773 lines
37 KiB
JavaScript
773 lines
37 KiB
JavaScript
// 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 mcp = require('../mcp');
|
||
const skills = require('./skills');
|
||
const rexec = require('../remote-exec');
|
||
const publicFiles = require('../public-files');
|
||
const scheduler = require('../scheduler'); // 只依赖 log/cron-parser,不成环(engine→executor→scheduler 单向)
|
||
|
||
// ---------- 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) || '.';
|
||
}
|
||
|
||
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 压缩后落盘到任务目录 .imgcache/,
|
||
// 返回 {text, image_rel};engine 拼成 image_ref 内容块落库,llm.buildMessages 发请求前
|
||
// 现签成公网外链(GET /p/<id>)进 image_url 块,base64 不再进 DB/请求体。
|
||
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)并落盘 .imgcache/。
|
||
// 成功返回 {text, image_rel};产物文件名带源文件指纹(mtime+size),同图重复读复用同一份不增殖。
|
||
// 无收益守卫(与浏览器端压缩同款):重编码 JPEG 反而 ≥ 原图时不落盘,image_rel 直指原文件。
|
||
// 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(ctx, 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 });
|
||
if (r.data.length >= size) {
|
||
// 低内容熵图(截图/线稿/纯色)JPEG 重编码可能比原图还大:不产出缓存,外链直接指向原文件
|
||
return {
|
||
text: '[图片文件] ' + rel + '(' + size + ' 字节,压缩无收益,按原图回传)',
|
||
image_rel: rel,
|
||
};
|
||
}
|
||
let st = null;
|
||
try { st = await fsp.stat(abs); } catch (_) {}
|
||
const stem = path.basename(rel).replace(/\.[^.]+$/, '').replace(/[^\w.\-\u4e00-\u9fff]/g, '_').slice(0, 60);
|
||
const cacheRel = '.imgcache/' + (stem || 'img') + '-' + size + '-' + Math.round(st ? st.mtimeMs : 0) + '.jpg';
|
||
const outAbs = path.join(ctx.taskDir, cacheRel);
|
||
if (!fs.existsSync(outAbs)) {
|
||
await fsp.mkdir(path.dirname(outAbs), { recursive: true });
|
||
await fsp.writeFile(outAbs, r.data);
|
||
}
|
||
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_rel: cacheRel,
|
||
};
|
||
} 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 terminal 执行 ls 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(ctx, 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。
|
||
// DP 工作集复用模块级 buffer:lineSimilarity 无递归/无重入(单线程下安全),
|
||
// 避免 fuzzyLocate m×n 内层每格新建两个 typed array(触顶时是千万次临时分配)。
|
||
let simPrev = null, simCurr = null;
|
||
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;
|
||
if (!simPrev || simPrev.length < lb + 1) {
|
||
simPrev = new Int32Array(lb + 1);
|
||
simCurr = new Int32Array(lb + 1);
|
||
}
|
||
let prev = simPrev, curr = simCurr; // 下面的交换只换局部引用,不动模块级变量
|
||
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;
|
||
// 升序排列并要求连续:不连续区间说明 DP 在中间跳过了无关的 buffer 行(left/13 代价 diag 路径),
|
||
// 整段 splice 会静默删掉这些行——这类候选直接丢弃(上层报找不到,模型补上下文重试)。
|
||
pairs.sort((a, b) => a - b);
|
||
let contiguous = true;
|
||
for (let k = 1; k < pairs.length; k++) {
|
||
if (pairs[k] !== pairs[k - 1] + 1) { contiguous = false; break; }
|
||
}
|
||
if (!contiguous) continue;
|
||
const startRow = pairs[0], endRow = pairs[pairs.length - 1];
|
||
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 (afterCRLF) {
|
||
// 尾随 \r\n 要整对切;只切 \n 会在拼接处留下孤立的 \r
|
||
if (repl.endsWith('\r\n')) repl = repl.slice(0, -2);
|
||
else if (repl.endsWith('\n')) repl = repl.slice(0, -1);
|
||
} else if (after === '\n' && 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');
|
||
// 空 new_text 的整行删除:区间恰好行对齐(前邻行首/换行、后随换行)时把尾随行尾一并吃掉,
|
||
// 对齐 Zed 空 new_text=删零行(否则 JS 语义会留下一个空行)。
|
||
if (repl === '' && (after === '\n' || afterCRLF) &&
|
||
(idx === 0 || text.charAt(idx - 1) === '\n' || text.charAt(idx - 1) === '\r')) {
|
||
return text.slice(0, idx) + text.slice(idx + pat.length + (afterCRLF ? 2 : 1));
|
||
}
|
||
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;
|
||
let usedFallback = false;
|
||
if (oldLines.length * curLines.length <= FUZZY_DP_CELL_LIMIT) {
|
||
loc = fuzzyLocate(curLines, oldTrim);
|
||
} else {
|
||
usedFallback = true;
|
||
// 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)。' + (usedFallback
|
||
? '文件过大,本次仅走了降级匹配(逐行 trim 精确相等、无相似度容忍):请核对 old_text 与原文是否逐字一致,或改用 read_file 按行号定位后用 terminal 的 sed 编辑,不要以同一文本反复重试。'
|
||
: '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;
|
||
// 空 new_text=整行删除:对齐 Zed "".lines() 的零行语义(JS ''.split('\n') 是 1 个空行,会留白);
|
||
// new_text='\n' 仍是「替换为一个空行」,语义不合并。
|
||
let newLines = new_text === '' ? [] : 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 select_terminal_output_lines:head/tail 各自取行;同时给时中间空一行;不去重叠;0 合法。
|
||
function selectOutput(text, head, tail) {
|
||
const hasH = typeof head === 'number' && isFinite(head);
|
||
const hasT = typeof tail === 'number' && isFinite(tail);
|
||
if (!hasH && !hasT) return text;
|
||
const lines = text.split('\n');
|
||
const parts = [];
|
||
if (hasH) parts.push(lines.slice(0, Math.max(0, Math.floor(head))).join('\n'));
|
||
if (hasT) parts.push(lines.slice(Math.max(0, lines.length - Math.floor(tail))).join('\n'));
|
||
return parts.join('\n\n');
|
||
}
|
||
|
||
// terminal:spawn + 进程组 kill + 超时 + 输出截断
|
||
async function toolTerminal(ctx, args) {
|
||
const cfg = ctx.cfg;
|
||
const timeout = args.timeout_ms || cfg.terminal.default_timeout_ms;
|
||
const maxChars = cfg.terminal.max_output_chars;
|
||
let cwd = ctx.taskDir;
|
||
if (args.cd) cwd = resolveInTask(ctx.taskDir, args.cd);
|
||
if (!fs.existsSync(cwd)) throw new Error('工作目录不存在:' + cwd);
|
||
|
||
// 配了 executor.url(或 env EXECUTOR_URL):命令交给远端 agent-executor 执行,文本拼回与本地完全一致。
|
||
if (rexec.isRemote(cfg)) {
|
||
const r = await rexec.run(cfg, { command: args.command, cwd, timeout_ms: timeout, run: ctx.run });
|
||
let text = selectOutput(r.out.replace(/\n+$/, ''), args.head_lines, args.tail_lines);
|
||
if (r.killed_by === 'timeout') {
|
||
text += (text ? '\n' : '') + '(命令超时 ' + timeout + 'ms,已终止;这是超时不是命令本身失败)';
|
||
} else if (r.killed_by === 'cancelled') {
|
||
text += (text ? '\n' : '') + '(用户终止任务,进程被 kill)';
|
||
} else {
|
||
text += (text ? '\n' : '') + '[exit code: ' + r.code + ']';
|
||
}
|
||
if (r.truncated) text += '\n(输出超长,已在 executor 侧截断)';
|
||
return truncate(text || '(无输出)', maxChars);
|
||
}
|
||
|
||
return new Promise((resolve) => {
|
||
const child = spawn('/bin/sh', ['-c', args.command], { cwd, detached: true, env: process.env });
|
||
ctx.run.procs.add(child);
|
||
let out = '';
|
||
let killedBy = null; // 'timeout' | 'cancelled'
|
||
let settled = false;
|
||
|
||
const timer = setTimeout(() => {
|
||
killedBy = 'timeout';
|
||
killGroup(child);
|
||
}, timeout);
|
||
|
||
const onData = (d) => { out += d.toString(); };
|
||
child.stdout.on('data', onData);
|
||
child.stderr.on('data', onData);
|
||
|
||
const finish = (code, extra) => {
|
||
if (settled) return;
|
||
settled = true;
|
||
clearTimeout(timer);
|
||
clearInterval(cancelCheck);
|
||
ctx.run.procs.delete(child);
|
||
// head_lines/tail_lines:对齐 Zed,在原始输出上选段(之前被完全忽略)。
|
||
let text = selectOutput(out.replace(/\n+$/, ''), args.head_lines, args.tail_lines);
|
||
if (killedBy === 'timeout') {
|
||
// 超时/终止不是“执行失败”,不报 exit code(避免 -1 误导模型)。
|
||
text += (text ? '\n' : '') + '(命令超时 ' + timeout + 'ms,已终止;这是超时不是命令本身失败)';
|
||
} else if (killedBy === 'cancelled') {
|
||
text += (text ? '\n' : '') + '(用户终止任务,进程被 kill)';
|
||
} else {
|
||
text += (text ? '\n' : '') + '[exit code: ' + code + ']';
|
||
}
|
||
if (extra) text += '\n' + extra;
|
||
resolve(truncate(text || '(无输出)', maxChars));
|
||
};
|
||
child.on('error', err => finish(-1, 'spawn 失败:' + err.message));
|
||
child.on('close', code => finish(killedBy ? -1 : (code === null ? -1 : code)));
|
||
|
||
const cancelCheck = setInterval(() => {
|
||
if (ctx.run.cancelled) { killedBy = 'cancelled'; killGroup(child); }
|
||
}, 200);
|
||
});
|
||
}
|
||
|
||
function killGroup(child) {
|
||
if (!child.pid) return;
|
||
try { process.kill(-child.pid, 'SIGKILL'); } catch (_) {
|
||
try { child.kill('SIGKILL'); } catch (__) {}
|
||
}
|
||
}
|
||
|
||
// fetch:抓取并转 Markdown。超时/取消覆盖到 body 读取完成(非仅响应头);
|
||
// 流式读取至多 cfg.fetch.max_bytes 字节;按响应 charset 解码(gbk 等依赖内置 full-icu)。
|
||
// undici 网络错误的 message 恒为 "fetch failed",真因窝在 err.cause 里——不解包模型分不清
|
||
// 域名已死/被拒连/TLS 问题(评审实测:只能靠另跑 curl 才能诊断)。常见码翻成中文直述。
|
||
const FETCH_NET_ERRORS = {
|
||
ENOTFOUND: '域名无法解析(ENOTFOUND,域名可能已不存在)',
|
||
EAI_AGAIN: 'DNS 暂时不可用(EAI_AGAIN,可稍后重试)',
|
||
ECONNREFUSED: '连接被拒绝(ECONNREFUSED,端口未监听)',
|
||
ECONNRESET: '连接被对端重置(ECONNRESET,多为服务端/中间设备拒服务)',
|
||
ETIMEDOUT: '连接超时(ETIMEDOUT)',
|
||
EHOSTUNREACH: '主机不可达(EHOSTUNREACH)',
|
||
ENETUNREACH: '网络不可达(ENETUNREACH)',
|
||
EPIPE: '管道断裂(EPIPE)',
|
||
EPROTO: '协议错误(EPROTO)',
|
||
};
|
||
function describeFetchError(err) {
|
||
const c = err && err.cause;
|
||
const code = c && c.code ? String(c.code).toUpperCase() : '';
|
||
const msg = (c && c.message) ? String(c.message) : String((err && err.message) || err);
|
||
if (FETCH_NET_ERRORS[code]) return FETCH_NET_ERRORS[code];
|
||
if (/self.?signed|certificate|CERT_|UNABLE_TO_/i.test(code + ' ' + msg)) return 'TLS 证书校验失败:' + msg;
|
||
return msg + (code ? '(' + code + ')' : '');
|
||
}
|
||
|
||
function isTextualContentType(ct) {
|
||
if (!ct) return true; // 未声明:先按文本读,读完用 NUL 字节兜底判定
|
||
if (ct.startsWith('text/')) return true;
|
||
if (/^application\/(?:[\w.-]+\+)?(?:json|xml|yaml|x-ndjson|ndjson)\b/.test(ct)) return true;
|
||
if (/^application\/(?:javascript|ecmascript|xhtml\+xml)\b/.test(ct)) return true;
|
||
return false;
|
||
}
|
||
|
||
function charsetOf(ct) {
|
||
const m = /charset\s*=\s*["']?([\w-]+)/i.exec(ct || '');
|
||
return m ? m[1].toLowerCase() : '';
|
||
}
|
||
|
||
// 流式读到 maxBytes 即停;超限 reader.cancel() 主动断开,不再继续下载。
|
||
async function readBodyLimited(stream, maxBytes) {
|
||
const reader = stream.getReader();
|
||
const chunks = [];
|
||
let total = 0;
|
||
let truncated = false;
|
||
try {
|
||
for (;;) {
|
||
const { done, value } = await reader.read();
|
||
if (done) break;
|
||
if (total + value.byteLength > maxBytes) {
|
||
if (maxBytes > total) chunks.push(value.subarray(0, maxBytes - total));
|
||
truncated = true;
|
||
break;
|
||
}
|
||
chunks.push(value);
|
||
total += value.byteLength;
|
||
}
|
||
} finally {
|
||
if (truncated) { try { await reader.cancel(); } catch (_) {} }
|
||
}
|
||
return { buf: Buffer.concat(chunks.map(c => Buffer.from(c))), truncated };
|
||
}
|
||
|
||
async function toolFetch(ctx, args) {
|
||
if (ctx.run.cancelled) throw new Error('已终止');
|
||
let url = String(args.url || '').trim();
|
||
if (!url) throw new Error('fetch 缺少参数 url');
|
||
// 对齐 Zed normalize_url:缺 scheme 时默认补 https://;非 http(s) scheme 仍拒绝。
|
||
if (!/^https?:\/\//i.test(url)) {
|
||
if (/^[a-z0-9+.-]+:\/\//i.test(url)) throw new Error('只支持 http/https URL:' + url);
|
||
url = 'https://' + url;
|
||
}
|
||
const fcfg = (ctx.cfg && ctx.cfg.fetch) || {};
|
||
const timeoutMs = fcfg.timeout_ms || 30000;
|
||
const maxBytes = fcfg.max_bytes || 1024 * 1024;
|
||
|
||
const ctrl = new AbortController();
|
||
const timer = setTimeout(() => ctrl.abort(), timeoutMs);
|
||
// 与 toolTerminal 同模式:等待中用户取消也能 abort 掉进行中的请求/读流。
|
||
const cancelWatch = setInterval(() => { if (ctx.run.cancelled) ctrl.abort(); }, 200);
|
||
try {
|
||
const res = await fetch(url, {
|
||
signal: ctrl.signal,
|
||
redirect: 'follow',
|
||
headers: { 'User-Agent': 'Mozilla/5.0 (agent-app)' },
|
||
});
|
||
const ct = (res.headers.get('content-type') || '').toLowerCase();
|
||
// 非文本:不读 body,直接回报元信息(避免二进制白下载 + 乱码)。
|
||
if (!isTextualContentType(ct)) {
|
||
const len = Number(res.headers.get('content-length')) || 0;
|
||
return 'HTTP ' + res.status + '\n[非文本响应] content-type=' + ct +
|
||
(len ? ',' + len + ' 字节' : '') + ',未读取内容。';
|
||
}
|
||
if (!res.body) return 'HTTP ' + res.status + '\n(无响应体)';
|
||
|
||
const { buf, truncated } = await readBodyLimited(res.body, maxBytes);
|
||
// content-type 未声明的兜底:前 8KB 含 NUL 判为二进制(与 read_file 的 isProbablyBinary 同口径)。
|
||
if (!ct && buf.subarray(0, 8192).includes(0)) {
|
||
return 'HTTP ' + res.status + '\n[非文本响应] content-type 未声明,body 疑似二进制(' + buf.length + ' 字节),未返回内容。';
|
||
}
|
||
// 按 charset 解码;未知/非法编码名退回 UTF-8,不抛错。
|
||
const charset = charsetOf(ct);
|
||
let text;
|
||
if (!charset || charset === 'utf-8' || charset === 'utf8') {
|
||
text = buf.toString('utf8');
|
||
} else {
|
||
try { text = new TextDecoder(charset, { fatal: false }).decode(buf); }
|
||
catch (_) { text = buf.toString('utf8'); }
|
||
}
|
||
let body = ct.includes('html') ? htmlToMarkdown(text) : text;
|
||
if (truncated) body += '\n(响应超过 ' + maxBytes + ' 字节,已截断)';
|
||
// 输出字符数兜底:engine 不截工具结果,这是返回值上最后一道护栏。
|
||
return 'HTTP ' + res.status + '\n' + truncate(body, 1024 * 1024);
|
||
} catch (err) {
|
||
// 超时与取消都表现为 signal.aborted(body 读流阶段的报错形态不定,按信号判定最稳)。
|
||
if (ctrl.signal.aborted) {
|
||
if (ctx.run.cancelled) throw new Error('已终止');
|
||
throw new Error('fetch 超时(' + timeoutMs + 'ms):' + url);
|
||
}
|
||
throw new Error('fetch 失败:' + describeFetchError(err));
|
||
} finally {
|
||
clearTimeout(timer);
|
||
clearInterval(cancelWatch);
|
||
}
|
||
}
|
||
|
||
// 单遍实体解码:旧版链式 replace 会把 &lt; 双重解码成 <;HTML5 全量表 2000+ 条,
|
||
// 这里只收网页文本真正常见的命名实体,数字实体(十进制/十六进制)通吃,认不出的原样保留。
|
||
const HTML_ENTITIES = {
|
||
nbsp: ' ', ensp: '\u2002', emsp: '\u2003', thinsp: '\u2009', zwnj: '\u200c', zwj: '\u200d', shy: '\u00ad',
|
||
amp: '&', lt: '<', gt: '>', quot: '"', apos: "'",
|
||
copy: '\u00a9', reg: '\u00ae', trade: '\u2122', deg: '\u00b0', hellip: '\u2026',
|
||
mdash: '\u2014', ndash: '\u2013', lsquo: '\u2018', rsquo: '\u2019', sbquo: '\u2018',
|
||
ldquo: '\u201c', rdquo: '\u201d', bdquo: '\u201e', laquo: '\u00ab', raquo: '\u00bb',
|
||
lsaquo: '\u2039', rsaquo: '\u203a', prime: '\u2032', Prime: '\u2033', bull: '\u2022', middot: '\u00b7',
|
||
times: '\u00d7', divide: '\u00f7', plusmn: '\u00b1', frac12: '\u00bd', frac14: '\u00bc', frac34: '\u00be',
|
||
sup1: '\u00b9', sup2: '\u00b2', sup3: '\u00b3', micro: '\u00b5', para: '\u00b6', sect: '\u00a7',
|
||
dagger: '\u2020', Dagger: '\u2021', permil: '\u2030',
|
||
euro: '\u20ac', pound: '\u00a3', yen: '\u00a5', cent: '\u00a2', curren: '\u00a4',
|
||
agrave: '\u00e0', aacute: '\u00e1', acirc: '\u00e2', atilde: '\u00e3', auml: '\u00e4', aring: '\u00e5', ccedil: '\u00e7',
|
||
egrave: '\u00e8', eacute: '\u00e9', ecirc: '\u00ea', euml: '\u00eb',
|
||
igrave: '\u00ec', iacute: '\u00ed', icirc: '\u00ee', iuml: '\u00ef',
|
||
ograve: '\u00f2', oacute: '\u00f3', ocirc: '\u00f4', otilde: '\u00f5', ouml: '\u00f6',
|
||
ugrave: '\u00f9', uacute: '\u00fa', ucirc: '\u00fb', uuml: '\u00fc',
|
||
yacute: '\u00fd', yuml: '\u00ff', ntilde: '\u00f1', szlig: '\u00df',
|
||
ne: '\u2260', le: '\u2264', ge: '\u2265', infin: '\u221e', sup: '\u2283', sub: '\u2286',
|
||
};
|
||
function decodeEntities(s) {
|
||
return s.replace(/&#x([0-9a-f]+);|&#(\d+);|&([a-zA-Z][a-zA-Z0-9]*);/gi, (m, hx, dec, name) => {
|
||
try {
|
||
if (hx) return String.fromCodePoint(parseInt(hx, 16));
|
||
if (dec) return String.fromCodePoint(parseInt(dec, 10));
|
||
// hasOwnProperty:防 &constructor; 这类名字踩到原型链。
|
||
return Object.prototype.hasOwnProperty.call(HTML_ENTITIES, name) ? HTML_ENTITIES[name] : m;
|
||
} catch (_) { return m; } // 码点越界等原样保留
|
||
});
|
||
}
|
||
|
||
function htmlToMarkdown(html) {
|
||
let s = html
|
||
.replace(/<(script|style|noscript|svg|head)[\s\S]*?<\/\1>/gi, '')
|
||
.replace(/<!--[\s\S]*?-->/g, '');
|
||
s = s
|
||
.replace(/<h([1-6])[^>]*>([\s\S]*?)<\/h\1>/gi, (m, lv, t) => '\n' + '#'.repeat(Number(lv)) + ' ' + strip(t).trim() + '\n')
|
||
.replace(/<li[^>]*>([\s\S]*?)<\/li>/gi, (m, t) => '\n- ' + strip(t).trim())
|
||
.replace(/<a[^>]*href=["']?([^"'\s>]+)["']?[^>]*>([\s\S]*?)<\/a>/gi, (m, href, t) => '[' + strip(t).trim() + '](' + href + ')')
|
||
.replace(/<(br|hr)[^>]*>/gi, '\n')
|
||
.replace(/<\/(p|div|section|article|tr|ul|ol|table|blockquote)>/gi, '\n')
|
||
.replace(/<[^>]+>/g, '');
|
||
// 实体解码放在剥标签之后:<p> 这类转义文本不会被当成标签剥掉。
|
||
return decodeEntities(s).replace(/\n{3,}/g, '\n\n').trim();
|
||
}
|
||
|
||
function strip(s) { return s.replace(/<[^>]+>/g, ''); }
|
||
|
||
function recordFile(ctx, abs) {
|
||
try {
|
||
const st = fs.statSync(abs);
|
||
if (!st.isFile()) return;
|
||
ctx.db.prepare(
|
||
'INSERT INTO files (session_id, filename, size, created_at) VALUES (?,?,?,?)'
|
||
).run(ctx.sessionId, relOf(ctx.taskDir, abs), st.size, Date.now());
|
||
} catch (_) { /* 文件表仅用于展示,失败不影响主流程 */ }
|
||
}
|
||
|
||
// mint_public_link:把当前会话目录里的文件签成一条短命公网链接(由 GET /p/<id> 取用)。
|
||
// 这是唯一的签发入口:与 core 同进程,直接调 publicFiles.mint,不必绕任何 HTTP;
|
||
// 而 ctx.taskDir 就是会话目录,relOf 的结果正是取用侧(routes/public-files.js)要的 rel。
|
||
async function toolMintPublicLink(ctx, args) {
|
||
const pl = (ctx.cfg && ctx.cfg.public_link) || {};
|
||
if (!pl.base_url) throw new Error('public_link.base_url 未配置,无法签发公网链接');
|
||
const abs = resolveInTask(ctx.taskDir, args.path);
|
||
const rel = relOf(ctx.taskDir, abs);
|
||
if (rel === '.') throw new Error('path 要指向一个文件,不能是任务目录本身');
|
||
let st;
|
||
try { st = fs.statSync(abs); } catch (_) { throw new Error('文件不存在:' + rel); }
|
||
if (!st.isFile()) throw new Error('不是文件:' + rel);
|
||
// 符号链接外泄防护:签出去的链接是公网可拉的,realpath 必须仍在任务目录内。
|
||
const real = fs.realpathSync(abs);
|
||
const rootReal = fs.realpathSync(ctx.taskDir);
|
||
if (!real.startsWith(rootReal + path.sep)) {
|
||
throw new Error('path 经符号链接指向了任务目录之外,拒绝签发');
|
||
}
|
||
const id = publicFiles.mint({ sessionId: ctx.sessionId, rel, ttlMs: pl.ttl_ms });
|
||
return pl.base_url + '/p/' + id;
|
||
}
|
||
|
||
// schedule_task:为当前会话创建定时任务。sessionId 取自 ctx(AI 无法指定别的会话);
|
||
// 参数与会话唯一性校验均在 scheduler.createTask,失败抛中文 Error 回给 AI 自行纠正。
|
||
// 生命周期:只能由 AI 创建,用户端只读 + 删除(改 = 删后让 AI 重建),故无修改/删除入口。
|
||
function toolScheduleTask(ctx, args) {
|
||
const cronExpr = String(args.cron_expr || '').trim();
|
||
const prompt = String(args.prompt || '').trim();
|
||
if (!cronExpr) throw new Error('cron_expr 不能为空');
|
||
if (!prompt) throw new Error('prompt 不能为空');
|
||
const t = scheduler.createTask(ctx.db, ctx.sessionId, cronExpr, prompt);
|
||
return '定时任务已创建:' + t.cron_expr + ',下次执行 ' + t.next_text + '(Asia/Shanghai)。' +
|
||
'注意:每次触发前本会话的历史消息会被存档(AI 只看到任务 prompt),因此 prompt 需自包含;' +
|
||
'如需修改或删除,请让用户在会话页操作。';
|
||
}
|
||
|
||
const HANDLERS = {
|
||
read_file: toolReadFile,
|
||
write_file: toolWriteFile,
|
||
edit_file: toolEditFile,
|
||
mint_public_link: toolMintPublicLink,
|
||
terminal: toolTerminal,
|
||
fetch: toolFetch,
|
||
schedule_task: toolScheduleTask,
|
||
list_skills: async (ctx, args) => JSON.stringify(await skills.listSkills(ctx), null, 2),
|
||
get_skill_info: (ctx, args) => skills.getSkillInfo(ctx, args.skill_name),
|
||
run_skill: (ctx, args) => skills.runSkill(ctx, args.skill_name, args.input_args),
|
||
};
|
||
|
||
// ctx: {sessionId, taskDir, run, cfg, db, vision}
|
||
// 返回值一般为字符串;read_file 读图且 ctx.vision 为 true 时返回 {text, image_rel}。
|
||
// 内置工具优先;未命中时交给 MCP 注册表(启动时发现的服务工具)。
|
||
async function executeTool(baseCtx, name, args) {
|
||
const handler = HANDLERS[name];
|
||
if (handler) return handler(baseCtx, args || {});
|
||
if (mcp.hasTool(name)) return mcp.callTool(name, args || {});
|
||
return 'Error: 未知工具 ' + name;
|
||
}
|
||
|
||
module.exports = { executeTool, resolveInTask, recordFile };
|