Current product intent after 1.3.1. Historical design notes live in plans/ and are not the backlog.
Bundled evals under evals/skills/*/basic/ are short task stubs. Example: evals/skills/create-service-object/basic/task.md is four bullets and has no fixture tree for the agent to edit. The engine's claim is that skill context beats baseline. These evals cannot measure that.
Do: turn a small set of those evals into fixture-backed tasks with task.md, criteria.json, and a starting codebase. Do not add providers in the same change.
Shipped: Clients::Providers::Xai, skill-bench init --xai.
Shipped: Clients::Providers::Bedrock (Runtime OpenAI-compatible path + Bedrock API key). IAM SigV4 is still deferred.
Shipped: golden parser fixture test/fixtures/judge_traces/canonical.json, Judge::Variance, and inter-run stats on the existing trend history (same eval+skill). Verdict math unchanged. IAM SigV4 for Bedrock is still deferred (item 3).