Feature Request: Expose step-by-step action/reasoning data in ActResult

[eran-tookii]

Problem
The Nova Act SDK currently returns an ActResult with high-level metadata (num_steps_executed, time_worked_s, etc.), but does not expose the detailed step-by-step execution data that appears in console logs during execution.

Console output shows valuable structured data like:

👀 ...
💭 ...
think("The cookie popup is now closed...")
agentClick("43,695,62,743")
This data is currently not accessible programmatically through the Python SDK.

Requested Feature
Expose step-level data in ActResult, for example:

result = nova.act("Navigate to pricing page")

for step in result.steps: # New property
print(step.observation) # What the agent saw
print(step.thinking) # Agent reasoning (think())
print(step.action) # Action taken (agentClick, agentType, etc.)
print(step.timestamp) # When it occurred
Use Cases
Debugging - Understanding why an automation failed at a specific step
Audit trails - Recording what actions were taken for compliance
Analytics - Analyzing agent behavior patterns
Building guides/tutorials - Turning automation runs into step-by-step documentation
Quality assurance - Validating agent reasoning matches expected behavior
Current Workarounds
Parsing stdout (fragile, format not guaranteed)
Manually checking logs_directory output files
Using the observability console (not programmatic)
Additional Context
The data appears to already exist internally - it would be valuable to expose it through the SDK's public API.

[ebaizel]

hey @eran-tookii Act records the entire trajectory locally in a json file. This contains artifacts including the prompts, think statements, and screenshots base64-encoded.

Is this what you are looking for?

[the-rpc]

Hi @eran-tookii, the latest SDK release (v3.3.35.0) should provide you with this functionality.

Set replayable=True in the NovaAct constructor and you get a trajectory_file_path on ActResult pointing to a structured JSON file with the step-by-step data you're looking for. The SDK also ships a Trajectory Pydantic model you can use to parse it:

from nova_act import NovaAct
from nova_act.impl.trajectory.types import Trajectory

with NovaAct(starting_page="https://example.com", replayable=True) as nova:
    result = nova.act("Click learn more")

    trajectory = Trajectory.model_validate_json(open(result.trajectory_file_path).read())

    for i, step in enumerate(trajectory.steps):
        print(f"Step {i+1}: {step.active_url}")
        for call in step.program.calls:
            print(f"  {call.name}({call.kwargs})")

See the Trajectory type definitions for more details on all accessible properties.

Closing this as the functionality is available today. If having it as an in-memory object directly on ActResult (rather than via file) is important, feel free to file a follow-up feature request. Thanks!