dsh-tap 0.20.0

package/scripts/probe-codebuddy-efforts.mjs

@@ -0,0 +1,355 @@
+#!/usr/bin/env node
+/**
+ * probe-codebuddy-efforts.mjs — CodeBuddy 通道「模型思考强度」现状探测。
+ *
+ * 目的：cordis.patch.yml 的 `reasoningEfforts` 档位表是静态清单的一部分
+ * （网关 /v3/config 只给**默认档位**不给档位清单），所以要更新档位表只能实测。
+ * 本脚本一次跑完两件事：
+ *
+ *   1) catalog 面：GET /v3/config，逐模型打印网关自述的 reasoning 声明
+ *      （`reasoning.effort` 默认档 + 是否有该键），作为"档位表该不该有"的
+ *      第一手依据。
+ *   2) chat 面：对每个目标模型逐档发一次流式请求（omit/off/low/medium/high/max），
+ *      记录 HTTP 结果、reasoning_content 字符数、content 字符数、finish_reason、
+ *      响应回显 model 与 usage。判据：
+ *        - 某档 HTTP 报错 / 带内错误帧 → 该档**网关不接受**（档位表不应含它）；
+ *        - omit（不传参）就有 reasoning_content → 模型"无参也会想"，
+ *          档位表不该给 `off`；
+ *        - omit 无 reasoning_content 而 off 也无 → `off` = 省略参数语义成立；
+ *        - 各档长度**非单调**属正常（推理长度自适应，见 patch 头注释）。
+ *
+ * 证据落盘：docs/probes/codebuddy-efforts-<YYYY-MM-DD>.json（预注册 expect 一并写入）。
+ *
+ * 用法：
+ *   node scripts/probe-codebuddy-efforts.mjs --catalog         # 只打目录面（快）
+ *   node scripts/probe-codebuddy-efforts.mjs                   # 目录 + 全档矩阵
+ *   node scripts/probe-codebuddy-efforts.mjs --models glm-5.3,kimi-k2.8-preview
+ *   node scripts/probe-codebuddy-efforts.mjs --levels omit,off,high
+ *   node scripts/probe-codebuddy-efforts.mjs --interval 800    # 每发间隔 ms（默认 600）
+ *
+ * 凭据：CODEBUDDY_API_KEY（env）或 ~/.dsh/.credentials.yaml（与插件同解析顺序）。
+ * 红线：单账号、只读优先、发问间隔 ≥600ms（STATE.md「探测纪律」）。
+ */
+
+import { readFileSync, writeFileSync, existsSync, mkdirSync } from 'node:fs'
+import { homedir } from 'node:os'
+import { join, dirname } from 'node:path'
+import { fileURLToPath } from 'node:url'
+
+const ROOT = dirname(dirname(fileURLToPath(import.meta.url)))
+const BASE = 'https://copilot.tencent.com'
+const PATCH_FILE = join(ROOT, 'cordis.patch.yml')
+
+// 目录面方言（catalog.js / verify-models 同款）：CLI 形态 UA + x-api-key。
+const CATALOG_HEADERS = (key) => ({
+  accept: 'application/json',
+  'x-api-key': key,
+  'user-agent': 'CLI/unknown CodeBuddy/2.136.0',
+  'x-product': 'SaaS',
+})
+
+// 聊天面方言：与 cordis.patch.yml 的 codebuddy provider 头一致（不含 Authorization，
+// 由 probe 逐请求拼 ck_ key）。
+const CHAT_HEADERS = {
+  'User-Agent': 'CodeBuddyCode/1.0',
+  'X-IDE-Type': 'CLI',
+  'X-IDE-Name': 'CLI',
+  'X-IDE-Version': '2.133.1',
+  'X-Product-Version': '2.133.1',
+  'X-Requested-With': 'XMLHttpRequest',
+  'X-Private-Data': 'false',
+  'Content-Type': 'application/json',
+}
+
+/**
+ * 档位 → 线值；omit = 不传 reasoning_effort（基线段）；off = 插件现行语义
+ * （off 的线值是 null ⇒ 省略参数）。offwire/alias 臂用来判定"目录声明
+ * canDisableThinking 的模型到底该发什么线值才能真的关思考"。
+ */
+const LEVEL_WIRE = {
+  omit: undefined,
+  off: null,
+  low: 'low',
+  medium: 'medium',
+  high: 'high',
+  max: 'max',
+  // 别名/关闭思考候选拼写（判定用，不是候选档位表的默认成员）
+  offwire: 'off',
+  none: 'none',
+  minimal: 'minimal',
+  disabled: 'disabled',
+  auto: 'auto',
+  xhigh: 'xhigh',
+}
+
+const argv = process.argv.slice(2)
+const argValue = (name) => {
+  const i = argv.indexOf(name)
+  return i === -1 ? null : argv[i + 1]
+}
+const CATALOG_ONLY = argv.includes('--catalog')
+const INTERVAL = Number(argValue('--interval') ?? 600)
+const REPEAT = Math.max(1, Number(argValue('--repeat') ?? 1))
+const SUMMARY = argValue('--summarize')
+const OUT = argValue('--out') ?? join(ROOT, 'docs', 'probes', `codebuddy-efforts-${new Date().toISOString().slice(0, 10)}.json`)
+
+/** 预注册预期（跑前写死，跑后逐条判定命中/推翻）。 */
+const EXPECT = {
+  P1: '无参（omit）就产 reasoning_content 的模型：kimi-k2.7/k3/k3-1/minimax-m2.7 —— 这四家档位表不含 off',
+  P2: 'off/low/medium/high/max 五种拼写对思考型模型全部 HTTP 200 可接受（不出现 400/11102 类拒绝）',
+  P3: '目录新 CLI 模型（glm-5.3 / glm-5.3-flash / kimi-k2.8-preview / hy3-x / hy4-preview / deepseek-v4.1-flash）同样接受五档拼写',
+  P4: '各档 reasoning 长度自适应、非严格单调（off < max 不必然成立）',
+  P5: '目录 reasoning 只给默认档，不含档位清单（即档位表无法从目录派生）——2026-09-22 复查已被推翻：新模型带 supportedEfforts/canDisableThinking/defaultEffort',
+}
+
+function loadApiKey() {
+  if (process.env.CODEBUDDY_API_KEY) return process.env.CODEBUDDY_API_KEY
+  const credFile = join(homedir(), '.dsh', '.credentials.yaml')
+  if (existsSync(credFile)) {
+    const m = readFileSync(credFile, 'utf8').match(/^\s*CODEBUDDY_API_KEY:\s*["']?([^"'\s]+)["']?\s*$/m)
+    if (m) return m[1]
+  }
+  return null
+}
+
+/** patch 静态清单的 id 顺序（YAML 行扫描，布局变了就要改这里）。 */
+function parsePatchIds() {
+  const ids = []
+  let inModels = false
+  for (const line of readFileSync(PATCH_FILE, 'utf8').split('\n')) {
+    if (/^\s*models:\s*$/.test(line)) { inModels = true; continue }
+    if (!inModels) continue
+    const m = line.match(/^\s*-\s*id:\s*(\S+)\s*$/)
+    if (m) { ids.push(m[1]); continue }
+    if (/^\s{0,10}\S/.test(line) && !/^\s{10,}/.test(line) && line.trim()) break
+  }
+  return ids
+}
+
+async function fetchCatalog(key) {
+  const res = await fetch(`${BASE}/v3/config`, { headers: CATALOG_HEADERS(key) })
+  const body = await res.json().catch(() => null)
+  if (!res.ok || body?.code !== 0) {
+    throw new Error(`catalog HTTP ${res.status} code=${body?.code} ${body?.msg ?? ''}`)
+  }
+  const data = body.data ?? {}
+  const cliEnabled = new Set((data.agents ?? []).find((a) => a.name === 'cli')?.models ?? [])
+  return {
+    fetchedAt: new Date().toISOString(),
+    models: (data.models ?? [])
+      .filter((m) => typeof m?.id === 'string')
+      .map((m) => ({
+        id: m.id,
+        name: m.name ?? m.id,
+        maxInputTokens: m.maxInputTokens ?? null,
+        maxOutputTokens: m.maxOutputTokens ?? null,
+        images: m.supportsImages === true,
+        cli: cliEnabled.has(m.id),
+        reasoning: m.reasoning ?? null,
+      })),
+  }
+}
+
+const sleep = (ms) => new Promise((r) => setTimeout(r, ms))
+
+/** 单档单发：流式聚合 reasoning/content 长度 + 响应回显 model + usage。 */
+async function probeArm(id, level, key) {
+  const started = Date.now()
+  const body = {
+    model: id,
+    messages: [{ role: 'user', content: '9.11和9.9哪个大？先想清楚再给一句话结论。' }],
+    max_tokens: 400,
+    stream: true,
+  }
+  const wire = LEVEL_WIRE[level]
+  if (wire) body.reasoning_effort = wire
+  let res
+  try {
+    res = await fetch(`${BASE}/v2/chat/completions`, {
+      method: 'POST',
+      headers: { ...CHAT_HEADERS, Authorization: `Bearer ${key}` },
+      body: JSON.stringify(body),
+    })
+  } catch (err) {
+    return { level, error: `network: ${err.message}`, ms: Date.now() - started }
+  }
+  if (!res.ok) {
+    const text = await res.text().catch(() => '')
+    return { level, http: res.status, error: text.slice(0, 200), ms: Date.now() - started }
+  }
+  let buf = ''
+  let reasonChars = 0
+  let contentChars = 0
+  let finish = null
+  let echoModel = null
+  let usage = null
+  let streamErr = null
+  let chunks = 0
+  const dec = new TextDecoder()
+  try {
+    const reader = res.body.getReader()
+    for (;;) {
+      const { done, value } = await reader.read()
+      if (done) break
+      buf += dec.decode(value, { stream: true })
+      let nl
+      while ((nl = buf.indexOf('\n')) >= 0) {
+        const line = buf.slice(0, nl).trim()
+        buf = buf.slice(nl + 1)
+        if (!line.startsWith('data:')) continue
+        const payload = line.slice(5).trim()
+        if (payload === '[DONE]') continue
+        let j
+        try { j = JSON.parse(payload) } catch { continue }
+        chunks++
+        if (j.error || (j.code != null && j.code !== 0)) streamErr = payload.slice(0, 200)
+        if (typeof j.model === 'string') echoModel = j.model
+        if (j.usage) usage = j.usage
+        const ch = j.choices?.[0]
+        if (ch?.delta?.reasoning_content) reasonChars += ch.delta.reasoning_content.length
+        if (ch?.delta?.content) contentChars += ch.delta.content.length
+        if (ch?.finish_reason) finish = ch.finish_reason
+      }
+      if (reasonChars + contentChars > 4000) break
+    }
+  } catch {
+    // 部分流也是证据
+  }
+  return {
+    level,
+    http: res.status,
+    reasonChars,
+    contentChars,
+    finish,
+    echoModel,
+    streamErr,
+    chunks,
+    reasoningTokens: usage?.completion_tokens_details?.reasoning_tokens ?? null,
+    totalTokens: usage?.total_tokens ?? null,
+    ms: Date.now() - started,
+  }
+}
+
+/**
+ * 证据 → 档位表决策（--summarize <evidence.json>）。
+ *
+ * 判据（与 patch 头注释同一套）：
+ *  1. 全臂 11102 → 目录里有、/v2 不可路由 → **不进静态清单**（routing.md R-R3）；
+ *  2. 某档 HTTP/带内报错 → 该档不进表（11150 = invalid_reasoning_effort）；
+ *  3. off 当且仅当**省略参数**（omit/off 两臂，等价请求）实测不产
+ *     reasoning_content —— 否则目录/宿主选择器上的 "Off" 就是个假开关；
+ *  4. 其余实测可接受档位全部进表（顺序按 EFFORT_TIER_ORDER）。
+ */
+if (SUMMARY) {
+  const ev = JSON.parse(readFileSync(SUMMARY, 'utf8'))
+  const median = (xs) => {
+    const s = [...xs].sort((a, b) => a - b)
+    return s.length ? s[Math.floor(s.length / 2)] : null
+  }
+  const byModel = new Map()
+  for (const arm of ev.arms) {
+    if (!byModel.has(arm.id)) byModel.set(arm.id, [])
+    byModel.get(arm.id).push(arm)
+  }
+  const TARGET_LEVELS = ['low', 'medium', 'high', 'max']
+  console.log(`# ${SUMMARY}`)
+  console.log(`# catalog fetchedAt=${ev.catalog?.fetchedAt ?? '?'}`)
+  console.log('')
+  for (const [id, arms] of byModel) {
+    const errors = arms.filter((a) => a.error)
+    const codes = [...new Set(errors.map((a) => (/"(?:code)":\s*(\d+)/.exec(a.error ?? '') ?? [])[1]).filter(Boolean))]
+    const notRouted = errors.length === arms.length && codes.includes('11102')
+    const omitArms = arms.filter((a) => (a.level === 'omit' || a.level === 'off') && !a.error && !a.streamErr)
+    const baseline = omitArms.length ? median(omitArms.map((a) => a.reasonChars)) : null
+    const detail = arms
+      .filter((a) => a.level !== 'omit' && a.level !== 'off')
+      .map((a) => `${a.level}=${a.error ? `ERR${(/"(?:code)":\s*(\d+)/.exec(a.error) ?? [])[1] ?? '?'}` : a.reasonChars}`)
+      .join(' ')
+    if (notRouted) {
+      console.log(`${id.padEnd(20)} omit=${baseline ?? '-'} ${detail}`)
+      console.log(`${' '.repeat(20)} → 不可路由（11102）：不进静态清单`)
+      continue
+    }
+    const accepted = TARGET_LEVELS.filter((level) => arms.some((a) => a.level === level && !a.error && !a.streamErr))
+    const rejected = TARGET_LEVELS.filter((level) => !accepted.includes(level))
+    const tiers = [...(baseline === 0 ? ['off'] : []), ...accepted]
+    console.log(`${id.padEnd(20)} omit=${baseline ?? '-'} ${detail}`)
+    console.log(`${' '.repeat(20)} → 档位表 ${JSON.stringify(tiers)}${baseline === 0 ? '' : `（省略参数会思考 → 不给 off）`}${rejected.length ? `｜拒绝档位 ${rejected.join('/')}` : ''}`)
+  }
+  process.exit(0)
+}
+
+const key = loadApiKey()
+if (!key) {
+  console.error('CODEBUDDY_API_KEY not found (env or ~/.dsh/.credentials.yaml)')
+  process.exit(1)
+}
+
+console.log('== catalog 面：GET /v3/config ==')
+const catalog = await fetchCatalog(key)
+console.log(`fetchedAt=${catalog.fetchedAt} models=${catalog.models.length}`)
+console.log('id'.padEnd(26), 'cli', 'ctx/max'.padEnd(18), 'reasoning')
+for (const m of catalog.models) {
+  console.log(
+    m.id.padEnd(26),
+    (m.cli ? ' ✓ ' : '   '),
+    `${m.maxInputTokens ?? '-'}/${m.maxOutputTokens ?? '-'}`.padEnd(18),
+    m.reasoning ? JSON.stringify(m.reasoning) : '(none)',
+  )
+}
+const withReasoning = catalog.models.filter((m) => m.reasoning)
+console.log(`\nreasoning 键存在：${withReasoning.length}/${catalog.models.length}`)
+const effortValues = [...new Set(withReasoning.map((m) => JSON.stringify(m.reasoning?.effort ?? m.reasoning?.defaultEffort ?? null)))]
+console.log(`reasoning 默认档取值集合：${effortValues.join(', ')}`)
+const tierListLike = withReasoning.filter((m) => Array.isArray(m.reasoning?.supportedEfforts) || Array.isArray(m.reasoning?.efforts) || Array.isArray(m.reasoning?.levels))
+console.log(`目录带档位清单的模型：${tierListLike.length}${tierListLike.length ? ' → ' + tierListLike.map((m) => `${m.id}=${JSON.stringify(m.reasoning.supportedEfforts ?? m.reasoning.efforts ?? m.reasoning.levels)}`).join(' ') : ''}（P5 判定依据）`)
+const disableLike = withReasoning.filter((m) => 'canDisableThinking' in (m.reasoning ?? {}))
+console.log(`目录带 canDisableThinking 的模型：${disableLike.length}${disableLike.length ? ' → ' + disableLike.map((m) => `${m.id}=${m.reasoning.canDisableThinking}`).join(' ') : ''}`)
+
+if (CATALOG_ONLY && !SUMMARY) {
+  mkdirSync(dirname(OUT), { recursive: true })
+  writeFileSync(OUT, JSON.stringify({ expect: EXPECT, catalog, arms: [] }, null, 2))
+  console.log(`\n证据落盘：${OUT}`)
+  process.exit(0)
+}
+
+const patchIds = parsePatchIds()
+const explicit = argValue('--models')
+const targets = explicit
+  ? explicit.split(',').map((s) => s.trim()).filter(Boolean)
+  : [...new Set([...patchIds, ...catalog.models.filter((m) => m.cli).map((m) => m.id)])]
+const levels = (argValue('--levels') ?? 'omit,off,low,medium,high,max').split(',').map((s) => s.trim()).filter(Boolean)
+for (const l of levels) {
+  if (!(l in LEVEL_WIRE)) {
+    console.error(`未知档位 ${l}（可选 ${Object.keys(LEVEL_WIRE).join('/')}）`)
+    process.exit(1)
+  }
+}
+
+console.log(`\n== chat 面：${targets.length} 模型 × ${levels.length} 档（间隔 ${INTERVAL}ms）==`)
+const arms = []
+for (const id of targets) {
+  console.log(`\n${id}`)
+  const catalogEntry = catalog.models.find((m) => m.id === id) ?? null
+  for (const level of levels) {
+    for (let rep = 0; rep < REPEAT; rep++) {
+      const label = REPEAT > 1 ? `${level}#${rep + 1}` : level
+      process.stdout.write(`  ${label.padEnd(10)}`)
+      const r = await probeArm(id, level, key)
+      if (r.error) console.log(` ✗ ${r.error}`)
+      else {
+        const sentinel = r.echoModel && r.echoModel !== id ? ` ⚠echo=${r.echoModel}` : ''
+        console.log(
+          ` reason=${String(r.reasonChars).padStart(5)} content=${String(r.contentChars).padStart(4)}` +
+          ` finish=${(r.finish ?? '-').padEnd(14)} ${String(r.ms).padStart(6)}ms${r.streamErr ? ` ⚠${r.streamErr}` : ''}${sentinel}`,
+        )
+      }
+      arms.push({ id, catalogReasoning: catalogEntry?.reasoning ?? null, cli: catalogEntry?.cli ?? null, rep: rep + 1, ...r })
+      await sleep(INTERVAL)
+    }
+  }
+}
+
+mkdirSync(dirname(OUT), { recursive: true })
+writeFileSync(OUT, JSON.stringify({ expect: EXPECT, catalog, arms }, null, 2))
+console.log(`\n证据落盘：${OUT}（${arms.length} 臂）`)

package/scripts/probe-codebuddy-tier-wiring.mjs

@@ -0,0 +1,244 @@
+#!/usr/bin/env node
+/**
+ * probe-codebuddy-tier-wiring.mjs — 「思考强度」全链路**真实上游**联调。
+ *
+ * 分工：verify-bridge [17] 是离线 mock 回归（可随时跑）；本脚本是真实联调，
+ * 证明线上真这么走，跑一次即可、不进 CI。
+ *
+ * 拓扑：本脚本 → 插件桥（临时 DSH_HOME，不碰用户 settings.yaml / 文件层 / 令牌）
+ *       → 本地捕获代理（原样转发并记录出站体）→ https://copilot.tencent.com
+ *
+ * 断言（每条都有真实字节或真实响应作依据）：
+ *   [1] 真实 /v3/config → model-sync 成功；model-list 的 `efforts` 档位表与
+ *       目录声明一致（glm-5.3-flash/kimi-k2.8-preview = low/high/max、
+ *       hy4-preview = high；legacy 模型走 patch 静态表；无表模型不出现）
+ *   [2] settings.yaml 镜像（宿主 Model/Effort 选择器的数据源）带 reasoningEfforts
+ *   [3] 真实 chat：捕获代理看到的出站体含被注入的 reasoning_effort，且上游 200 出正文
+ *   [4] 负例：未声明档位 / 无表模型 → 出站体不带 reasoning_effort
+ *
+ * 用法：
+ *   node scripts/probe-codebuddy-tier-wiring.mjs                 # 默认全套
+ *   node scripts/probe-codebuddy-tier-wiring.mjs --skip-chat     # 只验目录/镜像接线
+ *
+ * 凭据：CODEBUDDY_API_KEY（env）或 ~/.dsh/.credentials.yaml（与插件同解析顺序）。
+ */
+
+import { createServer } from 'node:http'
+import { Readable } from 'node:stream'
+import { mkdtempSync, writeFileSync, readFileSync, existsSync, mkdirSync } from 'node:fs'
+import { tmpdir, homedir } from 'node:os'
+import { join, dirname } from 'node:path'
+import { fileURLToPath } from 'node:url'
+
+const ROOT = dirname(dirname(fileURLToPath(import.meta.url)))
+const UPSTREAM = 'https://copilot.tencent.com'
+const argv = process.argv.slice(2)
+const SKIP_CHAT = argv.includes('--skip-chat')
+const OUT = join(ROOT, 'docs', 'probes', `codebuddy-tier-wiring-${new Date().toISOString().slice(0, 10)}.json`)
+
+function loadApiKey() {
+  if (process.env.CODEBUDDY_API_KEY) return process.env.CODEBUDDY_API_KEY
+  const credFile = join(homedir(), '.dsh', '.credentials.yaml')
+  if (existsSync(credFile)) {
+    const m = readFileSync(credFile, 'utf8').match(/^\s*CODEBUDDY_API_KEY:\s*["']?([^"'\s]+)["']?\s*$/m)
+    if (m) return m[1]
+  }
+  return null
+}
+
+const sleep = (ms) => new Promise((r) => setTimeout(r, ms))
+let failures = 0
+const check = (label, cond, detail = '') => {
+  if (cond) console.log(`  ok   ${label}`)
+  else { failures++; console.log(`  FAIL ${label}${detail ? ` — ${detail}` : ''}`) }
+}
+
+async function freePort() {
+  const probe = createServer()
+  await new Promise((r) => probe.listen(0, '127.0.0.1', r))
+  const { port } = probe.address()
+  await new Promise((r) => probe.close(r))
+  return port
+}
+
+const key = loadApiKey()
+if (!key) {
+  console.error('CODEBUDDY_API_KEY not found (env or ~/.dsh/.credentials.yaml)')
+  process.exit(1)
+}
+
+// ------------------------------------------------------------ 隔离的运行环境
+process.env.DSH_HOME = mkdtempSync(join(tmpdir(), 'dsh-tap-tier-'))
+mkdirSync(process.env.DSH_HOME, { recursive: true })
+process.env.CODEBUDDY_API_KEY = key
+process.env.CODEBUDDY_BRIDGE_LOG = join(process.env.DSH_HOME, 'bridge-log.jsonl')
+
+// ------------------------------------------------- 本地捕获代理（原样转发 + 记录）
+const captured = []
+const proxy = createServer((req, res) => {
+  const chunks = []
+  req.on('data', (c) => chunks.push(c))
+  req.on('end', async () => {
+    const body = Buffer.concat(chunks).toString('utf8')
+    const url = `${UPSTREAM}${req.url}`
+    const headers = { ...req.headers }
+    delete headers.host
+    delete headers['content-length']
+    let upstream
+    try {
+      upstream = await fetch(url, {
+        method: req.method,
+        headers,
+        body: req.method === 'GET' || req.method === 'HEAD' ? undefined : body,
+      })
+    } catch (err) {
+      captured.push({ url: req.url, body, proxyError: String(err.message ?? err) })
+      res.writeHead(502, { 'Content-Type': 'application/json' })
+      res.end('{"code":1,"msg":"proxy upstream failed"}')
+      return
+    }
+    captured.push({
+      url: req.url,
+      status: upstream.status,
+      body,
+      reasoningEffort: (() => { try { return JSON.parse(body).reasoning_effort ?? null } catch { return null } })(),
+      model: (() => { try { return JSON.parse(body).model ?? null } catch { return null } })(),
+    })
+    const out = {}
+    for (const [k, v] of upstream.headers) if (k !== 'content-encoding' && k !== 'content-length') out[k] = v
+    res.writeHead(upstream.status, out)
+    if (upstream.body) Readable.fromWeb(upstream.body).pipe(res)
+    else res.end()
+  })
+})
+
+async function main() {
+  const proxyPort = await freePort()
+  await new Promise((r) => proxy.listen(proxyPort, '127.0.0.1', r))
+  const bridgePort = await freePort()
+
+  const { apply } = await import(new URL('../index.js', import.meta.url).href)
+  const routes = {}
+  const ctx = {
+    inject: (_deps, cb) => cb({ webServer: { register: (r) => { routes[r.path] = r.handler } } }),
+    web: { registerSearchProvider: () => () => {}, registerFetchProvider: () => () => {} },
+    on: () => {},
+  }
+  apply(ctx, {
+    baseURL: `http://127.0.0.1:${proxyPort}`,
+    bridgePort,
+    authMode: 'api-key',
+  })
+  await sleep(300)
+
+  // 设置路由用事件发射器驱动（同 verify-bridge 的 callRoute）。
+  const { EventEmitter } = await import('node:events')
+  const call = (body) => new Promise((resolve, reject) => {
+    const handler = routes['/dsh-tap/settings']
+    if (!handler) return reject(new Error('settings route not registered'))
+    const req = new EventEmitter()
+    req.method = body ? 'POST' : 'GET'
+    req.headers = { host: '127.0.0.1:3080', origin: 'http://127.0.0.1:3080' }
+    const res = {
+      status: 0,
+      writeHead(s) { this.status = s },
+      end(b) { let json = null; try { json = JSON.parse(b) } catch { /* non-JSON */ } resolve({ status: this.status, json }) },
+    }
+    handler(req, res)
+    if (body) {
+      const bytes = Buffer.from(JSON.stringify(body))
+      req.emit('data', bytes.subarray(0, 3))
+      req.emit('data', bytes.subarray(3))
+      req.emit('end')
+    }
+  })
+
+  console.log(`[1] 真实目录 → 档位表（proxy :${proxyPort}，bridge :${bridgePort}）`)
+  let res = await call({ action: 'model-sync' })
+  check('model-sync 成功且拿到网关目录', res.json?.sync?.ok === true && res.json?.sync?.count > 20,
+    JSON.stringify(res.json?.sync))
+  const catalogCount = res.json?.sync?.count ?? 0
+
+  res = await call({ action: 'model-list' })
+  const listed = res.json
+  const efforts = listed?.efforts ?? {}
+  const expect = {
+    'glm-5.3-flash': ['low', 'high', 'max'],
+    'kimi-k2.8-preview': ['low', 'high', 'max'],
+    'hy4-preview': ['high'],
+  }
+  for (const [id, tiers] of Object.entries(expect)) {
+    check(`目录声明 → 档位表 ${id} = ${JSON.stringify(tiers)}`,
+      JSON.stringify(efforts[id]) === JSON.stringify(tiers), JSON.stringify(efforts[id]))
+  }
+  check('静态 legacy 模型仍由 patch 表驱动（deepseek-v4-pro）',
+    JSON.stringify(efforts['deepseek-v4-pro']) === '["low","medium","high","max"]', JSON.stringify(efforts['deepseek-v4-pro']))
+  check('hy3 的 off 已撤（省略参数会思考）',
+    JSON.stringify(efforts['hy3']) === '["low","medium","high","max"]', JSON.stringify(efforts['hy3']))
+  check('无表模型不出现（auto / deepseek-v3）',
+    efforts['auto'] === undefined && efforts['deepseek-v3'] === undefined)
+  const flash = (listed?.catalog?.models ?? []).find((m) => m.id === 'glm-5.3-flash')
+  check('目录条目带完整声明（supportedEfforts/canDisableThinking/defaultEffort）',
+    JSON.stringify(flash?.supportedEfforts) === '["low","high","max"]' && flash?.canDisableThinking === true
+      && flash?.defaultEffort === 'high', JSON.stringify(flash))
+
+  console.log('\n[2] settings.yaml 镜像（宿主 Effort 选择器的数据源）')
+  const mirror = readFileSync(join(process.env.DSH_HOME, 'settings.yaml'), 'utf8')
+  check('镜像含目录模型的 reasoningEfforts',
+    /glm-5\.3-flash[\s\S]{0,400}?reasoningEfforts/.test(mirror))
+
+  const chats = []
+  if (!SKIP_CHAT) {
+    console.log('\n[3] 真实 chat：出站注入 + 上游接受')
+    const layerPath = join(process.env.DSH_HOME, 'codebuddy-plugin.json')
+    const chat = async (model) => {
+      const r = await fetch(`http://127.0.0.1:${bridgePort}/v2/chat/completions`, {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify({ model, stream: true, max_tokens: 24, messages: [{ role: 'user', content: '只回一个字：好' }] }),
+      })
+      const text = await r.text()
+      const last = [...captured].reverse().find((c) => c.url?.endsWith('/chat/completions'))
+      chats.push({ model, status: r.status, effort: last?.reasoningEffort ?? null, bodyPreview: text.slice(0, 120) })
+      return r.status
+    }
+
+    writeFileSync(layerPath, JSON.stringify({ effortByModel: { 'glm-5.3-flash': 'max' } }) + '\n')
+    let status = await chat('glm-5.3-flash')
+    check('已声明档位 max → 出站体带 reasoning_effort:"max"，上游 200',
+      status === 200 && chats.at(-1).effort === 'max', JSON.stringify(chats.at(-1)))
+
+    writeFileSync(layerPath, JSON.stringify({ effortByModel: { 'glm-5.3-flash': 'medium' } }) + '\n')
+    status = await chat('glm-5.3-flash')
+    check('未声明档位 medium → 不注入（仍 200）',
+      status === 200 && chats.at(-1).effort === null, JSON.stringify(chats.at(-1)))
+
+    writeFileSync(layerPath, JSON.stringify({ effortByModel: { auto: 'high' } }) + '\n')
+    status = await chat('auto')
+    check('无表模型 auto → 不注入（仍 200）',
+      status === 200 && chats.at(-1).effort === null, JSON.stringify(chats.at(-1)))
+
+    writeFileSync(layerPath, JSON.stringify({}) + '\n')
+  }
+
+  const evidence = {
+    fetchedAt: new Date().toISOString(),
+    catalogCount,
+    efforts,
+    mirrorHasReasoningEfforts: /glm-5\.3-flash[\s\S]{0,400}?reasoningEfforts/.test(mirror),
+    chats,
+    captured: captured.map((c) => ({ url: c.url, status: c.status, model: c.model, reasoningEffort: c.reasoningEffort })),
+  }
+  mkdirSync(dirname(OUT), { recursive: true })
+  writeFileSync(OUT, JSON.stringify(evidence, null, 2))
+
+  console.log(failures === 0 ? '\nall wiring checks passed' : `\n${failures} check(s) FAILED`)
+  console.log(`证据落盘：${OUT}`)
+  proxy.close()
+  process.exit(failures === 0 ? 0 : 1)
+}
+
+main().catch((err) => {
+  console.error(`probe crashed: ${err.stack ?? err.message}`)
+  process.exit(1)
+})