dsh-web-fetch-playwright 0.2.20 → 0.2.23

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -60,7 +60,8 @@
60
60
  * collecting after each, until the walk can hold AT THE BOTTOM through a
61
61
  * {@link SCROLL_BOTTOM_PATIENCE_MS} window with the height unchanged
62
62
  * (growth restarts the clock — that is what pages load-more sentinels),
63
- * or the step/budget caps close it.
63
+ * or the step/budget caps close it. A page that never reacted and issued
64
+ * no fetch/XHR holds only {@link SCROLL_INERT_PATIENCE_MS}.
64
65
  *
65
66
  * A MutationObserver watches for the walk's whole reason to exist: the
66
67
  * page REACTING to scroll by adding or removing elements. Most of the web
@@ -71,6 +72,14 @@
71
72
  * jumped over may hold content that never rendered, so the walk rewinds
72
73
  * to where it jumped from and finishes carefully. Correctness never
73
74
  * rests on the warm-up guess; only the cost of a scroll-inert page does.
75
+ *
76
+ * The same observer HARVESTS UNKEYED ROWS — the third shape: react-window
77
+ * lists and X-style timelines give rows no DOM identity at all (the
78
+ * framework key never reaches the DOM), and framework-generated ids
79
+ * (`ember123`, React useId's `:r1:`) change per render so they are not
80
+ * keys either. When the engine removes such a row, the record still
81
+ * holds it whole plus the sibling nodes it sat between; that is the row's
82
+ * stash and its splice anchor, with no layout read at all.
74
83
  * 4. MERGE into a structural clone (`document.cloneNode(true)`: no
75
84
  * serialize/re-parse round trip — a runtime-rewritten data script with a
76
85
  * literal `</script>` in its text cannot corrupt it — and, having no
@@ -84,11 +93,21 @@
84
93
  * spliced parent is already present and skipped. Placement is
85
94
  * structural, so it is right whatever order the walk SAW rows in (an
86
95
  * anchored open starts mid-document).
96
+ * - RESTORE (unkeyed rows): in every list the engine both removed text-
97
+ * bearing rows from and added to during the walk, each removed row the
98
+ * list no longer holds goes back beside the sibling it was removed
99
+ * next to (chains through removed anchors; rows with no surviving
100
+ * anchor at all — a wholesale innerHTML re-render — go above the final
101
+ * window, which is the only place a top-down walk loses them). Rows
102
+ * still present are recognised by a per-list identity attribute
103
+ * (data-index, aria-rowindex, …) or by text; in-place stubs are
104
+ * absorbed.
87
105
  * - SWAP: every keyed element emptier than its snapshot (fewer child
88
106
  * elements, or meaningfully less text — the Discourse placeholder) is
89
107
  * replaced by that snapshot, except containers (ancestors of another
90
108
  * element being swapped), which are assembled from those descendants
91
- * instead: a container's own snapshot is one window's partial view.
109
+ * instead: a container's own snapshot is one window's partial view. A
110
+ * restored list and its ancestors are containers too.
92
111
  * The clone is then stripped of scripting elements and serialized; the
93
112
  * provider reads that string instead of the live page. The live document
94
113
  * is never rewritten, so no late scroll/rAF/timer reaction can touch the
@@ -98,13 +117,16 @@
98
117
  * error, or an engine that rejects the walk just keeps the legacy
99
118
  * single-instant capture — a walk can only add content, never fail a fetch.
100
119
  * A walk that restored nothing returns no HTML, and the provider reads the
101
- * live page exactly as before.
120
+ * live page exactly as before — scrolled back to where it opened when the
121
+ * page reacted, so a virtualizer the merge could not rebuild hands over its
122
+ * opening window rather than the walk's last one.
102
123
  *
103
124
  * The walk runs in BOTH denoise modes (the raw-HTML path serializes the same
104
125
  * cloaked placeholders) but never for non-HTML documents. It is bounded by
105
126
  * the `scrollBudgetMs` setting (`0` restores the pre-walk legacy behavior)
106
- * and always by the remaining fetch deadline minus a decode reserve — the
107
- * provider's abort closes the tab, which rejects the pending evaluation.
127
+ * and always by the fetch's remaining waiting budget: a walk cut short there
128
+ * still merges and returns what it stashed, and the content read plus decode
129
+ * run in the provider's extraction grace window after it.
108
130
  *
109
131
  * @module dsh-web-fetch-playwright/autoscroll
110
132
  */
@@ -135,6 +157,16 @@ export declare const SCROLL_BOTTOM_POLL_MS = 450;
135
157
  * two 450ms checks), so this is sized over a slow round trip.
136
158
  */
137
159
  export declare const SCROLL_BOTTOM_PATIENCE_MS = 2500;
160
+ /**
161
+ * Bottom hold for a page that has given no sign of loading more: it never
162
+ * added or removed an element in response to the walk's scrolling, and it
163
+ * issued no fetch/XHR since the walk began. The full
164
+ * {@link SCROLL_BOTTOM_PATIENCE_MS} exists for feeds whose next chunk is in
165
+ * flight; holding it on every static article made the walk cost ~4s on
166
+ * pages it could never help. Any reaction or data request (even one that
167
+ * lands during this shorter hold) restores the full patience.
168
+ */
169
+ export declare const SCROLL_INERT_PATIENCE_MS = 900;
138
170
  /**
139
171
  * Careful steps without any DOM reaction before the walk treats the page as
140
172
  * scroll-inert and probes the rest with a single jump (rewinding if the
@@ -162,11 +194,6 @@ export declare const SCROLL_STASH_MAX_CHARS = 2000000;
162
194
  * rehydrate. Such keys are remembered and never serialized again.
163
195
  */
164
196
  export declare const SCROLL_STASH_MAX_ENTRY_CHARS = 64000;
165
- /**
166
- * Tail of the fetch budget the walk never spends, keeping room for the
167
- * content read and the synchronous denoise decode.
168
- */
169
- export declare const SCROLL_FINISH_RESERVE_MS = 4000;
170
197
  /**
171
198
  * A walk budget below this cannot complete even the first window plus one
172
199
  * stability check, so the walk is skipped outright.
@@ -195,6 +222,14 @@ export interface DeepCaptureStats {
195
222
  rehydrated: number;
196
223
  /** Removed keyed elements spliced back (container walks only). */
197
224
  spliced?: number;
225
+ /** Unkeyed rows a virtualizer removed, restored into their list. */
226
+ restored?: number;
227
+ /**
228
+ * True when nothing could be restored on a page that reacted, so the walk
229
+ * scrolled back to where the page opened before handing it to the
230
+ * provider's live read.
231
+ */
232
+ rewound?: boolean;
198
233
  /**
199
234
  * Whether the page reacted to scrolling (added/removed elements). False
200
235
  * means the walk took the scroll-inert fast path to the bottom.
@@ -224,6 +259,7 @@ export interface DeepCaptureParams {
224
259
  stepSettleMs: number;
225
260
  bottomPollMs: number;
226
261
  bottomPatienceMs: number;
262
+ inertPatienceMs: number;
227
263
  inertSteps: number;
228
264
  maxSteps: number;
229
265
  stashMaxEntries: number;
@@ -38,6 +38,15 @@ export interface RawCandidate {
38
38
  kind: CandidateKind;
39
39
  /** HTML of the extraction; sanitized by the caller. */
40
40
  html: string;
41
+ /**
42
+ * The extraction as a live DOM node, when the generator already holds one
43
+ * (`doc`/`docc`/`ns_doc` serialize `body`, `main` the container itself).
44
+ * DOMPurify sanitizes a node by importing a copy — the caller's re-parse
45
+ * of `html` (hundreds of ms on large pages, measured on the Angular docs
46
+ * badcase: 555ms for the whole-body candidate alone) is skipped; `html`
47
+ * stays the identity for the dedup below.
48
+ */
49
+ node?: Element;
41
50
  /** Title to prepend when the body does not open with a heading. */
42
51
  title?: string;
43
52
  /** Whether the 0.2.18 pipeline would have returned exactly this candidate. */
@@ -109,6 +118,13 @@ export declare function childElements(element: Element): Element[];
109
118
  export declare function unwrapNoscripts(document: Document): number;
110
119
  /** Visible text inside `<noscript>` elements (parsed either way). */
111
120
  export declare function noscriptTextLength(document: Document): number;
121
+ /**
122
+ * Cookie/consent banner wording (EN/DE/FR/ES/IT/NL/ZH), as a regex source
123
+ * shared with the selector's `consent_hits` feature so both read the same
124
+ * languages. Chinese banners say "Cookie" or 隐私; 同意 ("agree") alone is
125
+ * everyday prose and only counts as the banner's ACTION below.
126
+ */
127
+ export declare const CONSENT_TERMS: string;
112
128
  /**
113
129
  * Strip page chrome from a cloned body: chrome-named/roled subtrees, hidden
114
130
  * subtrees, and consent banners recognised by their wording. A subtree is
@@ -146,7 +162,46 @@ export interface GenerationFacts {
146
162
  hasPreloaded: boolean;
147
163
  /** The semantic main container, when found. */
148
164
  main: Element | null;
165
+ /**
166
+ * True when the Readability family was SKIPPED because nearly all body
167
+ * text is custom-wrapped ({@link customWrappedTextShare}) — every
168
+ * extraction that would carry it sanitizes to nothing. The orchestrator
169
+ * then seeds the selector with the empty-`art` reference stats such a run
170
+ * would have measured, keeping the surviving candidates' features
171
+ * (`words_vs_art_log`, the block diffs) exactly what a dead Readability
172
+ * pick yields.
173
+ */
174
+ readabilitySkipped?: boolean;
149
175
  }
176
+ /**
177
+ * At or above this share of custom-wrapped body text the Readability family
178
+ * is skipped (see {@link customWrappedTextShare}).
179
+ */
180
+ export declare const CUSTOM_WRAP_SKIP_SHARE = 0.9;
181
+ /**
182
+ * Share of the body's visible text that sits inside custom-element subtrees
183
+ * (tag names containing `-`: `<ms-app>`, `<my-doc>` — the HTML spec reserves
184
+ * them for custom elements; no standard HTML or MathML tag and no tag the
185
+ * denoise pipeline's DOMPurify step keeps contains a dash). DOMPurify's
186
+ * allowlist removes such tags wholesale — `KEEP_CONTENT: false` takes their
187
+ * subtrees with them — so once nearly ALL the page's text is custom-wrapped,
188
+ * every extraction that would carry it sanitizes to a corpse. Measured on
189
+ * the badcase (aistudio.google.com/docs/computer-use, an Angular app whose
190
+ * entire DOM lives under `<ms-app>`): 99% of body text custom-wrapped;
191
+ * Readability's pick sanitized to a root with ZERO words, the whole-document
192
+ * candidate to 39 of ~4.5k words — while the three Readability passes over
193
+ * the 1.1MB document cost ~6.5s of pure latency on candidates that could
194
+ * never win. The DOM-backed fallbacks, the noscript paths and the block
195
+ * classifier — the candidates that survive DOMPurify on this shape, and the
196
+ * ones the selector picks on such pages anyway — carry the extraction.
197
+ *
198
+ * Noise subtrees (script/style/nav/…) count as neither wrapped nor unwrapped;
199
+ * the share is over CONTENT-bearing text only. A partial share (a web
200
+ * component widget inside a normal article) stays below the threshold and
201
+ * keeps the full pipeline: only the shape where the family is provably dead
202
+ * pays for it.
203
+ */
204
+ export declare function customWrappedTextShare(body: Element): number;
150
205
  /**
151
206
  * Propose every candidate extraction for one parsed page.
152
207
  * @param document - the parsed page (data URIs already elided); not mutated.
@@ -41,8 +41,8 @@
41
41
  export declare const CHALLENGE_POLL_INTERVAL_MS = 500;
42
42
  /**
43
43
  * Tail of the fetch budget held back from the challenge wait so a cleared
44
- * challenge still has time for its settle (5s networkidle cap) plus the
45
- * content read and the denoise pipeline.
44
+ * challenge still has time for its settle (the `DEFAULT_SETTLE_MS` cap)
45
+ * plus the content read and the denoise pipeline.
46
46
  */
47
47
  export declare const CHALLENGE_FINISH_RESERVE_MS = 8000;
48
48
  /**
@@ -37,16 +37,6 @@ export interface PlaywrightSettings {
37
37
  denoise?: boolean;
38
38
  /** How many fetches may render at once (1–200); blank = backend default. */
39
39
  maxConcurrency?: number;
40
- /**
41
- * Bounded wait (ms) for a Cloudflare challenge to clear naturally in the
42
- * same tab (0–60000); 0 = off (return the first response as-is).
43
- */
44
- challengeWaitMs?: number;
45
- /**
46
- * Bounded budget (ms) for the deep-capture scroll walk (0–30000); 0 = off
47
- * (capture the document as it rests, the legacy single-instant behavior).
48
- */
49
- scrollBudgetMs?: number;
50
40
  }
51
41
  /** What the Playwright card renders. */
52
42
  export interface PlaywrightCardState extends CardShell {
@@ -63,10 +53,6 @@ export interface PlaywrightCardState extends CardShell {
63
53
  denoise: CardFieldState;
64
54
  /** Concurrency input (draft decimal integer). */
65
55
  maxConcurrency: CardFieldState;
66
- /** Challenge wait input (draft decimal integer of milliseconds). */
67
- challengeWaitMs: CardFieldState;
68
- /** Scroll-walk budget input (draft decimal integer of milliseconds). */
69
- scrollBudgetMs: CardFieldState;
70
56
  }
71
57
  /** The registration-side face the card's slot entry injects. */
72
58
  export interface PlaywrightCardFace extends CardActions {
@@ -5,7 +5,7 @@
5
5
  * @module dsh-web-fetch-playwright/client/locales
6
6
  */
7
7
  /** Locale keys this card renders. */
8
- export type PlaywrightCardLocaleKey = 'description' | 'backendLabel' | 'backendLocal' | 'backendLocalHint' | 'backendCdp' | 'backendCdpHint' | 'playwrightPath' | 'playwrightPathHint' | 'playwrightPathPlaceholder' | 'cdpEndpoint' | 'cdpEndpointHint' | 'shareBrowserContext' | 'shareBrowserContextHint' | 'contextMode' | 'isolateBrowserContext' | 'isolateBrowserContextHint' | 'contextIsolationScope' | 'contextScopeSession' | 'contextScopeSessionHint' | 'contextScopeGlobal' | 'contextScopeGlobalHint' | 'denoise' | 'denoiseHint' | 'maxConcurrency' | 'maxConcurrencyHint' | 'maxConcurrencyPlaceholder' | 'challengeWaitMs' | 'challengeWaitMsHint' | 'challengeWaitMsPlaceholder' | 'scrollBudgetMs' | 'scrollBudgetMsHint' | 'scrollBudgetMsPlaceholder' | 'overridden' | 'reset' | 'readOnly' | 'save' | 'saving' | 'discard' | 'saveFailed' | 'invalidText';
8
+ export type PlaywrightCardLocaleKey = 'description' | 'backendLabel' | 'backendLocal' | 'backendLocalHint' | 'backendCdp' | 'backendCdpHint' | 'playwrightPath' | 'playwrightPathHint' | 'playwrightPathPlaceholder' | 'cdpEndpoint' | 'cdpEndpointHint' | 'shareBrowserContext' | 'shareBrowserContextHint' | 'contextMode' | 'isolateBrowserContext' | 'isolateBrowserContextHint' | 'contextIsolationScope' | 'contextScopeSession' | 'contextScopeSessionHint' | 'contextScopeGlobal' | 'contextScopeGlobalHint' | 'denoise' | 'denoiseHint' | 'maxConcurrency' | 'maxConcurrencyHint' | 'maxConcurrencyPlaceholder' | 'overridden' | 'reset' | 'readOnly' | 'save' | 'saving' | 'discard' | 'saveFailed' | 'invalidText';
9
9
  /** This plugin's dictionary namespace, merged into the locale key map. */
10
10
  declare module '@deepseek-ai/dsh-client-ui-slots' {
11
11
  interface LocaleNamespaceMap {
@@ -15,16 +15,24 @@
15
15
  import type { Volatile } from '@deepseek-ai/cordis';
16
16
  import z from '@deepseek-ai/schemastery';
17
17
  /**
18
- * Default bounded budget (ms) for the deep-capture scroll walk
19
- * (`src/autoscroll.ts`). Sized so long topics stay fully covered at the
20
- * careful cadence (300ms motion + 300ms rest per ~0.7-viewport step, plus
21
- * the bottom patience window): 20s walks ~25kpx with pagination, covering
22
- * the measured linux.do badcase (~18kpx, 51 replies) with headroom.
23
- * Scroll-inert pages take the walk's fast path and spend a few seconds.
18
+ * Default CAP (ms) on the deep-capture scroll walk (`src/autoscroll.ts`).
19
+ * It bounds only pages that keep reacting to scrolling — virtualized feeds
20
+ * and load-more streams — where 20s covers ~25kpx at the careful cadence
21
+ * (`SCROLL_STEP_MOTION_MS` + `SCROLL_STEP_SETTLE_MS` per ~0.7-viewport
22
+ * step). It is not what a typical page pays: a page that never reacts takes
23
+ * `SCROLL_INERT_STEPS` careful steps, one jump to the bottom and a
24
+ * `SCROLL_INERT_PATIENCE_MS` hold, and a page that does not scroll is
25
+ * skipped outright.
24
26
  */
25
27
  export declare const DEFAULT_SCROLL_BUDGET_MS = 20000;
26
- /** Ceiling the schema accepts for `scrollBudgetMs`. */
27
- export declare const MAX_SCROLL_BUDGET_MS = 30000;
28
+ /**
29
+ * Default cap (ms) on the post-navigation settle — the wait for SPA
30
+ * rendering to finish once the document has loaded (see the settle in
31
+ * `src/provider.ts`). A CAP, not a fixed cost: the settle ends as soon as
32
+ * the page's network goes quiet. Code-level only (tests, demo scripts):
33
+ * not part of the settings schema. `0` disables the settle entirely.
34
+ */
35
+ export declare const DEFAULT_SETTLE_MS = 5000;
28
36
  /** Default CDP endpoint when the settings section leaves it blank. */
29
37
  export declare const DEFAULT_CDP_ENDPOINT = "127.0.0.1:9222";
30
38
  /**
@@ -45,16 +53,24 @@ export declare const MAX_CONCURRENCY_CEILING = 200;
45
53
  /**
46
54
  * Default bounded wait (ms) for a Cloudflare challenge to clear naturally —
47
55
  * the user's real browser passes the verification on its own while the fetch
48
- * holds the same page and context. Sized to fit the 45s per-fetch deadline
49
- * with the settle/decode tail (and one retry) still inside it.
56
+ * holds the same page and context. Bounded further by the 25s per-fetch
57
+ * waiting budget, which keeps a settle/decode tail after the wait.
50
58
  */
51
59
  export declare const DEFAULT_CHALLENGE_WAIT_MS = 15000;
52
- /** Ceiling the schema accepts for `challengeWaitMs`. */
53
- export declare const MAX_CHALLENGE_WAIT_MS = 60000;
54
60
  /** Default same-page re-navigation attempts after a challenge wait runs out. */
55
61
  export declare const DEFAULT_CHALLENGE_RETRIES = 1;
56
62
  /** Ceiling the schema accepts for `challengeRetries`. */
57
63
  export declare const MAX_CHALLENGE_RETRIES = 3;
64
+ /**
65
+ * Default for `challengeHttpFallback`: when the browser cannot clear a
66
+ * challenge inside its window (the interactive checkbox was never offered,
67
+ * or verification refused), refetch the URL once through the page's
68
+ * plain-HTTP request context — the same cookie jar, no render decision —
69
+ * because many Cloudflare edges serve the real document to non-browser
70
+ * clients while withholding the interactive challenge from the browser
71
+ * session (measured live: v2ex.com from a datacenter IP).
72
+ */
73
+ export declare const DEFAULT_CHALLENGE_HTTP_FALLBACK = true;
58
74
  /** Which browser backend serves a fetch. */
59
75
  export type PlaywrightBackend = 'local' | 'cdp';
60
76
  /**
@@ -98,7 +114,9 @@ export interface Config {
98
114
  * Bounded wait (ms) for a Cloudflare challenge to clear naturally inside
99
115
  * the same page/context. `0` (or any config leaving this at 0) restores the
100
116
  * legacy behavior: the first response is final, no waiting — the A/B
101
- * baseline. Schema default: {@link DEFAULT_CHALLENGE_WAIT_MS}.
117
+ * baseline. Code-level only (tests, demo scripts): not part of the
118
+ * settings schema, so users cannot configure it. Default:
119
+ * {@link DEFAULT_CHALLENGE_WAIT_MS}.
102
120
  */
103
121
  challengeWaitMs?: number;
104
122
  /**
@@ -107,15 +125,32 @@ export interface Config {
107
125
  * Schema default: {@link DEFAULT_CHALLENGE_RETRIES}.
108
126
  */
109
127
  challengeRetries?: number;
128
+ /**
129
+ * When the browser cannot clear a challenge inside its window (no
130
+ * checkbox offered, or verification refused), refetch the URL once
131
+ * through the page's plain-HTTP request context (same cookie jar) and
132
+ * return that document when it is not itself a challenge. `false`
133
+ * restores the browser-only posture.
134
+ */
135
+ challengeHttpFallback?: boolean;
110
136
  /**
111
137
  * Bounded budget (ms) for the deep-capture scroll walk — the in-page pass
112
138
  * that defeats viewport virtualization on feed-shaped pages (Discourse
113
139
  * topic streams: replies outside the render window are empty height-only
114
140
  * placeholders carrying the post's id; see `src/autoscroll.ts`). `0`
115
- * restores the legacy single-instant capture.
116
- * Schema default: {@link DEFAULT_SCROLL_BUDGET_MS}.
141
+ * restores the legacy single-instant capture. Code-level only (tests,
142
+ * demo scripts): not part of the settings schema, so users cannot
143
+ * configure it. Default: {@link DEFAULT_SCROLL_BUDGET_MS}.
117
144
  */
118
145
  scrollBudgetMs?: number;
146
+ /**
147
+ * Cap (ms) on the post-navigation settle (load baseline + bounded
148
+ * network-quiet window with stale-request discount). Code-level only
149
+ * (tests, demo scripts): not part of the settings schema, so users cannot
150
+ * configure it. `0` disables the settle. Default:
151
+ * {@link DEFAULT_SETTLE_MS}.
152
+ */
153
+ settleMs?: number;
119
154
  /**
120
155
  * How many fetches may render at once. Blank = backend default (4 for
121
156
  * local — each slot launches a browser; 50 for CDP — each slot is a tab in
@@ -131,9 +166,8 @@ export declare const Config: z<Schemastery.ObjectS<NoInfer<{
131
166
  contextIsolationScope: z<"session" | "global", "session" | "global", "volatile-defined">;
132
167
  denoise: z<boolean, boolean, "volatile-defined">;
133
168
  maxConcurrency: z<number, number, "volatile">;
134
- challengeWaitMs: z<number, number, "volatile-defined">;
135
169
  challengeRetries: z<number, number, "volatile-defined">;
136
- scrollBudgetMs: z<number, number, "volatile-defined">;
170
+ challengeHttpFallback: z<boolean, boolean, "volatile-defined">;
137
171
  }>>, Schemastery.ObjectT<NoInfer<{
138
172
  backend: z<"local" | "cdp", "local" | "cdp", "volatile-defined">;
139
173
  playwrightPath: z<string, string, "volatile-defined">;
@@ -142,9 +176,8 @@ export declare const Config: z<Schemastery.ObjectS<NoInfer<{
142
176
  contextIsolationScope: z<"session" | "global", "session" | "global", "volatile-defined">;
143
177
  denoise: z<boolean, boolean, "volatile-defined">;
144
178
  maxConcurrency: z<number, number, "volatile">;
145
- challengeWaitMs: z<number, number, "volatile-defined">;
146
179
  challengeRetries: z<number, number, "volatile-defined">;
147
- scrollBudgetMs: z<number, number, "volatile-defined">;
180
+ challengeHttpFallback: z<boolean, boolean, "volatile-defined">;
148
181
  }>>, "plain">;
149
182
  /**
150
183
  * The config `apply` receives once the schema resolves: every volatile field
@@ -169,12 +202,10 @@ export interface PlaywrightPluginConfig {
169
202
  * field owns no default — the backend resolves one).
170
203
  */
171
204
  maxConcurrency: Volatile<number | undefined>;
172
- /** Challenge wait budget (ms) as a live reference. */
173
- challengeWaitMs: Volatile<number>;
174
205
  /** Challenge retry count as a live reference. */
175
206
  challengeRetries: Volatile<number>;
176
- /** Deep-capture scroll-walk budget (ms) as a live reference. */
177
- scrollBudgetMs: Volatile<number>;
207
+ /** Challenge dead-end plain-HTTP refetch toggle as a live reference. */
208
+ challengeHttpFallback: Volatile<boolean>;
178
209
  }
179
210
  /**
180
211
  * Read the live references `apply` holds into one plain input for
@@ -198,14 +229,16 @@ export declare function snapshotsOf(config: PlaywrightPluginConfig): Config;
198
229
  export declare function resolveConfig(config: Partial<Config>): ResolvedConfig;
199
230
  /**
200
231
  * Complete config after schemastery applies the field defaults it owns.
201
- * `maxConcurrency` and `scrollBudgetMs` stay optional — the former because
202
- * its effective default depends on `backend`, the latter so hand-built
203
- * resolved configs (tests) may omit it and let
204
- * {@link effectiveScrollBudgetMs} supply the default the schema would have.
232
+ * `maxConcurrency` stays optional because its effective default depends on
233
+ * `backend`; `challengeWaitMs`, `scrollBudgetMs` and `settleMs` stay
234
+ * optional because they are code-level overrides outside the schema —
235
+ * absent, the effective* helpers supply the built-in defaults.
205
236
  */
206
- export type ResolvedConfig = Omit<Required<Config>, 'maxConcurrency' | 'scrollBudgetMs'> & {
237
+ export type ResolvedConfig = Omit<Required<Config>, 'maxConcurrency' | 'challengeWaitMs' | 'scrollBudgetMs' | 'settleMs'> & {
207
238
  maxConcurrency?: number;
239
+ challengeWaitMs?: number;
208
240
  scrollBudgetMs?: number;
241
+ settleMs?: number;
209
242
  };
210
243
  /**
211
244
  * The concurrency limit a fetch actually runs with: an explicit setting
@@ -218,7 +251,7 @@ export type ResolvedConfig = Omit<Required<Config>, 'maxConcurrency' | 'scrollBu
218
251
  export declare function effectiveMaxConcurrency(config: Pick<Config, 'backend' | 'maxConcurrency'>): number;
219
252
  /**
220
253
  * The challenge wait budget a fetch actually runs with: an explicit setting
221
- * wins, else the schema default. `0` means the challenge path is off.
254
+ * wins, else the built-in default. `0` means the challenge path is off.
222
255
  *
223
256
  * @param config - the resolved settings section (or any partial of it).
224
257
  * @returns the bounded wait in milliseconds.
@@ -232,14 +265,31 @@ export declare function effectiveChallengeWaitMs(config: Pick<Config, 'challenge
232
265
  * @returns retries after a wait window runs out.
233
266
  */
234
267
  export declare function effectiveChallengeRetries(config: Pick<Config, 'challengeRetries'>): number;
268
+ /**
269
+ * Whether a challenge dead-end may refetch through plain HTTP: an explicit
270
+ * setting wins (absent = on, the default).
271
+ *
272
+ * @param config - the resolved settings section (or any partial of it).
273
+ * @returns true when the plain-HTTP fallback may run.
274
+ */
275
+ export declare function effectiveChallengeHttpFallback(config: Pick<Config, 'challengeHttpFallback'>): boolean;
235
276
  /**
236
277
  * The scroll-walk budget a fetch actually runs with: an explicit setting
237
- * wins, else the schema default. `0` means the deep-capture walk is off.
278
+ * wins, else the built-in default. `0` means the deep-capture walk is off.
238
279
  *
239
280
  * @param config - the resolved settings section (or any partial of it).
240
281
  * @returns the bounded walk budget in milliseconds.
241
282
  */
242
283
  export declare function effectiveScrollBudgetMs(config: Pick<Config, 'scrollBudgetMs'>): number;
284
+ /**
285
+ * The settle cap a fetch actually runs with: an explicit value wins, else
286
+ * the built-in default. `0` means the settle is off (domcontentloaded
287
+ * content is read as-is).
288
+ *
289
+ * @param config - the resolved settings section (or any partial of it).
290
+ * @returns the bounded settle cap in milliseconds.
291
+ */
292
+ export declare function effectiveSettleMs(config: Pick<Config, 'settleMs'>): number;
243
293
  /**
244
294
  * The context mode a fetch actually runs with: profile (a tab in the remote
245
295
  * browser's real profile) requires the CDP backend — a local launch has no
@@ -15,7 +15,7 @@
15
15
  */
16
16
  import type { Context } from '@deepseek-ai/cordis';
17
17
  import type { PlaywrightPluginConfig } from './config.ts';
18
- export { Config, DEFAULT_CDP_ENDPOINT, DEFAULT_CHALLENGE_RETRIES, DEFAULT_CHALLENGE_WAIT_MS, DEFAULT_SCROLL_BUDGET_MS, MAX_SCROLL_BUDGET_MS, effectiveChallengeRetries, effectiveChallengeWaitMs, effectiveContextMode, effectiveScrollBudgetMs, normalizeCdpEndpoint, resolveConfig, snapshotsOf, } from './config.ts';
18
+ export { Config, DEFAULT_CDP_ENDPOINT, DEFAULT_CHALLENGE_HTTP_FALLBACK, DEFAULT_CHALLENGE_RETRIES, DEFAULT_CHALLENGE_WAIT_MS, DEFAULT_SCROLL_BUDGET_MS, DEFAULT_SETTLE_MS, effectiveChallengeHttpFallback, effectiveChallengeRetries, effectiveChallengeWaitMs, effectiveContextMode, effectiveScrollBudgetMs, effectiveSettleMs, normalizeCdpEndpoint, resolveConfig, snapshotsOf, } from './config.ts';
19
19
  export type { Config as PlaywrightFetchConfig, CdpContextMode, ContextIsolationScope, PlaywrightBackend, PlaywrightPluginConfig, ResolvedConfig, } from './config.ts';
20
20
  export { CdpConnectionPool } from './cdp-pool.ts';
21
21
  export type { CdpAcquireMode, CdpConnect, CdpLease } from './cdp-pool.ts';
@@ -18,6 +18,21 @@
18
18
  * refetch through the page's API request context
19
19
  * ({@link recoverDownloadResponse}) — instead of a `WEB_PROVIDER_ERROR`.
20
20
  *
21
+ * Transient transport failures (issue: `net::ERR_NETWORK_CHANGED` and its
22
+ * family) are retried, not terminal: Chromium's NetworkChangeNotifier kills
23
+ * every in-flight request with `ERR_NETWORK_CHANGED` the moment the HOST's
24
+ * network configuration shifts under the browser — a route/interface/DNS/
25
+ * proxy change, a VPN or DHCP flap, a Docker embedded-DNS rewrite; the page
26
+ * and its target are perfectly fine, only that one navigation died, and the
27
+ * very next `goto` usually succeeds (measured live: source.android.com
28
+ * failed once mid-navigation in one DSH session and fetched clean from the
29
+ * same host seconds later). Each navigation therefore retries the
30
+ * transport-blip family ({@link isTransientNavigationFailure} — deliberately
31
+ * NOT `ERR_ABORTED`, which is overloaded with the deterministic shapes the
32
+ * recoveries above already own) a bounded number of times with a short
33
+ * backoff, inside the same page and the remaining waiting budget; a blip
34
+ * that survives the attempt cap still fails as before.
35
+ *
21
36
  * Lifecycle: the local backend launches a browser per fetch and closes it —
22
37
  * nothing outlives the call. CDP session isolation keeps one connection per
23
38
  * DSH session until its context becomes idle; this also isolates gateways that
@@ -55,9 +70,14 @@
55
70
  * behavior). The wait tracks the LAST main-frame navigation response (the
56
71
  * real page reloads in after the challenge clears) and watches the live DOM
57
72
  * so SPA-style clears are caught too. It never clicks, never injects
58
- * answers, never touches cookies itself; when the budget runs out the fetch
59
- * fails with {@link WEB_FETCH_CHALLENGE_CODE} instead of returning the
60
- * interstitial as content.
73
+ * answers, never touches cookies itself; when the budget runs out, ONE
74
+ * plain-HTTP refetch of the same URL through the page's request context
75
+ * (same cookie jar, the browser's UA, none of Chromium's render decisions
76
+ * — `challengeHttpFallback`, default on) tries to retrieve the real
77
+ * document, gated by the same response-level and content-level challenge
78
+ * classifiers so an interstitial can never pass as content. When that too
79
+ * fails, the fetch fails with {@link WEB_FETCH_CHALLENGE_CODE} instead of
80
+ * returning the interstitial as content.
61
81
  *
62
82
  * Deep capture (viewport-virtualized feeds and app-shell documents): feed-
63
83
  * shaped pages — Discourse topic streams, comment threads, timelines —
@@ -84,6 +104,17 @@
84
104
  * degrades to the legacy capture (see `src/autoscroll.ts` for the measured
85
105
  * linux.do and Feishu badcases).
86
106
  *
107
+ * Deadline (soft, then hard): each fetch has a {@link DEFAULT_TIMEOUT_MS}
108
+ * budget for WAITING — navigation, settle, challenge wait, scroll walk. When
109
+ * it runs out the page is NOT torn down: every wait is bounded by the same
110
+ * budget and simply ends, a navigation still loading is stopped
111
+ * (`window.stop()`) once its main-frame response has arrived, and whatever
112
+ * the page holds by then goes through the normal extraction — a slow or
113
+ * endless page still hands the agent the content it already rendered
114
+ * (flagged `truncated` when the document never finished loading). Only when
115
+ * nothing arrived at all, or extraction itself hangs past
116
+ * {@link EXTRACT_GRACE_MS}, does the fetch fail with `WEB_FETCH_TIMEOUT`.
117
+ *
87
118
  * @module dsh-web-fetch-playwright/provider
88
119
  */
89
120
  import type { WebFetchProvider, WebFetchRequest, WebFetchResult } from '@deepseek-ai/dsh-web';
@@ -119,19 +150,25 @@ export interface BrowserSession {
119
150
  release?: () => Promise<void>;
120
151
  }
121
152
  /**
122
- * A single abort deadline composing the caller's signal with this provider's
123
- * time budget. The timeout reason is kept apart from an outer abort so the
124
- * error translation can pick `WEB_FETCH_TIMEOUT` over `WEB_ABORTED`.
153
+ * The fetch's two-stage deadline composed with the caller's signal. The SOFT
154
+ * stage (`timeoutMs`) only bounds waiting: {@link remainingMs} counts down to
155
+ * it and {@link expired} reports it, so every wait ends there and the fetch
156
+ * extracts what the page holds. The HARD stage (`timeoutMs + graceMs`) aborts
157
+ * {@link signal}, which closes the page. The timeout reason is kept apart
158
+ * from an outer abort so the error translation can pick `WEB_FETCH_TIMEOUT`
159
+ * over `WEB_ABORTED`.
125
160
  */
126
161
  declare class Deadline {
127
162
  readonly signal: AbortSignal;
128
163
  private readonly controller;
129
164
  private readonly expiresAt;
130
165
  private timedOut;
131
- constructor(outer: AbortSignal | undefined, timeoutMs: number);
132
- /** Whether OUR timer fired (vs an outer cancellation). */
166
+ constructor(outer: AbortSignal | undefined, timeoutMs: number, graceMs?: number);
167
+ /** Whether OUR hard timer fired (vs an outer cancellation). */
133
168
  get isTimeout(): boolean;
134
- /** Milliseconds left on the budget, floored at 1 for Playwright options. */
169
+ /** Whether the soft (waiting) budget has run out. */
170
+ get expired(): boolean;
171
+ /** Milliseconds left on the waiting budget, floored at 1 for Playwright options. */
135
172
  remainingMs(): number;
136
173
  }
137
174
  /**
@@ -144,6 +181,16 @@ export declare class PlaywrightFetchProvider implements WebFetchProvider {
144
181
  private readonly sessionIdSource;
145
182
  readonly id = "playwright";
146
183
  private readonly semaphore;
184
+ /** Waiting budget per fetch (ms); code-level, overridable by subclasses (tests). */
185
+ protected readonly fetchTimeoutMs: number;
186
+ /** Extraction grace after the waiting budget (ms); code-level, overridable by subclasses. */
187
+ protected readonly extractGraceMs: number;
188
+ /** Settle quiet window (ms); code-level, overridable by subclasses (tests). */
189
+ protected readonly settleQuietMs: number;
190
+ /** Request age (ms) past which it stops gating the settle; overridable by subclasses (tests). */
191
+ protected readonly settleStaleMs: number;
192
+ /** Settle re-check cadence (ms); overridable by subclasses (tests). */
193
+ protected readonly settlePollMs: number;
147
194
  /** Shared CDP connection; injectable so the suite can substitute a fake. */
148
195
  protected readonly cdpPool: CdpConnectionPool;
149
196
  /**
@@ -174,6 +221,17 @@ export declare class PlaywrightFetchProvider implements WebFetchProvider {
174
221
  * budget before the final document is read (see the class docs).
175
222
  */
176
223
  private retrieve;
224
+ /**
225
+ * The challenge dead-end's plain-HTTP refetch: one request through the
226
+ * page's own request context (the SAME cookie jar — any clearance the
227
+ * profile holds rides along — but none of Chromium's render decisions),
228
+ * wearing the browser's own User-Agent. Returns the finished result when
229
+ * the answer is a NON-challenge HTML document, undefined otherwise (the
230
+ * caller throws the honest WEB_FETCH_CHALLENGE). A challenge served to
231
+ * this request too, a non-HTML body, or any transport failure keeps the
232
+ * browser-only verdict.
233
+ */
234
+ private challengeHttpFallback;
177
235
  /**
178
236
  * One tracked `goto`: the navigation response on success, or — when
179
237
  * Chromium failed the navigation for a reason that is not a transport
@@ -182,6 +240,11 @@ export declare class PlaywrightFetchProvider implements WebFetchProvider {
182
240
  * status-only 204/205 — {@link recoverUnrenderedStatus}), and the response
183
241
  * being a download the browser refuses to commit as a document
184
242
  * (`Content-Disposition: attachment` — {@link recoverDownloadResponse}).
243
+ * A TRANSIENT transport failure (`net::ERR_NETWORK_CHANGED` and family)
244
+ * is neither: it carries no response and no render decision, only a
245
+ * momentary host-network blip that killed this one navigation — so it is
246
+ * retried on the SAME page, bounded by {@link MAX_TRANSIENT_NAV_RETRIES}
247
+ * attempts, a short backoff, and the remaining waiting budget.
185
248
  *
186
249
  * @returns the response `goto` resolved with, plus `recovered` when the
187
250
  * navigation failure was recovered into a finished result the caller must
@@ -207,5 +270,34 @@ export declare class PlaywrightFetchProvider implements WebFetchProvider {
207
270
  * @returns true when the challenge cleared; false when the budget ran out.
208
271
  */
209
272
  private waitForChallengeClear;
273
+ /**
274
+ * Best-effort settle for client-rendered content — a Lighthouse-style
275
+ * bounded network-quiet window replacing Playwright's `networkidle`
276
+ * (officially discouraged): worker-script fetches and cross-origin auth
277
+ * frames hold one request open on the page session FOREVER (their
278
+ * completion events are attributed to other CDP targets), so strict
279
+ * networkidle never arrives on exactly the modern pages the settle exists
280
+ * for, and the legacy wait burned its whole cap on them every fetch.
281
+ *
282
+ * Sequence: a `load` baseline (usually already past — domcontentloaded
283
+ * resolved earlier), then poll until no request has STARTED within
284
+ * {@link settleQuietMs} and every still-open request is older than
285
+ * {@link settleStaleMs} — a request that old is an attribution artifact
286
+ * or a long-poll whose completion no longer gates content. All waits
287
+ * share the {@link effectiveSettleMs} cap and the fetch deadline; a page
288
+ * that exposes no request events degrades to the legacy bounded
289
+ * networkidle wait. Never throws: any exit path keeps what the document
290
+ * already holds.
291
+ */
292
+ private settleForContent;
293
+ /**
294
+ * Settle outcome observation hook — silent in production by design (the
295
+ * settle is best-effort), but subclasses (tests, diagnostics) override it
296
+ * to see which exit path every fetch took: `off` (config), `quiet`
297
+ * (network went quiet), `stale-discounted` (only stale requests remained
298
+ * open), `cap` (window exhausted), `fallback` (no request events — the
299
+ * legacy networkidle wait ran), `aborted` (fetch deadline fired).
300
+ */
301
+ protected noteSettleOutcome(outcome: 'off' | 'quiet' | 'stale-discounted' | 'cap' | 'fallback' | 'aborted', elapsedMs: number): void;
210
302
  }
211
303
  export {};