Find the commit that broke each GitHub Actions workflow

**Objective:** Diagnose the failed GitHub Actions runs in `{repository}` over the last `{lookback-days}` days (default 7), optionally only for the workflow named `{workflow-name}`, and identify where each failure started.

**Step 1: Find the failures**

Call `github-get-workflow-run` with company `{repository}`, conclusion `failure`, created-date.from set to today minus the lookback, name set to `{workflow-name}` when it is given, and max-records 100. Group the results by workflow name and branch.

**Step 2: Find where each failure started**

For each workflow and branch pair, call `github-get-workflow-run` again with the same name and branch and no conclusion filter, over the same window. Order by run-number and find the last run with conclusion `success` before the first failure. The run straight after it is the breaking run.

* If the latest run on that branch succeeded, the failure is already fixed; say so and deprioritise it.
* If failures alternate with successes on the same sha, treat it as a flaky workflow, not a code break.

**Step 3: Identify the breaking change**

For the breaking run, take its sha and call `github-get-commit` with sha set to it to get the author-name, author-email and full message. Also call `github-get-workflow` with name set to the workflow to report its path and whether its status is still active.

**Step 4: Report**

For each workflow and branch still failing:

* workflow name, path, and branch
* failing run count, first failing run-number and its url
* the breaking commit: sha, author, message, commit-date
* verdict: broken by a commit, flaky, or already fixed

Finish with the list of workflows that need a person to look at them now.

**Rules & Guidelines:**

* This prompt is read-only. Do NOT call `github-rerun-workflow-run`, `github-cancel-workflow-run`, `github-enable-workflow`, `github-disable-workflow`, `github-add-issue` or `github-add-issue-comment`.
* The tools do not return job logs. Do not invent an error message; say that the log must be opened from the run url.

**Validation:**

* If no failed runs are returned, report that there were no failures in the window and stop.
* If there is no successful run in the window before the first failure, say the break is older than the window and suggest a longer `{lookback-days}`.