jameskomo/config-drift-checker

CI for your coding-agent setup: evals on every PR and Claude Code release, pinned baselines, self-diagnosing reports

View on GitHub

Trust Signals

Scorecard Score
not yet scored
Maintenance Recency
Stale
License
None
namedescriptionrequireddefault
plugin-dirPath to the plugin (contains .claude-plugin/plugin.json, evals/ and optionally .cdc.yml)yes—
trackpinned (baseline, deterministic) or canary (alias model, latest Claude Code). Empty = .cdc.yml `track`—""
runsRuns per case (overrides case default and .cdc.yml)—""
modelAgent model alias or id (overrides the track)—""
judge-modelLLM grader model (overrides .cdc.yml judge_model)—""
ablationwith-without (baseline proof) or none (cheaper regression runs)—none
scaffoldRun case.yaml scaffold_script (only for suites you authored)—true
thresholdScore drop that counts as a regression (overrides .cdc.yml thresholds.score)—""
coverage-minFail the check when agent-config coverage (rules with an eval case) is under this percent. Empty = report only—""
report-base-urlPublic base URL of the results-branch Pages history folder (e.g. https://<you>.github.io/<repo>/history). When set, PR-comment case names deep-link into this run's HTML report—""
claude-code-versionClaude Code version to install (overrides the track's harness pin), or "latest"—""
results-branchOrphan branch that stores baseline.json, canary/, history/, spend.json and the release-watch state—eval-results
promote-baselineOverwrite baseline.json with this run (first run, or after an intentional change). Pinned track only—false
forceSkip the budget and interval gates (manual dispatch only)—false
open-prsOpen a bump PR after canary.promote_after green canaries, and a pin PR when the pinned track has no pin—true
repairOn a red result, let the agent propose a config fix, re-run the failing cases and open a repair PR (spends API credit, budget-capped)—false
slack-webhook-urlIncoming-webhook URL to notify on regression (pass a secret)—""
github-tokenToken for pushing results, opening PRs and commenting—${{ github.token }}
pagesAlso write docs/ (dashboard, report, latest.json) to the results branch for GitHub Pages (source: results branch, /docs)—true
preflightStatic checks before any model run: the skill linter and the suite format doctor (including a free load check on the installed Claude Code). warn = report in the job summary; fail = stop before spending model runs; off—warn
namedescription
regressedNumber of regressed/missing cases
redNumber of cases that turned the check red (regressions + fail_on efficiency drifts)
flaggedNumber of cases with efficiency drift (warning or red)
trackpinned or canary
skippedtrue when the budget or interval gate skipped the run
skip-reasonbudget | interval | (empty)
spent-monthUSD spent this month on this suite after this run (from the ledger)
runnerofficial or shim
claude-versionClaude Code version that was exercised
result-pathPath to aggregate-result.json for this run
pr-kindbump | pin | repair | none — which PR this run opened
pr-urlURL of the PR opened by this run, if any