// 工具执行器(内置工具)。所有文件路径强制解析到任务目录内(防路径穿越)。 // 工具名与参数 schema 来自 definitions.json;MCP 工具不在这里,见 ../mcp.js。 const fs = require('fs'); const path = require('path'); const fsp = require('fs').promises; const { spawn } = require('child_process'); const { Minimatch } = require('minimatch'); const mcp = require('../mcp'); const skills = require('./skills'); const rexec = require('../remote-exec'); function makeMatcher(pattern, opts) { const mm = new Minimatch(pattern, opts || {}); return function (p) { return mm.match(p); }; } const IGNORED_DIRS = new Set(['.git', 'node_modules', '__pycache__', '.venv', 'venv', 'dist', 'build', '.next', '.cache']); // ---------- 路径沙箱 ---------- function resolveInTask(taskDir, p) { if (typeof p !== 'string' || !p.trim()) throw new Error('路径不能为空'); let abs; if (path.isAbsolute(p)) { abs = path.normalize(p); } else { abs = path.resolve(taskDir, p); } const root = path.resolve(taskDir); if (abs !== root && !abs.startsWith(root + path.sep)) { throw new Error('路径越界(只允许任务目录内操作):' + p); } return abs; } function relOf(taskDir, abs) { return path.relative(path.resolve(taskDir), abs) || '.'; } // ---------- 文件系统遍历 ---------- function walk(taskDir, opts) { const out = []; const maxFiles = (opts && opts.maxFiles) || 20000; function rec(dir) { let entries; try { entries = fs.readdirSync(dir, { withFileTypes: true }); } catch (_) { return; } for (const ent of entries) { if (out.length >= maxFiles) return; if (ent.name.startsWith('.') && IGNORED_DIRS.has(ent.name)) continue; if (IGNORED_DIRS.has(ent.name)) continue; const abs = path.join(dir, ent.name); if (ent.isDirectory()) { rec(abs); } else if (ent.isFile()) { out.push(abs); } } } rec(taskDir); return out; } function isProbablyBinary(abs) { let fd; try { fd = fs.openSync(abs, 'r'); const buf = Buffer.alloc(8192); const bytes = fs.readSync(fd, buf, 0, buf.length, null); if (bytes === 0) return false; const slice = buf.slice(0, bytes); return slice.includes(0); } catch (_) { return true; } finally { if (fd !== undefined) try { fs.closeSync(fd); } catch (_) {} } } function truncate(text, max) { if (text.length <= max) return text; return text.slice(0, max) + '\n...(输出已截断,共 ' + text.length + ' 字符)'; } // ---------- 各工具实现 ---------- // 对齐 Zed:整读时文件字节数超过该阈值则返回大纲(outline),而非截断内容。 const AUTO_OUTLINE_SIZE = 16384; const IMAGE_EXTS = new Set(['.png', '.jpg', '.jpeg', '.webp', '.gif', '.bmp', '.tiff', '.tif']); // read_file 读图(仅当会话绑定的模型 vision: true):sharp 压缩后转 base64 data URL, // 由 engine 拼成下一轮请求里 tool 消息的 image_url 内容块。 const IMG_MAX_INPUT_BYTES = 50 * 1024 * 1024; // 与上传上限一致 const IMG_MAX_SIDE = 2000; // 长边压缩阈值(px) const IMG_JPEG_QUALITY = 80; // 简化版符号大纲(Zed 用 tree-sitter,这里用正则提取常见语言的顶层符号 + Markdown 标题)。 // 返回 "# File outline ..." 文本;无法提取时返回 null(调用方回退到首 1KB)。 const OUTLINE_RES = [ /^\s*(?:export\s+default\s+|export\s+|default\s+|pub(?:\([^)]*\))?\s+)?(?:pub\s+)?(?:async\s+)?(?:function|fn|func|def|sub)\s+\*?([A-Za-z0-9_]+)/, /^\s*(?:export\s+)?(?:default\s+)?(?:pub(?:\([^)]*\))?\s+)?(?:pub\s+)?(?:abstract\s+|final\s+|open\s+|sealed\s+|data\s+|implicit\s+)*(?:class|struct|trait|interface|enum|impl|namespace|module|type|record|union)\s+([A-Za-z0-9_<>,\[\] ]{1,60}?)(?:\s*[{:<(]|\s*$)/, /^\s*(?:export\s+)?(?:const|let|var)\s+([A-Za-z0-9_]+)\s*(?::[^=]+)?=\s*(?:async\s*)?(?:function\b|\([^)]*\)\s*=>|[A-Za-z0-9_]+\s*=>)/, ]; function buildOutline(absPath, text) { const ext = path.extname(absPath).toLowerCase(); const lines = text.split('\n'); const items = []; if (ext === '.md' || ext === '.markdown' || ext === '.rst') { for (let i = 0; i < lines.length; i++) { const m = /^(#{1,6})\s+(.*)$/.exec(lines[i]); if (m) items.push({ depth: m[1].length - 1, name: m[2].trim().slice(0, 80), line: i + 1 }); } } else { for (let i = 0; i < lines.length; i++) { for (const re of OUTLINE_RES) { const m = re.exec(lines[i]); if (m) { const indent = (/^[ \t]*/.exec(lines[i])[0] || '').length; items.push({ depth: Math.min(8, Math.floor(indent / 2)), name: m[1].trim(), line: i + 1 }); break; } } if (items.length >= 500) break; } } if (!items.length) return null; const body = items.map(it => ' '.repeat(it.depth) + it.name + ' [L' + it.line + ']').join('\n'); return '# File outline for ' + absPath + '\n\n' + body + '\n\nShowing symbols 1-' + items.length + ' (total symbols: ' + items.length + ')\n'; } // Zed 同款的大纲包裹文案(模型对该格式有既成预期)。 function outlineWrapper(outlineText) { return 'SUCCESS: File outline retrieved. This file is too large to read all at once, so the outline below shows the file\'s structure with line numbers.\n\n' + 'IMPORTANT: Do NOT retry this call without line numbers - you will get the same outline.\n' + 'Instead, use the line numbers below to read specific sections by calling this tool again with start_line and end_line parameters.\n\n' + outlineText + '\n' + 'NEXT STEPS: To read a specific symbol\'s implementation, call read_file with the same path plus start_line and end_line from the outline above.\n' + 'For example, to read a function shown as [L100-150], use start_line: 100 and end_line: 150.'; } // 用 sharp 压缩(按 EXIF 摆正、长边 ≤2000px、JPEG 质量 80)并转 base64 data URL。 // 成功返回 {text, image_data_url},engine 侧拼成 tool 消息的多模态内容块; // sharp 缺失或解码失败时退回纯文本提示,不让整个工具回合报错。 async function readImageAsBlocks(abs, rel, size) { let sharp; try { sharp = require('sharp'); } catch (_) { return '[图片文件] ' + rel + '(' + size + ' 字节)。服务端未安装 sharp,无法压缩回传图片内容。'; } try { const r = await sharp(abs) .rotate() .resize({ width: IMG_MAX_SIDE, height: IMG_MAX_SIDE, fit: 'inside', withoutEnlargement: true }) .jpeg({ quality: IMG_JPEG_QUALITY }) .toBuffer({ resolveWithObject: true }); return { text: '[图片文件] ' + rel + '(原 ' + size + ' 字节,已压缩为 ' + r.info.width + '×' + r.info.height + ' JPEG ' + r.data.length + ' 字节)', image_data_url: 'data:image/jpeg;base64,' + r.data.toString('base64'), }; } catch (err) { return '[图片文件] ' + rel + '(' + size + ' 字节),图片解码/压缩失败:' + (err && err.message || err); } } async function toolReadFile(ctx, args) { const abs = resolveInTask(ctx.taskDir, args.path); let st; try { st = await fsp.stat(abs); } catch (_) { throw new Error('文件不存在:' + args.path); } if (st.isDirectory()) throw new Error(args.path + ' 是目录不是文件,请用 list_directory 工具查看目录内容。'); if (IMAGE_EXTS.has(path.extname(abs).toLowerCase())) { const rel = relOf(ctx.taskDir, abs); if (!ctx.vision) { return '[图片文件] ' + rel + '(' + st.size + ' 字节)。图片文件,当前模型不支持视觉识别,无法回传图片内容。'; } if (st.size > IMG_MAX_INPUT_BYTES) { return '[图片文件] ' + rel + '(' + st.size + ' 字节)过大(上限 ' + IMG_MAX_INPUT_BYTES + ' 字节),已跳过回传。'; } return await readImageAsBlocks(abs, rel, st.size); } if (isProbablyBinary(abs)) return '[二进制文件,跳过文本读取] ' + relOf(ctx.taskDir, abs); const text = await fsp.readFile(abs, 'utf8'); let lines = text.split('\n'); const total = lines.length; const startLine = Math.max(1, args.start_line || 1); let endLine = args.end_line || total; if (startLine > total) return '(文件共 ' + total + ' 行,start_line 超出范围)'; // 对齐 Zed:未指定范围且文件超过 16KB 时返回大纲;提取不到大纲则回退回首 1KB。 if (!args.start_line && !args.end_line && Buffer.byteLength(text, 'utf8') > AUTO_OUTLINE_SIZE) { const outline = buildOutline(abs, text); if (outline) return outlineWrapper(outline); let head = text.slice(0, 1024); try { while (Buffer.byteLength(head, 'utf8') > 1024) head = head.slice(0, -1); } catch (_) {} return outlineWrapper('# First 1KB of ' + abs + ' (file too large to show full content, and no outline available)\n\n' + head); } lines = lines.slice(startLine - 1, Math.min(endLine, total)); return lines.map((l, i) => String(startLine + i).padStart(6, ' ') + '\t' + l).join('\n'); } async function toolWriteFile(ctx, args) { const abs = resolveInTask(ctx.taskDir, args.path); await fsp.mkdir(path.dirname(abs), { recursive: true }); await fsp.writeFile(abs, args.content ?? '', 'utf8'); recordFile(ctx, abs); return '已成功写入文件 ' + relOf(ctx.taskDir, abs) + '(' + Buffer.byteLength(args.content || '') + ' 字节)'; } // ---------- edit_file 匹配(对齐 Zed edit_session/streaming_fuzzy_matcher 的终止态语义) ---------- // Zed 常量:替换代价 1,插入(buffer 多余行)3,删除(query 行未命中)10;模糊相似阈值 0.8。 const COST_REP = 1, COST_INS = 3, COST_DEL = 10, FUZZY_THRESHOLD = 0.8; const FUZZY_DP_CELL_LIMIT = 8_000_000; // 超大文件跳过 DP,退回逐行 trim 扫描 // normalized Levenshtein 相似度(1 - lev/max_len),先做长度粗筛。 function lineSimilarity(a, b) { if (a === b) return 1; const la = a.length, lb = b.length, max = Math.max(la, lb); if (max === 0) return 1; if (1 - Math.abs(la - lb) / max < FUZZY_THRESHOLD) return 0; let prev = new Int32Array(lb + 1), curr = new Int32Array(lb + 1); for (let j = 0; j <= lb; j++) prev[j] = j; for (let i = 1; i <= la; i++) { curr[0] = i; const ca = a.charCodeAt(i - 1); for (let j = 1; j <= lb; j++) { const cost = ca === b.charCodeAt(j - 1) ? 0 : 1; curr[j] = Math.min(prev[j] + 1, curr[j - 1] + 1, prev[j - 1] + cost); } const t = prev; prev = curr; curr = t; } return 1 - prev[lb] / max; } function diagCost(trimOld, trimCur) { if (trimOld === trimCur) return 0; if (lineSimilarity(trimOld, trimCur) >= FUZZY_THRESHOLD) return COST_REP; return COST_DEL + COST_INS; } function lineNumberOf(text, idx) { let n = 1; for (let i = 0; i < idx; i++) if (text.charCodeAt(i) === 10) n++; return n; } // 全局 DP 对齐(query 行 × buffer 行,逐行 trim 比较)。 // 返回 { startRow, endRow };找不到返回 null;多个同分候选返回 { ambiguous: [rows] }。 function fuzzyLocate(curLines, oldLinesTrim) { const m = oldLinesTrim.length, n = curLines.length; const curTrim = new Array(n); for (let c = 0; c < n; c++) curTrim[c] = curLines[c].trim(); const W = n + 1; const INF = Number.POSITIVE_INFINITY; let dpPrev = new Float64Array(W), dpCur = new Float64Array(W); const dir = new Uint8Array((m + 1) * W); // bit1=diag bit2=up(删 query 行) bit4=left(跳过 buffer 行) for (let c = 0; c < W; c++) { dpPrev[c] = c * COST_INS; dir[c] = 4; } for (let r = 1; r <= m; r++) { dpCur[0] = dpPrev[0] + COST_DEL; dir[r * W] = 2; const qt = oldLinesTrim[r - 1]; for (let c = 1; c <= n; c++) { const d = dpPrev[c - 1] + diagCost(qt, curTrim[c - 1]); const u = dpPrev[c] + COST_DEL; const l = dpCur[c - 1] + COST_INS; let best = d, bits = 1; if (u < best) { best = u; bits = 2; } if (l < best) { best = l; bits = 4; } dpCur[c] = best; dir[r * W + c] = bits; } const t = dpPrev; dpPrev = dpCur; dpCur = t; } // 末行最小代价的所有列 → 逐列回溯成候选,按 (startRow,endRow) 去重。 let min = INF; for (let c = 1; c <= n; c++) if (dpPrev[c] < min) min = dpPrev[c]; if (!isFinite(min)) return null; const seen = new Set(); const candidates = []; for (let c0 = 1; c0 <= n; c0++) { if (dpPrev[c0] !== min) continue; let r = m, c = c0; const pairs = []; while (r > 0 && c > 0) { const bits = dir[r * W + c]; if (bits & 1) { if (diagCost(oldLinesTrim[r - 1], curTrim[c - 1]) <= COST_REP) pairs.push(c - 1); r--; c--; } else if (bits & 2) { r--; } else { c--; } } if (!pairs.length) continue; // 回溯是从下往上收集 pairs(降序),用 min/max 取区间,不依赖收集顺序。 const startRow = Math.min(...pairs), endRow = Math.max(...pairs); const span = endRow - startRow + 1; const ratio = pairs.length / Math.max(span, m); if (ratio < FUZZY_THRESHOLD) continue; const key = startRow + ':' + endRow; if (seen.has(key)) continue; seen.add(key); candidates.push({ startRow, endRow }); } if (!candidates.length) return null; if (candidates.length > 1) return { ambiguous: candidates.map(x => x.startRow + 1) }; return candidates[0]; } function indentOf(l) { return /^[ \t]*/.exec(l)[0]; } // 按“缓冲行缩进 - query 行缩进”的差值重排 new_text 缩进(空白行不动)。 function applyReindent(newLines, delta, ch) { if (delta === 0) return newLines; return newLines.map(l => { if (!l.trim()) return l; const ind = indentOf(l); const n = Math.max(0, ind.length + delta); if (n <= ind.length) return ' '.repeat(n) + l.slice(ind.length); return ind + ch.repeat(n - ind.length) + l.slice(ind.length); }); } // 应用单条 edit;失败抛错(含 edit 序号)。对齐 Zed:先精确(含重叠计数,≥2 歧义), // 再模糊;匹配区间不含行尾换行;new_text 末尾换行仅当缓冲区该位置后有 \n 时吃掉一个。 function applyEdit(text, edit, i) { const tag = 'edit #' + (i + 1); const old_text = edit.old_text, new_text = edit.new_text == null ? '' : String(edit.new_text); if (!old_text) throw new Error(tag + ': old_text 不能为空'); // 尾部行尾换行不计入匹配区间(Zed trailing_line_ending_len)。 let pat = String(old_text); if (pat.endsWith('\r\n')) pat = pat.slice(0, -2); else if (pat.endsWith('\n')) pat = pat.slice(0, -1); if (!pat) pat = String(old_text); // 1) 精确匹配(含重叠) const hits = []; let from = 0; while (hits.length < 2) { const idx = text.indexOf(pat, from); if (idx < 0) break; hits.push(idx); from = idx + 1; } if (hits.length >= 2) { throw new Error(tag + ': old_text 在文件中匹配到多处(第 ' + hits.map(h => lineNumberOf(text, h)).join(', ') + ' 行附近)。请在 old_text 中包含更多上下文以唯一定位。'); } const newLinesRaw = new_text.split('\n'); if (hits.length === 1) { // 精确匹配:原样替换,不重排缩进(Zed first_line_delta=0,精确下 rest 差值必然 0)。 const idx = hits[0]; let repl = new_text; // 尾换行摘除:匹配区间后紧跟 \n(或 CRLF 的 \r\n)时吃掉 new_text 的一个尾换行。 const after = text.charAt(idx + pat.length); const afterCRLF = after === '\r' && text.charAt(idx + pat.length + 1) === '\n'; if ((after === '\n' || afterCRLF) && repl.endsWith('\n')) repl = repl.slice(0, -1); // CRLF 文件:新内容里的行尾统一成 \r\n。 if (text.includes('\r\n')) repl = repl.replace(/\r?\n/g, '\r\n'); return text.slice(0, idx) + repl + text.slice(idx + pat.length); } // 2) 模糊匹配(逐行 trim) const curLines = text.split('\n'); const oldLines = pat.split('\n'); const oldTrim = oldLines.map(l => l.trim()); let loc = null; if (oldLines.length * curLines.length <= FUZZY_DP_CELL_LIMIT) { loc = fuzzyLocate(curLines, oldTrim); } else { // 超大文件降级:逐行 trim 首次命中(旧行为)。 for (let s = 0; s + oldLines.length <= curLines.length; s++) { let ok2 = true; for (let j = 0; j < oldLines.length; j++) { if (curLines[s + j].trim() !== oldTrim[j]) { ok2 = false; break; } } if (ok2) { loc = { startRow: s, endRow: s + oldLines.length - 1 }; break; } } } if (!loc) { throw new Error(tag + ': 在文件中找不到 old_text(可能已被之前的 edit 修改)。请重新 read_file 获取最新内容后再试。'); } if (loc.ambiguous) { throw new Error(tag + ': old_text 模糊匹配到多处(第 ' + loc.ambiguous.join(', ') + ' 行附近)。请在 old_text 中包含更多上下文以唯一定位。'); } const { startRow, endRow } = loc; let newLines = newLinesRaw; // 匹配区间后紧跟换行时,吃掉 new_text 的尾换行(防止吞掉行尾)。 const hasNewlineAfter = endRow < curLines.length - 1 || text.endsWith('\n'); if (hasNewlineAfter && newLines.length > 1 && newLines[newLines.length - 1] === '') { newLines = newLines.slice(0, -1); } // 缩进差值重排(以首条匹配行为基准)。 const bufInd = indentOf(curLines[startRow]); const qInd = indentOf(oldLines[0]); const delta = bufInd.length - qInd.length; if (delta !== 0) newLines = applyReindent(newLines, delta, bufInd.includes('\t') ? '\t' : ' '); // CRLF 文件:新增行补齐 \r。 if (text.includes('\r\n')) newLines = newLines.map(l => (l.endsWith('\r') || l === '' ? l : l + '\r')); curLines.splice(startRow, endRow - startRow + 1, ...newLines); return curLines.join('\n'); } // 对齐 Zed:edits 顺序应用到内存态;若第 N 条失败,前 N-1 条仍保存落盘, // 错误信息告知模型已应用了多少、需重新 read_file。 async function toolEditFile(ctx, args) { const abs = resolveInTask(ctx.taskDir, args.path); let st; try { st = await fsp.stat(abs); } catch (_) { throw new Error('文件不存在:' + args.path); } if (st.isDirectory()) throw new Error('无法编辑:路径是目录 ' + args.path); const edits = args.edits || []; let text = await fsp.readFile(abs, 'utf8'); const original = text; let applied = 0; let failure = null; for (let i = 0; i < edits.length; i++) { try { text = applyEdit(text, edits[i], i); applied++; } catch (err) { failure = err; break; } } if (text !== original) { await fsp.writeFile(abs, text, 'utf8'); recordFile(ctx, abs); } if (failure) { throw new Error(failure.message + (applied > 0 ? '(前 ' + applied + ' 处编辑已应用并保存,文件已变化,请重新 read_file 后再继续)' : '(文件未改动)')); } return '已成功应用 ' + applied + ' 处编辑到 ' + relOf(ctx.taskDir, abs); } // 对齐 Zed list_directory 的输出格式: // # Folders: //