j)dZddlmZddlZddlmZmZmZmZd%d Z d&d Z d'dZ d(dZ d)dZ d(dZd*dZd+dZd,d-dZd.dZd(dZd/d Zd0d#Zd1d$ZdS)2uTurning protocol values into Python ones, and back. ⛔ THESE ARE LEAVES, WHICH IS WHY THEY COULD MOVE AT ALL. Not one of them mentions a dispatcher class: they take a value and return a value. `server.py` holds a mutually recursive object graph - a page makes frames, a frame makes element handles, a handle points back at its frame - and splitting THAT across modules buys navigation with import cycles, which is a bad trade. These do not have the problem, and they were scattered BETWEEN the classes: reading `ElementHandleDispatcher` you fell into `_serialize` on the way to `FrameDispatcher`. They are re-exported from `server.py`, so nothing that imported them from there has to move. ) annotationsN)AnyDictListOptional expressionstrargumentrreturnc(dt||zS)aHWrap an expression so it can always be CALLED with the argument. PLAYWRIGHT SENDS BOTH FORMS DOWN THE SAME FIELD. `page.evaluate("1+1")` and `page.evaluate("() => document.title")` arrive as the same `expression` string, and the client does not say which is which. Evaluating the raw string works for the first and, for the second, produces a FUNCTION OBJECT: no error, no exception, and the caller gets None back for a call that looked fine. Measured on 2026-08-27: `page.evaluate("() => !!x")` answered None, and the assertion that caught it read like the browser had changed behaviour. So the string is wrapped and the result called if it turned out to be a function. Deciding by shape - does it start with `(` or `function` or `async` - is the version that gets it wrong: `(1+2)` starts with a parenthesis and is not a function. THE ARGUMENT GOES IN HERE, IN THE SAME PASS, AND THAT IS THE WHOLE POINT. It used to be spelled `r(ARG)` and the caller substituted it afterwards with `.replace("ARG", ...)` - over a string that already held the caller's expression. So every uppercase `ARG` a caller wrote was rewritten to the argument's JSON, inside their own code, silently. Measured on 2026-09-14 by driving it: `'CARGO'.length` answered 6 instead of 5, because the source that ran was `'CnullO'.length`; a selector `[data-role='TARGET']` became `[data-role='TnullET']` and matched nothing; `/ARGH/` became `/nullH/`; and `const ARG = 5` was a SyntaxError. The three letters need only appear inside a longer word - MARGIN, LARGE, CHARGE - and three of those four failures are silent, which is the bad kind. A format string is only scanned where the format string itself is, so the expression is an ARGUMENT to `%` and never a place `%` looks. z(() => { %s })()) _invocation)rr s VC:\Projects\Muse2API\.venv\Lib\site-packages\invisible_playwright/_juggler/_marshal.py _as_callablersD  J A A AA receiversc td|dd|tj|tfzdS)aThe statements that evaluate the caller's expression and, when it is a function, call it with `receivers` first and the argument last. THE ONLY PLACE THAT SPELLS THE CALL. The main-world wrapper above and the three element-scoped ones (`handle.evaluate`, `eval_on_selector`, `eval_on_selector_all`) used to write it out each on its own, and that is how the argument reached `handle.evaluate` and was dropped by the other two: a callback reading its second parameter got `undefined` there. The main-world output is pinned byte for byte by `test_the_argument_does_not_rewrite_the_expression.py`, because that one runs as the page. z const r = (z'); return typeof r === 'function' ? r(z, )defaultz) : r;)joinjsondumpsr )rr rs rr r :sEzzyydj3&G&G&G%IIJJJJ LMrparamsrreceivercpt|dt|d|S)zThe body of an element-scoped wrapper: the caller's function called with `receiver` and the caller's argument, Playwright's `fn(element, arg)`. rarg)r _deserializeget)rrs r _called_onrMs6 vl+\&**U:K:K-L-L ! !!rc(dt|dzS)zThe declaration `handle.evaluate` and `eval_on_selector` both send: one element, then the argument. Both run in the utility world.z(injected, el) => { %s }el)rrs r_element_functionr!Us & 64(@(@ @@rvaluect|ts|Sd|vrd|vrt|dSddfddfddffD]\}}||vr|||cSd |vrHd d td td td d|d Sd|vrd|dDSd|vrd|dDSd S)zCThe tagged union Playwright sends for an ARGUMENT, back to a value.r"handlesnc|SNvs rz_deserialize..arsc|Sr'r(r)s rr+z_deserialize..asQrbc|Sr'r(r)s rr+z_deserialize..br,rr*Nnaninfz-inf)null undefinedNaNInfinityz -Infinityac,g|]}t|Sr(r).0xs r z _deserialize..js444A Q444rocFi|]}|dt|dSkr*r9)r:es r z _deserialize..ls*AAA# QsV,,AAAr) isinstancedictrfloatr)r"tagconverts rrr) eT " " %I..E'N+++{{+c;;-?{{+-'' W %<<75:&& & & &  e||4e !%LL"6]],,,/Cc OO < e||44s4444 e||AAeCjAAAA 4rcnt|dt|dS)zThe expression, callable, with the caller's argument placed in it. THE ONLY PLACE THAT BUILDS THIS. `wait_for_function` used to repeat the two lines inline, which is how one of the two could have been fixed and the other left alone. rr)rrrr s r_with_argumentrIps- |,l6::e;L;L.M.M N NNrc*tj|S)ujA Python string as a JavaScript literal. ⛔ `json.dumps` and never manual quoting: the html handed to `set_content` is arbitrary, and a single quote, a backslash or a line separator inside it would close the literal early. That exact defect - an apostrophe closing a single-quoted JavaScript string - broke the driver bundle on 2026-08-24. )rrr"s r _js_stringrLzs :e  r Optional[str]cHt|tr d|vr|dSdS)Nguid)rCrDrKs r_guid_ofrPs+%6U??V} 4rcounterOptional[Dict]cddi|ddiSt|trd|iSt|ttfrd|iSt|trd|iSt|t t fr'dxxdz cc<fd |Ddd St|tr9dxxdz cc<fd |Ddd Sdt |iS) uGA Python value in the shape `parse_result` expects. ⛔ Playwright does not send bare JSON: it sends a tagged union, and `_connection.parse_value` reads the TAG. A bare `True` where `{"b": true}` is expected does not raise - it falls through to the object branch and comes back as something else entirely. Nr%rr*r3r/r-c0g|]}t|Sr( _serialize)r:r;rQs rr<z_serialize..s#<<<jG,,<<.s?111!QJq'$:$:;;111r)r=rX) rCboolintrEr listtuplerDitems)r"rQs `rrWrWsP( }V}%U|%#u&&U|%U|%$'' R   <<<.s*EEEASVV,,EEErrcr"rb)rCrDr^appendr r)rawoutentrys r_headers_arrayris#tFEEEEEE C?? eT " " ?v JJf !$UYYw%;%;!> ? ? ? Jrc |dp|dpd}dddddddd d d d d d |d S)u"⛔ Juggler says `cause`, Playwright says `resourceType`, and the two vocabularies only partly overlap. An unmapped cause becomes `other`, which is what upstream does too - guessing a nicer name would make `request.resource_type` disagree with itself between the two transports.cause internalCauserddocument stylesheetscriptimagefontmediaxhrfetch websocketother) rm subdocumentrnrorpimagesetrqrrxmlhttprequestrtrubeacon)rlower)rrks r_resource_typer|su ZZ EFJJ$?$? E2 L L N NEz"hg' G    c%rrcr[c:dddd|pddS)NrrT)leftmiddlerightrr)rcs r_buttonrs&Q / / 3 3DNFA F FFrargsr\cng}|D]}t|ts#|t|;d|vr)|t|dh|dr)|t|d|dr)|t|d|t|dpdd|S)u⛔ The console arguments arrive as remote objects, not as text. Reading `value` works for a primitive and gives nothing for an object, so the preview is used when there is one - which is what the driver shows too.r"previewunserializableValuetypeobject )rCrDrer rr)rpiecesr7s r _console_textrsF  : :!T"" : MM#a&& ! ! ! ! \\ MM#aj// * * * * UU9   : MM#a l++ , , , , UU( ) ) : MM#a 5677 8 8 8 8 MM#aeeFmm7x88 9 9 9 9 88F  rc|pi}|dpd|dp|dpd|dp|dpddS) u⛔ ALWAYS a location, never None: `_browser_context.py` casts it and falls back only on a missing KEY, so a null here becomes an AttributeError somewhere else entirely.urlrd lineNumberliner columnNumbercolumn)rrrr)rfs r _locationrsy )C775>>'R'',//G3776??GaGGN33Mswwx7H7HMA O OOr)rr r rr r )rr r rrr r r )rrrr r r )rrr r )r"rr r)r"r r r )r"rr rMr')r"rrQrRr r)r r_)rcrMr r[)rr\r r )r r)__doc__ __future__rrtypingrrrrrr rr!rrIrLrPrWrir|rrrr(rrrs  #""""" ,,,,,,,,,,,,"B"B"B"BJMMMM&!!!!AAAA *OOOO     H&     GGGG&OOOOOOr