feat: web_search via Mojeek · cursor editing in prompt · REASONIX_UI=plain · ticker consolidation (0.4.15)

CHANGELOG.md

@@ -3,6 +3,89 @@
 All notable changes to Reasonix. The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/);
 this project uses [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.4.15] — 2026-04-22
+
+**Headline:** Web search + fetch tools (on by default, zero
+configuration) plus real cursor editing in the prompt box (←/→,
+Backspace/Delete mid-string, multi-line ↑/↓ navigation).
+
+### Fixed
+
+- **PromptInput was append-only** — cursor was always pinned to
+  the end of the buffer, so the only way to fix a typo was
+  backspacing back through everything after it. Now:
+  - `←` / `→` move the cursor one column (clamped to buffer).
+  - `↑` / `↓` move across lines in a multi-line buffer, preserving
+    column when possible, clamping when the target line is shorter.
+  - `Ctrl+A` / `Ctrl+E` jump to start / end of the current line.
+  - `Backspace` deletes the char before the cursor; `Delete`
+    deletes the char under the cursor.
+  - Printable input inserts at the cursor (including multi-char
+    paste bursts).
+  - `Shift+Enter` / `Ctrl+J` insert a newline at the cursor.
+- **History recall no longer steals arrow keys from mid-edit.**
+  `↑` / `↓` only trigger prior-prompt recall when the buffer is
+  empty. A non-empty buffer keeps the arrows for cursor motion so
+  typed text isn't clobbered.
+
+### Added
+
+Web search + fetch tools are registered by default on `reasonix
+chat` and `reasonix code`. The model calls `web_search` /
+`web_fetch` on its own whenever a question needs fresher info than
+its training data. Backed by **Mojeek**'s public search page — no
+API key, no signup. Same Cache-First + repair + context-safety
+plumbing as every other tool.
+
+Implementation note: the first cut of this feature used DuckDuckGo,
+but a live probe from the dev machine confirmed DDG now serves
+HTTP 202 anti-bot pages for every unauthenticated POST regardless
+of UA. Mojeek is an independent-index engine that's been stable
+against the same probe (3/3 success on three queries spaced 3s
+apart). Real-browser `User-Agent` string avoids Mojeek's
+fast-path scraper filter.
+
+
+- **`src/tools/web.ts`** — two functions + one registration helper:
+  - `webSearch(query, opts?)` — fetches DDG's HTML endpoint, parses
+    ranked results (title + url + snippet). `topK` is clamped to
+    [1, 10]. Parser decodes DDG's `uddg=<url>` redirect wrapper and
+    common HTML entities.
+  - `webFetch(url, opts?)` — HTTP GET + HTML-to-text extraction
+    (scripts/styles/nav/footer/aside/svg stripped, paragraph breaks
+    preserved, entities decoded). 15s timeout, 32k-char cap (matches
+    tool-result budget), forwards caller's AbortSignal so Esc during
+    a long fetch is honored.
+  - `registerWebTools(registry, opts?)` — registers both as
+    ToolRegistry entries the model can invoke. Tool descriptions
+    guide the model to call search whenever training data might be
+    stale.
+- **`ReasonixConfig.search`** + **`searchEnabled()`** — a simple
+  boolean. Default on. Turn off with `search: false` in config or
+  `REASONIX_SEARCH=off|false|0` in env. No API keys, no provider
+  picker — one switch.
+- **Auto-registered in chat/code.** `reasonix chat` and
+  `reasonix code` register `web_search` + `web_fetch` by default.
+  Zero setup: after the normal wizard, the model can already reach
+  the web.
+
+### Tests (+18, suite 444→462)
+
+- `tests/web-tools.test.ts` (+13) — htmlToText strips
+  scripts/styles/nav/footer + decodes entities + collapses
+  whitespace; `parseDuckDuckGoResults` decodes redirect URLs + entities
+  + returns empty on unexpected markup; `webSearch` hits the DDG
+  endpoint with a browsery UA, respects topK, clamps to [1, 10],
+  throws on non-2xx; `formatSearchResults` renders the expected
+  layout; `registerWebTools` registers both verbs; `web_fetch` refuses
+  non-http(s) URLs; `webFetch` extracts title + body, truncates at
+  the cap with a visible marker, surfaces 404s.
+- `tests/config.test.ts` (+5) — `searchEnabled` defaults to true;
+  honors `search: false` in file; honors `REASONIX_SEARCH=off|false|0`;
+  stays true for unrelated env values; env off beats config true.
+
+---
+
 ## [0.4.14] — 2026-04-22
 
 **Headline:** Render-load reductions for Windows terminals where

README.md

@@ -322,6 +322,72 @@ don't need this opt-out.
 
 ---
 
+## Web search — on by default
+
+The model has two web tools the moment you launch: `web_search` and
+`web_fetch`. No flag, no API key, no signup. When you ask about
+something the model wasn't trained on (new releases, current events,
+obscure APIs), it decides to call `web_search` on its own; if a
+snippet isn't enough it follows up with `web_fetch`.
+
+```
+you › Flutter 3.19 新加了什么？
+assistant
+  ▸ tool<web_search> → query: "Flutter 3.19 new features"
+  ▸ tool<web_fetch> → https://docs.flutter.dev/release/3-19
+  ▸ 3.19 主要新增了 …
+```
+
+Backed by **Mojeek**'s public search page — an independent web index,
+no API key, no signup, bot-friendly. Coverage on niche or very recent
+queries can be thinner than Google/Bing, but it's reliable from
+scripts and doesn't gate on cookies or sessions. (DDG was the original
+backend but it started serving anti-bot pages in 2026.)
+
+**Turn it off** (offline mode / privacy / CI):
+
+```json
+// ~/.reasonix/config.json
+{ "apiKey": "sk-…", "search": false }
+```
+
+```bash
+# Or one env var (wins over config):
+REASONIX_SEARCH=off npx reasonix
+```
+
+**Bring your own provider** (Kagi, SearXNG, Serper, an internal
+cache) — implement the two tools however you want and register them
+manually:
+
+```ts
+import { ToolRegistry } from "reasonix";
+// Register your own `web_search` / `web_fetch` on a ToolRegistry,
+// then pass it to CacheFirstLoop (or `reasonix chat --no-config`
+// with seedTools via library API).
+```
+
+Inside the session:
+
+```
+reasonix › Flutter 3.19 引入了什么新的 Navigator API？
+assistant
+  ▸ tool<web_search> → query: "Flutter 3.19 new Navigator API"
+    answer: Flutter 3.19 introduces the NavigatorObserver changes …
+    1. Flutter 3.19 Release Notes — https://docs.flutter.dev/…
+    2. What's new in Flutter 3.19 — https://medium.com/…
+  ▸ tool<web_fetch> → https://docs.flutter.dev/release/release-notes/3-19-0
+    (full page text, clipped at 32k)
+  ▸ 3.19 新增了 …
+```
+
+For advanced / self-hosted search (Kagi, SearXNG, internal caches)
+implement the `WebSearchProvider` interface and call
+`registerWebTools(registry, { provider })` yourself, or bridge an
+existing MCP search server via `--mcp`.
+
+---
+
 ## MCP — bring your own tools
 
 Any [MCP](https://spec.modelcontextprotocol.io/) server works. Wizard

package.json

@@ -1,6 +1,6 @@
 {
   "name": "reasonix",
-  "version": "0.4.14",
+  "version": "0.4.15",
   "description": "DeepSeek-native agent framework: cache-first loop, R1 thought harvesting, tool-call repair.",
   "type": "module",
   "bin": {

src/cli/commands/chat.tsx

@@ -1,6 +1,6 @@
 import { render } from "ink";
 import React, { useState } from "react";
-import { loadApiKey } from "../../config.js";
+import { loadApiKey, searchEnabled } from "../../config.js";
 import { loadDotenv } from "../../env.js";
 import { McpClient } from "../../mcp/client.js";
 import { type InspectionReport, inspectMcpServer } from "../../mcp/inspect.js";
@@ -9,6 +9,7 @@ import { parseMcpSpec } from "../../mcp/spec.js";
 import { SseTransport } from "../../mcp/sse.js";
 import { type McpTransport, StdioTransport } from "../../mcp/stdio.js";
 import { ToolRegistry } from "../../tools.js";
+import { registerWebTools } from "../../tools/web.js";
 import { App } from "../ui/App.js";
 import { Setup } from "../ui/Setup.js";
 import type { McpServerSummary } from "../ui/slash.js";
@@ -189,6 +190,14 @@ export async function chatCommand(opts: ChatOptions): Promise<void> {
   }
   const mcpSpecs = successfulSpecs;
 
+  // Register web search/fetch tools unless explicitly disabled. DDG
+  // backs them with no key required; the model invokes them whenever
+  // a question needs info fresher than its training data.
+  if (searchEnabled()) {
+    if (!tools) tools = new ToolRegistry();
+    registerWebTools(tools);
+  }
+
   const { waitUntilExit } = render(
     <Root
       initialKey={initialKey}

src/cli/ui/App.tsx

@@ -328,21 +328,27 @@ export function App({
       }
     }
 
-    // Outside slash mode: ↑/↓ recall prior prompts.
-    const hist = promptHistory.current;
-    if (key.upArrow) {
-      if (hist.length === 0) return;
-      const nextCursor = Math.min(historyCursor.current + 1, hist.length - 1);
-      historyCursor.current = nextCursor;
-      setInput(hist[hist.length - 1 - nextCursor] ?? "");
-      return;
-    }
-    if (key.downArrow) {
-      if (historyCursor.current < 0) return;
-      const nextCursor = historyCursor.current - 1;
-      historyCursor.current = nextCursor;
-      setInput(nextCursor < 0 ? "" : (hist[hist.length - 1 - nextCursor] ?? ""));
-      return;
+    // Outside slash mode: ↑/↓ recall prior prompts — but ONLY when the
+    // buffer is empty. A non-empty buffer hands those keys to
+    // PromptInput for cursor movement (multi-line navigation, or a
+    // no-op on single-line) so the user doesn't accidentally clobber
+    // typed text with a recalled prompt.
+    if (input.length === 0) {
+      const hist = promptHistory.current;
+      if (key.upArrow) {
+        if (hist.length === 0) return;
+        const nextCursor = Math.min(historyCursor.current + 1, hist.length - 1);
+        historyCursor.current = nextCursor;
+        setInput(hist[hist.length - 1 - nextCursor] ?? "");
+        return;
+      }
+      if (key.downArrow) {
+        if (historyCursor.current < 0) return;
+        const nextCursor = historyCursor.current - 1;
+        historyCursor.current = nextCursor;
+        setInput(nextCursor < 0 ? "" : (hist[hist.length - 1 - nextCursor] ?? ""));
+        return;
+      }
     }
   });
 

src/cli/ui/EventLog.tsx

@@ -323,17 +323,25 @@ function StreamingAssistant({ event }: { event: DisplayEvent }) {
       ) : null}
       {tail ? (
         <Text dimColor>▸ {tail}</Text>
+      ) : preFirstByte ? (
+        <Text dimColor italic>
+          {"  connection open, first byte typically in 5-60s depending on model + load"}
+        </Text>
       ) : reasoningOnly ? (
         <Text color="yellow" dimColor>
           {
             "  R1 is thinking before it speaks — body text starts when reasoning completes (typically 20-90s)."
           }
         </Text>
-      ) : (
-        <Text dimColor italic>
-          {"  connection open, first byte typically in 5-60s depending on model + load"}
+      ) : toolCallOnly ? (
+        <Text color="magenta" dimColor>
+          {"  tool-call arguments streaming — the model is about to dispatch a tool"}
         </Text>
-      )}
+      ) : event.reasoning ? (
+        <Text color="yellow" dimColor>
+          {"  R1 still reasoning — body text or tool call arrives when thinking completes"}
+        </Text>
+      ) : null}
     </Box>
   );
 }

src/cli/ui/PromptInput.tsx

@@ -1,6 +1,6 @@
 import { Box, Text, useInput } from "ink";
-import React from "react";
-import { type MultilineKey, processMultilineKey } from "./multiline-keys.js";
+import React, { useRef, useState } from "react";
+import { type MultilineKey, lineAndColumn, processMultilineKey } from "./multiline-keys.js";
 import { useTick } from "./ticker.js";
 
 export interface PromptInputProps {
@@ -12,19 +12,16 @@ export interface PromptInputProps {
 }
 
 /**
- * Input box with multi-line composition support. Behaves like the
- * old single-line prompt by default, but:
- *   - Ctrl+J inserts a newline (universal across terminals).
- *   - Shift+Enter inserts a newline (only on terminals that report
- *     the modifier via CSI-u — many don't).
- *   - `\<Enter>` at end of line acts as bash-style continuation —
- *     always works, even on terminals without modifier reporting.
- *   - Pasted multi-line text lands as-is instead of submitting on
- *     the first '\r'.
+ * Input box with real cursor support. ←/→ move one column, ↑/↓ move
+ * across lines in multi-line buffers, Ctrl+A / Ctrl+E jump to
+ * start/end of the current line. Backspace deletes before cursor,
+ * Delete deletes under cursor. Multi-line composition via Ctrl+J,
+ * Shift+Enter, or bash-style `\<Enter>`.
  *
- * No mid-string insertion cursor — edits happen at the end, like
- * a readline prompt in raw mode. Simple, predictable, enough for
- * 95% of prompts and any code paste.
+ * Cursor state lives locally. When the parent replaces `value` out
+ * of band (history recall, slash completion, setup wizard) the
+ * cursor jumps to end; the `lastLocalValueRef` guards distinguishes
+ * that case from our own edits.
  */
 export function PromptInput({
   value,
@@ -33,15 +30,29 @@ export function PromptInput({
   disabled,
   placeholder,
 }: PromptInputProps) {
-  // Blink from the shared ticker (TICK_MS ≈ 120ms) scaled by 4 so the
-  // visible on/off toggles land around 500ms — standard cursor blink.
-  // When disabled the cursor is hidden entirely.
+  const [cursor, setCursor] = useState(value.length);
+  // Tracks the last `value` we ourselves produced via onChange. If the
+  // incoming `value` prop diverges from this, the parent (or some other
+  // source) replaced the buffer — we reset the cursor to end.
+  const lastLocalValueRef = useRef(value);
+  if (value !== lastLocalValueRef.current) {
+    lastLocalValueRef.current = value;
+    if (cursor !== value.length) {
+      // Conditional setState during render is the "derived state" pattern;
+      // React schedules the re-render and the else branch of the `if`
+      // prevents infinite loops.
+      setCursor(value.length);
+    }
+  }
+
+  // Shared ticker drives the cursor blink. Dividing the tick by 4 lands
+  // the visible on/off cycle around 480ms — standard cursor cadence.
   const tick = useTick();
   const showCursor = disabled ? false : Math.floor(tick / 4) % 2 === 0;
 
   useInput(
     (input, key) => {
-      const keyEvent: MultilineKey = {
+      const ke: MultilineKey = {
         input,
         return: key.return,
         shift: key.shift,
@@ -58,8 +69,14 @@ export function PromptInput({
         pageUp: key.pageUp,
         pageDown: key.pageDown,
       };
-      const action = processMultilineKey(value, keyEvent);
-      if (action.next !== null) onChange(action.next);
+      const action = processMultilineKey(value, cursor, ke);
+      if (action.next !== null) {
+        lastLocalValueRef.current = action.next;
+        onChange(action.next);
+      }
+      if (action.cursor !== null) {
+        setCursor(action.cursor);
+      }
       if (action.submit) onSubmit(action.submitValue ?? value);
     },
     { isActive: !disabled },
@@ -71,17 +88,14 @@ export function PromptInput({
 
   const lines = value.length > 0 ? value.split("\n") : [""];
   const borderColor = disabled ? "gray" : "cyan";
+  const { line: cursorLine, col: cursorCol } = lineAndColumn(value, cursor);
 
   return (
     <Box borderStyle="round" borderColor={borderColor} paddingX={1} flexDirection="column">
       {lines.map((line, i) => {
-        const isLast = i === lines.length - 1;
         const isFirst = i === 0;
         const showPlaceholder = isFirst && value.length === 0;
-        // Line content never survives a re-keying here: the value buffer
-        // fully re-renders on every keystroke, so the array index IS the
-        // stable identity. Biome's rule targets list reordering, which
-        // doesn't apply to an append-only splits-by-newline view.
+        const isCursorLine = i === cursorLine;
         return (
           // biome-ignore lint/suspicious/noArrayIndexKey: stable by construction — lines are derived from `value.split("\n")` and never reordered
           <Box key={i}>
@@ -92,21 +106,59 @@ export function PromptInput({
             ) : (
               <Text dimColor>{"     "}</Text>
             )}
-            {/* When showing the placeholder, the cursor is at position 0 —
-                put it BEFORE the dimmed hint text so it visually matches
-                "you're about to type here," not "you typed the placeholder."
-                When showing real content, the cursor follows the last char
-                of the last line (append-only edit model). */}
-            {showPlaceholder && isLast && !disabled ? (
-              <Text color={borderColor}>{showCursor ? "▌" : " "}</Text>
-            ) : null}
-            {showPlaceholder ? <Text dimColor>{effectivePlaceholder}</Text> : <Text>{line}</Text>}
-            {!showPlaceholder && isLast && !disabled ? (
-              <Text color={borderColor}>{showCursor ? "▌" : " "}</Text>
-            ) : null}
+            {showPlaceholder ? (
+              <>
+                {isCursorLine && !disabled ? (
+                  <Text color={borderColor}>{showCursor ? "▌" : " "}</Text>
+                ) : null}
+                <Text dimColor>{effectivePlaceholder}</Text>
+              </>
+            ) : isCursorLine && !disabled ? (
+              <LineWithCursor
+                line={line}
+                col={cursorCol}
+                showCursor={showCursor}
+                borderColor={borderColor}
+              />
+            ) : (
+              <Text>{line}</Text>
+            )}
           </Box>
         );
       })}
     </Box>
   );
 }
+
+function LineWithCursor({
+  line,
+  col,
+  showCursor,
+  borderColor,
+}: {
+  line: string;
+  col: number;
+  showCursor: boolean;
+  borderColor: "cyan" | "gray";
+}) {
+  const before = line.slice(0, col);
+  const atCursor = line.slice(col, col + 1);
+  const after = line.slice(col + 1);
+  if (atCursor.length === 0) {
+    // Cursor sits past the last char of this line (end-of-line). Render
+    // a trailing block so the user sees where they're typing next.
+    return (
+      <>
+        <Text>{before}</Text>
+        <Text color={borderColor}>{showCursor ? "▌" : " "}</Text>
+      </>
+    );
+  }
+  return (
+    <>
+      <Text>{before}</Text>
+      <Text inverse={showCursor}>{atCursor}</Text>
+      <Text>{after}</Text>
+    </>
+  );
+}

src/cli/ui/multiline-keys.ts

@@ -1,30 +1,34 @@
 /**
- * Pure keystroke → action reducer for MultilineInput.
+ * Pure keystroke → action reducer for PromptInput.
  *
- * Kept separate from the React component so the keyboard
- * semantics are easy to unit-test. MultilineInput is otherwise
- * a thin Ink wrapper that threads its `useInput` callback
- * through this function and applies the result.
+ * Kept separate from the React component so the keyboard semantics
+ * are easy to unit-test. The component threads `useInput` through
+ * this function and applies the returned action.
  *
- * Terminal reality check: on most terminals, pressing Shift+Enter
- * sends the exact same byte as plain Enter — the modifier is lost.
- * That's why the reliable newline insertion path is Ctrl+J (which
- * really is LF, distinct from CR), with Shift+Enter only working
- * on terminals that opt into CSI-u modifier encoding. The `\<Enter>`
- * bash-style continuation is the portable fallback.
+ * Edit model:
+ *   - Full cursor support. ←/→ move one column. ↑/↓ move across
+ *     lines in a multi-line buffer (column preserved when possible).
+ *     Ctrl+A / Ctrl+E jump to start / end of the current line.
+ *   - Backspace deletes the char before the cursor; Delete deletes
+ *     the char under the cursor.
+ *   - Printable chars (including multi-char paste bursts) insert
+ *     at the cursor.
+ *   - Enter submits unless Shift is held (newline), the line ends
+ *     with '\' and cursor is at end (bash-style continuation), or
+ *     the input is Ctrl+J (LF, terminal-universal newline).
+ *   - Parent owns Tab, Esc, PageUp/Down (slash-complete, abort,
+ *     unused). Arrow keys are split: empty buffer → parent (history
+ *     recall); non-empty → child (cursor movement).
  */
 
 export interface MultilineKey {
-  /** Printable character(s) delivered by the key event, may be empty. */
   input: string;
-  /** Modifier + named-key flags as Ink reports them. */
   return?: boolean;
   shift?: boolean;
   ctrl?: boolean;
   meta?: boolean;
   backspace?: boolean;
   delete?: boolean;
-  /** Navigation/editing keys owned by the parent App (slash-nav, history). */
   tab?: boolean;
   upArrow?: boolean;
   downArrow?: boolean;
@@ -36,87 +40,161 @@ export interface MultilineKey {
 }
 
 export interface MultilineAction {
-  /** New value to set on the text buffer. `null` = unchanged. */
+  /** New buffer value. `null` = unchanged. */
   next: string | null;
-  /** When `true`, the caller should fire `onSubmit(next ?? value)`. */
+  /** New cursor position (0..value.length). `null` = unchanged. */
+  cursor: number | null;
+  /** When `true`, fire `onSubmit(submitValue ?? value)`. */
   submit: boolean;
-  /** When submitting, which value to submit (after any final trims). */
   submitValue?: string;
 }
 
 const BACKSLASH_SUFFIX = /\\$/;
 
-/**
- * Decide what to do with a single keystroke against the current buffer.
- *
- * Decisions:
- *   Ctrl+J              → insert '\n' (terminal-universal newline)
- *   Shift+Return        → insert '\n' (works where CSI-u is on)
- *   Enter with "\" end  → replace trailing '\' with '\n' (bash continuation)
- *   Enter               → submit the current buffer
- *   Backspace/Delete    → drop one char from the end (cursor-at-end model)
- *   Printable char(s)   → append at the end
- *
- * We don't model an insertion cursor inside the string: always-at-end
- * keeps the reducer tiny and matches how readline-in-raw-mode feels. The
- * user can still edit by backspacing — for deep edits, exit and compose
- * in $EDITOR.
- */
-export function processMultilineKey(value: string, key: MultilineKey): MultilineAction {
-  // Parent App owns these keys: Tab for slash auto-complete, arrows
-  // for slash-nav/history, Esc for abort. Bowing out here keeps the
-  // text buffer from eating a stray "\t" or arrow escape sequence
-  // when both the parent and child useInput fire on the same event.
-  if (
-    key.tab ||
-    key.upArrow ||
-    key.downArrow ||
-    key.leftArrow ||
-    key.rightArrow ||
-    key.escape ||
-    key.pageUp ||
-    key.pageDown
-  ) {
-    return { next: null, submit: false };
+const NOOP: MultilineAction = { next: null, cursor: null, submit: false };
+
+export function processMultilineKey(
+  value: string,
+  cursor: number,
+  key: MultilineKey,
+): MultilineAction {
+  // Parent-owned keys: Tab (slash-complete), Esc (abort), PageUp/Down.
+  if (key.tab || key.escape || key.pageUp || key.pageDown) {
+    return NOOP;
   }
 
-  // Reliable newline: Ctrl+J arrives as literal '\n' in key.input on
-  // every terminal we've checked. We also accept key.ctrl && input=='j'
-  // as a belt-and-braces match against terminals that normalize the
-  // control-char into its ASCII letter form.
+  // Empty buffer + ↑/↓ → parent handles history recall.
+  if (value.length === 0 && (key.upArrow || key.downArrow)) {
+    return NOOP;
+  }
+
+  // Cursor motion.
+  if (key.leftArrow) {
+    return { next: null, cursor: Math.max(0, cursor - 1), submit: false };
+  }
+  if (key.rightArrow) {
+    return { next: null, cursor: Math.min(value.length, cursor + 1), submit: false };
+  }
+  if (key.upArrow) {
+    const moved = moveCursorUp(value, cursor);
+    return moved === cursor ? NOOP : { next: null, cursor: moved, submit: false };
+  }
+  if (key.downArrow) {
+    const moved = moveCursorDown(value, cursor);
+    return moved === cursor ? NOOP : { next: null, cursor: moved, submit: false };
+  }
+
+  // Emacs-style line jumps (universal across terminals; Home/End aren't
+  // reliably reported by Ink so we don't depend on them).
+  if (key.ctrl && key.input === "a") {
+    return { next: null, cursor: startOfLine(value, cursor), submit: false };
+  }
+  if (key.ctrl && key.input === "e") {
+    return { next: null, cursor: endOfLine(value, cursor), submit: false };
+  }
+
+  // Newline: Ctrl+J (LF literal) or ctrl+'j' normalized form.
   if (key.input === "\n" || (key.ctrl && key.input === "j")) {
-    return { next: `${value}\n`, submit: false };
+    return insertAt(value, cursor, "\n");
   }
 
   if (key.return) {
-    // Shift+Return → newline (when terminal reports the modifier).
-    if (key.shift) {
-      return { next: `${value}\n`, submit: false };
+    if (key.shift) return insertAt(value, cursor, "\n");
+    // Bash-style line continuation: trailing '\' + Enter (only when the
+    // cursor sits at end-of-buffer, so a stray '\' mid-line doesn't
+    // trigger it).
+    if (cursor === value.length && BACKSLASH_SUFFIX.test(value)) {
+      const replaced = `${value.slice(0, -1)}\n`;
+      return { next: replaced, cursor: replaced.length, submit: false };
     }
-    // Bash-style line continuation: trailing '\' + Enter.
-    if (BACKSLASH_SUFFIX.test(value)) {
-      return { next: `${value.slice(0, -1)}\n`, submit: false };
-    }
-    // Plain Enter → submit.
-    return { next: null, submit: true, submitValue: value };
+    return { next: null, cursor: null, submit: true, submitValue: value };
   }
 
-  if (key.backspace || key.delete) {
-    if (value.length === 0) return { next: null, submit: false };
-    return { next: value.slice(0, -1), submit: false };
+  if (key.backspace) {
+    if (cursor === 0) return NOOP;
+    return {
+      next: value.slice(0, cursor - 1) + value.slice(cursor),
+      cursor: cursor - 1,
+      submit: false,
+    };
+  }
+  if (key.delete) {
+    if (cursor === value.length) return NOOP;
+    return {
+      next: value.slice(0, cursor) + value.slice(cursor + 1),
+      cursor,
+      submit: false,
+    };
   }
 
-  // Ignore bare modifier keys (ctrl/meta without accompanying printable).
-  if ((key.ctrl || key.meta) && key.input.length === 0) {
-    return { next: null, submit: false };
+  // Bare modifier events (Ctrl/Meta with no printable) and unhandled
+  // Ctrl-<letter> chords are dropped so a stray Ctrl+L doesn't insert "l".
+  if ((key.ctrl || key.meta) && key.input.length === 0) return NOOP;
+  if (key.ctrl || key.meta) return NOOP;
+
+  // Printable input (may be a multi-char paste; pasted newlines land
+  // inside the buffer rather than triggering submit on the first line).
+  if (key.input.length > 0) {
+    return insertAt(value, cursor, key.input);
   }
 
-  // Printable input (may be a multi-char paste burst — accept as-is;
-  // pasted newlines become part of the buffer rather than triggering
-  // submit on the first line).
-  if (key.input.length > 0 && !key.ctrl && !key.meta) {
-    return { next: value + key.input, submit: false };
+  return NOOP;
+}
+
+function insertAt(value: string, cursor: number, insert: string): MultilineAction {
+  return {
+    next: value.slice(0, cursor) + insert + value.slice(cursor),
+    cursor: cursor + insert.length,
+    submit: false,
+  };
+}
+
+/**
+ * Line + column of a cursor inside a buffer. Exported because the
+ * renderer needs the same mapping for drawing the cursor block on the
+ * right line.
+ */
+export function lineAndColumn(value: string, cursor: number): { line: number; col: number } {
+  let line = 0;
+  let col = 0;
+  const n = Math.min(cursor, value.length);
+  for (let i = 0; i < n; i++) {
+    if (value[i] === "\n") {
+      line++;
+      col = 0;
+    } else {
+      col++;
+    }
   }
+  return { line, col };
+}
+
+function startOfLine(value: string, cursor: number): number {
+  return value.lastIndexOf("\n", cursor - 1) + 1;
+}
+
+function endOfLine(value: string, cursor: number): number {
+  const nl = value.indexOf("\n", cursor);
+  return nl === -1 ? value.length : nl;
+}
+
+function moveCursorUp(value: string, cursor: number): number {
+  const curStart = startOfLine(value, cursor);
+  if (curStart === 0) return cursor; // already on the first line
+  const col = cursor - curStart;
+  const prevEnd = curStart - 1; // the '\n' between the two lines
+  const prevStart = value.lastIndexOf("\n", prevEnd - 1) + 1;
+  const prevLen = prevEnd - prevStart;
+  return prevStart + Math.min(col, prevLen);
+}
 
-  return { next: null, submit: false };
+function moveCursorDown(value: string, cursor: number): number {
+  const nextNl = value.indexOf("\n", cursor);
+  if (nextNl === -1) return cursor; // already on the last line
+  const curStart = startOfLine(value, cursor);
+  const col = cursor - curStart;
+  const nextStart = nextNl + 1;
+  const followingNl = value.indexOf("\n", nextStart);
+  const nextLen = (followingNl === -1 ? value.length : followingNl) - nextStart;
+  return nextStart + Math.min(col, nextLen);
 }

src/config.ts

@@ -44,6 +44,12 @@ export interface ReasonixConfig {
   session?: string | null;
   /** Marks that `reasonix setup` has completed at least once. */
   setupCompleted?: boolean;
+  /**
+   * Whether `web_search` + `web_fetch` tools are registered. Default:
+   * enabled (no key required — backed by DuckDuckGo's public HTML
+   * endpoint). Set to `false` to keep the session offline.
+   */
+  search?: boolean;
 }
 
 export function defaultConfigPath(): string {
@@ -78,6 +84,19 @@ export function loadApiKey(path: string = defaultConfigPath()): string | undefin
   return readConfig(path).apiKey;
 }
 
+/**
+ * Resolve whether web-search tools should be registered. Default: on.
+ * Env `REASONIX_SEARCH=off` or config `search: false` turns it off.
+ * Any other value falls through to enabled.
+ */
+export function searchEnabled(path: string = defaultConfigPath()): boolean {
+  const env = process.env.REASONIX_SEARCH;
+  if (env === "off" || env === "false" || env === "0") return false;
+  const cfg = readConfig(path).search;
+  if (cfg === false) return false;
+  return true;
+}
+
 export function saveApiKey(key: string, path: string = defaultConfigPath()): void {
   const cfg = readConfig(path);
   cfg.apiKey = key.trim();

src/index.ts

@@ -33,6 +33,21 @@ export { ToolRegistry } from "./tools.js";
 export type { ToolDefinition, ToolCallContext } from "./tools.js";
 export { registerFilesystemTools } from "./tools/filesystem.js";
 export type { FilesystemToolsOptions } from "./tools/filesystem.js";
+export {
+  formatSearchResults,
+  htmlToText,
+  parseMojeekResults,
+  registerWebTools,
+  webFetch,
+  webSearch,
+} from "./tools/web.js";
+export type {
+  PageContent,
+  SearchResult,
+  WebFetchOptions,
+  WebSearchOptions,
+  WebToolsOptions,
+} from "./tools/web.js";
 
 export {
   SessionStats,
@@ -186,4 +201,4 @@ export type {
   JSONSchema,
 } from "./types.js";
 
-export const VERSION = "0.4.14";
+export const VERSION = "0.4.15";

src/tools/web.ts

@@ -0,0 +1,299 @@
+/**
+ * Built-in web search + fetch tools.
+ *
+ *   - `web_search(query, topK?)` — Mojeek's public search page. No API
+ *     key, no signup. We originally shipped this backed by DuckDuckGo's
+ *     HTML endpoint, but DDG started serving anti-bot interstitials
+ *     (HTTP 202 with a challenge page) for every unauthenticated POST.
+ *     Mojeek runs its own independent index, is bot-friendly, and
+ *     returns parseable HTML.
+ *   - `web_fetch(url)` — HTTP GET + naïve HTML-to-text extraction.
+ *
+ * Both are registered by default on `reasonix chat` / `reasonix code`;
+ * set `search: false` in config (or `REASONIX_SEARCH=off`) to turn
+ * them off. The model decides when to call them based on the query —
+ * no slash command required.
+ */
+
+import type { ToolRegistry } from "../tools.js";
+
+export interface SearchResult {
+  title: string;
+  url: string;
+  snippet: string;
+}
+
+export interface PageContent {
+  url: string;
+  title?: string;
+  text: string;
+  /** True when the extracted text was clipped to fit the cap. */
+  truncated: boolean;
+}
+
+export interface WebFetchOptions {
+  /** Max bytes of extracted text. Defaults to 32_000 to match tool-result cap. */
+  maxChars?: number;
+  /** Timeout in ms. Defaults to 15_000. */
+  timeoutMs?: number;
+  signal?: AbortSignal;
+}
+
+export interface WebSearchOptions {
+  topK?: number;
+  signal?: AbortSignal;
+}
+
+const DEFAULT_FETCH_MAX_CHARS = 32_000;
+const DEFAULT_FETCH_TIMEOUT_MS = 15_000;
+const DEFAULT_TOPK = 5;
+// Real-browser UA. Servers like Mojeek are bot-friendly but still gate
+// obvious scraper UAs; a stock Chrome string avoids the fast-path block.
+const USER_AGENT =
+  "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/120.0.0.0 Safari/537.36";
+const MOJEEK_ENDPOINT = "https://www.mojeek.com/search";
+
+/**
+ * Search the public web via Mojeek. Returns up to `topK` ranked
+ * results with title, url, snippet.
+ *
+ * Mojeek is an independent index (not a Google/Bing front-end) which
+ * means coverage on niche or very recent topics can be thinner, but
+ * it's reliable from scripts and doesn't gate on cookies or sessions.
+ * If the response has 0 results we distinguish "truly empty" from
+ * "layout changed or blocked" so the caller isn't left guessing.
+ */
+export async function webSearch(
+  query: string,
+  opts: WebSearchOptions = {},
+): Promise<SearchResult[]> {
+  const topK = Math.max(1, Math.min(10, opts.topK ?? DEFAULT_TOPK));
+  const resp = await fetch(`${MOJEEK_ENDPOINT}?q=${encodeURIComponent(query)}`, {
+    headers: {
+      "User-Agent": USER_AGENT,
+      Accept: "text/html,application/xhtml+xml,application/xml;q=0.9",
+      "Accept-Language": "en-US,en;q=0.9",
+    },
+    signal: opts.signal,
+    redirect: "follow",
+  });
+  if (!resp.ok) throw new Error(`web_search ${resp.status}`);
+  const html = await resp.text();
+  const results = parseMojeekResults(html).slice(0, topK);
+  if (results.length === 0) {
+    if (/no results found|did not match any documents/i.test(html)) return [];
+    if (/captcha|verify you are human|access denied|forbidden/i.test(html)) {
+      throw new Error("web_search: Mojeek anti-bot page — rate-limited or blocked");
+    }
+    throw new Error(
+      `web_search: 0 results but response doesn't look like a real empty page (${html.length} chars, first 120: ${html.slice(0, 120).replace(/\s+/g, " ")})`,
+    );
+  }
+  return results;
+}
+
+/**
+ * Extract results from a Mojeek search page.
+ *
+ * Mojeek's stable shape (as of April 2026):
+ *   <a … class="ob" href="URL"> … breadcrumb … </a>
+ *   <h2><a class="title" href="URL">Title</a></h2>
+ *   <p class="s">snippet text …</p>
+ *
+ * We do two tolerant passes — title anchors, then snippet paragraphs —
+ * and pair them positionally. Attribute order inside a tag varies
+ * between versions, so each pass captures the whole element and we
+ * re-extract href / inner text with a second regex. Exported for
+ * unit testing against a fixture.
+ */
+export function parseMojeekResults(html: string): SearchResult[] {
+  const titles: string[] = [];
+  const titleAnchorRe = /<a\b[^>]*\bclass="title"[^>]*>[\s\S]*?<\/a>/g;
+  let m: RegExpExecArray | null;
+  while (true) {
+    m = titleAnchorRe.exec(html);
+    if (m === null) break;
+    titles.push(m[0]);
+  }
+
+  const snippets: string[] = [];
+  const snippetRe = /<p\b[^>]*\bclass="s"[^>]*>([\s\S]*?)<\/p>/g;
+  while (true) {
+    m = snippetRe.exec(html);
+    if (m === null) break;
+    snippets.push(m[1] ?? "");
+  }
+
+  const hrefRe = /href="([^"]+)"/;
+  const innerRe = /<a\b[^>]*>([\s\S]*?)<\/a>/;
+  const results: SearchResult[] = [];
+  for (let i = 0; i < titles.length; i++) {
+    const anchor = titles[i]!;
+    const hrefMatch = anchor.match(hrefRe);
+    const innerMatch = anchor.match(innerRe);
+    if (!hrefMatch?.[1]) continue;
+    results.push({
+      title: decodeHtmlEntities(stripHtml(innerMatch?.[1] ?? "")).trim(),
+      url: hrefMatch[1],
+      snippet: decodeHtmlEntities(stripHtml(snippets[i] ?? ""))
+        .replace(/\s+/g, " ")
+        .trim(),
+    });
+  }
+  return results;
+}
+
+/**
+ * Download a URL, strip HTML down to readable text, return it. Times
+ * out at 15s, caps extracted text at 32k chars to fit the tool-result
+ * budget.
+ */
+export async function webFetch(url: string, opts: WebFetchOptions = {}): Promise<PageContent> {
+  const maxChars = opts.maxChars ?? DEFAULT_FETCH_MAX_CHARS;
+  const timeoutMs = opts.timeoutMs ?? DEFAULT_FETCH_TIMEOUT_MS;
+  const ctl = new AbortController();
+  const timer = setTimeout(() => ctl.abort(), timeoutMs);
+  // Forward the caller's abort too so an Esc during a long fetch is respected.
+  const cancel = () => ctl.abort();
+  opts.signal?.addEventListener("abort", cancel, { once: true });
+  let resp: Response;
+  try {
+    resp = await fetch(url, {
+      headers: { "User-Agent": USER_AGENT, Accept: "text/html,text/plain,*/*" },
+      signal: ctl.signal,
+      redirect: "follow",
+    });
+  } finally {
+    clearTimeout(timer);
+    opts.signal?.removeEventListener("abort", cancel);
+  }
+  if (!resp.ok) throw new Error(`web_fetch ${resp.status} for ${url}`);
+  const contentType = resp.headers.get("content-type") ?? "";
+  const raw = await resp.text();
+  const title = extractTitle(raw);
+  const text = contentType.includes("text/html") ? htmlToText(raw) : raw;
+  const truncated = text.length > maxChars;
+  const finalText = truncated
+    ? `${text.slice(0, maxChars)}\n\n[… truncated ${text.length - maxChars} chars …]`
+    : text;
+  return { url, title, text: finalText, truncated };
+}
+
+/**
+ * Strip HTML to readable text. Removes scripts/styles/nav/footer/aside
+ * blocks first, then tags, then collapses whitespace. Not a Readability
+ * clone — purpose-built to keep the extracted text small enough for the
+ * tool-result budget while preserving paragraph breaks.
+ */
+export function htmlToText(html: string): string {
+  let s = html;
+  s = s.replace(/<script[\s\S]*?<\/script>/gi, "");
+  s = s.replace(/<style[\s\S]*?<\/style>/gi, "");
+  s = s.replace(/<noscript[\s\S]*?<\/noscript>/gi, "");
+  s = s.replace(/<nav[\s\S]*?<\/nav>/gi, "");
+  s = s.replace(/<footer[\s\S]*?<\/footer>/gi, "");
+  s = s.replace(/<aside[\s\S]*?<\/aside>/gi, "");
+  s = s.replace(/<svg[\s\S]*?<\/svg>/gi, "");
+  // Preserve paragraph breaks by turning common block tags into newlines.
+  s = s.replace(/<\/?(p|div|br|h[1-6]|li|tr|section|article)\b[^>]*>/gi, "\n");
+  s = s.replace(/<[^>]+>/g, "");
+  s = decodeHtmlEntities(s);
+  s = s.replace(/[ \t]+/g, " ");
+  s = s.replace(/\n[ \t]+/g, "\n");
+  s = s.replace(/\n{3,}/g, "\n\n");
+  return s.trim();
+}
+
+function stripHtml(s: string): string {
+  return s.replace(/<[^>]+>/g, "");
+}
+
+function decodeHtmlEntities(s: string): string {
+  return s
+    .replace(/&nbsp;/g, " ")
+    .replace(/&amp;/g, "&")
+    .replace(/&lt;/g, "<")
+    .replace(/&gt;/g, ">")
+    .replace(/&quot;/g, '"')
+    .replace(/&#39;/g, "'");
+}
+
+function extractTitle(html: string): string | undefined {
+  const m = html.match(/<title[^>]*>([\s\S]*?)<\/title>/i);
+  if (!m?.[1]) return undefined;
+  return m[1].replace(/\s+/g, " ").trim() || undefined;
+}
+
+export interface WebToolsOptions {
+  /** Default top-K for `web_search` when the model doesn't specify. */
+  defaultTopK?: number;
+  /** Byte cap for `web_fetch` extracted text. */
+  maxFetchChars?: number;
+}
+
+/**
+ * Register `web_search` + `web_fetch` on a ToolRegistry. The model
+ * invokes them automatically when a question needs current info —
+ * no slash command from the user is required.
+ */
+export function registerWebTools(registry: ToolRegistry, opts: WebToolsOptions = {}): ToolRegistry {
+  const defaultTopK = opts.defaultTopK ?? DEFAULT_TOPK;
+  const maxFetchChars = opts.maxFetchChars ?? DEFAULT_FETCH_MAX_CHARS;
+
+  registry.register({
+    name: "web_search",
+    description:
+      "Search the public web. Returns ranked results with title, url, and snippet. Use this when the question needs information more current than your training data, when you're unsure of a factual detail, or when the user asks about a specific webpage/library/release you haven't seen.",
+    parameters: {
+      type: "object",
+      properties: {
+        query: { type: "string", description: "Natural-language search query." },
+        topK: {
+          type: "integer",
+          description: `Number of results to return (1..10). Default ${defaultTopK}.`,
+        },
+      },
+      required: ["query"],
+    },
+    fn: async (args: { query: string; topK?: number }, ctx) => {
+      const results = await webSearch(args.query, {
+        topK: args.topK ?? defaultTopK,
+        signal: ctx?.signal,
+      });
+      return formatSearchResults(args.query, results);
+    },
+  });
+
+  registry.register({
+    name: "web_fetch",
+    description:
+      "Download a URL and return its visible text content (HTML pages get scripts/styles/nav stripped). Truncated at the tool-result cap. Use after web_search when a snippet isn't enough.",
+    parameters: {
+      type: "object",
+      properties: {
+        url: { type: "string", description: "Absolute http:// or https:// URL." },
+      },
+      required: ["url"],
+    },
+    fn: async (args: { url: string }, ctx) => {
+      if (!/^https?:\/\//i.test(args.url)) {
+        throw new Error("web_fetch: url must start with http:// or https://");
+      }
+      const page = await webFetch(args.url, { maxChars: maxFetchChars, signal: ctx?.signal });
+      const header = page.title ? `${page.title}\n${page.url}` : page.url;
+      return `${header}\n\n${page.text}`;
+    },
+  });
+
+  return registry;
+}
+
+export function formatSearchResults(query: string, results: SearchResult[]): string {
+  const lines: string[] = [`query: ${query}`, `\nresults (${results.length}):`];
+  results.forEach((r, i) => {
+    lines.push(`\n${i + 1}. ${r.title}`);
+    lines.push(`   ${r.url}`);
+    if (r.snippet) lines.push(`   ${r.snippet}`);
+  });
+  return lines.join("\n");
+}

tests/config.test.ts

@@ -8,6 +8,7 @@ import {
   readConfig,
   redactKey,
   saveApiKey,
+  searchEnabled,
   writeConfig,
 } from "../src/config.js";
 
@@ -15,12 +16,15 @@ describe("config", () => {
   let dir: string;
   let path: string;
   const originalEnv = process.env.DEEPSEEK_API_KEY;
+  const originalSearch = process.env.REASONIX_SEARCH;
 
   beforeEach(() => {
     dir = mkdtempSync(join(tmpdir(), "reasonix-test-"));
     path = join(dir, "config.json");
     // biome-ignore lint/performance/noDelete: the string "undefined" leaks into process.env otherwise
     delete process.env.DEEPSEEK_API_KEY;
+    // biome-ignore lint/performance/noDelete: same reason
+    delete process.env.REASONIX_SEARCH;
   });
 
   afterEach(() => {
@@ -31,6 +35,12 @@ describe("config", () => {
     } else {
       process.env.DEEPSEEK_API_KEY = originalEnv;
     }
+    if (originalSearch === undefined) {
+      // biome-ignore lint/performance/noDelete: same reason
+      delete process.env.REASONIX_SEARCH;
+    } else {
+      process.env.REASONIX_SEARCH = originalSearch;
+    }
   });
 
   it("readConfig returns {} when file is missing", () => {
@@ -106,4 +116,33 @@ describe("config", () => {
     const loaded = readConfig(path);
     expect(loaded.session).toBeNull();
   });
+
+  it("searchEnabled defaults to true with no config and no env", () => {
+    expect(searchEnabled(path)).toBe(true);
+  });
+
+  it("searchEnabled honours `search: false` in the config file", () => {
+    writeConfig({ apiKey: "sk-test123abcdefghijkl", search: false }, path);
+    expect(searchEnabled(path)).toBe(false);
+  });
+
+  it("searchEnabled honours REASONIX_SEARCH=off/false/0", () => {
+    process.env.REASONIX_SEARCH = "off";
+    expect(searchEnabled(path)).toBe(false);
+    process.env.REASONIX_SEARCH = "false";
+    expect(searchEnabled(path)).toBe(false);
+    process.env.REASONIX_SEARCH = "0";
+    expect(searchEnabled(path)).toBe(false);
+  });
+
+  it("searchEnabled stays true for unrelated env values", () => {
+    process.env.REASONIX_SEARCH = "on";
+    expect(searchEnabled(path)).toBe(true);
+  });
+
+  it("env off beats config true", () => {
+    writeConfig({ apiKey: "sk-test123abcdefghijkl", search: true }, path);
+    process.env.REASONIX_SEARCH = "off";
+    expect(searchEnabled(path)).toBe(false);
+  });
 });

tests/multiline-keys.test.ts

@@ -1,112 +1,233 @@
 import { describe, expect, it } from "vitest";
-import { type MultilineKey, processMultilineKey } from "../src/cli/ui/multiline-keys.js";
+import {
+  type MultilineKey,
+  lineAndColumn,
+  processMultilineKey,
+} from "../src/cli/ui/multiline-keys.js";
 
 function key(overrides: Partial<MultilineKey> = {}): MultilineKey {
   return { input: "", ...overrides };
 }
 
-describe("processMultilineKey", () => {
-  it("appends printable input at the end", () => {
-    const r = processMultilineKey("hel", key({ input: "l" }));
-    expect(r).toEqual({ next: "hell", submit: false });
+describe("processMultilineKey — inserts at cursor", () => {
+  it("inserts a printable char at the cursor (mid-string)", () => {
+    // "heo", cursor after "he" → insert "ll" in the middle
+    const r = processMultilineKey("heo", 2, key({ input: "ll" }));
+    expect(r.next).toBe("hello");
+    expect(r.cursor).toBe(4);
   });
 
-  it("appends a multi-char paste burst as-is", () => {
-    const r = processMultilineKey("", key({ input: "line1\nline2" }));
-    expect(r.next).toBe("line1\nline2");
+  it("appends when cursor is at end", () => {
+    const r = processMultilineKey("hel", 3, key({ input: "l" }));
+    expect(r.next).toBe("hell");
+    expect(r.cursor).toBe(4);
+  });
+
+  it("inserts at position 0", () => {
+    const r = processMultilineKey("bc", 0, key({ input: "a" }));
+    expect(r.next).toBe("abc");
+    expect(r.cursor).toBe(1);
+  });
+
+  it("multi-char paste burst inserts as a block", () => {
+    const r = processMultilineKey("ab", 1, key({ input: "XYZ" }));
+    expect(r.next).toBe("aXYZb");
+    expect(r.cursor).toBe(4);
+  });
+
+  it("pasted newline lands inline (not a submit)", () => {
+    const r = processMultilineKey("a", 1, key({ input: "\nmore" }));
+    expect(r.next).toBe("a\nmore");
     expect(r.submit).toBe(false);
   });
+});
 
-  it("Enter on a plain buffer submits", () => {
-    const r = processMultilineKey("hi", key({ return: true }));
+describe("processMultilineKey — submit + newline", () => {
+  it("Enter submits by default", () => {
+    const r = processMultilineKey("hi", 2, key({ return: true }));
     expect(r.submit).toBe(true);
     expect(r.submitValue).toBe("hi");
-    expect(r.next).toBeNull();
   });
 
-  it("Enter with trailing backslash converts to newline (bash-style continuation)", () => {
-    const r = processMultilineKey("line1\\", key({ return: true }));
+  it("Shift+Enter inserts a newline at the cursor", () => {
+    const r = processMultilineKey("abc", 1, key({ return: true, shift: true }));
+    expect(r.next).toBe("a\nbc");
+    expect(r.cursor).toBe(2);
+    expect(r.submit).toBe(false);
+  });
+
+  it("Ctrl+J inserts a newline at the cursor (ASCII LF form)", () => {
+    const r = processMultilineKey("abc", 2, key({ input: "\n" }));
+    expect(r.next).toBe("ab\nc");
+    expect(r.cursor).toBe(3);
+  });
+
+  it("Ctrl+J normalized as ctrl+'j' also inserts a newline", () => {
+    const r = processMultilineKey("abc", 2, key({ input: "j", ctrl: true }));
+    expect(r.next).toBe("ab\nc");
+  });
+
+  it("Enter with trailing \\\\ at end-of-buffer → bash continuation", () => {
+    const v = "line1\\";
+    const r = processMultilineKey(v, v.length, key({ return: true }));
     expect(r.submit).toBe(false);
     expect(r.next).toBe("line1\n");
+    expect(r.cursor).toBe(6);
   });
 
-  it("Shift+Enter inserts a newline without submitting", () => {
-    const r = processMultilineKey("abc", key({ return: true, shift: true }));
-    expect(r.submit).toBe(false);
-    expect(r.next).toBe("abc\n");
-  });
-
-  it("Ctrl+J (ASCII LF) inserts a newline", () => {
-    const r = processMultilineKey("abc", key({ input: "\n" }));
-    expect(r.next).toBe("abc\n");
-    expect(r.submit).toBe(false);
-  });
-
-  it("Ctrl+J reported as ctrl+'j' (normalized form) also inserts a newline", () => {
-    const r = processMultilineKey("abc", key({ input: "j", ctrl: true }));
-    expect(r.next).toBe("abc\n");
-    expect(r.submit).toBe(false);
-  });
-
-  it("Backspace drops one char from the end", () => {
-    const r = processMultilineKey("abcd", key({ backspace: true }));
-    expect(r.next).toBe("abc");
-  });
-
-  it("Backspace on empty buffer is a no-op", () => {
-    const r = processMultilineKey("", key({ backspace: true }));
-    expect(r.next).toBeNull();
-    expect(r.submit).toBe(false);
-  });
-
-  it("Backspace can delete across newlines", () => {
-    const r = processMultilineKey("a\nb", key({ backspace: true }));
-    expect(r.next).toBe("a\n");
-    const r2 = processMultilineKey(r.next!, key({ backspace: true }));
-    expect(r2.next).toBe("a");
-  });
-
-  it("Delete behaves like Backspace (cursor-at-end model)", () => {
-    const r = processMultilineKey("xyz", key({ delete: true }));
-    expect(r.next).toBe("xy");
-  });
-
-  it("Tab is ignored (parent handles slash auto-complete)", () => {
-    const r = processMultilineKey("/he", key({ tab: true, input: "" }));
-    expect(r.next).toBeNull();
-    expect(r.submit).toBe(false);
-  });
-
-  it("Tab with '\\t' in input is still ignored", () => {
-    const r = processMultilineKey("x", key({ tab: true, input: "\t" }));
+  it("Enter with \\\\ mid-buffer (cursor not at end) does NOT trigger continuation", () => {
+    // User has "foo\\bar" and hits Enter with cursor after "foo\\" — that's
+    // a real edit, not a continuation marker. Submit instead.
+    const r = processMultilineKey("foo\\bar", 4, key({ return: true }));
+    expect(r.submit).toBe(true);
     expect(r.next).toBeNull();
   });
 
-  it("Arrow keys are ignored (parent handles slash-nav + history)", () => {
-    expect(processMultilineKey("x", key({ upArrow: true })).next).toBeNull();
-    expect(processMultilineKey("x", key({ downArrow: true })).next).toBeNull();
-    expect(processMultilineKey("x", key({ leftArrow: true })).next).toBeNull();
-    expect(processMultilineKey("x", key({ rightArrow: true })).next).toBeNull();
-  });
-
-  it("Escape is ignored (parent handles abort)", () => {
-    const r = processMultilineKey("x", key({ escape: true }));
-    expect(r.next).toBeNull();
-  });
-
-  it("Ctrl+<letter> that isn't j is dropped (no accidental text insert)", () => {
-    const r = processMultilineKey("x", key({ input: "c", ctrl: true }));
-    expect(r.next).toBeNull();
-  });
-
-  it("Meta (Alt) key events are dropped", () => {
-    const r = processMultilineKey("x", key({ input: "a", meta: true }));
-    expect(r.next).toBeNull();
-  });
-
-  it("Plain return on an empty buffer still submits (empty message case handled upstream)", () => {
-    const r = processMultilineKey("", key({ return: true }));
+  it("plain Enter on an empty buffer still submits", () => {
+    const r = processMultilineKey("", 0, key({ return: true }));
     expect(r.submit).toBe(true);
     expect(r.submitValue).toBe("");
   });
 });
+
+describe("processMultilineKey — deletion", () => {
+  it("Backspace deletes the char BEFORE the cursor, cursor moves back", () => {
+    const r = processMultilineKey("abcd", 2, key({ backspace: true }));
+    expect(r.next).toBe("acd");
+    expect(r.cursor).toBe(1);
+  });
+
+  it("Backspace at cursor 0 is a no-op", () => {
+    const r = processMultilineKey("abc", 0, key({ backspace: true }));
+    expect(r.next).toBeNull();
+    expect(r.cursor).toBeNull();
+  });
+
+  it("Delete removes the char AT the cursor, cursor stays", () => {
+    const r = processMultilineKey("abcd", 1, key({ delete: true }));
+    expect(r.next).toBe("acd");
+    expect(r.cursor).toBe(1);
+  });
+
+  it("Delete at end-of-buffer is a no-op", () => {
+    const r = processMultilineKey("abc", 3, key({ delete: true }));
+    expect(r.next).toBeNull();
+  });
+
+  it("Backspace across a newline removes the newline", () => {
+    const r = processMultilineKey("a\nb", 2, key({ backspace: true }));
+    expect(r.next).toBe("ab");
+    expect(r.cursor).toBe(1);
+  });
+});
+
+describe("processMultilineKey — cursor motion", () => {
+  it("←/→ clamp to the buffer", () => {
+    expect(processMultilineKey("abc", 2, key({ leftArrow: true })).cursor).toBe(1);
+    expect(processMultilineKey("abc", 0, key({ leftArrow: true })).cursor).toBe(0);
+    expect(processMultilineKey("abc", 2, key({ rightArrow: true })).cursor).toBe(3);
+    expect(processMultilineKey("abc", 3, key({ rightArrow: true })).cursor).toBe(3);
+  });
+
+  it("↑/↓ on a single-line non-empty buffer is a no-op (doesn't eat for history)", () => {
+    // Single-line buffer: moveCursorUp has nowhere to go. Returns NOOP so
+    // the parent can decide what to do — but parent only does history on
+    // empty buffer, so the net effect is "nothing happens."
+    const up = processMultilineKey("hello", 3, key({ upArrow: true }));
+    expect(up).toEqual({ next: null, cursor: null, submit: false });
+    const down = processMultilineKey("hello", 3, key({ downArrow: true }));
+    expect(down).toEqual({ next: null, cursor: null, submit: false });
+  });
+
+  it("↑/↓ on an empty buffer defers to the parent (history recall)", () => {
+    expect(processMultilineKey("", 0, key({ upArrow: true }))).toEqual({
+      next: null,
+      cursor: null,
+      submit: false,
+    });
+    expect(processMultilineKey("", 0, key({ downArrow: true }))).toEqual({
+      next: null,
+      cursor: null,
+      submit: false,
+    });
+  });
+
+  it("↑ moves cursor to the previous line, preserving column when possible", () => {
+    //  line 0: "hello" (cols 0-5)
+    //  line 1: "world" (cols 0-5)
+    //  cursor at col 3 on line 1 = index 5 (for "\n") + 3 + 1 = 9
+    const v = "hello\nworld";
+    const up = processMultilineKey(v, 9, key({ upArrow: true }));
+    // target: line 0 col 3 → index 3
+    expect(up.cursor).toBe(3);
+  });
+
+  it("↑ clamps column when the previous line is shorter", () => {
+    const v = "hi\nworld";
+    // cursor at line 1 col 4 = index 3 + 4 = 7
+    const up = processMultilineKey(v, 7, key({ upArrow: true }));
+    // target: line 0 col min(4, 2) = 2 → index 2
+    expect(up.cursor).toBe(2);
+  });
+
+  it("↓ moves cursor to the next line, preserving column", () => {
+    const v = "hello\nworld";
+    // cursor at line 0 col 2 = index 2
+    const down = processMultilineKey(v, 2, key({ downArrow: true }));
+    // target: line 1 col 2 → index 6 + 2 = 8
+    expect(down.cursor).toBe(8);
+  });
+
+  it("↓ clamps column when the next line is shorter", () => {
+    const v = "world\nhi";
+    // cursor at line 0 col 4 = index 4
+    const down = processMultilineKey(v, 4, key({ downArrow: true }));
+    // target: line 1 col min(4, 2) = 2 → index 6 + 2 = 8
+    expect(down.cursor).toBe(8);
+  });
+
+  it("Ctrl+A jumps to start of current line, Ctrl+E to end", () => {
+    const v = "one\ntwo\nthree";
+    // cursor mid-"two" at index 5 (o in two)
+    expect(processMultilineKey(v, 5, key({ input: "a", ctrl: true })).cursor).toBe(4);
+    expect(processMultilineKey(v, 5, key({ input: "e", ctrl: true })).cursor).toBe(7);
+  });
+});
+
+describe("processMultilineKey — parent-owned keys are ignored", () => {
+  it("Tab / Escape / PageUp / PageDown are dropped", () => {
+    expect(processMultilineKey("x", 1, key({ tab: true }))).toEqual({
+      next: null,
+      cursor: null,
+      submit: false,
+    });
+    expect(processMultilineKey("x", 1, key({ escape: true })).next).toBeNull();
+    expect(processMultilineKey("x", 1, key({ pageUp: true })).next).toBeNull();
+    expect(processMultilineKey("x", 1, key({ pageDown: true })).next).toBeNull();
+  });
+
+  it("unhandled Ctrl-<letter> chords are dropped (no accidental insert)", () => {
+    const r = processMultilineKey("x", 1, key({ input: "c", ctrl: true }));
+    expect(r.next).toBeNull();
+    expect(r.cursor).toBeNull();
+  });
+
+  it("Meta (Alt) key events are dropped", () => {
+    const r = processMultilineKey("x", 1, key({ input: "a", meta: true }));
+    expect(r.next).toBeNull();
+  });
+});
+
+describe("lineAndColumn", () => {
+  it("maps a cursor offset to {line, col}", () => {
+    expect(lineAndColumn("abc", 2)).toEqual({ line: 0, col: 2 });
+    expect(lineAndColumn("abc\ndef", 4)).toEqual({ line: 1, col: 0 });
+    expect(lineAndColumn("abc\ndef", 6)).toEqual({ line: 1, col: 2 });
+    expect(lineAndColumn("a\n\nb", 2)).toEqual({ line: 1, col: 0 });
+    expect(lineAndColumn("a\n\nb", 3)).toEqual({ line: 2, col: 0 });
+  });
+
+  it("clamps cursor values past value.length", () => {
+    expect(lineAndColumn("abc", 99)).toEqual({ line: 0, col: 3 });
+  });
+});

tests/web-tools.test.ts

@@ -0,0 +1,292 @@
+import { describe, expect, it, vi } from "vitest";
+import { ToolRegistry } from "../src/tools.js";
+import {
+  formatSearchResults,
+  htmlToText,
+  parseMojeekResults,
+  registerWebTools,
+  webFetch,
+  webSearch,
+} from "../src/tools/web.js";
+
+describe("htmlToText", () => {
+  it("strips script/style/nav/footer and preserves paragraph breaks", () => {
+    const html = `
+      <html><head><title>x</title><style>body{color:red}</style></head>
+      <body>
+        <nav><a>skip</a></nav>
+        <p>Hello <strong>world</strong>.</p>
+        <p>Second paragraph.</p>
+        <script>evil()</script>
+        <footer>fine print</footer>
+      </body></html>
+    `;
+    const out = htmlToText(html);
+    expect(out).toContain("Hello world.");
+    expect(out).toContain("Second paragraph.");
+    expect(out).not.toContain("evil");
+    expect(out).not.toContain("skip");
+    expect(out).not.toContain("fine print");
+    expect(out).not.toContain("color:red");
+    expect(out).toMatch(/Hello world\.\n\nSecond paragraph\./);
+  });
+
+  it("decodes the common entities", () => {
+    expect(htmlToText("<p>a &amp; b &lt;c&gt; &quot;d&quot;</p>")).toContain('a & b <c> "d"');
+  });
+
+  it("collapses whitespace runs but keeps paragraph breaks", () => {
+    const out = htmlToText("<p>one    two</p><p>three</p>");
+    expect(out).toBe("one two\n\nthree");
+  });
+});
+
+describe("parseMojeekResults", () => {
+  // Fixture mirrors the shape Mojeek actually returns as of April 2026.
+  const sampleHtml = `
+    <ul class="results">
+      <li>
+        <a title="https://example.com/a" href="https://example.com/a" class="ob">
+          <p class="i"><span class="url">https://example.com</span></p>
+        </a>
+        <h2>
+          <a class="title" title="https://example.com/a" href="https://example.com/a">
+            Flutter 3.19 release notes
+          </a>
+        </h2>
+        <p class="s">
+          Flutter 3.19 introduces <strong>new Navigator</strong>&nbsp;APIs &amp; more.
+        </p>
+      </li>
+      <li>
+        <a href="https://medium.com/flutter/x" class="ob">
+          <p class="i"><span class="url">medium.com</span></p>
+        </a>
+        <h2>
+          <a class="title" href="https://medium.com/flutter/x">What's new in 3.19</a>
+        </h2>
+        <p class="s">An overview post.</p>
+      </li>
+    </ul>
+  `;
+
+  it("extracts title/url/snippet from the expected markup", () => {
+    const items = parseMojeekResults(sampleHtml);
+    expect(items).toHaveLength(2);
+    expect(items[0]).toEqual({
+      title: "Flutter 3.19 release notes",
+      url: "https://example.com/a",
+      snippet: "Flutter 3.19 introduces new Navigator APIs & more.",
+    });
+    expect(items[1]).toEqual({
+      title: "What's new in 3.19",
+      url: "https://medium.com/flutter/x",
+      snippet: "An overview post.",
+    });
+  });
+
+  it("returns empty on markup that doesn't match the expected shape", () => {
+    expect(parseMojeekResults("<html><body>nothing here</body></html>")).toEqual([]);
+  });
+
+  it("tolerates attribute-order swaps (href before class)", () => {
+    const html = `
+      <a href="https://example.com/z" class="title">Title Z</a>
+      <p class="s">Snippet Z.</p>
+    `;
+    const items = parseMojeekResults(html);
+    expect(items).toHaveLength(1);
+    expect(items[0]).toEqual({
+      title: "Title Z",
+      url: "https://example.com/z",
+      snippet: "Snippet Z.",
+    });
+  });
+
+  it("handles a title with no snippet sibling (empty snippet)", () => {
+    const html = `<a class="title" href="https://example.com/s">Solo</a>`;
+    const items = parseMojeekResults(html);
+    expect(items).toHaveLength(1);
+    expect(items[0]?.snippet).toBe("");
+  });
+});
+
+describe("webSearch", () => {
+  const twoResultsHtml = `
+    <a class="title" href="https://example.com/a">A</a>
+    <p class="s">snippet A</p>
+    <a class="title" href="https://example.com/b">B</a>
+    <p class="s">snippet B</p>`;
+
+  it("GETs Mojeek with a browser UA and query string", async () => {
+    const captured: { url: string; method: string; ua: string } = { url: "", method: "", ua: "" };
+    const originalFetch = globalThis.fetch;
+    globalThis.fetch = vi.fn(async (url: string | URL, init?: RequestInit) => {
+      captured.url = String(url);
+      captured.method = init?.method ?? "GET";
+      const headers = (init?.headers ?? {}) as Record<string, string>;
+      captured.ua = headers["User-Agent"] ?? "";
+      return new Response(twoResultsHtml, {
+        status: 200,
+        headers: { "Content-Type": "text/html" },
+      });
+    }) as unknown as typeof fetch;
+    try {
+      const out = await webSearch("flutter 3.19", { topK: 2 });
+      expect(captured.url).toContain("mojeek.com/search");
+      expect(captured.url).toContain("q=flutter%203.19");
+      expect(captured.method).toBe("GET");
+      expect(captured.ua).toMatch(/Mozilla\/5.0/);
+      expect(out).toHaveLength(2);
+      expect(out[0]?.title).toBe("A");
+    } finally {
+      globalThis.fetch = originalFetch;
+    }
+  });
+
+  it("clamps topK to [1, 10]", async () => {
+    const originalFetch = globalThis.fetch;
+    globalThis.fetch = vi.fn(
+      async () => new Response(twoResultsHtml, { status: 200 }),
+    ) as unknown as typeof fetch;
+    try {
+      const outMax = await webSearch("x", { topK: 99 });
+      expect(outMax.length).toBeLessThanOrEqual(10);
+      const outMin = await webSearch("x", { topK: 0 });
+      expect(outMin.length).toBe(1);
+    } finally {
+      globalThis.fetch = originalFetch;
+    }
+  });
+
+  it("throws on non-2xx", async () => {
+    const originalFetch = globalThis.fetch;
+    globalThis.fetch = vi.fn(
+      async () => new Response("blocked", { status: 429 }),
+    ) as unknown as typeof fetch;
+    try {
+      await expect(webSearch("q")).rejects.toThrow(/web_search 429/);
+    } finally {
+      globalThis.fetch = originalFetch;
+    }
+  });
+
+  it("returns [] on a legitimately empty 'No results' page", async () => {
+    const originalFetch = globalThis.fetch;
+    globalThis.fetch = vi.fn(
+      async () =>
+        new Response("<html><body>Your search did not match any documents.</body></html>", {
+          status: 200,
+        }),
+    ) as unknown as typeof fetch;
+    try {
+      const out = await webSearch("zzyzx nothing here");
+      expect(out).toEqual([]);
+    } finally {
+      globalThis.fetch = originalFetch;
+    }
+  });
+
+  it("surfaces a clear error on an anti-bot interstitial", async () => {
+    const originalFetch = globalThis.fetch;
+    globalThis.fetch = vi.fn(
+      async () =>
+        new Response("<html>Please solve the captcha to continue.</html>", { status: 200 }),
+    ) as unknown as typeof fetch;
+    try {
+      await expect(webSearch("q")).rejects.toThrow(/anti-bot|rate-limited|blocked/);
+    } finally {
+      globalThis.fetch = originalFetch;
+    }
+  });
+
+  it("surfaces a diagnostic error when the response looks like neither empty-nor-results", async () => {
+    const originalFetch = globalThis.fetch;
+    globalThis.fetch = vi.fn(
+      async () =>
+        new Response("<html><body>totally unexpected shape</body></html>", { status: 200 }),
+    ) as unknown as typeof fetch;
+    try {
+      await expect(webSearch("q")).rejects.toThrow(/doesn't look like a real empty page/);
+    } finally {
+      globalThis.fetch = originalFetch;
+    }
+  });
+});
+
+describe("formatSearchResults", () => {
+  it("renders a query header + numbered list", () => {
+    const out = formatSearchResults("hello", [
+      { title: "One", url: "https://one", snippet: "first" },
+      { title: "Two", url: "https://two", snippet: "second" },
+    ]);
+    expect(out).toMatch(/query: hello/);
+    expect(out).toMatch(/results \(2\)/);
+    expect(out).toMatch(/1\. One\n\s+https:\/\/one\n\s+first/);
+    expect(out).toMatch(/2\. Two/);
+  });
+});
+
+describe("registerWebTools", () => {
+  it("registers web_search and web_fetch", () => {
+    const registry = new ToolRegistry();
+    registerWebTools(registry);
+    expect(registry.size).toBe(2);
+  });
+
+  it("web_fetch refuses non-http(s) urls", async () => {
+    const registry = new ToolRegistry();
+    registerWebTools(registry);
+    const out = await registry.dispatch("web_fetch", JSON.stringify({ url: "file:///etc/passwd" }));
+    expect(out).toMatch(/must start with http/);
+  });
+});
+
+describe("webFetch", () => {
+  it("extracts title + body text from an html response", async () => {
+    const originalFetch = globalThis.fetch;
+    globalThis.fetch = vi.fn(
+      async () =>
+        new Response(
+          "<html><head><title>Demo</title></head><body><p>Hello world.</p></body></html>",
+          { status: 200, headers: { "Content-Type": "text/html; charset=utf-8" } },
+        ),
+    ) as unknown as typeof fetch;
+    try {
+      const page = await webFetch("https://example.com/demo");
+      expect(page.title).toBe("Demo");
+      expect(page.text).toContain("Hello world.");
+      expect(page.text).not.toContain("<title>");
+      expect(page.truncated).toBe(false);
+    } finally {
+      globalThis.fetch = originalFetch;
+    }
+  });
+
+  it("truncates long pages and flags the result", async () => {
+    const originalFetch = globalThis.fetch;
+    const big = `<html><body><p>${"a".repeat(50_000)}</p></body></html>`;
+    globalThis.fetch = vi.fn(
+      async () => new Response(big, { status: 200, headers: { "Content-Type": "text/html" } }),
+    ) as unknown as typeof fetch;
+    try {
+      const page = await webFetch("https://example.com/big", { maxChars: 1000 });
+      expect(page.truncated).toBe(true);
+      expect(page.text).toMatch(/\[… truncated \d+ chars …\]$/);
+    } finally {
+      globalThis.fetch = originalFetch;
+    }
+  });
+
+  it("surfaces non-2xx as a thrown error", async () => {
+    const originalFetch = globalThis.fetch;
+    globalThis.fetch = vi.fn(
+      async () => new Response("nope", { status: 404 }),
+    ) as unknown as typeof fetch;
+    try {
+      await expect(webFetch("https://example.com/missing")).rejects.toThrow(/web_fetch 404/);
+    } finally {
+      globalThis.fetch = originalFetch;
+    }
+  });
+});