Every promise, proven.
Declare what your system must do. Temper fails the build when a promise has no proof.
Software development is moving to LLMs. Temper is a framework for guiding LLMs to build resilient, bug-free software.
People specify the functional surface area of the system and its invariants: what it must do, and what it must never do.
Temper associates each test with the functionality it proves. It builds a coverage report from those links. People and CI use the report to audit systems that LLMs generate.
Temper works with Go and Python (pytest).
Install
brew install turbolytics/tap/temper
Or go install github.com/turbolytics/temper/cmd/temper@latest. For Python, pip install temper-python.
In 30 seconds
Declare what the system does in temper/features.yml:
features:
- id: cart.total
description: Totals a cart, applying the best discount.
requires: [unit]
- id: cart.checkout
description: Charges the card and empties the cart.
requires: [unit]
Each test states what it proves:
func TestTotal(t *testing.T) {
temper.Covers(t, "cart.total") // this test proves cart.total
if got := Total([]int{250, 100}); got != 350 {
t.Fatalf("Total = %d, want 350", got)
}
}
Run the tests through Temper, then build the matrix:
temper run --level unit -- go test ./...
temper matrix --all-levels
temper/features.yml: cart.checkout is missing at unit
summary: 1 covered, 0 exempt, 0 failing, 1 missing, 0 unenforced
cart.checkout has no passing test, so the build fails. A skipped test does not count.
Next, follow the tutorial, which designs Temper itself with Temper. The README covers setup, and the specification has every rule.
Findings so far
Temper's findings span 9 open-source projects with 24.8k GitHub stars. We reported each bug to the project's maintainers.
- 9 issues: 1 confirmed, 8 reported
- 8 pull requests: 5 open, 3 draft
Work with us
Contact us to get onboarded to Temper, or to train your team to build better software with LLMs.