@push.rocks/smartpuppeteer 2.6.0 → 2.7.0

package/dist_ts/smartpuppeteer.interfaces.livebrowser.d.ts

@@ -19,6 +19,8 @@ export interface ILiveBrowserScreencastOptions {
     maxOutstandingFrames?: number;
     firstFrameTimeoutMs?: number;
 }
+export interface ILiveBrowserScreencastUpdateOptions extends Pick<ILiveBrowserScreencastOptions, 'quality' | 'maxWidth' | 'maxHeight' | 'everyNthFrame'> {
+}
 export interface ILiveBrowserSecurityOptions {
     denyDownloads?: boolean;
     denyFileChoosers?: boolean;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@push.rocks/smartpuppeteer",
-  "version": "2.6.0",
+  "version": "2.7.0",
   "private": false,
   "description": "Provides simplified access to Puppeteer for automation and testing purposes.",
   "main": "dist_ts/index.js",
@@ -15,7 +15,7 @@
     "@git.zone/tsbuild": "^4.4.2",
     "@git.zone/tsrun": "^2.0.6",
     "@git.zone/tstest": "^4.0.0",
-    "@types/node": "^26.2.0"
+    "@types/node": "^26.4.1"
   },
   "dependencies": {
     "@push.rocks/smartdelay": "^3.0.1",

package/readme.hints.md

@@ -5,6 +5,8 @@
 - CDP is private to the live runtime. Public contracts contain transport-neutral values and `Uint8Array` image data, never `CDPSession`, raw CDP frame IDs, or base64 image strings.
 - Every published screencast frame has one private sequence-to-CDP acknowledgement entry. Public acknowledgement requires matching tab ID, sequence, generation, and viewport revision. `screencast.maxOutstandingFrames` bounds those entries independently of any application transport window; overflow retires oldest-first. Drops and all stream invalidation paths must issue each CDP acknowledgement at most once and await in-flight acknowledgements before detaching the CDP session.
 - `refreshScreencast()` installs its exact-generation waiter before starting CDP, retires the previous stream, and resolves only after the first validated frame is published. Pre-aborted and queued calls reject without changing the stream. Active caller cancellation settles only after restorative restart; shutdown, tab replacement, crash, or closure revokes restart authority and must never resurrect the stream.
+- A refresh timeout or restart failure on a tab that is still restorable is not runtime-fatal. Revoke the restart authority, wait for the in-flight `startScreencast()` to unwind, stop whatever it established, and emit non-fatal `screencast_refresh_failed` with the tab left open, non-streaming, and invalidated. `refreshScreencast()` therefore also accepts a non-streaming active tab; `updateScreencastOptions()` shares that restart core after storing validated options. Browser loss stays fatal and is normally reported by the disconnect listener first.
+- Raw input is dispatched through a per-tab queue. Chromium does not keep a pipelined mouse or key command behind an earlier `mouseWheel` command, so anything queued behind wheel input waits until that wheel CDP call settles. Wheel input arriving while a wheel dispatch is in flight merges into the trailing wheel entry for the same CDP session, generation, and viewport revision; a non-wheel entry ends that batch.
 - Activation, viewport changes, navigation, tab closure, snapshots, observations, semantic actions, and shutdown share one bounded operation scheduler. Raw input and frame acknowledgement remain direct, but must validate active tab, generation, and viewport revision. Repeated internal navigation/load state updates are coalesced per tab, and shutdown cancels queued work.
 - Retain the scheduler-owned launch `AbortController` for the full browser lifetime. Shutdown aborts both the active operation and Chromium itself so a non-signal-aware Puppeteer command or disabled protocol timeout cannot retain the browser ahead of queued cleanup.
 - Viewport revision starts at 1 and advances only after `Page.setViewport()` succeeds. Stop and flush the active screencast before applying a viewport or navigation mutation, then restart it with a new generation.

package/readme.md

@@ -173,20 +173,37 @@ try {
 
 Only the active tab is streamed. Frames carry a session-monotonic sequence, tab/CDP generation, viewport revision, viewport, MIME type, encoded dimensions, screencast metadata, and binary `Uint8Array` data. Every delivered frame must be acknowledged with all four identity fields. A delivered frame remains pending until it is acknowledged, dropped, or retired by the runtime; mismatched, duplicate, stale, retired, or operationally failed acknowledgements return `{ accepted: false }`. Operational acknowledgement failures also emit an `error` event whose code is `frame_acknowledgement_failed`. `screencast.maxOutstandingFrames` is an integer from 1 through `liveBrowserMaxOutstandingFrames` (64) and defaults to `liveBrowserDefaultMaxOutstandingFrames` (3). At capacity, SmartPuppeteer retires and CDP-acknowledges the oldest pending frame before publishing the next one. Tab switches, navigation, resize, page cleanup, disconnect, and shutdown also retire pending frames and await their single CDP acknowledgement attempt before detaching the screencast session.
 
-The SmartPuppeteer outstanding-frame bound covers only the private frame-to-`Page.screencastFrameAck` lifecycle. It is not a binary transport window, and `acknowledgeFrame()` is not an application transport acknowledgement. A higher-level runtime must maintain and bound its application frame window separately. `refreshScreencast()` retires the active stream and resolves with an `ILiveBrowserFrameIdentity` for the exact first validated frame from a new generation. Its `tabId`, `sequence`, `generation`, and `viewportRevision` fields can be passed directly to `acknowledgeFrame()`. `screencast.firstFrameTimeoutMs` bounds both the CDP restart and first-frame arrival, accepts integers from 100 through 60,000, and defaults to 5,000; timeout or another refresh failure emits fatal `screencast_refresh_failed` and shuts the session down. Screencast `format` accepts `jpeg` or `png`, `quality` accepts integers from 0 through 100, `maxWidth` and `maxHeight` accept integers from 1 through 4096 subject to an 8,294,400-pixel combined ceiling, and `everyNthFrame` accepts integers from 1 through 100.
+The SmartPuppeteer outstanding-frame bound covers only the private frame-to-`Page.screencastFrameAck` lifecycle. It is not a binary transport window, and `acknowledgeFrame()` is not an application transport acknowledgement. A higher-level runtime must maintain and bound its application frame window separately. `refreshScreencast()` retires the active stream and resolves with an `ILiveBrowserFrameIdentity` for the exact first validated frame from a new generation. Its `tabId`, `sequence`, `generation`, and `viewportRevision` fields can be passed directly to `acknowledgeFrame()`. `screencast.firstFrameTimeoutMs` bounds both the CDP restart and first-frame arrival, accepts integers from 100 through 60,000, and defaults to 5,000. Screencast `format` accepts `jpeg` or `png`, `quality` accepts integers from 0 through 100, `maxWidth` and `maxHeight` accept integers from 1 through 4096 subject to an 8,294,400-pixel combined ceiling, and `everyNthFrame` accepts integers from 1 through 100.
+
+A refresh timeout or restart failure on an otherwise healthy tab is recoverable. The `refreshScreencast()` promise rejects, an `error` event with code `screencast_refresh_failed` and `fatal: false` is emitted, and the tab stays open with `streaming: false` while the session keeps running; raw input for that tab is rejected until a new generation starts. Any of `refreshScreencast()`, `setViewport()`, or `activateTab()` on that tab then starts a new screencast generation. Only browser loss keeps the failure fatal: a disconnected Chromium reports `browser_disconnected` (or a fatal `screencast_refresh_failed` if the refresh observes the loss first) and shuts the session down, and page or CDP-session loss remains tab-scoped as described below.
+
+`updateScreencastOptions({ quality, maxWidth, maxHeight, everyNthFrame })` changes the encoder settings of a session at runtime. Each supplied field is validated with the same bounds as the constructor, the merged result must still satisfy the pixel-area ceiling, and unknown keys such as `format` are rejected. Accepted values are stored for every later screencast start. When the active tab is streaming, the call restarts the screencast with a new generation exactly like `refreshScreencast()` and resolves with the same `ILiveBrowserFrameIdentity`; otherwise it stores the values and resolves with `null`.
+
+```typescript
+const identity = await session.updateScreencastOptions({
+  quality: 45,
+  maxWidth: 640,
+  maxHeight: 360,
+});
+if (identity) {
+  await session.acknowledgeFrame(identity);
+}
+```
 
 The live API includes:
 
-- Lifecycle and state: `start()`, `stop()`, `terminate()`, `refreshScreencast()`, `getProcessState()`, `onEvent()`, and `getState()`
+- Lifecycle and state: `start()`, `stop()`, `terminate()`, `refreshScreencast()`, `updateScreencastOptions()`, `getProcessState()`, `onEvent()`, and `getState()`
 - Tabs and navigation: `createTab()`, `activateTab()`, `closeTab()`, `navigate()`, `back()`, `forward()`, and `reload()`
 - Viewport and raw input: `setViewport()`, `dispatchMouse()`, `dispatchWheel()`, `dispatchKey()`, and `insertText()`
 - Agent-oriented actions: `click()`, `fill()`, and `press()` with bounded selectors and timeouts
 - Capture and observation: `captureSnapshot()` returns viewport-only JPEG or PNG bytes; `observe()` returns bounded URL, title, tab state, and textual accessibility content without image bytes
 - Optional evaluation: `evaluate()` returns bounded JSON values when the session explicitly sets `allowEvaluation: true`
 
-Coordinate, keyboard, text, and semantic input messages include `tabId`, `generation`, and `viewportRevision`. This rejects input derived from an old stream generation, resize, or tab state. Snapshot, observation, and semantic operations are serialized with lifecycle mutations; inactive tabs receive the current session viewport before use. `viewport` takes precedence over `launchOptions.defaultViewport`; `null` falls back to 800x600. The runtime canonicalizes every page to a desktop, non-touch viewport because mobile emulation flags are outside the public viewport contract. Viewport dimensions, device scale factor, and physical pixel area are bounded, and full-page snapshots are intentionally unsupported. CDP sessions and CDP frame identifiers remain private implementation details. `LiveBrowserSession` owns launch cancellation, so callers cannot supply `launchOptions.signal`. It supports only Chromium over CDP and rejects Firefox or WebDriver BiDi launch selections.
+Coordinate, keyboard, text, and semantic input messages include `tabId`, `generation`, and `viewportRevision`. This rejects input derived from an old stream generation, resize, or tab state. Snapshot, observation, and semantic operations are serialized with lifecycle mutations; inactive tabs receive the current session viewport before use.
+
+`dispatchWheel()` merges bursts. While one `mouseWheel` CDP dispatch for a tab is in flight, further `dispatchWheel()` calls for the same tab and stream accumulate: `deltaX` and `deltaY` are summed and clamped to ±1,000,000, and the latest `x`, `y`, and `modifiers` win. The accumulated input is sent as one CDP event once the in-flight dispatch settles, and every caller's promise resolves when its deltas have been sent or rejects if that merged send fails. Ordering with other raw input on the same tab is preserved: a `dispatchMouse()`, `dispatchKey()`, or `insertText()` call issued after wheel input is sent only after the preceding wheel dispatch has been acknowledged by Chromium, and wheel input issued after such a call starts a new batch behind it. `viewport` takes precedence over `launchOptions.defaultViewport`; `null` falls back to 800x600. The runtime canonicalizes every page to a desktop, non-touch viewport because mobile emulation flags are outside the public viewport contract. Viewport dimensions, device scale factor, and physical pixel area are bounded, and full-page snapshots are intentionally unsupported. CDP sessions and CDP frame identifiers remain private implementation details. `LiveBrowserSession` owns launch cancellation, so callers cannot supply `launchOptions.signal`. It supports only Chromium over CDP and rejects Firefox or WebDriver BiDi launch selections.
 
-`start()`, `refreshScreencast()`, tab and navigation methods, `setViewport()`, `captureSnapshot()`, `observe()`, semantic actions, and `evaluate()` accept a trailing `{ signal }` operation argument. A pre-aborted operation is never admitted. An operation aborted while queued is removed immediately. An active operation receives cancellation when its Puppeteer or CDP primitive supports it; otherwise its promise rejects only after the underlying work settles, and it continues to occupy the serialized queue until then. An active `refreshScreencast()` completes the restorative restart before settling caller cancellation, unless session shutdown or page invalidation revokes that restart. Cancellation therefore does not promise that an already-started browser side effect did not occur. `stop()`, frame acknowledgement, event/state access, and direct raw input are intentionally not caller-cancellable.
+`start()`, `refreshScreencast()`, `updateScreencastOptions()`, tab and navigation methods, `setViewport()`, `captureSnapshot()`, `observe()`, semantic actions, and `evaluate()` accept a trailing `{ signal }` operation argument. A pre-aborted operation is never admitted. An operation aborted while queued is removed immediately. An active operation receives cancellation when its Puppeteer or CDP primitive supports it; otherwise its promise rejects only after the underlying work settles, and it continues to occupy the serialized queue until then. An active `refreshScreencast()` or `updateScreencastOptions()` completes the restorative restart before settling caller cancellation, unless session shutdown or page invalidation revokes that restart. Cancellation therefore does not promise that an already-started browser side effect did not occur. `stop()`, frame acknowledgement, event/state access, and direct raw input are intentionally not caller-cancellable.
 
 Optional browser guards can be enabled when composing a higher-level runtime:
 

package/ts/00_commitinfo_data.ts

@@ -3,6 +3,6 @@
  */
 export const commitinfo = {
   name: '@push.rocks/smartpuppeteer',
-  version: '2.6.0',
+  version: '2.7.0',
   description: 'Provides simplified access to Puppeteer for automation and testing purposes.'
 }