Skip to content

G2A CSV-artifact journey flakes ~13% — model sometimes skips artifact creation #403

Description

@0x-copilot-dev

tools/desktop-journeys/generative-workflows/g2a_csv_artifact_surface.py fails intermittently because the model does not call the artifact tool at all.

Measured over 15 consecutive real runs against OpenAI (2026-07-27): 13 passed, 2 failed — runs 6 and 13, both with the identical assertion:

File "g2_csv_lifecycle.py", line 578, in _dataset_artifact_from_run
    assert artifacts, "agent did not create a dataset artifact"
AssertionError: agent did not create a dataset artifact

No run errors, no provider errors — the run completes, it just produces no dataset artifact. This is model non-determinism, not a product defect.

Why it matters: this journey is a verification harness for artifact work. A ~13% false-red rate means roughly one run in seven fails for reasons unrelated to the code under test, which erodes trust in the signal and encourages re-running until green — exactly the habit that hides real regressions.

Options:

  • Retry once on this specific assertion before failing (cheapest; keeps the signal honest about repeated non-creation).
  • Firm up CREATE_PROMPT so artifact creation is unambiguous.
  • Assert on a terminal run state first, so the failure distinguishes "model chose not to" from "artifact pipeline broken" — currently both surface identically.

Raw data: 15/15 runs also showed zero occurrences of the empty-call_id defect tracked in #402, so this flake is the journey's dominant failure mode today.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions