EvalGlass Core v0.2.2

[0.2.2] - 2026-08-20 (pre-alpha)

Plugin-storefront alignment — no runtime, public-contract, authority, or verdict-behaviour change:
the vendored runtime and all typed artifacts are behaviourally identical to 0.2.1.

Changed

• Plugin-facing docs align with the shipped capability (no false confidence, applied to our own
  storefront). Removed every reference that implied an "advanced discovery engine" — the
  Core executes · Discovery finds · Intelligence explains three-tier framing and the pointers to a
  separate (private, unreachable) evalglass-discovery repo — from the README, architecture doc,
  AGENTS.md, the plugin/marketplace descriptions, and the ADR index. What evalglass-core actually
  ships is a single, conservative, read-only candidate call-site inventory (installer discover);
  automated metric discovery (deriving what to measure from an app's traces, prompts, and schemas)
  is deliberately out of scope — EvalGlass runs the checks the host authors and derives none on
  its own. No runtime, public-contract, authority, or verdict-behaviour change.

EvalGlass Core v0.2.1

EvalGlass Core v0.2.1 — the final pre-public cut. A vendored, local-first, host-directed evaluation framework for agentic LLM apps, over an effect-free Evaluation Core and a single Verdict Engine, operated through your Claude Code / Codex agent. Governing rule: no false confidence — a green or non-failing result never implies more than the evidence supports.

This release is public-launch preparation. No runtime, public-contract, authority, or verdict-behaviour change since 0.2.0.

Added

• CodeQL static analysis (visibility-gated — activates automatically when the repository is made public).
• Contributor Covenant 2.1 Code of Conduct.

Removed

• The internal docs/plugin/ folder (acceptance runbooks, implementation-lessons log, marketplace notes, release checklist) — development/QA process material, not external docs. The release gates are enforced by the tests/plugin/ suite and summarised in CONTRIBUTING.md.

Fixed / Security

• Scrubbed a local path and internal tool versions; fixed broken doc links and a private-repo reference. Enabled repository secret scanning + push protection and Dependabot alerts.

Full changelog: CHANGELOG.md · Compare: v0.2.0...v0.2.1

Install — /plugin marketplace add EvalGlass/evalglass-core then /plugin install evalglass-core@evalglass.

EvalGlass Core v0.2.0

EvalGlass Core v0.2.0 — the pure, host-directed evaluation framework for agentic LLM apps: a vendored, local-first runtime over an effect-free Evaluation Core and a single Verdict Engine, operated through your Claude Code / Codex agent. Its governing rule is no false confidence — a green or non-failing result never implies more than the evidence supports, and a fresh run is informational by design.

This release refines packaging and positioning on top of the complete 0.1.0 core; the runtime, public contracts, authority, and verdict behaviour are unchanged.

Added

• Modular optional-dependency extras — lean by default, complete by composition. evalglass[traces] installs all three trace connectors (Langfuse + Phoenix + LangSmith); evalglass[all] the full optional surface. The grouped extras compose the granular ones by self-reference and pin no SDK directly, so the required tier stays PyYAML-only and SDK-free.

Changed

• Canonical Core positioning — Core executes · Discovery finds · Intelligence explains. Core is the open, host-directed evaluation runtime: you tell it what to evaluate and it executes honestly. Deriving what to measure (metric discovery) is a separate concern that lives in the evalglass-discovery repo.

Notes

• No runtime / public-contract / authority / verdict-behaviour change since 0.1.0.
• Full hermetic suite green on Python 3.12 and 3.13; version alignment, honesty audit, and the deletion-invariant all pass.

Full changelog: CHANGELOG.md · Compare: v0.1.0...v0.2.0

Install — /plugin marketplace add EvalGlass/evalglass-core then /plugin install evalglass-core@evalglass.

EvalGlass Core v0.1.0

EvalGlass Core v0.1.0 — initial public release of the pure evaluation framework: a small, vendored, local-first evaluation framework for LLM applications, delivered as a Claude Code / Codex plugin over an effect-free Evaluation Core with a single Verdict Engine.

Its governing rule is no false confidence: a green or non-failing result never implies more evidence, authority, comparability, calibration, or safety than the run actually has, and a fresh run is informational by design. EvalGlass is the framework, not the oracle — it supplies the machinery and lets you decide what to measure; automated metric discovery lives in the separate evalglass-discovery repo.

See CHANGELOG.md for the complete, itemised feature list (Evaluation Core, Runtime Harness, judges & rubrics, optional lanes, host-directed authoring, dual-runtime plugin, and repository trust tooling).