Skip to main content
The CI gate replays your eval set against every pull request that touches your agent, and reports a Mirrors eval gate status check on the PR. Make the check required and regressions block the merge — before they reach a customer.

​
How it works

On each PR against your base branch, Mirrors fetches your agent’s code at the configured path, runs the bound eval set against your mirror with the candidate agent, and diffs the results against the baseline (the current agent). The check lands as in-progress → awaiting review, with a candidate-vs-baseline diff in the dashboard (/ci/<check-id>) where you approve or reject — flipping the GitHub check.

​
Set it up

1

Connect GitHub

In the dashboard, go to Quickstart → CI eval gate and click Connect GitHub. This installs the Mirrors GitHub App; GitHub redirects you back with the installation already linked to your workspace.
2

Bind a repo

Pick the repository and bind it to:
  • a mirror (the twin to replay against),
  • a validation eval set (what runs on each PR),
  • the agent path in the repo (e.g. agent/agent.py) — the webhook fetches this file on each PR,
  • the base branch (default main).
3

Make it required

Click Make required on the repo row to turn Mirrors eval gate into a required status check (needs the App’s Administration permission), or add it manually under the repo’s Settings → Branches. Merges now block until the gate passes review.

​
Reviewing a check

Each check links to a dashboard page showing the candidate-vs-baseline diff per eval case. Approve marks the check green; Reject fails it. A regression is a decision, not a surprise.
An installation can only be linked to one workspace. If you see “already linked to another workspace”, the GitHub account was claimed elsewhere — connect from that workspace or use a different GitHub org.