What to keep
- Keep your current test or eval runner.
- Keep your current CI workflow.
- Keep your existing datasets and assertions.
What to add
- Install
@icodenet/eval-dashboards. - Write eval results into
.evals_output/*.jsonaseval-report/v1. - Run
eval-dashboards lint,check, andreport. - Publish the static report directory in CI.
Copy-paste path (simple)
pnpm add -D @icodenet/eval-dashboards
# optional starter scaffolding
eval-dashboards init --preset=agent-quality --write --playbook
# keep your existing runner command here
pnpm test
# eval-dashboards layer
eval-dashboards lint --input=.evals_output
eval-dashboards check --input=.evals_output --max-new-failures=0 --zero-critical --json-out=eval-report/check-result.json
eval-dashboards report --input=.evals_output --reporter=html --reporter=json-summary --report-dir=eval-report
If your runner is Promptfoo, DeepEval, or AgentEvals
eval-dashboards import --from=promptfoo --input=./promptfoo-results.json --out=.evals_output/import-promptfoo.json
# or
eval-dashboards import --from=deepeval --input=./deepeval-results.json --out=.evals_output/import-deepeval.json
# or
eval-dashboards import --from=agentevals --input=./agentevals-results.json --out=.evals_output/import-agentevals.json
# then run the same lint/check/report flow
eval-dashboards lint --input=.evals_output
eval-dashboards check --input=.evals_output --max-new-failures=0 --zero-critical
eval-dashboards report --input=.evals_output --reporter=html --report-dir=eval-report
Definition of done for adoption
- Your CI job fails on real regressions (not just test command failures).
- Your team can open one static report URL and see row-level evidence.
- You can compare current run vs baseline with stable row ids.
Adoption map (candidate repo targets)
See docs/adoption-map.md for three candidate existing-runner adoption targets and PR-style change sets. Not endorsed or accepted by those projects.