feat(tools): 复活 list_directory/copy_path 开 @<会话id> 只读历史通道;recall 收敛纯记忆三 op

- recall 删 files/link 两 op(沙箱外克隆必漂移:link 漏抄 mint_public_link 的
  realpath 符号链接防线,整删即消洞);description 重写:去跨工具越权指引与
  重复细节,补纯文本返回契约与空态语义;count 补 minimum/maximum/default
- list_directory:单层 ls、≤200 条、隐藏项不列;'@<sid>/<rel>' 列历史会话,
  条目输出 @sid/ 全形路径可直抄进 copy_path
- copy_path:历史成果唯一进口。resolvePastPath 校验单点(形状+存在+同用户);
  目标拒 @(历史只读、写只进本会话)、撞名拒覆盖、预算 200 文件/200MB、
  符号链接整趟拒绝防逃逸;导入后 read_file/vision/mint 全量生效,零核心沙箱改动
- advising:阶段硬闸名单加入 copy_path,白名单点名放行 list_directory
- definitions 保持按 name 字母序;ARCHITECTURE 记部分复活与撤销理由
  (生产 terminal 在远端 executor 容器,host-tasks 被 wrapper unmount,
  当年「统一走 terminal」的删除理由不覆盖跨会话场景)
- 验证:harness 23/23(@ 下钻/跨用户/撞名/symlink/退役 op 回归);node --check 全过

版本 2.2.0 → 2.3.0
This commit is contained in:
2026-09-30 09:44:14 +08:00
parent ba4b1149aa
commit 419008597d
7 changed files with 209 additions and 65 deletions
+1 -1
View File
@@ -1,6 +1,6 @@
{
"name": "agent-app",
"version": "2.2.0",
"version": "2.3.0",
"description": "自用通用 Agent 系统:多模型调度 + 任务目录 + 工具调用 + H5 Web UI",
"main": "server/server.js",
"scripts": {
+3 -2
View File
@@ -113,12 +113,13 @@ const DEFAULTS = {
// 模型 profile(prompt + tools 名单组合):platforms 里的模型用 profile: <名> 引用。
model_profiles: {
// advising:顾问阶段(sessions.phase)专用组合——不由模型配置引用,而是 resolveProfileName 在
// advising 时强制命中:顾问只聊和读、不写(write_file 已移除,批内幻觉调用另由引擎挡下),
// advising 时强制命中:顾问只聊和读、不写(write_file/edit_file/copy_path 不在名单,批内幻觉调用另由引擎挡下;
// list_directory 入选——翻阅历史目录是顾问高频刚需),
// 例外是 profile(画像元信息,不碰任务目录;顾问问答恰是收集称谓/偏好密度最高的时刻)
// 与 recall(日记只读翻阅;用户和顾问聊「上次那个事」恰是回忆查询的高峰场景);
// 唯一出口是 handoff;tools_deny 必须显式清空:不配 = 继承全局 ['lab','handoff'],
// allow 滤完再叠 deny 会把 handoff 自己禁掉(同 lab 组合的教训)。
advising: { tools_allow: ['read_file', 'fetch', 'handoff', 'profile', 'recall'], tools_deny: [], system_prompt_file: 'advising-system.txt' },
advising: { tools_allow: ['read_file', 'list_directory', 'fetch', 'handoff', 'profile', 'recall'], tools_deny: [], system_prompt_file: 'advising-system.txt' },
// lab:统一入口元工具自测组合——tools 收缩到 1 个,其余能力经 lab(help) 现场发现。
// tools_deny 必须显式清空:不配 = 继承全局 ['lab'],deny 后叠加会把 lab 自己禁掉 → 零工具。
lab: { tools_allow: ['lab'], tools_deny: [], system_prompt_file: 'lab-system.txt' },
+3 -2
View File
@@ -345,7 +345,8 @@ function effectiveTools(profileName) {
// ---------- 阶段(sessions.phase:advising 顾问 / executing 执行) ----------
// advising:新建会话与「清空」的落点。profile 被 resolveProfileName 强制为 'advising',
// 只聊和读不写(write_file/edit_file 不在其 schema,批内幻觉调用也在 runLoopInner 挡下);
// 只聊和读不写(write_file/edit_file/copy_path 不在其 schema,批内幻觉调用也在 runLoopInner 挡下;
// list_directory 在白名单里——顾问翻阅历史目录正当其用);
// 唯一出口 = handoff 工具成功返回(brief 由 handler 落盘任务目录根 task.md,见 tools/executor.js)。
// executing:一路走到底,只有清空才回 advising;压缩不碰 phase(压完仍回原模式)。
@@ -783,7 +784,7 @@ async function runLoopInner(sessionId, run) {
output = chk.error;
isError = true;
} else if (session.phase === 'advising'
&& (tc.function.name === 'write_file' || tc.function.name === 'edit_file')) {
&& (tc.function.name === 'write_file' || tc.function.name === 'edit_file' || tc.function.name === 'copy_path')) {
// 顾问单入口硬约束:写类工具不在其 schema,但「名单外照跑」是全局策略(executeTool 全表),
// 幻觉调用会真写文件、绕过出口制造假承诺——阶段契约优先,挡下并指向 handoff。
output = 'Error: 顾问阶段不能写文件。任务书请通过 handoff 工具的 brief 参数提交。';
+47 -14
View File
@@ -1,4 +1,28 @@
[
{
"type": "function",
"function": {
"name": "copy_path",
"description": "Copies a file or directory, preserving the original; directories are copied recursively. More efficient than reading then writing contents. Returns confirmation with the destination path.\n\nImport past work: source_path accepts '@<session-id>/<rel>' to copy from your OWN past session (id from a diary fragment's 【…】 heading; use recall to find it, list_directory to browse it). destination_path never accepts '@' — past stays read-only, imports land in the current session directory. The destination must not exist yet; nothing is overwritten (pick a new name, or delete yours first via terminal).\n\nAfter import the file is fully local: read_file (including images via vision), terminal, mint_public_link all work on it.\n\nLimits per call: ≤200 MB total, ≤200 files for a directory tree; symlinks are refused (as a source or inside a copied tree) to prevent escaping the directory.",
"parameters": {
"type": "object",
"required": [
"source_path",
"destination_path"
],
"properties": {
"source_path": {
"type": "string",
"description": "File or directory to copy, relative to the session working directory (absolute paths inside it accepted). '@<session-id>/<rel>' imports from a same-user past session."
},
"destination_path": {
"type": "string",
"description": "Where to place the copy inside the CURRENT session directory (relative, or absolute inside it). '@' paths are rejected here. Must not already exist."
}
}
}
}
},
{
"type": "function",
"function": {
@@ -173,6 +197,22 @@
}
}
},
{
"type": "function",
"function": {
"name": "list_directory",
"description": "Lists the entries directly inside one directory (one level, like `ls`): folders and files, sorted; hidden entries (starting with '.') omitted; max 200 per call — go deeper by passing a deeper path.\n\npath is relative to the session working directory (absolute paths inside it accepted); empty/omitted lists the current task directory.\n\nPast sessions: '@<session-id>/<rel>' lists a PAST session's own directory, same user only (read-only view). The id is the 【…】 heading of a diary fragment — find it with recall op=search. In '@' mode entries are printed as complete '@<session-id>/...' paths — copy them straight into copy_path's source_path.",
"parameters": {
"type": "object",
"properties": {
"path": {
"type": "string",
"description": "Directory to list, relative to the session working directory. Absolute paths accepted. '@<session-id>/<rel>' lists that past session's directory (same user only; id from a diary fragment's 【…】 heading)."
}
}
}
}
},
{
"type": "function",
"function": {
@@ -246,7 +286,7 @@
"type": "function",
"function": {
"name": "recall",
"description": "Access your long-term memory: the work diary. Each finished session leaves one short fragment (a few dozen characters, headed by the source session id), stored permanently; only the newest fragments were injected at session start. When the user vaguely recalls something from the past, search is your primary entry — you cannot know what you have forgotten. Read-only.\n\nop=search Requires query (keywords; space-separated words rank better, e.g. '身份证 水印'). Fuzzy fragment-level match across ALL dates (Chinese tolerant to partial phrases), best matches first. Optional count (default 10, max 30).\nop=days No arguments. Index of diary days (date + fragment count) — use when the user remembers roughly WHEN but not what; then op=day.\nop=day Requires date (YYYY-MM-DD, as listed by op=days). Full fragments of that day.\nop=files Requires id (the session id from a diary fragment's 【…】 heading). Files of that past session that still exist on disk right now.\nop=link Requires id and path (a name from op=files). Mint a temporary download link.",
"description": "Access your long-term memory: the work diary. Each finished session leaves one short fragment (a few dozen chars) there, headed by its source session id in 【…】; fragments are kept permanently, but only the newest were injected at session start — recall reads the whole archive on demand. Read-only.\n\nop=search Fuzzy keyword match over ALL fragments, best first. Primary entry when the user vaguely recalls something — you cannot know what you have forgotten.\nop=days Index of days that have entries (newest first) — when the user remembers roughly WHEN but not what.\nop=day Full fragments of one day.\n\nReturns are plain text: search hits print '--- date (match%) ---' followed by 【id】 and the fragment; days list '- date(n 条)'; day lists fragments. A miss comes back as a short 'no records / no diary' line — a normal answer, not an error.",
"parameters": {
"type": "object",
"required": [
@@ -258,31 +298,24 @@
"enum": [
"search",
"days",
"day",
"files",
"link"
"day"
],
"description": "What to recall: search = keyword fuzzy over all diary fragments (primary); days = which dates have entries; day = one date's full diary; files = a past session's current artifacts; link = a download link for one of them."
"description": "Recall operation; what each op does is in the tool description."
},
"query": {
"type": "string",
"description": "op=search: keywords to look for (space-separated words rank best, e.g. '身份证 水印'). Chinese and English both supported, case-insensitive."
"description": "op=search: keywords; space-separated words rank best (e.g. '身份证 水印'). Chinese and English, case-insensitive; Chinese tolerates partial phrases."
},
"count": {
"type": "integer",
"minimum": 1,
"maximum": 30,
"default": 10,
"description": "op=search: max fragments to return (default 10, max 30)."
},
"date": {
"type": "string",
"description": "op=day: the day to open, YYYY-MM-DD exactly as shown by op=days."
},
"id": {
"type": "string",
"description": "op=files / op=link: the source session id as written inside a diary fragment's 【…】 heading (e.g. 'task-20260929-003')."
},
"path": {
"type": "string",
"description": "op=link: file name inside that session's task directory, exactly as listed by op=files."
}
}
}
+152 -44
View File
@@ -37,6 +37,57 @@ function relOf(taskDir, abs) {
return path.relative(path.resolve(taskDir), abs) || '.';
}
// ---------- 跨会话路径:@<会话id>/<rel>(历史只读,数据过墙只走 copy_path 一道门) ----------
// 背景:terminal 在远端 executor 容器里跑(compose:wrapper 启动即 unmount host-tasks),
// 别的会话目录在 shell 物理视图里根本不存在;core 进程内的文件工具又被 resolveInTask
// 锁死本会话。跨会话取旧成果只剩这一个口子,同用户校验单点在此(承自 recall 旧 files/link,
// 那两个 op 是 ls/mint 的沙箱外克隆,已整删——克隆会漂移:link 漏抄 mint 的 realpath 防线)。
const PAST_RE = /^@([^\s/]+)(?:\/([\s\S]*))?$/;
function resolvePastPath(ctx, p) {
const m = PAST_RE.exec(p);
if (!m) throw new Error('@ 路径格式:@<会话id>/<相对路径>,如 @task-20260929-003/out/a.png');
const sid = m[1];
const rel = m[2] || '.';
if (!/^task-[0-9A-Za-z_\-]+$/.test(sid)) throw new Error('@ 后面需要会话 id(形如 task-YYYYMMDD-NNN,来自日记碎片【编号】):' + sid);
const s = ctx.db.prepare('SELECT id, user, title FROM sessions WHERE id = ?').get(sid);
if (!s) throw new Error('会话 ' + sid + ' 不存在(可能已删除,目录随会话一并销毁)');
const me = ctx.db.prepare('SELECT user FROM sessions WHERE id = ?').get(ctx.sessionId);
if (String(s.user || '') !== String((me && me.user) || '')) throw new Error('无权访问该会话(非当前用户)');
const root = String((ctx.cfg && ctx.cfg.workspace_root) || '');
if (!root) throw new Error('workspace_root 未配置');
const dir = path.join(root, sid);
return { sid: sid, title: s.title || sid, dir: dir, abs: resolveInTask(dir, rel) };
}
// list_directory 单层展示上限 / copy_path 单次导入预算(文件数 + 字节)
const LIST_MAX_ENTRIES = 200;
const COPY_MAX_FILES = 200;
const COPY_MAX_BYTES = 200 * 1048576;
// 目录源先整趟预算:计数、累字节、遇符号链接即拒(防沿链接跳出源目录把 core 文件系统拽进任务目录)。
async function walkCopyBudget(srcAbs) {
let n = 0; let bytes = 0;
const stack = [srcAbs];
while (stack.length) {
const d = stack.pop();
const entries = await fsp.readdir(d, { withFileTypes: true });
for (const e of entries) {
const abs = path.join(d, e.name);
if (e.isSymbolicLink()) {
throw new Error('源包含符号链接(拒拷贝,防沿链接跳出源目录):' + path.relative(srcAbs, abs));
}
if (e.isDirectory()) stack.push(abs);
else if (e.isFile()) {
n++; bytes += (await fsp.stat(abs)).size;
if (n > COPY_MAX_FILES || bytes > COPY_MAX_BYTES) {
throw new Error('源过大(上限 ' + COPY_MAX_FILES + ' 文件 / ' + Math.round(COPY_MAX_BYTES / 1048576) + ' MB);先用 list_directory 看清结构,分批挑');
}
}
}
}
return { files: n, bytes: bytes };
}
function isProbablyBinary(abs) {
let fd;
try {
@@ -163,7 +214,7 @@ async function toolReadFile(ctx, args) {
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 (st.isDirectory()) throw new Error(args.path + ' 是目录不是文件,请用 list_directory 查看(历史会话传 @<会话id>/<路径>)。');
if (IMAGE_EXTS.has(path.extname(abs).toLowerCase())) {
const rel = relOf(ctx.taskDir, abs);
if (!ctx.vision) {
@@ -202,6 +253,98 @@ async function toolReadFile(ctx, args) {
return lines.map((l, i) => String(startLine + i).padStart(6, ' ') + '\t' + l).join('\n');
}
// 对齐 Zed list_directory 的输出格式:
// # Folders:
// <dir>/<name>
//
// # Files:
// <dir>/<name>
// 本会话路径:条目相对任务目录(同旧版)。'@<sid>/<rel>' 历史模式:条目带 '@<sid>/' 全形前缀,
// 可直接抄进 copy_path 的 source_path——列得出来就导得进来。隐藏项('.' 开头的 .imgcache/
// .task-history/.tmp 等内部态)不列;确切知道名字时仍可直接把它喂给 copy_path。
async function toolListDirectory(ctx, args) {
const p = typeof args.path === 'string' ? args.path.trim() : '';
let abs; let prefix; let title = '';
if (p.startsWith('@')) {
const past = resolvePastPath(ctx, p);
abs = past.abs;
const inner = relOf(past.dir, abs);
prefix = '@' + past.sid + (inner === '.' ? '/' : '/' + inner + '/');
title = '「' + past.title + '」(历史会话,只读;条目可直接作为 copy_path 的 source_path 导入)\n';
} else {
abs = resolveInTask(ctx.taskDir, p || '.');
const inner = relOf(ctx.taskDir, abs);
prefix = inner === '.' ? '' : inner + '/';
}
let entries;
try {
entries = await fsp.readdir(abs, { withFileTypes: true });
} catch (e) {
if (e.code === 'ENOENT') throw new Error('目录不存在:' + p + (p.startsWith('@') ? '(先用 list_directory 逐层确认该历史会话的文件结构)' : ''));
throw e;
}
const dirs = []; const files = [];
for (const e of entries) {
if (e.name.startsWith('.')) continue;
(e.isDirectory() ? dirs : files).push(prefix + e.name);
}
dirs.sort(); files.sort();
const total = dirs.length + files.length;
if (total === 0) return title + '目录为空';
const shownDirs = dirs.slice(0, LIST_MAX_ENTRIES);
const shownFiles = files.slice(0, LIST_MAX_ENTRIES - shownDirs.length);
const sections = [];
if (shownDirs.length) sections.push('# Folders:\n' + shownDirs.join('\n'));
if (shownFiles.length) sections.push('# Files:\n' + shownFiles.join('\n'));
let out = title + sections.join('\n\n');
if (shownDirs.length + shownFiles.length < total) out += '\n\n...(共 ' + total + ' 项,本层只展示前 ' + LIST_MAX_ENTRIES + ' 项;传更深路径下钻)';
return out;
}
// copy_path:普通复制 + 历史成果的唯一进口。源可 '@<sid>/<rel>'(同用户校验在 resolvePastPath);
// 目标永不接受 @——历史只读、写只进本会话,「导入后文件即本地」让 read_file/vision/mint 全量生效,
// 不必给它们开跨墙特例。撞名拒绝覆盖(保用户现作)。
async function toolCopyPath(ctx, args) {
const sp = String(args.source_path || '').trim();
const dp = String(args.destination_path || '').trim();
if (!sp || !dp) throw new Error('source_path 与 destination_path 都必填');
if (dp.startsWith('@')) throw new Error('destination_path 不支持 @ 记号——历史只读,导入只会落进本会话目录');
const dst = resolveInTask(ctx.taskDir, dp);
if (path.resolve(dst) === path.resolve(ctx.taskDir)) throw new Error('destination_path 不能是本任务目录本身');
let src; let srcTitle = '';
if (sp.startsWith('@')) {
const past = resolvePastPath(ctx, sp);
src = past.abs;
srcTitle = '(来自「' + past.title + '」)';
} else {
src = resolveInTask(ctx.taskDir, sp);
}
if (path.resolve(src) === path.resolve(dst)) throw new Error('源与目标是同一个路径');
let lst;
try {
lst = await fsp.lstat(src);
} catch (_) {
throw new Error('源不存在:' + sp + (sp.startsWith('@') ? '(历史会话目录以 list_directory 实际列举为准)' : ''));
}
if (lst.isSymbolicLink()) throw new Error('源是符号链接,拒绝拷贝(防沿链接跳出任务目录):' + sp);
let budget = null;
if (lst.isFile()) {
if (lst.size > COPY_MAX_BYTES) throw new Error('源文件 ' + Math.round(lst.size / 1048576) + ' MB,超过单次 ' + Math.round(COPY_MAX_BYTES / 1048576) + ' MB 上限');
} else if (lst.isDirectory()) {
budget = await walkCopyBudget(src);
} else {
throw new Error('源不是普通文件或目录:' + sp);
}
let exists = true;
try { await fsp.stat(dst); } catch (_) { exists = false; }
if (exists) throw new Error('目标已存在,拒绝覆盖:' + relOf(ctx.taskDir, dst) + '(换个目标名,或先 terminal 删除本会话的同名旧物)');
await fsp.mkdir(path.dirname(dst), { recursive: true });
await fsp.cp(src, dst, { recursive: true });
recordFile(ctx, dst); // 文件表只记普通文件(目录自然跳过),供前端/微信附件展示
const sizeNote = budget ? '(' + budget.files + ' 个文件,' + Math.round(budget.bytes / 1024) + ' KB)' : '';
return '已复制 ' + sp + srcTitle + ' → ' + relOf(ctx.taskDir, dst) + sizeNote;
}
async function toolWriteFile(ctx, args) {
const abs = resolveInTask(ctx.taskDir, args.path);
// 纵深防御:引擎 checkToolArgs 已把 null 判缺,但直调 executor(测试等)不经过它;
@@ -1382,25 +1525,16 @@ async function toolHandoff(ctx, args) {
return '已接收交接(reason: ' + reason + '),本批工具跑完后系统将切换到执行阶段。';
}
// recall:长期记忆(工作日记)的回忆入口(定稿 B 方案:极简三 op,描述短不挤占工具位)。
// 日记按会话收口持久存储(memory.js),开场只注入最近 N 条;更早的条目与某次工作的成果文件由本工具按需取。
// 只读数据面:仅能碰本用户的日记与会话(id 同用户校验防互探);link 产出的外链与 mint_public_link 同机制同时效。
// recall:长期记忆(工作日记)的回忆入口——纯记忆三 op。
// 日记按会话收口持久存储(memory.js),开场只注入最近 N 条;更早的条目由本工具按需取。
// 历史成果文件不再归日记管:浏览/导入走 list_directory / copy_path 的 @<会话id>/<路径> 记号
// (校验单点 resolvePastPath),导入后就是本会话普通文件,read_file/mint 照常处理。
// 旧 files/link 两 op 是 ls/mint 的沙箱外克隆,已整删(克隆必漂移:link 漏抄 mint 的 realpath 防线)。
function toolRecall(ctx, args) {
const op = String(args.op || '').trim().toLowerCase();
const db = ctx.db;
const me = db.prepare('SELECT user FROM sessions WHERE id = ?').get(ctx.sessionId);
const user = String((me && me.user) || '');
// files/link 共用:解析并校验目标会话(存在 + 同用户 + id 形态),返回 {sid, title, dir}
const resolveSession = () => {
const sid = String(args.id || '').trim();
if (!/^task-[0-9A-Za-z_\-]+$/.test(sid)) throw new Error('需要 id:日记碎片【编号】里的会话 id(形如 task-YYYYMMDD-NNN)');
const s = db.prepare('SELECT id, user, title FROM sessions WHERE id = ?').get(sid);
if (!s) throw new Error('会话 ' + sid + ' 不存在(可能已删除)');
if (String(s.user || '') !== user) throw new Error('无权访问该会话(非当前用户)');
const root = String((ctx.cfg && ctx.cfg.workspace_root) || '');
if (!root) throw new Error('workspace_root 未配置');
return { sid: sid, title: s.title || sid, dir: path.join(root, sid) };
};
switch (op) {
case 'search': {
const q = String(args.query || '').trim();
@@ -1419,41 +1553,15 @@ function toolRecall(ctx, args) {
const txt = memory.dayEntry(db, user, date);
return txt ? date + ' 的日记:\n' + txt : date + ' 无日记。';
}
case 'files': {
const t = resolveSession();
let names = [];
try {
names = fs.readdirSync(t.dir, { withFileTypes: true })
.filter(d => d.isFile() && !d.name.startsWith('.'))
.map(d => d.name)
.sort();
} catch (_) {
return '「' + t.title + '」任务目录不存在,无成果文件。';
}
if (!names.length) return '「' + t.title + '」无成果文件。';
const shown = names.slice(0, 80);
return '「' + t.title + '」现存文件(顶层 ' + shown.length + (names.length > shown.length ? '/共 ' + names.length : '') + ' 个):\n'
+ shown.map(n => '- ' + n).join('\n');
}
case 'link': {
const t = resolveSession();
const rel = String(args.path || '').trim();
if (!rel) throw new Error('op=link 需要 path(op=files 清单里的文件名)');
const abs = resolveInTask(t.dir, rel); // 路径沙箱:只允许该会话任务目录内
let st;
try { st = fs.statSync(abs); } catch (_) { throw new Error('文件不存在:' + rel + '(先 op=files 看现存清单)'); }
if (!st.isFile()) throw new Error('不是普通文件:' + rel);
const pl = ctx.cfg && ctx.cfg.public_link;
if (!pl || !pl.base_url) throw new Error('public_link 未配置,无法生成外链');
return pl.base_url + '/p/' + publicFiles.mint({ sessionId: t.sid, rel: rel, ttlMs: pl.ttl_ms });
}
default:
throw new Error('未知的 op:"' + op + '"。可用:search(需 query,可选 count)/ days(日期索引)/ day(需 date)/ files(需 id)/ link(需 id+path)');
throw new Error('未知的 op:"' + op + '"。可用:search(需 query,可选 count)/ days(日期索引)/ day(需 date)。历史会话文件请用 list_directory / copy_path 的 @<会话id>/<路径> 记号');
}
}
const HANDLERS = {
read_file: toolReadFile,
list_directory: toolListDirectory,
copy_path: toolCopyPath,
handoff: toolHandoff,
write_file: toolWriteFile,
edit_file: toolEditFile,
+2 -1
View File
@@ -694,7 +694,8 @@ async function deliverConclusion(bid, m, sessionId) {
}
// 扫回复文本里的 /p/<22位id> 链 → 可卡片化的本地附件({filename, b64},最多 5 个,单文件限 20MB)。
// 归属校验按「同用户」放行(B 案定稿):recall op=link 合法取旧会话文件,原「同会话」规矩会打断
// 归属校验按「同用户」放行(B 案定稿:当年 recall op=link 合法取旧会话文件——该 op 现已退役,
// 链接只剩 mint_public_link 从本会话签发;同用户放宽保留,兼容存量会话里已签出的旧形条目),原「同会话」规矩会打断
// 记忆链最后一公里;pid 是 22 位随机不可猜的 bearer 能力,真正要防的是跨用户外流——签发会话查不到
// 主人(已删)也拒。过期(resolve 空)、文件已删、跳出任务目录、超限全部静默跳过——链接文字留在正文可点。
const ATTACH_MAX_BYTES = 20 * 1024 * 1024;
+1 -1
View File
@@ -336,7 +336,7 @@ stateDiagram-v2
|---|---|
| `read_file` | 走 `resolveInTask` 防路径穿越;图像(PNG/JPEG/WebP/GIF/BMP/TIFF)按 vision 能力走 `sharp` 压缩后落盘 `.imgcache/` 并回传 `image_rel`(长边 ≤ 2000px,JPEG q=80;文件名带源文件 mtime+size 指纹,重读复用不增殖;重编码无收益时不落盘、直指原文件),发送前由 `llm.buildMessages` 现签成公网外链(engine 侧 mint 前 realpath 复核);大文本/二进制给 outline,超 32MB 拒读并引导 terminal;行号前缀格式 "`<6位右对齐行号>\\t<原文>`"(`edit_file` 用) |
| `write_file` / `edit_file` | `edit_file` 用模糊定位(Zed 风格):先精确匹配,失败后用 Levenshtein DP 找最相似行;超 32MB 拒绝整文件读写(OOM 防护);DP 候选区间不连续(中间跳过无关 buffer 行)即弃用,防静默吞行;空 `new_text` = 整行干净删除(精确/模糊两层统一,对齐 `"".lines()` 零行语义);`old_text`/`new_text` 协议见 `definitions.json` |
| 已移除的工具 | `list_directory` / `create_directory` / `delete_path` / `move_path` / `copy_path` / `find_path` / `grep`:环境固定,文件操作与检索统一走 `terminal`(ls / mkdir / rm / mv / cp / find / rg),内置工具 17 → 10;技能三件套又并为 `skill` 后为 8;再增元工具 `lab` 为 9(默认被全局 `tools_deny` 屏蔽);后续增 `db`、`profile` 各一 → 现 12 项 |
| 已移除的工具(部分复活) | `list_directory` / `copy_path` 于 2026-09 复活并加 `@<session-id>/<rel>` 只读记号:当年删除理由是「环境固定,统一走 terminal」,但生产 terminal 在远端 executor 容器(compose:wrapper 启动即 unmount host-tasks),其他会话目录在 shell 物理视图里不存在——跨会话取旧成果只剩 core 进程一个口子。校验单点 `resolvePastPath`(形状+存在+同用户,承自 recall 旧 files/link);list_directory 单层 ls(≤200 条、隐藏项不列,@ 模式条目带 '@<sid>/' 全形前缀可直接抄进 copy_path);copy_path 源可 @、目标拒 @(历史只读,导入落本会话后 read_file/vision/mint 全量生效,不给存量工具开跨墙特例),预算 200 文件/200MB,符号链接拒收防逃逸(mint 同款防线)。同批拆除 recall 的 files/link 两 op(沙箱外克隆必漂移:link 漏抄 mint 的 realpath 防线),recall 收敛为纯记忆 search/days/day。仍移除:`create_directory` / `delete_path` / `move_path` / `find_path` / `grep`(写删检索继续走 terminal);17→10 大盘不变,复活两件+后续 db/profile → 现 15 项 |
| `terminal` | **唯一一个走 executor 的内置工具**(当 `rexec.isRemote(cfg)`);本地分支用 `spawn` + `detached` + 进程组 kill;远端分支 `rexec.run` SSE 收流,结果套上超时/截断/取消注解后 `truncate(text, cfg.terminal.max_output_chars)` 返回 |
| `fetch` | `AbortController` 覆盖全生命周期(响应头之后的 body 读流同样被超时/取消打断,取消用与 terminal 同款的 200ms 轮询);content-type 先行,非文本只回状态/类型/尺寸不读 body;**下载/返回双上限解耦**:流式读到 `cfg.fetch.max_bytes`(默认 1MB)即断开(下载不占上下文可宽松),返回窗默认 `max_chars`(12000)字符,模型显式传 `max_chars` 夹在 `max_chars_hard`(64000)硬帽内;超窗尾部拼 `offset` 续读提示,转换后全文缓在 `ctx.run.fetchCache`(回合级 LRU 3 条),续读纯本地切片不重抓;**SSRF 防护**:主机名解析结果含回环/私有/链路本地地址即拒(`allow_private: true` 才放行),`redirect:'manual'` 逐跳复查至多 5 次(残余风险:undici 建连时二次解析,DNS 重绑定不可根治);按响应 charset 解码(full-icu 支持 gbk/gb18030 等,未知编码退回 UTF-8);HTML 用 `htmlToMarkdown` 转(剥 script/style/head 及 nav/header/footer/aside/form 整块,class/id 样板词启发式丢块,`<title>` 提成首行);HTML 实体单遍解码(常见命名实体+数字实体通吃,修 `&amp;` 双重解码);字节截断切在多字节字符中时尾 U+FFFD 去噪;网络错误解包 `err.cause` 直述 ENOTFOUND/TLS 等真因,SSRF/重定向类错误带 `fetchUserError` 标记直接冒泡;engine 不截工具结果,返回值末尾另有 `truncate` 1MB 字符兜底护栏 |
| `mint_public_link` | **唯一的 mint 入口**:与 core 同进程,直接 `publicFiles.mint()`,不绕任何 HTTP;`ctx.taskDir` 就是会话目录,`relOf` 的结果正是取用侧要的 `rel`。校验:`public_link.base_url` 已配 + `resolveInTask` 沙箱 + 确认是文件 + `realpath` 未沿符号链接跳出任务目录;返回单行 URL |