Colors: oklch()/oklab() and color-mix(); linear-gradient backgrounds.
Grid: named-line placement, repeat() expansion, drop whitespace-only
grid items; upgrade taffy 0.7 -> 0.12 so aspect-ratio images in
auto-height rows no longer collapse to slivers.
CSS: @layer, nesting, @property, :is()/:where(), clamp(), @media not all;
source/track/param/area and svg desc/metadata as display:none.
Images: webp + intrinsic sizing, srcset/picture, browser UA + Accept for
CDN fetches, grey placeholder on failure; width/height HTML attributes
rank below author CSS with aspect-ratio preserved for auto axes.
End-to-end rasterization behind the render feature (which now enables
obscura-render/paint):
- obscura-render: screenshot_png(dom, viewport) paints and encodes to PNG
via tiny-skia's encode_png (no extra dep).
- obscura-browser: Page::screenshot renders the live DOM at a viewport.
- CLI: obscura fetch --screenshot out.png writes the settled page to a PNG.
- CDP: Page.captureScreenshot returns the PNG base64-encoded, so
Puppeteer/Playwright page.screenshot() works. captureSnapshot still
errors (it is a DOM/layer snapshot, not an image).
Verified: both paths produce correct 1280x720 RGBA PNGs with elements
laid out and colored as the CSS specifies (block stacking, backgrounds,
white page). Default build unchanged; obstacle course 33/33.
Text, borders, and images are not yet painted; the box layout and
background colors are rendered.
Connection threads are detached and never joined, so `save_cookies()` ran on
the main thread while they were still serving: a connection writing a
Set-Cookie after that point lost it, and the process then exited and killed the
thread mid-flight. Before the per-connection move this could not happen -- the
single processor saved on its own way out, ordered against all connection work
on one LocalSet. Draining on the live-connection counter restores that ordering.
Scope, honestly: this bounds the loss window rather than closing it. The drain
does not finish early today because a connection's frame reader is parked in a
socket read and never observes `notify_waiters`, so a client that holds a
connection open costs the full 3s deadline and then gets the warning. Making
readers wake on shutdown is a separate change and not #430's business.
Measured: 2ms to shut down with no connections (unchanged), 3005ms with one
held open, against the 10s `docker stop` allows before SIGKILL. Cookies are
persisted in both cases.
Refs #430, #333.
Each connection runs on its own OS thread with its own V8 isolates, so
connection count is what drives the server's thread footprint.
MAX_PENDING_WS_HANDOFFS bounds only the handoff queue: once a stream is handed
off, nothing limited how many could be live at once, and a client could grow
the server without limit.
Reserve a slot in the accept loop and release it from a guard that runs when
the connection thread unwinds, on every path including panics and the early
returns before the runtime is built -- a plain decrement at the end of the
closure would leak slots until the cap wedged the server shut. Connections
past the limit get an explicit 503 with X-Obscura-Reason: max-connections
rather than a bare reset, so a client can tell a refusal from a crash. The
refusal happens in the WS dispatch loop, so the HTTP control plane (#62) stays
reachable at the cap.
Default 128, matching the handoff bound and well above any real client fan-out
(Playwright/Puppeteer use one connection per browser). At the cap, with arenas
capped by the previous commit: 146 threads, 33.2 GiB of reserved address space
and 51 MiB resident -- nearly all of that reservation being V8's process-wide
sandbox, which is present at zero connections.
Threads are the honest cost here, not memory: at 100 connections each running
JS this server reserves 83.5 GiB against `main`'s 83.6 GiB, because `main`
pays the same V8 sandbox as soon as it runs any JS. What it does not pay is
118 threads against 17.
The test covers all three properties, and specifically that a closed
connection frees its slot: a leak there would only surface as a server that
silently stops accepting after N lifetime connections.
Refs #430.
glibc hands each new thread its own 64 MiB arena (up to 8x cores). With one
thread per connection that arena is the dominant per-connection memory term,
and freed chunks stay reserved in it rather than returning to the OS: resident
memory tracks the *peak* number of concurrent connections and never comes back
down, which reads as a leak although nothing is leaked and no thread is stuck.
Measured with reliability/conn-scale.py in obscura-benchmark. 100 connections
each running JS reserve 90.0 GiB of address space uncapped versus 83.5 GiB
capped; per idle connection, 65 MiB of reserved address space drops to 2.0 MiB.
`main` at that same 100-connection JS workload reserves 83.6 GiB, so with the
cap this server is level with it -- most of the total is V8's process-wide
sandbox, which `main` pays too the moment it runs any JS.
The resident-set effect is the one that bites in production. Measured in the
container image against Google Maps, four concurrent connections per round:
350 / 619 / 826 MiB over three rounds uncapped and still climbing linearly,
versus 166 / 235 / 269 MiB capped, on a decelerating curve.
Two arenas cost no measurable throughput here -- 8 concurrent connections x 12
navigations, 1.53s uncapped against 1.50s capped over three runs. V8 allocates
the JS heap through its own allocator and the Rust side is dominated by network
I/O rather than malloc traffic. Applied only from `serve`, which owns the
process, and skipped when the caller has already set MALLOC_ARENA_MAX.
Refs #430.
- Delete the now-unused CdpContext::new_full (its only caller switched to
new_with_shared_context); have _new_inner build the shared BrowserContext and
delegate, so the struct's field list lives in one place.
- Extract drive_module_eval, shared verbatim by load_module and
load_inline_module.
- Update comments that still described the deleted process-wide V8 lock and the
single-LocalSet model to the per-connection reality, drop the dangling
reference to the removed obscura-js/src/v8_lock.rs, and soften the
ISOLATE_CREATE_LOCK note now that the main-thread warmup is the real guard
against the InitializeBuiltinJSDispatchTable segfault.
Refs #430.
With every connection on its own thread (previous commit), the process-wide V8
lock served no safety purpose across connections yet forced all of them to
serialize on one mutex, so throughput did not scale with concurrency. Move the
lock into CdpContext so each connection has its own: a connection's nav task and
command dispatch still coordinate on one page's isolate at a time on its thread,
but different connections no longer block each other. Delete the now-unused
process-wide obscura_js::v8_lock.
Measured on a local CPU-bound fixture, page throughput stays flat (~75/s from
concurrency 1 to 8) with the global lock and scales to about 2.5x at concurrency
8 with the per-connection lock, with no crash at any concurrency. Obstacle
course 33/33, obscura-cdp 76/76.
Refs #430.
The CDP server ran every connection's pages on one LocalSet thread, so two
pages' V8 isolates shared an OS thread. Under concurrent navigation a cancelled
event-loop future could leave one isolate entered while another page's
execute_script ran, tripping V8's heap->isolate() == Isolate::TryGetCurrent()
check and aborting the process (#430). No single-thread serialization fixes
this reliably: the cancellation-leak pattern is pervasive and partly inside
deno_core.
Give each WebSocket connection its own OS thread with its own current_thread
runtime, LocalSet, and CdpContext, so every page's isolate is confined to one
thread. V8's invariant is per-thread, so two connections' isolates can never
collide. This is the "Option 2" fix named in obscura-js/src/v8_lock.rs, and it
also lets connections run in parallel instead of serializing on one thread.
- CdpContext::new_with_shared_context reuses one Arc<BrowserContext> across all
connections, so they share one cookie jar and one HTTP client.
- A single signal watcher flips the accept flag and wakes every connection's
processor via notify_waiters(); the accept side persists cookies once on
shutdown.
- Serialize isolate construction with a process-wide lock: the leaptiering
JSDispatchTable setup is not safe to run from several threads at once.
- Warm up V8 once on the main thread before any connection thread creates an
isolate; creating the first isolate off the main thread segfaults in
InitializeBuiltinJSDispatchTable.
obstacle course 33/33, obscura-cdp 75/75 including the concurrent-navigation and
control-plane tests. Concurrency 2 and 4 against heavy pages no longer abort.
Fixes#430.
The keyDown/keyUp paths interpolated `key`/`code` into the generated
KeyboardEvent snippet with single-quote escaping only. Chrome sends
`key: "\"` (U+005C) for the backslash key, which became `key:'\'` — the
backslash escaped the closing quote, ran the string literal on, and made
the whole page.evaluate a syntax error, so the keydown was silently
dropped. Any key/code carrying a backslash broke dispatch.
Escape backslash before single-quote (backslash first, so the quote-escape
backslashes are not themselves doubled), matching the sibling text path in
the same function. Adds a regression test asserting the backslash key
arrives at a keydown listener alongside an ordinary key.
Requests made by page JS (fetch/XHR/dynamic script and stylesheet loads)
run in the V8 op layer and never produced a NetworkEvent, so CDP emitted
Network.requestWillBeSent / responseReceived only for the static
navigation subresources. Puppeteer/Playwright page.on('request') and
on('response') therefore missed every XHR/JSON response, which breaks
response capture (and is the root cause of the Aviasales half of #394).
op_fetch_url now records a JsNetworkEvent keyed by the same fetch-{N} id
under which it already stores the response body. The Page drains these
into its network_events before the CDP layer emits, so they flow through
emit_navigation_events as requestWillBeSent / responseReceived /
loadingFinished alongside the document subresources, and
Network.getResponseBody resolves for them via the shared id. Applied to
both navigation paths (server process_navigation and in-process
do_navigate). The queue is capped at 4096 events per page.
Closes#406.
Runtime.executionContextsCleared on navigation told clients every prior
context was gone, but valid_context_ids was insert-only, so a
Runtime.callFunctionOn targeting a pre-navigation context id still ran
instead of being rejected with Chrome's 'Cannot find context with
specified id'. On a long-lived CDP connection the set also grew by one
id per navigation and was never pruned.
Clear valid_context_ids when executionContextsCleared is emitted, then
re-register the default world (id 2) and the isolated worlds as their
executionContextCreated events are emitted. Closes#407.
* fix(net): default cookie Path to the request directory (RFC 6265 5.1.4)
When a Set-Cookie carried no Path attribute, the jar stored the full request
path as the cookie's path. RFC 6265 5.1.4 default-path is the request URI's
directory — the path up to but not including the right-most '/'. Storing the
full path scoped a session cookie to the exact URL that set it: a cookie set on
GET /app/login was stored with path=/app/login and then, because path-match
(5.1.4, fixed earlier) requires a '/' boundary, was not sent to /app/dashboard.
The user appeared logged out on the very next navigation.
Add default_cookie_path() implementing the RFC algorithm and use it as the
initial path in both set_cookie (HTTP) and set_cookie_from_js (document.cookie).
An explicit Path attribute still overrides it. Tests cover the algorithm and the
end-to-end 'cookie set on /app/login is sent to /app/dashboard' path for both
the HTTP and JS entry points; both failed before this change.
* fix(cdp): apply RFC 6265 default cookie Path to CDP-injected cookies
Network.setCookie/setCookies (the Puppeteer page.setCookie and
Playwright addCookies surface) derived the cookie Path from the full
request URL when the caller omitted path, the same bug the parent
commit fixed for Set-Cookie responses. Reuse default_cookie_path (now
pub) so a cookie set under /v1/things scopes to /v1 and is resent on
sibling paths. Fixes the CDP half of #401.
Correct the test that locked in the old full-path behavior and add a
case confirming an explicit Path still wins.
---------
Co-authored-by: SGavrl <[email protected]>
* fix(js): HTMLFormElement.submit() must not fire a cancelable submit event; add requestSubmit()
Per the HTML standard, the submit() method submits without firing a submit
event and cannot be canceled by a page submit listener; only requestSubmit()
and user-initiated submits fire the cancelable event. bootstrap.js dispatched a
cancelable submit event from submit() and aborted navigation when canceled,
which breaks the common pattern where a submit listener preventDefault()s the
native submit and a callback later calls form.submit() to actually send it
(e.g. an invisible-reCAPTCHA data-callback): the listener also canceled the
programmatic submit, so no navigation was ever issued.
- submit(): navigate unconditionally, no submit event
- requestSubmit(): fire the cancelable submit event, navigate only if not canceled
- submit-button clicks route through requestSubmit() (a real click fires the event)
Adds a regression test; full obscura-cdp suite (56 tests) passes.
* fix(cdp): route synthetic submit-button clicks and Enter through requestSubmit
A CDP Input.dispatchMouseEvent on a submit button and an Enter key in a
form field are user-initiated submits, so per the HTML spec they must
fire the cancelable submit event. input.rs called form.submit() (the
method) directly, bypassing any preventDefault() submit listener, which
mirrors the bootstrap.js bug fixed in the parent commit. Use
requestSubmit() with a submit() fallback so the listener can veto.
---------
Co-authored-by: SGavrl <[email protected]>
server.rs logged previews of attacker-controlled strings with raw byte
slicing: `&text[..text.len().min(200)]` for a malformed CDP frame (inside a
`warn!`, which is active at the worker's default log level) and the same for
an intercepted event JSON. `&s[..n]` panics when byte `n` falls inside a
multi-byte UTF-8 character, so a single WebSocket frame that is invalid JSON,
>=201 bytes, with a multi-byte char straddling byte 200 panics the CDP
processor task — an unauthenticated one-frame crash. page.rs had the same
pattern (behind --verbose debug logging) for evaluated JS expressions.
Add truncate_on_char_boundary (str::floor_char_boundary is still unstable) and
route all four sites through it. Regression test constructs the exact 'a'*199 +
'€' input, asserts the old raw slice panics (via catch_unwind) and the helper
truncates on the boundary instead.
DOM.getDocument{depth:-1} (the standard full-tree call from Puppeteer/
Playwright/browser-use) arrived as u32::MAX and drove serialize_node to
recurse once per DOM nesting level. A page can nest nodes tens of thousands
deep (trivially scriptable; html5ever caps no generic nesting), so the
recursion overflowed the stack — a SIGABRT that catch_unwind cannot recover,
crashing the whole worker (AGENTS.md: "one page must never crash a worker").
Deriving the builder alone is not enough: the response is inherently nested
JSON, and serde_json's own serialization and the Value's Drop recurse over
that nesting too, overflowing the ~2 MiB tokio worker stack. So this both
rewrites serialize_node as an explicit heap worklist and clamps the effective
depth to MAX_SERIALIZE_DEPTH (256), mirroring DOMSnapshot's MAX_NODES guard.
Real DOMs are far shallower; clients wanting a deeper subtree re-request it
with DOM.requestChildNodes / describeNode.
Adds a regression test that builds a 50k-deep tree and serializes with
depth:-1; it SIGABRTs before this change and passes after (run on nextest's
2 MiB test thread, mirroring production).
* fix(js): fire trusted input/change from setFileInputFiles
DOM.setFileInputFiles dispatched plain input/change events, so
event.isTrusted was false. Upload flows and anti-bot code that gate their
change handler on isTrusted ignored them, which silently broke the exact
case the feature targets. The Input domain already marks its synthesized
clicks/keys trusted via __obscura_markTrusted; do the same here.
Fixes#384.
* fix(cdp): bound IO stream buffers so downloads cannot leak
Fetch.takeResponseBodyAsStream moved a response body out of the page's
LRU-bounded cache into CdpContext.io_streams, a server-lifetime map freed only
by an explicit IO.close. A client that opens a stream and disconnects, errors
mid-download, or never closes it leaked the full buffer forever; no teardown
path (WS close, Browser.close, Target.closeTarget) touched io_streams. That
reintroduced the unbounded memory growth streaming was meant to avoid (#360).
Replace the raw map + counter with an IoStreamStore that caps the number of
open streams and their total bytes (OBSCURA_IO_STREAM_MAX_ENTRIES / _BYTES),
evicting the oldest first and always keeping the just-opened stream. Reading an
evicted handle fails cleanly, which is the right trade against an OOM. Read
chunking is unchanged.
Fixes#386.
Network.getResponseBody returns the whole body in one base64 blob and the
cache drops anything over 2 MiB, so large downloads (images, PDFs,
archives) either failed or forced the long-running server to hold the
full body in memory (issue #360).
Add Fetch.takeResponseBodyAsStream + IO.read + IO.close. The body is moved
out of the page cache into a stream handle and delivered in bounded chunks
via IO.read, then freed on IO.close, so it is held once and released after
the download. Binary bodies are byte-exact (stored base64). To stream a
large body, raise OBSCURA_NETWORK_BODY_BUFFER_BYTES so it is cached.
This is entirely opt-in: nothing runs unless a client calls the new
methods, so the default fetch/render/scrape path is unchanged (verified:
obstacle course 33/33, obscura-vs-Chrome head-to-head within noise of the
pre-change baseline).
Fixes#360.
handle_fetch_resolution built InterceptResolution::Continue with url, method,
headers, and body hardcoded to None, so Fetch.continueRequest silently sent the
request unmodified. Playwright route.continue({ headers, url, postData }) and
Puppeteer request.continue reported success while their overrides were dropped.
op_fetch_url already applies all four (and re-validates a rewritten URL through
the SSRF gate); only the CDP layer was discarding them. Read url, method,
headers, and postData from the continueRequest params and pass them through.
Fixes#365.
The cdp_processor decided whether to take the spawn-and-defer navigation path
with `text.contains("Page.navigate")`, which also matches
`Page.navigateToHistoryEntry`. goBack / goForward (both Puppeteer and
Playwright drive Page.navigateToHistoryEntry) were misrouted into the raw
navigate path, which reads an absent `url` param and fails with
"Invalid URL: relative URL without a base", so history navigation never
reached its real handler. Any frame merely embedding the literal
"Page.navigate" (e.g. a Runtime.evaluate expression) was misrouted the same
way.
Match on the parsed CDP method instead, so only an exact Page.navigate takes
the navigation path.
Fixes#363.
Puppeteer's ElementHandle.uploadFile and Playwright's setInputFiles drive
an <input type=file> through CDP DOM.setFileInputFiles, which obscura
rejected as an unknown method, so any upload flow failed (issue #359).
Implement it: resolve the node, read each local file, and hand the bytes
(base64) to the JS layer, which builds real File objects (backed by the
bytes, so page code can read them via FileReader or upload them via
fetch/FormData) and fires input+change like a genuine selection. The file
input exposes a FileList via .files, and .value reads C:\fakepath\<name>
to match Chrome.
Fixes#359.
Typing over CDP into a React controlled input never fired onChange, so
the value never reached component state (issue #324). Two root causes,
both needed:
1. Value tracker. React/Preact/Vue install a value tracker by redefining
`value` on the element instance to record the last value they wrote.
obscura's CDP typing assigned `el.value = ...` directly, which runs
that wrapper and keeps the tracker in sync, so React saw no change on
the input event. Synthetic value/backspace writes now go through a new
__obscura_setFieldValue helper that writes via the element's prototype
setter, bypassing the tracker. The tracker stays stale, so the input
event reads as a real edit.
2. Event handler feature detection. React decides whether to use the
modern `input` event via `("oninput" in document)`. obscura exposed the
GlobalEventHandlers on* attributes only on window, so that check failed
and React fell back to a legacy change-detection path that ignores
input events. The on* attributes are now also present on Document and
Element, matching the HTML spec (GlobalEventHandlers is mixed into
Window, Document, and HTMLElement).
Verified end to end over CDP against the real React 18 UMD build: typing
into controlled inputs updates state, the component re-renders, and a
submit handler reads the captured values.
Closes#324.
The DOMSnapshot walk recursed one frame per nesting level. The MAX_NODES
cap bounds the node count but not the depth, so a deeply nested linear
chain (script can build thousands of nested elements) could overflow the
stack before the count guard triggered. Rewrite it with an explicit stack
so depth lives on the heap. Pre-order traversal and the MAX_NODES cap are
preserved.
Fixes#341.
These Chrome CDP methods logged "Unknown method" warnings. obscura runs
one page per session and does not download to disk, so there is nothing
to do, but clients that call them should get a success ack instead of a
warning. Added as no-op stubs.
Refs #340.
Navigating straight to an image, PDF, or other resource could not read the
main response body over CDP, and Page.frameNavigated always reported
text/html.
Two problems: the main resource's body was stored under its internal
request id, but the client sees it as the navigation loaderId (Chrome's
requestId === loaderId convention), so Network.getResponseBody(loaderId)
missed. And binary bodies were stored as UTF-8-lossy text, corrupting the
bytes. Alias the stored body to the loaderId, store non-text-like content
types (images, PDF, octet-stream) base64, and report the response's real
Content-Type in frameNavigated. HTML pages are unchanged.
Fixes#340.
serve --storage-dir saved cookies only on Ctrl-C (SIGINT). docker stop
and kill send SIGTERM, which was not caught, so the save_cookies() call
at shutdown never ran and cookies were lost on restart. Watch SIGTERM as
well as Ctrl-C on Unix.
Backspace always trimmed the last character and typing always appended,
ignoring any selection. So "select-all, then Backspace to clear" left
all but one character, and typing over a selection accumulated text
instead of replacing it.
Backspace now deletes the selected range when one is set, and typed
text replaces it, matching the browser. selectionStart is null during
ordinary typing, so the previous append/trim behavior is preserved when
no selection is tracked.
Pairs with the triple-click select-all in #320 to make the common
"triple-click, Backspace, retype" clear pattern work end to end. Also
fixes the same pattern when a page sets the selection via select() or
setSelectionRange().
Passing detail:clickCount to MouseEvent is necessary but not sufficient
for triple-click to select all text. The select-all-on-triple-click is a
browser native behavior that synthetic MouseEvent dispatch doesn't trigger.
Add manual selection after click dispatch: when clickCount >= 3 and the
target is an INPUT or TEXTAREA, set selectionStart=0 and
selectionEnd=value.length.
Added Network.setBlockedURLs dispatch handling.
Stores blocked URL patterns on each page.
Applies blocking to page subresources and JS fetch / XHR requests.
Supports wildcard CDP-style patterns like *://*.gstatic.com/*.woff2.
Bug 1: clickCount parameter was read with underscore prefix and silently
discarded. MouseEvent constructor never received the detail field, so
triple-click (clickCount:3) was always treated as a single click by the
DOM — inputs never selected their full text.
Fix: rename _click_count to click_count, pass it as detail in both
mousedown and click MouseEvent constructors.
Bug 2: page.type() intermittently dropped the last character because
the char handler's quote-escaping order was wrong (single-quote before
backslash), and the V8 event loop wasnt pumped after dispatching the
input event.
Fix: fix escaping order (backslash first, then single-quote), and add
page.settle(50) after the char handler's evaluate to pump the event
loop.
isTrusted was hardcoded to true for every Event, so new Event('x').isTrusted returned true. Real browsers return false for script-constructed events and true only for user-agent dispatched ones, so this was a trivial bot-detection tell (Closes#303).
Track trusted events in a closure-private WeakSet that page JS cannot read or forge; default isTrusted to false. The CDP input pipeline marks its synthesized mouse/key events trusted via a non-enumerable __obscura_markTrusted helper, so automation clicks stay trusted like real Chrome over CDP while page-built events report false.
isTrusted was hardcoded to true for every Event, so new Event('x').isTrusted
returned true. Real browsers return false for script-constructed events and
true only for user-agent dispatched ones, so this was a trivial bot-detection
tell (issue #303).
Track trusted events in a closure-private WeakSet that page JS cannot read or
forge; default isTrusted to false. The CDP input pipeline marks its synthesized
mouse/key events trusted via a non-enumerable __obscura_markTrusted helper, so
automation clicks stay trusted like real Chrome over CDP while page-built
events report false.
CDP clients (browser-use and similar) focus an input via DOM.focus before
typing; Input.dispatchKeyEvent then targets document.activeElement. Without a
DOM.focus handler the focus call errored, keystrokes landed on nothing, and the
field stayed empty. Resolve the node and call its JS focus(), which sets
document.activeElement. Completes the type path for DOM-agent frameworks.
CDP DOM-agent frameworks (browser-use and similar) could connect but saw a
blank page: navigation never refreshed their cached target url/title, and their
DOM build aborted with "Unknown domain: DOMSnapshot".
- Emit Target.targetInfoChanged after navigation (both CDP Page.navigate and
JS-initiated) carrying the post-nav url/title. Strict clients refresh a target
only on this event; it also fixes Puppeteer/Playwright page.url() tracking.
- Implement DOMSnapshot.captureSnapshot, built from the DOM with synthetic
geometry (no paint engine) and plausible computed styles, keyed by the same
backendNodeId as DOM.getDocument. Enough for tag/ARIA/accessibility element
detection; it runs only when a client requests it, so the fetch path is
unaffected. Coordinate clicking still requires a real layout engine.
A hung page (a runaway Runtime.evaluate, a synchronous DOM op) holds the
process-wide V8 lock and wedges every other session forever. The one-shot CLI has
a process-level hard deadline for this; the long-running server cannot force-exit,
so terminate just the offending isolate instead.
dispatch() arms a watchdog on the session page's isolate after taking the V8 lock
and disarms it after the handler; on overrun it terminates the isolate and clears
the termination flag so the session keeps working. OBSCURA_CDP_COMMAND_TIMEOUT_MS
tunes the bound (default 60s, 0 disables); navigation already self-bounds well
under it.
The watchdog is a single long-lived shared thread (cdp_watchdog), not a thread
per command: the V8 lock serializes commands so at most one is ever armed, and
arm/disarm are a mutex plus notify with no measurable per-command cost. Also
fixes a lost-wakeup in the per-call watchdog (spawn_watchdog) where disarming a
fast call before the watchdog thread started blocked for the full budget.
Verified: a hung Runtime.evaluate is terminated at the budget and the session
recovers; CDP integration 37/37, obstacle course 33/33; per-command latency
unchanged (watchdog on equals off within noise).
* feat(api): add obscura-api crate with Browser + Page API
Adds a Rust library crate exposing Obscura as a programmable browser:
- Browser: build with config (proxy, stealth, UA, storage_dir)
- Page: goto, url, evaluate, content
- BrowserBuilder with fluent API
- examples/basic.rs
Part of #169
* feat(api): add Element API + CookieStore + wait_for_selector
- Element: text(), attribute(), click() via JS evaluation
- CookieStore: set/get_all/get_for_url/save_to_file/load_from_file
- Page::wait_for_selector with 100ms polling
- Page::query_selector for single element lookup
- Updated examples/basic.rs with full interaction demo
Part of #169
* feat(api): add Browser::cookies() method
* refactor(api): move to obscura crate with feature flag
- Renamed obscura-api → obscura (per maintainer feedback #169)
- API enabled by default via 'api' feature flag
- Usage: cargo add obscura (default=api) or cargo add obscura --features api
- Removed separate obscura-api crate
* chore: exclude docker workflow from fork
* fix(api): make the obscura crate build, run, and document install correctly
- examples/basic.rs and the lib.rs doctest imported `obscura_api`, but the crate
is named `obscura`, so both failed to compile. Use the real crate name.
- wait_for_selector / query_selector read the node id with `Value::as_u64`, but
`evaluate` returns JS numbers as f64, so it was always None and every
wait_for_selector timed out. Accept an integer or a non-negative finite float.
- restore .github/workflows/docker.yml, which the PR removed (a fork-specific
change that should not land upstream).
- add crates/obscura/README.md with honest install instructions: the crate is not
on crates.io, so depend on it via git (and it builds the embedded V8 from
source). The `cargo add obscura` form does not work yet.
Verified: workspace + example build, the lib doctest passes, the example runs end
to end (goto, wait_for_selector, text, attribute), and the integration suite is
28/28 with no regressions.
---------
Co-authored-by: zhangyang-crazy-one <[email protected]>
Co-authored-by: SGavrl <[email protected]>
The LocalSet accept loop blocked on ws_rx.recv() forever because the
detached accept thread kept a ws_tx clone alive, so the channel never
closed and start_with_full_serve_options never returned. cdp_processor
caught Ctrl-C and saved cookies, but that only ended its own task.
Add a shared shutdown flag and Notify. On Ctrl-C cdp_processor sets the
flag and notifies; the accept loop selects on the notify and breaks, so
run_until returns and the process exits. The accept thread checks the
flag between connections so it stops and drops ws_tx on the slow path.
Closes#225
* perf(cdp): TCP_NODELAY + zero-buffer WS writes shave 130ms off newPage
CDP traffic during browser.newPage() and Page.navigate is many small
(~100-byte) request/response frames. Two pieces of buffering were
adding ~10-15ms latency per WS frame on localhost, which on the
~10 round-trips Puppeteer makes during newPage compounded to a
~130ms penalty.
1. TCP_NODELAY on every accepted WS stream. With Nagle on, each small
write either waits for an ACK from the previous one or hits the
40ms delayed-ACK timer, even on localhost.
2. WebSocketConfig.write_buffer_size = 0 on accept_async. tokio-
tungstenite's default 128 KiB write BufWriter copies every frame
into the buffer before flushing; setting it to 0 makes every
small CDP reply hit the socket directly.
Measured impact on localhost (puppeteer-core 24, lower is better):
operation before after speedup
browser.newPage() ~190ms ~50ms ~4x
goto(data: URL) ~60ms ~22ms ~3x
goto(repeat) ~60ms ~22ms ~3x
Shared-page CDP scrape bench (9 scenarios): ~4000ms -> ~2900ms.
28/28 integration tests pass.
* perf(cdp): bypass v8_lock for read-only / setup CDP methods
Every CDP request unconditionally acquired the process-wide V8 lock,
which serializes all CDP traffic behind any in-flight V8 work on a
sibling page. During Puppeteer's browser.newPage() the ~8 setup
calls (Page.enable, Runtime.enable, Page.getFrameTree,
Page.addScriptToEvaluateOnNewDocument, Page.createIsolatedWorld, etc.)
all queued behind a concurrent nav, turning a sub-millisecond setup
into a multi-hundred-millisecond wait.
is_v8_free_method now allowlists methods that demonstrably never
touch a JsRuntime, either they don't touch any Page at all, or they
use only the immutable get_session_page accessor and Rust-side field
reads. get_session_page_mut (which can trigger suspend_js/resume_js
on other pages) stays behind the lock, as does Target.createTarget
(which calls page.navigate).
Allowlisted: most of Target/Browser/Storage, plus Page.enable,
Page.getFrameTree, Page.addScriptToEvaluateOnNewDocument,
Page.createIsolatedWorld, Runtime.enable, Network.* cookie ops,
Fetch.continue/fulfill/fail.
Measured on the puppeteer-core 24 CDP scrape bench:
browser.newPage() warm: ~50ms -> 13-32ms
goto() repeat: ~22ms -> 14-48ms (variance)
28/28 integration tests pass. 18/18 realistic scrape scenarios pass.
* engine: document.currentScript, data: subresources, dynamic module scripts
document.currentScript:
- ScriptInfo now carries the script element's nid.
- Before each script body runs, set globalThis.__currentScriptNid; clear after.
- Document.currentScript getter wraps the current nid. Next.js / Turbopack
reads document.currentScript.src to derive the chunk base path. Without
this the loader saw null, threw 'chunk path empty', and hydration stalled.
- Also threaded through the dynamic-script appendChild path so synchronously
injected <script> elements see their own element as currentScript.
data: URIs as subresources:
- subresource_allowed permits 'data:' alongside http/https. The bytes are
inline in the URI, no network fetch, no SSRF.
- Script-fetch loop short-circuits data: URLs and decodes them locally
instead of routing through the HTTP client.
- Real Chrome allows this by default; Instagram and most Meta properties
ship their bootstrap as <script src=data:application/x-javascript;base64,...>.
Dynamic <script type=module>:
- appendChild SCRIPT handler now accepts type='module' instead of dropping it.
- For external module src, uses dynamic import() instead of fetch + eval.
- For inline modules, wraps the source in a data: URL and import()s that
so module syntax (import/export) parses correctly.
Before / after on previously broken sites (HTML bytes rendered):
vercel.com 0 -> 960 KB
instagram.com 0 -> 669 KB
nextjs.org 0 -> 290 KB
music.apple.com 0 -> 1.88 MB
tailwindcss.com 0 -> 900 KB
linear.app 0 -> 2.4 MB
github.com PR 0 -> 468 KB
28/28 integration tests pass.
* engine: customElements.define upgrades existing matches + connectedCallback
The old customElements was a registry stub: define() recorded the class,
but never instantiated it on matching elements, never called the lifecycle
callbacks, and whenDefined() resolved synchronously with whatever was
already registered (or undefined forever for late-defined components).
New behavior matches the Web Components spec well enough for lit /
MusicKit / Polymer to wire up:
- define(): on register, querySelectorAll the document for matching tags,
copy the prototype methods onto each element instance, call the class
constructor with the element as 'this' for init logic, then fire
connectedCallback if the element is currently in the document.
- whenDefined(): returns a real Promise that resolves when the class is
later defined, not before.
- upgrade(root): walks the subtree, upgrades any matching unupgraded
element. Used by frameworks that mount components asynchronously.
We don't truly subclass HTMLElement per spec (Element is shared across
the engine), so the upgrade copies prototype descriptors onto the
element instance instead of swapping its [[Prototype]]. Behavior is
identical for the common patterns (custom methods, shadow DOM attach
in constructor, connectedCallback render).
28/28 integration tests pass.
* scraping: fix puppeteer page.$, click, page.$$ + form / radio / checkbox actions
These bugs broke every realistic Puppeteer scraping flow. Repro: type into
a search box, click a button, scrape rows. Before this commit the very
first step (page.$('selector')) returned null on every site.
Runtime.evaluate (and the two related wrappers) put the user expression
on a single line as '({expr})'. Puppeteer's bundled injected.js ends with
a '//# sourceURL=pptr:internal' line comment; the comment swallowed our
closing paren and the wrapped function failed to parse with a SyntaxError
that never propagated back. Result: every page.$, page.$$, type(),
hover(), click(), waitForSelector(), getByRole(), etc. silently returned
null. Fix: newline before the trailing ')' so the comment can't eat the
wrapper. Applied to all three wrap paths (eval, store_object_with_meta,
wrap_expression).
documentElement and body now return viewport dimensions for clientWidth /
clientHeight / scrollWidth / scrollHeight / offsetWidth / offsetHeight and
for getBoundingClientRect itself. Puppeteer's #clickableBox clips each
candidate quad to document.documentElement.clientWidth/Height; with the
old 100x20 stub every synthesized element bbox was clipped to empty and
clickablePoint threw 'Node is either not clickable or not an Element'.
Also aligned innerWidth / innerHeight to 1280x720 to match
Page.getLayoutMetrics.
document.elementFromPoint now does real hit testing against the synthetic
bboxes (depth-first, deepest match wins) instead of always returning
<body>. Input.dispatchMouseEvent uses it to pick the click target, which
removed the dependency on a __obscura_click_target side-channel that
the previous getBoundingClientRect call had to set right before the
dispatch.
Element.dispatchEvent now fires inline event handlers
(onclick='...' content attribute and the IDL onclick property) in
addition to addEventListener registrations. Without this, every site
that uses inline onclick / onsubmit / onchange was a no-op.
Input.dispatchMouseEvent's click path now triggers the right default
action by element kind: submit button -> form.submit(); checkbox /
radio -> toggle checked + dispatch 'change'; anchor with href -> nav.
Previously only anchor was handled, so checking a checkbox or submitting
a form by clicking its button did nothing through CDP.
Runtime.getProperties now allocates a stable child objectId for every
object-valued property and stashes the value in __obscura_objects.
Items whose value has a numeric nodeType also get subtype:'node' +
className so Puppeteer's JSHandle.asElement() wraps them as
ElementHandle. Without this, page.$$('selector') returned an array
of plain JSHandles and checkboxes[0].click() raised
'click is not a function'.
Real-world Puppeteer scrape coverage (test_all.py style scenarios):
- HN top stories scrape: fail -> pass
- GitHub repo file rows scrape: fail -> pass
- Wikipedia search input type: fail -> pass
- httpbin form fill + check: fail -> pass
- Cookies persistence across nav: pass -> pass
28/28 integration tests pass.
* scraping: inline scripts run on DCL, JS nav emits CDP events, hit testing
Three remaining gaps from realistic Puppeteer scraping flows:
1. DCL path skipped execute_scripts entirely. Returning early on
waitUntil=domcontentloaded meant every inline <script> was silently
dropped: addEventListener calls in onload bootstraps never ran,
form submit handlers never registered, page.click() hit no-ops.
Spec: DOMContentLoaded fires AFTER scripts. Now scripts run before
the lifecycle flips to DCL regardless of waitUntil.
To keep heavy SPAs from blowing the 25s Puppeteer goto timeout
now that scripts always run, added a soft deadline on the
script-execution phase (default 10s, OBSCURA_SCRIPT_DEADLINE_MS
to override). Both fetch and per-script execution check it.
2. JS-initiated navigation (form.submit, location.assign) did not
emit CDP frameNavigated / lifecycleEvent. Puppeteer's
waitForNavigation hung forever. Runtime.evaluate and
Runtime.callFunctionOn now drain pending_navigation through a
helper that fires the same post-nav event sequence Page.navigate
does.
3. Document.elementFromPoint walked the tree only descending into
ancestors containing (x,y). Our synthetic per-nid bboxes do not
form a containment hierarchy, so a deep <input> inside <label><p>
was never reached and clicks resolved to <body>. Switched to a
flat sweep, deepest match wins (highest nid as a proxy for tree
depth).
Realistic Puppeteer scrape scenarios: 18 of 18 assertions pass across
HN, Wikipedia, GitHub, httpbin forms, cookie persistence, and inline
JS submit handlers. 28/28 integration tests pass.
* scraping: fetch in page context, textarea, select, basic history
Four more gaps from the broad API sweep:
1. fetch() and XHR from page JS hung forever. set_intercept_tx() in
the runtime auto-set state.intercept_enabled = true as a side
effect, so every navigation silently enabled Fetch interception
even when the CDP client never called Fetch.enable. op_fetch_url
then routed every request through the intercept channel waiting
for a Fetch.continueRequest that the client wasn't going to send.
Split the flag from the channel: set_intercept_tx wires the
channel without enabling; new set_intercept_enabled toggles the
flag. Fetch.enable / Fetch.disable now call both, and the nav
path only wires the channel.
2. page.type into textarea lost everything past the first newline.
The dispatchKeyEvent handler skipped text=='\n' entirely, so
'multi\nline' inserted only 'multi'. Also Enter on a textarea
tried to submit the nearest form instead of inserting a newline.
Now Enter inserts \n inside textareas and only submits forms
from input fields, matching Chrome.
3. <select>.value did not reflect the selected option. Generic value
getter returned the select's value attribute (empty) instead of
the matching option. Added select-aware getter that walks options
for the selected one and falls back to the first; setter accepts
page.select() and dispatches a 'change' event.
4. Basic navigation history. Pages now track a visited-URL stack
and history_index. Page.getNavigationHistory returns the real
history; Page.navigateToHistoryEntry honors entryId and emits the
same CDP nav-event sequence as Page.navigate so the response and
lifecycleEvent reach waiting clients. resetNavigationHistory
clears the stack. Push happens after every successful nav (skips
consecutive duplicates so Page.reload doesn't grow the stack).
Broad API sweep across 13 feature groups: 43/44 PASS.
All categories solid (navigation, selectors, typing, waitForSelector,
evaluate, cookies, keyboard, attributes, events, fetch/XHR, storage,
timers). goBack reaches its target URL but Puppeteer client-side
waitForNavigation still hangs on it; the CDP frameNavigated +
lifecycleEvent fire correctly, so the remaining gap is in event
shape rather than logic. 28/28 integration tests pass.
CDP / navigation:
- Route all Page.navigate through process_with_interception so the
parent dispatcher stays responsive during long navs (fixes
Target.createTarget timeouts after a handful of newPage calls).
- Default waitUntil to domcontentloaded. Puppeteer and Playwright
don't pass it server-side and wait client-side via lifecycleEvent,
so the old Load default blocked nearly every nav on slow sites.
- 30s navigation hard ceiling, OBSCURA_NAV_TIMEOUT_MS to override.
- execute_scripts loop now checks wall-clock deadline per iteration.
- Emit Network.requestWillBeSent before Page.frameNavigated so
Playwright's FrameManager attaches the navigation response (#189).
- Use loaderId as the Document request id so page.goto() resolves
to a Response instead of null (#189).
- Fresh executionContextId per navigation, no more 'Cannot find
context with specified id' after the first nav (#192).
- POST navigation params plumbed through the intercept path.
DOM / Web API surface:
- MutationObserver: subtree filter walks parent chain, fires on
textContent / CharacterData.data / innerHTML, type filter respected,
oldValue passed.
- IntersectionObserver: multi-fire (initial + 120/500/1500ms),
disconnect / unobserve / takeRecords, root / rootMargin / thresholds.
- ResizeObserver: per-target firing from getBoundingClientRect, drop
the no-op override that was clobbering the real class.
- crypto.subtle.digest: real SHA-1/256/384/512 via op_subtle_digest.
- DOMParser / document.implementation.createHTMLDocument /
createDocument: detached doc with body, head, documentElement,
querySelector*, getElementById, getElementsByTagName.
- WebSocket / EventSource / BroadcastChannel: real EventTarget,
open fires on microtask.
- indexedDB: request, object store, transaction, database with
async onsuccess and in-memory Map storage.
- getComputedStyle: width/height/top/left from bbox so React
virtualization libs render items.
Site compatibility went from 3/40 to 175/175 reachable sites in
the sweep. Perf is equal or faster than pristine main on every
site tested (DCL default removes the block on long-tail subresources).
28/28 integration tests pass.
* fix(js): coerce URL constructor input/base to string
The polyfill at bootstrap.js called `base.match(...)` directly, which
threw `TypeError: base.match is not a function` whenever callers passed
a Location object (e.g. `new URL(href, window.location)`), a URL
instance, or any other non-string with a sensible toString. Per the
WHATWG URL Standard both `url` and `base` are stringified before
parsing.
Coerce explicitly so the existing regex paths keep working and the
constructor matches real-browser behaviour. AngularJS, jQuery, and
anything that resolves a relative URL against `window.location` now
loads under Obscura.
* fix(browser,cdp): run preload scripts before page scripts
`Page.addScriptToEvaluateOnNewDocument` was executed AFTER the page's
own scripts (and even after the load event), which silently broke
every CDP contract that depends on early injection — most visibly
puppeteer's `page.exposeFunction`, whose user-facing wrapper has to
override `globalThis[name]` before any `<script>` tag in the document
runs.
Move the registered preload sources onto `Page` and execute them
inside `navigate_single`, right after `__documentReadyState__ =
'loading'` and before the parser-discovered scripts. The CDP layer
hands the current preload set to the page via `set_preload_scripts`
before kicking off navigation, replacing the two post-navigate
`execute_preload_script` loops in `domains/page.rs` and `server.rs`.
* feat(cdp,js): emit Runtime.bindingCalled for addBinding shims
`Runtime.addBinding` was a no-op stub that installed an empty
`globalThis[name] = function () { return null; }` placeholder. Page JS
calls were silently swallowed and the Node-side callback registered
via puppeteer's `page.exposeFunction` never fired.
Wire the binding all the way through:
1. New `op_binding_called(name, payload)` op forwards every page-side
call into a `pending_binding_calls` queue on `ObscuraState`.
2. `Page::take_pending_binding_calls` drains the queue.
3. `dispatch()` empties every page's queue after each CDP request and
pushes one `Runtime.bindingCalled` event per entry into
`ctx.pending_events`, so the writer task forwards them to the
client like any other CDP event.
4. The `addBinding` handler installs a real shim that calls the op,
registers it as a preload script (keyed `__obscura_binding__<name>`)
so it survives navigations, and evaluates it now so the binding is
usable on the current document.
5. Implement `removeBinding` to mirror `addBinding`.
The shim coerces non-string arguments via `JSON.stringify`, matching
the wire contract that puppeteer's `exposeFunction` wrapper depends
on (`Runtime.bindingCalled.params.payload` is always a string).
* fix(cdp): emit canAccessOpener in all TargetInfo payloads, add aarch64 cross-compile CI
Merges PR #165 and PR #172.
PR #165: The CDP spec marks canAccessOpener as a required boolean on
TargetInfo but Obscura was omitting it. Strict Rust clients like
chromiumoxide reject the event entirely and panic with "Created target
not present" when the field is absent. Added canAccessOpener: false
to every TargetInfo emitted by setDiscoverTargets, getTargets,
createTarget, attachToBrowserTarget, attachToTarget, and getTargetInfo.
PR #172: Added a cross-compile matrix entry for aarch64-unknown-linux-gnu
running on ubuntu-22.04 x86_64, producing obscura-aarch64-linux-cross.
This is a fallback for when GitHub ARM runners are unavailable without
removing the native ARM build.
* fix(cdp): keep HTTP control plane reachable while V8 JS evaluation blocks (#173)
fix(cdp): unblock HTTP control plane during synchronous V8 execution
Fixes#62.
Previously, the TCP accept loop and CDP processor shared a single-threaded
tokio LocalSet. Long synchronous JS evaluations starved the accept loop,
causing incoming connections to hang in the kernel backlog and effectively
killing the HTTP control plane.
This moves the TCP listener and `/json/*` endpoints to a dedicated OS
thread using `std::net::TcpListener`. HTTP requests are now handled
directly via blocking I/O on this thread, completely bypassing V8.
WebSocket connections are handed off to the LocalSet via a bounded mpsc
channel (capacity 128). If V8 stalls and the queue fills, new WS
connections are dropped (ECONNRESET) rather than buffering unboundedly
to prevent OOMs.
Includes a new integration test to verify the control plane stays
responsive during a 5s JS loop, plus inline documentation clarifying
thread lifetime semantics and a minor test TOCTOU race.
* fix(js,dom): add Storage global, CharacterData ChildNode methods, fix remove/isConnected (#176)
fix(dom): add missing DOM/JS shims for modern frameworks
Closes#163.
Adds missing DOM and JS shims required by modern frameworks (Svelte 5,
Vue, Lit) and consent-management SDKs to execute without errors:
- Define the `Storage` global constructor to resolve `ReferenceError`s
and fix `Storage.prototype` access.
- Apply standard ChildNode methods (`before`, `after`, `replaceWith`,
`remove`) to `CharacterData.prototype`. This prevents `TypeError`s when
frameworks call these on Text or Comment nodes during hydration.
- Update `Element.prototype.isConnected` to perform a genuine parent-walk
rather than returning a hardcoded `true`.
- Fix `Element.prototype.remove` to properly clean up the internal ID
index so removed elements are no longer queryable.
- Remove outdated no-op stubs on the `Element` class that were shadowing
the proper polyfill implementations.
* feat(cli): add --dump assets to inventory page sub-resources (124) (#174)
feat: add `--dump assets` to enumerate page sub-resources
Adds a `--dump assets` flag to output all page-referenced sub-resources
(scripts, stylesheets, images, iframes, media, embeds, objects) as NDJSON,
allowing users to easily replay the asset graph.
- Output format: `{"url":"https://example.test/app.js","type":"script"}`
- URLs are resolved against the page origin (matching `--dump links`).
- The `type` tag is derived from the source element. `<link>` elements map
their `rel` attribute to specific types (stylesheet, icon, manifest, preload,
prefetch, modulepreload, dns-prefetch, preconnect, alternate), falling back
to "link" if missing or unknown.
- Adds 7 tests covering URL resolution, rel-mapping, supported selectors,
and empty-attribute skipping.
- Updates README dump-format table and examples.
* fix(cdp,js): playwright actionability stubs and objectId resolve (#178)
Playwright fails locator actions on Obscura for two avoidable reasons
that are separate from the missing layout engine. Both have trivial
fixes that unblock the actionability polling loop.
- Element.checkVisibility(opts) was undefined. Playwright >= 1.25 calls
it before every input event and retries until its action timeout if
the method isn't a function. Add a stub that returns true. Without a
layout pass we can't compute it for real, but a permissive stub still
beats the undefined-function path.
- ARIA reflection properties were missing on Element (role, ariaLabel,
ariaSelected, ariaExpanded, ariaHidden, ariaChecked, ariaDisabled,
ariaRoleDescription). Playwright's getByRole and getByLabel locators
read them. Reflect through to the underlying aria-* attributes for
both read and write.
- DOM.getBoxModel and DOM.getContentQuads ignored the objectId param.
Playwright passes only objectId (from a prior DOM.resolveNode), and
the handlers silently defaulted to node 0, so dispatchMouseEvent
always clicked the wrong element. Extract a resolve_node_id helper
that honors nodeId, backendNodeId, and objectId in that order.
- Page.captureScreenshot and Page.captureSnapshot now return a clear
"not supported, use a real Chromium for the screenshot leg" error
instead of the generic "Unknown Page method" fallback. Matches the
pattern already in place for Page.printToPDF.
Partial fix for #45. The architectural blocker (no layout engine, so
multi-element click coordinates are wrong) is unchanged.
* fix(mcp): allow browser MCP clients through CORS preflight (#179)
The OPTIONS preflight hardcoded Access-Control-Allow-Headers to just
Content-Type. Any browser MCP client sending mcp-protocol-version (per
the MCP spec) or Authorization / X-API-Key (typical for hosted
deployments) failed the preflight and the browser blocked the actual
request.
Add those three to the allow-list and a 24h Max-Age so browsers don't
re-preflight every call. Adds an integration test that drives the
server with a real preflight request.
Closes#175.
* fix(js,dom): scope element queries, real timers, form/prepend/equal (#180)
- Element.querySelector / querySelectorAll scope to the receiver subtree
via a new query_selector_scoped Rust op, not always document root.
Same for DocumentFragment.
- setTimeout respects its delay argument via a new async op_sleep.
setInterval is a real repeating timer instead of an alias to setTimeout.
- document.forms / images / links return live querySelectorAll results
instead of hardcoded [].
- HTMLFormElement subclasses Element with .elements / .length, and form
nodes are wrapped as HTMLFormElement so instanceof works. submit()
stays on Element.prototype to preserve the existing submit-event
dispatch and form-data navigation path.
- Element.prepend has a real implementation.
- Node.isEqualNode does a structural compare (nodeType, nodeName,
nodeValue, attributes, recursive children) instead of nid identity.
- Fix long-standing insert_before arg-order bug: bootstrap.js was
passing (parent, new, ref) but _dom only forwards two args, so ref
was silently dropped. Pass (new, ref) directly. Same fix in
replaceChild.
Adds 3 selector unit tests and 6 runtime regression tests covering
the JS shim fixes.
Closes#105.
* fix(cdp): unblock puppeteer connect, bound Runtime.evaluate, full TargetInfo (#181)
Three independent gaps that hurt browser-based MCP clients connecting
to obscura's CDP endpoint:
- Runtime.enable returned "No page" when called before a session was
attached. puppeteer-extra's FrameManager.initialize fires it on the
browser-level WebSocket BEFORE any page target exists, so the
standard puppeteer.connect / browser.newPage flow died on connect.
Now reply with {} when no session is attached and let the next
Target.attachToTarget emit executionContextCreated.
- Runtime.evaluate had no per-call timeout. A long synchronous JS
expression held the global V8 lock indefinitely and starved every
other CDP command on the session. Honor the CDP `timeout` parameter
and default to 30s (Chrome's protocolTimeout). Returns a clear error
on expiry so client-side deadlines fire cleanly instead of hanging.
- Target.getTargetInfo for the browser target omitted canAccessOpener.
Strict CDP clients (chromiumoxide) panic when the field is missing.
Fill it in (false for the browser target itself).
Partial fix for #122. The remaining items (networkIdle hard-deadline
on SPAs, /json/list hardcoded data, docs on cheap-vs-queued methods)
need bigger or non-code changes and are deferred.
* fix(net,browser): detect response charset, decode non-UTF-8 pages (#182)
obscura was calling String::from_utf8_lossy on every response body, which
silently corrupts every non-UTF-8 page (GBK, Big5, Shift-JIS, Windows-125x,
EUC-KR, ISO-8859-x). Whole language families came through as replacement
characters.
Add an encoding module (obscura-net::encoding) that picks the decoder
the way real browsers do per HTML5 § 8.2.2.4:
1. HTTP Content-Type charset=... parameter (authoritative when present)
2. <meta charset> or <meta http-equiv=Content-Type> sniffed from the
first 1KB of the body (HTML only)
3. UTF-8 fallback
Use it in:
- obscura-net Response::text (signature changed: String, not Result)
- obscura-browser page.rs HTML body decode (the primary scrape path)
- obscura-browser page.rs JS body decode
- obscura-browser page.rs CSS body decode
For JS / CSS we deliberately skip the meta sniff: those bodies are not
HTML and any '<meta charset=...>' substring inside them is not a charset
declaration.
Adds 7 unit tests for the encoding module covering header precedence,
quoted charset values, meta charset sniff, http-equiv form, the 1KB
sniff window, and a round-trip decode of GBK bytes. Adds a Python
feature test exercising file:// + meta charset and a live HTTP server
serving Shift_JIS with the wrong meta tag.
Closes#113.
* feat(cdp): implement Network.setCookie + deleteCookies, full field fidelity (#170)
* feat(cdp): implement Network.setCookie + deleteCookies, full field fidelity
Closes#168.
Adds the missing Network.setCookie (singular) and Network.deleteCookies
methods and upgrades the existing Network.setCookies/getCookies handlers
so sameSite, expires, httpOnly, secure, and the url-fallback parameter
all round-trip correctly.
Network and Storage cookie handlers share a single parse_cdp_cookie /
parse_delete_cookies_params helper to avoid drift.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* fix(cdp): fall back to default cookie jar, add Network.getAllCookies
Addresses @SGavrl review on PR #170:
- Network.{set,get,delete}Cookie(s)/clearBrowserCookies now resolve the
cookie jar via a shared helper that prefers the session page's jar and
falls back to ctx.default_context.cookie_jar when no target has been
attached. Mirrors how Storage.setCookies already routes when there is
no session, so Puppeteer/Playwright flows that issue cookie methods
before Target.attachToTarget no longer hit "No page".
- Add Network.getAllCookies (separate CDP method per spec); aliases the
existing getCookies handler so both arms return the full jar.
Tests: 6 new #[tokio::test] cases in domains::network::tests cover the
no-session fallback for set/get/setMany/delete/clear and the new
getAllCookies arm. RED on prod-reverted file (6/6 fail with "No page" /
"Unknown Network method: getAllCookies"), GREEN on patched
(22 -> 30 obscura-cdp lib tests, obscura-net 22/22 baseline preserved).
* fix(net): use bundled webpki roots in stealth client (#184) (#185)
set_default_paths only resolves OpenSSL's compile-time CA dirs, which
exist on Linux but are empty on Windows. The stealth client's cert store
ended up empty there and every TLS handshake failed with
CERTIFICATE_VERIFY_FAILED inside BoringSSL handshake.cc.
CertStore::default() uses wreq's bundled webpki-root-certs (Mozilla's
trust list), which behaves the same on every platform and removes the
dependency on a system CA store.
* fix(js): Element.attributes, on* event defaults, per-nid getBoundingClientRect (#45) (#186)
Three DOM-completeness gaps that broke jQuery 1.11.3 init and Playwright
actionability:
1. Element.attributes was undefined, so legacy IE feature-detect code
like div.attributes[eventName].expando threw TypeError. Added a
NamedNodeMap-like backed by a new attribute_names DOM op (length,
numeric index, string name lookup, getNamedItem/setNamedItem,
iterable).
2. Standard on* event handler properties were missing from window, so
('on' + ev) in window returned false and forced jQuery into the
legacy IE branch. Initialised the 79 HTML-spec on* handlers to null.
3. getBoundingClientRect returned the same {x:8,y:8,width:100,height:20}
literal for every element, so Playwright's hit-testing couldn't tell
elements apart. Replaced with a deterministic synthesis from _nid
mapped into a 12-column grid sized to a 1280x720 viewport. Still not
pixel-accurate (no layout engine), but stable across reads and
distinct per node.
* feat: cookie jar persistence + --storage-dir flag + fix HTTP cookie header filtering (#114)
* feat: cookie jar persistence + storage-dir + filtering fix (clean v4)
* fix: resolve build errors, add storage_dir param throughout server chain
* fix: storage_dir end-to-end, plus two test-literal field additions
Five issues found while testing PR 114 end-to-end.
Build fixes (two test literals out of sync with new struct fields):
- crates/obscura-net/src/cookies.rs: CookieInfo literal in
test_domain_matches_subdomain_without_leading_dot was missing the
same_site and expires fields added in #170.
- crates/obscura-cli/src/main.rs: Command::Fetch literal in
matcher_still_uses_fetch_variant was missing the new storage_dir
field. Without these cargo test --workspace would not compile.
run_fetch was branching between BrowserContext::with_storage (which
hardcodes proxy=None, stealth=false, user_agent=None) and
with_options (which has no storage_dir slot), so --storage-dir +
--proxy + --stealth silently dropped the proxy and stealth flags.
Same bug SGavrl flagged on dispatch.rs in the previous round, just on
the other code path. Switched to with_storage_full which carries all
four. Verified the fetch now errors out against an invalid proxy.
run_fetch --eval short-circuited the function before
context.save_cookies() ran, so cookies set during an eval session
were never persisted to {storage_dir}/cookies.json. Moved the save
above the return.
obscura serve loaded cookies on startup but never saved them
anywhere, so a session that set cookies via Network.setCookie or
navigation dropped them on shutdown. Subscribed to Ctrl-C once in
put a single save_cookies() after the loop so both the Ctrl-C path
and the channel-closed path persist cookies. Verified end-to-end.
---------
Co-authored-by: zhangyang-crazy-one <[email protected]>
Co-authored-by: SGavrl <[email protected]>
* fix(cdp): page.goto returns Response over Puppeteer and Playwright (#189) (#190)
Two correlation invariants real Chrome holds were missing, so the
nav Response was never attached to goto().
requestId === loaderId for the main Document event. Both clients
gate isNavigationRequest() on this; we were emitting a synthetic
per-page id.
Network.requestWillBeSent for the Document fires before
Page.frameNavigated. Playwright's FrameManager needs the request
present as pendingDocument before the frame commits, otherwise
currentDocument lands with request=undefined.
Subresource events and the Document's responseReceived /
loadingFinished stay where they were.
Verified end-to-end on both clients against example.com (200),
google.com (200), httpbin/404 (404).
* fix(browser,cdp): run preloads on about:blank, tighten binding shim arity
Two follow-ups from PR #166 review:
- navigate_single's about: branch returned before any preload sources
ran, so addScriptToEvaluateOnNewDocument scripts and the
Runtime.addBinding shim never installed on about:blank. Puppeteer
hits this every session: browser.newPage() lands on about:blank and
exposeFunction is dead until the first real navigation. Now the
about: branch runs the preload set after init_js(), guarded by the
same 5s timeout as regular preloads.
- The Runtime.addBinding shim used JSON.stringify on non-string args
and silently accepted any arity. Chromium's V8InspectorImpl rejects
calls without exactly one argument and ToString-coerces that arg
before emitting Runtime.bindingCalled. Match that: guard on
arguments.length and use String(arg) for the non-string branch.
Single-string calls (the path puppeteer's exposeFunction wrapper
uses) round-trip unchanged.
* fix(cdp): allocate fresh execution context ids per isolated world (#192) (#200)
Page.createIsolatedWorld and do_navigate both hardcoded context id 100
for every isolated world emission. After the second navigation Playwright
expected a fresh incrementing id (101, 102) while obscura kept re-using
100. Runtime.evaluate then failed with 'Cannot find context with
specified id: 101' because the id was never registered in
valid_context_ids and the client's bookkeeping had diverged from the
server's.
Added next_isolated_context_id counter on CdpContext (starts at 100) and
a next_isolated_context() helper that claims an id, registers it in
valid_context_ids, and increments. Both call sites now use it.
* feat(cli): default V8 heap to 4 GB on 64-bit so heavy SPAs don't OOM (#199) (#202)
V8 embedded defaults to ~1.4 GB max-old-space on 64-bit. Pages with
heavy fingerprinting or analytics bundles (e.g. demo.fingerprint.com/
web-scraping) blow past it and the worker dies with
'Fatal JavaScript out of memory' and SIGTRAP, taking the CDP session
with it.
Always set --max-old-space-size=4096 (1024 on 32-bit targets) before
V8 startup, matching headless Chrome's default. User-supplied
--v8-flags are appended after the default so they win (V8 parses
left-to-right and the last value wins), which keeps the existing
escape hatch intact for memory-constrained environments.
---------
Co-authored-by: SGavrl <[email protected]>
Co-authored-by: Ben Younes <[email protected]>
Co-authored-by: zhangyang-crazy-one <[email protected]>
Picks up @Timtech4u's PR #33, which has been waiting since April. The
design (CLI flag + threading through BrowserContext + http client) is
theirs and is correct; this is a clean rewrite on top of current main
because repeated rebase rounds had been reverting unrelated work that
landed in the meantime.
Issue #4 blocked all loopback / RFC1918 / link-local fetches to prevent
SSRF. The only escape was OBSCURA_ALLOW_PRIVATE_NETWORK=1, which is
process-wide and clunky in pipelines. Multiple users (@Timtech4u in
the original PR, then @duartemvix and @necrogami in the comment
thread) reported they cannot use obscura for development against
http://localhost:N without it.
Adds --allow-private-network as a global CLI flag, works on fetch,
serve, and scrape. Threads through BrowserContext::with_storage_and_network
and the http client's validate_url. CdpContext gets a new_full
builder and obscura_cdp::start_with_full_serve_options entry point so
'obscura serve --allow-private-network' lets Puppeteer/Playwright
navigate to localhost too. Default remains blocked.
Backward compat: the env var still works (OR'd with the flag inside
validate_url). All existing constructors and entry points unchanged.
Verified end-to-end against a local http.server on 127.0.0.1: fetch
and CDP Page.navigate both blocked by default, both succeed with the
flag. Public URLs unaffected. Integration suite 28/28, full workspace
cargo test all green.
Page.createIsolatedWorld and do_navigate both hardcoded context id 100
for every isolated world emission. After the second navigation Playwright
expected a fresh incrementing id (101, 102) while obscura kept re-using
100. Runtime.evaluate then failed with 'Cannot find context with
specified id: 101' because the id was never registered in
valid_context_ids and the client's bookkeeping had diverged from the
server's.
Added next_isolated_context_id counter on CdpContext (starts at 100) and
a next_isolated_context() helper that claims an id, registers it in
valid_context_ids, and increments. Both call sites now use it.
SGavrl