dsh-tap 0.20.0

package/providers/trae/remote.js

@@ -0,0 +1,365 @@
+/**
+ * providers/trae/remote.js — Trae remote 会话传输（chat_sessions 协议）。
+ *
+ * 存在理由（2026-08-24 探测定论，证据 docs/probes/trae-model-routing[234]-*）：
+ * raw llm_utils_chat 的模型路由被 function 位钉死——inline_chat 恒走账户默认
+ * 模型（非默认模型名一律 3003 "all models failed"），solo_agent_lite 恒
+ * seed-code-lite、solo_work_lite 恒 glm-5.2、chat_v3 恒 seed-code-lite；
+ * model/custom_model 字段均被忽略。**唯一真实的模型选择机制是 remote 会话
+ * 协议**：POST /api/remote/v1/chat_sessions 的 initial_message.model_name +
+ * model_selection_strategy:"manual"（solo_agent_remote 组全部模型可选，
+ * model_config/done 事件的 config_name 双重证实）。
+ *
+ * 代价（诚实标注）：每请求起一个云端沙箱 agent（25k 系统提示，跨会话有前缀
+ * 缓存），消耗 **work 额度池**（available_endpoint=1；raw 通道消耗 IDE 池
+ * endpoint=0）；并发受套餐 solo_agent_parallel_limit 限制（超额排队，
+ * queuing/notification 事件）；dsh 侧 OpenAI 工具无法原生传递（远端 agent
+ * 自持有工具，参照实现用 XML 提示模拟——本传输 v1 不做，带 tools 的请求
+ * 由网关侧明确拒绝）。
+ *
+ * 事件语法（2026-08-24 真实流校准，证据 docs/probes/trae-remote-stream-glm53.txt）：
+ *   metadata / heartbeat / status_changed / platform_timing / timing_events → 忽略
+ *   queuing / notification → 排队提示
+ *   model_config {config_name} → 实际路由模型（真值源）
+ *   plan_item {id, thought, reasoning_content, tool_call_info} → thought 为可见
+ *     文本、reasoning_content 为思考，**均为累计快照**（按 id 分槽前缀差分）；
+ *     tool_call_info.name==="finish" 时 params.summary 为最终答复
+ *   token_usage {prompt_tokens, completion_tokens, total_tokens, ...} → usage
+ *   done {status:"completed", user_message_context.model_info.config_name} → 终结
+ *   error → 错误
+ * 参照：github.com/autumnsentiment/Trae2api-cn（trae_remote_client.py / sse.py）。
+ */
+
+const DEFAULT_REMOTE_ORIGIN = 'https://solo.trae.cn'
+const WEB_UA =
+  'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 ' +
+  '(KHTML, like Gecko) Chrome/149.0.0.0 Safari/537.36'
+
+/** remote 面请求头（Web 客户端形态；与 raw IDE 头组完全不同的指纹）。 */
+export function remoteWebHeaders(token, { stream = true } = {}) {
+  return {
+    Authorization: `Cloud-IDE-JWT ${token}`,
+    'Content-Type': 'application/json',
+    'X-Trae-Client-Type': 'web',
+    'X-Preferenced-Language': 'zh-CN',
+    'x-user-region': 'CN',
+    Origin: DEFAULT_REMOTE_ORIGIN,
+    Referer: DEFAULT_REMOTE_ORIGIN + '/',
+    'User-Agent': WEB_UA,
+    Accept: stream ? 'text/event-stream' : 'application/json',
+  }
+}
+
+/** OpenAI content（字符串或分段数组）→ 纯文本。 */
+function contentToText(content) {
+  if (typeof content === 'string') return content
+  if (Array.isArray(content)) {
+    return content
+      .filter((p) => p && typeof p === 'object' && p.type === 'text' && typeof p.text === 'string')
+      .map((p) => p.text)
+      .join('\n')
+  }
+  return ''
+}
+
+/**
+ * OpenAI messages → remote query 字符串（JSON 数组包一个 text 块）。
+ * 多轮历史带角色标记扁平化（镜像参照 flatten_query）：[System]/[Assistant]/
+ * [Client Tool Call]/[Client Tool Result]——remote 会话每请求新建，历史只能
+ * 以文本形式随 query 携带。
+ */
+export function flattenQuery(messages) {
+  const parts = []
+  for (const m of Array.isArray(messages) ? messages : []) {
+    if (!m || typeof m !== 'object') continue
+    const role = String(m.role ?? 'user')
+    const content = contentToText(m.content)
+    if (role === 'system' || role === 'developer') {
+      if (content) parts.push(`[System]\n${content}`)
+      continue
+    }
+    if (role === 'assistant') {
+      const segs = []
+      if (content) segs.push(content)
+      for (const c of Array.isArray(m.tool_calls) ? m.tool_calls : []) {
+        const fn = c?.function ?? {}
+        const id = c?.id ?? 'unknown'
+        const args = typeof fn.arguments === 'string' ? fn.arguments : JSON.stringify(fn.arguments ?? {})
+        segs.push(`[Client Tool Call: ${id} ${fn.name ?? 'unknown'}]\nArguments: ${args}`)
+      }
+      if (segs.length) parts.push('[Assistant]\n' + segs.join('\n\n'))
+      continue
+    }
+    if (role === 'tool') {
+      const id = m.tool_call_id ?? 'unknown'
+      const name = m.name ?? 'tool'
+      parts.push(`[Client Tool Result: ${id} ${name}]\n${content}`)
+      continue
+    }
+    if (content) parts.push(content)
+  }
+  return JSON.stringify([{ type: 'text', data: { content: parts.join('\n\n') } }])
+}
+
+function commonParams() {
+  return JSON.stringify({
+    language: 'zh-cn', app_language: 'zh-CN', quality: 'stable',
+    app_version: '1.0.0.1229', web_id: '', user_identity: 'Free',
+    is_freshman: '0', biz_user_id: '', user_unique_id: '',
+    scope: 'marscode-cn', tenant: 'marscode', region: 'cn', aiRegion: 'cn',
+    is_privacy_mode: 0, privacy_mode: 'off', solo_chat_mode: 'code',
+  })
+}
+
+/** chat_sessions 创建体（9router 变体：content 空数组，query 承载全部）。
+ *  opts.reasoningEffort：目录声明的档位拼写 → initial_message.custom_model
+ *  .reasoning_effort（官方客户端线缆形态，@byted-icube/solo-lite 551.*.mjs
+ *  custom_model 合并链取证 + 2026-10-04 probe-trae-thinking-scene 证实接受面；
+ *  preset 模型 config_source=1）。不带/空串 = 不出 custom_model（既有路径
+ *  零变化）。 */
+export function buildRemoteCreateBody(model, messages, { reasoningEffort } = {}) {
+  const initialMessage = {
+    chat_session_id: '',
+    content: [],
+    query: flattenQuery(messages),
+    model_name: model,
+    agent_type: 'solo_agent_remote',
+    model_selection_strategy: 'manual',
+    common_params: commonParams(),
+  }
+  if (typeof reasoningEffort === 'string' && reasoningEffort) {
+    initialMessage.custom_model = {
+      model_name: model,
+      config_name: model,
+      config_source: 1,
+      is_preset: true,
+      use_remote_service: true,
+      multimodal: false,
+      reasoning_effort: reasoningEffort,
+    }
+  }
+  return {
+    mode: 'code',
+    environment_id: 'default',
+    initial_message: initialMessage,
+    env: 'remote',
+    auto_create_project: false,
+    origin: 'web',
+  }
+}
+
+/** 创建 remote 会话。返回 {sessionId, messageId}；失败抛带 HTTP/业务码的 Error。
+ *  边缘韧性（2026-08-24 故障取证，docs/diagnosis-trae-3003.md）：TLB/nginx 存在
+ *  节点间路由表漂移——同一时刻带凭据请求可能命中缺 /api/remote/v1 路由的节点，
+ *  返回**裸文本 404 "Not Found"（或 WAF 空体 403）**，而业务级错误恒为 JSON。
+ *  对这两种"裸非 JSON 拒绝"做一次短退避重试（创建失败不产生会话，幂等安全）。 */
+const EDGE_RETRY_DELAY_MS = 900
+const isEdgeSkewReject = (status, text) =>
+  (status === 404 || status === 403) && !text.trimStart().startsWith('{')
+
+/** create 会话是短请求（非流式 JSON），必须限时——边缘/本地代理死态时快速
+ *  失败而非无限挂起（2026-08-24 本地代理抽风实测）。测试可用 opts.timeoutMs 调短。 */
+const CREATE_TIMEOUT_MS = 20_000
+
+export async function createRemoteSession(baseURL, token, model, messages, { timeoutMs, reasoningEffort } = {}) {
+  const limit = Number(timeoutMs) > 0 ? Number(timeoutMs) : CREATE_TIMEOUT_MS
+  const attemptOnce = async () => {
+    const resp = await fetch(`${baseURL}/api/remote/v1/chat_sessions`, {
+      method: 'POST',
+      headers: remoteWebHeaders(token, { stream: false }),
+      body: JSON.stringify(buildRemoteCreateBody(model, messages, { reasoningEffort })),
+      signal: AbortSignal.timeout(limit),
+    })
+    return { resp, text: await resp.text() }
+  }
+  let resp
+  let text
+  try {
+    ;({ resp, text } = await attemptOnce())
+    if (isEdgeSkewReject(resp.status, text)) {
+      await new Promise((r) => setTimeout(r, EDGE_RETRY_DELAY_MS))
+      ;({ resp, text } = await attemptOnce())
+    }
+  } catch (e) {
+    if (e?.name === 'TimeoutError' || /timeout|timed?\s?out/i.test(String(e?.message ?? ''))) {
+      throw new Error(`trae remote create_session 超时（${limit}ms 无响应）——边缘/WAF 拦截或本地代理异常；可稍后重试或改用 inline 通道`)
+    }
+    throw e
+  }
+  if (resp.status >= 400) {
+    const skew = isEdgeSkewReject(resp.status, text)
+    const err = new Error(`trae remote create_session [${resp.status}]: ${text.slice(0, 300)}${skew ? '（边缘节点路由漂移/WAF 拦截，通常数分钟内自愈；可稍后重试或改用 inline 通道）' : ''}`)
+    err.status = resp.status
+    throw err
+  }
+  let data
+  try { data = JSON.parse(text) } catch {
+    throw new Error(`trae remote create_session: non-json response: ${text.slice(0, 200)}`)
+  }
+  if (data.code != null && data.code !== 0 && !data.data) {
+    const err = new Error(`trae remote create_session: ${data.code} ${data.message ?? ''}`.trim())
+    err.code = data.code
+    throw err
+  }
+  const payload = data.data ?? data
+  const sessionId = String(payload.chat_session_id ?? '')
+  const messageId = String(payload.message_id ?? '')
+  if (!sessionId || !messageId) throw new Error('trae remote create_session: missing chat_session_id/message_id')
+  return { sessionId, messageId }
+}
+
+/** 拉取事件流（调用方负责读 response.body 到 EOF）。边缘「收下不回应」时
+ *  不能永久挂起——与 inline 面（gateway.js 首字节护栏）同口径：护栏只约束
+ *  「连上却不出响应头」的窗口，响应头到达即解绑（AbortController + 手动计时，
+ *  而非 AbortSignal.timeout——后者会约束整条 SSE 长流，把正常长会话误杀）。
+ *  测试可用 opts.timeoutMs 调短。 */
+const OPEN_EVENTS_TIMEOUT_MS = 20_000
+
+export async function openRemoteEvents(baseURL, token, sessionId, messageId, { timeoutMs } = {}) {
+  const limit = Number(timeoutMs) > 0 ? Number(timeoutMs) : OPEN_EVENTS_TIMEOUT_MS
+  const inbound = new AbortController()
+  const timer = setTimeout(() => {
+    const err = new Error(`trae remote events 超时（${limit}ms 无首字节）——边缘/WAF 拦截或本地代理异常；可稍后重试或改用 inline 通道`)
+    err.name = 'TimeoutError'
+    inbound.abort(err)
+  }, limit)
+  const resp = await fetch(
+    `${baseURL}/api/remote/v1/chat_sessions/${sessionId}/events?reply_to_message_id=${messageId}`,
+    { headers: remoteWebHeaders(token), signal: inbound.signal },
+  ).finally(() => clearTimeout(timer))
+  if (resp.status >= 400) {
+    const text = await resp.text().catch(() => '')
+    const err = new Error(`trae remote events [${resp.status}]: ${text.slice(0, 300)}`)
+    err.status = resp.status
+    throw err
+  }
+  return resp
+}
+
+/** 终止会话（best-effort，绝不抛出）。 */
+export async function stopRemoteSession(baseURL, token, sessionId, messageId) {
+  if (!sessionId || !messageId) return
+  try {
+    await fetch(`${baseURL}/api/remote/v1/chat_sessions/${sessionId}/stop`, {
+      method: 'POST',
+      headers: remoteWebHeaders(token, { stream: false }),
+      body: JSON.stringify({ chat_session_id: sessionId, user_message_id: messageId }),
+      signal: AbortSignal.timeout(10_000),
+    })
+  } catch { /* best-effort */ }
+}
+
+/** 累计快照 → 增量（前缀扩展，否则回退公共前缀——镜像 Trae2api-cn
+ *  ProtocolTextAccumulator）。实现唯一归此文件；gateway.js import 并 re-export。 */
+export function cumulativeDelta(previous, current) {
+  if (typeof current !== 'string' || current === '') return ''
+  if (typeof previous !== 'string' || previous === '') return current
+  if (current.startsWith(previous)) return current.slice(previous.length)
+  if (previous.startsWith(current)) return ''
+  let i = 0
+  const limit = Math.min(previous.length, current.length)
+  while (i < limit && previous[i] === current[i]) i += 1
+  return current.slice(i)
+}
+
+/** 去尾部空白（finish summary 覆盖判定用）。 */
+function rstripEnd(s) {
+  return typeof s === 'string' ? s.replace(/\s+$/, '') : ''
+}
+
+/**
+ * remote 事件流解析器。吃 (eventName, dataObj)，产 {text?, reasoning?, usage?,
+ * actualModel?, queue?, finish?, error?}。文本语义：
+ * - 每个 plan_item id 一槽，thought=可见文本、reasoning_content=思考（累计快照差分）
+ * - finish 工具调用的 params.summary=最终答复：流结束时若可见文本未覆盖则补发
+ * - done/EOF 终结；finish_reason 恒 'stop'
+ */
+export function createRemoteEventParser() {
+  const state = {
+    slots: new Map(), // planItemId -> {thought, reasoning}
+    order: [],
+    finalSummary: '',
+    usage: null,
+    actualModel: null,
+    done: false,
+    queueSeen: false,
+  }
+  const streamedAll = () => state.order.map((id) => state.slots.get(id)?.thought ?? '').join('')
+  /** finish summary 兜底：可见文本为空→补全文；未以 summary 结尾→追加。 */
+  const summaryTail = () => {
+    const full = streamedAll()
+    if (!state.finalSummary) return ''
+    if (!full.trim()) return state.finalSummary
+    if (!rstripEnd(full).endsWith(rstripEnd(state.finalSummary))) return '\n\n' + state.finalSummary
+    return ''
+  }
+
+  return {
+    isDone: () => state.done,
+    usage: () => state.usage,
+    actualModel: () => state.actualModel,
+    /** 非流式聚合用的最终可见文本（finish summary 兜底）。 */
+    finalText: () => streamedAll() + summaryTail(),
+    handle(eventName, obj) {
+      if (!obj || typeof obj !== 'object') return {}
+      const event = typeof obj.event === 'string' && obj.event ? obj.event : eventName
+      if (event === 'error') {
+        state.done = true
+        return { error: obj }
+      }
+      if (event === 'model_config') {
+        if (typeof obj.config_name === 'string' && obj.config_name) state.actualModel = obj.config_name
+        return {}
+      }
+      if (event === 'token_usage') {
+        const u = obj.usage ?? obj
+        const num = (k) => (typeof u[k] === 'number' && Number.isFinite(u[k]) ? u[k] : null)
+        const usage = {
+          prompt_tokens: num('prompt_tokens'),
+          completion_tokens: num('completion_tokens'),
+          total_tokens: num('total_tokens'),
+        }
+        if (usage.total_tokens != null || usage.prompt_tokens != null) state.usage = usage
+        return {}
+      }
+      if (event === 'queuing' || event === 'notification') {
+        if (!state.queueSeen) {
+          state.queueSeen = true
+          return { queue: true }
+        }
+        return {}
+      }
+      if (event === 'done') {
+        state.done = true
+        const info = obj.user_message_context?.model_info
+        if (!state.actualModel && typeof info?.config_name === 'string' && info.config_name) {
+          state.actualModel = info.config_name
+        }
+        const tail = summaryTail()
+        return tail ? { text: tail, finish: 'stop' } : { finish: 'stop' }
+      }
+      if (event !== 'plan_item') return {}
+      const out = {}
+      const tci = obj.tool_call_info
+      if (tci && typeof tci === 'object' && tci.name === 'finish') {
+        const summary = tci.params?.summary
+        if (typeof summary === 'string' && summary) state.finalSummary = summary
+      }
+      const id = typeof obj.id === 'string' && obj.id ? obj.id : '(anon)'
+      if (!state.slots.has(id)) {
+        state.slots.set(id, { thought: '', reasoning: '' })
+        state.order.push(id)
+      }
+      const slot = state.slots.get(id)
+      const thoughtSnap = typeof obj.thought === 'string' ? obj.thought : ''
+      const reasoningSnap = typeof obj.reasoning_content === 'string' ? obj.reasoning_content : ''
+      const td = cumulativeDelta(slot.thought, thoughtSnap)
+      const rd = cumulativeDelta(slot.reasoning, reasoningSnap)
+      if (thoughtSnap) slot.thought = thoughtSnap
+      if (reasoningSnap) slot.reasoning = reasoningSnap
+      if (td) out.text = td
+      if (rd) out.reasoning = rd
+      return out
+    },
+  }
+}

package/scripts/capture-cache.mjs

@@ -0,0 +1,65 @@
+#!/usr/bin/env node
+/**
+ * Cache-hit reproducer: two fresh dsh sessions get the SAME user prompt
+ * verbatim (plus one second turn in the first), all pinned to the same model
+ * (deepseek-v4-pro) via session.selectModel. The bridge dump/log then shows
+ * exactly which prefix tokens survive between identical prompts.
+ *
+ * Requires: dsh web on :3080 with CODEBUDDY_BRIDGE_LOG + CODEBUDDY_BRIDGE_DUMP.
+ * Usage: node scripts/capture-cache.mjs [port]
+ */
+
+const PORT = process.argv[2] ?? '3080'
+const BASE = `http://127.0.0.1:${PORT}/api/`
+const MODEL = { provider: 'codebuddy', model: 'deepseek-v4-pro' }
+
+async function rpc(method, payload) {
+  const res = await fetch(BASE + method, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify({ type: 'client-request', rpcId: crypto.randomUUID(), method, payload }),
+  })
+  const body = await res.json().catch(() => null)
+  if (!body?.result?.ok) throw new Error(`${method} failed: ${JSON.stringify(body).slice(0, 200)}`)
+  return body.result.value
+}
+
+const sleep = (ms) => new Promise((r) => setTimeout(r, ms))
+
+async function waitIdle(sessionId, timeoutMs = 180000) {
+  const deadline = Date.now() + timeoutMs
+  await sleep(1500)
+  while (Date.now() < deadline) {
+    const { items } = await rpc('session.list', {})
+    const item = items.find((i) => i.sessionId === sessionId)
+    if (item && item.running === false) return item
+    await sleep(1000)
+  }
+  throw new Error(`session ${sessionId} still running after ${timeoutMs}ms`)
+}
+
+async function prompt(sessionId, text) {
+  await rpc('session.prompt', { sessionId, mode: 'queue', content: [{ type: 'text', text }] })
+  const item = await waitIdle(sessionId)
+  console.log(`  done: turns=${item.projections.values.sessionStats.turns} steps=${item.projections.values.sessionStats.steps}`)
+}
+
+const Q1 = '请用一句话介绍深井水循环系统的维护要点。'
+const Q2 = '再补充一句冬季防冻注意事项。'
+
+console.log('[D] session D, model=v4-pro, prompt Q1 then Q2')
+const d = await rpc('session.create', { cwd: process.cwd() })
+console.log('  session D:', d.sessionId)
+await rpc('session.selectModel', { sessionId: d.sessionId, ...MODEL })
+await prompt(d.sessionId, Q1)
+await sleep(3500) // let the title call settle
+await prompt(d.sessionId, Q2)
+
+console.log('[E] fresh session E, same model, Q1 verbatim (identical resend)')
+const e = await rpc('session.create', { cwd: process.cwd() })
+console.log('  session E:', e.sessionId)
+await rpc('session.selectModel', { sessionId: e.sessionId, ...MODEL })
+await prompt(e.sessionId, Q1)
+await sleep(3500)
+
+console.log('\ncapture done — analyze /tmp/bridge-cache.jsonl and /tmp/bridge-dump/')

package/scripts/capture-traffic.mjs

@@ -0,0 +1,83 @@
+#!/usr/bin/env node
+/**
+ * Traffic driver for bridge forensics: talks to a RUNNING `dsh web` over its
+ * loopback RPC API and produces a controlled main-chat request sequence, so
+ * the bridge log (CODEBUDDY_BRIDGE_LOG) can be classified afterwards:
+ *
+ *   session A: three short turns       → full-history resend growth + title call
+ *   session B: A's first prompt verbatim → "manual resend" instance
+ *   session C: asks for a subagent      → parent/child request streams
+ *
+ * Real gateway cost: ~8-12 tiny completions (answers are short by prompt
+ * design; title calls cap at 64 tokens). Requires: dsh web on 127.0.0.1:3080
+ * with this plugin loaded and the bridge log enabled.
+ *
+ * Usage: node scripts/capture-traffic.mjs [port]
+ */
+
+const PORT = process.argv[2] ?? '3080'
+const BASE = `http://127.0.0.1:${PORT}/api/`
+
+async function rpc(method, payload) {
+  const res = await fetch(BASE + method, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify({ type: 'client-request', rpcId: crypto.randomUUID(), method, payload }),
+  })
+  const body = await res.json().catch(() => null)
+  if (!body?.result?.ok) throw new Error(`${method} failed: ${JSON.stringify(body).slice(0, 200)}`)
+  return body.result.value
+}
+
+const sleep = (ms) => new Promise((r) => setTimeout(r, ms))
+
+/** Poll session.list until the session reports running=false. */
+async function waitIdle(sessionId, timeoutMs = 120000) {
+  const deadline = Date.now() + timeoutMs
+  await sleep(1500) // let the turn register first
+  while (Date.now() < deadline) {
+    const { items } = await rpc('session.list', {})
+    const item = items.find((i) => i.sessionId === sessionId)
+    if (item && item.running === false) return item
+    await sleep(1000)
+  }
+  throw new Error(`session ${sessionId} still running after ${timeoutMs}ms`)
+}
+
+async function prompt(sessionId, text) {
+  await rpc('session.prompt', {
+    sessionId,
+    mode: 'queue',
+    content: [{ type: 'text', text }],
+  })
+  const item = await waitIdle(sessionId)
+  const usage = item?.projections?.values?.tokenUsage ?? null
+  console.log(`  done: turns=${item.projections.values.sessionStats.turns} steps=${item.projections.values.sessionStats.steps}`
+    + ` tokenUsage=${JSON.stringify(usage)}`)
+  return item
+}
+
+const Q1 = '只回复两个字：好的'
+const Q2 = '再只回复两个字：收到'
+const Q3 = '最后只回复两个字：完成'
+
+console.log('[A] session with three short turns')
+const a = await rpc('session.create', { cwd: process.cwd() })
+console.log('  session A:', a.sessionId)
+await prompt(a.sessionId, Q1)
+await sleep(3000) // the title call fires after the first turn settles
+await prompt(a.sessionId, Q2)
+await prompt(a.sessionId, Q3)
+
+console.log('[B] new session, A’s first prompt verbatim (manual resend)')
+const b = await rpc('session.create', { cwd: process.cwd() })
+console.log('  session B:', b.sessionId)
+await prompt(b.sessionId, Q1)
+await sleep(3000)
+
+console.log('[C] new session, subagent attempt')
+const c = await rpc('session.create', { cwd: process.cwd() })
+console.log('  session C:', c.sessionId)
+await prompt(c.sessionId, '请使用 subagent 工具启动一个子代理，让它回答"1+1等于几"并报告结果。你自己不要回答。')
+
+console.log('\ncapture done — now classify the bridge log (CODEBUDDY_BRIDGE_LOG).')