Skip to content

test-priority

untested says which symbols no test exercises. This command says which of those to test next: the untested symbols ranked by downstream × churn × (1 + coupled partners), the same product risk uses, with file-level dependents replaced by symbol-level downstream — the files that import that symbol specifically. A file ranking tells you where to look; a symbol ranking tells you which test to write next, and that is the only refinement this join buys over ranking by file risk alone.

Terminal window
sephera test-priority --path . --top 10 # the ten most dangerous untested symbols
sephera test-priority --path . --top 10 --format json # the same ranking, machine-readable
sephera test-priority --path . --window 1000 --top 10 # the same ranking over a longer history window

Run from inside the repository whose history you want to read.

Terminal window
sephera test-priority --path .
sephera test-priority --path . --top 10
sephera test-priority --path . --window 1000 --format json

For each untested symbol:

priority = symbol_downstream × file_churn × (1 + file_coupled_partners)
  • symbol_downstream: the number of direct consumer files importing that symbol specifically, read with breaking’s consumer scan. Namespace users are excluded — a namespace import names nothing, the same rule untested itself applies.
  • file_churn: the commits touching the symbol’s file inside the window, under coupling’s rule.
  • file_coupled_partners: the files coupled to the symbol’s file under coupling’s defaults.

Both file-level inputs are shared by every symbol in a file; only downstream refines within a file. The score is a ranking heuristic, not a probability. Churn window truncation is intended: recently-churned untested code is the danger.

  • The output lists the FULL untested set, ranked, never filtered. A symbol with zero downstream scores 0 and ranks last, but stays listed — dropping it would hide work.
  • Tested symbols never appear, direct and indirect tiers alike.
  • The order is total: score desc, file asc, symbol asc. The same tree twice gives the same ranking.
  • There is no gate in v1: ranking is advisory, and a gate would need a threshold semantic of its own.

The same conventions as untested: a test, tests, __tests__, spec, or specs directory anywhere in the path, or a file named test_*, *_test.*, *.test.*, or *.spec.*. Inline unit tests do not cross an import edge, so a Rust #[cfg(test)] module’s own symbols read as untested unless a separate test file names them — the same limit, stated plainly.

  • A namespace import names nothing, so it is never evidence and never downstream.
  • Symbols declared in a test file itself are not code under test.
  • Only Rust and Python collect declarations, so only their symbols can be ranked.
  • The analysis needs a git history. A tree that is not a git repository is an error, not an empty ranking.

Which test should come first rank gaps by downstream times churn most dangerous first