vue-tui: Instructions file for Codex

Common Pitfalls & Best Practices

• AGENTS.md is the source of truth. CLAUDE.md is a symlink to it.
• Always use Vue's shallowRef over ref by default. Using ref requires a solid justification and a code comment explaining why deep reactivity is needed.
• Always use defineComponent() to define components. Never use bare { setup() {} } objects — they lack component scope, so inject, watch, and onScopeDispose won't work correctly.
• Vue SFCs must use <script setup> unless there's an explicit reason not to.
• When passing function-valued props into composables, do not pass a one-time prop value like useInput(props.onInput). Pass a live source with toRef(props, "onInput"), or pass a wrapper closure that reads props.onInput(...) at event time. For composable APIs that accept function handlers, prefer MaybeRef<Handler> + unref() over MaybeRefOrGetter<Handler>: handler functions and getter functions are ambiguous.
• For ordinary Vue UI, start with template syntax. Use JSX/TSX when it makes test fixtures or highly dynamic structures clearer, and reserve h() for renderer internals or genuinely programmatic vnode construction where template/JSX would be the wrong tool.
• Prefer kebab-case for new file names, including Vue SFCs and JSX/TSX files. Keep local consistency when touching existing areas, and do not rename existing files only for casing unless the task is explicitly about naming.
• When code must deviate from normal/idiomatic style because the situation genuinely requires it (e.g. a control-character regex in a terminal parser, a deliberate string code-point spread, a lint rule suppressed for a justified reason), add a comment explaining why it has to be written that way. Don't silence a linter or write surprising code without a note — the next reader should not have to guess whether it's intentional.
• Bug fixes must follow test-first: write a failing test that reproduces the bug, then fix the code and verify the test passes.
• Runtime behavior is proved by end-to-end and integration tests; unit tests are supplementary. Never change internal code, or add code to it, for the sake of a unit test. Vouched in Architecture decisions.
• Tests must simulate the output capability they assert. Ordinary Runtime and @vue-tui/testing tests select the public color option instead of mutating process color variables; focused Runtime unit tests pass an internal color capability. Resolver tests may model FORCE_COLOR / NO_COLOR precedence directly, and PTY tests exercise automatic detection against a real TTY. A spawned tool whose own CLI color behavior is in scope may use one deliberate color-control state, but must never set NO_COLOR together with FORCE_COLOR.
• When debugging color/ANSI output in a non-TTY shell, remove NO_COLOR and use FORCE_COLOR=3 for that command. To debug suppression, remove FORCE_COLOR and use non-empty NO_COLOR. Never encode “unset” as an empty string in shared test configuration.
• A package's src/ contains build inputs only. Put tests, type-test fixtures, and other test-only inputs under that package's tests/ tree. Use tests, never a singular test directory.
• Keep source-owned tests under packages/*/tests. Cross-package Runtime behavior belongs in the private tests/runtime workspace, and full Vite plugin system tests belong in the private tests/vite workspace.
• Performance measurements belong in private top-level benchmarks/* workspaces and use *.bench.*. Keep their workload inputs under fixtures/; correctness and lifecycle coverage remain ordinary tests in the owning package or system-test workspace. check typechecks benchmarks but does not run host-sensitive measurements.
• Package-local tests/ trees mirror source or contract ownership without a redundant unit/ layer. System-test workspaces classify top-level suites by verification kind; keep each suite's fixtures and harness beside that suite rather than mixing support directories with suite kinds.
• An ordinary unit test mirrors its target's source path and file stem. When one target has several substantial test facets, put them under a directory named for the target rather than creating one giant file or a pile of ambiguous compound names. Integration and end-to-end tests instead live in explicit integration / e2e directories and are named for the contract or user-visible claim they prove.
• Use *.test.* repository-wide, never *.spec.*. The standardized *.sequential.test.* form is the execution-constraint exception, not a behavior name.
• Test files run in parallel (fileParallelism: true), but tests WITHIN a file run serially. We deliberately do NOT set sequence.concurrent: many render tests assert timing-sensitive counts driven by the renderer's ~32ms commit throttle, and in-file concurrency starves them of wall-clock on a 4-core CI runner (it passes on higher-core dev machines — the classic local-vs-CI trap). PTY tests need pool: "forks" (node-pty requires child_process.fork, not worker threads).
• Tests that depend on process-global state — fake timers (vi.useFakeTimers, which mutates global setTimeout/performance) or assertions on shared globals (process.listenerCount, live yoga-node counts) — live in *.sequential.test.* files. Even file-level parallelism can perturb them, and grouping them by name documents the constraint. Add a header comment saying which global forces it.
• Tests must not implicitly depend on the host environment. If a test needs a specific TTY/color behavior, model it through the official test host, a focused fake stream, an internal color capability, or a sanitized child environment rather than relying on ambient values or a Vitest-wide color override.
• After completing any task, run vp run check (or vpr check) to verify: lint, type-check, test all packages, and build.
• Commit messages and PR titles must follow Conventional Commits, e.g. fix(runtime): preserve inline history on resize.
• Never commit anything under docs/. That directory is for local working notes and specs — it must stay out of git.
• vue-tui is an independent product. Peer frameworks provide bounded behavioral evidence; they do not define product scope, package membership, public API, or a default behavior. Start from local product intent and contracts, then use the standing set and selection rules in .agents/docs/peer-frameworks.md when a comparison would reduce uncertainty.
• A terminal-dependent peer behavior claim is a hypothesis until a real harness verifies it against the exact upstream version or commit examined. For each new or reopened decision, pin load-bearing peer evidence in the local record that uses it. Source reading, shared names, and memory are not proof of observable output, input, timing, or cleanup.
• Record accepted vue-tui behavior in its owning design record or decision ledger, never in a parallel parity or divergence catalog. A surprising difference from relevant peers deserves investigation and a local reason; it is not automatically a bug.