Skip to content

1.0.5

Latest

Choose a tag to compare

@github-actions github-actions released this 20 Nov 14:02
· 36 commits to main since this release
8f220d5

Behavior changes

  • PandaScript: page.evaluate() now returns a promise, like page.goto(). Saved scripts need await in front of every page.evaluate(...). Several evaluates can now run at once with Promise.all.
  • --v8-gc-hint-floor-mb is renamed --v8-gc-request-min-mb. The old name still works but is deprecated.
  • New --v8-gc-auto-min-mb, default 32: V8 doesn't start a full GC on its own below this heap size. 0 keeps V8's default.
  • classify, acceptCookies and login can use Gemini, but only with LIGHTPANDA_DECISION_PROVIDER=gemini. TypeSafe stays the default.
  • agent --save replays a read-only script once before writing it. If the replay fails or returns nothing, the model gets one turn to fix the script, and the first script is kept, with a warning, when the fix doesn't replay. Scripts that act on a page are written unchecked.
  • A fetch or XHR whose connection drops before the end of a Content-Length or chunked body now fails, instead of resolving as a complete 200.
  • history.pushState/replaceState to a URL of another origin throws a SecurityError.
  • Responses served from the HTTP cache now go through CORS checks.
  • document.currentScript is null while a module runs.
  • The goto timeout (script API and MCP tool) now bounds the navigation wait, even while a parser-blocking script is loading.
  • Mouse hit testing skips inert subtrees.

Layout

  • More realistic top and height, with awareness of positioned boxes, tables and text
  • display: grid support
  • em, rem and basic calc() lengths; min-height and max-height
  • Text wrapping uses the computed width
  • Hit testing: positioned boxes are above the normal flow, negative z-index below it
  • Layout, hit testing and bounding rects are much faster (style caches are cleared only where a change can reach)

New Web APIs

  • NavigateEvent.intercept()
  • ReadableStream.tee() and values()
  • XMLHttpRequest.upload progress events
  • Attr.specified

Web API behavior / fixes

  • XHR: re-opening a request in flight doesn't fire abort/loadend; responseType = 'document' is ignored in a worker
  • WebSocket close() while connecting moves to CLOSING and queues a close event
  • Request keeps the case of methods other than the six standard ones
  • fetch with redirect: 'manual' keeps the URL
  • The formdata event gets its own copy of the FormData
  • ResizeObserver starts from a size of -1×-1
  • IndexedDB: a pending request keeps its JS object until its event fires (fixes Zone.js apps)

CDP

  • Playwright can use the default context before any page: context.addCookies() works, and page.close() on the startup page no longer hangs
  • DOM.getBoxModel returns Chrome's error for an element under display: none
  • Accessibility.getFullAXTree reports the selected option's label as a combobox value
  • LP.dump withFrames option: include iframe content in the HTML dump
  • Crash fix in request interception on an aborted transfer

WebDriver

  • /alert commands answer no such alert
  • Element /displayed, /computedrole and /computedlabel
  • Shadow root endpoints: /element/{id}/shadow, /shadow/{id}/element, /shadow/{id}/elements
  • Sessions accept acceptInsecureCerts

BiDi

  • script.addPreloadScript and script.removePreloadScript (no arguments yet)

Tools / MCP / Agent

  • lightpanda mcp --tools goto,tree,... exposes only the named tools; --list-tools prints every name
  • MCP over HTTP: memory is returned to the OS when a session ends, and the idle default session closes its pages after --http-session-timeout
  • PandaScript page.getCookies()
  • lightpanda run and agent now apply --cookie and --cookie-jar
  • Markdown: definition list terms and details go on their own lines

Performance / memory

  • Stylesheets with --load-resources stylesheet are preloaded in parallel
  • Smaller CSS DOM: rules are created only when JS reads them, @media/@layer/@supports parsed once
  • Smaller nodes and elements (a div goes from 80 to 64 bytes)
  • The page HTML is held once instead of three times (CNN peak memory 65 MB → 54 MB)
  • fetch() takes the response body instead of copying it
  • Semantic tree uses less memory (HTML spec: 249 MB → 154 MB)
  • Faster getComputedStyle, nodeList[i] and object wrapping