Quality baseline
The checks that every project in this family runs, and where to copy them from. The files are in baseline/; linked-codebook itself is the strictest example.
| Concern | Baseline | Where |
|---|---|---|
| One entry point | make check = lint, types, tests, size; CI runs exactly that | baseline/Makefile |
| Python lint and dead code | ruff, pinned to one minor version; unused names, unused arguments, commented-out code | baseline/ruff.toml |
| Python types | mypy --strict on the cleaned modules; ty as an advisory job until it leaves beta | Makefile |
| TypeScript lint and dead code | eslint flat config with complexity, depth, function length and --max-warnings; tsc with noUnusedLocals and noUnusedParameters | baseline/eslint.config.js, baseline/tsconfig.base.json |
| Size and complexity | per-module statement budgets that only go down | scripts/budget.py, scripts/check_size.py in semsynth |
| Versions | one script fails when README, docs and packages disagree | scripts/check_versions.py |
| Dependencies | a lock file in CI (uv sync --locked, npm ci); monthly grouped Dependabot; a weekly job that installs the newest of everything | semsynth .github/workflows/canary.yml |
| Agent notes | AGENTS.md of at most 30 lines; CLAUDE.md containing @AGENTS.md | baseline/AGENTS.template.md |
Two tiers keep this affordable. New code takes the strict settings. Legacy code takes the curated settings, with today's violations recorded as an ignore list or a warning count that may only shrink.
Left out on purpose: dead-code scanners such as vulture and knip, and mutation testing. ruff's unused-name and unused-argument rules and tsc's unused flags find most of what they would, and the rest is found by running vulture once by hand on a legacy project to seed a deletion list.
Pins to linked-codebook are the one dependency Dependabot cannot follow, because they are URLs. The semsynth canary job compares its pin with the latest release and warns; copy that step into other projects that depend on a release.