feat: add Browser.cookies with a CookieJar

[Lyesbcb]

Fixes #27.

Implements Browser.cookies and the CookieJar behind it, filling in the scaffolding that Browser and DefaultBrowser already carried in comments.

browser.cookies.setAll(listOf(Network.CookieParam(name = "session", value = "abc", domain = "example.com")))
browser.cookies.getAll()
browser.cookies.save(Path("session.json"))
browser.cookies.load(Path("session.json"))
browser.cookies.clear()

CookieJar is an interface with a DefaultCookieJar implementation, matching how Tab/DefaultTab and Browser/DefaultBrowser are already split here. OpenTelemetryBrowser picks it up for free through its by browser delegation.

Where this departs from the Zendriver original, and why

Cookies go over the browser connection, not through a tab. The original walks browser.tabs, takes the first non-closed one and falls back to browser.connection. Cookies are browser-scoped through the Storage domain, so the tab makes no difference to the result. It also cannot be expressed here: closed is private to DefaultConnection and not on the Connection interface. Using browser.connection directly is both simpler and equivalent.

Sessions are JSON, not pickle. pickle has no multiplatform equivalent, and Network.Cookie is already @Serializable, so kotlinx.serialization does the job and leaves a file you can actually read.

pattern now filters what is written. This is a real bug in the code being ported:

included_cookies = []
for cookie in cookies:
    for match in compiled_pattern.finditer(str(cookie.__dict__)):
        included_cookies.append(cookie)
        break
pickle.dump(cookies, save_path.open("w+b"))   # dumps everything, ignores the filter

included_cookies is built and then discarded, so save(pattern=...) always wrote every cookie. testSavePatternOnlyKeepsMatchingCookies covers it. The pattern is matched against the serialized cookie, which is the closest well-defined equivalent to the original's str(cookie.__dict__).

save and load return the cookies they selected, so a caller can see what a pattern actually matched rather than reopening the file to find out.

requests_cookie_format is not ported, being specific to the Python requests library.

Restoring cookies

getAll yields Network.Cookie while setAll takes Network.CookieParam, so load converts between them through an internal Cookie.toParam(). size and session are dropped deliberately: both are derived by the browser, and session is merely the absence of an expiry, which CookieParam.expires already expresses.

Tests

Four integration tests in BrowserTest, against a real headless Chrome:

• setAll then getAll round trip, checking name, value and domain;
• clear empties the jar;
• save then clear then load restores the cookie, which is the actual use case of the pair;
• save with a pattern keeps only the matching cookie, the upstream bug above.

./gradlew jvmTest is green, 183 tests.

Note that ./gradlew detekt already fails on main for unrelated missing-documentation findings in cdp; nothing here adds to it, and the CI does not run it.

[codecov]

Codecov Report

❌ Patch coverage is 88.57143% with 4 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...otlin/dev/kdriver/core/browser/DefaultCookieJar.kt	87.50%	1 Missing and 3 partials ⚠️

📢 Thoughts on this report? Let us know!

[Lyesbcb]

Follow-up on the save(pattern=...) bug mentioned above: reported and fixed upstream in cdpdriver/zendriver#263 and cdpdriver/zendriver#264, so the two implementations should end up agreeing.

[Lyesbcb]

Pushed a follow-up for the Windows failure.

It was testUpdateTargetSetsTargetTitle, not one of the cookie tests: it reads tab.targetInfo?.title straight after get() with nothing waiting for the load, and got about:blank. The race is pre-existing and has failed on main before, but adding four browser launches to the same class was clearly enough to tip it over, since #80 passes on all three platforms with the same runners.

Two changes:

• the cookie tests no longer navigate anywhere. Cookies are browser-scoped and setAll carries the domain itself, so the browser.get("https://example.com") was never needed. That is four page loads out of the suite.
• testUpdateTargetSetsTargetTitle now waits for the ready state before reading the title.

Happy to drop the second one if you would rather keep it out of this PR.