True,
|
||||
cwd=resolved_cwd,
|
||||
env=command_env,
|
||||
timeout=self.config.command_timeout_seconds,
|
||||
encoding="utf-8",
|
||||
errors="replace",
|
||||
stdout=subprocess.PIPE,
|
||||
stderr=subprocess.STDOUT,
|
||||
)
|
||||
output = result.stdout
|
||||
returncode = result.returncode
|
||||
exception_info = ""
|
||||
except Exception as exc:
|
||||
raw_output = getattr(exc, "output", None)
|
||||
output = raw_output.decode("utf-8", errors="replace") if isinstance(raw_output, bytes) else (raw_output or "")
|
||||
returncode = -1
|
||||
exception_info = f"An error occurred while executing the command: {exc}"
|
||||
|
||||
log_path = self._write_step_log(output)
|
||||
observation = self._capture_observation(
|
||||
command=command,
|
||||
cwd=resolved_cwd,
|
||||
output=output,
|
||||
returncode=returncode,
|
||||
exception_info=exception_info,
|
||||
log_path=log_path,
|
||||
)
|
||||
return {
|
||||
"output": output,
|
||||
"returncode": returncode,
|
||||
"exception_info": exception_info,
|
||||
"observation": observation,
|
||||
}
|
||||
|
||||
def _capture_observation(
|
||||
self,
|
||||
*,
|
||||
command: str,
|
||||
cwd: Path,
|
||||
output: str,
|
||||
returncode: int,
|
||||
exception_info: str,
|
||||
log_path: Path | None,
|
||||
) -> dict[str, Any]:
|
||||
final_script_path = self._final_script_path()
|
||||
recent_screenshot_paths = self._recent_screenshots()
|
||||
latest_screenshot = recent_screenshot_paths[0] if recent_screenshot_paths else None
|
||||
final_script_preview = ""
|
||||
if final_script_path.exists():
|
||||
final_script_preview = self._truncate(
|
||||
final_script_path.read_text(encoding="utf-8", errors="replace"),
|
||||
self.config.final_script_preview_chars,
|
||||
)
|
||||
|
||||
workspace_dir = self._workspace_dir()
|
||||
recent_screenshots = [str(path.relative_to(workspace_dir)) for path in recent_screenshot_paths[:10]]
|
||||
return {
|
||||
"success": returncode == 0 and not exception_info,
|
||||
"exception": exception_info,
|
||||
"command": command,
|
||||
"returncode": returncode,
|
||||
"workspace_dir": str(workspace_dir),
|
||||
"cwd": str(cwd),
|
||||
"url": self.config.start_url or "",
|
||||
"title": "",
|
||||
"aria_snapshot": "",
|
||||
"console_output": "",
|
||||
"recent_console": "",
|
||||
"command_output": self._truncate(output, self.config.output_truncation_chars),
|
||||
"log_path": str(log_path) if log_path is not None else "",
|
||||
"task_metadata_path": str(self._task_metadata_path()),
|
||||
"final_script_path": str(final_script_path) if final_script_path.exists() else "",
|
||||
"final_script_exists": final_script_path.exists(),
|
||||
"final_script_preview": final_script_preview,
|
||||
"screenshot_path": str(latest_screenshot) if latest_screenshot is not None else "",
|
||||
"recent_screenshots": recent_screenshots,
|
||||
"workspace_files": self._recent_workspace_files(),
|
||||
}
|
||||
|
||||
def get_template_vars(self, **kwargs) -> dict[str, Any]:
|
||||
return {
|
||||
"start_url": self.config.start_url or "",
|
||||
"output_dir": str(self._workspace_dir()),
|
||||
"workspace_dir": str(self._workspace_dir()),
|
||||
"task_metadata_path": str(self._task_metadata_path()),
|
||||
"final_script_path": str(self._final_script_path()),
|
||||
"browser_mode": self.config.browser_mode,
|
||||
**kwargs,
|
||||
}
|
||||
|
||||
def serialize(self) -> dict:
|
||||
return {
|
||||
"environment": {
|
||||
"config": self.config.model_dump(mode="json"),
|
||||
"environment_type": f"{self.__class__.__module__}.{self.__class__.__name__}",
|
||||
"workspace_dir": str(self._workspace_dir()),
|
||||
}
|
||||
}
|
||||
|
||||
def close(self) -> None:
|
||||
return None
|
||||
@@ -0,0 +1,19 @@
|
||||
from __future__ import annotations
|
||||
|
||||
|
||||
class InterruptAgentFlow(Exception):
|
||||
def __init__(self, *messages: dict):
|
||||
super().__init__()
|
||||
self.messages = list(messages)
|
||||
|
||||
|
||||
class LimitsExceeded(InterruptAgentFlow):
|
||||
pass
|
||||
|
||||
|
||||
class Submitted(InterruptAgentFlow):
|
||||
pass
|
||||
|
||||
|
||||
class FormatError(InterruptAgentFlow):
|
||||
pass
|
||||
@@ -0,0 +1,25 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import copy
|
||||
import importlib
|
||||
|
||||
from webwright import Model
|
||||
|
||||
_MODEL_MAPPING = {
|
||||
"openai": "webwright.models.openai_model.OpenAIModel",
|
||||
"anthropic": "webwright.models.anthropic_model.AnthropicModel",
|
||||
"openrouter": "webwright.models.openrouter_model.OpenRouterModel",
|
||||
}
|
||||
|
||||
|
||||
def get_model_class(spec: str) -> type[Model]:
|
||||
full_path = _MODEL_MAPPING.get(spec, spec)
|
||||
module_name, class_name = full_path.rsplit(".", 1)
|
||||
module = importlib.import_module(module_name)
|
||||
return getattr(module, class_name)
|
||||
|
||||
|
||||
def get_model(config: dict, *, default_type: str = "openai") -> Model:
|
||||
copied = copy.deepcopy(config)
|
||||
model_class = copied.pop("model_class", default_type)
|
||||
return get_model_class(model_class)(**copied)
|
||||
@@ -0,0 +1,192 @@
|
||||
"""Anthropic (Claude) Messages API model backend."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
import random
|
||||
from typing import Any
|
||||
|
||||
from webwright.models.base import (
|
||||
BaseModel,
|
||||
BaseModelConfig,
|
||||
OptStr,
|
||||
_safe_int,
|
||||
)
|
||||
|
||||
# Anthropic-specific retry limits (independent of OpenAI defaults). The Claude
|
||||
# Opus org-level ITPM cap can saturate for minutes at high concurrency, so we
|
||||
# allow many more retries and longer backoffs.
|
||||
MAX_RATE_LIMIT_RETRIES = 50
|
||||
MAX_TRANSIENT_GATEWAY_RETRIES = 20
|
||||
RATE_LIMIT_BACKOFF_MIN_SECONDS = 30.0
|
||||
RATE_LIMIT_BACKOFF_MAX_SECONDS = 60.0
|
||||
TRANSIENT_BACKOFF_BASE_SECONDS = 1.5
|
||||
TRANSIENT_BACKOFF_CAP_SECONDS = 60.0
|
||||
|
||||
|
||||
def _retry_after_seconds(exc: BaseException) -> float | None:
|
||||
response = getattr(exc, "response", None)
|
||||
if response is None:
|
||||
return None
|
||||
header = response.headers.get("retry-after") if getattr(response, "headers", None) else None
|
||||
if not header:
|
||||
return None
|
||||
try:
|
||||
return max(0.0, float(header))
|
||||
except (TypeError, ValueError):
|
||||
return None
|
||||
|
||||
|
||||
def _image_source_from_url(image_url: str) -> dict[str, Any]:
|
||||
if image_url.startswith("data:"):
|
||||
header, _, encoded = image_url.partition(",")
|
||||
media_type = header.split(";")[0].removeprefix("data:") or "image/png"
|
||||
return {"type": "base64", "media_type": media_type, "data": encoded}
|
||||
return {"type": "url", "url": image_url}
|
||||
|
||||
|
||||
def _serialize_anthropic_content_part(part: dict[str, Any]) -> dict[str, Any]:
|
||||
if part.get("type") == "input_image":
|
||||
return {"type": "image", "source": _image_source_from_url(part.get("image_url", ""))}
|
||||
return {"type": "text", "text": part.get("text", "")}
|
||||
|
||||
|
||||
def _serialize_anthropic_messages(
|
||||
messages: list[dict[str, Any]],
|
||||
) -> tuple[str | None, list[dict[str, Any]]]:
|
||||
system_chunks: list[str] = []
|
||||
serialized: list[dict[str, Any]] = []
|
||||
for message in messages:
|
||||
role = message["role"]
|
||||
if role == "exit":
|
||||
continue
|
||||
content = message.get("content", "")
|
||||
if role == "system":
|
||||
if isinstance(content, str):
|
||||
if content:
|
||||
system_chunks.append(content)
|
||||
else:
|
||||
for part in content:
|
||||
if isinstance(part, dict) and part.get("type") != "input_image":
|
||||
text = part.get("text", "")
|
||||
if text:
|
||||
system_chunks.append(text)
|
||||
continue
|
||||
|
||||
if isinstance(content, str):
|
||||
serialized.append({"role": role, "content": content})
|
||||
continue
|
||||
parts = [_serialize_anthropic_content_part(p) for p in content if isinstance(p, dict)]
|
||||
if parts and all(p.get("type") == "text" for p in parts):
|
||||
serialized.append({"role": role, "content": "\n".join(p["text"] for p in parts)})
|
||||
else:
|
||||
serialized.append({"role": role, "content": parts})
|
||||
|
||||
system_prompt = "\n\n".join(system_chunks) if system_chunks else None
|
||||
return system_prompt, serialized
|
||||
|
||||
|
||||
def _extract_anthropic_text(payload: dict[str, Any]) -> str:
|
||||
texts: list[str] = []
|
||||
for block in payload.get("content") or []:
|
||||
if isinstance(block, dict) and block.get("type") == "text":
|
||||
text = block.get("text", "")
|
||||
if text:
|
||||
texts.append(text)
|
||||
return "\n".join(texts)
|
||||
|
||||
|
||||
def _usage_from_anthropic_payload(payload: dict[str, Any]) -> dict[str, int]:
|
||||
usage = payload.get("usage") or {}
|
||||
input_tokens = _safe_int(usage.get("input_tokens"))
|
||||
output_tokens = _safe_int(usage.get("output_tokens"))
|
||||
cached_input_tokens = _safe_int(usage.get("cache_read_input_tokens"))
|
||||
return {
|
||||
"input_tokens": input_tokens,
|
||||
"output_tokens": output_tokens,
|
||||
"total_tokens": input_tokens + output_tokens,
|
||||
"cached_input_tokens": cached_input_tokens,
|
||||
"reasoning_output_tokens": 0,
|
||||
}
|
||||
|
||||
|
||||
def _metrics_input_from_anthropic(
|
||||
system_prompt: str | None, anthropic_messages: list[dict[str, Any]]
|
||||
) -> list[dict[str, Any]]:
|
||||
items: list[dict[str, Any]] = []
|
||||
if system_prompt:
|
||||
items.append({"content": [{"type": "input_text", "text": system_prompt}]})
|
||||
for msg in anthropic_messages:
|
||||
content = msg.get("content", "")
|
||||
if isinstance(content, str):
|
||||
items.append({"content": [{"type": "input_text", "text": content}]})
|
||||
continue
|
||||
parts: list[dict[str, Any]] = []
|
||||
for part in content:
|
||||
if not isinstance(part, dict):
|
||||
continue
|
||||
if part.get("type") == "text":
|
||||
parts.append({"type": "input_text", "text": part.get("text", "")})
|
||||
elif part.get("type") == "image":
|
||||
parts.append({"type": "input_image"})
|
||||
items.append({"content": parts})
|
||||
return items
|
||||
|
||||
|
||||
class AnthropicModelConfig(BaseModelConfig):
|
||||
model_name: OptStr = "claude-opus-4-7"
|
||||
anthropic_api_key: OptStr = ""
|
||||
anthropic_endpoint: OptStr = "https://api.anthropic.com/v1/messages"
|
||||
anthropic_version: OptStr = "2023-06-01"
|
||||
max_output_tokens: int = 8000
|
||||
|
||||
|
||||
class AnthropicModel(BaseModel):
|
||||
_API_KEY_FIELD = "anthropic_api_key"
|
||||
_ENV_VAR = "ANTHROPIC_API_KEY"
|
||||
_LOG_SOURCE = "anthropic"
|
||||
_MAX_RATE_LIMIT_RETRIES = MAX_RATE_LIMIT_RETRIES
|
||||
_MAX_TRANSIENT_RETRIES = MAX_TRANSIENT_GATEWAY_RETRIES
|
||||
_DEFAULT_CONFIG_CLASS = AnthropicModelConfig
|
||||
|
||||
def _request_headers(self) -> dict[str, str]:
|
||||
return {
|
||||
"Content-Type": "application/json",
|
||||
"x-api-key": self.config.anthropic_api_key,
|
||||
"anthropic-version": self.config.anthropic_version,
|
||||
}
|
||||
|
||||
def _post_url(self) -> str:
|
||||
return self.config.anthropic_endpoint
|
||||
|
||||
async def _rate_limit_backoff(self, attempt: int, exc: BaseException) -> None:
|
||||
delay = random.uniform(RATE_LIMIT_BACKOFF_MIN_SECONDS, RATE_LIMIT_BACKOFF_MAX_SECONDS)
|
||||
retry_after = _retry_after_seconds(exc)
|
||||
if retry_after is not None and retry_after > delay:
|
||||
delay = min(retry_after, RATE_LIMIT_BACKOFF_MAX_SECONDS * 2)
|
||||
await asyncio.sleep(delay)
|
||||
|
||||
async def _transient_backoff(self, attempt: int, exc: BaseException) -> None:
|
||||
await asyncio.sleep(
|
||||
min(TRANSIENT_BACKOFF_BASE_SECONDS * (2 ** attempt), TRANSIENT_BACKOFF_CAP_SECONDS)
|
||||
)
|
||||
|
||||
def _build_payload(self, messages: list[dict[str, Any]]) -> dict[str, Any]:
|
||||
system_prompt, anth_messages = _serialize_anthropic_messages(messages)
|
||||
payload: dict[str, Any] = {
|
||||
"model": self.config.model_name,
|
||||
"messages": anth_messages,
|
||||
"max_tokens": self.config.max_output_tokens,
|
||||
}
|
||||
if system_prompt:
|
||||
payload["system"] = system_prompt
|
||||
return payload
|
||||
|
||||
def _request_metrics_input(self, payload: dict[str, Any]) -> list[dict[str, Any]]:
|
||||
return _metrics_input_from_anthropic(payload.get("system"), payload.get("messages") or [])
|
||||
|
||||
def _extract_text(self, payload: dict[str, Any]) -> str:
|
||||
return _extract_anthropic_text(payload)
|
||||
|
||||
def _usage_metrics_from_payload(self, payload: dict[str, Any]) -> dict[str, int]:
|
||||
return _usage_from_anthropic_payload(payload)
|
||||
@@ -0,0 +1,587 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
import base64
|
||||
import json
|
||||
import mimetypes
|
||||
import os
|
||||
import subprocess
|
||||
from pathlib import Path
|
||||
from typing import Annotated, Any
|
||||
|
||||
import httpx
|
||||
from jinja2 import StrictUndefined, Template
|
||||
from pydantic import BaseModel as PydanticBaseModel, BeforeValidator, field_validator
|
||||
|
||||
from webwright.exceptions import FormatError
|
||||
from webwright.utils.logging import append_runtime_log
|
||||
from webwright.utils.runtime import run_async
|
||||
|
||||
|
||||
def _none_to_str(value: Any) -> str:
|
||||
return "" if value is None else str(value)
|
||||
|
||||
|
||||
# String field that coerces None -> "" and any value -> str via pydantic.
|
||||
OptStr = Annotated[str, BeforeValidator(_none_to_str)]
|
||||
|
||||
MAX_JSON_PARSE_RETRIES = 3
|
||||
DEFAULT_OBSERVATION_TEMPLATE = """Observation:
|
||||
Status: {{ 'ok' if observation.success else 'error' }}
|
||||
URL: {{ observation.url }}
|
||||
Title: {{ observation.title }}
|
||||
{% if observation.exception %}Exception:
|
||||
{{ observation.exception }}
|
||||
{% endif %}{% if observation.console_output %}Console output:
|
||||
{{ observation.console_output }}
|
||||
{% endif %}{% if observation.aria_snapshot %}ARIA snapshot:
|
||||
{{ observation.aria_snapshot }}
|
||||
{% endif %}{% if observation.screenshot_path %}Screenshot path: {{ observation.screenshot_path }}
|
||||
{% endif %}"""
|
||||
DEFAULT_FORMAT_ERROR_TEMPLATE = """Format error:
|
||||
|
||||
{{ error }}
|
||||
|
||||
Please respond with strict JSON using exactly these fields:
|
||||
- thought: short reasoning about the next step
|
||||
- bash_command: exactly one shell command for local-workspace tasks
|
||||
- python_code: exactly one async Python browser step for local-browser tasks
|
||||
- done: boolean indicating whether the task is complete
|
||||
- final_response: final natural-language answer when done, otherwise empty
|
||||
"""
|
||||
|
||||
ACTION_FIELDS = {"bash_command", "python_code"}
|
||||
|
||||
|
||||
def _is_rate_limit_error(exc: BaseException | None) -> bool:
|
||||
current: BaseException | None = exc
|
||||
while current is not None:
|
||||
status_code = getattr(current, "status_code", None)
|
||||
if status_code == 429:
|
||||
return True
|
||||
response = getattr(current, "response", None)
|
||||
if getattr(response, "status_code", None) == 429:
|
||||
return True
|
||||
text = str(current).lower()
|
||||
if "rate limit" in text or "ratelimit" in text or "too many requests" in text:
|
||||
return True
|
||||
current = current.__cause__ if isinstance(current.__cause__, BaseException) else None
|
||||
return False
|
||||
|
||||
|
||||
def _is_transient_http_error(exc: BaseException | None) -> bool:
|
||||
"""True for retryable transient HTTP failures (timeouts, 5xx, conn resets, ...).
|
||||
|
||||
Applies to any HTTP backend, not just gateway/proxy setups.
|
||||
"""
|
||||
current: BaseException | None = exc
|
||||
while current is not None:
|
||||
if isinstance(current, (httpx.TimeoutException, httpx.NetworkError, httpx.RemoteProtocolError)):
|
||||
return True
|
||||
status_code = getattr(current, "status_code", None)
|
||||
if isinstance(status_code, int) and status_code in {408, 409, 425, 500, 502, 503, 504}:
|
||||
return True
|
||||
response = getattr(current, "response", None)
|
||||
response_status = getattr(response, "status_code", None)
|
||||
if isinstance(response_status, int) and response_status in {408, 409, 425, 500, 502, 503, 504}:
|
||||
return True
|
||||
text = str(current).lower()
|
||||
if any(
|
||||
needle in text
|
||||
for needle in (
|
||||
"bad gateway",
|
||||
"gateway timeout",
|
||||
"server disconnected",
|
||||
"temporary failure",
|
||||
"temporarily unavailable",
|
||||
"connection reset",
|
||||
"connection aborted",
|
||||
"timed out",
|
||||
)
|
||||
):
|
||||
return True
|
||||
current = current.__cause__ if isinstance(current.__cause__, BaseException) else None
|
||||
return False
|
||||
|
||||
|
||||
def parse_json_output(raw: str, *, action_field: str = "bash_command") -> dict[str, Any]:
|
||||
try:
|
||||
parsed = json.loads(raw)
|
||||
except json.JSONDecodeError as exc:
|
||||
raise ValueError(f"Unable to parse JSON output: {exc}") from exc
|
||||
if not isinstance(parsed, dict):
|
||||
raise ValueError("Model output was JSON but not a JSON object.")
|
||||
# Strict-schema responses cannot have done=true with a non-empty action;
|
||||
# tolerate it from non-strict callers by demoting `done`.
|
||||
action_text = str(parsed.get(action_field, "") or "").strip()
|
||||
if action_text and bool(parsed.get("done", False)):
|
||||
parsed = dict(parsed)
|
||||
parsed["done"] = False
|
||||
return parsed
|
||||
|
||||
|
||||
def _validate_bash_command(command: str) -> None:
|
||||
result = subprocess.run(
|
||||
["/bin/bash", "-n"],
|
||||
input=command,
|
||||
text=True,
|
||||
capture_output=True,
|
||||
encoding="utf-8",
|
||||
errors="replace",
|
||||
check=False,
|
||||
)
|
||||
if result.returncode == 0:
|
||||
return
|
||||
error = (result.stderr or result.stdout or "bash syntax check failed").strip()
|
||||
raise ValueError(f"Invalid bash_command syntax: {error}")
|
||||
|
||||
|
||||
def text_part(text: str) -> dict[str, Any]:
|
||||
return {"type": "input_text", "text": text}
|
||||
|
||||
|
||||
def image_part_from_path(path: Path) -> dict[str, Any]:
|
||||
mime_type, _ = mimetypes.guess_type(str(path))
|
||||
encoded = base64.b64encode(path.read_bytes()).decode("ascii")
|
||||
return {
|
||||
"type": "input_image",
|
||||
"image_url": f"data:{mime_type or 'image/png'};base64,{encoded}",
|
||||
"detail": "high",
|
||||
}
|
||||
|
||||
|
||||
def _safe_int(value: Any) -> int:
|
||||
try:
|
||||
return int(value)
|
||||
except (TypeError, ValueError):
|
||||
return 0
|
||||
|
||||
|
||||
def _request_metrics_from_serialized_input(serialized_input: list[dict[str, Any]]) -> dict[str, int]:
|
||||
message_count = len(serialized_input)
|
||||
text_part_count = 0
|
||||
image_part_count = 0
|
||||
text_chars = 0
|
||||
|
||||
for item in serialized_input:
|
||||
for content in item.get("content") or []:
|
||||
if not isinstance(content, dict):
|
||||
continue
|
||||
part_type = content.get("type")
|
||||
if part_type in {"input_text", "output_text"}:
|
||||
text_part_count += 1
|
||||
text_chars += len(str(content.get("text", "") or ""))
|
||||
elif part_type == "input_image":
|
||||
image_part_count += 1
|
||||
|
||||
serialized_chars = len(json.dumps(serialized_input, ensure_ascii=False))
|
||||
|
||||
return {
|
||||
"message_count": message_count,
|
||||
"text_part_count": text_part_count,
|
||||
"image_part_count": image_part_count,
|
||||
"text_chars": text_chars,
|
||||
"serialized_chars": serialized_chars,
|
||||
}
|
||||
|
||||
|
||||
_REQUEST_METRIC_KEYS = (
|
||||
"message_count",
|
||||
"text_part_count",
|
||||
"image_part_count",
|
||||
"text_chars",
|
||||
"serialized_chars",
|
||||
)
|
||||
_USAGE_METRIC_KEYS = (
|
||||
"input_tokens",
|
||||
"output_tokens",
|
||||
"total_tokens",
|
||||
"cached_input_tokens",
|
||||
"reasoning_output_tokens",
|
||||
)
|
||||
|
||||
|
||||
class BaseModelConfig(PydanticBaseModel):
|
||||
"""Fields common to every model backend (OpenAI, Anthropic, ...)."""
|
||||
|
||||
model_name: OptStr = ""
|
||||
max_output_tokens: int = 4000
|
||||
request_timeout_seconds: int = 120
|
||||
error_log_path: Path | None = None
|
||||
observation_template: OptStr = DEFAULT_OBSERVATION_TEMPLATE
|
||||
format_error_template: OptStr = DEFAULT_FORMAT_ERROR_TEMPLATE
|
||||
attach_observation_screenshot: bool = True
|
||||
action_field: str = "bash_command"
|
||||
|
||||
@field_validator("action_field")
|
||||
@classmethod
|
||||
def validate_action_field(cls, value: str) -> str:
|
||||
normalized = value.strip()
|
||||
if normalized not in ACTION_FIELDS:
|
||||
raise ValueError(f"action_field must be one of: {', '.join(sorted(ACTION_FIELDS))}")
|
||||
return normalized
|
||||
|
||||
|
||||
class BaseModel:
|
||||
"""Provider-agnostic model backend.
|
||||
|
||||
Subclasses must override:
|
||||
- class constants ``_API_KEY_FIELD``, ``_ENV_VAR``, ``_LOG_SOURCE``,
|
||||
``_DEFAULT_CONFIG_CLASS`` (and optionally ``_MAX_RATE_LIMIT_RETRIES``,
|
||||
``_MAX_TRANSIENT_RETRIES``)
|
||||
- ``_request_headers``, ``_post_url``
|
||||
- ``_build_payload``, ``_request_metrics_input``
|
||||
- ``_extract_text``, ``_usage_metrics_from_payload``
|
||||
|
||||
Optionally override ``_rate_limit_backoff`` / ``_transient_backoff`` for
|
||||
custom retry timing.
|
||||
"""
|
||||
|
||||
_API_KEY_FIELD: str = ""
|
||||
_ENV_VAR: str = ""
|
||||
_LOG_SOURCE: str = ""
|
||||
_MAX_RATE_LIMIT_RETRIES: int = 5
|
||||
_MAX_TRANSIENT_RETRIES: int = 5
|
||||
_DEFAULT_CONFIG_CLASS: type = BaseModelConfig
|
||||
|
||||
def __init__(self, *, config_class: type | None = None, **kwargs):
|
||||
self.config = (config_class or self._DEFAULT_CONFIG_CLASS)(**kwargs)
|
||||
self._last_request_metrics: dict[str, int] = {k: 0 for k in _REQUEST_METRIC_KEYS}
|
||||
self._last_usage_metrics: dict[str, int] = {k: 0 for k in _USAGE_METRIC_KEYS}
|
||||
self._cumulative_request_metrics: dict[str, int] = dict(self._last_request_metrics)
|
||||
self._cumulative_usage_metrics: dict[str, int] = dict(self._last_usage_metrics)
|
||||
|
||||
if self._API_KEY_FIELD:
|
||||
if not getattr(self.config, self._API_KEY_FIELD, ""):
|
||||
setattr(self.config, self._API_KEY_FIELD, os.environ.get(self._ENV_VAR, ""))
|
||||
if not getattr(self.config, self._API_KEY_FIELD, ""):
|
||||
raise RuntimeError(f"Missing {self._ENV_VAR}.")
|
||||
|
||||
# ---- subclass extension points ------------------------------------------------
|
||||
|
||||
def _request_headers(self) -> dict[str, str]:
|
||||
raise NotImplementedError
|
||||
|
||||
def _post_url(self) -> str:
|
||||
raise NotImplementedError
|
||||
|
||||
def _build_payload(self, messages: list[dict[str, Any]]) -> dict[str, Any]:
|
||||
raise NotImplementedError
|
||||
|
||||
def _build_text_payload(self, messages: list[dict[str, Any]]) -> dict[str, Any]:
|
||||
return self._build_payload(messages)
|
||||
|
||||
def _request_metrics_input(self, payload: dict[str, Any]) -> list[dict[str, Any]]:
|
||||
raise NotImplementedError
|
||||
|
||||
def _extract_text(self, payload: dict[str, Any]) -> str:
|
||||
raise NotImplementedError
|
||||
|
||||
def _usage_metrics_from_payload(self, payload: dict[str, Any]) -> dict[str, int]:
|
||||
raise NotImplementedError
|
||||
|
||||
async def _rate_limit_backoff(self, attempt: int, exc: BaseException) -> None:
|
||||
await asyncio.sleep(min(5 * (attempt + 1), 30))
|
||||
|
||||
async def _transient_backoff(self, attempt: int, exc: BaseException) -> None:
|
||||
await asyncio.sleep(min(2 * (attempt + 1), 10))
|
||||
|
||||
# ---- shared infrastructure ----------------------------------------------------
|
||||
|
||||
def get_template_vars(self, **kwargs) -> dict[str, Any]:
|
||||
vars: dict[str, Any] = {
|
||||
"action_field": self.config.action_field,
|
||||
"model_name": self.config.model_name,
|
||||
}
|
||||
for k, v in self._last_request_metrics.items():
|
||||
vars[f"last_request_{k}"] = v
|
||||
for k, v in self._last_usage_metrics.items():
|
||||
vars[f"last_request_{k}"] = v
|
||||
for k, v in self._cumulative_request_metrics.items():
|
||||
vars[f"cumulative_request_{k}"] = v
|
||||
for k, v in self._cumulative_usage_metrics.items():
|
||||
vars[f"cumulative_{k}"] = v
|
||||
vars.update(kwargs)
|
||||
return vars
|
||||
|
||||
def _response_schema(self) -> dict[str, Any]:
|
||||
action_field = self.config.action_field
|
||||
return {
|
||||
"type": "object",
|
||||
"additionalProperties": False,
|
||||
"properties": {
|
||||
"thought": {"type": "string"},
|
||||
action_field: {"type": "string"},
|
||||
"done": {"type": "boolean"},
|
||||
"final_response": {"type": "string"},
|
||||
},
|
||||
"required": ["thought", action_field, "done", "final_response"],
|
||||
}
|
||||
|
||||
def _usage_snapshot(self) -> dict[str, dict[str, int]]:
|
||||
return {
|
||||
"last_request": {
|
||||
"message_count": self._last_request_metrics["message_count"],
|
||||
"text_part_count": self._last_request_metrics["text_part_count"],
|
||||
"image_part_count": self._last_request_metrics["image_part_count"],
|
||||
"input_tokens": self._last_usage_metrics["input_tokens"],
|
||||
"cached_input_tokens": self._last_usage_metrics["cached_input_tokens"],
|
||||
},
|
||||
"last_response": dict(self._last_usage_metrics),
|
||||
"cumulative_request": {
|
||||
"message_count": self._cumulative_request_metrics["message_count"],
|
||||
"text_part_count": self._cumulative_request_metrics["text_part_count"],
|
||||
"image_part_count": self._cumulative_request_metrics["image_part_count"],
|
||||
"input_tokens": self._cumulative_usage_metrics["input_tokens"],
|
||||
"cached_input_tokens": self._cumulative_usage_metrics["cached_input_tokens"],
|
||||
},
|
||||
"cumulative_response": dict(self._cumulative_usage_metrics),
|
||||
}
|
||||
|
||||
def _log_gateway_error(self, *, event: str, attempt: int, error: BaseException) -> None:
|
||||
response = getattr(error, "response", None)
|
||||
response_text = ""
|
||||
if response is not None:
|
||||
try:
|
||||
response_text = str(getattr(response, "text", "") or "")
|
||||
except Exception:
|
||||
response_text = ""
|
||||
if len(response_text) > 4000:
|
||||
response_text = response_text[:4000]
|
||||
|
||||
append_runtime_log(
|
||||
self.config.error_log_path,
|
||||
source=self._LOG_SOURCE,
|
||||
event=event,
|
||||
model_name=self.config.model_name,
|
||||
endpoint=self._post_url(),
|
||||
attempt=attempt,
|
||||
error_type=type(error).__name__,
|
||||
error=str(error),
|
||||
status_code=getattr(response, "status_code", None),
|
||||
response_text=response_text,
|
||||
)
|
||||
|
||||
def _raw_response_log_path(self) -> Path | None:
|
||||
if self.config.error_log_path is None:
|
||||
return None
|
||||
return self.config.error_log_path.parent / "raw_responses.jsonl"
|
||||
|
||||
def format_message(self, **kwargs) -> dict[str, Any]:
|
||||
role = kwargs["role"]
|
||||
content = kwargs.get("content", "")
|
||||
extra = kwargs.get("extra", {})
|
||||
return {"role": role, "content": content, "extra": extra}
|
||||
|
||||
def format_observation_messages(
|
||||
self,
|
||||
message: dict[str, Any],
|
||||
outputs: list[dict[str, Any]],
|
||||
template_vars: dict[str, Any] | None = None,
|
||||
) -> list[dict[str, Any]]:
|
||||
observation_messages: list[dict[str, Any]] = []
|
||||
for output in outputs:
|
||||
observation = output.get("observation", {})
|
||||
content = Template(self.config.observation_template, undefined=StrictUndefined).render(
|
||||
output=output,
|
||||
observation=observation,
|
||||
**(template_vars or {}),
|
||||
)
|
||||
|
||||
parts: list[dict[str, Any]] = [text_part(content)]
|
||||
screenshot_path = observation.get("screenshot_path")
|
||||
if self.config.attach_observation_screenshot and screenshot_path:
|
||||
parts.append(image_part_from_path(Path(screenshot_path)))
|
||||
|
||||
observation_messages.append(
|
||||
self.format_message(role="user", content=parts, extra={"observation": observation})
|
||||
)
|
||||
return observation_messages
|
||||
|
||||
def _format_error(self, *, raw_text: str, error: str) -> FormatError:
|
||||
return FormatError(
|
||||
self.format_message(
|
||||
role="user",
|
||||
content=Template(self.config.format_error_template, undefined=StrictUndefined).render(
|
||||
error=error,
|
||||
model_response=raw_text,
|
||||
**self.get_template_vars(),
|
||||
),
|
||||
extra={
|
||||
"interrupt_type": "FormatError",
|
||||
"model_response": raw_text,
|
||||
},
|
||||
)
|
||||
)
|
||||
|
||||
def _format_repair_message(self, *, raw_text: str, error: str) -> dict[str, Any]:
|
||||
return self.format_message(
|
||||
role="user",
|
||||
content=Template(self.config.format_error_template, undefined=StrictUndefined).render(
|
||||
error=error,
|
||||
model_response=raw_text,
|
||||
**self.get_template_vars(),
|
||||
),
|
||||
extra={
|
||||
"interrupt_type": "FormatErrorRetry",
|
||||
"model_response": raw_text,
|
||||
},
|
||||
)
|
||||
|
||||
async def _post_with_retries(self, payload: dict[str, Any]) -> dict[str, Any]:
|
||||
headers = self._request_headers()
|
||||
url = self._post_url()
|
||||
for attempt in range(max(self._MAX_RATE_LIMIT_RETRIES, self._MAX_TRANSIENT_RETRIES) + 1):
|
||||
try:
|
||||
async with httpx.AsyncClient(timeout=self.config.request_timeout_seconds) as client:
|
||||
response = await client.post(url, headers=headers, json=payload)
|
||||
response.raise_for_status()
|
||||
return response.json()
|
||||
except Exception as exc:
|
||||
if _is_rate_limit_error(exc):
|
||||
self._log_gateway_error(
|
||||
event="rate_limit_error", attempt=attempt + 1, error=exc
|
||||
)
|
||||
if attempt >= self._MAX_RATE_LIMIT_RETRIES:
|
||||
raise
|
||||
await self._rate_limit_backoff(attempt, exc)
|
||||
continue
|
||||
if _is_transient_http_error(exc):
|
||||
self._log_gateway_error(
|
||||
event="transient_http_error", attempt=attempt + 1, error=exc
|
||||
)
|
||||
if attempt >= self._MAX_TRANSIENT_RETRIES:
|
||||
raise
|
||||
await self._transient_backoff(attempt, exc)
|
||||
continue
|
||||
self._log_gateway_error(
|
||||
event="fatal_gateway_error", attempt=attempt + 1, error=exc
|
||||
)
|
||||
raise
|
||||
raise RuntimeError("Exceeded retry budget without exception or success.")
|
||||
|
||||
async def _query_async(self, messages: list[dict[str, Any]]) -> dict[str, Any]:
|
||||
last_error: ValueError | None = None
|
||||
raw_text = ""
|
||||
request_messages = list(messages)
|
||||
for attempt_index in range(MAX_JSON_PARSE_RETRIES + 1):
|
||||
payload = self._build_payload(request_messages)
|
||||
request_metrics = _request_metrics_from_serialized_input(self._request_metrics_input(payload))
|
||||
self._last_request_metrics = dict(request_metrics)
|
||||
for key, value in request_metrics.items():
|
||||
self._cumulative_request_metrics[key] += value
|
||||
|
||||
response_payload = await self._post_with_retries(payload)
|
||||
|
||||
usage_metrics = self._usage_metrics_from_payload(response_payload)
|
||||
self._last_usage_metrics = dict(usage_metrics)
|
||||
for key, value in usage_metrics.items():
|
||||
self._cumulative_usage_metrics[key] += value
|
||||
|
||||
raw_text = self._extract_text(response_payload)
|
||||
append_runtime_log(
|
||||
self._raw_response_log_path(),
|
||||
source="model",
|
||||
event="raw_text",
|
||||
attempt=attempt_index + 1,
|
||||
raw_text=raw_text,
|
||||
)
|
||||
try:
|
||||
parsed = parse_json_output(raw_text, action_field=self.config.action_field)
|
||||
break
|
||||
except ValueError as exc:
|
||||
last_error = exc
|
||||
if attempt_index < MAX_JSON_PARSE_RETRIES:
|
||||
request_messages.append(
|
||||
self._format_repair_message(raw_text=raw_text, error=str(exc))
|
||||
)
|
||||
else:
|
||||
raise self._format_error(
|
||||
raw_text=raw_text,
|
||||
error=str(last_error or ValueError("Unable to parse model output.")),
|
||||
)
|
||||
|
||||
actions: list[dict[str, Any]] = []
|
||||
action_field = self.config.action_field
|
||||
action_text = str(parsed.get(action_field, "") or "").strip()
|
||||
if action_text:
|
||||
action = {action_field: action_text, "command": action_text}
|
||||
if action_field == "bash_command":
|
||||
action["bash_command"] = action_text
|
||||
try:
|
||||
_validate_bash_command(action_text)
|
||||
except ValueError as exc:
|
||||
raise self._format_error(raw_text=raw_text, error=str(exc))
|
||||
else:
|
||||
action["python_code"] = action_text
|
||||
actions.append(action)
|
||||
|
||||
return self.format_message(
|
||||
role="assistant",
|
||||
content=parsed.get("thought", ""),
|
||||
extra={
|
||||
"actions": actions,
|
||||
"done": bool(parsed.get("done", False)),
|
||||
"final_response": parsed.get("final_response", ""),
|
||||
"raw_response": parsed,
|
||||
"usage": self._usage_snapshot(),
|
||||
},
|
||||
)
|
||||
|
||||
async def _complete_text_async(
|
||||
self,
|
||||
messages: list[dict[str, Any]],
|
||||
*,
|
||||
max_output_tokens: int | None = None,
|
||||
) -> str:
|
||||
original_max_output_tokens = self.config.max_output_tokens
|
||||
if max_output_tokens is not None:
|
||||
self.config.max_output_tokens = max_output_tokens
|
||||
try:
|
||||
payload = self._build_text_payload(messages)
|
||||
request_metrics = _request_metrics_from_serialized_input(self._request_metrics_input(payload))
|
||||
self._last_request_metrics = dict(request_metrics)
|
||||
for key, value in request_metrics.items():
|
||||
self._cumulative_request_metrics[key] += value
|
||||
|
||||
response_payload = await self._post_with_retries(payload)
|
||||
|
||||
usage_metrics = self._usage_metrics_from_payload(response_payload)
|
||||
self._last_usage_metrics = dict(usage_metrics)
|
||||
for key, value in usage_metrics.items():
|
||||
self._cumulative_usage_metrics[key] += value
|
||||
|
||||
raw_text = self._extract_text(response_payload)
|
||||
append_runtime_log(
|
||||
self._raw_response_log_path(),
|
||||
source="model",
|
||||
event="raw_text",
|
||||
raw_text=raw_text,
|
||||
)
|
||||
return raw_text
|
||||
finally:
|
||||
self.config.max_output_tokens = original_max_output_tokens
|
||||
|
||||
def __call__(
|
||||
self,
|
||||
messages: list[dict[str, Any]],
|
||||
**kwargs: Any,
|
||||
) -> str:
|
||||
return run_async(self._complete_text_async(messages, **kwargs))
|
||||
|
||||
def query(self, messages: list[dict[str, Any]], **kwargs) -> dict[str, Any]:
|
||||
return run_async(self._query_async(messages))
|
||||
|
||||
def serialize(self) -> dict[str, Any]:
|
||||
config_dump = self.config.model_dump(mode="json")
|
||||
if self._API_KEY_FIELD:
|
||||
config_dump[self._API_KEY_FIELD] = "<redacted>"
|
||||
return {
|
||||
"model": {
|
||||
"config": config_dump,
|
||||
"usage": {
|
||||
**self._usage_snapshot(),
|
||||
},
|
||||
"model_type": f"{self.__class__.__module__}.{self.__class__.__name__}",
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,157 @@
|
||||
"""OpenAI Responses API model backend."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
from webwright.models.base import (
|
||||
BaseModel,
|
||||
BaseModelConfig,
|
||||
OptStr,
|
||||
_safe_int,
|
||||
image_part_from_path,
|
||||
text_part,
|
||||
)
|
||||
|
||||
__all__ = [
|
||||
"OpenAIModel",
|
||||
"OpenAIModelConfig",
|
||||
"_extract_response_text",
|
||||
"text_part",
|
||||
]
|
||||
|
||||
|
||||
def _serialize_response_content_part(part: dict[str, Any], *, role: str) -> dict[str, Any]:
|
||||
if part.get("type") == "input_image":
|
||||
return {
|
||||
"type": "input_image",
|
||||
"image_url": part.get("image_url", ""),
|
||||
"detail": part.get("detail", "high"),
|
||||
}
|
||||
text = part.get("text", "")
|
||||
if role == "assistant":
|
||||
return {"type": "output_text", "text": text}
|
||||
return {"type": "input_text", "text": text}
|
||||
|
||||
|
||||
def _serialize_response_input(messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
|
||||
serialized: list[dict[str, Any]] = []
|
||||
for message in messages:
|
||||
role = message["role"]
|
||||
if role == "exit":
|
||||
continue
|
||||
content = message.get("content", "")
|
||||
if isinstance(content, str):
|
||||
serialized_content = [text_part(content)]
|
||||
else:
|
||||
serialized_content = [part for part in content if isinstance(part, dict)]
|
||||
# The Responses API accepts "developer" for system-style instructions.
|
||||
mapped_role = "developer" if role == "system" else role
|
||||
serialized.append(
|
||||
{
|
||||
"type": "message",
|
||||
"role": mapped_role,
|
||||
"content": [
|
||||
_serialize_response_content_part(part, role=mapped_role)
|
||||
for part in serialized_content
|
||||
],
|
||||
}
|
||||
)
|
||||
return serialized
|
||||
|
||||
|
||||
def _extract_response_text(payload: dict[str, Any]) -> str:
|
||||
output_text = payload.get("output_text")
|
||||
if isinstance(output_text, str) and output_text:
|
||||
return output_text
|
||||
|
||||
texts: list[str] = []
|
||||
for item in payload.get("output") or []:
|
||||
if not isinstance(item, dict):
|
||||
continue
|
||||
if item.get("type") != "message":
|
||||
continue
|
||||
for content in item.get("content") or []:
|
||||
if not isinstance(content, dict):
|
||||
continue
|
||||
if isinstance(content.get("text"), str):
|
||||
texts.append(content["text"])
|
||||
elif isinstance(content.get("output_text"), str):
|
||||
texts.append(content["output_text"])
|
||||
return "\n".join(texts)
|
||||
|
||||
|
||||
def _usage_metrics_from_response_payload(payload: dict[str, Any]) -> dict[str, int]:
|
||||
usage = payload.get("usage")
|
||||
if not isinstance(usage, dict):
|
||||
usage = {}
|
||||
input_details = usage.get("input_tokens_details")
|
||||
if not isinstance(input_details, dict):
|
||||
input_details = {}
|
||||
output_details = usage.get("output_tokens_details")
|
||||
if not isinstance(output_details, dict):
|
||||
output_details = {}
|
||||
|
||||
return {
|
||||
"input_tokens": _safe_int(usage.get("input_tokens")),
|
||||
"output_tokens": _safe_int(usage.get("output_tokens")),
|
||||
"total_tokens": _safe_int(usage.get("total_tokens")),
|
||||
"cached_input_tokens": _safe_int(input_details.get("cached_tokens")),
|
||||
"reasoning_output_tokens": _safe_int(output_details.get("reasoning_tokens")),
|
||||
}
|
||||
|
||||
|
||||
class OpenAIModelConfig(BaseModelConfig):
|
||||
model_name: OptStr = "gpt-4o"
|
||||
openai_api_key: OptStr = ""
|
||||
openai_endpoint: OptStr = "https://api.openai.com/v1/responses"
|
||||
|
||||
|
||||
class OpenAIModel(BaseModel):
|
||||
_API_KEY_FIELD = "openai_api_key"
|
||||
_ENV_VAR = "OPENAI_API_KEY"
|
||||
_LOG_SOURCE = "openai"
|
||||
_MAX_RATE_LIMIT_RETRIES = 5
|
||||
_MAX_TRANSIENT_RETRIES = 5
|
||||
_DEFAULT_CONFIG_CLASS = OpenAIModelConfig
|
||||
|
||||
def _request_headers(self) -> dict[str, str]:
|
||||
return {
|
||||
"Content-Type": "application/json",
|
||||
"Authorization": f"Bearer {self.config.openai_api_key}",
|
||||
}
|
||||
|
||||
def _post_url(self) -> str:
|
||||
return self.config.openai_endpoint
|
||||
|
||||
def _build_payload(self, messages: list[dict[str, Any]]) -> dict[str, Any]:
|
||||
return {
|
||||
"model": self.config.model_name,
|
||||
"input": _serialize_response_input(messages),
|
||||
"max_output_tokens": self.config.max_output_tokens,
|
||||
"text": {
|
||||
"format": {
|
||||
"type": "json_schema",
|
||||
"name": "playwright_step",
|
||||
"schema": self._response_schema(),
|
||||
"strict": True,
|
||||
}
|
||||
},
|
||||
}
|
||||
|
||||
def _build_text_payload(self, messages: list[dict[str, Any]]) -> dict[str, Any]:
|
||||
return {
|
||||
"model": self.config.model_name,
|
||||
"input": _serialize_response_input(messages),
|
||||
"max_output_tokens": self.config.max_output_tokens,
|
||||
}
|
||||
|
||||
def _request_metrics_input(self, payload: dict[str, Any]) -> list[dict[str, Any]]:
|
||||
return payload.get("input") or []
|
||||
|
||||
def _extract_text(self, payload: dict[str, Any]) -> str:
|
||||
return _extract_response_text(payload)
|
||||
|
||||
def _usage_metrics_from_payload(self, payload: dict[str, Any]) -> dict[str, int]:
|
||||
return _usage_metrics_from_response_payload(payload)
|
||||
@@ -0,0 +1,206 @@
|
||||
"""OpenRouter chat completions model backend."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from typing import Any
|
||||
from urllib.parse import urlparse
|
||||
|
||||
from webwright.models.base import (
|
||||
BaseModel,
|
||||
BaseModelConfig,
|
||||
OptStr,
|
||||
_safe_int,
|
||||
)
|
||||
|
||||
__all__ = [
|
||||
"OpenRouterModel",
|
||||
"OpenRouterModelConfig",
|
||||
]
|
||||
|
||||
|
||||
def _serialize_chat_content_part(part: dict[str, Any]) -> dict[str, Any] | None:
|
||||
part_type = part.get("type")
|
||||
if part_type in {"input_text", "output_text"}:
|
||||
return {"type": "text", "text": str(part.get("text", "") or "")}
|
||||
if part_type == "input_image":
|
||||
return {
|
||||
"type": "image_url",
|
||||
"image_url": {
|
||||
"url": str(part.get("image_url", "") or ""),
|
||||
"detail": str(part.get("detail", "high") or "high"),
|
||||
},
|
||||
}
|
||||
return None
|
||||
|
||||
|
||||
def _serialize_chat_messages(messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
|
||||
serialized: list[dict[str, Any]] = []
|
||||
for message in messages:
|
||||
role = message["role"]
|
||||
if role == "exit":
|
||||
continue
|
||||
mapped_role = "system" if role == "system" else ("assistant" if role == "assistant" else "user")
|
||||
content = message.get("content", "")
|
||||
if isinstance(content, str):
|
||||
serialized.append({"role": mapped_role, "content": content})
|
||||
continue
|
||||
parts = [
|
||||
serialized_part
|
||||
for part in content
|
||||
if isinstance(part, dict)
|
||||
for serialized_part in [_serialize_chat_content_part(part)]
|
||||
if serialized_part is not None
|
||||
]
|
||||
if mapped_role == "assistant" or all(part.get("type") == "text" for part in parts):
|
||||
serialized.append(
|
||||
{
|
||||
"role": mapped_role,
|
||||
"content": "\n".join(str(part.get("text", "") or "") for part in parts),
|
||||
}
|
||||
)
|
||||
else:
|
||||
serialized.append({"role": mapped_role, "content": parts})
|
||||
return serialized
|
||||
|
||||
|
||||
def _metrics_input_from_chat_messages(chat_messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
|
||||
metrics_input: list[dict[str, Any]] = []
|
||||
for message in chat_messages:
|
||||
content = message.get("content", "")
|
||||
if isinstance(content, str):
|
||||
metrics_input.append({"content": [{"type": "input_text", "text": content}]})
|
||||
continue
|
||||
parts: list[dict[str, Any]] = []
|
||||
for part in content:
|
||||
if not isinstance(part, dict):
|
||||
continue
|
||||
if part.get("type") == "text":
|
||||
parts.append({"type": "input_text", "text": str(part.get("text", "") or "")})
|
||||
elif part.get("type") == "image_url":
|
||||
parts.append({"type": "input_image"})
|
||||
metrics_input.append({"content": parts})
|
||||
return metrics_input
|
||||
|
||||
|
||||
def _extract_chat_completions_text(payload: dict[str, Any]) -> str:
|
||||
choices = payload.get("choices")
|
||||
if not isinstance(choices, list) or not choices:
|
||||
return ""
|
||||
first_choice = choices[0]
|
||||
if not isinstance(first_choice, dict):
|
||||
return ""
|
||||
message = first_choice.get("message", {})
|
||||
if not isinstance(message, dict):
|
||||
return ""
|
||||
content = message.get("content", "")
|
||||
if isinstance(content, str):
|
||||
return content
|
||||
if isinstance(content, list):
|
||||
return "\n".join(
|
||||
str(part.get("text", "") or "")
|
||||
for part in content
|
||||
if isinstance(part, dict) and part.get("type") == "text"
|
||||
)
|
||||
return ""
|
||||
|
||||
|
||||
def _usage_metrics_from_chat_completions(payload: dict[str, Any]) -> dict[str, int]:
|
||||
usage = payload.get("usage")
|
||||
if not isinstance(usage, dict):
|
||||
usage = {}
|
||||
return {
|
||||
"input_tokens": _safe_int(usage.get("prompt_tokens")),
|
||||
"output_tokens": _safe_int(usage.get("completion_tokens")),
|
||||
"total_tokens": _safe_int(usage.get("total_tokens")),
|
||||
"cached_input_tokens": 0,
|
||||
"reasoning_output_tokens": 0,
|
||||
}
|
||||
|
||||
|
||||
def _endpoint_host(endpoint: str) -> str:
|
||||
return (urlparse(endpoint).hostname or "").lower()
|
||||
|
||||
|
||||
def _is_openai_endpoint(endpoint: str) -> bool:
|
||||
return _endpoint_host(endpoint) == "api.openai.com"
|
||||
|
||||
|
||||
def _is_openrouter_endpoint(endpoint: str) -> bool:
|
||||
return _endpoint_host(endpoint).endswith("openrouter.ai")
|
||||
|
||||
|
||||
class OpenRouterModelConfig(BaseModelConfig):
|
||||
model_name: OptStr = "openai/gpt-5.4"
|
||||
openrouter_api_key: OptStr = ""
|
||||
openrouter_endpoint: OptStr = "https://openrouter.ai/api/v1/chat/completions"
|
||||
http_referer: OptStr = ""
|
||||
app_title: OptStr = ""
|
||||
provider_require_parameters: bool = True
|
||||
|
||||
|
||||
class OpenRouterModel(BaseModel):
|
||||
_API_KEY_FIELD = "openrouter_api_key"
|
||||
_ENV_VAR = "OPENROUTER_API_KEY"
|
||||
_LOG_SOURCE = "openrouter"
|
||||
_MAX_RATE_LIMIT_RETRIES = 5
|
||||
_MAX_TRANSIENT_RETRIES = 5
|
||||
_DEFAULT_CONFIG_CLASS = OpenRouterModelConfig
|
||||
|
||||
def _request_headers(self) -> dict[str, str]:
|
||||
headers = {
|
||||
"Content-Type": "application/json",
|
||||
"Authorization": f"Bearer {self.config.openrouter_api_key}",
|
||||
}
|
||||
if _is_openrouter_endpoint(self.config.openrouter_endpoint) and self.config.http_referer:
|
||||
headers["HTTP-Referer"] = self.config.http_referer
|
||||
if _is_openrouter_endpoint(self.config.openrouter_endpoint) and self.config.app_title:
|
||||
headers["X-Title"] = self.config.app_title
|
||||
return headers
|
||||
|
||||
def _post_url(self) -> str:
|
||||
return self.config.openrouter_endpoint
|
||||
|
||||
def _build_payload(self, messages: list[dict[str, Any]]) -> dict[str, Any]:
|
||||
payload: dict[str, Any] = {
|
||||
"model": self.config.model_name,
|
||||
"messages": _serialize_chat_messages(messages),
|
||||
"stream": False,
|
||||
"response_format": {
|
||||
"type": "json_schema",
|
||||
"json_schema": {
|
||||
"name": "playwright_step",
|
||||
"strict": True,
|
||||
"schema": self._response_schema(),
|
||||
},
|
||||
},
|
||||
}
|
||||
if _is_openai_endpoint(self.config.openrouter_endpoint) and self.config.model_name.startswith("gpt-5"):
|
||||
payload["max_completion_tokens"] = self.config.max_output_tokens
|
||||
else:
|
||||
payload["max_tokens"] = self.config.max_output_tokens
|
||||
if self.config.provider_require_parameters and _is_openrouter_endpoint(self.config.openrouter_endpoint):
|
||||
payload["provider"] = {"require_parameters": True}
|
||||
return payload
|
||||
|
||||
def _build_text_payload(self, messages: list[dict[str, Any]]) -> dict[str, Any]:
|
||||
payload: dict[str, Any] = {
|
||||
"model": self.config.model_name,
|
||||
"messages": _serialize_chat_messages(messages),
|
||||
"stream": False,
|
||||
}
|
||||
if _is_openai_endpoint(self.config.openrouter_endpoint) and self.config.model_name.startswith("gpt-5"):
|
||||
payload["max_completion_tokens"] = self.config.max_output_tokens
|
||||
else:
|
||||
payload["max_tokens"] = self.config.max_output_tokens
|
||||
if self.config.provider_require_parameters and _is_openrouter_endpoint(self.config.openrouter_endpoint):
|
||||
payload["provider"] = {"require_parameters": True}
|
||||
return payload
|
||||
|
||||
def _request_metrics_input(self, payload: dict[str, Any]) -> list[dict[str, Any]]:
|
||||
return _metrics_input_from_chat_messages(payload.get("messages") or [])
|
||||
|
||||
def _extract_text(self, payload: dict[str, Any]) -> str:
|
||||
return _extract_chat_completions_text(payload)
|
||||
|
||||
def _usage_metrics_from_payload(self, payload: dict[str, Any]) -> dict[str, int]:
|
||||
return _usage_metrics_from_chat_completions(payload)
|
||||
@@ -0,0 +1,175 @@
|
||||
from __future__ import annotations
|
||||
|
||||
from datetime import datetime
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
import typer
|
||||
from rich.console import Console
|
||||
|
||||
from webwright.agents import get_agent
|
||||
from webwright.config import get_config_from_spec, snapshot_config_specs
|
||||
from webwright.environments import get_environment
|
||||
from webwright.models import get_model
|
||||
from webwright.utils.serialize import UNSET, recursive_merge
|
||||
from webwright.run.doctor import run_doctor
|
||||
|
||||
|
||||
DEFAULT_CONFIGS = ["base.yaml", "model_openai.yaml"]
|
||||
|
||||
app = typer.Typer(no_args_is_help=True)
|
||||
console = Console(highlight=False)
|
||||
|
||||
|
||||
def _timestamped_output_dir(base_dir: str | Path | None, task_id: str | None) -> Path:
|
||||
base = Path(base_dir or "outputs").expanduser()
|
||||
stamp = datetime.now().strftime("%Y%m%d_%H%M%S")
|
||||
suffix = task_id or "adhoc"
|
||||
return base / f"{suffix}_{stamp}"
|
||||
|
||||
|
||||
def run_one(
|
||||
*,
|
||||
task: str | None = None,
|
||||
task_id: str | None = None,
|
||||
start_url: str | None = None,
|
||||
config_spec: list[str] | None = None,
|
||||
output_dir: Path | None = None,
|
||||
resolved_output_dir: Path | None = None,
|
||||
debug: bool = False,
|
||||
snapshot_config: bool = True,
|
||||
) -> Any:
|
||||
config_spec = config_spec or DEFAULT_CONFIGS
|
||||
configs = [get_config_from_spec(spec) for spec in config_spec]
|
||||
config = recursive_merge(*configs)
|
||||
|
||||
run_config = config.get("run", {})
|
||||
resolved_task_id = task_id or run_config.get("task_id")
|
||||
resolved_task = task or run_config.get("task")
|
||||
resolved_start_url = start_url or run_config.get("start_url")
|
||||
|
||||
if not resolved_task:
|
||||
raise ValueError("A task is required. Use --task.")
|
||||
|
||||
resolved_output_dir = resolved_output_dir or _timestamped_output_dir(
|
||||
output_dir or config.get("environment", {}).get("output_dir") or "outputs",
|
||||
resolved_task_id,
|
||||
)
|
||||
if snapshot_config:
|
||||
snapshot_config_specs(config_spec, resolved_output_dir, merged_config=config)
|
||||
|
||||
config = recursive_merge(
|
||||
config,
|
||||
{
|
||||
"run": {
|
||||
"task": resolved_task,
|
||||
"task_id": resolved_task_id or UNSET,
|
||||
"start_url": resolved_start_url or UNSET,
|
||||
},
|
||||
"environment": {
|
||||
"output_dir": str(resolved_output_dir),
|
||||
"start_url": resolved_start_url or UNSET,
|
||||
"headless": False if debug else UNSET,
|
||||
"devtools": True if debug else UNSET,
|
||||
"keep_open_on_exit": True if debug else UNSET,
|
||||
"prompt_before_close": True if debug else UNSET,
|
||||
"slow_mo_ms": 250 if debug else UNSET,
|
||||
},
|
||||
"model": {
|
||||
"error_log_path": str(resolved_output_dir / "runtime_errors.jsonl"),
|
||||
},
|
||||
"agent": {
|
||||
"output_path": str(resolved_output_dir / "trajectory.json"),
|
||||
},
|
||||
},
|
||||
)
|
||||
|
||||
model = get_model(config.get("model", {}))
|
||||
env = get_environment(config.get("environment", {}))
|
||||
agent = get_agent(model, env, config.get("agent", {}), default_type="default")
|
||||
|
||||
console.print(f"Running task in [bold green]{resolved_output_dir}[/bold green]")
|
||||
run_exception: Exception | None = None
|
||||
close_exception: Exception | None = None
|
||||
result: dict[str, Any] = {}
|
||||
try:
|
||||
env.prepare(
|
||||
task=resolved_task,
|
||||
task_id=resolved_task_id,
|
||||
start_url=resolved_start_url,
|
||||
)
|
||||
result = agent.run(
|
||||
resolved_task,
|
||||
task_id=resolved_task_id or "",
|
||||
start_url=resolved_start_url or "",
|
||||
)
|
||||
except Exception as exc:
|
||||
run_exception = exc
|
||||
if getattr(agent, "messages", None):
|
||||
result = dict(agent.messages[-1].get("extra", {}))
|
||||
result.setdefault("exit_status", type(exc).__name__)
|
||||
result.setdefault("submission", "")
|
||||
result.setdefault("final_response", "")
|
||||
result["run_exception"] = str(exc)
|
||||
finally:
|
||||
try:
|
||||
env.close()
|
||||
except Exception as exc:
|
||||
close_exception = exc
|
||||
result.setdefault("exit_status", type(exc).__name__)
|
||||
result.setdefault("submission", "")
|
||||
result.setdefault("final_response", "")
|
||||
result.setdefault("run_exception", str(exc))
|
||||
result["close_exception"] = str(exc)
|
||||
if run_exception is None:
|
||||
run_exception = exc
|
||||
result["_output_dir"] = str(resolved_output_dir)
|
||||
if close_exception is not None:
|
||||
result["_close_exception"] = str(close_exception)
|
||||
console.print(
|
||||
result.get("final_response") or result.get("submission") or "Task finished."
|
||||
)
|
||||
if run_exception is not None:
|
||||
raise run_exception
|
||||
return result
|
||||
|
||||
|
||||
@app.command()
|
||||
def main(
|
||||
task: str = typer.Option(
|
||||
..., "-t", "--task", help="Natural language task description."
|
||||
),
|
||||
task_id: str | None = typer.Option(
|
||||
None, "--task-id", help="Optional identifier used in the output directory name."
|
||||
),
|
||||
start_url: str | None = typer.Option(
|
||||
None, "--start-url", help="Optional starting URL for the task."
|
||||
),
|
||||
config_spec: list[str] = typer.Option(DEFAULT_CONFIGS, "-c", "--config"),
|
||||
output_dir: Path | None = typer.Option(None, "-o", "--output-dir"),
|
||||
debug: bool = typer.Option(
|
||||
False,
|
||||
"--debug",
|
||||
help="Launch headed local Playwright with devtools and keep it open for inspection.",
|
||||
),
|
||||
) -> Any:
|
||||
return run_one(
|
||||
task=task,
|
||||
task_id=task_id,
|
||||
start_url=start_url,
|
||||
config_spec=config_spec,
|
||||
output_dir=output_dir,
|
||||
debug=debug,
|
||||
)
|
||||
|
||||
|
||||
@app.command()
|
||||
def doctor():
|
||||
"""
|
||||
Validate local Webwright setup.
|
||||
"""
|
||||
run_doctor()
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
app()
|
||||
@@ -0,0 +1,147 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import os
|
||||
import subprocess
|
||||
import sys
|
||||
from importlib.util import find_spec
|
||||
from pathlib import Path
|
||||
from rich.console import Console
|
||||
from rich.table import Table
|
||||
|
||||
console = Console()
|
||||
|
||||
|
||||
def check_python():
|
||||
version = sys.version_info
|
||||
|
||||
if version >= (3, 10):
|
||||
return True, f"Python {version.major}.{version.minor}"
|
||||
|
||||
return False, ("Python 3.10+ required\nFix: install Python 3.10 or newer")
|
||||
|
||||
|
||||
def check_playwright():
|
||||
if find_spec("playwright") is not None:
|
||||
return True, "playwright installed"
|
||||
|
||||
return False, ("playwright not installed\nFix: pip install playwright")
|
||||
|
||||
|
||||
def check_chromium():
|
||||
try:
|
||||
result = subprocess.run(
|
||||
["playwright", "install", "--dry-run"],
|
||||
capture_output=True,
|
||||
text=True,
|
||||
)
|
||||
|
||||
if result.returncode == 0:
|
||||
return True, "chromium available"
|
||||
|
||||
return False, ("chromium missing\nFix: playwright install chromium")
|
||||
|
||||
except Exception as e:
|
||||
return False, str(e)
|
||||
|
||||
|
||||
def check_screenshot():
|
||||
try:
|
||||
from playwright.sync_api import sync_playwright
|
||||
|
||||
screenshot_path = Path("doctor_test.png")
|
||||
|
||||
with sync_playwright() as p:
|
||||
browser = p.chromium.launch(headless=True)
|
||||
|
||||
page = browser.new_page()
|
||||
|
||||
page.set_content("<h1>Webwright Doctor</h1>")
|
||||
|
||||
page.screenshot(path=str(screenshot_path))
|
||||
|
||||
browser.close()
|
||||
|
||||
if screenshot_path.exists():
|
||||
screenshot_path.unlink(missing_ok=True)
|
||||
|
||||
return True, "screenshot capture working"
|
||||
|
||||
return False, "screenshot file was not created"
|
||||
|
||||
except Exception:
|
||||
return False, (
|
||||
"unable to launch Chromium for screenshot validation\n"
|
||||
"Fix: playwright install"
|
||||
)
|
||||
|
||||
|
||||
def check_openai_key():
|
||||
if os.getenv("OPENAI_API_KEY"):
|
||||
return True, "OPENAI_API_KEY found"
|
||||
|
||||
return False, (
|
||||
"OPENAI_API_KEY missing\nFix: set the OPENAI_API_KEY environment variable"
|
||||
)
|
||||
|
||||
|
||||
def check_plugin_manifests():
|
||||
claude = Path(".claude-plugin/plugin.json")
|
||||
codex = Path(".codex-plugin/plugin.json")
|
||||
|
||||
missing = []
|
||||
|
||||
if not claude.exists():
|
||||
missing.append("Claude")
|
||||
|
||||
if not codex.exists():
|
||||
missing.append("Codex")
|
||||
|
||||
if not missing:
|
||||
return True, "plugin manifests found"
|
||||
|
||||
return False, (
|
||||
f"missing plugin manifests: {', '.join(missing)}\n"
|
||||
"Fix: configure Claude/Codex plugins"
|
||||
)
|
||||
|
||||
|
||||
CHECKS = [
|
||||
("Python", check_python),
|
||||
("Playwright", check_playwright),
|
||||
("Chromium", check_chromium),
|
||||
("Screenshot", check_screenshot),
|
||||
("OpenAI Key", check_openai_key),
|
||||
("Plugins", check_plugin_manifests),
|
||||
]
|
||||
|
||||
|
||||
def run_doctor():
|
||||
table = Table(title="Webwright Doctor")
|
||||
|
||||
table.add_column("Check")
|
||||
table.add_column("Status")
|
||||
table.add_column("Details")
|
||||
|
||||
passed = 0
|
||||
|
||||
for name, fn in CHECKS:
|
||||
ok, message = fn()
|
||||
|
||||
status = "PASS" if ok else "FAIL"
|
||||
|
||||
table.add_row(
|
||||
name,
|
||||
status,
|
||||
message,
|
||||
)
|
||||
|
||||
if ok:
|
||||
passed += 1
|
||||
|
||||
console.print(table)
|
||||
|
||||
console.print(f"\n{passed}/{len(CHECKS)} checks passed")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
run_doctor()
|
||||
@@ -0,0 +1,6 @@
|
||||
python -m webwright.run.cli \
|
||||
-c base.yaml -c model_openai.yaml \
|
||||
-t "Search for flights from SEA to JFK on 2026-08-15 to 2026-08-20" \
|
||||
--start-url https://www.google.com/flights \
|
||||
--task-id demo_openai \
|
||||
-o outputs/default
|
||||
@@ -0,0 +1,77 @@
|
||||
"""Resolve the model client used by inner tools (image_qa, self_reflection).
|
||||
|
||||
The CLI snapshots the fully merged run config to
|
||||
``<workspace_dir>/config_snapshot/merged_config.yaml``; the tools read that file
|
||||
(or an explicit ``--model-config`` override) and instantiate the same model the
|
||||
agent uses.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
import yaml
|
||||
|
||||
from webwright.models import get_model
|
||||
|
||||
DEFAULT_MERGED_CONFIG_RELPATH = Path("config_snapshot") / "merged_config.yaml"
|
||||
|
||||
|
||||
def _load_structured_config(path: Path) -> dict[str, Any]:
|
||||
text = path.read_text(encoding="utf-8")
|
||||
if path.suffix.lower() in {".yaml", ".yml"}:
|
||||
loaded = yaml.safe_load(text)
|
||||
else:
|
||||
loaded = json.loads(text)
|
||||
if not isinstance(loaded, dict):
|
||||
raise ValueError(f"Model config must be an object: {path}")
|
||||
return loaded
|
||||
|
||||
|
||||
def _extract_model_block(config: dict[str, Any]) -> dict[str, Any]:
|
||||
model_block = config.get("model")
|
||||
if not isinstance(model_block, dict):
|
||||
raise ValueError(
|
||||
"Model config is missing a top-level `model:` block; "
|
||||
"stack a model_*.yaml (e.g. model_claude.yaml) or pass --model-config <path>."
|
||||
)
|
||||
return model_block
|
||||
|
||||
|
||||
def resolve_model_config_path(model_config_arg: str, *, workspace_dir: str) -> Path:
|
||||
"""Return the path to a config containing a top-level ``model:`` block.
|
||||
|
||||
Resolution order:
|
||||
1. ``model_config_arg`` (absolute or relative to ``workspace_dir``).
|
||||
2. ``<workspace_dir>/config_snapshot/merged_config.yaml`` (written by the CLI).
|
||||
"""
|
||||
candidates: list[Path] = []
|
||||
if model_config_arg:
|
||||
configured = Path(model_config_arg)
|
||||
candidates.append(configured)
|
||||
if workspace_dir and not configured.is_absolute():
|
||||
candidates.append(Path(workspace_dir) / configured)
|
||||
if workspace_dir:
|
||||
candidates.append(Path(workspace_dir) / DEFAULT_MERGED_CONFIG_RELPATH)
|
||||
for candidate in candidates:
|
||||
if candidate.exists():
|
||||
return candidate.resolve()
|
||||
raise FileNotFoundError(
|
||||
"No tool model config found. Pass --model-config <path> or run via the agent so "
|
||||
f"<workspace-dir>/{DEFAULT_MERGED_CONFIG_RELPATH} is available."
|
||||
)
|
||||
|
||||
|
||||
def load_tool_model(
|
||||
*,
|
||||
model_config_arg: str,
|
||||
workspace_dir: str,
|
||||
timeout_seconds: int,
|
||||
) -> Any:
|
||||
config_path = resolve_model_config_path(model_config_arg, workspace_dir=workspace_dir)
|
||||
config = _load_structured_config(config_path)
|
||||
model_block = dict(_extract_model_block(config))
|
||||
model_block["request_timeout_seconds"] = timeout_seconds
|
||||
return get_model(model_block)
|
||||
@@ -0,0 +1,141 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
import base64
|
||||
import json
|
||||
import mimetypes
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
from webwright.models.base import text_part
|
||||
from webwright.tools._model_config import load_tool_model
|
||||
|
||||
|
||||
def _build_prompt(question: str) -> str:
|
||||
return (
|
||||
"Answer the user's question using only visible evidence from the provided image or images. "
|
||||
"If the answer is not visible, say so instead of guessing. Return only a JSON object with "
|
||||
"string `answer`, string array `evidence`, boolean `unknown`, and number `confidence`.\n\n"
|
||||
f"Question: {question.strip()}"
|
||||
)
|
||||
|
||||
|
||||
def _high_detail_image_part_from_path(image_path: Path) -> dict[str, Any]:
|
||||
mime_type, _ = mimetypes.guess_type(str(image_path))
|
||||
encoded = base64.b64encode(image_path.read_bytes()).decode("ascii")
|
||||
return {
|
||||
"type": "input_image",
|
||||
"image_url": f"data:{mime_type or 'image/png'};base64,{encoded}",
|
||||
"detail": "high",
|
||||
}
|
||||
|
||||
|
||||
def _resolve_image_path(image_path: str, workspace_dir: str = "") -> Path:
|
||||
path = Path(image_path)
|
||||
if not path.is_absolute():
|
||||
base_dir = Path(workspace_dir) if workspace_dir else Path.cwd()
|
||||
path = base_dir / path
|
||||
path = path.resolve()
|
||||
if not path.exists():
|
||||
raise FileNotFoundError(f"Image path does not exist: {path}")
|
||||
return path
|
||||
|
||||
|
||||
def _normalize_image_paths(
|
||||
*,
|
||||
image_path: Path | None = None,
|
||||
image_paths: list[Path] | tuple[Path, ...] | None = None,
|
||||
) -> list[Path]:
|
||||
normalized = list(image_paths or [])
|
||||
if image_path is not None:
|
||||
normalized.insert(0, image_path)
|
||||
if not normalized:
|
||||
raise ValueError("At least one image path is required.")
|
||||
return normalized
|
||||
|
||||
|
||||
def _parse_json_response(raw_text: str) -> dict[str, Any]:
|
||||
try:
|
||||
parsed = json.loads(raw_text)
|
||||
except json.JSONDecodeError:
|
||||
start = raw_text.find("{")
|
||||
end = raw_text.rfind("}")
|
||||
if start == -1 or end == -1 or end <= start:
|
||||
raise
|
||||
parsed = json.loads(raw_text[start : end + 1])
|
||||
if not isinstance(parsed, dict):
|
||||
raise ValueError("image_qa model response must be a JSON object.")
|
||||
return parsed
|
||||
|
||||
|
||||
def run_image_qa(
|
||||
*,
|
||||
image_path: Path | None = None,
|
||||
image_paths: list[Path] | tuple[Path, ...] | None = None,
|
||||
question: str,
|
||||
model_client: Any,
|
||||
) -> dict[str, Any]:
|
||||
resolved_image_paths = _normalize_image_paths(image_path=image_path, image_paths=image_paths)
|
||||
raw_text = model_client(
|
||||
[
|
||||
{
|
||||
"role": "user",
|
||||
"content": [text_part(_build_prompt(question))]
|
||||
+ [_high_detail_image_part_from_path(path) for path in resolved_image_paths],
|
||||
}
|
||||
],
|
||||
max_output_tokens=32000,
|
||||
).strip()
|
||||
parsed = _parse_json_response(raw_text)
|
||||
result = {
|
||||
"image_paths": [str(path) for path in resolved_image_paths],
|
||||
"question": question,
|
||||
**parsed,
|
||||
}
|
||||
if len(resolved_image_paths) == 1:
|
||||
result["image_path"] = str(resolved_image_paths[0])
|
||||
return result
|
||||
|
||||
|
||||
def build_parser() -> argparse.ArgumentParser:
|
||||
parser = argparse.ArgumentParser(description="Ask a visual question about a local image and print JSON.")
|
||||
parser.add_argument(
|
||||
"--image",
|
||||
required=True,
|
||||
action="append",
|
||||
help="Path to an image file. Repeat --image to include multiple images.",
|
||||
)
|
||||
parser.add_argument("--question", required=True, help="Question to answer from the image.")
|
||||
parser.add_argument("--workspace-dir", default="", help="Optional base directory for relative image paths.")
|
||||
parser.add_argument(
|
||||
"--model-config",
|
||||
default="",
|
||||
help=(
|
||||
"Path to a JSON/YAML config containing a top-level `model:` block. "
|
||||
"If omitted, reads <workspace-dir>/config_snapshot/merged_config.yaml."
|
||||
),
|
||||
)
|
||||
parser.add_argument("--timeout-seconds", type=int, default=60, help="HTTP request timeout.")
|
||||
return parser
|
||||
|
||||
|
||||
def main(argv: list[str] | None = None) -> int:
|
||||
parser = build_parser()
|
||||
args = parser.parse_args(argv)
|
||||
image_paths = [_resolve_image_path(image_path, workspace_dir=args.workspace_dir) for image_path in args.image]
|
||||
model_client = load_tool_model(
|
||||
model_config_arg=args.model_config,
|
||||
workspace_dir=args.workspace_dir,
|
||||
timeout_seconds=args.timeout_seconds,
|
||||
)
|
||||
result = run_image_qa(
|
||||
image_paths=image_paths,
|
||||
question=args.question,
|
||||
model_client=model_client,
|
||||
)
|
||||
print(json.dumps(result, ensure_ascii=True, indent=2))
|
||||
return 0
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
raise SystemExit(main())
|
||||
@@ -0,0 +1,314 @@
|
||||
"""CLI for managing a long-lived, reusable local Chromium browser session.
|
||||
|
||||
Local-browser counterpart to ``browserbase_session.py``. Launches a
|
||||
detached headless Chromium subprocess via the Playwright-bundled binary
|
||||
with ``--remote-debugging-port=0`` and a per-session ``--user-data-dir``,
|
||||
parses the printed ``DevTools listening on ws://...`` URL, and persists
|
||||
``{id, pid, connectUrl, userDataDir}`` to a JSON file on disk so any
|
||||
later Python/bash step can attach via
|
||||
``playwright.chromium.connect_over_cdp(connectUrl)`` and end with
|
||||
``await browser.disconnect()`` (NEVER ``browser.close()``) to keep the
|
||||
browser alive across steps.
|
||||
|
||||
Subcommands:
|
||||
|
||||
* ``create`` -> spawn detached Chromium, write JSON, print id.
|
||||
* ``info`` -> print whether the saved PID is still alive plus
|
||||
the persisted JSON.
|
||||
* ``release`` -> SIGTERM (then SIGKILL) the PID, optionally remove
|
||||
the user-data-dir and the JSON file.
|
||||
|
||||
Usage:
|
||||
python -m webwright.tools.persistent_local_browser create --workspace-dir <ws> --out .lb_session.json
|
||||
python -m webwright.tools.persistent_local_browser info --workspace-dir <ws> --session-file .lb_session.json
|
||||
python -m webwright.tools.persistent_local_browser release --workspace-dir <ws> --session-file .lb_session.json --delete-file
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
import json
|
||||
import os
|
||||
import re
|
||||
import shutil
|
||||
import signal
|
||||
import subprocess
|
||||
import sys
|
||||
import time
|
||||
import uuid
|
||||
from pathlib import Path
|
||||
|
||||
DEFAULT_SESSION_FILE = ".lb_session.json"
|
||||
DEFAULT_USER_DATA_SUBDIR = ".lb_user_data"
|
||||
_DEVTOOLS_RE = re.compile(r"DevTools listening on (ws://\S+)")
|
||||
|
||||
|
||||
def _resolve_path(path_str: str, workspace_dir: str = "") -> Path:
|
||||
path = Path(path_str)
|
||||
if not path.is_absolute():
|
||||
base = Path(workspace_dir) if workspace_dir else Path.cwd()
|
||||
path = base / path
|
||||
return path
|
||||
|
||||
|
||||
def _chromium_executable() -> str:
|
||||
"""Locate the Playwright-bundled Chromium executable."""
|
||||
try:
|
||||
from playwright.sync_api import sync_playwright
|
||||
except ImportError as exc: # pragma: no cover - import guard
|
||||
raise SystemExit(f"error: playwright is not installed: {exc}")
|
||||
with sync_playwright() as p:
|
||||
path = p.chromium.executable_path
|
||||
if not path or not Path(path).exists():
|
||||
raise SystemExit(
|
||||
"error: Playwright chromium binary not found. Run `playwright install chromium`."
|
||||
)
|
||||
return path
|
||||
|
||||
|
||||
def _pid_alive(pid: int) -> bool:
|
||||
if pid <= 0:
|
||||
return False
|
||||
try:
|
||||
os.kill(pid, 0)
|
||||
except ProcessLookupError:
|
||||
return False
|
||||
except PermissionError:
|
||||
return True
|
||||
return True
|
||||
|
||||
|
||||
def _wait_for_devtools_url(proc: subprocess.Popen, timeout: float) -> str:
|
||||
"""Read Chromium's stderr until the DevTools ws:// URL appears."""
|
||||
assert proc.stderr is not None
|
||||
deadline = time.monotonic() + timeout
|
||||
while time.monotonic() < deadline:
|
||||
if proc.poll() is not None:
|
||||
tail = ""
|
||||
try:
|
||||
tail = proc.stderr.read() or ""
|
||||
except Exception: # noqa: BLE001
|
||||
pass
|
||||
raise SystemExit(
|
||||
f"error: chromium exited early (code={proc.returncode}); stderr tail:\n{tail}"
|
||||
)
|
||||
line = proc.stderr.readline()
|
||||
if not line:
|
||||
time.sleep(0.05)
|
||||
continue
|
||||
match = _DEVTOOLS_RE.search(line)
|
||||
if match:
|
||||
return match.group(1).strip()
|
||||
raise SystemExit(
|
||||
f"error: timed out after {timeout:.1f}s waiting for 'DevTools listening on ws://...' line"
|
||||
)
|
||||
|
||||
|
||||
def _cmd_create(args: argparse.Namespace) -> int:
|
||||
workspace_dir = args.workspace_dir or str(Path.cwd())
|
||||
out_path = _resolve_path(args.out, workspace_dir)
|
||||
user_data_dir = _resolve_path(
|
||||
args.user_data_dir or DEFAULT_USER_DATA_SUBDIR, workspace_dir
|
||||
)
|
||||
out_path.parent.mkdir(parents=True, exist_ok=True)
|
||||
user_data_dir.mkdir(parents=True, exist_ok=True)
|
||||
|
||||
chromium = _chromium_executable()
|
||||
chromium_args = [
|
||||
chromium,
|
||||
"--remote-debugging-port=0",
|
||||
f"--user-data-dir={user_data_dir}",
|
||||
"--no-first-run",
|
||||
"--no-default-browser-check",
|
||||
"--disable-features=TranslateUI,MediaRouter",
|
||||
f"--window-size={args.window_width},{args.window_height}",
|
||||
]
|
||||
if args.headless:
|
||||
chromium_args.append("--headless=new")
|
||||
if args.no_sandbox:
|
||||
chromium_args.append("--no-sandbox")
|
||||
chromium_args.extend(args.chromium_arg or [])
|
||||
|
||||
# Detach into its own process group so it survives the parent shell exit.
|
||||
popen_kwargs: dict = {
|
||||
"stdout": subprocess.DEVNULL,
|
||||
"stderr": subprocess.PIPE,
|
||||
"stdin": subprocess.DEVNULL,
|
||||
"text": True,
|
||||
"bufsize": 1,
|
||||
"close_fds": True,
|
||||
}
|
||||
if os.name == "posix":
|
||||
popen_kwargs["start_new_session"] = True
|
||||
|
||||
proc = subprocess.Popen(chromium_args, **popen_kwargs) # noqa: S603
|
||||
try:
|
||||
connect_url = _wait_for_devtools_url(proc, args.startup_timeout)
|
||||
except SystemExit:
|
||||
try:
|
||||
proc.terminate()
|
||||
except Exception: # noqa: BLE001
|
||||
pass
|
||||
raise
|
||||
|
||||
session = {
|
||||
"id": uuid.uuid4().hex,
|
||||
"pid": proc.pid,
|
||||
"connectUrl": connect_url,
|
||||
"userDataDir": str(user_data_dir),
|
||||
"executablePath": chromium,
|
||||
"headless": bool(args.headless),
|
||||
"createdAt": int(time.time()),
|
||||
}
|
||||
out_path.write_text(json.dumps(session, indent=2) + "\n", encoding="utf-8")
|
||||
|
||||
print(f"LB_SESSION_ID={session['id']}")
|
||||
print(f"LB_SESSION_PID={session['pid']}")
|
||||
print(f"LB_SESSION_FILE={out_path}")
|
||||
print(f"LB_CONNECT_URL={connect_url}")
|
||||
return 0
|
||||
|
||||
|
||||
def _cmd_info(args: argparse.Namespace) -> int:
|
||||
session_path = _resolve_path(args.session_file, args.workspace_dir)
|
||||
if not session_path.exists():
|
||||
print(f"LB_INFO_MISSING file={session_path}")
|
||||
return 1
|
||||
session = json.loads(session_path.read_text(encoding="utf-8"))
|
||||
session["alive"] = _pid_alive(int(session.get("pid", 0)))
|
||||
print(json.dumps(session, indent=2))
|
||||
return 0
|
||||
|
||||
|
||||
def _terminate_pid(pid: int, kill_timeout: float) -> str:
|
||||
if pid <= 0 or not _pid_alive(pid):
|
||||
return "not_running"
|
||||
try:
|
||||
os.kill(pid, signal.SIGTERM)
|
||||
except ProcessLookupError:
|
||||
return "already_gone"
|
||||
deadline = time.monotonic() + kill_timeout
|
||||
while time.monotonic() < deadline:
|
||||
if not _pid_alive(pid):
|
||||
return "terminated"
|
||||
time.sleep(0.1)
|
||||
try:
|
||||
os.kill(pid, signal.SIGKILL)
|
||||
except ProcessLookupError:
|
||||
return "already_gone"
|
||||
time.sleep(0.2)
|
||||
return "killed" if not _pid_alive(pid) else "still_alive"
|
||||
|
||||
|
||||
def _cmd_release(args: argparse.Namespace) -> int:
|
||||
session_path = _resolve_path(args.session_file, args.workspace_dir)
|
||||
if not session_path.exists():
|
||||
print(f"LB_RELEASE_SKIPPED missing={session_path}")
|
||||
return 0
|
||||
session = json.loads(session_path.read_text(encoding="utf-8"))
|
||||
pid = int(session.get("pid", 0))
|
||||
status = _terminate_pid(pid, args.kill_timeout)
|
||||
print(f"LB_RELEASE_REQUESTED pid={pid} status={status}")
|
||||
|
||||
if args.delete_user_data:
|
||||
udd = session.get("userDataDir", "")
|
||||
if udd and Path(udd).exists():
|
||||
try:
|
||||
shutil.rmtree(udd)
|
||||
print(f"LB_USER_DATA_DELETED {udd}")
|
||||
except OSError as exc:
|
||||
print(f"LB_USER_DATA_DELETE_FAILED {udd} {exc}")
|
||||
|
||||
if args.delete_file:
|
||||
try:
|
||||
session_path.unlink()
|
||||
print(f"LB_SESSION_FILE_DELETED {session_path}")
|
||||
except OSError as exc:
|
||||
print(f"LB_SESSION_FILE_DELETE_FAILED {session_path} {exc}")
|
||||
return 0
|
||||
|
||||
|
||||
def _build_parser() -> argparse.ArgumentParser:
|
||||
parser = argparse.ArgumentParser(
|
||||
prog="python -m webwright.tools.persistent_local_browser",
|
||||
description="Manage a keep-alive local Chromium session shared across bash steps.",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--workspace-dir",
|
||||
default="",
|
||||
help="Resolve --out / --session-file / --user-data-dir relative to this directory.",
|
||||
)
|
||||
|
||||
sub = parser.add_subparsers(dest="command", required=True)
|
||||
|
||||
create = sub.add_parser("create", help="Launch a detached Chromium and persist its connectUrl.")
|
||||
create.add_argument("--out", default=DEFAULT_SESSION_FILE, help="Where to write the session JSON.")
|
||||
create.add_argument(
|
||||
"--user-data-dir",
|
||||
default="",
|
||||
help=f"Per-session Chromium user-data-dir (default: <workspace>/{DEFAULT_USER_DATA_SUBDIR}).",
|
||||
)
|
||||
create.add_argument(
|
||||
"--headless",
|
||||
action=argparse.BooleanOptionalAction,
|
||||
default=True,
|
||||
help="Launch Chromium headless (default: True).",
|
||||
)
|
||||
create.add_argument(
|
||||
"--no-sandbox",
|
||||
action=argparse.BooleanOptionalAction,
|
||||
default=True,
|
||||
help="Pass --no-sandbox (often required in containers/CI).",
|
||||
)
|
||||
create.add_argument("--window-width", type=int, default=1280)
|
||||
create.add_argument("--window-height", type=int, default=1800)
|
||||
create.add_argument(
|
||||
"--startup-timeout",
|
||||
type=float,
|
||||
default=30.0,
|
||||
help="Seconds to wait for the DevTools URL to appear on stderr.",
|
||||
)
|
||||
create.add_argument(
|
||||
"--chromium-arg",
|
||||
action="append",
|
||||
default=[],
|
||||
help="Extra Chromium command-line argument; repeat for multiple.",
|
||||
)
|
||||
create.set_defaults(func=_cmd_create)
|
||||
|
||||
info = sub.add_parser("info", help="Print the persisted session JSON and liveness.")
|
||||
info.add_argument("--session-file", default=DEFAULT_SESSION_FILE)
|
||||
info.set_defaults(func=_cmd_info)
|
||||
|
||||
release = sub.add_parser("release", help="Terminate the persisted session.")
|
||||
release.add_argument("--session-file", default=DEFAULT_SESSION_FILE)
|
||||
release.add_argument(
|
||||
"--delete-file",
|
||||
action=argparse.BooleanOptionalAction,
|
||||
default=True,
|
||||
help="Also delete the session JSON file after release.",
|
||||
)
|
||||
release.add_argument(
|
||||
"--delete-user-data",
|
||||
action=argparse.BooleanOptionalAction,
|
||||
default=True,
|
||||
help="Also remove the per-session Chromium user-data-dir.",
|
||||
)
|
||||
release.add_argument(
|
||||
"--kill-timeout",
|
||||
type=float,
|
||||
default=10.0,
|
||||
help="Seconds to wait for SIGTERM before sending SIGKILL.",
|
||||
)
|
||||
release.set_defaults(func=_cmd_release)
|
||||
|
||||
return parser
|
||||
|
||||
|
||||
def main(argv: list[str] | None = None) -> int:
|
||||
args = _build_parser().parse_args(argv)
|
||||
return args.func(args)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
sys.exit(main())
|
||||
@@ -0,0 +1,611 @@
|
||||
"""Self-reflection two-stage screenshot judge CLI.
|
||||
|
||||
Previously named ``two_stage_judge``; renamed to ``self_reflection``.
|
||||
|
||||
Stage 1: for each screenshot, send a (system, user + image) pair to the
|
||||
configured model and parse a 1-5 ``Score`` with a short ``Reasoning``.
|
||||
|
||||
Stage 2: drop every per-image ``Reasoning`` into the caller-provided final
|
||||
user prompt template (via ``{image_reasonings}``), attach EVERY screenshot,
|
||||
and make one final call that must end with ``Status: success`` or
|
||||
``Status: failure``.
|
||||
|
||||
The CLI reads all of its config from a single JSON file so the agent can
|
||||
prepare it in one turn and invoke the tool in the next.
|
||||
|
||||
Usage::
|
||||
|
||||
python -m webwright.tools.self_reflection --config self_reflect_config.json
|
||||
|
||||
JSON schema (paths relative to ``--workspace-dir`` or the CWD)::
|
||||
|
||||
{
|
||||
"images": ["final_runs/run_001/screenshots/final_execution_1.png", ...],
|
||||
"image_judge_system_prompt": "...",
|
||||
"image_judge_user_prompt": "...", // sent verbatim with each image
|
||||
"final_verdict_system_prompt": "...",
|
||||
"final_verdict_user_prompt": "...{action_history_log}...{image_reasonings}..."
|
||||
}
|
||||
|
||||
Any of the four prompt fields may instead be supplied via
|
||||
``<field>_file`` variants pointing to a text file on disk (recommended when
|
||||
prompts contain many literal braces or newlines).
|
||||
|
||||
The output JSON written to ``--output`` (or stdout) contains the per-image
|
||||
records, the image path list, the final response, and
|
||||
``predicted_label`` (``1`` for success, ``0`` for failure, ``null`` if the
|
||||
``Status:`` line could not be parsed). Exit code: 0 if PASS, 1 otherwise.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
import asyncio
|
||||
import base64
|
||||
import json
|
||||
import mimetypes
|
||||
import re
|
||||
import sys
|
||||
from dataclasses import dataclass
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
from webwright.models.base import text_part
|
||||
from webwright.tools._model_config import load_tool_model
|
||||
|
||||
DEFAULT_IMAGE_PARSE_MAX_RETRIES = 3
|
||||
|
||||
_PROMPT_FIELDS = (
|
||||
("image_judge_system_prompt", True),
|
||||
("image_judge_user_prompt", True),
|
||||
("final_verdict_system_prompt", True),
|
||||
("final_verdict_user_prompt", True),
|
||||
)
|
||||
|
||||
_IMAGE_SUFFIXES = frozenset({".png", ".jpg", ".jpeg", ".webp"})
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Image helpers
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
def _resolve_image_path(image_path: str, workspace_dir: str = "") -> Path:
|
||||
path = Path(image_path)
|
||||
if not path.is_absolute():
|
||||
base_dir = Path(workspace_dir) if workspace_dir else Path.cwd()
|
||||
path = base_dir / path
|
||||
path = path.resolve()
|
||||
if not path.exists():
|
||||
raise FileNotFoundError(f"Image path does not exist: {path}")
|
||||
return path
|
||||
|
||||
|
||||
def _final_execution_sort_key(name: str) -> tuple[int, str]:
|
||||
match = re.match(r"final_execution_(\d+)_", name)
|
||||
if match:
|
||||
return (int(match.group(1)), name)
|
||||
nums = re.findall(r"\d+", name)
|
||||
return (int(nums[0]) if nums else 0, name)
|
||||
|
||||
|
||||
def _run_id_sort_key(name: str) -> tuple[int, str]:
|
||||
match = re.search(r"run_(\d+)", name)
|
||||
if match:
|
||||
return (int(match.group(1)), name)
|
||||
return (0, name)
|
||||
|
||||
|
||||
def _sorted_image_paths(image_dir: Path) -> list[Path]:
|
||||
if not image_dir.is_dir():
|
||||
return []
|
||||
return sorted(
|
||||
[path for path in image_dir.iterdir() if path.is_file() and path.suffix.lower() in _IMAGE_SUFFIXES],
|
||||
key=lambda path: _final_execution_sort_key(path.name),
|
||||
)
|
||||
|
||||
|
||||
def _discover_latest_run_screenshots(
|
||||
final_runs_dir: Path,
|
||||
) -> tuple[Path | None, list[Path]]:
|
||||
"""Find the highest-numbered ``final_runs/run_<id>/screenshots`` dir and its images.
|
||||
|
||||
Returns ``(run_dir_or_None, sorted_image_paths)``. Empty list if no images found.
|
||||
"""
|
||||
if not final_runs_dir.exists() or not final_runs_dir.is_dir():
|
||||
return None, []
|
||||
candidates = sorted(
|
||||
(d for d in final_runs_dir.iterdir() if d.is_dir() and re.fullmatch(r"run_\d+", d.name)),
|
||||
key=lambda p: _run_id_sort_key(p.name),
|
||||
)
|
||||
# Walk from highest-numbered run downward and pick the first one with any screenshots.
|
||||
for run_dir in reversed(candidates):
|
||||
screenshots_dir = run_dir / "screenshots"
|
||||
images = _sorted_image_paths(screenshots_dir)
|
||||
if images:
|
||||
return run_dir, images
|
||||
return None, []
|
||||
|
||||
|
||||
def _infer_run_dir_from_images(images: list[Path]) -> Path | None:
|
||||
run_dirs = {
|
||||
path.parent.parent.resolve()
|
||||
for path in images
|
||||
if path.parent.name == "screenshots"
|
||||
}
|
||||
if len(run_dirs) == 1:
|
||||
return next(iter(run_dirs))
|
||||
return None
|
||||
|
||||
|
||||
def _resolve_artifact_dir(
|
||||
*,
|
||||
images: list[Path],
|
||||
discovered_run_dir: Path | None,
|
||||
output_path: str,
|
||||
workspace_dir: str,
|
||||
) -> Path | None:
|
||||
candidates: list[Path] = []
|
||||
|
||||
inferred_run_dir = _infer_run_dir_from_images(images)
|
||||
if inferred_run_dir is not None:
|
||||
candidates.append(inferred_run_dir)
|
||||
|
||||
if discovered_run_dir is not None:
|
||||
candidates.append(discovered_run_dir.resolve())
|
||||
|
||||
if output_path:
|
||||
candidates.append(Path(output_path).resolve().parent)
|
||||
|
||||
base_dir = Path(workspace_dir).resolve() if workspace_dir else Path.cwd().resolve()
|
||||
candidates.append(base_dir)
|
||||
|
||||
seen: set[Path] = set()
|
||||
ordered_candidates: list[Path] = []
|
||||
for candidate in candidates:
|
||||
if candidate in seen:
|
||||
continue
|
||||
seen.add(candidate)
|
||||
ordered_candidates.append(candidate)
|
||||
|
||||
for candidate in ordered_candidates:
|
||||
if (candidate / "final_script_log.txt").is_file():
|
||||
return candidate
|
||||
|
||||
return ordered_candidates[0] if ordered_candidates else None
|
||||
|
||||
|
||||
def _load_action_history_log(artifact_dir: Path | None) -> str:
|
||||
if artifact_dir is None:
|
||||
return ""
|
||||
log_path = artifact_dir / "final_script_log.txt"
|
||||
if not log_path.is_file():
|
||||
return ""
|
||||
return log_path.read_text(encoding="utf-8").rstrip()
|
||||
|
||||
|
||||
def _render_final_verdict_user_prompt(
|
||||
template: str,
|
||||
*,
|
||||
image_reasonings: str,
|
||||
action_history_log: str,
|
||||
) -> str:
|
||||
rendered = template
|
||||
if "{image_reasonings}" in template or "{action_history_log}" in template:
|
||||
try:
|
||||
rendered = template.format(
|
||||
image_reasonings=image_reasonings,
|
||||
action_history_log=action_history_log,
|
||||
)
|
||||
except KeyError as exc:
|
||||
raise ValueError(
|
||||
"Unknown placeholder in final_verdict_user_prompt: "
|
||||
f"{exc.args[0]!r}. Supported placeholders are "
|
||||
"{image_reasonings} and {action_history_log}; double any literal "
|
||||
"braces as {{ and }}."
|
||||
) from exc
|
||||
return rendered
|
||||
|
||||
|
||||
def _high_detail_image_part_from_path(image_path: Path) -> dict[str, Any]:
|
||||
mime_type, _ = mimetypes.guess_type(str(image_path))
|
||||
encoded = base64.b64encode(image_path.read_bytes()).decode("ascii")
|
||||
return {
|
||||
"type": "input_image",
|
||||
"image_url": f"data:{mime_type or 'image/png'};base64,{encoded}",
|
||||
"detail": "high",
|
||||
}
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Model call: plain message list -> text
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
def _call_model(
|
||||
*,
|
||||
model_client: Any,
|
||||
system_prompt: str,
|
||||
user_content: list[dict[str, Any]],
|
||||
max_new_tokens: int,
|
||||
) -> str:
|
||||
return model_client(
|
||||
[
|
||||
{"role": "system", "content": system_prompt},
|
||||
{"role": "user", "content": user_content},
|
||||
],
|
||||
max_output_tokens=max_new_tokens,
|
||||
).strip()
|
||||
|
||||
|
||||
def _model_endpoint(model_client: Any) -> str:
|
||||
config = getattr(model_client, "config", None)
|
||||
for key in ("openai_endpoint", "anthropic_endpoint", "openrouter_endpoint"):
|
||||
value = getattr(config, key, "")
|
||||
if value:
|
||||
return str(value)
|
||||
return ""
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Parsing helpers (ported from webjudge_online_mind2web_sandbox.py)
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
def _parse_image_judge_response(response: str) -> tuple[str, int]:
|
||||
score_match = re.search(r"(?is)\bscore\b[^1-5]*([1-5])\b", response)
|
||||
reasoning_match = re.search(
|
||||
r"(?is)(?:\*\*?\s*reasoning\s*\*\*?|reasoning)\s*[:\-]\s*"
|
||||
r"(.*?)(?=\n\s*(?:\d+\.\s*)?(?:\*\*?\s*score\s*\*\*?|score)\s*[:\-]|\Z)",
|
||||
response,
|
||||
)
|
||||
|
||||
if score_match and reasoning_match:
|
||||
reasoning = re.sub(r"\s+", " ", reasoning_match.group(1)).strip()
|
||||
return reasoning, int(score_match.group(1))
|
||||
|
||||
try:
|
||||
payload = json.loads(response)
|
||||
except Exception:
|
||||
payload = None
|
||||
|
||||
if isinstance(payload, dict):
|
||||
score = payload.get("Score", payload.get("score"))
|
||||
reasoning = payload.get("Reasoning", payload.get("reasoning"))
|
||||
if (
|
||||
isinstance(score, int)
|
||||
and 1 <= score <= 5
|
||||
and isinstance(reasoning, str)
|
||||
and reasoning.strip()
|
||||
):
|
||||
return re.sub(r"\s+", " ", reasoning).strip(), score
|
||||
|
||||
raise ValueError("Could not parse image judge response")
|
||||
|
||||
|
||||
def _parse_final_verdict(response: str) -> int | None:
|
||||
matches = list(re.finditer(r"(?i)status:\s*", response))
|
||||
if not matches:
|
||||
return None
|
||||
tail = response[matches[-1].end():].strip()
|
||||
m = re.match(r"""^[\'\"“”‘’\s]*(success|failure)\b""", tail, re.IGNORECASE)
|
||||
if not m:
|
||||
return None
|
||||
return 1 if m.group(1).lower() == "success" else 0
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Per-image scoring
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
async def _judge_one_image(
|
||||
*,
|
||||
image_path: Path,
|
||||
image_judge_system_prompt: str,
|
||||
image_judge_user_prompt: str,
|
||||
model_client: Any,
|
||||
max_new_tokens: int,
|
||||
max_parse_retries: int,
|
||||
) -> dict[str, Any]:
|
||||
user_content = [
|
||||
text_part(image_judge_user_prompt),
|
||||
_high_detail_image_part_from_path(image_path),
|
||||
]
|
||||
|
||||
last_response = ""
|
||||
last_error: BaseException | None = None
|
||||
for attempt in range(1, max_parse_retries + 1):
|
||||
last_response = await asyncio.to_thread(
|
||||
_call_model,
|
||||
model_client=model_client,
|
||||
system_prompt=image_judge_system_prompt,
|
||||
user_content=user_content,
|
||||
max_new_tokens=max_new_tokens,
|
||||
)
|
||||
try:
|
||||
reasoning, score = _parse_image_judge_response(last_response)
|
||||
return {
|
||||
"image_path": str(image_path),
|
||||
"Response": last_response,
|
||||
"Score": score,
|
||||
"Reasoning": reasoning,
|
||||
"Attempts": attempt,
|
||||
"ParseFailed": False,
|
||||
}
|
||||
except Exception as exc: # noqa: BLE001
|
||||
last_error = exc
|
||||
print(
|
||||
f"[self_reflection] parse attempt {attempt}/{max_parse_retries} failed for "
|
||||
f"{image_path}: {exc}",
|
||||
file=sys.stderr,
|
||||
)
|
||||
|
||||
return {
|
||||
"image_path": str(image_path),
|
||||
"Response": last_response,
|
||||
"Score": 0,
|
||||
"Reasoning": "",
|
||||
"Attempts": max_parse_retries,
|
||||
"ParseFailed": True,
|
||||
"ParseError": str(last_error) if last_error is not None else "unknown",
|
||||
}
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Orchestrator
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
@dataclass
|
||||
class SelfReflectionResult:
|
||||
image_records: list[dict[str, Any]]
|
||||
image_paths: list[str]
|
||||
final_user_text: str
|
||||
final_system_msg: str
|
||||
final_response: str
|
||||
predicted_label: int | None # 1 success, 0 failure, None unparsed
|
||||
model: str = ""
|
||||
endpoint: str = ""
|
||||
|
||||
def to_dict(self) -> dict[str, Any]:
|
||||
return {
|
||||
"model": self.model,
|
||||
"endpoint": self.endpoint,
|
||||
"predicted_label": self.predicted_label,
|
||||
"final_response": self.final_response,
|
||||
"final_user_text": self.final_user_text,
|
||||
"final_system_msg": self.final_system_msg,
|
||||
"image_paths": self.image_paths,
|
||||
"image_records": self.image_records,
|
||||
}
|
||||
|
||||
|
||||
async def run_self_reflection_async(
|
||||
*,
|
||||
images: list[Path],
|
||||
image_judge_system_prompt: str,
|
||||
image_judge_user_prompt: str,
|
||||
final_verdict_system_prompt: str,
|
||||
final_verdict_user_prompt: str,
|
||||
action_history_log: str,
|
||||
max_image_parse_retries: int,
|
||||
final_max_new_tokens: int,
|
||||
image_max_new_tokens: int,
|
||||
model_client: Any,
|
||||
) -> SelfReflectionResult:
|
||||
model_name = str(getattr(model_client.config, "model_name", ""))
|
||||
endpoint = _model_endpoint(model_client)
|
||||
|
||||
if images:
|
||||
per_image = await asyncio.gather(
|
||||
*(
|
||||
_judge_one_image(
|
||||
image_path=path,
|
||||
image_judge_system_prompt=image_judge_system_prompt,
|
||||
image_judge_user_prompt=image_judge_user_prompt,
|
||||
model_client=model_client,
|
||||
max_new_tokens=image_max_new_tokens,
|
||||
max_parse_retries=max_image_parse_retries,
|
||||
)
|
||||
for path in images
|
||||
)
|
||||
)
|
||||
else:
|
||||
per_image = []
|
||||
|
||||
image_paths = [record["image_path"] for record in per_image]
|
||||
reasonings = [record["Reasoning"] or "" for record in per_image]
|
||||
|
||||
reasonings_block = "\n".join(
|
||||
f"{i + 1}. {text}" for i, text in enumerate(reasonings)
|
||||
)
|
||||
|
||||
final_user_text = _render_final_verdict_user_prompt(
|
||||
final_verdict_user_prompt,
|
||||
image_reasonings=reasonings_block,
|
||||
action_history_log=action_history_log,
|
||||
)
|
||||
|
||||
user_content: list[dict[str, Any]] = [text_part(final_user_text)]
|
||||
for path_str in image_paths:
|
||||
user_content.append(_high_detail_image_part_from_path(Path(path_str)))
|
||||
|
||||
final_response = await asyncio.to_thread(
|
||||
_call_model,
|
||||
model_client=model_client,
|
||||
system_prompt=final_verdict_system_prompt,
|
||||
user_content=user_content,
|
||||
max_new_tokens=final_max_new_tokens,
|
||||
)
|
||||
predicted_label = _parse_final_verdict(final_response)
|
||||
|
||||
return SelfReflectionResult(
|
||||
image_records=list(per_image),
|
||||
image_paths=image_paths,
|
||||
final_user_text=final_user_text,
|
||||
final_system_msg=final_verdict_system_prompt,
|
||||
final_response=final_response,
|
||||
predicted_label=predicted_label,
|
||||
model=model_name,
|
||||
endpoint=endpoint,
|
||||
)
|
||||
|
||||
|
||||
def run_self_reflection(**kwargs: Any) -> SelfReflectionResult:
|
||||
return asyncio.run(run_self_reflection_async(**kwargs))
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# CLI
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
def _resolve_prompt(cfg: dict[str, Any], key: str, *, required: bool) -> str | None:
|
||||
inline = cfg.get(key)
|
||||
file_key = f"{key}_file"
|
||||
file_path = cfg.get(file_key)
|
||||
if inline is not None and file_path is not None:
|
||||
raise ValueError(f"Provide only one of {key!r} or {file_key!r}, not both.")
|
||||
if file_path is not None:
|
||||
return Path(file_path).read_text(encoding="utf-8")
|
||||
if inline is not None:
|
||||
return inline
|
||||
if required:
|
||||
raise ValueError(f"Missing required prompt: {key} (or {file_key}).")
|
||||
return None
|
||||
|
||||
|
||||
def _load_config(config_arg: str) -> dict[str, Any]:
|
||||
if config_arg == "-":
|
||||
return json.loads(sys.stdin.read())
|
||||
return json.loads(Path(config_arg).read_text(encoding="utf-8"))
|
||||
|
||||
|
||||
def build_parser() -> argparse.ArgumentParser:
|
||||
parser = argparse.ArgumentParser(
|
||||
description=(
|
||||
"Two-stage screenshot judge. Reads a JSON config describing images and "
|
||||
"prompts, calls the configured model, and prints a "
|
||||
"JSON result with per-image records and the final verdict."
|
||||
)
|
||||
)
|
||||
parser.add_argument("--config", required=True, help="Path to JSON config, or '-' for stdin.")
|
||||
parser.add_argument("--workspace-dir", default="", help="Base directory for relative image paths.")
|
||||
parser.add_argument("--output", default="", help="Write JSON result to this path instead of stdout.")
|
||||
parser.add_argument(
|
||||
"--auto-latest-run",
|
||||
default="final_runs",
|
||||
help=(
|
||||
"When the config has no 'images' list, auto-discover screenshots from the "
|
||||
"highest-numbered `<workspace-dir>/<this-value>/run_<id>/screenshots` folder. "
|
||||
"Default: 'final_runs'. Pass '' (empty string) to disable auto-discovery."
|
||||
),
|
||||
)
|
||||
parser.add_argument("--max-image-parse-retries", type=int, default=DEFAULT_IMAGE_PARSE_MAX_RETRIES)
|
||||
parser.add_argument("--image-max-new-tokens", type=int, default=1024)
|
||||
parser.add_argument("--final-max-new-tokens", type=int, default=8192)
|
||||
parser.add_argument(
|
||||
"--model-config",
|
||||
default="",
|
||||
help=(
|
||||
"Path to a JSON/YAML config containing a top-level `model:` block. "
|
||||
"If omitted, reads <workspace-dir>/config_snapshot/merged_config.yaml."
|
||||
),
|
||||
)
|
||||
parser.add_argument("--timeout-seconds", type=int, default=120)
|
||||
return parser
|
||||
|
||||
|
||||
def main(argv: list[str] | None = None) -> int:
|
||||
parser = build_parser()
|
||||
args = parser.parse_args(argv)
|
||||
base_dir = Path(args.workspace_dir).resolve() if args.workspace_dir else Path.cwd().resolve()
|
||||
|
||||
cfg = _load_config(args.config)
|
||||
|
||||
prompts = {
|
||||
key: _resolve_prompt(cfg, key, required=required)
|
||||
for key, required in _PROMPT_FIELDS
|
||||
}
|
||||
|
||||
images_config = cfg.get("images") or cfg.get("images_path") or []
|
||||
resolved_images = [
|
||||
_resolve_image_path(p, workspace_dir=args.workspace_dir) for p in images_config
|
||||
]
|
||||
discovered_run_dir = _infer_run_dir_from_images(resolved_images)
|
||||
|
||||
# If config did not provide images, fall back to the latest run's screenshots.
|
||||
if not resolved_images:
|
||||
discovered: list[Path] = []
|
||||
discovered_source = ""
|
||||
if args.auto_latest_run:
|
||||
auto_root = Path(args.auto_latest_run)
|
||||
if not auto_root.is_absolute():
|
||||
auto_root = base_dir / auto_root
|
||||
auto_root = auto_root.resolve()
|
||||
discovered_run_dir, discovered = _discover_latest_run_screenshots(auto_root)
|
||||
if discovered_run_dir is not None:
|
||||
discovered_source = str(discovered_run_dir / "screenshots")
|
||||
if discovered:
|
||||
resolved_images = discovered
|
||||
print(
|
||||
f"[self_reflection] auto-discovered {len(resolved_images)} screenshots from {discovered_source}",
|
||||
file=sys.stderr,
|
||||
)
|
||||
|
||||
artifact_dir = _resolve_artifact_dir(
|
||||
images=resolved_images,
|
||||
discovered_run_dir=discovered_run_dir,
|
||||
output_path=args.output,
|
||||
workspace_dir=args.workspace_dir,
|
||||
)
|
||||
action_history_log = _load_action_history_log(artifact_dir)
|
||||
|
||||
if not resolved_images:
|
||||
print(
|
||||
"[self_reflection] warning: no images provided; final stage will run without screenshot attachments.",
|
||||
file=sys.stderr,
|
||||
)
|
||||
|
||||
if not action_history_log:
|
||||
print(
|
||||
"[self_reflection] warning: no final_script_log.txt found; final prompt will omit action history content.",
|
||||
file=sys.stderr,
|
||||
)
|
||||
|
||||
model_client = load_tool_model(
|
||||
model_config_arg=args.model_config,
|
||||
workspace_dir=args.workspace_dir,
|
||||
timeout_seconds=args.timeout_seconds,
|
||||
)
|
||||
|
||||
result = run_self_reflection(
|
||||
images=resolved_images,
|
||||
image_judge_system_prompt=prompts["image_judge_system_prompt"],
|
||||
image_judge_user_prompt=prompts["image_judge_user_prompt"],
|
||||
final_verdict_system_prompt=prompts["final_verdict_system_prompt"],
|
||||
final_verdict_user_prompt=prompts["final_verdict_user_prompt"],
|
||||
action_history_log=action_history_log,
|
||||
max_image_parse_retries=args.max_image_parse_retries,
|
||||
final_max_new_tokens=args.final_max_new_tokens,
|
||||
image_max_new_tokens=args.image_max_new_tokens,
|
||||
model_client=model_client,
|
||||
)
|
||||
|
||||
payload = result.to_dict()
|
||||
serialized = json.dumps(payload, indent=2, ensure_ascii=False)
|
||||
if args.output:
|
||||
Path(args.output).write_text(serialized, encoding="utf-8")
|
||||
print(f"Wrote result to {args.output}", file=sys.stderr)
|
||||
else:
|
||||
sys.stdout.write(serialized)
|
||||
sys.stdout.write("\n")
|
||||
|
||||
label = result.predicted_label
|
||||
if label == 1:
|
||||
print("JUDGE VERDICT: PASS", file=sys.stderr)
|
||||
return 0
|
||||
if label == 0:
|
||||
print("JUDGE VERDICT: FAIL", file=sys.stderr)
|
||||
return 1
|
||||
print("JUDGE VERDICT: UNPARSED (treating as FAIL)", file=sys.stderr)
|
||||
return 1
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
raise SystemExit(main())
|
||||
@@ -0,0 +1,21 @@
|
||||
from __future__ import annotations
|
||||
|
||||
from datetime import datetime, timezone
|
||||
import json
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
|
||||
def append_runtime_log(path: Path | None, *, source: str, event: str, **data: Any) -> None:
|
||||
if path is None:
|
||||
return
|
||||
path.parent.mkdir(parents=True, exist_ok=True)
|
||||
payload = {
|
||||
"timestamp": datetime.now(timezone.utc).isoformat(),
|
||||
"source": source,
|
||||
"event": event,
|
||||
**data,
|
||||
}
|
||||
with path.open("a", encoding="utf-8") as handle:
|
||||
handle.write(json.dumps(payload, ensure_ascii=True))
|
||||
handle.write("\n")
|
||||
@@ -0,0 +1,14 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
from typing import TypeVar
|
||||
|
||||
T = TypeVar("T")
|
||||
|
||||
|
||||
def run_async(coro) -> T:
|
||||
try:
|
||||
asyncio.get_running_loop()
|
||||
except RuntimeError:
|
||||
return asyncio.run(coro)
|
||||
raise RuntimeError("Webwright does not support running inside an active event loop.")
|
||||
@@ -0,0 +1,22 @@
|
||||
from __future__ import annotations
|
||||
|
||||
from typing import Any
|
||||
|
||||
UNSET = object()
|
||||
|
||||
|
||||
def recursive_merge(*dictionaries: dict | None) -> dict[str, Any]:
|
||||
result: dict[str, Any] = {}
|
||||
for dictionary in dictionaries:
|
||||
if dictionary is None:
|
||||
continue
|
||||
for key, value in dictionary.items():
|
||||
if value is UNSET:
|
||||
continue
|
||||
if key in result and isinstance(result[key], dict) and isinstance(value, dict):
|
||||
result[key] = recursive_merge(result[key], value)
|
||||
elif isinstance(value, dict):
|
||||
result[key] = recursive_merge(value)
|
||||
else:
|
||||
result[key] = value
|
||||
return result
|
||||
Reference in New Issue
Block a user