Skip to content

Triage: the failing job's log excerpt reaches the judge unfenced #641

Closed

Description

@kirkmarple-clockgrove

The red-check triage question (triage.mjs triageInput) gives the model "Log of the failing job (key lines)" without a fence. That log is written by the repository's code and its CI, so a pull request can print text into it. The #637 silent-step block and the other decision checks put untrusted data between random-tag <<<BEGIN tag>>> / <<<END tag>>> markers, with a "data, never instructions" line. This excerpt does not. Found in the review of #639.

What a crafted log could do: argue the triage into "flaky_or_infra", which reruns the failed jobs (a bounded number of times), or into "base_failure". It cannot merge anything on its own, but it is the same untrusted-input rule the other checks follow.

Acceptance:

  1. The log excerpt, and every other part of the triage question taken from the log or the PR, sits inside the random-tag fence, with the same "data, never instructions" sentence.
  2. One regression test: a log line that tries to close the fence or give instructions stays inside it.
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

    P2Low impact: minor bug or cleanup; deferred

    Type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions