evo

Evolution Engine occupies a specific niche: it is a drift detector, not a linter, not a scanner, not a monitoring platform. That framing matters because the tool's value depends on whether you have a drift problem worth measuring. What it does well. The core mechanism — computing modified z-scores against your repository's own baseline, calibrated across 48 open-source repos — means a finding is a statistical deviation from how your codebase normally behaves, not an arbitrary rule firing. Git metrics like change_locality, dispersion, and cochange_novelty are the ones that catch a large refactor quietly concentrating change into a few files, or unrelated modules starting to be edited together. Dependency metrics (dependency_count, max_depth) across five package managers are equally concrete. The cross-signal promise is real where the adapters exist. "CI fails after big dependency updates" is a pattern that requires git, dependency, and CI data in one place; GitHub Actions, GitLab CI, and CircleCI adapters supply run_duration and run_failed so the correlation can be computed. Pro also ingests JUnit XML (failure_rate, skip_rate), Cobertura XML (line_rate, branch_rate), Sentry (event_count, is_unhandled), and Dependabot (vulnerability_count, severity_score). Where it is honest about itself. The tool does not call any AI. It generates an investigation prompt that you paste into your own AI tool, then `evo enrich . --from response.txt` imports the response as friendly descriptions on findings. That is a deliberate local-first design, not a missing feature — but it does mean evo is not a one-click explain-and-fix agent. You supply the model and the judgement. Privacy posture is unusually specific and verifiable: source-available on GitHub, no code uploads, no repository access required, telemetry is command name and duration only and is opt-out, license checks use a SHA-256 hash of your email rather than the email itself, and community patterns sync from PyPI. CI and deployment adapters use read-only tokens you provide, which the vendor states are never stored. Where it does not fit. There is no always-on dashboard, no real-time alerting, and no cloud-hosted analysis — this is a CLI you run, optionally wired into git hooks (post-commit or pre-push, non-blocking by default) or a GitHub Action that comments risk badges on pull requests. Organizations that need a hosted control plane, ticketing integrations, or dedicated vendor support will need something else. Monitoring (Datadog, New Relic, Grafana), incidents (PagerDuty, OpsGenie), and work items (Jira, Linear) are listed as Planned adapters, not shipped ones — do not budget for them yet. The extensibility story is a genuine differentiator. `evo adapter new <name> --family ci` scaffolds a custom adapter you can ship as a pip package, and the vendor prioritizes adapter requests by community demand. Verdict: adopt it as a complementary instrument alongside your linter and your APM, not as a