dsh-loop-rescue 0.0.0-stage → 0.2.0

package/tools/scan-session.mjs

@@ -0,0 +1,95 @@
+#!/usr/bin/env node
+/**
+ * 离线扫描：用插件同一份检测器复跑 DSH 会话文件，回答「这些消息会不会被切、切在哪」。
+ * 用于调阈值：先扫一遍自己的历史会话，看命中是不是都是真循环。
+ *
+ * 用法：
+ *   node tools/scan-session.mjs <会话文件或目录> [--json] [--all]
+ *
+ * - 支持纯 jsonl 和 DSH 的 zstd 压缩会话（按内容识别，需要 Node 22.15+）；
+ * - 目录会递归扫描其中的 *.jsonl / *.zst / *.zstd；
+ * - 只读 `assistant/message` 的已提交内容（即模型下一轮会看到的内容）；
+ * - `--all` 连没命中的消息也输出统计；`--json` 输出机器可读结果。
+ *
+ * 本脚本只导入 ../src/detector.js，不依赖 @deepseek-ai/*，普通 node 即可运行。
+ */
+import { readFileSync, readdirSync, statSync } from 'node:fs'
+import { join } from 'node:path'
+import * as zlib from 'node:zlib'
+import { detectInText } from '../src/detector.js'
+import { decodeLog } from './zstd-log.mjs'
+
+const args = process.argv.slice(2)
+const asJson = args.includes('--json')
+const showAll = args.includes('--all')
+const targets = args.filter((a) => !a.startsWith('--'))
+if (targets.length === 0) {
+  console.error('用法: node tools/scan-session.mjs <会话文件或目录> [--json] [--all]')
+  process.exit(2)
+}
+
+/** @param {string} p */
+function* files(p) {
+  const st = statSync(p)
+  if (st.isDirectory()) {
+    for (const e of readdirSync(p)) yield* files(join(p, e))
+  } else if (/\.(jsonl|zst|zstd)$/.test(p)) {
+    yield p
+  }
+}
+
+/** @param {string} path */
+function readText(path) {
+  return decodeLog(readFileSync(path))
+}
+
+const results = []
+let scanned = 0
+for (const target of targets) {
+  for (const path of files(target)) {
+    let text
+    try {
+      text = readText(path)
+    } catch (error) {
+      console.error(`跳过 ${path}: ${error.message}`)
+      continue
+    }
+    for (const line of text.split('\n')) {
+      if (!line.trim()) continue
+      let ev
+      try { ev = JSON.parse(line) } catch { continue }
+      if (ev?.type !== 'assistant/message') continue
+      const blocks = ev.data?.message?.content ?? []
+      for (const block of blocks) {
+        if (block.type !== 'reasoning' && block.type !== 'text') continue
+        scanned++
+        const hit = detectInText(block.type, block.text ?? '')
+        if (hit === undefined && !showAll) continue
+        results.push({
+          file: path,
+          seq: ev.seq,
+          turn: ev.data.turn,
+          step: ev.data.step,
+          kind: block.type,
+          chars: (block.text ?? '').length,
+          rule: hit?.rule ?? null,
+          repeatedChars: hit ? (block.text ?? '').length - hit.cut : 0,
+          cut: hit?.cut ?? null,
+          preview: hit ? (block.text ?? '').slice(hit.cut, hit.cut + 80).replace(/\s+/g, ' ') : '',
+        })
+      }
+    }
+  }
+}
+
+if (asJson) {
+  console.log(JSON.stringify({ scanned, results }, null, 2))
+} else {
+  const hits = results.filter((r) => r.rule)
+  console.log(`扫描 ${scanned} 个 block，命中 ${hits.length} 个\n`)
+  for (const r of results) {
+    const tag = r.rule ? `${r.rule.padEnd(5)} 剪 ${r.repeatedChars}/${r.chars}` : `未命中 ${r.chars}`
+    console.log(`${r.file}  seq=${r.seq} t${r.turn}s${r.step} ${r.kind.padEnd(9)} ${tag}`)
+    if (r.preview) console.log(`    从这里开始剪: ${r.preview}`)
+  }
+}

package/tools/zstd-log.mjs

@@ -0,0 +1,54 @@
+/**
+ * DSH 会话日志的 zstd 读写。
+ *
+ * session.v4.jsonl.zstd 是多帧文件：头部一帧，之后每次追加一帧。
+ *   - zlib.zstdDecompressSync 只解第一帧 —— 直接用它只能读到头部一行，必须逐帧解；
+ *   - DSH 列会话时只解第一帧并要求它「恰好一行头部」，否则把会话当损坏、静默隐藏 ——
+ *     写回时头部必须单独成帧。
+ */
+import * as zlib from 'node:zlib'
+
+const MAGIC = 0xfd2fb528
+const OPTS = () => ({ params: { [zlib.constants.ZSTD_c_checksumFlag]: 1 } })
+
+/** RFC 8878 帧边界：返回从 start 开始那一帧的结束偏移。 */
+export function frameEnd(buf, start = 0) {
+  if (buf.readUInt32LE(start) !== MAGIC) return undefined
+  const fhd = buf[start + 4]
+  const fcsFlag = fhd >> 6; const single = (fhd >> 5) & 1; const checksum = (fhd >> 2) & 1; const dictFlag = fhd & 3
+  let o = start + 5 + (single ? 0 : 1) + [0, 1, 2, 4][dictFlag] + [single ? 1 : 0, 2, 4, 8][fcsFlag]
+  for (;;) {
+    if (o + 3 > buf.length) return undefined
+    const h = buf[o] | (buf[o + 1] << 8) | (buf[o + 2] << 16)
+    o += 3 + (((h >> 1) & 3) === 1 ? 1 : h >>> 3)
+    if (h & 1) break
+  }
+  o += checksum ? 4 : 0
+  return o <= buf.length ? o : undefined
+}
+
+export function decodeLog(buf) {
+  if (buf.length < 4 || buf.readUInt32LE(0) !== MAGIC) return buf.toString('utf8')
+  if (typeof zlib.zstdDecompressSync !== 'function') throw new Error('需要 Node 22.15+（内置 zstd）。')
+  const parts = []
+  let o = 0
+  while (o < buf.length) {
+    const magic = buf.readUInt32LE(o)
+    if ((magic & 0xfffffff0) === 0x184d2a50) { o += 8 + buf.readUInt32LE(o + 4); continue }
+    const end = frameEnd(buf, o)
+    if (end === undefined) throw new Error(`zstd 帧不完整（偏移 ${o}）`)
+    parts.push(zlib.zstdDecompressSync(buf.subarray(o, end)))
+    o = end
+  }
+  return Buffer.concat(parts).toString('utf8')
+}
+
+export function encodeLog(text, compressed = true) {
+  if (!compressed) return Buffer.from(text, 'utf8')
+  const nl = text.indexOf('\n')
+  const head = nl < 0 ? `${text}\n` : text.slice(0, nl + 1)
+  const rest = nl < 0 ? '' : text.slice(nl + 1)
+  const frames = [zlib.zstdCompressSync(Buffer.from(head, 'utf8'), OPTS())]
+  if (rest.length) frames.push(zlib.zstdCompressSync(Buffer.from(rest, 'utf8'), OPTS()))
+  return Buffer.concat(frames)
+}