dsh-tap 0.20.0

package/scripts/probe-trae-agent-v3.mjs

@@ -0,0 +1,399 @@
+#!/usr/bin/env node
+/**
+ * probe-trae-agent-v3.mjs — Trae agent 面（llm_utils_chat 的 agent_v3 function 家族）
+ * 真实凭据探针（M1：证据先于代码）。
+ *
+ * 背景（docs/diagnosis-trae-3003.md §12 + 本 goal 立项）：inline_chat 面 3003 已
+ * 6 周未愈（scene 级故障），remote 面（chat_sessions）架构性拒绝 OpenAI tools
+ * （remote-no-tools）。社区两个独立实现（dsh-connect-trae / Trae2api-cn）主张
+ * llm_utils_chat 存在「agent 用法」：接受 OpenAI 风格 tools、返回结构化
+ * tool_calls、工具由调用端本地执行。按踩坑 #22，一切以本探针实测为准。
+ *
+ * 臂位（≥6，expect 预注册；--arms 子集跑）：
+ *   A1 纯聊天（无 tools）                    —— 决策门臂 1
+ *   A2 tools + tool_choice=auto（提示触发工具）—— 决策门臂 2 前置
+ *   A3 两轮工具闭环（assistant tool_calls + role:tool 回传 → 最终回答）—— 决策门臂 2
+ *   A4 并行调用提示（期望 0..N 个 tool_calls，如实记录）
+ *   A5 reasoning_effort 字段注入（对照 A2 同工具；记录是否被忽略/拒绝）
+ *   E1 错模型名（不存在的模型 id）
+ *   E2 坏参数形态（temperature 字符串）
+ *   Q  额度差分观测（payload 级提示，默认跳过）：A1 前后各读一次
+ *      ide_user_ent_usage，比对 IDE 池 used 增量（<0.01 分辨率不可见则如实记录）
+ *
+ * 用法：
+ *   node scripts/probe-trae-agent-v3.mjs [--fn agent_chat] [--model glm-5.3]
+ *     [--arms A1,A2,A3] [--sweep] [--quota] [--raw]
+ *   --sweep：只跑 FN_SWEEP 候选 function 名扫描（每候选一发 A1 形态，间隔 3s），
+ *     用于定位「agent 用法」的真实 function 值（2001=no function config 即否决）。
+ * 决策门（goal 明文）：A1 与 A3 任一失败（含 3003）→ 停止实装，goal 以前提被证伪
+ *   收束。脚本报答案在末尾 GATE 行给出。
+ *
+ * 纪律：臂间隔 ≥2s（raw 面历史 4011 限流阈 20s 的保守化：本探针臂数少、并发恒 1）；
+ * 单账号（~/.dsh/trae-plugin-auth.json）；写操作为零（除 Q 臂只读额度接口）；
+ * 证据落 docs/probes/trae-agent-v3-<date>.jsonl；token 永不落盘（JSONL 无凭据字段）。
+ */
+import { writeFileSync, mkdirSync } from 'node:fs'
+import { homedir } from 'node:os'
+import { join, dirname } from 'node:path'
+
+import { createTraeOAuth } from '../providers/trae/oauth.js'
+import { buildChatRequest, traeOutboundHeaders } from '../providers/trae/gateway.js'
+import { readJson } from '../core/json-store.js'
+
+const args = process.argv.slice(2)
+const argOf = (n, d) => { const i = args.indexOf(n); return i >= 0 && args[i + 1] !== undefined ? args[i + 1] : d }
+const FN = argOf('--fn', 'solo_work_lite')
+const MODEL = argOf('--model', 'glm-5.3')
+const ARMS = new Set(argOf('--arms', 'A1,A2,A3,A4,A5,E1,E2').split(',').map((s) => s.trim()))
+const WITH_QUOTA = args.includes('--quota')
+const DUMP_RAW = args.includes('--raw')
+const SWEEP = args.includes('--sweep')
+const SWEEP2 = args.includes('--sweep2')
+const SWEEP_MODELS = args.includes('--sweep-models')
+// function 位候选扫描清单（--sweep）：2001=no function config 即否决该名字。
+// 依据：历史实测 inline_chat/chat_v3/solo_agent_lite/solo_work_lite 均注册
+// （docs/reverse/trae-cloud-api.md §5.1）；社区实现主张的 agent 用法 function 名
+// 未公开钉死（踩坑 #22：以实测为准）。
+const FN_SWEEP = ['agent_chat', 'agent_v3', 'solo_agent_chat', 'agent', 'chat', 'solo_agent', 'assistant_chat']
+// sweep 第二轮（--sweep2）：chat 是注册 function（4023=模型未知而非 2001=function 未注册），
+// 对它补目录真实模型名 + tools 两发，判定「chat 即 agent 用法」假设。
+const FN_SWEEP2 = ['chat', 'chat_v3']
+
+const DSH_HOME = process.env.DSH_HOME ?? join(homedir(), '.dsh')
+const AUTH_PATH = join(DSH_HOME, 'trae-plugin-auth.json')
+const MCHOST = 'https://trae-api-cn.mchost.guru'
+const SETTINGS = { traeAuthBaseURL: 'https://api.trae.cn', traeChatBaseURL: MCHOST, traeLoginHost: 'https://www.trae.cn' }
+
+const oauth = createTraeOAuth({ readAuth: () => readJson(AUTH_PATH), writeAuth: () => { /* 探针不写回 */ } })
+const cred = await oauth.resolveTraeCredential(SETTINGS)
+if (!cred) { console.error('未登录 Trae（先跑 node scripts/probe-trae-live.mjs --login）'); process.exit(2) }
+const token = String(cred.authorization).replace(/^Cloud-IDE-JWT\s+/, '')
+const authStore = readJson(AUTH_PATH)
+const device = authStore.device ?? {}
+const uid = String(authStore.account?.uid ?? '')
+const ideHeaders = (requestId) => ({
+  ...traeOutboundHeaders(device, uid ? Number(uid) : undefined, requestId),
+  Authorization: `Cloud-IDE-JWT ${token}`,
+  'X-Cloudide-Token': token,
+})
+
+// ---------------------------------------------------------------------------
+// SSE 帧摘要（原文可选 --raw 落盘；默认只落结构化摘要，token 与长文本不进证据）
+
+function summarizeSse(raw) {
+  const events = []
+  let lastEvt = null
+  for (const l of raw.split('\n')) {
+    const t = l.trim()
+    if (!t) continue
+    if (t.startsWith('event:')) { lastEvt = t.slice(6).trim(); continue }
+    if (!t.startsWith('data:')) continue
+    const data = t.slice(5).trim()
+    if (data === '[DONE]') { events.push({ ev: lastEvt ?? null, done: true }); continue }
+    let j
+    try { j = JSON.parse(data) } catch { events.push({ ev: lastEvt ?? null, nonJson: data.slice(0, 80) }); continue }
+    const ev = (typeof j.event === 'string' && j.event) ? j.event : lastEvt
+    const rec = { ev: ev ?? null }
+    if (j.code != null) rec.code = j.code
+    if (typeof j.message === 'string') rec.message = j.message.slice(0, 200)
+    if (typeof j.provider_model_name === 'string') rec.providerModel = j.provider_model_name
+    if (typeof j.response === 'string') rec.responseLen = j.response.length
+    if (typeof j.reasoning_content === 'string') rec.reasoningLen = j.reasoning_content.length
+    if (Array.isArray(j.tool_calls)) rec.toolCalls = j.tool_calls.map((c) => ({
+      index: c.index, id: c.id, hasFnCall: Boolean(c.function_call), hasFn: Boolean(c.function),
+      name: c.function_call?.name ?? c.function?.name,
+      argsLen: String(c.function_call?.arguments ?? c.function?.arguments ?? '').length,
+    }))
+    if (j.tool_call_info && typeof j.tool_call_info === 'object') {
+      rec.toolCallInfo = { name: j.tool_call_info.name, id: j.tool_call_info.tool_call_id ?? j.tool_call_info.id }
+    }
+    if (j.finish_reason) rec.finish = j.finish_reason
+    // token_usage 事件形态（dsh-connect-trae sse.ts 全字段表）：顶层即计数
+    if (ev === 'token_usage') {
+      rec.usage = {
+        prompt_tokens: j.prompt_tokens, completion_tokens: j.completion_tokens,
+        total_tokens: j.total_tokens, reasoning_tokens: j.reasoning_tokens,
+        cache_read_input_tokens: j.cache_read_input_tokens,
+      }
+    }
+    events.push(rec)
+  }
+  return events
+}
+
+/** 从摘要重建可见文本（response 为累计快照时取最后一帧）。 */
+function finalTextOf(events) {
+  let t = ''
+  for (const e of events) if (typeof e.responseLen === 'number') t = e // 只取长度不足以重建文本；真值由 --raw 提供
+  return null // 摘要模式不重建文本（文本含模型输出，证据只记长度与事件形态）
+}
+
+// ---------------------------------------------------------------------------
+
+const sleep = (ms) => new Promise((r) => setTimeout(r, ms))
+
+async function postLlmUtilsChat(payload, { fn, fnKey, timeoutMs = 120_000 } = {}) {
+  const { body, requestId } = buildChatRequest(payload, `agentv3-${Date.now().toString(36)}`, { fnKey })
+  body.function = fn
+  const t0 = Date.now()
+  const resp = await fetch(`${MCHOST}/api/agent/v3/llm_utils_chat`, {
+    method: 'POST',
+    headers: ideHeaders(requestId),
+    body: JSON.stringify(body),
+    redirect: 'follow',
+    signal: AbortSignal.timeout(timeoutMs),
+  })
+  const raw = await resp.text()
+  return { status: resp.status, raw, ms: Date.now() - t0, sentBody: body }
+}
+
+const TOOL_GET_TIME = {
+  type: 'function',
+  function: {
+    name: 'get_current_time',
+    description: '获取当前时间。当用户询问时间、日期时必须调用。',
+    parameters: { type: 'object', properties: {} },
+  },
+}
+
+// expect 预注册（跑前钉死，跑后对照）
+const EXPECT = {
+  A1: '若 function 有效：HTTP 200 + SSE 文本事件 + done；若 fn 未知：响亮错误（码入证据）',
+  A2: '接受 tools 不 400；若支持工具则出 tool_calls 事件或文本回答',
+  A3: '第二轮（带 role:tool）不报错且产出最终文本回答',
+  A4: '并行提示：0..N 个 tool_calls（如实记录，N≥2 即证实并行）',
+  A5: 'reasoning_effort 被接受（不报错）或被忽略（行为同 A2）；若 400 则证实该面拒此字段',
+  E1: '错模型名：业务错误码（预期 3003/4001/4011 家族之一）',
+  E2: '坏参数：HTTP 4xx 或 SSE error（预期 4001 unmarshal 家族）',
+}
+
+const arms = []
+
+async function runArm(label, fn, payload, { expect, fnKey, timeoutMs } = {}) {
+  const out = { arm: label, fn, expect, at: new Date().toISOString() }
+  try {
+    const r = await postLlmUtilsChat(payload, { fn, fnKey, timeoutMs })
+    out.http = r.status
+    out.ms = r.ms
+    out.sentBodyShape = {
+      model: r.sentBody.model, function: r.sentBody.function,
+      messages: r.sentBody.messages.length,
+      roles: r.sentBody.messages.map((m) => m.role),
+      assistantToolCallKey: r.sentBody.messages.find((m) => m.role === 'assistant')?.tool_calls?.[0]
+        ? Object.keys(r.sentBody.messages.find((m) => m.role === 'assistant').tool_calls[0]).join('+') : null,
+      tools: Array.isArray(r.sentBody.tools) ? r.sentBody.tools.length : 0,
+      tool_choice: r.sentBody.tool_choice ?? null,
+      reasoning_effort: r.sentBody.reasoning_effort ?? null,
+    }
+    const events = summarizeSse(r.raw)
+    out.sseBytes = r.raw.length
+    out.eventCount = events.length
+    out.errors = events.filter((e) => e.code != null).slice(0, 3)
+    out.finishReasons = [...new Set(events.map((e) => e.finish ?? e.stopReason).filter(Boolean))]
+    out.providerModels = [...new Set(events.map((e) => e.providerModel).filter(Boolean))]
+    out.toolCallEvents = events.filter((e) => e.toolCalls || e.toolCallInfo).slice(0, 6)
+    out.maxResponseLen = Math.max(0, ...events.map((e) => e.responseLen ?? 0))
+    out.usage = events.map((e) => e.usage).filter(Boolean).at(-1) ?? null
+    out.eventNames = [...new Set(events.map((e) => e.ev).filter(Boolean))]
+    if (DUMP_RAW) out.rawSse = r.raw.slice(0, 20_000)
+    out.ok = out.http === 200 && out.errors.length === 0
+  } catch (err) {
+    out.ok = false
+    out.error = String(err?.message ?? err).slice(0, 300)
+  }
+  arms.push(out)
+  console.log(`[${label}] http=${out.http ?? '-'} ok=${out.ok} events=${out.eventCount ?? 0} errors=${JSON.stringify(out.errors ?? out.error ?? null)}`)
+  return out
+}
+
+// ---------------------------------------------------------------------------
+
+console.log(`fn=${FN} model=${MODEL} arms=${[...ARMS].join(',')}${SWEEP ? ' SWEEP=' + FN_SWEEP.join('/') : ''}`)
+console.log('expect（预注册）:', JSON.stringify(EXPECT))
+
+if (SWEEP) {
+  for (const fn of FN_SWEEP) {
+    await runArm(`sweep:${fn}`, fn, {
+      model: MODEL, stream: true,
+      messages: [{ role: 'user', content: '只回复两个字：成功' }],
+    }, { expect: 'function 名注册判定：2001=no function config 即否决', timeoutMs: 60_000 })
+    await sleep(3_000)
+  }
+  const date0 = new Date().toISOString().slice(0, 10)
+  const sweepPath = join(process.cwd(), 'docs/probes', `trae-agent-v3-fnsweep-${date0}.jsonl`)
+  mkdirSync(dirname(sweepPath), { recursive: true })
+  writeFileSync(sweepPath, arms.map((a) => JSON.stringify(a)).join('\n') + '\n')
+  console.log(`证据 → ${sweepPath}`)
+  process.exit(0)
+}
+
+let quotaBefore = null
+async function quotaSnapshot(tag) {
+  const resp = await fetch(`${SETTINGS.traeAuthBaseURL}/trae/api/v2/pay/ide_user_ent_usage`, {
+    method: 'POST',
+    headers: ideHeaders(`quota-${tag}`),
+    body: JSON.stringify({ require_usage: true, req_source: 0 }),
+    signal: AbortSignal.timeout(15_000),
+  })
+  const body = await resp.json().catch(() => null)
+  const packs = body?.user_entitlement_pack_list
+  if (!Array.isArray(packs)) return { tag, error: `code ${body?.code ?? `http ${resp.status}`}` }
+  const mk = () => ({ limit: 0, used: 0 })
+  const pools = { ide: mk(), work: mk() }
+  for (const p of packs) {
+    const ep = p?.entitlement_base_info?.available_endpoint
+    const pool = ep === 1 ? pools.work : ep === 0 ? pools.ide : null
+    if (!pool) continue
+    pool.limit += p?.entitlement_base_info?.quota?.credits_limit ?? 0
+    pool.used += p?.usage?.credits_amount ?? 0
+  }
+  const r = (n) => Math.round(n * 1000) / 1000
+  return { tag, ideUsed: r(pools.ide.used), ideLimit: pools.ide.limit, workUsed: r(pools.work.used), workLimit: pools.work.limit }
+}
+
+if (WITH_QUOTA) quotaBefore = await quotaSnapshot('before')
+
+if (SWEEP2) {
+  for (const fn of FN_SWEEP2) {
+    await runArm(`sweep2:${fn}:plain`, fn, {
+      model: MODEL, stream: true,
+      messages: [{ role: 'user', content: '只回复两个字：成功' }],
+    }, { expect: '已注册 function + 真实模型名：期望 200 出文本', timeoutMs: 90_000 })
+    await sleep(3_000)
+    await runArm(`sweep2:${fn}:tools`, fn, {
+      model: MODEL, stream: true, tool_choice: 'auto',
+      messages: [{ role: 'user', content: '现在几点？请调用 get_current_time 工具获取。' }],
+      tools: [TOOL_GET_TIME],
+    }, { expect: '带 tools：接受且出 tool_calls 即证实工具面', timeoutMs: 90_000 })
+    await sleep(3_000)
+  }
+  const date0 = new Date().toISOString().slice(0, 10)
+  const p = join(process.cwd(), 'docs/probes', `trae-agent-v3-fnsweep2-${date0}.jsonl`)
+  mkdirSync(dirname(p), { recursive: true })
+  writeFileSync(p, arms.map((a) => JSON.stringify(a)).join('\n') + '\n')
+  console.log(`证据 → ${p}`)
+  process.exit(0)
+}
+
+if (ARMS.has('A1')) {
+  await runArm('A1', FN, {
+    model: MODEL, stream: true,
+    messages: [{ role: 'user', content: '只回复两个字：成功' }],
+  }, { expect: EXPECT.A1 })
+  await sleep(2_000)
+}
+
+let a2 = null
+if (ARMS.has('A2')) {
+  a2 = await runArm('A2', FN, {
+    model: MODEL, stream: true, tool_choice: 'auto',
+    messages: [{ role: 'user', content: '现在几点？请调用 get_current_time 工具获取。' }],
+    tools: [TOOL_GET_TIME],
+  }, { expect: EXPECT.A2 })
+  await sleep(2_000)
+}
+
+if (ARMS.has('A3')) {
+  // 两轮闭环（Trae2api-cn 参照形态）：assistant 历史消息的 tool_calls 用
+  // **function_call 键**（与该面 SSE 出站的 tool_calls[i].function_call 同构——
+  // 2026-10-05 A3 臂实测：OpenAI 风格 function 键被 proto 层拒绝「required
+  // field Name is not set」，说明该面把 function_call.Name 映射到 proto 字段）。
+  // tool_calls 形态以 A2 实测为准（A2 若未出 tool_calls，则构造等价形态——
+  // 该臂目的是验证「role:tool 回传被接受且产出最终回答」这一协议面）。
+  const callId = a2?.toolCallEvents?.[0]?.toolCalls?.[0]?.id
+    ?? a2?.toolCallEvents?.[0]?.toolCallInfo?.id
+    ?? 'call_probe_time_1'
+  const callName = a2?.toolCallEvents?.[0]?.toolCalls?.[0]?.name
+    ?? a2?.toolCallEvents?.[0]?.toolCallInfo?.name
+    ?? 'get_current_time'
+  await runArm('A3', FN, {
+    model: MODEL, stream: true, tool_choice: 'auto',
+    messages: [
+      { role: 'user', content: '现在几点？请调用 get_current_time 工具获取。' },
+      { role: 'assistant', content: '', tool_calls: [{ id: callId, type: 'function', function: { name: callName, arguments: '{}' } }] },
+      { role: 'tool', tool_call_id: callId, name: callName, content: '2026-10-05 14:30:00 CST' },
+    ],
+    tools: [TOOL_GET_TIME],
+  }, { expect: EXPECT.A3, fnKey: 'function_call' })
+  await sleep(2_000)
+}
+
+if (ARMS.has('A4')) {
+  await runArm('A4', FN, {
+    model: MODEL, stream: true, tool_choice: 'auto',
+    messages: [{ role: 'user', content: '请同时调用两次 get_current_time（并行），然后告诉我时间。' }],
+    tools: [TOOL_GET_TIME],
+  }, { expect: EXPECT.A4 })
+  await sleep(2_000)
+}
+
+if (ARMS.has('A5')) {
+  await runArm('A5', FN, {
+    model: MODEL, stream: true, tool_choice: 'auto', reasoning_effort: 'high',
+    messages: [{ role: 'user', content: '现在几点？请调用 get_current_time 工具获取。' }],
+    tools: [TOOL_GET_TIME],
+  }, { expect: EXPECT.A5 })
+  await sleep(2_000)
+}
+
+if (ARMS.has('E1')) {
+  await runArm('E1', FN, {
+    model: 'no-such-model-agentv3-probe', stream: true,
+    messages: [{ role: 'user', content: 'hi' }],
+  }, { expect: EXPECT.E1 })
+  await sleep(2_000)
+}
+
+if (ARMS.has('E2')) {
+  await runArm('E2', FN, {
+    model: MODEL, stream: true, temperature: 'hot',
+    messages: [{ role: 'user', content: 'hi' }],
+  }, { expect: EXPECT.E2 })
+}
+
+let quotaAfter = null
+if (WITH_QUOTA) { await sleep(2_000); quotaAfter = await quotaSnapshot('after') }
+
+// M1 轮次 A6：多模型对照（单发/模型，证明 agent 面模型路由非钉死）。
+// 历史事实：raw 面「模型路由被 function 位钉死」（trae-cloud-api.md §5.1，
+// solo_work_lite 恒 glm-5.2）；timing_cost.provider_model_name 是唯一真值源。
+if (ARMS.has('A6')) {
+  for (const m of ['kimi-k2.6', 'DeepSeek-V4-Flash']) {
+    await runArm(`A6:${m}`, FN, {
+      model: m, stream: true,
+      messages: [{ role: 'user', content: '只回复两个字：成功' }],
+    }, { expect: 'provider_model_name 跟随请求模型即证实真路由；钉死则恒 glm-5.2' })
+    await sleep(2_000)
+  }
+}
+
+// ---------------------------------------------------------------------------
+
+const gate = {
+  A1: arms.find((a) => a.arm === 'A1'),
+  A3: arms.find((a) => a.arm === 'A3'),
+}
+const gatePass = (gate.A1 ? gate.A1.ok : null) && (gate.A3 ? gate.A3.ok : null)
+
+const evidence = {
+  probe: 'trae-agent-v3',
+  at: new Date().toISOString(),
+  fn: FN,
+  model: MODEL,
+  accountUidMasked: uid.slice(0, 3) + '***',
+  expect: EXPECT,
+  arms,
+  quota: WITH_QUOTA ? {
+    before: quotaBefore, after: quotaAfter,
+    deltaIde: quotaBefore?.ideUsed != null && quotaAfter?.ideUsed != null ? Math.round((quotaAfter.ideUsed - quotaBefore.ideUsed) * 1000) / 1000 : null,
+    deltaWork: quotaBefore?.workUsed != null && quotaAfter?.workUsed != null ? Math.round((quotaAfter.workUsed - quotaBefore.workUsed) * 1000) / 1000 : null,
+  } : undefined,
+  gate: gatePass == null ? 'SKIPPED（A1/A3 未跑）' : gatePass ? 'PASS（A1 与 A3 均成功）' : 'FAIL（决策门臂失败，停止实装）',
+}
+
+const date = new Date().toISOString().slice(0, 10)
+const outPath = join(process.cwd(), 'docs/probes', `trae-agent-v3-${date}.jsonl`)
+mkdirSync(dirname(outPath), { recursive: true })
+writeFileSync(outPath, arms.map((a) => JSON.stringify(a)).join('\n') + '\n' + JSON.stringify({ summary: evidence.gate, quota: evidence.quota ?? null }) + '\n')
+console.log(`GATE: ${evidence.gate}`)
+console.log(`证据 → ${outPath}`)

package/scripts/probe-trae-efforts.mjs

@@ -0,0 +1,149 @@
+#!/usr/bin/env node
+/**
+ * probe-trae-efforts.mjs — Trae 通道 `reasoning_effort` 方言判定（真实上游，小额）。
+ *
+ * 为什么需要它：Trae 路由的 `compat` 一直没开（cordis.patch.yml:369-371 注明
+ * "reasoning_effort dialect unverified"），所以该通道在宿主「推理等级」里从不出档。
+ * 目录（state.vscdb / 远端模型表）其实**声明了逐模型档位**
+ * （`reasoning_effort_config.options`，实测 Doubao-Seed-2.1-Pro = ["light","high"]，
+ * default_level "high"），只要方言可用就能同 Qoder 一样自动出档。
+ *
+ * 判据（踩坑 #42 纪律）：自适应思考下推理长度是非单调弱信号 ⇒ 必须**多臂多采样**，
+ * 只有"每个重复都落在基线的同一侧"才判 DIRECTION-CONSISTENT；否则 INCONCLUSIVE，
+ * 不下结论、不改出站。
+ *
+ * 用法：
+ *   node scripts/probe-trae-efforts.mjs [--model <id>] [--levels light,high]
+ *        [--repeat 2] [--prompt "…"]
+ * 证据 → docs/probes/trae-efforts-<ts>.json（不含任何令牌）
+ */
+import { writeFileSync, mkdirSync } from 'node:fs'
+import { homedir } from 'node:os'
+import { join } from 'node:path'
+
+import { createTraeOAuth } from '../providers/trae/oauth.js'
+import { buildChatRequest, createTraeStreamParser, TRAE_APP_ID, TRAE_IDE_VERSION, TRAE_IDE_VERSION_CODE } from '../providers/trae/gateway.js'
+import { readJson } from '../core/json-store.js'
+
+const DSH_HOME = process.env.DSH_HOME ?? join(homedir(), '.dsh')
+const AUTH_PATH = join(DSH_HOME, 'trae-plugin-auth.json')
+const SETTINGS = {
+  traeAuthBaseURL: process.env.TRAE_AUTH_BASE ?? 'https://api.trae.cn',
+  traeChatBaseURL: process.env.TRAE_CHAT_BASE ?? 'https://trae-api-cn.mchost.guru',
+  traeLoginHost: process.env.TRAE_LOGIN_HOST ?? 'https://www.trae.cn',
+}
+
+const args = process.argv.slice(2)
+const argOf = (name, dflt) => {
+  const i = args.indexOf(name)
+  return i >= 0 && args[i + 1] !== undefined ? args[i + 1] : dflt
+}
+const MODEL = argOf('--model', 'Doubao-Seed-2.1-Pro')
+const LEVELS = String(argOf('--levels', 'light,high')).split(',').map((s) => s.trim()).filter(Boolean)
+const REPEAT = Number(argOf('--repeat', '2')) || 2
+const PROMPT = argOf('--prompt', '一个笼子里有鸡和兔共 10 只，脚共 28 只。鸡兔各几只？只给最终数字。')
+
+const oauth = createTraeOAuth({ readAuth: () => readJson(AUTH_PATH), writeAuth: () => {} })
+const cred = await oauth.resolveTraeCredential(SETTINGS)
+if (!cred) {
+  console.error('未登录 Trae（先跑 node scripts/probe-trae-live.mjs --login）')
+  process.exit(2)
+}
+const token = String(cred.authorization).replace(/^Cloud-IDE-JWT\s+/, '')
+const authStore = readJson(AUTH_PATH)
+
+async function arm(label, level) {
+  const payload = {
+    model: MODEL, stream: true,
+    messages: [{ role: 'user', content: PROMPT }],
+    ...(level === undefined ? {} : { reasoning_effort: level }),
+  }
+  const { body, requestId } = buildChatRequest(payload, 'probe-efforts')
+  const res = await fetch(`${SETTINGS.traeChatBaseURL}/api/agent/v3/llm_utils_chat`, {
+    method: 'POST',
+    headers: {
+      'Content-Type': 'application/json',
+      Accept: 'text/event-stream',
+      Authorization: `Cloud-IDE-JWT ${token}`,
+      'X-Cloudide-Token': token,
+      'x-ide-token': token,
+      'x-app-id': TRAE_APP_ID,
+      'x-ide-version': TRAE_IDE_VERSION,
+      'x-ide-version-code': TRAE_IDE_VERSION_CODE,
+      'x-request-id': requestId,
+      ...(authStore.account?.uid ? { 'x-uid': String(authStore.account.uid) } : {}),
+      ...(authStore.device?.deviceId ? { 'x-device-id': authStore.device.deviceId } : {}),
+      ...(authStore.device?.machineId ? { 'x-machine-id': authStore.device.machineId } : {}),
+    },
+    body: JSON.stringify(body),
+  })
+  const raw = await res.text()
+  const parser = createTraeStreamParser()
+  let reasoning = ''
+  let text = ''
+  let inband = null
+  for (const line of raw.split('\n')) {
+    const s = line.trim()
+    if (!s.startsWith('data:')) continue
+    const data = s.slice(5).trim()
+    if (!data || data === '[DONE]') continue
+    let chunk
+    try { chunk = JSON.parse(data) } catch { continue }
+    const ev = parser.handle(null, chunk)
+    if (ev.error) inband = ev.error
+    if (typeof ev.reasoning === 'string') reasoning += ev.reasoning
+    if (typeof ev.text === 'string') text += ev.text
+  }
+  const usage = parser.usage()
+  return {
+    arm: label,
+    level: level ?? null,
+    http: res.status,
+    reasoningChars: reasoning.length,
+    textChars: text.length,
+    reasoningHead: reasoning.slice(0, 40),
+    providerModel: parser.providerModel(),
+    finish: parser.finish(),
+    usage,
+    inband,
+    bytes: raw.length,
+    rawHead: raw.slice(0, 700),
+  }
+}
+
+console.log(`model=${MODEL} levels=${JSON.stringify(LEVELS)} repeat=${REPEAT}`)
+const arms = []
+for (let i = 0; i < REPEAT; i++) {
+  arms.push(await arm(`baseline#${i + 1}`, undefined))
+  for (const level of LEVELS) arms.push(await arm(`${level}#${i + 1}`, level))
+}
+
+console.log('\narm                 http  reasoning  text  providerModel         finish  inband')
+for (const a of arms) {
+  console.log(`${a.arm.padEnd(18)} ${String(a.http).padEnd(5)} ${String(a.reasoningChars).padStart(9)} ${String(a.textChars).padStart(5)}  ${String(a.providerModel ?? '-').padEnd(20)} ${String(a.finish ?? '-').padEnd(7)} ${a.inband ? JSON.stringify(a.inband).slice(0, 60) : ''}`)
+}
+
+const mean = (xs) => (xs.length ? xs.reduce((a, b) => a + b, 0) / xs.length : 0)
+const baseVals = arms.filter((a) => a.level === null).map((a) => a.reasoningChars)
+const baseMean = mean(baseVals)
+console.log(`\nbaseline reasoning 均值 ${Math.round(baseMean)}（n=${baseVals.length}）`)
+const verdicts = {}
+for (const level of LEVELS) {
+  const vals = arms.filter((a) => a.level === level).map((a) => a.reasoningChars)
+  const m = mean(vals)
+  const allAbove = vals.every((v) => v > baseMean)
+  const allBelow = vals.every((v) => v < baseMean)
+  const verdict = allAbove || allBelow ? 'DIRECTION-CONSISTENT' : 'INCONCLUSIVE'
+  verdicts[level] = { vals, mean: Math.round(m), verdict }
+  console.log(`  ${level.padEnd(8)} 均值 ${String(Math.round(m)).padStart(5)} 各次 ${JSON.stringify(vals)} → ${verdict}`)
+}
+
+mkdirSync('docs/probes', { recursive: true })
+const out = `docs/probes/trae-efforts-${Date.now()}.json`
+writeFileSync(out, JSON.stringify({
+  at: new Date().toISOString(),
+  model: MODEL, levels: LEVELS, repeat: REPEAT, prompt: PROMPT,
+  baselineMean: Math.round(baseMean), verdicts, arms,
+}, null, 2))
+console.log('\n证据 →', out)
+console.log('判读纪律：DIRECTION-CONSISTENT 才说明该档位**可能**改变上游思考量；INCONCLUSIVE 不下结论（自适应非单调 + 小样本）。')