Changelog

No tagged releases yet next

EvalGlass is pre-alpha. There are no version tags yet; the current version is 0.1.0, and nothing has been cut as a release. The work below has merged to main — the honest record of what changed, not a tagged release.

Unreleased — landed on main

EvalGlass was repackaged from a single setup skill into a Claude Code plugin, with two framework follow-ups. This is delivery and packaging — EvalGlass itself (the Evaluation Core, the single Verdict Engine, the vendoring boundary, typed authority, no false confidence) is unchanged.

• Claude Code plugin (0.1.0). A publisher marketplace and a two-command install; the /evalglass umbrella (setup, connect, run, view, explain, compare, baseline, ci) with an always-on honesty guardrail and a natural-language router; a bundled launcher so a marketplace-only user needs no pip install. The direct CLI path is unchanged and still supported. ADR 0022 →
• Authoring tier + advanced extensions (v1.1). add-metric, add-judge, calibrate plus the backing skills. Everything scaffolds proposed/uncalibrated; gate activation stays a host YAML edit with no verb. connect --live is a deletable opt-in lane (no provider SDK); connect --synth is governance-only (no generator). ADR 0025 →
• Score subject identity + view --by-call. Score now carries optional example_id/unit_id (additive provenance; the engine stamps it; old artifacts still parse), so view can group by explicit per-call identity — never by list order. Mapping a score to its source function remains an unbuilt extension. ADR 0024 →
• Codex second runtime next Packaged from the same repo and the same canonical skills tree — thin per-runtime manifests, a deterministic sync, runtime independence proven across runtimes. A public Codex marketplace listing remains a maintainer step. ADR 0023 →
• Installer package rename. The integration package evalglass.skill was renamed to evalglass.installer and the legacy in-package SKILL.md removed — the direct CLI is now python -m evalglass.installer install --root .. Pre-release, so it's a clean rename, not a break; hosts run the vendored _evalglass and are unaffected. ADR 0026 →
• Install command renamed. The marketplace catalog is now named after the publisher, so the install reads /plugin install evalglass@syntelesis-lab (was evalglass@evalglass) — evalglass is the plugin, syntelesis-lab the catalog. marketplace add Syntelesis-Lab/evalglass (the repo slug) is unchanged; no host-facing impact. ADR 0027 →

What you'll find here after 0.1

Once releases exist, each entry will carry the things a host repo needs to upgrade with confidence.

• Version and date. The tagged version and when it was cut, traceable back to a specific commit.
• Trust-boundary changes. Any change to the verdict, authority, or the Scorecard and RunRecord schema — called out explicitly, because those are the load-bearing contracts.
• Re-vendoring guidance. What host repos need to do to move managed files forward safely, and what stays host-owned. Re-vendor safely →

GitHub Releases will be the canonical source.