Skip to content

Tell the model that [REDACTED] was inserted by marginal #178

Open

Description

@exactml

Context

On #176, a doc explained redaction with the example "https://user:s3cr3t@host becomes https://[REDACTED]@host". marginal redacted the example's user:s3cr3t@ before the model saw it. So the model read "https://[REDACTED]@host becomes https://[REDACTED]@host" and flagged the example as a no-op (Medium, 85%), a false positive created by marginal itself.

The same will happen to any docs, tests or fixtures that show a secret-shaped example: the model sees [REDACTED] and has no way to know that marginal put it there.

Behavior to implement

When redaction changed anything in the prompt, tell the model so: [REDACTED] marks a value marginal removed before review, the original text had a real-looking value there, and the placeholder itself must never be flagged or quoted as the code's content.

Acceptance criteria

  • With a redacted value in the diff, the prompt contains the note; without one, the prompt is unchanged.
  • An eval case: a diff documenting a redaction example gets no "no-op" or "placeholder" finding.

Activity

  1. added
    bugSomething isn't working
    good first issueIssues suitable for contributors who are new to the our codebase
    on Oct 25, 2026
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

    backendbackend devbugSomething isn't workinggood first issueIssues suitable for contributors who are new to the our codebase

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions