add first step logging to take step #2927
Conversation
Agent Task Evaluation Results: 1/3 (33%)View detailed results
Check the evaluate-tasks job for detailed task execution logs. |
| # First step | ||
| self._log_first_step_startup() | ||
| await self._execute_initial_actions() | ||
|
|
There was a problem hiding this comment.
Bug: Redundant Logging and Actions on First Step
The startup message logs multiple times, and initial actions can run redundantly. This is because _log_first_step_startup() and _execute_initial_actions() are called in run(), take_step(), and step(), all using len(self.history.history) == 0 to identify the first step. As history is only populated after a step completes, this check is true in multiple places during initial execution.
Additional Locations (1)
| self.step_start_time = time.time() | ||
|
|
||
| # Show startup message on first step | ||
| self._log_first_step_startup() |
There was a problem hiding this comment.
Startup message will be logged twice on the first step: here and again in step(); remove this call or guard to avoid duplicate logs.
Prompt for AI agents
Address the following comment on browser_use/agent/service.py at line 620:
<comment>Startup message will be logged twice on the first step: here and again in step(); remove this call or guard to avoid duplicate logs.</comment>
<file context>
@@ -617,6 +616,9 @@ async def step(self, step_info: AgentStepInfo | None = None) -> None:
self.step_start_time = time.time()
+ # Show startup message on first step
+ self._log_first_step_startup()
+
browser_state_summary = None
</file context>
| input_metadata = await event.event_result(raise_if_any=True, raise_if_none=False) | ||
| msg = f"Input '{params.text}' into element {params.index}." | ||
| logger.info(msg) | ||
| logger.debug(msg) |
There was a problem hiding this comment.
Logging raw input text to debug can expose sensitive data; redact or avoid logging the actual text (e.g., respect has_sensitive_data) to prevent leaks.
(Based on your team's feedback about matching the project's tab-based indentation, the fix suggestion preserves tabs.)
Prompt for AI agents
Address the following comment on browser_use/tools/service.py at line 328:
<comment>Logging raw input text to debug can expose sensitive data; redact or avoid logging the actual text (e.g., respect has_sensitive_data) to prevent leaks.
(Based on your team's feedback about matching the project's tab-based indentation, the fix suggestion preserves tabs.)</comment>
<file context>
@@ -325,7 +325,7 @@ async def input_text(params: InputTextAction, browser_session: BrowserSession, h
input_metadata = await event.event_result(raise_if_any=True, raise_if_none=False)
msg = f"Input '{params.text}' into element {params.index}."
- logger.info(msg)
+ logger.debug(msg)
# Include input coordinates in metadata if available
</file context>
| logger.debug(msg) | |
| logger.debug(f"Input '{'[REDACTED]' if has_sensitive_data else params.text}' into element {params.index}.") |
| if len(self.history.history) == 0: | ||
| # First step | ||
| self._log_first_step_startup() | ||
| await self._execute_initial_actions() |
There was a problem hiding this comment.
First-step handling is duplicated across run(), take_step(), and step() using the history-length check, which can trigger the startup log and initial actions multiple times. Consolidate this into a single entry point or guard with a one-time flag to ensure first-step initialization runs exactly once.
Prompt for AI agents
Address the following comment on browser_use/agent/service.py at line 1135:
<comment>First-step handling is duplicated across run(), take_step(), and step() using the history-length check, which can trigger the startup log and initial actions multiple times. Consolidate this into a single entry point or guard with a one-time flag to ensure first-step initialization runs exactly once.</comment>
<file context>
@@ -1122,6 +1129,11 @@ async def take_step(self, step_info: AgentStepInfo | None = None) -> tuple[bool,
+ if len(self.history.history) == 0:
+ # First step
+ self._log_first_step_startup()
+ await self._execute_initial_actions()
+
await self.step(step_info)
</file context>
Auto-generated PR for: add first step logging to take step
Summary by cubic
Log the startup message only when the agent takes its first step, instead of during initialization, to ensure a single, accurate "starting" log. Also reduce log noise by changing input_text logs from info to debug.