codeply-cli 0.3.1 → 0.4.0

package/lib/edit-engine.js

@@ -1,364 +1,364 @@
-/**
- * Codeply CLI - surgical SEARCH/REPLACE edit engine
- *
- * Adapted from Codeply-App/main.js's computeInstructionEdits/
- * applySearchReplace/applyEditsToContent - same approach (small verified
- * diffs instead of a full-file rewrite), trimmed for a stateless single-shot
- * CLI: no local usage-history file, no .codeply/ response cache, no
- * multi-turn conversation context (each CLI invocation is its own process).
- */
-const fs = require('fs');
-const path = require('path');
-const ai = require('./ai');
-
-// Apply ONE search/replace edit to a string. Exact match first, then
-// whitespace-tolerant (ignore leading/trailing space, then collapse internal
-// runs), skipping blank lines on both sides. Re-indents the replacement to fit.
-// Returns { ok, content } or { ok:false, error:'notfound'|'multiple'|'empty' }.
-function applySearchReplace(fileContent, searchBlock, replaceBlock) {
-  // Matching works on LF text; a CRLF file is written back as CRLF so an edit
-  // never rewrites every line ending in the file (a whole-file diff, and a
-  // mixed-ending file once a later write adds LF lines).
-  const res = applySearchReplaceLF(fileContent, searchBlock, replaceBlock);
-  if (res.ok && /\r\n/.test(fileContent || '')) res.content = res.content.replace(/\r?\n/g, '\r\n');
-  return res;
-}
-
-function applySearchReplaceLF(fileContent, searchBlock, replaceBlock) {
-  const nl = (s) => (s || '').replace(/\r\n/g, '\n');
-  fileContent = nl(fileContent); searchBlock = nl(searchBlock); replaceBlock = nl(replaceBlock);
-  if (!searchBlock.trim()) return { ok: false, error: 'empty' };
-
-  const fileLines = fileContent.split('\n');
-  const searchLines = searchBlock.split('\n');
-
-  const lineStartOffset = (idx) => { let o = 0; for (let k = 0; k < idx; k++) o += fileLines[k].length + 1; return o; };
-
-  const reindentBlock = (block, indent) => {
-    const ls = block.split('\n');
-    const firstNonEmpty = ls.find(l => l.trim() !== '') || '';
-    const base = (firstNonEmpty.match(/^[ \t]*/) || [''])[0];
-    return ls.map(l => {
-      if (l.trim() === '') return '';
-      const stripped = l.startsWith(base) ? l.slice(base.length) : l.replace(/^[ \t]*/, '');
-      return indent + stripped;
-    }).join('\n');
-  };
-
-  const locate = (normalize) => {
-    const s = searchLines.map(normalize).filter(x => x !== '');
-    if (!s.length) return [];
-    const found = [];
-    for (let start = 0; start < fileLines.length; start++) {
-      if (normalize(fileLines[start]) !== s[0]) continue;
-      let fi = start, si = 0, end = start;
-      while (si < s.length && fi < fileLines.length) {
-        const nf = normalize(fileLines[fi]);
-        if (nf === '') { fi++; continue; }       // skip blank lines in the file
-        if (nf !== s[si]) break;
-        end = fi; fi++; si++;
-      }
-      if (si === s.length) found.push([start, end]);
-    }
-    return found;
-  };
-
-  const exact = fileContent.indexOf(searchBlock);
-  if (exact !== -1) {
-    if (exact !== fileContent.lastIndexOf(searchBlock)) return { ok: false, error: 'multiple' };
-    return { ok: true, content: fileContent.slice(0, exact) + replaceBlock + fileContent.slice(exact + searchBlock.length) };
-  }
-
-  let matches = locate(l => l.trim());
-  if (matches.length === 0) matches = locate(l => l.trim().replace(/\s+/g, ' '));
-
-  if (matches.length === 0) {
-    // Last resort: anchor on the first AND last non-blank lines of the search
-    // block (recovers when the AI dropped/added a line in the middle). Kept
-    // conservative - both anchors must be unique and the span must be modest.
-    const sNon = searchLines.map(l => l.trim()).filter(x => x !== '');
-    if (sNon.length >= 2) {
-      const firstHits = [], lastHits = [];
-      fileLines.forEach((l, i) => {
-        const t = l.trim();
-        if (t === sNon[0]) firstHits.push(i);
-        if (t === sNon[sNon.length - 1]) lastHits.push(i);
-      });
-      if (firstHits.length === 1 && lastHits.length === 1 && lastHits[0] >= firstHits[0]) {
-        const sL = firstHits[0], eL = lastHits[0];
-        if (eL - sL + 1 <= sNon.length * 3 + 5) {   // don't grab a huge region
-          const indent = (fileLines[sL].match(/^[ \t]*/) || [''])[0];
-          const reindented = reindentBlock(replaceBlock, indent);
-          return { ok: true, content: fileContent.slice(0, lineStartOffset(sL)) + reindented + fileContent.slice(lineStartOffset(eL) + fileLines[eL].length) };
-        }
-      }
-    }
-    const fuzzy = fuzzyLocate(fileContent, searchBlock);
-    if (fuzzy.ok) {
-      const at = fileContent.indexOf(fuzzy.span);
-      // The model's replacement carries the indentation it wrongly assumed;
-      // move it to where the block really sits.
-      const text = fuzzy.reindent
-        ? reindentBlock(replaceBlock, (fuzzy.span.split('\n').find((l) => l.trim()) || '').match(/^[ \t]*/)[0])
-        : replaceBlock;
-      return { ok: true, content: fileContent.slice(0, at) + text + fileContent.slice(at + fuzzy.span.length), fuzzy: true };
-    }
-    return { ok: false, error: fuzzy.error === 'multiple' ? 'multiple' : 'notfound' };
-  }
-  if (matches.length > 1) return { ok: false, error: 'multiple' };
-
-  const [startLine, endLine] = matches[0];
-  const indent = (fileLines[startLine].match(/^[ \t]*/) || [''])[0];
-  const reindented = reindentBlock(replaceBlock, indent);
-  const startChar = lineStartOffset(startLine);
-  const endChar = lineStartOffset(endLine) + fileLines[endLine].length;
-  return { ok: true, content: fileContent.slice(0, startChar) + reindented + fileContent.slice(endChar) };
-}
-
-// ─── Fallback matchers ──────────────────────────────────────────────────────
-//
-// Adapted from opencode (packages/opencode/src/tool/edit.ts), MIT License,
-// Copyright (c) 2025 opencode - https://github.com/sst/opencode
-//
-// Each matcher yields candidate spans that exist VERBATIM in the file for a
-// search block the model copied imperfectly (escaped quotes, shifted
-// indentation, a middle line reworded). They only run after the matcher above
-// has found nothing, and a candidate is used only when it occurs exactly once
-// and is not far larger than what the model asked for.
-
-function levenshtein(a, b) {
-  if (a === '' || b === '') return Math.max(a.length, b.length);
-  let prev = Array.from({ length: b.length + 1 }, (_, j) => j);
-  for (let i = 1; i <= a.length; i++) {
-    const cur = [i];
-    for (let j = 1; j <= b.length; j++) {
-      cur[j] = Math.min(prev[j] + 1, cur[j - 1] + 1, prev[j - 1] + (a[i - 1] === b[j - 1] ? 0 : 1));
-    }
-    prev = cur;
-  }
-  return prev[b.length];
-}
-
-function spanOfLines(lines, start, end) {
-  return lines.slice(start, end + 1).join('\n');
-}
-
-const ANCHOR_SIMILARITY = 0.65;
-
-// First and last lines match exactly (trimmed); the middle only has to be
-// similar. Recovers a search block with one line misremembered.
-function* blockAnchorMatches(content, find) {
-  const lines = content.split('\n');
-  const want = find.split('\n');
-  if (want[want.length - 1] === '') want.pop();
-  if (want.length < 3) return;
-  const first = want[0].trim(), last = want[want.length - 1].trim();
-  const maxDelta = Math.max(1, Math.floor(want.length * 0.25));
-
-  const candidates = [];
-  for (let i = 0; i < lines.length; i++) {
-    if (lines[i].trim() !== first) continue;
-    for (let j = i + 2; j < lines.length; j++) {
-      if (lines[j].trim() !== last) continue;
-      if (Math.abs(j - i + 1 - want.length) <= maxDelta) candidates.push([i, j]);
-      break;
-    }
-  }
-
-  let best = null, bestScore = -1;
-  for (const [s, e] of candidates) {
-    const size = e - s + 1;
-    const middle = Math.min(want.length - 2, size - 2);
-    let score = 1;
-    if (middle > 0) {
-      score = 0;
-      for (let k = 1; k < want.length - 1 && k < size - 1; k++) {
-        const a = lines[s + k].trim(), b = want[k].trim();
-        const len = Math.max(a.length, b.length);
-        score += len ? 1 - levenshtein(a, b) / len : 1;
-      }
-      score /= middle;
-    }
-    if (score > bestScore) { bestScore = score; best = [s, e]; }
-  }
-  if (best && bestScore >= ANCHOR_SIMILARITY) yield spanOfLines(lines, best[0], best[1]);
-}
-
-// Same block with its indentation shifted as a whole.
-function* indentationFlexibleMatches(content, find) {
-  const dedent = (text) => {
-    const ls = text.split('\n');
-    const indents = ls.filter((l) => l.trim()).map((l) => l.match(/^\s*/)[0].length);
-    if (!indents.length) return text;
-    const min = Math.min(...indents);
-    return ls.map((l) => (l.trim() ? l.slice(min) : l)).join('\n');
-  };
-  const target = dedent(find);
-  const lines = content.split('\n');
-  const n = find.split('\n').length;
-  for (let i = 0; i + n <= lines.length; i++) {
-    const block = lines.slice(i, i + n).join('\n');
-    if (dedent(block) === target) yield block;
-  }
-}
-
-// The model wrote \n, \" or \` escapes where the file has the real characters.
-function* escapeNormalizedMatches(content, find) {
-  const unescape = (s) => s.replace(/\\(n|t|r|'|"|`|\\|\n|\$)/g, (m, c) =>
-    ({ n: '\n', t: '\t', r: '\r', "'": "'", '"': '"', '`': '`', '\\': '\\', '\n': '\n', $: '$' }[c] ?? m));
-  const target = unescape(find);
-  if (target !== find && content.includes(target)) yield target;
-  const lines = content.split('\n');
-  const n = target.split('\n').length;
-  for (let i = 0; i + n <= lines.length; i++) {
-    const block = lines.slice(i, i + n).join('\n');
-    if (block !== target && unescape(block) === target) yield block;
-  }
-}
-
-// Stray blank lines or spaces around the block.
-function* trimmedBoundaryMatches(content, find) {
-  const t = find.trim();
-  if (t === find) return;
-  if (content.includes(t)) yield t;
-}
-
-const FALLBACK_MATCHERS = [indentationFlexibleMatches, escapeNormalizedMatches, trimmedBoundaryMatches, blockAnchorMatches];
-
-// A "match" many times bigger than the search block means an anchor latched
-// onto the wrong closing line; replacing it would silently delete code.
-function isDisproportionate(span, find) {
-  const spanLines = span.split('\n').length, findLines = find.split('\n').length;
-  if (spanLines >= Math.max(findLines + 3, findLines * 2)) return true;
-  if (findLines === 1) return false;
-  return span.trim().length > Math.max(find.trim().length + 500, find.trim().length * 4);
-}
-
-/**
- * Find the one verbatim span in `content` that an imperfect search block most
- * likely meant. @returns {{ok:true, span:string} | {ok:false, error:'notfound'|'multiple'|'disproportionate'}}
- */
-function fuzzyLocate(content, find) {
-  let sawMultiple = false;
-  for (const matcher of FALLBACK_MATCHERS) {
-    for (const span of matcher(content, find)) {
-      if (!span.trim()) continue;
-      const at = content.indexOf(span);
-      if (at === -1) continue;
-      if (at !== content.lastIndexOf(span)) { sawMultiple = true; continue; }
-      if (isDisproportionate(span, find)) return { ok: false, error: 'disproportionate' };
-      return { ok: true, span, reindent: matcher === indentationFlexibleMatches || matcher === blockAnchorMatches };
-    }
-  }
-  return { ok: false, error: sawMultiple ? 'multiple' : 'notfound' };
-}
-
-// Apply an ordered list of verified {search, replace} hunks to file content,
-// all-or-nothing (pure - no disk write).
-function applyEditsToContent(fileContent, edits) {
-  let working = (fileContent || '').replace(/\r\n/g, '\n');
-  for (let i = 0; i < edits.length; i++) {
-    const { search, replace } = edits[i];
-    const res = applySearchReplace(working, search, replace);
-    if (!res.ok) return { ok: false, error: res.error, index: i };
-    working = res.content;
-  }
-  return { ok: true, content: working };
-}
-
-const SYSTEM_PROMPT = `You are a precise code editor. You receive an INSTRUCTION describing a small change, and the FULL current file. Make ONLY the specific change requested, as surgical SEARCH/REPLACE edits - do NOT reproduce the whole file.
-
-STRICT RULES:
-1. Return an "edits" array. Use the FEWEST edits that cleanly express the change.
-2. "search" must be copied VERBATIM from the file below - exact existing text being changed, with just enough surrounding lines to be unique.
-3. "replace" is that same block with ONLY the requested change applied - preserve everything else in it exactly (formatting, unrelated properties, comments, whitespace style).
-4. If the SAME property/selector/value appears more than once in the file (e.g. "top" set in both a "from" and a "to" block), do NOT pick the first or most obvious match - use every clue in the instruction to find the ONE correct occurrence: the CURRENT value it mentions, nearby selectors, rule/keyframe names, or ordering (first/last, start/end). If more than one occurrence still fits equally well after that, return an EMPTY "edits" array with a "reason" naming the ambiguity instead of guessing.
-5. If the instruction genuinely requires large new content that can't be expressed as a small edit (e.g. "build a whole new page/game/module from scratch"), return an EMPTY "edits" array - do not guess badly at a huge diff.
-6. Return ONLY valid JSON, no markdown, no commentary.
-
-Response format:
-{
-  "edits": [ { "search": "<exact verbatim block from the file>", "replace": "<that block with the change applied>" } ],
-  "reason": "<one short sentence>",
-  "confidence": <0-100>
-}`;
-
-/**
- * Asks the AI for surgical edits, self-corrects up to twice against bad
- * "search" matches, and returns only the edits that verify cleanly against
- * the real file content.
- * @returns {Promise<{success:boolean, edits?:Array, badCount?:number, reason?:string, confidence?:number, tokensUsed?:number, modelUsed?:string, error?:string}>}
- */
-async function computeInstructionEdits(instruction, filePath) {
-  if (!fs.existsSync(filePath)) return { success: false, error: `File not found: ${filePath}` };
-  const content = fs.readFileSync(filePath, 'utf8');
-
-  const probe = (search) => {
-    const res = applySearchReplace(content, search, search);
-    return res.ok ? 'ok' : res.error;
-  };
-
-  let totalTokens = 0;
-  let lastModelUsed = null;
-
-  const askModel = async (feedback) => {
-    const messages = [
-      { role: 'system', content: SYSTEM_PROMPT },
-      // Large/stable content (the file) first, small/always-different content
-      // (the instruction) last - the ai-proxy's underlying model caches a
-      // repeated PREFIX, so this ordering matters for repeated calls against
-      // the same file even when the instruction text differs each time.
-      { role: 'user', content: `FULL FILE (${path.basename(filePath)}):\n${content}\n\nINSTRUCTION:\n${instruction}` },
-    ];
-    if (feedback) messages.push({ role: 'user', content: feedback });
-
-    const aiResult = await ai.chat(messages, { json: true, meta: { promptText: instruction, filePath } });
-    if (!aiResult.success) throw new Error(aiResult.error);
-    if (aiResult.modelUsed) lastModelUsed = aiResult.modelUsed;
-    const data = aiResult.data;
-    totalTokens += (data.usage || {}).total_tokens || 0;
-
-    let parsed;
-    try { parsed = JSON.parse(data.choices[0].message.content); }
-    catch { return { edits: [], confidence: 0, reason: '' }; }
-
-    const edits = Array.isArray(parsed.edits)
-      ? parsed.edits.filter(e => e && e.search && e.replace != null).map(e => ({ search: e.search, replace: e.replace }))
-      : [];
-    return { edits, confidence: parsed.confidence == null ? 75 : parsed.confidence, reason: parsed.reason || '' };
-  };
-
-  let edits, confidence, reason;
-  try {
-    ({ edits, confidence, reason } = await askModel());
-  } catch (e) {
-    return { success: false, error: e.message };
-  }
-
-  const verify = (list) => list.map((e) => ({ e, status: probe(e.search) }));
-  let checked = verify(edits);
-  let bad = checked.filter(c => c.status !== 'ok');
-
-  let pass = 0;
-  while (bad.length && pass < 2) {
-    pass++;
-    const feedback =
-      `Some "search" blocks from your last answer are wrong. Return the COMPLETE corrected edits array again, fixing these:\n` +
-      bad.map(c => c.status === 'multiple'
-        ? `- Matched MULTIPLE places - include more surrounding lines so it is unique:\n${c.e.search}`
-        : `- NOT found in the file - copy it EXACTLY from the FULL FILE above, character for character:\n${c.e.search}`
-      ).join('\n') +
-      `\nEvery "search" must be copied verbatim from the file shown above.`;
-    let retry;
-    try { retry = await askModel(feedback); }
-    catch (e) { break; }
-    if (retry.edits.length) { edits = retry.edits; confidence = retry.confidence; reason = retry.reason || reason; }
-    checked = verify(edits);
-    bad = checked.filter(c => c.status !== 'ok');
-  }
-
-  const good = checked.filter(c => c.status === 'ok').map(c => c.e);
-  return { success: true, edits: good, badCount: bad.length, confidence, reason, tokensUsed: totalTokens, modelUsed: lastModelUsed };
-}
-
-module.exports = { applySearchReplace, applyEditsToContent, computeInstructionEdits, fuzzyLocate };
+/**
+ * Codeply CLI - surgical SEARCH/REPLACE edit engine
+ *
+ * Adapted from Codeply-App/main.js's computeInstructionEdits/
+ * applySearchReplace/applyEditsToContent - same approach (small verified
+ * diffs instead of a full-file rewrite), trimmed for a stateless single-shot
+ * CLI: no local usage-history file, no .codeply/ response cache, no
+ * multi-turn conversation context (each CLI invocation is its own process).
+ */
+const fs = require('fs');
+const path = require('path');
+const ai = require('./ai');
+
+// Apply ONE search/replace edit to a string. Exact match first, then
+// whitespace-tolerant (ignore leading/trailing space, then collapse internal
+// runs), skipping blank lines on both sides. Re-indents the replacement to fit.
+// Returns { ok, content } or { ok:false, error:'notfound'|'multiple'|'empty' }.
+function applySearchReplace(fileContent, searchBlock, replaceBlock) {
+  // Matching works on LF text; a CRLF file is written back as CRLF so an edit
+  // never rewrites every line ending in the file (a whole-file diff, and a
+  // mixed-ending file once a later write adds LF lines).
+  const res = applySearchReplaceLF(fileContent, searchBlock, replaceBlock);
+  if (res.ok && /\r\n/.test(fileContent || '')) res.content = res.content.replace(/\r?\n/g, '\r\n');
+  return res;
+}
+
+function applySearchReplaceLF(fileContent, searchBlock, replaceBlock) {
+  const nl = (s) => (s || '').replace(/\r\n/g, '\n');
+  fileContent = nl(fileContent); searchBlock = nl(searchBlock); replaceBlock = nl(replaceBlock);
+  if (!searchBlock.trim()) return { ok: false, error: 'empty' };
+
+  const fileLines = fileContent.split('\n');
+  const searchLines = searchBlock.split('\n');
+
+  const lineStartOffset = (idx) => { let o = 0; for (let k = 0; k < idx; k++) o += fileLines[k].length + 1; return o; };
+
+  const reindentBlock = (block, indent) => {
+    const ls = block.split('\n');
+    const firstNonEmpty = ls.find(l => l.trim() !== '') || '';
+    const base = (firstNonEmpty.match(/^[ \t]*/) || [''])[0];
+    return ls.map(l => {
+      if (l.trim() === '') return '';
+      const stripped = l.startsWith(base) ? l.slice(base.length) : l.replace(/^[ \t]*/, '');
+      return indent + stripped;
+    }).join('\n');
+  };
+
+  const locate = (normalize) => {
+    const s = searchLines.map(normalize).filter(x => x !== '');
+    if (!s.length) return [];
+    const found = [];
+    for (let start = 0; start < fileLines.length; start++) {
+      if (normalize(fileLines[start]) !== s[0]) continue;
+      let fi = start, si = 0, end = start;
+      while (si < s.length && fi < fileLines.length) {
+        const nf = normalize(fileLines[fi]);
+        if (nf === '') { fi++; continue; }       // skip blank lines in the file
+        if (nf !== s[si]) break;
+        end = fi; fi++; si++;
+      }
+      if (si === s.length) found.push([start, end]);
+    }
+    return found;
+  };
+
+  const exact = fileContent.indexOf(searchBlock);
+  if (exact !== -1) {
+    if (exact !== fileContent.lastIndexOf(searchBlock)) return { ok: false, error: 'multiple' };
+    return { ok: true, content: fileContent.slice(0, exact) + replaceBlock + fileContent.slice(exact + searchBlock.length) };
+  }
+
+  let matches = locate(l => l.trim());
+  if (matches.length === 0) matches = locate(l => l.trim().replace(/\s+/g, ' '));
+
+  if (matches.length === 0) {
+    // Last resort: anchor on the first AND last non-blank lines of the search
+    // block (recovers when the AI dropped/added a line in the middle). Kept
+    // conservative - both anchors must be unique and the span must be modest.
+    const sNon = searchLines.map(l => l.trim()).filter(x => x !== '');
+    if (sNon.length >= 2) {
+      const firstHits = [], lastHits = [];
+      fileLines.forEach((l, i) => {
+        const t = l.trim();
+        if (t === sNon[0]) firstHits.push(i);
+        if (t === sNon[sNon.length - 1]) lastHits.push(i);
+      });
+      if (firstHits.length === 1 && lastHits.length === 1 && lastHits[0] >= firstHits[0]) {
+        const sL = firstHits[0], eL = lastHits[0];
+        if (eL - sL + 1 <= sNon.length * 3 + 5) {   // don't grab a huge region
+          const indent = (fileLines[sL].match(/^[ \t]*/) || [''])[0];
+          const reindented = reindentBlock(replaceBlock, indent);
+          return { ok: true, content: fileContent.slice(0, lineStartOffset(sL)) + reindented + fileContent.slice(lineStartOffset(eL) + fileLines[eL].length) };
+        }
+      }
+    }
+    const fuzzy = fuzzyLocate(fileContent, searchBlock);
+    if (fuzzy.ok) {
+      const at = fileContent.indexOf(fuzzy.span);
+      // The model's replacement carries the indentation it wrongly assumed;
+      // move it to where the block really sits.
+      const text = fuzzy.reindent
+        ? reindentBlock(replaceBlock, (fuzzy.span.split('\n').find((l) => l.trim()) || '').match(/^[ \t]*/)[0])
+        : replaceBlock;
+      return { ok: true, content: fileContent.slice(0, at) + text + fileContent.slice(at + fuzzy.span.length), fuzzy: true };
+    }
+    return { ok: false, error: fuzzy.error === 'multiple' ? 'multiple' : 'notfound' };
+  }
+  if (matches.length > 1) return { ok: false, error: 'multiple' };
+
+  const [startLine, endLine] = matches[0];
+  const indent = (fileLines[startLine].match(/^[ \t]*/) || [''])[0];
+  const reindented = reindentBlock(replaceBlock, indent);
+  const startChar = lineStartOffset(startLine);
+  const endChar = lineStartOffset(endLine) + fileLines[endLine].length;
+  return { ok: true, content: fileContent.slice(0, startChar) + reindented + fileContent.slice(endChar) };
+}
+
+// ─── Fallback matchers ──────────────────────────────────────────────────────
+//
+// Adapted from opencode (packages/opencode/src/tool/edit.ts), MIT License,
+// Copyright (c) 2025 opencode - https://github.com/sst/opencode
+//
+// Each matcher yields candidate spans that exist VERBATIM in the file for a
+// search block the model copied imperfectly (escaped quotes, shifted
+// indentation, a middle line reworded). They only run after the matcher above
+// has found nothing, and a candidate is used only when it occurs exactly once
+// and is not far larger than what the model asked for.
+
+function levenshtein(a, b) {
+  if (a === '' || b === '') return Math.max(a.length, b.length);
+  let prev = Array.from({ length: b.length + 1 }, (_, j) => j);
+  for (let i = 1; i <= a.length; i++) {
+    const cur = [i];
+    for (let j = 1; j <= b.length; j++) {
+      cur[j] = Math.min(prev[j] + 1, cur[j - 1] + 1, prev[j - 1] + (a[i - 1] === b[j - 1] ? 0 : 1));
+    }
+    prev = cur;
+  }
+  return prev[b.length];
+}
+
+function spanOfLines(lines, start, end) {
+  return lines.slice(start, end + 1).join('\n');
+}
+
+const ANCHOR_SIMILARITY = 0.65;
+
+// First and last lines match exactly (trimmed); the middle only has to be
+// similar. Recovers a search block with one line misremembered.
+function* blockAnchorMatches(content, find) {
+  const lines = content.split('\n');
+  const want = find.split('\n');
+  if (want[want.length - 1] === '') want.pop();
+  if (want.length < 3) return;
+  const first = want[0].trim(), last = want[want.length - 1].trim();
+  const maxDelta = Math.max(1, Math.floor(want.length * 0.25));
+
+  const candidates = [];
+  for (let i = 0; i < lines.length; i++) {
+    if (lines[i].trim() !== first) continue;
+    for (let j = i + 2; j < lines.length; j++) {
+      if (lines[j].trim() !== last) continue;
+      if (Math.abs(j - i + 1 - want.length) <= maxDelta) candidates.push([i, j]);
+      break;
+    }
+  }
+
+  let best = null, bestScore = -1;
+  for (const [s, e] of candidates) {
+    const size = e - s + 1;
+    const middle = Math.min(want.length - 2, size - 2);
+    let score = 1;
+    if (middle > 0) {
+      score = 0;
+      for (let k = 1; k < want.length - 1 && k < size - 1; k++) {
+        const a = lines[s + k].trim(), b = want[k].trim();
+        const len = Math.max(a.length, b.length);
+        score += len ? 1 - levenshtein(a, b) / len : 1;
+      }
+      score /= middle;
+    }
+    if (score > bestScore) { bestScore = score; best = [s, e]; }
+  }
+  if (best && bestScore >= ANCHOR_SIMILARITY) yield spanOfLines(lines, best[0], best[1]);
+}
+
+// Same block with its indentation shifted as a whole.
+function* indentationFlexibleMatches(content, find) {
+  const dedent = (text) => {
+    const ls = text.split('\n');
+    const indents = ls.filter((l) => l.trim()).map((l) => l.match(/^\s*/)[0].length);
+    if (!indents.length) return text;
+    const min = Math.min(...indents);
+    return ls.map((l) => (l.trim() ? l.slice(min) : l)).join('\n');
+  };
+  const target = dedent(find);
+  const lines = content.split('\n');
+  const n = find.split('\n').length;
+  for (let i = 0; i + n <= lines.length; i++) {
+    const block = lines.slice(i, i + n).join('\n');
+    if (dedent(block) === target) yield block;
+  }
+}
+
+// The model wrote \n, \" or \` escapes where the file has the real characters.
+function* escapeNormalizedMatches(content, find) {
+  const unescape = (s) => s.replace(/\\(n|t|r|'|"|`|\\|\n|\$)/g, (m, c) =>
+    ({ n: '\n', t: '\t', r: '\r', "'": "'", '"': '"', '`': '`', '\\': '\\', '\n': '\n', $: '$' }[c] ?? m));
+  const target = unescape(find);
+  if (target !== find && content.includes(target)) yield target;
+  const lines = content.split('\n');
+  const n = target.split('\n').length;
+  for (let i = 0; i + n <= lines.length; i++) {
+    const block = lines.slice(i, i + n).join('\n');
+    if (block !== target && unescape(block) === target) yield block;
+  }
+}
+
+// Stray blank lines or spaces around the block.
+function* trimmedBoundaryMatches(content, find) {
+  const t = find.trim();
+  if (t === find) return;
+  if (content.includes(t)) yield t;
+}
+
+const FALLBACK_MATCHERS = [indentationFlexibleMatches, escapeNormalizedMatches, trimmedBoundaryMatches, blockAnchorMatches];
+
+// A "match" many times bigger than the search block means an anchor latched
+// onto the wrong closing line; replacing it would silently delete code.
+function isDisproportionate(span, find) {
+  const spanLines = span.split('\n').length, findLines = find.split('\n').length;
+  if (spanLines >= Math.max(findLines + 3, findLines * 2)) return true;
+  if (findLines === 1) return false;
+  return span.trim().length > Math.max(find.trim().length + 500, find.trim().length * 4);
+}
+
+/**
+ * Find the one verbatim span in `content` that an imperfect search block most
+ * likely meant. @returns {{ok:true, span:string} | {ok:false, error:'notfound'|'multiple'|'disproportionate'}}
+ */
+function fuzzyLocate(content, find) {
+  let sawMultiple = false;
+  for (const matcher of FALLBACK_MATCHERS) {
+    for (const span of matcher(content, find)) {
+      if (!span.trim()) continue;
+      const at = content.indexOf(span);
+      if (at === -1) continue;
+      if (at !== content.lastIndexOf(span)) { sawMultiple = true; continue; }
+      if (isDisproportionate(span, find)) return { ok: false, error: 'disproportionate' };
+      return { ok: true, span, reindent: matcher === indentationFlexibleMatches || matcher === blockAnchorMatches };
+    }
+  }
+  return { ok: false, error: sawMultiple ? 'multiple' : 'notfound' };
+}
+
+// Apply an ordered list of verified {search, replace} hunks to file content,
+// all-or-nothing (pure - no disk write).
+function applyEditsToContent(fileContent, edits) {
+  let working = (fileContent || '').replace(/\r\n/g, '\n');
+  for (let i = 0; i < edits.length; i++) {
+    const { search, replace } = edits[i];
+    const res = applySearchReplace(working, search, replace);
+    if (!res.ok) return { ok: false, error: res.error, index: i };
+    working = res.content;
+  }
+  return { ok: true, content: working };
+}
+
+const SYSTEM_PROMPT = `You are a precise code editor. You receive an INSTRUCTION describing a small change, and the FULL current file. Make ONLY the specific change requested, as surgical SEARCH/REPLACE edits - do NOT reproduce the whole file.
+
+STRICT RULES:
+1. Return an "edits" array. Use the FEWEST edits that cleanly express the change.
+2. "search" must be copied VERBATIM from the file below - exact existing text being changed, with just enough surrounding lines to be unique.
+3. "replace" is that same block with ONLY the requested change applied - preserve everything else in it exactly (formatting, unrelated properties, comments, whitespace style).
+4. If the SAME property/selector/value appears more than once in the file (e.g. "top" set in both a "from" and a "to" block), do NOT pick the first or most obvious match - use every clue in the instruction to find the ONE correct occurrence: the CURRENT value it mentions, nearby selectors, rule/keyframe names, or ordering (first/last, start/end). If more than one occurrence still fits equally well after that, return an EMPTY "edits" array with a "reason" naming the ambiguity instead of guessing.
+5. If the instruction genuinely requires large new content that can't be expressed as a small edit (e.g. "build a whole new page/game/module from scratch"), return an EMPTY "edits" array - do not guess badly at a huge diff.
+6. Return ONLY valid JSON, no markdown, no commentary.
+
+Response format:
+{
+  "edits": [ { "search": "<exact verbatim block from the file>", "replace": "<that block with the change applied>" } ],
+  "reason": "<one short sentence>",
+  "confidence": <0-100>
+}`;
+
+/**
+ * Asks the AI for surgical edits, self-corrects up to twice against bad
+ * "search" matches, and returns only the edits that verify cleanly against
+ * the real file content.
+ * @returns {Promise<{success:boolean, edits?:Array, badCount?:number, reason?:string, confidence?:number, tokensUsed?:number, modelUsed?:string, error?:string}>}
+ */
+async function computeInstructionEdits(instruction, filePath) {
+  if (!fs.existsSync(filePath)) return { success: false, error: `File not found: ${filePath}` };
+  const content = fs.readFileSync(filePath, 'utf8');
+
+  const probe = (search) => {
+    const res = applySearchReplace(content, search, search);
+    return res.ok ? 'ok' : res.error;
+  };
+
+  let totalTokens = 0;
+  let lastModelUsed = null;
+
+  const askModel = async (feedback) => {
+    const messages = [
+      { role: 'system', content: SYSTEM_PROMPT },
+      // Large/stable content (the file) first, small/always-different content
+      // (the instruction) last - the ai-proxy's underlying model caches a
+      // repeated PREFIX, so this ordering matters for repeated calls against
+      // the same file even when the instruction text differs each time.
+      { role: 'user', content: `FULL FILE (${path.basename(filePath)}):\n${content}\n\nINSTRUCTION:\n${instruction}` },
+    ];
+    if (feedback) messages.push({ role: 'user', content: feedback });
+
+    const aiResult = await ai.chat(messages, { json: true, meta: { promptText: instruction, filePath } });
+    if (!aiResult.success) throw new Error(aiResult.error);
+    if (aiResult.modelUsed) lastModelUsed = aiResult.modelUsed;
+    const data = aiResult.data;
+    totalTokens += (data.usage || {}).total_tokens || 0;
+
+    let parsed;
+    try { parsed = JSON.parse(data.choices[0].message.content); }
+    catch { return { edits: [], confidence: 0, reason: '' }; }
+
+    const edits = Array.isArray(parsed.edits)
+      ? parsed.edits.filter(e => e && e.search && e.replace != null).map(e => ({ search: e.search, replace: e.replace }))
+      : [];
+    return { edits, confidence: parsed.confidence == null ? 75 : parsed.confidence, reason: parsed.reason || '' };
+  };
+
+  let edits, confidence, reason;
+  try {
+    ({ edits, confidence, reason } = await askModel());
+  } catch (e) {
+    return { success: false, error: e.message };
+  }
+
+  const verify = (list) => list.map((e) => ({ e, status: probe(e.search) }));
+  let checked = verify(edits);
+  let bad = checked.filter(c => c.status !== 'ok');
+
+  let pass = 0;
+  while (bad.length && pass < 2) {
+    pass++;
+    const feedback =
+      `Some "search" blocks from your last answer are wrong. Return the COMPLETE corrected edits array again, fixing these:\n` +
+      bad.map(c => c.status === 'multiple'
+        ? `- Matched MULTIPLE places - include more surrounding lines so it is unique:\n${c.e.search}`
+        : `- NOT found in the file - copy it EXACTLY from the FULL FILE above, character for character:\n${c.e.search}`
+      ).join('\n') +
+      `\nEvery "search" must be copied verbatim from the file shown above.`;
+    let retry;
+    try { retry = await askModel(feedback); }
+    catch (e) { break; }
+    if (retry.edits.length) { edits = retry.edits; confidence = retry.confidence; reason = retry.reason || reason; }
+    checked = verify(edits);
+    bad = checked.filter(c => c.status !== 'ok');
+  }
+
+  const good = checked.filter(c => c.status === 'ok').map(c => c.e);
+  return { success: true, edits: good, badCount: bad.length, confidence, reason, tokensUsed: totalTokens, modelUsed: lastModelUsed };
+}
+
+module.exports = { applySearchReplace, applyEditsToContent, computeInstructionEdits, fuzzyLocate };