optimize-login
/optimize-login
Skraca testy z logowaniem do aplikacji webowej: wykonuje logowanie na żywo, redukuje je do najmniejszego działającego skryptu, zapisuje sesję przeglądarki i wczytuje ją w nowej przeglądarce.
- Autor
- mhmdelgazzar ↗
- Repozytorium
- agentex ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 26.10.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 3
- Skrypty
- 2
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 1
Brak sygnałów. Repozytorium zmienione 26.10.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- 2 skrypty w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: scripts/session.js, scripts/session.test.js
Co robi i kiedy
Skraca testy z logowaniem do aplikacji webowej: wykonuje logowanie na żywo, redukuje je do najmniejszego działającego skryptu, zapisuje sesję przeglądarki i wczytuje ją w nowej przeglądarce.
Oryginalny opis (EN)
Pay a web application's login cost once per session instead of once per test — drive the login live, reduce it to the smallest script that works, save the browser session, and reload that session into a fresh browser to continue. Use when login is the slow part of a run, when several scenarios each re-login, or whenever a run needs an authenticated browser.
Jak działa
- Role
- The loop
- Verify by landmark, never by URL
- When a step does not work
- Record what you learn, per application
- Limits
- Output
- Session files are credentials
Podgląd SKILL.md
---
name: optimize-login
description: >
Pay a web application's login cost once per session instead of once per test — drive the
login live, reduce it to the smallest script that works, save the browser session, and
reload that session into a fresh browser to continue. Use when login is the slow part of a
run, when several scenarios each re-login, or whenever a run needs an authenticated browser.
---
# Optimize Login — pay for the login once
## Role
You turn a web application's login into a one-time cost per session: discover the real login
live, script the smallest path through it, then save and reload that session for every later
run. You never build around a security gate — a captcha or OTP you cannot solve is handed to
a person, never bypassed.
Login is usually the most expensive step of a browser run and the least interesting. On a real
project it was **~197 seconds of agentic driving per scenario**, inside a ~12-minute preamble.
After applying this skill: **~38s once**, then **~8s** per later run.
The loop below is the same everywhere. **How much code it takes is not** — a username/password
form is three lines; a cross-origin iframe with a captcha and an OTP is a hundred and fifty.
Find out by looking, not by assuming.
## The loop
1. **Look at the real login.** Open it and read what is actually there: which fields, whether
the form sits in an iframe, what the submit control is, what stands between you and a
session. Never write the script from memory or from another application's login. Log in
with a user from the active environment's `users` (password: the user's `password` or
`defaults.password`; `login.mode` in `config/project.json` says whether to reuse saved
sessions (`"session"`) or log in fresh every run (`"fresh"`)).
2. **Write the smallest script that gets through it**, driving only what the page requires.
3. **Prove you are in** — by a landmark, never by the URL (see below).
4. **Save the session** (`storageState`). In a playwright-cli session (a test run,
define-flow): `login_state.js save`, never `state-save` by hand.
5. **Reload it into a fresh browser, prove you are in again**, then continue the real work.
In a playwright-cli session: `login_state.js load`, which loads into that session and
checks the landmark saved with the file.
6. **When the session dies, log in again.** Do not build renewal machinery; re-login is cheap
once step 2 exists.
Steps 3–5 are identical in every application and are bundled:
- **In a playwright-cli session** (a test run, define-flow):
`${CLAUDE_PLUGIN_ROOT}/skills/browser-driver/scripts/login_state.js`. Follow the browser-driver
skill's "Login mode" section and
`${CLAUDE_PLUGIN_ROOT}/skills/browser-driver/references/tools/login-state.md`.
- **In your own library script, or a standalone check:**
`${CLAUDE_PLUGIN_ROOT}/skills/optimize-login/scripts/session.js` —
`saveSession(page, {statePath, landmark})`, `resumeSession({statePath, url, landmark})`,
`isAuthenticated(page, landmark)`.
Steps 1–2 are yours to discover per application.
To check a saved session outside a run, without writing any project code (this launches its own
library browser, not the run's session):
node ${CLAUDE_PLUGIN_ROOT}/skills/optimize-login/scripts/session.js resume \
--state test/.auth/<environment>-<user>-state.json \
--url https://app.example.com/dashboard \
--absent "role=button[name='Login']"
Prints one `RESULT: RESUME_PASS|RESUME_FAIL` line.
`session.js` needs the `playwright` PACKAGE, because it drives a library browser of its own (a
playwright-cli session loads a saved `storageState` through `login_state.js` and needs no
library). It looks for it in the project you run it from — the working directory and its
parents, so a monorepo hoist works — falling back to `NODE_PATH` if you set one. If
the project has not got it: `npm i -D playwright`, then `npx playwright install chromium`.
The bundled Chromium is what it launches; add `--channel chrome` (or `msedge`, …) only when
the test needs that specific browser, and `--headed` to watch a resume fail with your own
eyes. A channel you ask for and Playwright cannot launch is an error — no other browser is
quietly used in its place.
## Verify by landmark, never by URL
A landmark is something true **only** when logged in: an element that appears (an account
menu), or one that disappears (the Login button). Pass `{present: sel}`, `{absent: sel}`, or
both.
URLs lie. A login page can carry `?returnUrl=/dashboard` and satisfy any path-based check while
the user is still logged out — observed live, reporting a successful login that had not
happened, with the real failure surfacing much later somewhere unrelated.
Verify **before** saving and again **after** loading. A state file written from a half-finished
login is well-formed and useless; and a session that has since expired leaves the file
untouched, so "the file exists" is never "the session is alive". Do not gate on the file's age
either — a 15-minute-old session was dead here while a 47-minute-old one restored cleanly.
## When a step does not work
Read the live page, not your assumptions: dump the fields, buttons, disabled states and any
error text **as they are at the moment of failure**, then fix that specific thing and re-run.
Most login scripts fail for a reason the page will state plainly if you ask it.
Two habits that pay for themselves:
- Give every login script a hard watchdog, so a hang always terminates with a result line.
- Make failures dump page state automatically — the next repair should start from evidence,
not from a fresh guess.
## Record what you learn, per application
Whatever you discover about a specific login — a control that ignores clicks on its label, a
field the server reads that the visible inputs never populate, a step that opens on a different
method than you expect — belongs in that application's own notes, not in this skill. In a
project using the AgenTeX page map, that means the page's `gotchas`.
This is deliberate. A shipped list of somebody else's surprises invites you to read the next
login through the wrong lens; the surprises are findings from exploring one application, and
the next one will be strange in its own way.
## Limits
- **Gates are surfaced, never defeated.** A captcha the page itself renders can be handed to a
person, or read by the agent. reCAPTCHA, hCaptcha and Turnstile are not that, and no attempt
to work around them belongs here — run headed, let a person complete the login, and the
session is still saved at the end.
- **An OTP you cannot obtain is a stop.** Static test codes work; a real SMS, email or
authenticator code needs a person, and the headed fallback covers it.
- **`storageState` is cookies and localStorage only.** An application holding its tokens in
IndexedDB cannot be resumed this way — the post-load landmark check surfaces that plainly
instead of leaving a confusing failure later.
- **Only for applications you are authorised to access.** This is a way to stop paying for your
own login repeatedly, not a way into anyone else's account.
## Output
This skill produces no report of its own — its output is state, consumed by whatever flow
invoked it:
- The bundled `session.js resume` check prints exactly one line, `RESULT: RESUME_PASS|RESUME_FAIL`,
and exits 0/1 to match.
- A successful run leaves a session file at `test/.auth/<environment>-<user>-state.json` for
later runs to reload.
- Anything else worth keeping — what the login looked like, gotchas found — goes to the
application's own notes (see "Record what you learn" above), never into this skill.
## Session files are credentials
A saved session is a bearer token in a file: whoever holds it is logged in as that user. Keep
these in `test/.auth/`, which /init-test and /update-agentex add to the project's `.gitignore`,
and never commit one. Never open, print, copy or edit one (or its `.meta.json` sidecar); in a
playwright-cli session only `login_state.js` loads or saves it.
Sessions are saved per environment and user — `test/.auth/<environment>-<user>-state.json`
(e.g. `qa-buyer-state.json`); a session saved on one environment is never resumed on another,
and one user's is never resumed as another's.
Pliki w paczce (3)
- SKILL.md 8 KB
- scripts/session.js 10 KB
- scripts/session.test.js 12 KB
Wersje
W bazie nie ma innych skilli o nazwie optimize-login.
Migawka bazy z 13.11.2026 · wersja paczki 137ed6ff1ed6 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 86real-test
Uruchamia scenariusze testów z prawdziwym logowaniem przez użytkownika i zapisuje wyniki w dokumentach akceptacji. - 100agent-browser
Wykonuje zadania w przeglądarce z poziomu CLI, takie jak nawigacja, logowanie, formularze, zrzuty ekranu, ekstrakcja danych i testy aplikacji webowych. - 100web-validate
Sprawdza wdrożoną stronę w prawdziwej przeglądarce, także za logowaniem, używając profilu zalogowanego przez użytkownika. - 100haven-auth
Loguje do stosu worktree jako admin lub zasiany użytkownik poleceniem haven browser login --as, zapisując plik storage-state Playwright i nigdy nie wypisując hasła. - 100browser-adapter
Znajduje w sesji dostępne narzędzie przeglądarki i wykonuje na żywo testy aplikacji własnej lub portalu, który tylko czyta, bez zmian danych. - 100autoweb
Steruje prawdziwą przeglądarką według pliku celów, z zapisanym logowaniem i równoległymi ścieżkami, raportując fakty ze źródłem.