Skip to content

Code-executing WTI agent can exhaust output budget and return empty response #180

Description

@mbrotos

Summary

build_wti_code_exec_config() can exhaust its response budget while loading skills/resources. The model then ends with STOP and no content, and AgentPredictor attempts json.loads(""), producing a misleading JSONDecodeError.

Reproduction

  1. Open implementations/energy_oil_forecasting/02_intro_agentic_predictor.ipynb.
  2. Run the Step 4 code-executing agent with the default lite model.
  3. The forecast call can fail with JSONDecodeError: Expecting value: line 1 column 1.

A trace showed 16,382 output tokens consumed against a 16,384 limit, with an empty final STOP response.

Expected

The agent should have enough budget to complete tool use and return structured forecast JSON; empty responses should not be parsed as JSON.

Possible fix

Raise the code-agent response budget (32,768 resolved this run) and handle empty runner output with a descriptive agent error.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions