Skip to main content
A correction supplies guidance for a particular situation. In parts of the dashboard it is called an override. Start with a concrete failure and describe both what should change and when the guidance should apply. You or your coding agent author the guidance. Mirrors retains the evidence, preview, publication controls, and runtime delivery.

​
From evidence to activation

  1. Inspect the original conversations and tool results behind the failure.
  2. Write a focused correction and scope its trigger to the relevant situation.
  3. Save the draft and preview that version against recorded runs.
  4. Review matching runs, exclusions, and conflicts. Include nearby situations where the correction should not apply.
  5. Choose Activate after reviewing the current version’s preview. Activation still checks for conflicts.
Changing a draft invalidates its earlier preview. Run a fresh preview before activating the changed version.

​
What preview tells you

Preview checks where the trigger matches stored evidence. It helps you inspect how broadly a correction could apply. It does not run a controlled experiment showing that the guidance improves the model’s answers. After activation, evaluate actual agent behavior on representative cases. Keep successful context delivery, prompt application, and task success as separate measurements. Use runtime controls to limit delivery while validating the change.

​
Keep guidance narrow

For example, a correction might tell an agent how to interpret a particular tool’s missing status field. Scope it to that tool and situation instead of adding an instruction that changes every conversation. Check examples where the field is present as well as examples where it is absent. Once the integration is deployed, activating eligible guidance can affect later context requests without an application redeploy. Context already returned to your application cannot be recalled; use fresh results at the intended turn.