Skip to content

v0.34.0

Choose a tag to compare

@github-actions github-actions released this 07 Jun 15:31
· 72 commits to main since this release
db42c1b

⚠️ Breaking changes

Three flagged breaking changes in this release. Full detail in the linked PRs.

BREAKING (CLI) — name@cwd variant syntax removed (#186)

The name@cwd string syntax for variant expressions is removed. Runtime context (cwd) is now declared structurally, decoupling artifact identity from runtime context (this also fixes the @-in-git-ref / @-in-path parsing bugs the old encoding caused).

# old
omk eval --control project-env@/proj --treatment ./skill.md@/proj
# new
omk eval --control project-env --control-cwd /proj --treatment ./skill.md --treatment-cwd /proj
  • CLI: use --control-cwd <dir> and --treatment-cwd <dir,...> (comma-separated, aligned by position with --treatment; an empty slot = no cwd for that treatment).
  • eval.yaml: the structural per-variant cwd: field (unchanged).
  • Hitting the old name@cwd form raises a migration error pointing at the new flags.
  • Not a comparability change — report-variant-key semantics are byte-identical; cwd only moved from string-encoded to structured.

BREAKING-SERIALIZATION — kind discriminator fields renamed (#210)

Persisted JSON discriminator fields were converged from bare kind to qualified names. Product behavior (running evals, viewing reports, doctor/observe) is unchanged; scripts/tools that read omk JSON or pin old report files are affected.

  • report.kindreport.reportKind (EvaluationReport / BatchEvaluationReport / DoctorReport, plus observe/inbox documents).
  • runtime / soft-standard / runtime-node discriminators → runtimeKind / standardKind / nodeKind.
  • DOCTOR_REPORT_SCHEMA_VERSION bumped to 2.0.0 — update any consumer that pins it.
  • Old report / observe / doctor JSON is not guaranteed to deserialize on the new version. Regenerate reports, or update your reader to the new field names; no general schema migration is provided.
  • Not a comparability change — judge prompts, the five-layer scoring pipeline, bootstrap CI, Krippendorff α, length-debias, and verdict computation are untouched. The same eval, re-generated, produces the same numbers.

BREAKING-COMPARABILITY — git-variant runtime fingerprint fix (#213)

Fixes a pre-existing Δ-contamination bug: a git: variant's runtime environment (and its recorded runtime fingerprint) was polluted by the run cwd's node_modules/.bin, asymmetric with baseline.

  • Affects only git: variants, and only when eval runs from a cwd containing node_modules/.bin. A clean cwd is unchanged.
  • The affected old fingerprint values were contaminated (baseline vs git diverged, not comparable); the fix restores within-run comparability. Cross-version, a git variant's fingerprint value and cache key change, so old git-variant cache entries re-compute.
  • No frozen invariant is touched (judge/observe prompt hashes, five-layer scoring, bootstrap/α, report schema field semantics).

What's Changed

  • refactor(eval)!: 移除 name@cwd,cwd 全程结构化(BREAKING CLI) by @lizhiyao in #186
  • fix(eval-workflows): batch 修 treatment 角色误标 control(#183) by @lizhiyao in #187
  • chore: 清理硬编码 model ID + 移除 sensitive-scan 文档 by @chenfengjw163 in #188
  • refactor(eval-workflows): 退役 runEvaluation 的 artifacts 注入(收成 spec 单一来源,#189) by @lizhiyao in #190
  • feat(authoring): evolve 加显著性接受门 + 锁定 test 集 + 抗过拟 by @lizhiyao in #191
  • docs(site): 用 VitePress 搭文档站,部署到 Cloudflare Pages by @lizhiyao in #192
  • fix(site): 锁 yarn 1 classic + 跳过 engine 校验,修 Cloudflare Pages 构建 by @lizhiyao in #193
  • docs(site): docs 全量双语化(教科书式 i18n)+ 删内部文档 + 拆 sample-design by @lizhiyao in #195
  • docs(site): docs 站系统性提升 —— executors 拆分 + 术语统一 + 准确性修复 + 覆盖缺口补全 by @lizhiyao in #196
  • docs(site): docs 站逐页精修 —— 准确性修复 + 导航/文案可读性提升 by @lizhiyao in #197
  • chore(assets): 按用途归整 assets 目录 + 加索引 README by @lizhiyao in #198
  • docs(site): sample-design 设计依据并入 spec 附录,对用户全公开 by @lizhiyao in #199
  • chore(release): 发布包不再带 source/declaration map,瘦身约 40% by @lizhiyao in #200
  • fix(observability): 修 12 处 test 类型漂移,恢复 typecheck 绿 by @lizhiyao in #201
  • docs(site): README/包指向文档站 + omk skill 单一来源迁到中性路径 .agents/skills/omk by @lizhiyao in #202
  • chore(release): 迁移到 Yarn 4(Berry),退役 yarn 1 经典 + ignore-engines 创可贴 by @lizhiyao in #204
  • chore(release): .yarnrc.yml 注释固化 enableScripts 安全默认 by @lizhiyao in #205
  • feat(cli): 支持安装内置 Agent Skill by @lizhiyao in #207
  • refactor(types)!: BREAKING-SERIALIZATION 收敛 kind 判别字段 by @lizhiyao in #210
  • feat(cli): install 补完整为证据门控管理入口 + managed 记录核心 by @lizhiyao in #211
  • feat(cli): install 支持 git 源 + source-resolver 抽象 by @lizhiyao in #212
  • fix(eval-core)!: git variant runtime fingerprint 消除 node_modules 污染 [BREAKING-COMPARABILITY] by @lizhiyao in #213
  • docs(cli): install 参考补 git/本地源 + 受管记录;修两处半角标点 by @lizhiyao in #215
  • chore(release): 发布 0.34.0 by @lizhiyao in #216

Full Changelog: v0.33.0...v0.34.0