ucode-agent 1.61.1 → 1.62.1

package/src/tools/files.js

@@ -1,865 +1,865 @@
-/**
- * files.js — reading and changing files.
- *
- * The rule that matters most in here: an edit never guesses. Zero matches or
- * two matches is an error with an explanation, never a silent partial change.
- * A wrong edit that reports success is the single most expensive thing a
- * coding agent can do, because everything after it is built on a lie.
- */
-
-import { promises as fs, existsSync } from 'node:fs';
-import { remember } from '../core/undo.js';
-import path from 'node:path';
-import { ToolFailure } from '../core/failure.js';
-import {
-  resolveIn, guard, result, fsFailure, looksBinary, toLines, bytes,
-  changedRegion, renderDiff, renderNewFile, READ_LINES, MAX_FILE_OUTPUT,
-  noteFile, writeTracked, assertUnchanged, getRoot,
-} from './shared.js';
-import { packageJsonWritten } from './shell.js';
-import { fuzzyReplace } from './fuzzy.js';
-import { parse as parseSource } from '@babel/parser';
-
-/**
- * Up to three names beside a missing file that look like what was meant —
- * "App.jsx" for "app.jsx", "index.html" for "index.htm". Named in the refusal
- * (as opencode's read tool does), the next step is the right read instead of
- * a list_dir to find out.
- */
-async function lookalikes(target) {
-  const dir = path.dirname(target.abs);
-  const base = path.basename(target.abs).toLowerCase();
-  const shown = path.dirname(target.show);
-  try {
-    return (await fs.readdir(dir))
-      .filter((n) => n.toLowerCase().includes(base) || (n.length > 2 && base.includes(n.toLowerCase())))
-      .slice(0, 3)
-      .map((n) => (shown === '.' ? n : `${shown}/${n}`));
-  } catch {
-    return [];
-  }
-}
-
-export async function readFile({ path: p, offset = 1, limit = READ_LINES }) {
-  const target = resolveIn(p, 'read_file');
-  await guard(target, `read ${target.abs}`);
-  const attempted = `reading ${target.show}`;
-
-  let stat;
-  try {
-    stat = await fs.stat(target.abs);
-  } catch (err) {
-    const failure = fsFailure(err, attempted, target.show);
-    // Outside the project the user said yes to one path, not to a listing of
-    // the folder around it — so no lookalikes there.
-    const near = err?.code === 'ENOENT' && target.inside ? await lookalikes(target) : [];
-    if (near.length) failure.fix = `Did you mean ${near.join(', ')}? ${failure.fix}`;
-    throw failure;
-  }
-
-  if (stat.isDirectory()) {
-    throw new ToolFailure({
-      kind: 'is_directory',
-      attempted,
-      failed: `${target.show} is a directory.`,
-      fix: `Use list_dir with path "${target.show}" to see what is in it.`,
-    });
-  }
-
-  let buf;
-  try {
-    buf = await fs.readFile(target.abs);
-  } catch (err) {
-    throw fsFailure(err, attempted, target.show);
-  }
-
-  if (looksBinary(buf.subarray(0, 8192))) {
-    throw new ToolFailure({
-      kind: 'binary',
-      attempted,
-      failed: `${target.show} is a binary file (${bytes(stat.size)}).`,
-      fix: 'ucode reads text only. Inspect it with run_command and a tool built for the format.',
-    });
-  }
-
-  // Read at this moment, so a later whole-file overwrite can tell its own
-  // change apart from somebody else's.
-  await noteFile(target.abs);
-
-  const lines = toLines(buf.toString('utf8'));
-  const from = Math.max(1, Math.floor(Number(offset) || 1));
-  const count = Math.max(1, Math.floor(Number(limit) || READ_LINES));
-  const slice = lines.slice(from - 1, from - 1 + count);
-
-  if (slice.length === 0) {
-    throw new ToolFailure({
-      kind: 'bad_args',
-      attempted,
-      failed: `offset ${from} is past the end of the file, which has ${lines.length} lines.`,
-      fix: `Read again with an offset between 1 and ${lines.length}.`,
-    });
-  }
-
-  const last = from + slice.length - 1;
-  const width = String(last).length;
-  // The numbers are a gutter for the model to reason about, and they are
-  // stated to be display-only in the tool description, because an edit whose
-  // old_string still carries them will never match.
-  const body = slice.map((line, i) => `${String(from + i).padStart(width)} | ${line}`).join('\n');
-
-  const more = last < lines.length
-    ? `\n\n[lines ${from}-${last} of ${lines.length}. Continue with offset=${last + 1}.]`
-    : '';
-
-  return result(
-    body + more,
-    more || from > 1 ? `lines ${from}-${last} of ${lines.length}` : `${lines.length} lines`,
-    MAX_FILE_OUTPUT
-  );
-}
-
-/**
- * Everything that writes a file ends here. A package.json with dependencies
- * starts its install in the background at once, while the rest of the app is
- * still being written.
- */
-function written(target, content) {
-  if (path.basename(target.abs) === 'package.json') packageJsonWritten(target.abs, content);
-}
-
-const PARSEABLE = /\.(?:[cm]?[jt]sx?)$/i;
-
-/**
- * Does this source parse? Checked the moment a file is written.
- *
- * Measured on a real build: a JSX typo sat unnoticed until `npm run build`,
- * which takes up to a minute, failed on it — then the fix, then another full
- * build. Parsing the file takes milliseconds, so the error comes back in the
- * same step that wrote it, with the line, while the model still has the file
- * in front of it. Syntax only: types are checked at the end of the turn.
- */
-export function syntaxProblem(file, text) {
-  if (!PARSEABLE.test(file)) return null;
-  const ext = path.extname(file).toLowerCase();
-  const plugins = ext === '.tsx'
-    ? ['typescript', 'jsx']
-    : /^\.[cm]?ts$/.test(ext) ? ['typescript'] : ['jsx'];
-  const where = (loc) => (loc ? `line ${loc.line}, column ${loc.column + 1}` : 'somewhere in the file');
-  const clean = (m) => String(m).replace(/\s*\(\d+:\d+\)\s*$/, '');
-  try {
-    const ast = parseSource(text, {
-      sourceType: 'unambiguous',
-      plugins: [...plugins, 'decorators-legacy'],
-      errorRecovery: true,
-      allowReturnOutsideFunction: true,
-      allowAwaitOutsideFunction: true,
-    });
-    const first = ast.errors?.[0];
-    return first ? `${where(first.loc)}: ${clean(first.message)}` : null;
-  } catch (err) {
-    return `${where(err.loc)}: ${clean(err.message)}`;
-  }
-}
-
-/**
- * How many times in a row each file has come back unparseable.
- *
- * A model that has broken a file once usually fixes it. A model that has
- * broken it three times is guessing at a structure it has lost track of, and
- * will keep guessing: one run spent fifteen minutes on a single line before
- * giving up and rewriting the file, which is what it should have been told to
- * do after the second try.
- */
-const brokenRuns = new Map();
-
-export function forgetBrokenRuns() { brokenRuns.clear(); }
-
-/**
- * The note appended to a result: empty when the file parses, and the run of
- * failures forgotten, so three good writes later a single slip is a slip again.
- */
-const parseNote = (show, problem) => {
-  if (!problem) { brokenRuns.delete(show); return ''; }
-  return brokenNote(show, problem);
-};
-
-/** The warning appended to a result when a file does not parse. */
-const brokenNote = (show, problem) => {
-  const runs = (brokenRuns.get(show) ?? 0) + 1;
-  brokenRuns.set(show, runs);
-  const base = `
-
-⚠ ${show} does not parse — ${problem}.`;
-  if (runs < 3) return `${base} Fix it now: the build will fail on it.`;
-  return `${base} That is ${runs} attempts in a row on this file. Stop editing it: ` +
-    `write the whole file again with write_file, in one piece, rather than patching ` +
-    `a structure you have lost track of.`;
-};
-
-/**
- * A module in a page with no build step that imports a stylesheet or an image.
- *
- * A bundler would take it; a browser refuses the whole module ("Failed to
- * load module script ... MIME type text/css"), so none of the app's script
- * runs and every button is dead. The syntax is fine, so nothing else notices
- * until the page is opened — a live build shipped exactly this as done.
- */
-const ASSET_IMPORT = /^[ \t]*import\s+(?:[\w$*{}\s,]+?\s+from\s+)?['"]([^'"]+\.(?:css|scss|sass|less|svg|png|jpe?g|gif|webp))['"](?!\s*(?:with|assert)\s*\{)/m;
-
-export function assetImport(abs, text) {
-  if (!/\.m?js$/i.test(abs)) return null;
-  // An import quoted inside a block comment is not an import.
-  const hit = ASSET_IMPORT.exec(String(text).replace(/\/\*[\s\S]*?\*\//g, ''));
-  if (!hit) return null;
-  // Anything with a package.json above it, up to the project root, may well be
-  // bundled. path.relative, not a string prefix: "ucode2" starts with "ucode",
-  // and Windows paths differ in case from one shell to the next.
-  const top = getRoot();
-  const within = (dir) => {
-    const rel = path.relative(top, dir);
-    return rel === '' || (!rel.startsWith('..') && !path.isAbsolute(rel));
-  };
-  for (let dir = path.dirname(abs); ; dir = path.dirname(dir)) {
-    if (existsSync(path.join(dir, 'package.json'))) return null;
-    if (!within(dir) || path.relative(top, dir) === '' || path.dirname(dir) === dir) break;
-  }
-  return `imports ${hit[1]}, which a page with no build step cannot do: the browser refuses the ` +
-    'whole module, so none of its script runs. Remove that import and load it from the page ' +
-    'instead — <link rel="stylesheet" href="…"> for a stylesheet, an <img> or a plain URL for an image.';
-}
-
-/**
- * A page styled with Tailwind classes that never loads Tailwind.
- *
- * A live build wrote a 27 KB page of "flex px-4 bg-[var(--bg)]" and a
- * stylesheet of colour tokens only: every class was ignored and the app came
- * out as bare browser defaults, with nothing failing to tell anyone. When a
- * page leans on those classes and has no Tailwind of its own, the browser
- * build is added to its head, so the page looks the way it was written.
- */
-const TAILWIND_CLASS = /^(?:-?(?:m|p)[trblxy]?-\S+|flex|grid|hidden|block|inline-flex|items-\S+|justify-\S+|gap-\S+|space-[xy]-\S+|(?:min-|max-)?[wh]-\S+|text-\S+|bg-\S+|rounded(?:-\S+)?|border(?:-\S+)?|shadow(?:-\S+)?|font-\S+|leading-\S+|tracking-\S+|(?:sm|md|lg|xl|hover|focus|dark):\S+)$/;
-export function withTailwind(html) {
-  if (/tailwindcss|cdn\.tailwind/i.test(html) || !/<\/head>/i.test(html)) return html;
-  let count = 0;
-  for (const [, list] of html.matchAll(/class="([^"]*)"/g)) {
-    for (const token of list.split(/\s+/)) if (TAILWIND_CLASS.test(token)) count++;
-    if (count >= 12) break;
-  }
-  if (count < 12) return html;
-  return html.replace(/<\/head>/i, '  <script src="https://cdn.jsdelivr.net/npm/@tailwindcss/browser@4"></script>\n</head>');
-}
-
-/** The note for that, appended to a write's result like a parse problem is. */
-const pageNote = (target, text) => {
-  const problem = assetImport(target.abs, text);
-  return problem ? `\n\n⚠ ${target.show} ${problem}` : '';
-};
-
-/** A file this short comes back whole after an edit; longer ones show the part around the change. */
-const SHOW_WHOLE = 250;
-const AROUND = 15;
-
-/**
- * The file as it stands after an edit, with the same line-number gutter as
- * read_file.
- *
- * Measured on a real build: the model read the same component thirteen times
- * in one turn, because an edit's result showed only the lines it replaced and
- * the next edit needed the file as it now was. Sending the current text back
- * with the edit costs the same tokens the re-read would have, and saves the
- * round trip every time.
- */
-function nowReads(show, text, at, span) {
-  const lines = toLines(text);
-  const whole = lines.length <= SHOW_WHOLE;
-  const from = whole ? 1 : Math.max(1, at - AROUND);
-  const to = whole ? lines.length : Math.min(lines.length, at + span + AROUND);
-  const width = String(to).length;
-  const body = lines.slice(from - 1, to).map((l, i) => `${String(from + i).padStart(width)} | ${l}`).join('\n');
-  const heading = whole
-    ? `${show} now reads (all ${lines.length} lines`
-    : `${show} now reads, lines ${from}-${to} of ${lines.length}`;
-  return `\n\n${heading} — this is the current text, so there is no need to read it again):\n${body}`;
-}
-
-/** At most this many files in one read_files call. */
-const MAX_BATCH = 20;
-
-/**
- * Several files in one call.
- *
- * Reading from disk takes a millisecond. What makes reading slow is the round
- * trip around it: every file read on its own is a whole request to the model,
- * and on a reasoning model that is several seconds of thinking before it even
- * asks for the next one. Reading the six files a change touches in one call
- * turns six of those into one.
- *
- * The files are read in parallel, and a missing one is reported in its place
- * rather than failing the rest — one wrong path should not cost the other five.
- */
-export async function readFiles({ paths, limit = READ_LINES }) {
-  if (!Array.isArray(paths) || paths.length === 0) {
-    throw new ToolFailure({
-      kind: 'bad_args',
-      attempted: 'reading several files',
-      failed: 'The "paths" argument must be a non-empty array of file paths.',
-      fix: 'Pass paths as ["src/app.js", "src/lib/api.js", ...].',
-    });
-  }
-
-  const wanted = [...new Set(paths.map((p) => String(p ?? '').trim()).filter(Boolean))];
-  const batch = wanted.slice(0, MAX_BATCH);
-  const dropped = wanted.slice(MAX_BATCH);
-
-  const readOne = (p) => readFile({ path: p, limit }).then(
-    (out) => ({ p, out }),
-    (err) => ({ p, err })
-  );
-
-  // Anything outside the project needs a yes, and two questions cannot be
-  // asked at once — so those go one at a time. Everything else goes together.
-  const outside = batch.some((p) => !resolveIn(p, 'read_files', 'paths').inside);
-  const settled = [];
-  if (outside) {
-    for (const p of batch) settled.push(await readOne(p));
-  } else {
-    settled.push(...(await Promise.all(batch.map(readOne))));
-  }
-
-  // Two whole files' worth of output between them. Past that, the rest are
-  // named rather than silently cut, so the model knows to ask again.
-  const budget = MAX_FILE_OUTPUT * 2;
-  let used = 0;
-  let read = 0;
-  let failed = 0;
-  let lines = 0;
-  const blocks = [];
-  const deferred = [];
-
-  for (const { p, out, err } of settled) {
-    if (err) {
-      failed++;
-      blocks.push(`=== ${p} — could not be read ===\n${err.forModel ? err.forModel() : err.message}`);
-      continue;
-    }
-    const block = `=== ${p} (${out.summary}) ===\n${out.content}`;
-    if (read > 0 && used + block.length > budget) {
-      deferred.push(p);
-      continue;
-    }
-    blocks.push(block);
-    used += block.length;
-    read++;
-    // "42 lines" for a whole file, "lines 1-600 of 900" for a page of one.
-    const whole = /^(\d+) lines$/.exec(out.summary);
-    const page = /^lines (\d+)-(\d+)/.exec(out.summary);
-    lines += whole ? Number(whole[1]) : page ? Number(page[2]) - Number(page[1]) + 1 : 0;
-  }
-
-  const leftOver = [...deferred, ...dropped];
-  if (leftOver.length) {
-    blocks.push(
-      `[not included, to stay inside one reply: ${leftOver.join(', ')}. ` +
-      'Read those with another read_files call.]'
-    );
-  }
-
-  return result(
-    blocks.join('\n\n'),
-    `${read} file${read === 1 ? '' : 's'} · ${lines} lines` +
-      (failed ? ` · ${failed} missing` : '') +
-      (leftOver.length ? ` · ${leftOver.length} deferred` : ''),
-    budget + 2_000
-  );
-}
-
-/** Write one file, returning the rows that show what changed. */
-async function put(target, content, { diffMax = 16 } = {}) {
-  const attempted = `writing ${target.show}`;
-  if (/\.html?$/i.test(target.abs)) content = withTailwind(content);
-
-  // Read what is there before clobbering it, so an overwrite can be shown as
-  // an actual diff rather than as a claim that something changed.
-  let previous = null;
-  try {
-    previous = await fs.readFile(target.abs, 'utf8');
-  } catch {
-    previous = null; // missing, or binary — either way it is treated as new
-  }
-
-  if (previous !== null) await assertUnchanged(target.abs, target.show);
-
-  try {
-    await fs.mkdir(path.dirname(target.abs), { recursive: true });
-    await remember(target.abs);
-    await writeTracked(target.abs, content);
-  } catch (err) {
-    throw fsFailure(err, attempted, target.show);
-  }
-  written(target, content);
-
-  const existed = previous !== null;
-  const lineCount = content === '' ? 0 : toLines(content).length;
-  const diff = existed
-    ? renderDiff(changedRegion(previous, content), { max: diffMax })
-    : (content === '' ? [] : renderNewFile(content, diffMax));
-
-  return {
-    existed,
-    lineCount,
-    diff,
-    problem: syntaxProblem(target.abs, content),
-    note: pageNote(target, content),
-    line: `${existed ? 'Overwrote' : 'Created'} ${target.show} ` +
-      `(${lineCount} lines, ${bytes(Buffer.byteLength(content))})`,
-  };
-}
-
-export async function writeFile({ path: p, content }) {
-  const target = resolveIn(p, 'write_file');
-  if (typeof content !== 'string') {
-    throw new ToolFailure({
-      kind: 'bad_args',
-      attempted: `writing ${target.show}`,
-      failed: 'The "content" argument was missing or was not a string.',
-      fix: 'Call write_file again with content set to the whole text of the file.',
-    });
-  }
-  await guard(target, `write ${target.abs}`);
-
-  const written = await put(target, content);
-  const out = result(
-    `${written.line}.${parseNote(target.show, written.problem)}${written.note}`,
-    `${written.existed ? 'overwrote' : 'created'} · ${written.lineCount} lines${written.problem ? ' · does not parse' : ''}`
-  );
-  out.diff = written.diff;
-  return out;
-}
-
-/**
- * Several files in one call.
- *
- * Scaffolding a project is twenty writes before anything can be run, and doing
- * that one round trip at a time is most of the wait.
- */
-export async function batchWrite({ files }) {
-  if (!Array.isArray(files) || files.length === 0) {
-    throw new ToolFailure({
-      kind: 'bad_args',
-      attempted: 'writing several files',
-      failed: 'The "files" argument must be a non-empty array.',
-      fix: 'Pass files as [{ path, content }, ...].',
-    });
-  }
-
-  const lines = [];
-  const diff = [];
-  let created = 0;
-  let broken = 0;
-
-  for (const [index, file] of files.entries()) {
-    const { path: p, content } = file ?? {};
-    if (typeof p !== 'string' || typeof content !== 'string') {
-      throw new ToolFailure({
-        kind: 'bad_args',
-        attempted: `writing file ${index + 1} of ${files.length}`,
-        failed: 'Every entry needs "path" and "content", both strings.',
-        fix: `Fix entry ${index + 1} and call batch_write again. ${index} file(s) were already written.`,
-      });
-    }
-
-    const target = resolveIn(p, 'batch_write');
-    await guard(target, `write ${target.abs}`);
-
-    // Per-file diffs are kept short here; twenty files at sixteen rows each
-    // would bury the reply under three hundred lines of gutter.
-    const written = await put(target, content, { diffMax: 6 });
-    if (!written.existed) created++;
-    lines.push(written.line + parseNote(target.show, written.problem) + written.note);
-    if (written.problem) broken++;
-    diff.push(`~${target.show}`, ...written.diff);
-  }
-
-  const out = result(
-    lines.join('\n'),
-    `${files.length} file${files.length === 1 ? '' : 's'} · ${created} new${broken ? ` · ${broken} do not parse` : ''}`
-  );
-  out.diff = diff;
-  return out;
-}
-
-/**
- * Why an old_string missed, worked out rather than guessed at.
- *
- * "not found" tells the model nothing it did not already know. Whether the
- * text is present with different whitespace, or present but only its first
- * line, is the difference between a fix on the next step and three more
- * failed attempts.
- */
-function explainMiss(original, oldString, show) {
-  const flatten = (s) => s.replace(/\s+/g, ' ').trim();
-  const firstLine = oldString.split('\n')[0].trim();
-
-  if (flatten(original).includes(flatten(oldString))) {
-    return {
-      failed: `old_string is not in ${show} as written — the text is there, but the whitespace differs.`,
-      fix: "Match the file's own indentation exactly: tabs versus spaces, and the line breaks.",
-    };
-  }
-
-  const nearby = firstLine.length > 3
-    ? original.split(/\r?\n/)
-        .map((line, i) => [i + 1, line])
-        .filter(([, line]) => line.includes(firstLine))
-        .slice(0, 3)
-    : [];
-
-  if (nearby.length) {
-    return {
-      failed:
-        `old_string is not in ${show}. Its first line does appear at ` +
-        `line${nearby.length > 1 ? 's' : ''} ${nearby.map(([n]) => n).join(', ')}, ` +
-        'so it is the lines after it that differ.',
-      fix: `Read ${show} around line ${nearby[0][0]} and copy the block exactly as it is.`,
-    };
-  }
-
-  return {
-    failed: `old_string does not appear anywhere in ${show}.`,
-    fix: `Read ${show} again and copy the text verbatim, without the line-number gutter.`,
-  };
-}
-
-/** Apply one replacement to a string, or explain precisely why it cannot. */
-function replaceOnce(text, { old_string, new_string, replace_all }, { show, attempted, label = '' }) {
-  const prefix = label ? `${label}: ` : '';
-
-  if (typeof old_string !== 'string' || typeof new_string !== 'string') {
-    throw new ToolFailure({
-      kind: 'bad_args', attempted,
-      failed: `${prefix}old_string and new_string must both be strings.`,
-      fix: 'Fix that entry and call again. Nothing was written.',
-    });
-  }
-  if (old_string === '') {
-    throw new ToolFailure({
-      kind: 'bad_args', attempted,
-      failed: `${prefix}old_string was empty.`,
-      fix: 'edit_file replaces existing text. Use write_file to create a file.',
-    });
-  }
-  // An edit whose two halves are the same asks for the file to stay as it is,
-  // which it will. Refusing that was a hard failure, and the model answered it
-  // by sending the same edit again — the stuck detector carries a special case
-  // for exactly this loop. Saying "already done" ends it in one step.
-  if (old_string === new_string) {
-    const found = text.indexOf(old_string);
-    return { text, at: found < 0 ? 1 : toLines(text.slice(0, found)).length, how: '', count: 0 };
-  }
-
-  // Models write \n. A file checked out on Windows is often \r\n, and then an
-  // otherwise perfect old_string can never match. Speak the file's dialect.
-  let oldText = old_string;
-  let newText = new_string;
-  if (text.includes('\r\n') && !oldText.includes('\r')) {
-    oldText = oldText.replace(/\r?\n/g, '\r\n');
-    newText = newText.replace(/\r?\n/g, '\r\n');
-  }
-
-  const ambiguous = (hits, how = '') => new ToolFailure({
-    kind: 'ambiguous', attempted,
-    failed: `${prefix}old_string appears ${hits} times in ${show}${how}. Refusing to guess which one you meant.`,
-    fix: 'Add surrounding lines to old_string until it matches exactly one place.',
-    detail: { hits },
-  });
-
-  // replace_all is for renames: every copy changes, and many copies is the point.
-  const all = replace_all === true || replace_all === 'true';
-  const hits = text.split(oldText).length - 1;
-  if (hits > 1 && !all) throw ambiguous(hits);
-  if (hits >= 1) {
-    const at = text.slice(0, text.indexOf(oldText)).split(/\r?\n/).length;
-    const out = all ? text.split(oldText).join(newText) : text.replace(oldText, () => newText);
-    return { text: out, at, how: '', count: hits };
-  }
-
-  // No exact match. The commonest reason by far is whitespace — tabs against
-  // spaces, a different indent depth, trailing spaces — with every word right.
-  // Match line by line ignoring that, and re-indent the replacement to fit.
-  // Still unique or nothing: a loose match found twice is refused like any other.
-  const loose = all ? null : looseReplace(text, old_string, new_string);
-  if (loose?.count === 1) {
-    return { text: loose.text, at: loose.at, how: 'ignoring whitespace and re-indented to fit', count: 1 };
-  }
-  if (loose?.count > 1) throw ambiguous(loose.count, ' once whitespace is ignored');
-
-  // Still nothing. The remaining slips — a middle line remembered slightly
-  // wrong, escapes written out, a blank line at either end — each have a
-  // matcher of their own (fuzzy.js). They are given the edit in the file's
-  // own line endings, and only the matched span changes: the rest of a file
-  // with mixed endings keeps every one it had.
-  const fuzzy = fuzzyReplace(text, oldText, newText, { all });
-  if (fuzzy?.ambiguous) throw ambiguous('several', ' once matched loosely');
-  if (fuzzy?.wide) {
-    throw new ToolFailure({
-      kind: 'no_match', attempted,
-      failed: `${prefix}old_string only matches ${show} loosely, across far more text than it contains. Refusing to replace that much.`,
-      fix: `Read ${show} again and copy the exact text you mean to replace.`,
-    });
-  }
-  if (fuzzy) {
-    const at = text.slice(0, fuzzy.index).split('\n').length;
-    return { text: fuzzy.text, at, how: fuzzy.how, count: fuzzy.count };
-  }
-
-  const { failed, fix } = explainMiss(text, old_string, show);
-  throw new ToolFailure({ kind: 'no_match', attempted, failed: prefix + failed, fix });
-}
-
-/**
- * Find old_string by its lines' content alone and swap in new_string, indented
- * the way the file is indented at that spot. Returns { count } when it finds
- * none or several, and { count: 1, text, at } when it finds exactly one.
- */
-function looseReplace(text, oldString, newString) {
-  const eol = text.includes('\r\n') ? '\r\n' : '\n';
-  const lines = text.split(/\r?\n/);
-
-  const want = oldString.replace(/\r/g, '').split('\n');
-  while (want.length > 1 && !want[want.length - 1].trim()) want.pop();
-  while (want.length > 1 && !want[0].trim()) want.shift();
-  const target = want.map((l) => l.trim());
-  if (target.every((t) => !t)) return null;
-
-  const starts = [];
-  for (let i = 0; i + want.length <= lines.length; i++) {
-    let same = true;
-    for (let j = 0; j < want.length; j++) {
-      if (lines[i + j].trim() !== target[j]) { same = false; break; }
-    }
-    if (same) starts.push(i);
-  }
-  if (starts.length !== 1) return { count: starts.length };
-
-  const start = starts[0];
-  const indent = (l) => /^[ \t]*/.exec(l)[0];
-  const first = want.findIndex((l) => l.trim());
-  const fileIndent = indent(lines[start + first]);
-  const wroteIndent = indent(want[first]);
-
-  const replacement = newString.replace(/\r/g, '').split('\n');
-  if (replacement.length > 1 && replacement[replacement.length - 1] === '') replacement.pop();
-  const reindented = replacement.map((l) => {
-    if (!l.trim()) return l.trim();
-    const body = l.startsWith(wroteIndent) ? l.slice(wroteIndent.length) : l.replace(/^[ \t]*/, '');
-    return fileIndent + body;
-  });
-
-  const out = [...lines.slice(0, start), ...reindented, ...lines.slice(start + want.length)];
-  return { count: 1, text: out.join(eol), at: start + 1 };
-}
-
-export async function editFile({ path: p, old_string, new_string, replace_all }) {
-  const target = resolveIn(p, 'edit_file');
-  const attempted = `editing ${target.show}`;
-  await guard(target, `edit ${target.abs}`);
-
-  let original;
-  try {
-    original = await fs.readFile(target.abs, 'utf8');
-  } catch (err) {
-    throw fsFailure(err, attempted, target.show);
-  }
-
-  const { text, at, how, count } = replaceOnce(original, { old_string, new_string, replace_all }, {
-    show: target.show, attempted,
-  });
-
-  try {
-    await remember(target.abs);
-    await writeTracked(target.abs, text);
-  } catch (err) {
-    throw fsFailure(err, attempted, target.show);
-  }
-  written(target, text);
-
-  const delta = toLines(text).length - toLines(original).length;
-  const change = delta === 0 ? 'same line count' : `${delta > 0 ? '+' : ''}${delta} lines`;
-  const where = count > 1
-    ? `${count} occurrences in ${target.show}, the first at line ${at}`
-    : `one occurrence in ${target.show} at line ${at}`;
-  const matched = how ? `, matched ${how}` : '';
-
-  const span = toLines(new_string).length;
-  const out = result(
-    `Replaced ${where} (${change}${matched}).` +
-      parseNote(target.show, syntaxProblem(target.abs, text)) + pageNote(target, text) +
-      nowReads(target.show, text, at, span),
-    `${count > 1 ? `${count} changes from line` : '1 change at line'} ${at} · ${change}` +
-      `${how ? ' · whitespace-tolerant' : ''}${syntaxProblem(target.abs, text) ? ' · does not parse' : ''}`,
-    MAX_FILE_OUTPUT
-  );
-  // The replacement is diffed on its own and offset to where it landed, so
-  // the gutter shows the file's line numbers rather than 1, 2, 3.
-  out.diff = renderDiff(changedRegion(old_string, new_string), { offset: at - 1 });
-  return out;
-}
-
-/**
- * Several replacements in one file, applied in order, each seeing the result
- * of the one before it.
- *
- * Everything is validated against a working copy first. If the third edit is
- * ambiguous, none of the three are written — a half-applied set of edits is a
- * file in a state nobody designed.
- */
-export async function multiEdit({ path: p, edits }) {
-  const target = resolveIn(p, 'multi_edit');
-  const attempted = `editing ${target.show}`;
-
-  if (!Array.isArray(edits) || edits.length === 0) {
-    throw new ToolFailure({
-      kind: 'bad_args',
-      attempted,
-      failed: 'The "edits" argument must be a non-empty array.',
-      fix: 'Pass edits as [{ old_string, new_string }, ...].',
-    });
-  }
-
-  await guard(target, `edit ${target.abs}`);
-
-  let original;
-  try {
-    original = await fs.readFile(target.abs, 'utf8');
-  } catch (err) {
-    throw fsFailure(err, attempted, target.show);
-  }
-
-  let text = original;
-  const diff = [];
-
-  for (const [index, edit] of edits.entries()) {
-    const applied = replaceOnce(text, edit ?? {}, {
-      show: target.show,
-      attempted,
-      label: `edit ${index + 1} of ${edits.length}`,
-    });
-    diff.push(
-      ...renderDiff(changedRegion(edit.old_string, edit.new_string), {
-        offset: applied.at - 1,
-        max: 8,
-      })
-    );
-    text = applied.text;
-  }
-
-  try {
-    await remember(target.abs);
-    await writeTracked(target.abs, text);
-  } catch (err) {
-    throw fsFailure(err, attempted, target.show);
-  }
-  written(target, text);
-
-  const delta = toLines(text).length - toLines(original).length;
-  const change = delta === 0 ? 'same line count' : `${delta > 0 ? '+' : ''}${delta} lines`;
-
-  const out = result(
-    `Applied ${edits.length} edits to ${target.show} (${change}).` +
-      parseNote(target.show, syntaxProblem(target.abs, text)) + pageNote(target, text) +
-      nowReads(target.show, text, 1, toLines(text).length),
-    `${edits.length} edits · ${change}`,
-    MAX_FILE_OUTPUT
-  );
-  out.diff = diff;
-  return out;
-}
-
-/**
- * Exact replacements across several files in one call.
- *
- * A change that touches the route, the component and the type together is one
- * round trip instead of three. Every edit in every file is applied to a copy
- * in memory first; if any of them fails, nothing is written anywhere — a
- * cross-file change that half landed leaves the project in a state that
- * compiles nowhere.
- */
-export async function editFiles({ files }) {
-  if (!Array.isArray(files) || files.length === 0) {
-    throw new ToolFailure({
-      kind: 'bad_args',
-      attempted: 'editing several files',
-      failed: 'The "files" argument must be a non-empty array.',
-      fix: 'Pass files as [{ path, edits: [{ old_string, new_string }, ...] }, ...].',
-    });
-  }
-
-  const planned = [];
-  const seen = new Set();
-
-  for (const [i, entry] of files.entries()) {
-    const target = resolveIn(entry?.path, 'edit_files');
-    const attempted = `editing ${target.show}`;
-
-    if (seen.has(target.abs)) {
-      throw new ToolFailure({
-        kind: 'bad_args', attempted,
-        failed: `${target.show} is listed twice.`,
-        fix: 'List each file once, with all of its edits together. Nothing was written.',
-      });
-    }
-    seen.add(target.abs);
-
-    if (!Array.isArray(entry.edits) || entry.edits.length === 0) {
-      throw new ToolFailure({
-        kind: 'bad_args', attempted,
-        failed: `File ${i + 1} (${target.show}) has no edits.`,
-        fix: 'Give every file a non-empty edits array. Nothing was written.',
-      });
-    }
-
-    await guard(target, `edit ${target.abs}`);
-
-    let original;
-    try {
-      original = await fs.readFile(target.abs, 'utf8');
-    } catch (err) {
-      throw fsFailure(err, attempted, target.show);
-    }
-
-    let text = original;
-    const diff = [`~${target.show}`];
-    for (const [j, edit] of entry.edits.entries()) {
-      const applied = replaceOnce(text, edit ?? {}, {
-        show: target.show,
-        attempted,
-        label: `${target.show}, edit ${j + 1} of ${entry.edits.length} (nothing was written)`,
-      });
-      diff.push(...renderDiff(changedRegion(edit.old_string, edit.new_string), { offset: applied.at - 1, max: 6 }));
-      text = applied.text;
-    }
-    planned.push({ target, text, diff, count: entry.edits.length });
-  }
-
-  for (const { target, text } of planned) {
-    try {
-      await remember(target.abs);
-    await writeTracked(target.abs, text);
-    } catch (err) {
-      throw fsFailure(err, `editing ${target.show}`, target.show);
-    }
-    written(target, text);
-  }
-
-  const edits = planned.reduce((n, p) => n + p.count, 0);
-  const out = result(
-    planned.map((p) => {
-      const problem = syntaxProblem(p.target.abs, p.text);
-      return `Edited ${p.target.show} (${p.count} change${p.count === 1 ? '' : 's'})` +
-        parseNote(p.target.show, problem) + pageNote(p.target, p.text);
-    }).join('\n'),
-    `${planned.length} files · ${edits} edits`
-  );
-  out.diff = planned.flatMap((p) => p.diff);
-  return out;
-}
+/**
+ * files.js — reading and changing files.
+ *
+ * The rule that matters most in here: an edit never guesses. Zero matches or
+ * two matches is an error with an explanation, never a silent partial change.
+ * A wrong edit that reports success is the single most expensive thing a
+ * coding agent can do, because everything after it is built on a lie.
+ */
+
+import { promises as fs, existsSync } from 'node:fs';
+import { remember } from '../core/undo.js';
+import path from 'node:path';
+import { ToolFailure } from '../core/failure.js';
+import {
+  resolveIn, guard, result, fsFailure, looksBinary, toLines, bytes,
+  changedRegion, renderDiff, renderNewFile, READ_LINES, MAX_FILE_OUTPUT,
+  noteFile, writeTracked, assertUnchanged, getRoot,
+} from './shared.js';
+import { packageJsonWritten } from './shell.js';
+import { fuzzyReplace } from './fuzzy.js';
+import { parse as parseSource } from '@babel/parser';
+
+/**
+ * Up to three names beside a missing file that look like what was meant —
+ * "App.jsx" for "app.jsx", "index.html" for "index.htm". Named in the refusal
+ * (as opencode's read tool does), the next step is the right read instead of
+ * a list_dir to find out.
+ */
+async function lookalikes(target) {
+  const dir = path.dirname(target.abs);
+  const base = path.basename(target.abs).toLowerCase();
+  const shown = path.dirname(target.show);
+  try {
+    return (await fs.readdir(dir))
+      .filter((n) => n.toLowerCase().includes(base) || (n.length > 2 && base.includes(n.toLowerCase())))
+      .slice(0, 3)
+      .map((n) => (shown === '.' ? n : `${shown}/${n}`));
+  } catch {
+    return [];
+  }
+}
+
+export async function readFile({ path: p, offset = 1, limit = READ_LINES }) {
+  const target = resolveIn(p, 'read_file');
+  await guard(target, `read ${target.abs}`);
+  const attempted = `reading ${target.show}`;
+
+  let stat;
+  try {
+    stat = await fs.stat(target.abs);
+  } catch (err) {
+    const failure = fsFailure(err, attempted, target.show);
+    // Outside the project the user said yes to one path, not to a listing of
+    // the folder around it — so no lookalikes there.
+    const near = err?.code === 'ENOENT' && target.inside ? await lookalikes(target) : [];
+    if (near.length) failure.fix = `Did you mean ${near.join(', ')}? ${failure.fix}`;
+    throw failure;
+  }
+
+  if (stat.isDirectory()) {
+    throw new ToolFailure({
+      kind: 'is_directory',
+      attempted,
+      failed: `${target.show} is a directory.`,
+      fix: `Use list_dir with path "${target.show}" to see what is in it.`,
+    });
+  }
+
+  let buf;
+  try {
+    buf = await fs.readFile(target.abs);
+  } catch (err) {
+    throw fsFailure(err, attempted, target.show);
+  }
+
+  if (looksBinary(buf.subarray(0, 8192))) {
+    throw new ToolFailure({
+      kind: 'binary',
+      attempted,
+      failed: `${target.show} is a binary file (${bytes(stat.size)}).`,
+      fix: 'ucode reads text only. Inspect it with run_command and a tool built for the format.',
+    });
+  }
+
+  // Read at this moment, so a later whole-file overwrite can tell its own
+  // change apart from somebody else's.
+  await noteFile(target.abs);
+
+  const lines = toLines(buf.toString('utf8'));
+  const from = Math.max(1, Math.floor(Number(offset) || 1));
+  const count = Math.max(1, Math.floor(Number(limit) || READ_LINES));
+  const slice = lines.slice(from - 1, from - 1 + count);
+
+  if (slice.length === 0) {
+    throw new ToolFailure({
+      kind: 'bad_args',
+      attempted,
+      failed: `offset ${from} is past the end of the file, which has ${lines.length} lines.`,
+      fix: `Read again with an offset between 1 and ${lines.length}.`,
+    });
+  }
+
+  const last = from + slice.length - 1;
+  const width = String(last).length;
+  // The numbers are a gutter for the model to reason about, and they are
+  // stated to be display-only in the tool description, because an edit whose
+  // old_string still carries them will never match.
+  const body = slice.map((line, i) => `${String(from + i).padStart(width)} | ${line}`).join('\n');
+
+  const more = last < lines.length
+    ? `\n\n[lines ${from}-${last} of ${lines.length}. Continue with offset=${last + 1}.]`
+    : '';
+
+  return result(
+    body + more,
+    more || from > 1 ? `lines ${from}-${last} of ${lines.length}` : `${lines.length} lines`,
+    MAX_FILE_OUTPUT
+  );
+}
+
+/**
+ * Everything that writes a file ends here. A package.json with dependencies
+ * starts its install in the background at once, while the rest of the app is
+ * still being written.
+ */
+function written(target, content) {
+  if (path.basename(target.abs) === 'package.json') packageJsonWritten(target.abs, content);
+}
+
+const PARSEABLE = /\.(?:[cm]?[jt]sx?)$/i;
+
+/**
+ * Does this source parse? Checked the moment a file is written.
+ *
+ * Measured on a real build: a JSX typo sat unnoticed until `npm run build`,
+ * which takes up to a minute, failed on it — then the fix, then another full
+ * build. Parsing the file takes milliseconds, so the error comes back in the
+ * same step that wrote it, with the line, while the model still has the file
+ * in front of it. Syntax only: types are checked at the end of the turn.
+ */
+export function syntaxProblem(file, text) {
+  if (!PARSEABLE.test(file)) return null;
+  const ext = path.extname(file).toLowerCase();
+  const plugins = ext === '.tsx'
+    ? ['typescript', 'jsx']
+    : /^\.[cm]?ts$/.test(ext) ? ['typescript'] : ['jsx'];
+  const where = (loc) => (loc ? `line ${loc.line}, column ${loc.column + 1}` : 'somewhere in the file');
+  const clean = (m) => String(m).replace(/\s*\(\d+:\d+\)\s*$/, '');
+  try {
+    const ast = parseSource(text, {
+      sourceType: 'unambiguous',
+      plugins: [...plugins, 'decorators-legacy'],
+      errorRecovery: true,
+      allowReturnOutsideFunction: true,
+      allowAwaitOutsideFunction: true,
+    });
+    const first = ast.errors?.[0];
+    return first ? `${where(first.loc)}: ${clean(first.message)}` : null;
+  } catch (err) {
+    return `${where(err.loc)}: ${clean(err.message)}`;
+  }
+}
+
+/**
+ * How many times in a row each file has come back unparseable.
+ *
+ * A model that has broken a file once usually fixes it. A model that has
+ * broken it three times is guessing at a structure it has lost track of, and
+ * will keep guessing: one run spent fifteen minutes on a single line before
+ * giving up and rewriting the file, which is what it should have been told to
+ * do after the second try.
+ */
+const brokenRuns = new Map();
+
+export function forgetBrokenRuns() { brokenRuns.clear(); }
+
+/**
+ * The note appended to a result: empty when the file parses, and the run of
+ * failures forgotten, so three good writes later a single slip is a slip again.
+ */
+const parseNote = (show, problem) => {
+  if (!problem) { brokenRuns.delete(show); return ''; }
+  return brokenNote(show, problem);
+};
+
+/** The warning appended to a result when a file does not parse. */
+const brokenNote = (show, problem) => {
+  const runs = (brokenRuns.get(show) ?? 0) + 1;
+  brokenRuns.set(show, runs);
+  const base = `
+
+⚠ ${show} does not parse — ${problem}.`;
+  if (runs < 3) return `${base} Fix it now: the build will fail on it.`;
+  return `${base} That is ${runs} attempts in a row on this file. Stop editing it: ` +
+    `write the whole file again with write_file, in one piece, rather than patching ` +
+    `a structure you have lost track of.`;
+};
+
+/**
+ * A module in a page with no build step that imports a stylesheet or an image.
+ *
+ * A bundler would take it; a browser refuses the whole module ("Failed to
+ * load module script ... MIME type text/css"), so none of the app's script
+ * runs and every button is dead. The syntax is fine, so nothing else notices
+ * until the page is opened — a live build shipped exactly this as done.
+ */
+const ASSET_IMPORT = /^[ \t]*import\s+(?:[\w$*{}\s,]+?\s+from\s+)?['"]([^'"]+\.(?:css|scss|sass|less|svg|png|jpe?g|gif|webp))['"](?!\s*(?:with|assert)\s*\{)/m;
+
+export function assetImport(abs, text) {
+  if (!/\.m?js$/i.test(abs)) return null;
+  // An import quoted inside a block comment is not an import.
+  const hit = ASSET_IMPORT.exec(String(text).replace(/\/\*[\s\S]*?\*\//g, ''));
+  if (!hit) return null;
+  // Anything with a package.json above it, up to the project root, may well be
+  // bundled. path.relative, not a string prefix: "ucode2" starts with "ucode",
+  // and Windows paths differ in case from one shell to the next.
+  const top = getRoot();
+  const within = (dir) => {
+    const rel = path.relative(top, dir);
+    return rel === '' || (!rel.startsWith('..') && !path.isAbsolute(rel));
+  };
+  for (let dir = path.dirname(abs); ; dir = path.dirname(dir)) {
+    if (existsSync(path.join(dir, 'package.json'))) return null;
+    if (!within(dir) || path.relative(top, dir) === '' || path.dirname(dir) === dir) break;
+  }
+  return `imports ${hit[1]}, which a page with no build step cannot do: the browser refuses the ` +
+    'whole module, so none of its script runs. Remove that import and load it from the page ' +
+    'instead — <link rel="stylesheet" href="…"> for a stylesheet, an <img> or a plain URL for an image.';
+}
+
+/**
+ * A page styled with Tailwind classes that never loads Tailwind.
+ *
+ * A live build wrote a 27 KB page of "flex px-4 bg-[var(--bg)]" and a
+ * stylesheet of colour tokens only: every class was ignored and the app came
+ * out as bare browser defaults, with nothing failing to tell anyone. When a
+ * page leans on those classes and has no Tailwind of its own, the browser
+ * build is added to its head, so the page looks the way it was written.
+ */
+const TAILWIND_CLASS = /^(?:-?(?:m|p)[trblxy]?-\S+|flex|grid|hidden|block|inline-flex|items-\S+|justify-\S+|gap-\S+|space-[xy]-\S+|(?:min-|max-)?[wh]-\S+|text-\S+|bg-\S+|rounded(?:-\S+)?|border(?:-\S+)?|shadow(?:-\S+)?|font-\S+|leading-\S+|tracking-\S+|(?:sm|md|lg|xl|hover|focus|dark):\S+)$/;
+export function withTailwind(html) {
+  if (/tailwindcss|cdn\.tailwind/i.test(html) || !/<\/head>/i.test(html)) return html;
+  let count = 0;
+  for (const [, list] of html.matchAll(/class="([^"]*)"/g)) {
+    for (const token of list.split(/\s+/)) if (TAILWIND_CLASS.test(token)) count++;
+    if (count >= 12) break;
+  }
+  if (count < 12) return html;
+  return html.replace(/<\/head>/i, '  <script src="https://cdn.jsdelivr.net/npm/@tailwindcss/browser@4"></script>\n</head>');
+}
+
+/** The note for that, appended to a write's result like a parse problem is. */
+const pageNote = (target, text) => {
+  const problem = assetImport(target.abs, text);
+  return problem ? `\n\n⚠ ${target.show} ${problem}` : '';
+};
+
+/** A file this short comes back whole after an edit; longer ones show the part around the change. */
+const SHOW_WHOLE = 250;
+const AROUND = 15;
+
+/**
+ * The file as it stands after an edit, with the same line-number gutter as
+ * read_file.
+ *
+ * Measured on a real build: the model read the same component thirteen times
+ * in one turn, because an edit's result showed only the lines it replaced and
+ * the next edit needed the file as it now was. Sending the current text back
+ * with the edit costs the same tokens the re-read would have, and saves the
+ * round trip every time.
+ */
+function nowReads(show, text, at, span) {
+  const lines = toLines(text);
+  const whole = lines.length <= SHOW_WHOLE;
+  const from = whole ? 1 : Math.max(1, at - AROUND);
+  const to = whole ? lines.length : Math.min(lines.length, at + span + AROUND);
+  const width = String(to).length;
+  const body = lines.slice(from - 1, to).map((l, i) => `${String(from + i).padStart(width)} | ${l}`).join('\n');
+  const heading = whole
+    ? `${show} now reads (all ${lines.length} lines`
+    : `${show} now reads, lines ${from}-${to} of ${lines.length}`;
+  return `\n\n${heading} — this is the current text, so there is no need to read it again):\n${body}`;
+}
+
+/** At most this many files in one read_files call. */
+const MAX_BATCH = 20;
+
+/**
+ * Several files in one call.
+ *
+ * Reading from disk takes a millisecond. What makes reading slow is the round
+ * trip around it: every file read on its own is a whole request to the model,
+ * and on a reasoning model that is several seconds of thinking before it even
+ * asks for the next one. Reading the six files a change touches in one call
+ * turns six of those into one.
+ *
+ * The files are read in parallel, and a missing one is reported in its place
+ * rather than failing the rest — one wrong path should not cost the other five.
+ */
+export async function readFiles({ paths, limit = READ_LINES }) {
+  if (!Array.isArray(paths) || paths.length === 0) {
+    throw new ToolFailure({
+      kind: 'bad_args',
+      attempted: 'reading several files',
+      failed: 'The "paths" argument must be a non-empty array of file paths.',
+      fix: 'Pass paths as ["src/app.js", "src/lib/api.js", ...].',
+    });
+  }
+
+  const wanted = [...new Set(paths.map((p) => String(p ?? '').trim()).filter(Boolean))];
+  const batch = wanted.slice(0, MAX_BATCH);
+  const dropped = wanted.slice(MAX_BATCH);
+
+  const readOne = (p) => readFile({ path: p, limit }).then(
+    (out) => ({ p, out }),
+    (err) => ({ p, err })
+  );
+
+  // Anything outside the project needs a yes, and two questions cannot be
+  // asked at once — so those go one at a time. Everything else goes together.
+  const outside = batch.some((p) => !resolveIn(p, 'read_files', 'paths').inside);
+  const settled = [];
+  if (outside) {
+    for (const p of batch) settled.push(await readOne(p));
+  } else {
+    settled.push(...(await Promise.all(batch.map(readOne))));
+  }
+
+  // Two whole files' worth of output between them. Past that, the rest are
+  // named rather than silently cut, so the model knows to ask again.
+  const budget = MAX_FILE_OUTPUT * 2;
+  let used = 0;
+  let read = 0;
+  let failed = 0;
+  let lines = 0;
+  const blocks = [];
+  const deferred = [];
+
+  for (const { p, out, err } of settled) {
+    if (err) {
+      failed++;
+      blocks.push(`=== ${p} — could not be read ===\n${err.forModel ? err.forModel() : err.message}`);
+      continue;
+    }
+    const block = `=== ${p} (${out.summary}) ===\n${out.content}`;
+    if (read > 0 && used + block.length > budget) {
+      deferred.push(p);
+      continue;
+    }
+    blocks.push(block);
+    used += block.length;
+    read++;
+    // "42 lines" for a whole file, "lines 1-600 of 900" for a page of one.
+    const whole = /^(\d+) lines$/.exec(out.summary);
+    const page = /^lines (\d+)-(\d+)/.exec(out.summary);
+    lines += whole ? Number(whole[1]) : page ? Number(page[2]) - Number(page[1]) + 1 : 0;
+  }
+
+  const leftOver = [...deferred, ...dropped];
+  if (leftOver.length) {
+    blocks.push(
+      `[not included, to stay inside one reply: ${leftOver.join(', ')}. ` +
+      'Read those with another read_files call.]'
+    );
+  }
+
+  return result(
+    blocks.join('\n\n'),
+    `${read} file${read === 1 ? '' : 's'} · ${lines} lines` +
+      (failed ? ` · ${failed} missing` : '') +
+      (leftOver.length ? ` · ${leftOver.length} deferred` : ''),
+    budget + 2_000
+  );
+}
+
+/** Write one file, returning the rows that show what changed. */
+async function put(target, content, { diffMax = 16 } = {}) {
+  const attempted = `writing ${target.show}`;
+  if (/\.html?$/i.test(target.abs)) content = withTailwind(content);
+
+  // Read what is there before clobbering it, so an overwrite can be shown as
+  // an actual diff rather than as a claim that something changed.
+  let previous = null;
+  try {
+    previous = await fs.readFile(target.abs, 'utf8');
+  } catch {
+    previous = null; // missing, or binary — either way it is treated as new
+  }
+
+  if (previous !== null) await assertUnchanged(target.abs, target.show);
+
+  try {
+    await fs.mkdir(path.dirname(target.abs), { recursive: true });
+    await remember(target.abs);
+    await writeTracked(target.abs, content);
+  } catch (err) {
+    throw fsFailure(err, attempted, target.show);
+  }
+  written(target, content);
+
+  const existed = previous !== null;
+  const lineCount = content === '' ? 0 : toLines(content).length;
+  const diff = existed
+    ? renderDiff(changedRegion(previous, content), { max: diffMax })
+    : (content === '' ? [] : renderNewFile(content, diffMax));
+
+  return {
+    existed,
+    lineCount,
+    diff,
+    problem: syntaxProblem(target.abs, content),
+    note: pageNote(target, content),
+    line: `${existed ? 'Overwrote' : 'Created'} ${target.show} ` +
+      `(${lineCount} lines, ${bytes(Buffer.byteLength(content))})`,
+  };
+}
+
+export async function writeFile({ path: p, content }) {
+  const target = resolveIn(p, 'write_file');
+  if (typeof content !== 'string') {
+    throw new ToolFailure({
+      kind: 'bad_args',
+      attempted: `writing ${target.show}`,
+      failed: 'The "content" argument was missing or was not a string.',
+      fix: 'Call write_file again with content set to the whole text of the file.',
+    });
+  }
+  await guard(target, `write ${target.abs}`);
+
+  const written = await put(target, content);
+  const out = result(
+    `${written.line}.${parseNote(target.show, written.problem)}${written.note}`,
+    `${written.existed ? 'overwrote' : 'created'} · ${written.lineCount} lines${written.problem ? ' · does not parse' : ''}`
+  );
+  out.diff = written.diff;
+  return out;
+}
+
+/**
+ * Several files in one call.
+ *
+ * Scaffolding a project is twenty writes before anything can be run, and doing
+ * that one round trip at a time is most of the wait.
+ */
+export async function batchWrite({ files }) {
+  if (!Array.isArray(files) || files.length === 0) {
+    throw new ToolFailure({
+      kind: 'bad_args',
+      attempted: 'writing several files',
+      failed: 'The "files" argument must be a non-empty array.',
+      fix: 'Pass files as [{ path, content }, ...].',
+    });
+  }
+
+  const lines = [];
+  const diff = [];
+  let created = 0;
+  let broken = 0;
+
+  for (const [index, file] of files.entries()) {
+    const { path: p, content } = file ?? {};
+    if (typeof p !== 'string' || typeof content !== 'string') {
+      throw new ToolFailure({
+        kind: 'bad_args',
+        attempted: `writing file ${index + 1} of ${files.length}`,
+        failed: 'Every entry needs "path" and "content", both strings.',
+        fix: `Fix entry ${index + 1} and call batch_write again. ${index} file(s) were already written.`,
+      });
+    }
+
+    const target = resolveIn(p, 'batch_write');
+    await guard(target, `write ${target.abs}`);
+
+    // Per-file diffs are kept short here; twenty files at sixteen rows each
+    // would bury the reply under three hundred lines of gutter.
+    const written = await put(target, content, { diffMax: 6 });
+    if (!written.existed) created++;
+    lines.push(written.line + parseNote(target.show, written.problem) + written.note);
+    if (written.problem) broken++;
+    diff.push(`~${target.show}`, ...written.diff);
+  }
+
+  const out = result(
+    lines.join('\n'),
+    `${files.length} file${files.length === 1 ? '' : 's'} · ${created} new${broken ? ` · ${broken} do not parse` : ''}`
+  );
+  out.diff = diff;
+  return out;
+}
+
+/**
+ * Why an old_string missed, worked out rather than guessed at.
+ *
+ * "not found" tells the model nothing it did not already know. Whether the
+ * text is present with different whitespace, or present but only its first
+ * line, is the difference between a fix on the next step and three more
+ * failed attempts.
+ */
+function explainMiss(original, oldString, show) {
+  const flatten = (s) => s.replace(/\s+/g, ' ').trim();
+  const firstLine = oldString.split('\n')[0].trim();
+
+  if (flatten(original).includes(flatten(oldString))) {
+    return {
+      failed: `old_string is not in ${show} as written — the text is there, but the whitespace differs.`,
+      fix: "Match the file's own indentation exactly: tabs versus spaces, and the line breaks.",
+    };
+  }
+
+  const nearby = firstLine.length > 3
+    ? original.split(/\r?\n/)
+        .map((line, i) => [i + 1, line])
+        .filter(([, line]) => line.includes(firstLine))
+        .slice(0, 3)
+    : [];
+
+  if (nearby.length) {
+    return {
+      failed:
+        `old_string is not in ${show}. Its first line does appear at ` +
+        `line${nearby.length > 1 ? 's' : ''} ${nearby.map(([n]) => n).join(', ')}, ` +
+        'so it is the lines after it that differ.',
+      fix: `Read ${show} around line ${nearby[0][0]} and copy the block exactly as it is.`,
+    };
+  }
+
+  return {
+    failed: `old_string does not appear anywhere in ${show}.`,
+    fix: `Read ${show} again and copy the text verbatim, without the line-number gutter.`,
+  };
+}
+
+/** Apply one replacement to a string, or explain precisely why it cannot. */
+function replaceOnce(text, { old_string, new_string, replace_all }, { show, attempted, label = '' }) {
+  const prefix = label ? `${label}: ` : '';
+
+  if (typeof old_string !== 'string' || typeof new_string !== 'string') {
+    throw new ToolFailure({
+      kind: 'bad_args', attempted,
+      failed: `${prefix}old_string and new_string must both be strings.`,
+      fix: 'Fix that entry and call again. Nothing was written.',
+    });
+  }
+  if (old_string === '') {
+    throw new ToolFailure({
+      kind: 'bad_args', attempted,
+      failed: `${prefix}old_string was empty.`,
+      fix: 'edit_file replaces existing text. Use write_file to create a file.',
+    });
+  }
+  // An edit whose two halves are the same asks for the file to stay as it is,
+  // which it will. Refusing that was a hard failure, and the model answered it
+  // by sending the same edit again — the stuck detector carries a special case
+  // for exactly this loop. Saying "already done" ends it in one step.
+  if (old_string === new_string) {
+    const found = text.indexOf(old_string);
+    return { text, at: found < 0 ? 1 : toLines(text.slice(0, found)).length, how: '', count: 0 };
+  }
+
+  // Models write \n. A file checked out on Windows is often \r\n, and then an
+  // otherwise perfect old_string can never match. Speak the file's dialect.
+  let oldText = old_string;
+  let newText = new_string;
+  if (text.includes('\r\n') && !oldText.includes('\r')) {
+    oldText = oldText.replace(/\r?\n/g, '\r\n');
+    newText = newText.replace(/\r?\n/g, '\r\n');
+  }
+
+  const ambiguous = (hits, how = '') => new ToolFailure({
+    kind: 'ambiguous', attempted,
+    failed: `${prefix}old_string appears ${hits} times in ${show}${how}. Refusing to guess which one you meant.`,
+    fix: 'Add surrounding lines to old_string until it matches exactly one place.',
+    detail: { hits },
+  });
+
+  // replace_all is for renames: every copy changes, and many copies is the point.
+  const all = replace_all === true || replace_all === 'true';
+  const hits = text.split(oldText).length - 1;
+  if (hits > 1 && !all) throw ambiguous(hits);
+  if (hits >= 1) {
+    const at = text.slice(0, text.indexOf(oldText)).split(/\r?\n/).length;
+    const out = all ? text.split(oldText).join(newText) : text.replace(oldText, () => newText);
+    return { text: out, at, how: '', count: hits };
+  }
+
+  // No exact match. The commonest reason by far is whitespace — tabs against
+  // spaces, a different indent depth, trailing spaces — with every word right.
+  // Match line by line ignoring that, and re-indent the replacement to fit.
+  // Still unique or nothing: a loose match found twice is refused like any other.
+  const loose = all ? null : looseReplace(text, old_string, new_string);
+  if (loose?.count === 1) {
+    return { text: loose.text, at: loose.at, how: 'ignoring whitespace and re-indented to fit', count: 1 };
+  }
+  if (loose?.count > 1) throw ambiguous(loose.count, ' once whitespace is ignored');
+
+  // Still nothing. The remaining slips — a middle line remembered slightly
+  // wrong, escapes written out, a blank line at either end — each have a
+  // matcher of their own (fuzzy.js). They are given the edit in the file's
+  // own line endings, and only the matched span changes: the rest of a file
+  // with mixed endings keeps every one it had.
+  const fuzzy = fuzzyReplace(text, oldText, newText, { all });
+  if (fuzzy?.ambiguous) throw ambiguous('several', ' once matched loosely');
+  if (fuzzy?.wide) {
+    throw new ToolFailure({
+      kind: 'no_match', attempted,
+      failed: `${prefix}old_string only matches ${show} loosely, across far more text than it contains. Refusing to replace that much.`,
+      fix: `Read ${show} again and copy the exact text you mean to replace.`,
+    });
+  }
+  if (fuzzy) {
+    const at = text.slice(0, fuzzy.index).split('\n').length;
+    return { text: fuzzy.text, at, how: fuzzy.how, count: fuzzy.count };
+  }
+
+  const { failed, fix } = explainMiss(text, old_string, show);
+  throw new ToolFailure({ kind: 'no_match', attempted, failed: prefix + failed, fix });
+}
+
+/**
+ * Find old_string by its lines' content alone and swap in new_string, indented
+ * the way the file is indented at that spot. Returns { count } when it finds
+ * none or several, and { count: 1, text, at } when it finds exactly one.
+ */
+function looseReplace(text, oldString, newString) {
+  const eol = text.includes('\r\n') ? '\r\n' : '\n';
+  const lines = text.split(/\r?\n/);
+
+  const want = oldString.replace(/\r/g, '').split('\n');
+  while (want.length > 1 && !want[want.length - 1].trim()) want.pop();
+  while (want.length > 1 && !want[0].trim()) want.shift();
+  const target = want.map((l) => l.trim());
+  if (target.every((t) => !t)) return null;
+
+  const starts = [];
+  for (let i = 0; i + want.length <= lines.length; i++) {
+    let same = true;
+    for (let j = 0; j < want.length; j++) {
+      if (lines[i + j].trim() !== target[j]) { same = false; break; }
+    }
+    if (same) starts.push(i);
+  }
+  if (starts.length !== 1) return { count: starts.length };
+
+  const start = starts[0];
+  const indent = (l) => /^[ \t]*/.exec(l)[0];
+  const first = want.findIndex((l) => l.trim());
+  const fileIndent = indent(lines[start + first]);
+  const wroteIndent = indent(want[first]);
+
+  const replacement = newString.replace(/\r/g, '').split('\n');
+  if (replacement.length > 1 && replacement[replacement.length - 1] === '') replacement.pop();
+  const reindented = replacement.map((l) => {
+    if (!l.trim()) return l.trim();
+    const body = l.startsWith(wroteIndent) ? l.slice(wroteIndent.length) : l.replace(/^[ \t]*/, '');
+    return fileIndent + body;
+  });
+
+  const out = [...lines.slice(0, start), ...reindented, ...lines.slice(start + want.length)];
+  return { count: 1, text: out.join(eol), at: start + 1 };
+}
+
+export async function editFile({ path: p, old_string, new_string, replace_all }) {
+  const target = resolveIn(p, 'edit_file');
+  const attempted = `editing ${target.show}`;
+  await guard(target, `edit ${target.abs}`);
+
+  let original;
+  try {
+    original = await fs.readFile(target.abs, 'utf8');
+  } catch (err) {
+    throw fsFailure(err, attempted, target.show);
+  }
+
+  const { text, at, how, count } = replaceOnce(original, { old_string, new_string, replace_all }, {
+    show: target.show, attempted,
+  });
+
+  try {
+    await remember(target.abs);
+    await writeTracked(target.abs, text);
+  } catch (err) {
+    throw fsFailure(err, attempted, target.show);
+  }
+  written(target, text);
+
+  const delta = toLines(text).length - toLines(original).length;
+  const change = delta === 0 ? 'same line count' : `${delta > 0 ? '+' : ''}${delta} lines`;
+  const where = count > 1
+    ? `${count} occurrences in ${target.show}, the first at line ${at}`
+    : `one occurrence in ${target.show} at line ${at}`;
+  const matched = how ? `, matched ${how}` : '';
+
+  const span = toLines(new_string).length;
+  const out = result(
+    `Replaced ${where} (${change}${matched}).` +
+      parseNote(target.show, syntaxProblem(target.abs, text)) + pageNote(target, text) +
+      nowReads(target.show, text, at, span),
+    `${count > 1 ? `${count} changes from line` : '1 change at line'} ${at} · ${change}` +
+      `${how ? ' · whitespace-tolerant' : ''}${syntaxProblem(target.abs, text) ? ' · does not parse' : ''}`,
+    MAX_FILE_OUTPUT
+  );
+  // The replacement is diffed on its own and offset to where it landed, so
+  // the gutter shows the file's line numbers rather than 1, 2, 3.
+  out.diff = renderDiff(changedRegion(old_string, new_string), { offset: at - 1 });
+  return out;
+}
+
+/**
+ * Several replacements in one file, applied in order, each seeing the result
+ * of the one before it.
+ *
+ * Everything is validated against a working copy first. If the third edit is
+ * ambiguous, none of the three are written — a half-applied set of edits is a
+ * file in a state nobody designed.
+ */
+export async function multiEdit({ path: p, edits }) {
+  const target = resolveIn(p, 'multi_edit');
+  const attempted = `editing ${target.show}`;
+
+  if (!Array.isArray(edits) || edits.length === 0) {
+    throw new ToolFailure({
+      kind: 'bad_args',
+      attempted,
+      failed: 'The "edits" argument must be a non-empty array.',
+      fix: 'Pass edits as [{ old_string, new_string }, ...].',
+    });
+  }
+
+  await guard(target, `edit ${target.abs}`);
+
+  let original;
+  try {
+    original = await fs.readFile(target.abs, 'utf8');
+  } catch (err) {
+    throw fsFailure(err, attempted, target.show);
+  }
+
+  let text = original;
+  const diff = [];
+
+  for (const [index, edit] of edits.entries()) {
+    const applied = replaceOnce(text, edit ?? {}, {
+      show: target.show,
+      attempted,
+      label: `edit ${index + 1} of ${edits.length}`,
+    });
+    diff.push(
+      ...renderDiff(changedRegion(edit.old_string, edit.new_string), {
+        offset: applied.at - 1,
+        max: 8,
+      })
+    );
+    text = applied.text;
+  }
+
+  try {
+    await remember(target.abs);
+    await writeTracked(target.abs, text);
+  } catch (err) {
+    throw fsFailure(err, attempted, target.show);
+  }
+  written(target, text);
+
+  const delta = toLines(text).length - toLines(original).length;
+  const change = delta === 0 ? 'same line count' : `${delta > 0 ? '+' : ''}${delta} lines`;
+
+  const out = result(
+    `Applied ${edits.length} edits to ${target.show} (${change}).` +
+      parseNote(target.show, syntaxProblem(target.abs, text)) + pageNote(target, text) +
+      nowReads(target.show, text, 1, toLines(text).length),
+    `${edits.length} edits · ${change}`,
+    MAX_FILE_OUTPUT
+  );
+  out.diff = diff;
+  return out;
+}
+
+/**
+ * Exact replacements across several files in one call.
+ *
+ * A change that touches the route, the component and the type together is one
+ * round trip instead of three. Every edit in every file is applied to a copy
+ * in memory first; if any of them fails, nothing is written anywhere — a
+ * cross-file change that half landed leaves the project in a state that
+ * compiles nowhere.
+ */
+export async function editFiles({ files }) {
+  if (!Array.isArray(files) || files.length === 0) {
+    throw new ToolFailure({
+      kind: 'bad_args',
+      attempted: 'editing several files',
+      failed: 'The "files" argument must be a non-empty array.',
+      fix: 'Pass files as [{ path, edits: [{ old_string, new_string }, ...] }, ...].',
+    });
+  }
+
+  const planned = [];
+  const seen = new Set();
+
+  for (const [i, entry] of files.entries()) {
+    const target = resolveIn(entry?.path, 'edit_files');
+    const attempted = `editing ${target.show}`;
+
+    if (seen.has(target.abs)) {
+      throw new ToolFailure({
+        kind: 'bad_args', attempted,
+        failed: `${target.show} is listed twice.`,
+        fix: 'List each file once, with all of its edits together. Nothing was written.',
+      });
+    }
+    seen.add(target.abs);
+
+    if (!Array.isArray(entry.edits) || entry.edits.length === 0) {
+      throw new ToolFailure({
+        kind: 'bad_args', attempted,
+        failed: `File ${i + 1} (${target.show}) has no edits.`,
+        fix: 'Give every file a non-empty edits array. Nothing was written.',
+      });
+    }
+
+    await guard(target, `edit ${target.abs}`);
+
+    let original;
+    try {
+      original = await fs.readFile(target.abs, 'utf8');
+    } catch (err) {
+      throw fsFailure(err, attempted, target.show);
+    }
+
+    let text = original;
+    const diff = [`~${target.show}`];
+    for (const [j, edit] of entry.edits.entries()) {
+      const applied = replaceOnce(text, edit ?? {}, {
+        show: target.show,
+        attempted,
+        label: `${target.show}, edit ${j + 1} of ${entry.edits.length} (nothing was written)`,
+      });
+      diff.push(...renderDiff(changedRegion(edit.old_string, edit.new_string), { offset: applied.at - 1, max: 6 }));
+      text = applied.text;
+    }
+    planned.push({ target, text, diff, count: entry.edits.length });
+  }
+
+  for (const { target, text } of planned) {
+    try {
+      await remember(target.abs);
+    await writeTracked(target.abs, text);
+    } catch (err) {
+      throw fsFailure(err, `editing ${target.show}`, target.show);
+    }
+    written(target, text);
+  }
+
+  const edits = planned.reduce((n, p) => n + p.count, 0);
+  const out = result(
+    planned.map((p) => {
+      const problem = syntaxProblem(p.target.abs, p.text);
+      return `Edited ${p.target.show} (${p.count} change${p.count === 1 ? '' : 's'})` +
+        parseNote(p.target.show, problem) + pageNote(p.target, p.text);
+    }).join('\n'),
+    `${planned.length} files · ${edits} edits`
+  );
+  out.diff = planned.flatMap((p) => p.diff);
+  return out;
+}