v0.3.0

The first eval-verified release. Behavior evals now run for real (isolated workspaces, tool-enabled arms, adjudicated judging), and every skill change ships with the run that motivated or confirmed it: snapshot handles cross-machine handoffs and reports before resuming, dig never skips its issue.md on repeated-failure entries, shaping treats candidates as evidence for recutting the frame, and rules/ keeps only what no skill owns. Full details in CHANGELOG.md.

v0.2.0

The four-skill working loop: shaping, coding, dig, snapshot. Breaking: spec-out, deep-thought, muse, test-first, best-practices, inversion, first-principles and figure-out are gone, their jobs absorbed by the four.

v0.1.0

Nine skills for Claude Code and Codex: socrates, muse, first-principles, inversion, figure-out, clarify, test-first, best-practices, snapshot.

• English-first skill bodies and references, with zh-CN translations alongside. Descriptions stay bilingual as the trigger surface.
• Dual-judge trigger evals and behavior evals across the family.
• Style, contract, and privacy gates wired into CI.

Install: npx skills add liustack/liustack -g