Writing tests for adept

Defaults

• Table-driven tests with github.com/stretchr/testify/require. One t.Run(tc.name, …)

per case; name cases for the behavior they pin.

• Put unit tests in the same package (package foo) for white-box access; use

package foo_test when you want to assert only the public surface.

• Use t.TempDir() for any filesystem work — never write into the repo or a shared path.
• No network in unit tests. Use net/http/httptest for HTTP clients, and the package's

interfaces + fakes for git, the registry, and the LLM provider (see how Deps is wired).

Golden fixtures

Renderers are pinned by golden files in each internal/render/<id>/testdata/. When you change rendered output on purpose:

1. Update the fixture in the same PR.
2. Explain why in the commit message.

Never loosen an assertion to make a diff pass — fix the code or update the golden deliberately.

End-to-end tests

cmd/adept/*_test.go build the real binary and drive commands against temp dirs with an isolated environment:

env := []string{
    "PATH=" + os.Getenv("PATH"),
    "HOME=" + t.TempDir(),          // isolate user config
    "ADEPT_LIBRARY=" + libRoot,     // isolate the library
}

Guard the slow ones so go test -short skips them:

if testing.Short() { t.Skip("skipping e2e under -short") }

Assert on observable behavior — exit codes, files on disk, stdout/JSON — not internals. Exit codes: 0 clean · 1 error · 2 drift/dirty or merge conflict.

Run them

go test ./...                  # fast
go test -race ./...            # CI gate
go test -run E2E ./cmd/adept   # e2e only
go test -cover ./...           # per-package coverage

Coverage gates

Keep ≥80% on internal/render, internal/status, internal/budget, internal/locks, and internal/canonical. Prefer tests that would actually catch a regression (error paths, edge cases, round-trips) over coverage-padding happy-path calls.