- Handle failures during HTML, Markdown, and link serialization.
- Return MCP internal errors when result serialization fails.
- Refactor resource reading logic for better clarity and consistency.
Replace the hardcoded stub with a working implementation that stores
registered scripts and evaluates them in each new document.
Changes:
- Add ScriptOnNewDocument struct and storage list on BrowserContext
- Store scripts with unique identifiers when addScript is called
- Evaluate all registered scripts in pageNavigated, after the execution
context is created but before frameNavigated/loadEventFired events
are sent to the CDP client
- Add removeScriptToEvaluateOnNewDocument for cleanup
- Return unique identifiers per the CDP spec (was hardcoded to "1")
Scripts are evaluated with error suppression (warns on failure) to
avoid breaking navigation if a script has issues.
This unblocks CDP clients that rely on auto-injected scripts (polyfills,
monitoring, test helpers) persisting across navigations. Previously
clients had to manually re-inject after every Page.navigate.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Fixes issues reported by polyglot-scanner:
- Removed explicit `return` keywords and trailing semicolons to resolve DEAD_CODE/DEAD_BRANCH warnings.
- Replaced `epoch::advance().unwrap()` and `stats::resident::read().unwrap()` with safer alternatives (`drop` and `unwrap_or(0)`) to resolve ERROR_SWALLOW warnings.
- Replaced `let _ = Box::from_raw(...)` with `drop(Box::from_raw(...))` to correctly drop the box while fixing the ERROR_SWALLOW warning.
This fixes a bug in MCP where interactive elements were not assigned
a backendNodeId, preventing agents from clicking or filling them. Also
extracts link collection to a shared browser module.
Add `Page.reload` to the CDP Page domain dispatch. Reuses the existing
`page.navigate()` path with `NavigationKind.reload`, matching what
`Location.reload` already does for the JS `location.reload()` API.
Accepts the standard CDP params (`ignoreCache`, `scriptToEvaluateOnLoad`)
per the Chrome DevTools Protocol spec.
The current page URL is copied to the stack before `replacePage()` to
avoid a use-after-free when the old page's arena is freed.
This unblocks CDP clients (Puppeteer, capybara-lightpanda, etc.) that
call `Page.reload` and currently get `UnknownMethod`.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Only look for @ within the first part of the url (up to the first possible
separator, i.e /, # or ?). This fixes potentially incorrect (and insecure)
getOrigin and getHost, both of which use the new helper.
Also make port parsing IPv6-aware.
This causes v8 to verify the receiver of a function, and prevents calling an
accessor or function with the wrong receiver, e.g.:
```
const g = Object.getOwnPropertyDescriptor(Window.prototype, 'document').get;
g.call(null);
```
A few other cleanups in this commit:
1 - Define any accessor with a getter as ReadOnly
2 - Ability to define an accessor with the DontDelete attribute
(window.document and window.location)
3 - Replace v8__ObjectTemplate__SetAccessorProperty__DEFAULTX overloads with
new v8__ObjectTemplate__SetAccessorProperty__Config
4 - Remove unnecessary @constCast for FunctionTemplate which can be const
everywhere.
