fusengine · fusengine · Jul 19, 2026 · Jul 19, 2026 · Jul 19, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -1,5 +1,25 @@
 # Changelog
 
+## [0.1.63] - 19-07-2026
+
+### Fixed
+
+- **`browser_fetch`/`browser_fetch_batch` `browserFallback` never escalating chrome-heavy SPAs** — `isThinShell` gated escalation on raw-HTML visible-text length (`< 600` chars); client-rendered pages with a lot of nav/footer chrome (e.g. `digitec.ch`, `fr.skoda.ch/apps/stock`) cleared that threshold while their real content still only existed after JS hydration, so the fast-path extraction silently returned `wordCount: 0` with `escalated: false`. `browserFallback` now also escalates on the extraction RESULT: when it comes back empty/near-empty (`wordCount < 15`, and not already fixed by hollow-extraction recovery) on HTML carrying JS-rendering markers (SPA mount point or heavy `<script>` count — `looksJsRendered`, `src/net/thin-shell.ts`), the fetch retries through a real browser render and re-extracts. Escalates at most once per fetch (loop guard). A further last-resort fallback (`recoverFromRawText`) covers the rare case where Defuddle's own content-scoring still finds nothing even on the real browser-rendered HTML (a known Defuddle limitation on dense, non-article layouts like an e-commerce grid homepage — no upstream config fixes it): the browser's own raw visible text is shipped instead of near-empty markdown, with a real word count and `extraction: "recovered"`, at no extra network/browser cost. New shared orchestrator (`src/agent/fetch-orchestrate.ts`) is now the single fetch-and-render path for both `browser_fetch` and `browser_fetch_batch`; server-rendered pages whose fast-path extraction already succeeds are unaffected (byte-identical, no new browser launch).
+- **`browser_fetch` silent hollow extraction** — on some pages (e.g. SMF forum threads) Defuddle's own boilerplate-removal deleted the element wrapping every post body, silently returning near-empty markdown (`status:200`, no error, no signal). The markdown path now detects a "hollow" extraction (nonzero-but-low word count and a low prose-capture ratio against the page's real **non-link** prose) and recovers the raw page text instead — gated on an absolute floor of recoverable non-link prose (`MIN_PROSE_CHARS`, tuned against real pages incl. `github.com/login`'s 505-char chrome footprint, not just the saved SMF fixtures), so a genuinely sparse or link-heavy page (listing/index, login, 404) is never dumped as raw link/nav soup, regardless of its link density. A **zero**-word-count Defuddle result (no content container found at all — typical of a heavy client-rendered SPA whose real content never reaches the raw HTML) is never treated as "hollow": that failure mode belongs to `browserFallback`, not raw-text recovery.
+- **`browser_fetch`/price extraction raw-CSS leak** — `htmlToText` (the raw-text fallback feeding both hollow-extraction recovery and price extraction) read `document.body.textContent`, which includes `<script>`/`<style>` tag source verbatim; a live CSS-in-JS SSR page injecting critical CSS into `<body>` leaked raw stylesheet source into the "recovered" text. `<script>`/`<style>` elements are now removed from the DOM before reading `textContent`.
+- **`browser_act`/`browser_fill` "Malformed value" on `<input type="range">`** — Playwright's `locator.fill()` sets `.value` then round-trip-asserts the raw string against the post-clamp DOM value; a native range input snaps to `min`/`max`/`step`, so the assertion always threw. `smartFill`/`actByRef` now detect a range input (`isRangeInput`) and set it via the native value setter after clamping+snapping (`fillRange`, `src/actions/fill-range.ts`), dispatching `input`/`change` so framework listeners observe it. Plain text/label/placeholder fills are unaffected.
+- **`browser_act`/`browser_fill` no-op on ARIA comboboxes** — a click-to-open combobox/listbox only renders its options after a click, so filling it directly silently no-oped. `smartFill`, `actByRef`, and `browser_act`'s `pick` kind now detect a combobox/listbox trigger (`isComboboxTrigger`, checked on the element or up to 2 ancestors) and open the trigger + click the matching option (`openComboboxAndPick`, `src/actions/combobox.ts`) instead.
+- **`pick` throwing on human-readable targets** — `browser_act`'s `pick` kind and the `perform`/`pick` action resolved their target via a raw `page.locator(target)`, which throws `Unexpected token "?"` on text like `"Où allez-vous ?"`. `pick` now resolves through the same ordered selector→role→text→label strategies as `browser_click` (`resolveClickTarget`, extracted from `smart-click.ts` into `src/actions/resolve-click-target.ts`), falling back to the raw selector only when nothing else matches.
+- **`detectChallenges` missed AWS WAF Bot Control** — `Challenges` now reports `awsWaf` (`chal_t=` query param or `window.awsWafCookieDomainList`), additive and optional in `challengesSchema`.
+- **`browser_click`/`browser_act` failing on obscured or sticky elements** — a target that hit-tests as covered (booking.com's `obscured:true` destination field, skoda.ch's sticky combobox panel) or is otherwise unclickable now escalates through a bounded click ladder (`robustClick`, `src/actions/robust-click.ts`): plain click → scroll-into-view → dismiss a covering consent/cookie overlay → force-click → raw mouse click at coordinates, each rung attempted only after the previous one throws or times out. A normally-clickable element still resolves on rung 1, byte-identical to the previous plain `locator.click()`. Point-based hit-testing (`isTopElement`, `src/actions/hit-test.ts`) confirms the true top element before trusting a bypass-actionability click; a probe failure (no DOM access, detached element) always defaults to "clear" so it never blocks a real click.
+- **Keyboard-stepped range/slider fills reporting false success** — `<input type=range>` and ARIA `role=slider` widgets that only re-render on a real trusted key event (not a synthetic setter+dispatch) now step via actual keyboard presses in bounded batches with a read-back after each batch (`stepSliderByKeyboard`, `src/actions/slider-keyboard.ts`), so a stalled widget reports an honest undershoot instead of a silent `ok:true` on an unset value. Also fixes a floating-point false negative where a clamped value (e.g. `0.30000000000000004`) read back as the equivalent parsed number (`0.3`) was reported as a mismatch.
+
+### Added
+
+- **`browser_fetch` `wordCount`/`extraction` signal** — every markdown result now reports the extractor's word count and which path produced the text (`"primary"` | `"recovered"`), both optional (unset on the `text`/non-HTML path).
+- **`browser_fetch` `contentSelector`** — optional CSS selector to pin the main content container yourself, bypassing auto-detection, for pages where recovery still isn't the right call.
+- **`browser_type`** (51st tool) — types text into whatever element currently has focus, no `ref`/`target` needed (drives `page.keyboard.type`, not a locator) — for elements a locator can't reach (e.g. a closed shadow-DOM input): `browser_click` the field first, then `browser_type`.
+
 ## [0.1.62] - 17-07-2026
 
 ### Fixed

diff --git a/README.md b/README.md
@@ -10,7 +10,7 @@ Shadow DOM + iframes), multi-step plans, structured extraction, visual diff, and
 guardrails** for payments and bookings. It drives real Chromium, so it reads **Next.js / SPA**
 pages after hydration — not just static HTML.
 
-> 50 MCP tools · stealth + rotating proxies · HTTP fast-path (single, batch & crawl) · full-site content + screenshot snapshots · structured per-card product extraction · form fill + file upload · encrypted credential vault + TOTP (fill-by-reference, origin-bound, never exposed to the LLM) · hover + drag&drop · PDF export · cookies / permissions · network mocking · clipboard · virtualized-list scraping + autoscroll · tabs / dialogs / downloads · console + network logs · MCP screenshot resources · `FUSE_CAPS` tool-group filtering · named auth profiles (cookies + localStorage + IndexedDB, saved at login) · `blockResources` · HAR record/replay · pixel visual-diff · human handoff + live view · typed structured output (`outputSchema` + `structuredContent`).
+> 51 MCP tools · stealth + rotating proxies · HTTP fast-path (single, batch & crawl) · full-site content + screenshot snapshots · structured per-card product extraction · form fill + file upload · encrypted credential vault + TOTP (fill-by-reference, origin-bound, never exposed to the LLM) · hover + drag&drop · PDF export · cookies / permissions · network mocking · clipboard · virtualized-list scraping + autoscroll · tabs / dialogs / downloads · console + network logs · MCP screenshot resources · `FUSE_CAPS` tool-group filtering · named auth profiles (cookies + localStorage + IndexedDB, saved at login) · `blockResources` · HAR record/replay · pixel visual-diff · human handoff + live view · typed structured output (`outputSchema` + `structuredContent`).
 
 ## Install
 
@@ -68,7 +68,7 @@ Full reference in **[`docs/`](./docs/README.md)**:
 
 [Installation](./docs/installation.md) ·
 [CLI](./docs/cli.md) ·
-[MCP tools (50)](./docs/mcp-tools.md) ·
+[MCP tools (51)](./docs/mcp-tools.md) ·
 [Configuration](./docs/configuration.md) ·
 [Sessions](./docs/sessions.md) ·
 [Extraction](./docs/extraction.md) ·

diff --git a/biome.json b/biome.json
@@ -1,7 +1,7 @@
 {
   "$schema": "https://biomejs.dev/schemas/2.4.16/schema.json",
   "files": {
-    "includes": ["src/**", "tests/**", "!dist/**", "!node_modules/**"]
+    "includes": ["src/**", "tests/**", "!dist/**", "!node_modules/**", "!tests/fixtures/**"]
   },
   "formatter": {
     "enabled": false

diff --git a/docs/README.md b/docs/README.md
@@ -7,7 +7,7 @@ New here? Start with the root [README](../README.md), then dive in:
 | --- | --- |
 | [Installation](./installation.md) | Requirements, install, Chromium, MCP registration, the three ways to get a browser |
 | [CLI](./cli.md) | `probe` / `fetch` / `fetch-batch` / `crawl` / `collect-batch` / `shots` / `shots-batch` / `site-shots` / `serp-batch` + one-shot page commands (`run` / `products` / `extract` / `snapshot` / `screenshot` / `inspect`) + every flag |
-| [MCP tools](./mcp-tools.md) | All 50 tools with parameters and examples |
+| [MCP tools](./mcp-tools.md) | All 51 tools with parameters and examples |
 | [Configuration](./configuration.md) | `AgentOptions`, `FUSE_*` env vars, identity, retry, output location |
 | [Sessions](./sessions.md) | Session lifecycle, auto crash recovery, `storageState` auto-save, HAR record/replay, CDP attach |
 | [Extraction](./extraction.md) | `browser_extract` / `extract_schema` / `collect` + the clean→validate→dedupe→emit pipeline |

diff --git a/docs/mcp-tools.md b/docs/mcp-tools.md
@@ -1,6 +1,6 @@
 # MCP tools
 
-Complete reference for the 50 `browser_*` tools exposed by the fuse-browser MCP server.
+Complete reference for the 51 `browser_*` tools exposed by the fuse-browser MCP server.
 
 Tools fall into two families:
 
@@ -18,11 +18,11 @@ Every tool declares an `outputSchema` (zod) and returns a **typed `structuredCon
 
 ## Capability groups (`FUSE_CAPS`)
 
-By default all 50 tools are registered. Set the `FUSE_CAPS` env var (comma-separated group names) to expose fewer tools — a lighter context for the LLM client:
+By default all 51 tools are registered. Set the `FUSE_CAPS` env var (comma-separated group names) to expose fewer tools — a lighter context for the LLM client:
 
 | Group | Tools |
 | --- | --- |
-| `core` | Session lifecycle (`browser_open`/`browser_status`/`browser_close`/`browser_connect`), navigation (`browser_navigate`/`browser_back`/`browser_forward`), actions (`browser_click`/`browser_fill`/`browser_login`/`browser_scroll`/`browser_press`/`browser_select`), `browser_tabs`, `browser_dialog`/`browser_downloads`, `browser_snapshot`/`browser_act`, `browser_wait`/`browser_wait_for`, `browser_screenshot`, `browser_autoscroll`, `browser_vault`. |
+| `core` | Session lifecycle (`browser_open`/`browser_status`/`browser_close`/`browser_connect`), navigation (`browser_navigate`/`browser_back`/`browser_forward`), actions (`browser_click`/`browser_fill`/`browser_login`/`browser_scroll`/`browser_press`/`browser_type`/`browser_select`), `browser_tabs`, `browser_dialog`/`browser_downloads`, `browser_snapshot`/`browser_act`, `browser_wait`/`browser_wait_for`, `browser_screenshot`, `browser_autoscroll`, `browser_vault`. |
 | `batch` | `browser_probe`, `browser_probe_html`, `browser_fetch`, `browser_fetch_batch`, `browser_crawl`, `browser_collect_batch`, `browser_shots_batch`, `browser_site_shots`, `browser_serp_batch`. |
 | `extract` | `browser_collect`, `browser_run`, `browser_extract`, `browser_extract_schema`, `browser_products`. |
 | `debug` | `browser_inspect`, `browser_console`, `browser_network`, `browser_visual_diff`, `browser_metrics`, `browser_pdf`, `browser_cookies`. |
@@ -94,10 +94,15 @@ Accepts the full [agentOptionShape](#browser_open) plus the same probe flags as
 
 HTTP fetch with browser TLS/HTTP2 impersonation — no browser launch, ~10x faster. For server-rendered HTML; for JS/SPA pages set `browserFallback: true` (or use `browser_probe`). Non-HTML responses (JSON APIs, `text/plain`) are returned **verbatim** — the markdown/HTML pipeline is skipped — so this also works as a fast JSON-API fetcher. The body download is capped at 10 MB.
 
+With `browserFallback: true`, escalation to a real browser render fires on **either** of two independent signals: (1) the raw HTML itself looks like an unrendered SPA shell (thin visible text + SPA/script markers), or (2) the raw HTML looks JS-rendered (SPA hydration mount point or heavy `<script>` count) **and** the fast-path extraction still came back empty/near-empty (`wordCount < 15`, not already recovered by the hollow-extraction path) — this covers chrome-heavy pages (nav/footer markup clearing 600+ visible chars) whose real content only exists after client-side hydration. Either path sets `escalated: true` and re-renders from the browser's rendered HTML; a body that already escalated is never escalated twice. If the content extractor still finds nothing on the real browser-rendered HTML (rare — a dense, non-article layout like an e-commerce grid homepage can defeat its content-scoring even post-render), the browser's own raw visible text is shipped instead of near-empty markdown, with `extraction: "recovered"` and a real `wordCount` — no extra browser launch.
+
+The markdown path auto-detects a **hollow extraction** — when the content extractor silently captures only a sliver of the page's real non-link prose (e.g. some forum threads whose posts sit inside a wrapper element the extractor drops as boilerplate) — and recovers the raw page text instead, gated so a genuinely sparse or link-heavy page (listing/index, login, 404) never gets dumped as raw link soup: recovery only fires when the page holds a substantial amount of reading content outside its `<a>` chrome that the extractor missed. The result carries `wordCount` and `extraction` (`"primary"` | `"recovered"`) so callers can see which path produced `text`. Pass `contentSelector` to pin the main content container yourself when auto-detection still picks the wrong element.
+
 | Param | Type | Required | Description |
 | --- | --- | --- | --- |
 | `url` | string | yes | URL to fetch. |
 | `format` | enum `markdown` \| `text` | no | Output format (default `markdown`: main content + YAML frontmatter). Forced to raw `text` for non-HTML bodies (JSON, plain text). |
+| `contentSelector` | string | no | CSS selector pinning the main content element, bypassing auto-detection. Markdown path only. |
 | `extractPrices` | boolean | no | Run the price extractor on the body. |
 | `extractContacts` | boolean | no | Extract `{ emails, phones, hasContactForm }` from the fetched HTML (no browser). |
 | `contactFilter` | enum `strict` \| `off` | no | Drop placeholder/template emails (default `strict`). |
@@ -110,6 +115,8 @@ HTTP fetch with browser TLS/HTTP2 impersonation — no browser launch, ~10x fast
 { "url": "https://example.com", "extractContacts": true }
 ```
 
+Returns `{ status, url, format, escalated, text, wordCount?, extraction? }` — `wordCount`/`extraction` are set on the markdown path only.
+
 ---
 
 ### browser_fetch_batch
@@ -521,6 +528,19 @@ Press a key or shortcut (e.g. `Enter`, `ArrowDown`, `Control+a`).
 { "sessionId": "s_abc123", "key": "Enter" }
 ```
 
+### browser_type
+
+Type text into whatever element currently has focus — no `ref`/`target` (it drives `page.keyboard.type`, not a locator). Use for elements a ref can't reach, e.g. a closed shadow-DOM search input: `browser_click` the field first, then `browser_type`.
+
+| Param | Type | Required | Description |
+| --- | --- | --- | --- |
+| `sessionId` | string | yes | Target session. |
+| `text` | string | yes | Text to type into the focused element. |
+
+```json
+{ "sessionId": "s_abc123", "text": "iphone" }
+```
+
 ### browser_select
 
 Select an option in a `<select>` by value, label, or index.

diff --git a/package.json b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@fusengine/browser-mcp",
-  "version": "0.1.62",
+  "version": "0.1.63",
   "description": "MCP server + CLI giving AI agents a real, stealth browser (Patchright/Playwright) — per-country identity, self-healing actions, snapshots, multi-step plans, structured extraction, CDP attach.",
   "license": "MIT",
   "author": "Fusengine",
@@ -55,7 +55,7 @@
     "typecheck": "tsc -p tsconfig.json --noEmit",
     "lint": "biome check src tests",
     "test": "bun test tests/unit",
-    "test:integration": "node --test --import tsx tests/integration/mcp.test.ts tests/integration/probe.test.ts tests/integration/snapshot.test.ts tests/integration/snapshot-prune.test.ts tests/integration/snapshot-frames.test.ts tests/integration/collect.test.ts tests/integration/collect-batch.test.ts tests/integration/selectors.test.ts tests/integration/visual-diff.test.ts tests/integration/session-state.test.ts tests/integration/pipeline.test.ts tests/integration/run.test.ts tests/integration/extract-schema.test.ts tests/integration/recovery.test.ts tests/integration/live-view.test.ts tests/integration/vault.test.ts",
+    "test:integration": "node --test --import tsx tests/integration/mcp.test.ts tests/integration/probe.test.ts tests/integration/snapshot.test.ts tests/integration/snapshot-prune.test.ts tests/integration/snapshot-frames.test.ts tests/integration/collect.test.ts tests/integration/collect-batch.test.ts tests/integration/selectors.test.ts tests/integration/visual-diff.test.ts tests/integration/session-state.test.ts tests/integration/pipeline.test.ts tests/integration/run.test.ts tests/integration/extract-schema.test.ts tests/integration/recovery.test.ts tests/integration/live-view.test.ts tests/integration/vault.test.ts tests/integration/fetch-extract.test.ts tests/integration/dismiss-overlay-selectors.test.ts",
     "browsers": "patchright install chromium",
     "mcp": "node --import tsx src/bin/mcp.ts",
     "cli": "node --import tsx src/bin/cli.ts"