[scarletkc]

Why

Rust has rustfmt in CI. The Python tools in core/tools and tools have no formatter, even though CI and agents rely on them: CI test selection, guards, cleanup, worktrees and the production LOC counter. Each author's and each agent's style landed as written, which costs something in review:

• Diffs hide the real change. A fix in a long, dense line rewrites the whole line, so reviewers and git blame point at more than what changed.
• Parallel branches conflict on lines nobody meant to touch. This repository runs many worktrees and agents at once.
• Style keeps coming back up. Without a config, each contributor has to guess the house style.

With one pinned formatter, a diff shows only the change, and style is settled by a single command.

Why it is safe

• No behavior change.
  • Every changed file parses to the same AST as before. Docstring indentation is the only allowed difference.
  • Production LOC is unchanged at 102160, because development tools are outside every counted scope.
• Reproducible. Check out commit 1 and run pipx run --spec ruff==0.17.2 ruff format. The result should be commit 2 exactly.
• Follows AGENTS.md §4. The check only reports: it is not a required check, nothing runs before a push, and there is no hook.
• Narrow scope. It touches core/tools and tools only. Docs, evidence trees and core/vendor/ are untouched.
• Cheap CI. One job of about 15 seconds, which runs only when those files or the config change.

What is in it

1. Config and check (cc1d509d)
   • ruff.toml: Python 3.11 (the code uses tomllib), width 120.
   • .github/workflows/format.yml: runs ruff format --diff. It does not check lint findings, so it starts green.
   • Two short additions to AGENTS.md §4.
2. One format pass (69d9cdc): 31 files, all generated by Ruff, with no hand edits.

Notes

• Rebuilt on current main. This PR now covers only the Python tools. The JavaScript/TypeScript part moved with dsh/ and is now Ephemeral-AI-Lab/dsh-layerfs#2.
• Not run: tests were not run locally. The existing ci.yml runs the core/tools unit tests on this PR.
• Reviewing: read commit 1, and check commit 2 by regenerating it.

[jiangkoumo]

在本地复现过，结论：可以合并。

复现（Debian 13 x86_64，node 22.23.3）：
• 在 52dd148 上跑 npx @biomejs/biome@2.5.15 format --write .（dsh 下两遍，第一遍 183 个、第二遍 2 个），再跑 ruff 0.17.2 format（32 个），结果和 f3d2d88 只差一个文件：dsh/sim/avatars/build.py。
• 差别：f3d2d88ca 里 VARIANT_POSES 和 ELLEN_POSES 两个列表拆成了每行一项（看起来是按默认 88 宽格式化的），按 ruff.toml 的 120 宽重跑会保留在一行。因为有尾逗号，64fcd0631 上 ruff format --diff 仍然判定已格式化，不影响新的 CI 检查，只是和 PR 里“结果应完全等于 commit 2”的说法差这一处。顺手重新生成一下就能对齐。
• 64fcd06 上 biome format 和 ruff format --diff 都干净；32 个 Python 文件格式化前后 AST 全部相同。

测试（base 8e2fce2 对比 head 64fcd06，结果完全一致）：
• dsh npm test：两边都是 16/17，失败的那项是 REBUILD.md §4 在我的 worktree 里找不到 core/docs/issues/352/fixes，属于环境问题，和这个 PR 无关。
• core/tools 和 tools 下 9 个 Python 测试文件：两边都是 8 个 OK，test_ci_tests 两边都因离线 cargo metadata 失败。

关于 #427：它是长期 Draft，不用等它，这个 PR 先合就行。我试合并过，#427 会有 8 个文件冲突（check_product_boundary.py、它的测试，以及 dsh/packages/host/src 下 6 个文件），我们同步 main 时用同版本 biome/ruff 重新格式化即可解决。

另外同意加 .git-blame-ignore-revs 收录格式化那个提交。