Cardinal — review changes and consolidate findings
Select the operation
For an orchestrated review, use core guidance and only the relevant role prompt: Rochefort for missed audit findings, Milady for regressions, Comte de Wardes for over-optimization, or Cardinal for coordination. Delegation, model calls and tests depend on the host and the existing task scope; the confrontation script itself does not launch agents, execute suggested commands, repair code or contact model providers.
To consolidate reports already produced:
python skills/cardinal/scripts/cardinal_confront.py <blue_reports_dir> <red_reports_dir>
| Required file | Minimum input accepted by the script |
|---|---|
Blue audit/audit-report.json |
Non-empty JSON object. |
Blue fix-report.json |
Non-empty JSON object. |
Blue optimize/optimization-plan.json |
Non-empty JSON object. |
Red counter-audit.json |
false_negatives and underestimated: lists of objects. |
Red counter-fix.json |
regressions and incomplete_fixes (or prompt alias incomplete): lists of objects. |
Red counter-optim.json |
over_optimizations and wrongly_excluded_skills (or prompt alias wrongly_excluded): lists of objects. |
If both aliases are supplied, their values must agree. Explicit empty finding
lists are accepted; do not invent an empty review to represent a stage that
did not run. A missing/malformed report or required findings field returns exit
code 2, with no new verdict or report write. An earlier confrontation.json
is preserved and must not be mistaken for this failed run's output.
Consequences, evidence and reuse
Read effects.json. A complete input set produces stdout and
replaces <red_reports_dir>/confrontation.json. Preserve a useful prior result
before rerunning. There is no multi-file snapshot, output transaction or lock.
Reports can contain source paths and finding descriptions; review their audience.
The score is a weighted count of six supplied finding categories. Blue reports receive only a presence/object check; the script does not compare their claims against source, authenticate reviewers or prove that tests ran. Other counter-report categories are not included in the score. A zero exit means consolidation completed, including for an unfavorable verdict; it is not a quality gate.
Use evidence from the reviewed checkout, actual tests and counter-review coverage before acting on the summary. On failure, reconcile inputs and any existing output before retrying. Reuse in another pipeline needs explicit field mapping, omission handling and target-specific acceptance criteria. Keep predictions, observations and unverified claims separate in the task report; see the common contract.
