rust-unit-tests
GitHub指导在 warp 项目中编写、改进和运行 Rust 单元测试。明确测试范围,强调确定性逻辑应使用单元测试,而涉及 IO 或组件交互的场景需升级至集成测试,并提供具体的代码结构和断言规范。
Trigger Scenarios
Install
npx skills add warpdotdev/warp --skill rust-unit-tests -g -y
SKILL.md
Frontmatter
{
"name": "rust-unit-tests",
"description": "Write, improve, and run Rust unit tests in the warp Rust codebase."
}
Rust Unit Tests in warp
Scope
- This skill focuses on crate-level unit tests.
- Favor incremental, well-scoped tests that exercise a single function or behavior per case.
- For how to structure and run tests, read on. For whether a change needs a test and at what level, start with "What to unit test".
What to unit test
Default to a unit test when the logic is deterministic and reachable without booting the app:
- Parsing, encoding, escaping, and any pure data transformation.
- Domain logic and state machines (block lifecycle, selection, ranges, diffing).
- Boundary and edge cases: empty input, single element, off-by-one at range ends, min/max, invalid UTF-8, zero-width and wide grapheme clusters.
- Error paths and fallbacks, not just the happy path.
- Bug fixes, where the bug is reachable at this level. A reproducible bug usually means a case was missing from the suite, so add the failing test first and then fix it. Not every bug is unit-testable — a rendering artifact, a PTY timing race, or a crash that needs a real window will not reduce to one. Cover those at the level where they actually reproduce, and don't reshape the code just to force a unit test.
Prefer unit tests for anything that fits: they are fast, deterministic, and point straight at the failure.
When a unit test is the wrong level
Be honest about this codebase. Warp is a terminal emulator with a GPU renderer, PTY and shell integration, and IPC. The common "80% unit tests" heuristic assumes a business-logic-heavy system where units exchange messages and transform data. Large parts of this repo are not that, and forcing them under a unit test usually means mocking away the only thing that could actually break.
Escalate to a higher level when any of these hold:
- The behavior needs a real PTY, shell, window, or display to exist at all.
- What can break is the wiring between components — a setting taking effect, a keybinding dispatching, focus moving across panes — rather than logic inside any one of them.
- Reproducing it requires real IO, spawned processes, or cross-process timing.
- The only way to make it unit-testable is to add a trait and an indirection layer that exist solely for the test.
- You find yourself stubbing so much that the test no longer exercises anything real.
Where to go instead:
gui-integration-test— GUI end-to-end behavior, terminal and shell integration, settings and keybinding wiring.tui-testing— TUI element and screen rendering.gui-integration-test-videoorcomputer_use— when the real question is visual and someone needs to look at it.
Before escalating, try splitting the problem. Most "untestable" code is a thin shell of IO wrapped around logic that tests fine once separated: extract the decision-making into a pure function, unit test that, and let an integration test cover the thin shell that remains. That is usually cheaper than either a heavily stubbed unit test or a full app-boot test.
That said, weigh the indirection on its own merits. Code that is hard to test is sometimes genuinely badly designed — if extracting the logic would make the code clearer regardless of testing, do it. If the seam would exist only to satisfy a test, don't.
When NOT to write a unit test
Tests cost real maintenance, and a bad test costs more than no test. Skip or delete these:
- Change-detector tests. A test that restates the implementation — inject two collaborators, assert they were called in order — fails on every refactor and catches no defects. It has negative value. Rewrite it as a state assertion or delete it.
- Trivial code with no logic. Getters,
From/Intopassthroughs,Defaultimpls, plain struct construction. There is nothing that can break independently. - Code you don't own. Don't test the standard library,
tokio, orwgpu. Test your usage of them. - Redundant tests. If a case is already covered, a near-identical test adds maintenance cost and no signal. Prune tests as ruthlessly as production code.
Don't chase a coverage number. Coverage says a line executed, not that anything was verified, and a target reliably turns into a ceiling.
Where unit tests live
- Put unit tests in separate files named
${filename}_tests.rsormod_test.rs. - Include the test module at the end of the corresponding source file:
#[cfg(test)]
#[path = "filename_tests.rs"] // or "mod_test.rs"
mod tests;
Writing good tests
Aim for a test you never touch again unless the behavior changes. Refactors, new features, and bug fixes should not require editing existing tests; only a deliberate behavior change should. If a refactor breaks your tests, that is usually a defect in the tests.
Test behavior through the public API
Exercise the unit the way its callers do. Reaching into private state makes the test fail on refactors no caller would notice. If a helper type exists only to serve one or two callers, test it through them rather than directly.
Assert on state, not on interactions
Assert what the system is after the action, not which functions it called to get there.
// Brittle: still passes if the entry is dropped right after insertion, and
// fails on an equivalent refactor that calls a different internal method.
assert!(recorder.saw_call_to_insert(id));
// Better: asserts the outcome the caller actually cares about.
store.insert(id, entry.clone());
assert_eq!(store.get(id), Some(&entry));
One behavior per test, named after that behavior
The test name is often the only thing visible in a failure report, so make it a sentence about behavior rather than about the method:
#[test]
fn parses_utf8_sequence_when_valid() { /* ... */ }
#[test]
fn returns_replacement_char_for_invalid_utf8() { /* ... */ }
If the name needs an "and", you are testing two behaviors — split it. Structure the body as arrange / act / assert, separated by blank lines.
Keep the test complete and concise
Everything a reader needs to understand the result belongs in the test body; everything irrelevant belongs out of it. Prefer a builder or helper constructor that takes only the fields the test cares about over one large shared fixture. If a test asserts on a specific value, set that value in the test rather than inheriting it from shared setup.
Prefer duplication over indirection
Test code has no tests of its own, so it has to be obviously correct on inspection. Some repetition is a fair price for a test that reads top to bottom. Extract a helper when it removes noise, not merely to remove repetition.
No logic in tests
No conditionals, loops, arithmetic, or string concatenation to compute an expected value. Write expected values literally — computing them re-implements the code under test and can reproduce the same bug in the assertion.
Make failures diagnosable
- Prefer
assert_eq!/assert_ne!overassert!for readable diffs. - Add a message when the values alone aren't self-explanatory:
assert_eq!(got, want, "cursor should clamp to line end for {input:?}"). - Use
#[should_panic]only when panicking is intended API, and pin the message withexpected = "...".
Repo-specific
- Minimize global state; inject dependencies via traits/constructors so logic is testable without heavy mocking.
- When adding an enum variant or expanding behavior, prefer exhaustive matches in the code under test and mirror the new cases in tests.
- Be mindful of terminal model locking: avoid patterns that acquire multiple
model.lock()calls in the same call stack from tests, and prefer passing an already-locked reference down.
Test doubles: prefer real code, then fakes
Work down this list and stop at the first option that is fast and deterministic:
- The real implementation. The default for value types and pure logic. Running real collaborator code is what makes a test meaningful, and a failure caused by a real dependency's bug is a true positive worth having.
- A fake the repo already provides. These are maintained alongside the real thing, so they don't drift the way hand-rolled doubles do:
warpui::App::test,VirtualFS,TerminalModel::mock(..),TestBlockListBuilder/TestBlockBuilder,Appearance::mock(), andFeatureFlag::X.override_enabled(..). See "Common helpers to use" below for usage. - A stubbed return value, only to push the unit into a state you cannot otherwise reach, such as a rare error branch. Each stub should map to an assertion in the same test. Needing many stubs is a signal the unit does too much.
- Asserting that a call happened, as a last resort, and only for state-changing effects (a write, a send, a spawn) whose result you cannot observe any other way. Never assert on calls to pure getters — the return value is already covered by whatever you assert next.
Keeping tests deterministic
A flaky test is worse than no test: once people learn to re-run a red test, they stop trusting every other test too. Fix the cause instead of adding retries.
- Time — never read the system clock from logic under test. Inject a clock or timestamp so the test can pin it.
- Async — never sleep to wait for something. Await the future, use a callback, or poll for the state transition with a generous timeout.
- Ordering — tests must pass in any order and in parallel. Watch for statics, singletons,
OnceCell, and environment variables. - Shared state — when a test must touch global or external state, use
serial_test's#[serial]or scope the state locally.
If you can't make a test deterministic quickly, quarantine it (#[ignore] with a linked issue) rather than leaving an intermittently red test in the suite — and treat that as debt to pay down, not a place to leave it.
Async and feature-gated code
- For async logic, use
#[tokio::test]when the code requires a runtime. - Prefer runtime feature checks (e.g.,
FeatureFlag::X.is_enabled()) over#[cfg(...)]so tests don’t require recompilation to toggle behavior.
Quickstart harness (UI/model tests)
- Prefer
warpui::App::testfor deterministic unit tests around views/models. - Initialize app models once, then mutate via
updateand assert viaread.
use warpui::App;
// In app crate tests prefer `crate::test_util::...`; from other crates use `warp::test_util::...`.
use warp::test_util::{terminal::initialize_app_for_terminal_view, add_window_with_terminal};
#[test]
fn example() {
App::test((), |mut app| async move {
// One-time app setup for terminal/view tests
initialize_app_for_terminal_view(&mut app); // includes settings init
let term = add_window_with_terminal(&mut app, None);
// Act
term.update(&mut app, |view, _ctx| {
view.model.lock().simulate_block("ls", "out");
});
// Assert
term.read(&app, |view, _ctx| {
assert!(view.model.lock().block_list().len() > 0);
});
})
}
TUI element tests
Tests for the headless TUI render an element tree to text lines rather than drawing pixels. Use warpui_core::elements::tui::test_support::render_to_lines and TuiBuffer::to_lines, and keep them in *_tests.rs files next to the source in crates/warp_tui and crates/warpui_core/src/elements/tui. They are plain unit tests and do NOT use the GUI integration / real-display / computer_use framework. See the tui-testing skill for details. The warpui::App::test harness above still applies to shared model logic that both front-ends use.
Common helpers to use
- Terminal model shortcuts:
TerminalModel::mock(..),.simulate_block(..),.finish_block(),.simulate_cmd(..). - Builders for focused tests:
terminal::model::test_utils::{TestBlockListBuilder, TestBlockBuilder}. - Virtual filesystem for IO-heavy code:
use virtual_fs::{VirtualFS, Stub};
VirtualFS::test("case", |_dirs, mut fs| {
fs.with_files(vec![Stub::FileWithContent("path/file.txt", "contents")]);
// run logic and assert
});
- Feature flags (scoped):
use warp::features::FeatureFlag; // or `use crate::features::FeatureFlag;` inside the app crate
let _flag = FeatureFlag::CreatingSharedSessions.override_enabled(true);
- UI numeric assertions (lines):
assert_lines_approx_eq!(actual_lines, INLINE_BANNER_HEIGHT);
- Concurrency: keep
model.lock()scopes minimal; avoid nested/re-entrant locks in the same call chain. - Don’t call
initialize_settings_for_testsdirectly when usinginitialize_app_for_terminal_view(it already calls it). - Async needs: use
#[tokio::test]when a real runtime is required; otherwise preferApp::test. - Tests touching global/external state: consider
serial_test's#[serial]or local mocking instead of parallelism.
Running unit tests
- Workspace (parallel):
cargo nextest run --no-fail-fast --workspace --exclude command-signatures-v2
- Single crate:
cargo nextest run -p <crate_name>
- Single test (filter by name):
cargo nextest run -E 'test(<substring>)'
- Doc tests:
cargo test --doc
Linting and formatting
Run before submitting changes:
./script/format
cargo clippy --workspace --all-targets --all-features --tests -- -D warnings
For a full local check before a PR, you can also run:
./script/presubmit
Version History
-
04a7f83
Current 2026-08-19 21:13
新增关于何时选择单元测试与集成测试的指导原则,强调避免过度 mocking,并增加对 bug 修复场景下测试编写的具体建议。
- 726c1b6 2026-07-24 20:21


