Pre-PR Checks

Run all PR-readiness checks in one pass. Designed to be spawned as a background-or-foreground subagent so the main agent does not absorb long test/coverage output into its context.

When to use

The main agent triggers this skill when the PreToolUse hook blocks a gh pr create call. The main agent spawns a subagent with this skill; the subagent runs the full routine and replies with either:

• "ready" — all checks pass, main agent retries `gh pr create …

  checks-passed`

• "not ready" — short summary of what failed and the precise next step the main agent should take

Routine

Run in this order. Stop and report on the first blocking failure, unless the failure is fixable in place (formatters, trivial lints).

1. Format

• task format if a Taskfile exists
• Otherwise language default (R: styler::style_pkg(), Python: ruff format ., Julia: JuliaFormatter.format("."))
• Commit any formatting changes with a style: prefix

2. Lint

• task lint if available
• Otherwise language default (R: lintr::lint_package(), Python: ruff check ., Julia: Aqua or project equivalents)
• Fix all issues. Commit fixes separately from feature work.

3. Tests

• task test if available
• Otherwise language default (R: devtools::test(), Python: pytest, Julia: Pkg.test())
• All tests must pass. Do not skip or xfail to get green.

4. Type or package checks

• task check if available (runs R CMD check, Pkg.test, etc.)
• R packages: devtools::check() must be clean (0 errors, 0 warnings, note-only is acceptable)
• Python: mypy / pyright if configured
• Julia: Aqua.test_all() if configured

5. Coverage on changed files only

• Identify files changed on this branch: git diff --name-only $(git merge-base HEAD origin/main)..HEAD
• Report uncovered lines in those files only. Ignore pre-existing gaps in untouched files.
• Tools:
  • R: covr::zero_coverage() filtered to changed files
  • Python: pytest --cov=<pkg> --cov-report=term-missing then filter
  • Julia: Coverage.process_folder() filtered to changed files

Skip rules

Skip a step silently if the project has no corresponding tooling (no Taskfile target and no language default configured). Do not invent checks. If the project has no tests at all, note it in the report and continue.

Reporting format

Return a short message to the main thread:

ready|not-ready

<one-line status per step — format, lint, tests, checks, coverage>

<if not-ready: the exact next action for the main agent>

Keep long test output out of the reply. The main agent should not need to see raw stack traces — only what changed and what to do next.

Rules

• Use run_in_background: false when the main agent is waiting to retry; true only if it explicitly wants to keep working.
• Do not run gh pr create yourself; that is the main agent's job once you report ready.
• Do not add # checks-passed anywhere; the main agent appends that on retry.
• Make commits for fixes where appropriate so the main agent's retry includes them.