Driving Cypress with cypress tap

cypress tap controls an already-running cypress open session. Use it to iterate on specs, inspect the reporter and command log, and read the app under test without GUI interaction. Use cypress run instead for a one-shot headless batch.

Core loop

npx cypress tap sessions                    # confirm a session; note pid if several exist
npx cypress tap status                      # baseline startedAt + spec
npx cypress tap run cypress/e2e/foo.cy.ts   # async — returns immediately
# poll status until passed|failed for your spec with a changed startedAt
npx cypress tap reporter
npx cypress tap reporter --test-id r3       # one test: log, network, stubs, error

run is asynchronous. A stale verdict matches a fresh one on every field except startedAt, and a watcher rerun of the selected spec can advance startedAt without running your file. Gate on startedAt, spec, and a terminal status. Do not wait to observe running — fast specs can finish between polls. Copy-paste poller: session-lifecycle.md. JSON field helpers: json-recipes.md.

In poll loops, prefer ./node_modules/.bin/cypress tap … over npx cypress tap … when the project pins Cypress.

Prerequisites

• Confirm from package metadata or the lockfile that the resolved Cypress is 15.21.0+ and contains tap. Do not use tap --help to probe an unknown older build; prereleases below the version floor may attempt session discovery instead of printing help.
• The session must use Electron, Chrome, Chromium, or Edge. Firefox and WebKit are unsupported.
• cypress open and any configured baseUrl dev server must already be running.
• The cwd chooses both the Cypress binary and the automatically selected session. When the target project pins an older Cypress, run tap from a compatible checkout and pass --session <pid> on every call.

Route by task

• Start, select, or poll a session: read session-lifecycle.md.
• Run a spec or read its results: read session-lifecycle.md and reading-results.md.
• Author or inspect a spec: read recipes.md and reading-the-app.md.
• Diagnose a failure: read recipes.md, reading-results.md, and reading-the-app.md.
• Command failure, hang, wrong project, or surprising output: read troubleshooting.md.
• Parse --json or script a poll: read json-recipes.md.
• One noninteractive batch: use cypress run, not tap.

Read only the references required for the current task.

Core commands

• sessions: reachable sessions, project roots, testing types, and browsers; JSON adds support and renderer health.
• status: lifecycle stage, selected spec, run identity, counts, build error, and active pin.
• specs: runnable project-relative spec paths for the session's testing type.
• run <spec>: dispatches a spec and returns immediately.
• reporter: spec overview and test ids; with --test-id, the complete test attempt.
• command: one command-log row with network data, snapshots, and console properties.
• pin: rewinds the app frame to a command snapshot.
• dom, aria, inspect: read the settled app or currently pinned snapshot.

All commands accept --session <pid>, --json, and --timeout <ms>. On a confirmed supported build, use npx cypress tap <command> --help for command-specific flags. Help is generated from the live session contract — when help and this skill disagree, trust help.

The non-negotiable verdict rule

run confirms dispatch, not execution, and returns before the new run starts. During that gap, status and app reads can still return the previous run's plausible verdict and page.

For every explicit run:

1. Read the current startedAt.
2. Dispatch exactly one spec.
3. Poll one status --json response at a time.
4. Accept only passed or failed for the expected spec with a non-empty, changed startedAt. startedAt is null only when no run has ever started for that spec in the session — a build failure on first selection. A build that fails on rerun or on a watcher rebuild still advances startedAt, so it takes the normal path. Keep the null fallback (a changed observable baseline, or a preceding loading/running observation) for the first-selection case only.
5. Bound the loop and fail if no matching fresh verdict arrives.

Saving the active spec triggers an automatic watcher run. After editing, either use that run or let it settle before taking a baseline and dispatching another. Never intentionally put two runs in flight.

Blank and partial payloads from a successful status call occur transiently. Treat missing fields as "keep waiting," not as a state change. A nonzero status exit is a command failure, not a partial read: stop polling and report it.

Only passed and failed are verdicts. A build failure is failed with the diagnostic in status.error, possibly before any tests exist. status is the only surface that carries that diagnostic — reporter renders a failed build as an empty spec.

Critical correctness rules

1. Target the intended session. If several sessions exist, or auto-selection behaves oddly, inspect sessions and pass --session <pid> on every call. Auto-selection can choose another project or an unresponsive session. E2e and component testing use separate sessions: they can share a projectRoot but expose different specs lists, and only status names which session answered — other commands carry no pid in their output.
2. Preserve the binary location. Cwd controls npx resolution on every call. When the project pins an older Cypress, run commands from a compatible checkout and pass --session <pid>.
3. Do not parse failed commands. Check the exit code before parsing JSON. Supported-build failures generally use stderr, but older compatibility failures may use stdout. An ambiguous selector is the intentional exception: it exits 1 and lists matches on stdout.
4. Do not discard dispatch stdout while checking compatibility. An older Cypress may print Unknown command "tap" and usage text to stdout; redirecting it hides the cause.
5. Redirect potentially large JSON. reporter --test-id --json and command --json can be hundreds of kilobytes. Save them to a file and parse the file.
6. Check truncation before concluding absence. dom and aria cap output. Narrow the selector or raise the limit when (output truncated) appears.
7. Sanity-check the live frame before concluding absence. A trailing pending/skipped test can leave the settled runner on a blank placeholder while app reads still exit 0. Confirm a known app anchor. If the frame is blank, pin a snapshot from the last real command and read that state instead.
8. Read results before editing or deleting the spec. Results and snapshots live in the Cypress app's memory and can disappear on rerun, restart, rename, or deletion.
9. Clear pins. After inspecting a command snapshot, run pin --clear; otherwise later app reads continue to describe the pinned past. An exit-0 cleared:false result is a benign no-op even if human output says FAILED TO CLEAR PIN.
10. Use the right reader for live state. Use aria for current form-control values: dom and inspect can show the initial HTML value attribute, and inspect may omit the accessibility value. Use dom for exact live-region, toast, status, and label text because aria may omit descendant text.
11. Chain broad selector probes with ;, not &&. For dom/aria/inspect, zero matches exit 0 while ambiguity exits 1. A broad probe that enumerates matches is the useful outcome; && drops every follow-up command after the first ambiguous selector.
12. A passing verdict is not a full proof. After green, read reporter --test-id and confirm intercept match counts, stub usage, and that requests you care about were stubbed. See reading-results.md and json-recipes.md.

Output contract

• Human output is for reading; --json is for parsing and may contain much more data. Default rendering is usually better for one-off debugging: it inlines code frames and interleaves network events in log order.
• Exit 0 is not success of your goal. tap does not start Cypress. With no session, status is not connected at exit 0; sessions prints guidance instead of JSON. Selector misses, empty specs, and unknown --session pids can also exit 0. Branch on fields, not silence.
• status exits 0 for known lifecycle stages, including not connected. Discovery, compatibility, unsupported-browser, and renderer failures exit 1, sometimes with no stdout; a poller must fail fast on that nonzero exit.
• dom, aria, and inspect require exactly one selected element. Ambiguity exits 1 with candidate selectors. A miss is not a CLI failure: dom and inspect report found:false; aria returns an empty tree both for a miss and for an element with no accessibility node.
• Failures are prose without stable error codes. Branch on exit status, not message text.

Performance defaults

• Prefer one reporter --test-id read over one command call per row.
• After a fresh verdict and live-frame sanity check, independent app reads may run concurrently.
• If bounded status polling fails, inspect sessions for rendererResponsive: false; restart a wedged renderer instead of increasing --timeout.

Route by symptom