dsh-tap 0.20.0

package/scripts/verify-host-config.mjs

@@ -0,0 +1,265 @@
+/**
+ * verify-host-config.mjs — 宿主配置层（host-config.js）离线回归。
+ *
+ * 锁定 dsh 0.1.7 的配置持久化迁移：settings.yaml（dsh-settings-file，已删除）
+ * → profile 的 cordis.patch.yml（dsh-settings 的 forms seam + dsh-config-editor）。
+ * 两代宿主的选路、写前比对、revision 冲突重试、不可写降级、legacy 回退的注释
+ * 保留，全部用假服务/临时文件断言，不碰真实 ~/.dsh。
+ *
+ * 跑法：node scripts/verify-host-config.mjs
+ */
+import { mkdtempSync, writeFileSync, readFileSync, existsSync, rmSync } from 'node:fs'
+import { tmpdir } from 'node:os'
+import { join } from 'node:path'
+import { createHostConfigLayer } from '../host-config.js'
+
+let pass = 0
+let fail = 0
+const ok = (desc, cond, extra = '') => {
+  if (cond) { pass++; console.log(`  ok   ${desc}`) }
+  else { fail++; console.log(`  FAIL ${desc}${extra ? ' — ' + extra : ''}`) }
+}
+const section = (t) => console.log(`\n[${t}]`)
+
+// --- 假宿主：0.1.7 形态的 Settings forms 服务 -------------------------------
+function makeFormsService(opts = {}) {
+  const state = { providers: structuredClone(opts.providers ?? {}) }
+  let revision = opts.revision ?? 1
+  let conflicts = opts.conflicts ?? 0
+  const calls = { mutate: [], configure: [], register: [], dispose: 0 }
+  const svc = {
+    calls,
+    get writable() { return opts.writable !== false },
+    documentPath: opts.documentPath ?? '/fake/.dsh/profiles/web/cordis.patch.yml',
+    prepareDocument: async () => svc.documentPath,
+    configure(presentation, owner) {
+      calls.configure.push({ presentation, owner })
+      return () => { calls.dispose++ }
+    },
+    describe() {
+      return [
+        { ns: 'llm-pi-ai', revision, value: { providers: state.providers }, autoGenerate: false, applies: 'live', schema: { uid: 'x' } },
+        { ns: 'ui-theme', revision: 9, value: { preference: 'dark' }, autoGenerate: true, applies: 'live', schema: {} },
+      ]
+    },
+    async mutate(ns, ops, expected) {
+      calls.mutate.push({ ns, ops: structuredClone(ops), expected })
+      if (expected !== revision) {
+        const e = new Error(`expected ${expected}, actual ${revision}`)
+        e.code = 'SETTINGS_CONFLICT'
+        throw e
+      }
+      if (conflicts > 0) {
+        conflicts--
+        revision++
+        const e = new Error('injected conflict')
+        e.code = 'SETTINGS_CONFLICT'
+        throw e
+      }
+      for (const op of ops) {
+        const path = op.path
+        if (op.op === 'unset') {
+          let cur = state
+          for (let i = 0; i < path.length - 1; i++) cur = cur?.[path[i]]
+          if (cur) delete cur[path[path.length - 1]]
+        } else {
+          let cur = state
+          for (let i = 0; i < path.length - 1; i++) { cur[path[i]] ??= {}; cur = cur[path[i]] }
+          cur[path[path.length - 1]] = structuredClone(op.value)
+        }
+      }
+      revision++
+    },
+    _state: state,
+    _revision: () => revision,
+  }
+  return svc
+}
+
+/** 只有 register() 的旧宿主（≤0.1.6）。 */
+function makeLegacyService() {
+  const calls = { register: [], configure: [] }
+  return { calls, register(ns, schema) { calls.register.push({ ns, schema }) } }
+}
+
+function fakeCtx(service, fiber = { id: 'dsh-tap-fiber' }) {
+  return {
+    fiber,
+    inject(deps, cb) {
+      if (Array.isArray(deps) && deps.includes('settings')) cb({ settings: service })
+    },
+  }
+}
+
+const TRAE_BLOCK = {
+  displayName: 'TraeWork CN',
+  api: 'openai-completions',
+  baseURL: 'http://127.0.0.1:3902/v1',
+  headers: { Authorization: 'Bearer dsh-trae-bridge' },
+  models: [{ id: 'tm1', name: 'Trae M1', contextWindow: 128000, maxTokens: 32000 }],
+}
+
+// === A. forms 选路（dsh 0.1.7+）============================================
+section('A] forms seam（0.1.7+：写落 profile patch）')
+{
+  const svc = makeFormsService()
+  const layer = createHostConfigLayer({ settingsPath: '/nonexistent/settings.yaml', schema: { fake: true }, log: () => {} })
+  layer.attach(fakeCtx(svc))
+
+  ok('A1 attach 走 configure({auto:false}) 而不是 register',
+    svc.calls.configure.length === 1 && svc.calls.configure[0].presentation.auto === false && svc.calls.register.length === 0,
+    JSON.stringify({ configure: svc.calls.configure.length, register: svc.calls.register.length }))
+  ok('A2 configure 带上调用方 fiber（策略归属本插件实例）',
+    svc.calls.configure[0]?.owner?.id === 'dsh-tap-fiber')
+  ok('A3 mode() = forms', layer.mode() === 'forms', layer.mode())
+  ok('A4 servicePresent() = true', layer.servicePresent() === true)
+
+  const r1 = await layer.applyOps([{ op: 'set', path: ['providers', 'trae'], value: TRAE_BLOCK }])
+  ok('A5 set 新块：ok 且 changed', r1.ok === true && r1.changed === true, JSON.stringify(r1))
+  ok('A6 mutate 命名空间 = profile entry id llm-pi-ai', svc.calls.mutate[0]?.ns === 'llm-pi-ai', svc.calls.mutate[0]?.ns)
+  ok('A7 mutate 带 describe 拿到的 expectedRevision', svc.calls.mutate[0]?.expected === 1, String(svc.calls.mutate[0]?.expected))
+  ok('A8 块真的进了有效配置', layer.readProviders().trae?.baseURL === 'http://127.0.0.1:3902/v1')
+
+  const before = svc.calls.mutate.length
+  const r2 = await layer.applyOps([{ op: 'set', path: ['providers', 'trae'], value: structuredClone(TRAE_BLOCK) }])
+  ok('A9 同值重写不再 mutate（避免 revision 抖动/reload 风暴）',
+    r2.ok === true && r2.changed === false && svc.calls.mutate.length === before, JSON.stringify(r2))
+
+  const r3 = await layer.applyOps([{ op: 'unset', path: ['providers', 'nope'] }])
+  ok('A10 unset 不存在的路径：不写', r3.ok === true && r3.changed === false && svc.calls.mutate.length === before)
+
+  const r4 = await layer.applyOps([{ op: 'unset', path: ['providers', 'trae'] }])
+  ok('A11 unset 存在的路径：删除生效', r4.ok === true && r4.changed === true && layer.readProviders().trae === undefined)
+
+  const r5 = await layer.applyOps([{ op: 'set', path: ['providers', 'codebuddy', 'models'], value: [{ id: 'deepseek-v3', name: 'DeepSeek V3', contextWindow: 131072, maxTokens: 32768 }] }])
+  ok('A12 深路径（providers.codebuddy.models）可写', r5.ok === true && layer.readProviders().codebuddy?.models?.[0]?.id === 'deepseek-v3')
+
+  const p = layer.probe()
+  ok('A13 probe 报 forms/可写/documentPath/entry/autoGenerate',
+    p.mode === 'forms' && p.writable === true && /cordis\.patch\.yml$/.test(p.documentPath) && p.entryNs === 'llm-pi-ai' && p.autoGenerate === false,
+    JSON.stringify({ mode: p.mode, writable: p.writable, doc: p.documentPath, ns: p.entryNs, auto: p.autoGenerate }))
+  ok('A14 probe 列出全部命名空间（升级排查用）', Array.isArray(p.allNamespaces) && p.allNamespaces.includes('ui-theme'))
+
+  const disposer = svc.calls.dispose
+  layer.dispose()
+  ok('A15 dispose 释放 configure 的 effect', svc.calls.dispose === disposer + 1)
+}
+
+section('A] forms seam：冲突与降级')
+{
+  const svc = makeFormsService({ conflicts: 1 })
+  const layer = createHostConfigLayer({ settingsPath: '/nonexistent/settings.yaml', schema: {}, log: () => {} })
+  layer.attach(fakeCtx(svc))
+  const r = await layer.applyOps([{ op: 'set', path: ['providers', 'trae'], value: TRAE_BLOCK }])
+  ok('A16 SETTINGS_CONFLICT 后重读 revision 重试成功',
+    r.ok === true && svc.calls.mutate.length === 2 && svc.calls.mutate[1].expected === 2,
+    JSON.stringify({ ok: r.ok, n: svc.calls.mutate.length, exp: svc.calls.mutate.map((c) => c.expected) }))
+}
+{
+  const svc = makeFormsService({ writable: false })
+  const layer = createHostConfigLayer({ settingsPath: '/nonexistent/settings.yaml', schema: {}, log: () => {} })
+  layer.attach(fakeCtx(svc))
+  const r = await layer.applyOps([{ op: 'set', path: ['providers', 'trae'], value: TRAE_BLOCK }])
+  ok('A17 不可写 profile：返回失败而不是抛（踩坑 #33）',
+    r.ok === false && /SETTINGS_NOT_WRITABLE/.test(r.error ?? '') && svc.calls.mutate.length === 0, JSON.stringify(r))
+  ok('A18 失败落 lastError', layer.lastError === r.error)
+}
+{
+  const svc = makeFormsService()
+  svc.mutate = async () => { throw new Error('boom upstream') }
+  const layer = createHostConfigLayer({ settingsPath: '/nonexistent/settings.yaml', schema: {}, log: () => {} })
+  layer.attach(fakeCtx(svc))
+  const r = await layer.applyOps([{ op: 'set', path: ['providers', 'trae'], value: TRAE_BLOCK }])
+  ok('A19 mutate 抛非冲突错：不 reject，带错误码前缀', r.ok === false && /boom upstream/.test(r.error ?? ''), JSON.stringify(r))
+  const r2 = await layer.applyOps([])
+  ok('A20 空 ops 直接成功且不触服务', r2.ok === true && r2.changed === false)
+}
+
+// === A2. 宿主替换 Settings 实例（0.1.7 profile 重载语义）====================
+section('A] 宿主替换 Settings 实例后仍能写（踩坑 #43 同批实测）')
+{
+  // 每次写 profile patch 都会触发重载，Settings 服务实例可能被**替换**；
+  // cordis 的 ctx.settings 是实时 getter，缓存旧引用会带着陈旧 revision 写，
+  // 实测症状 = SETTINGS_CONFLICT 连撞三次全败（镜像停在残缺清单）。
+  const makeSvc = (rev) => makeFormsService({ revision: rev })
+  let current = makeSvc(1)
+  const ctx = {
+    fiber: { id: 'dsh-tap-fiber' },
+    inject(deps, cb) {
+      if (Array.isArray(deps) && deps.includes('settings')) cb({ get settings() { return current } })
+    },
+  }
+  const layer = createHostConfigLayer({ settingsPath: '/nonexistent/settings.yaml', schema: {}, log: () => {} })
+  layer.attach(ctx)
+  const r1 = await layer.applyOps([{ op: 'set', path: ['providers', 'trae'], value: TRAE_BLOCK }])
+  ok('A21 首次写入成功', r1.ok === true, JSON.stringify(r1))
+
+  // 宿主重载：旧实例作废（再写即 SETTINGS_CLOSED），换上 revision 不同的新实例
+  const stale = current
+  stale.mutate = async () => { const e = new Error('service closed'); e.code = 'SETTINGS_CLOSED'; throw e }
+  current = makeSvc(20)
+  const r2 = await layer.applyOps([{ op: 'set', path: ['providers', 'qoder'], value: { displayName: 'Qoder CN', models: [{ id: 'qmodel' }] } }])
+  ok('A22 服务被替换后仍写成功（活取 ctx.settings，不吃陈旧引用）', r2.ok === true, JSON.stringify(r2))
+  ok('A23 写入落在新实例上、命名空间正确', current.calls.mutate.length === 1 && current.calls.mutate[0].ns === 'llm-pi-ai' && current.calls.mutate[0].expected === 20,
+    JSON.stringify(current.calls.mutate.map((c) => c.expected)))
+  ok('A24 陈旧实例没有被再碰过（否则会 SETTINGS_CLOSED）', stale.calls.mutate.length === 1, String(stale.calls.mutate.length))
+}
+
+// === B. legacy 回退（≤0.1.6：settings.yaml 文档编辑）========================
+section('B] legacy 回退（settings.yaml + 注释保留）')
+const dir = mkdtempSync(join(tmpdir(), 'dsh-tap-hostconfig-'))
+{
+  const path = join(dir, 'settings.yaml')
+  writeFileSync(path, [
+    '# 用户手写注释必须活下来',
+    'ui-theme:',
+    '  preference: dark',
+    'llm-pi-ai:',
+    '  providers:',
+    '    codebuddy:',
+    '      models:',
+    '        - id: deepseek-v3',
+    '          name: DeepSeek V3',
+    '',
+  ].join('\n'))
+  const svc = makeLegacyService()
+  const layer = createHostConfigLayer({ settingsPath: path, schema: { fake: 'schema' }, log: () => {} })
+  layer.attach(fakeCtx(svc))
+
+  ok('B1 旧宿主走 register(命名空间, schema)',
+    svc.calls.register.length === 1 && svc.calls.register[0].ns === 'dsh-tap' && svc.calls.register[0].schema?.fake === 'schema' && svc.calls.configure.length === 0,
+    JSON.stringify(svc.calls))
+  ok('B2 mode() = legacy', layer.mode() === 'legacy', layer.mode())
+  ok('B3 readProviders 读 settings.yaml', layer.readProviders().codebuddy?.models?.[0]?.id === 'deepseek-v3')
+
+  const r = await layer.applyOps([{ op: 'set', path: ['providers', 'trae'], value: TRAE_BLOCK }])
+  const text = readFileSync(path, 'utf8')
+  ok('B4 set 写进 llm-pi-ai.providers.trae', r.ok === true && /trae:/.test(text) && /TraeWork CN/.test(text))
+  ok('B5 用户注释与无关段保留', text.includes('# 用户手写注释必须活下来') && /ui-theme:/.test(text) && /preference: dark/.test(text))
+  ok('B6 既有 codebuddy 块未被整段重写掉', /codebuddy:/.test(text) && /deepseek-v3/.test(text))
+
+  const snapshot = readFileSync(path, 'utf8')
+  const r2 = await layer.applyOps([{ op: 'set', path: ['providers', 'trae'], value: structuredClone(TRAE_BLOCK) }])
+  ok('B7 同值不写文件', r2.ok === true && r2.changed === false && readFileSync(path, 'utf8') === snapshot)
+
+  await layer.applyOps([{ op: 'unset', path: ['providers', 'trae'] }])
+  ok('B8 unset 删块（路由存在性管理的旧路径）', !/trae:/.test(readFileSync(path, 'utf8')))
+  ok('B9 删块后注释仍在', readFileSync(path, 'utf8').includes('# 用户手写注释必须活下来'))
+}
+
+// === C. 完全没有 settings 服务（rc.6 等）===================================
+section('C] 无 settings 服务')
+{
+  const path = join(dir, 'bare.yaml')
+  const layer = createHostConfigLayer({ settingsPath: path, schema: {}, log: () => {} })
+  layer.attach({ fiber: null, inject: () => {} }) // inject 永不回调
+  ok('C1 servicePresent = false 且 mode = legacy', layer.servicePresent() === false && layer.mode() === 'legacy')
+  const r = await layer.applyOps([{ op: 'set', path: ['providers', 'qoder'], value: { displayName: 'Qoder CN', models: [{ id: 'qmodel' }] } }])
+  ok('C2 仍能把块写进文件（宿主无 settings 服务时不失能）',
+    r.ok === true && existsSync(path) && /qoder:/.test(readFileSync(path, 'utf8')))
+  ok('C3 probe 不炸且报 legacy', layer.probe().mode === 'legacy' && layer.probe().servicePresent === false)
+}
+
+rmSync(dir, { recursive: true, force: true })
+console.log(`\n=== verify-host-config: ${pass} 通过 / ${fail} 失败 ===`)
+process.exit(fail ? 1 : 0)

package/scripts/verify-models.mjs

@@ -0,0 +1,390 @@
+#!/usr/bin/env node
+/**
+ * Verify that every model declared in cordis.patch.yml is still served by
+ * the CodeBuddy gateway. Retired or unknown model ids fail with error
+ * 11102 ("service info not found").
+ *
+ * Usage:
+ *   node scripts/verify-models.mjs --list   # offline: print configured models
+ *   node scripts/verify-models.mjs          # online: probe each model
+ *   node scripts/verify-models.mjs --sync   # online: diff against the
+ *                                           # gateway's /v3/config catalog
+ *   node scripts/verify-models.mjs --efforts [id ...]
+ *                                           # online: probe reasoning_effort
+ *                                           # levels (omit/low/medium/high/
+ *                                           # max) per model; defaults to
+ *                                           # models lacking reasoningEfforts
+ *
+ * No npm dependencies. The API key is read from CODEBUDDY_API_KEY (env) or
+ * ~/.dsh/.credentials.yaml — the same resolution order dsh uses.
+ */
+
+import { readFileSync, existsSync } 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 PATCH_FILE = join(ROOT, 'cordis.patch.yml')
+const BASE_URL = 'https://copilot.tencent.com/v2'
+
+// The gateway has no OpenAI-style GET /models endpoint (404), which is why
+// dsh's built-in "fetch available models" cannot work. The catalog the
+// official CLI itself uses lives at /v3/config instead — with its own auth
+// dialect: x-api-key header, and a UA that must match the CLI shape or the
+// gateway answers error 12403 ("check ua"). The /v2 client UA is rejected.
+const CATALOG_URL = 'https://copilot.tencent.com/v3/config'
+const CATALOG_USER_AGENT = 'CLI/unknown CodeBuddy/2.136.0'
+
+// Same header set the provider config sends — the gateway routes and logs
+// CodeBuddy traffic by these client markers.
+const CLIENT_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',
+}
+
+/**
+ * Extract the `models:` list from cordis.patch.yml. The patch has a fixed,
+ * flat shape, so a line scanner scoped to the models block is enough —
+ * anything failing to parse here means the patch layout changed and the
+ * script must be updated.
+ */
+function parseModels(text) {
+  const models = []
+  let modelsIndent = null
+  let current = null
+  for (const line of text.split('\n')) {
+    if (modelsIndent === null) {
+      const header = line.match(/^(\s*)models:\s*$/)
+      if (header) modelsIndent = header[1].length
+      continue
+    }
+    if (!line.trim() || line.trim().startsWith('#')) continue
+    const indent = line.match(/^\s*/)[0].length
+    if (indent <= modelsIndent) break // models block ended
+    const entry = line.match(/^\s*-\s*id:\s*(\S+)\s*$/)
+    if (entry) {
+      current = { id: entry[1], contextWindow: null, maxTokens: null, input: [] }
+      models.push(current)
+      continue
+    }
+    const kv = line.match(/^\s*(contextWindow|maxTokens):\s*(\d+)/)
+    if (kv && current) current[kv[1]] = Number(kv[2])
+    if (/^\s*reasoningEfforts:/.test(line) && current) current.reasoningEfforts = true
+    const input = line.match(/^\s*input:\s*\[([^\]]*)\]/)
+    if (input && current) {
+      current.input = input[1].split(',').map((s) => s.trim()).filter(Boolean)
+    }
+  }
+  return models
+}
+
+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
+}
+
+async function fetchCatalog(apiKey) {
+  const res = await fetch(CATALOG_URL, {
+    headers: {
+      accept: 'application/json',
+      'x-api-key': apiKey,
+      'user-agent': CATALOG_USER_AGENT,
+      'x-product': 'SaaS',
+    },
+  })
+  if (!res.ok) throw new Error(`catalog endpoint returned HTTP ${res.status}`)
+  const body = await res.json()
+  if (body?.code !== 0) {
+    throw new Error(`catalog error: ${body?.msg ?? body?.code}`)
+  }
+  const data = body.data ?? {}
+  const models = new Map((data.models ?? []).map((m) => [m.id, m]))
+  const cliEnabled = new Set(
+    (data.agents ?? []).find((a) => a.name === 'cli')?.models ?? [],
+  )
+  return { models, cliEnabled }
+}
+
+async function probeModel(id, apiKey) {
+  let res
+  try {
+    res = await fetch(`${BASE_URL}/chat/completions`, {
+      method: 'POST',
+      headers: {
+        ...CLIENT_HEADERS,
+        Authorization: `Bearer ${apiKey}`,
+        'Content-Type': 'application/json',
+      },
+      body: JSON.stringify({
+        model: id,
+        messages: [{ role: 'user', content: 'hi' }],
+        max_tokens: 1,
+        stream: true,
+      }),
+    })
+  } catch (err) {
+    return { ok: false, detail: `network error: ${err.message}` }
+  }
+  if (!res.ok) {
+    const body = await res.text().catch(() => '')
+    return { ok: false, detail: `HTTP ${res.status} ${body.slice(0, 200)}` }
+  }
+  // The gateway may answer 200 and still stream an error payload (e.g.
+  // code 11102 for retired ids), so scan the stream instead of trusting
+  // the status code.
+  let streamed = ''
+  try {
+    const reader = res.body.getReader()
+    const decoder = new TextDecoder()
+    for (;;) {
+      const { done, value } = await reader.read()
+      if (done) break
+      streamed += decoder.decode(value, { stream: true })
+      if (streamed.length > 4096) break
+    }
+  } catch {
+    // Body already partially read; what we have is enough to judge.
+  }
+  const errMatch = streamed.match(/"code"\s*:\s*(\d+)|"error"\s*:/)
+  if (errMatch) {
+    return { ok: false, detail: `streamed error: ${streamed.slice(0, 200)}` }
+  }
+  return { ok: true, detail: 'ok' }
+}
+
+/**
+ * Diff the patch's model list against the /v3/config catalog: report
+ * parameter drift, ids the catalog no longer lists, and catalog ids the
+ * patch could adopt; emit corrected YAML entries for the drifted ones.
+ * Exits 1 only when a model present in both disagrees with the catalog —
+ * unlisted legacy ids and not-yet-adopted catalog ids are informational.
+ */
+async function runSync(models, apiKey) {
+  const { models: catalog, cliEnabled } = await fetchCatalog(apiKey)
+  const inPatch = new Set(models.map((m) => m.id))
+  const drifted = []
+  const aligned = []
+  const unlisted = []
+
+  for (const m of models) {
+    const c = catalog.get(m.id)
+    if (!c) {
+      unlisted.push(m.id)
+      continue
+    }
+    const issues = []
+    if (c.maxInputTokens != null && m.contextWindow !== c.maxInputTokens) {
+      issues.push(`contextWindow ${m.contextWindow} → ${c.maxInputTokens}`)
+    }
+    if (c.maxOutputTokens != null && m.maxTokens !== c.maxOutputTokens) {
+      issues.push(`maxTokens ${m.maxTokens} → ${c.maxOutputTokens}`)
+    }
+    const wantsImage = c.supportsImages === true
+    const hasImage = m.input.includes('image')
+    if (wantsImage !== hasImage) {
+      issues.push(`input ${hasImage ? '[text, image]' : '[text]'} → ${wantsImage ? '[text, image]' : '[text]'}`)
+    }
+    issues.length ? drifted.push({ m, c, issues }) : aligned.push(m.id)
+  }
+
+  console.log(`catalog: ${catalog.size} models, ${cliEnabled.size} CLI-enabled`)
+  console.log(`\naligned with catalog (${aligned.length}): ${aligned.join(', ')}`)
+
+  console.log(`\nparameter drift (${drifted.length}):`)
+  if (drifted.length === 0) {
+    console.log('  none')
+  } else {
+    for (const { m, issues } of drifted) {
+      console.log(`  ${m.id}`)
+      for (const issue of issues) console.log(`    ${issue}`)
+    }
+    console.log('\ncorrected entries (replace the matching ids in the patch):')
+    for (const { m, c } of drifted) {
+      const lines = [
+        `          - id: ${m.id}`,
+        `            name: ${m.name ?? m.id}`,
+      ]
+      if (c.maxInputTokens != null) lines.push(`            contextWindow: ${c.maxInputTokens}`)
+      if (c.maxOutputTokens != null) lines.push(`            maxTokens: ${c.maxOutputTokens}`)
+      if (c.supportsImages === true) lines.push('            input: [text, image]')
+      console.log(lines.join('\n'))
+    }
+  }
+
+  console.log(`\nnot in catalog, still serving as legacy ids (${unlisted.length}):`)
+  for (const id of unlisted) console.log(`  ${id}`)
+
+  const adoptable = [...catalog.keys()].filter((id) => !inPatch.has(id))
+  console.log(`\nin catalog but not in the patch (${adoptable.length}):`)
+  for (const id of adoptable) {
+    const c = catalog.get(id)
+    const cli = cliEnabled.has(id) ? ' [CLI]' : ''
+    const sizes = c.maxInputTokens != null
+      ? `ctx=${c.maxInputTokens} max=${c.maxOutputTokens}`
+      : 'no token fields'
+    const img = c.supportsImages === true ? ' img' : ''
+    console.log(`  ${id}${cli}  ${sizes}${img}`)
+  }
+
+  return drifted.length > 0 ? 1 : 0
+}
+
+/**
+ * Probe which reasoning_effort levels a model accepts: one streaming request
+ * per level (omitting the parameter is the baseline). reasoning_content vs
+ * content length tells whether the level actually changed the thinking
+ * budget — identical lengths across all levels mean the gateway ignores
+ * the parameter for that model.
+ */
+const EFFORT_LEVELS = [null, 'low', 'medium', 'high', 'max']
+
+async function probeEffort(id, apiKey, effort) {
+  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: 800,
+    stream: true,
+  }
+  if (effort) body.reasoning_effort = effort
+  let res
+  try {
+    res = await fetch(`${BASE_URL}/chat/completions`, {
+      method: 'POST',
+      headers: {
+        ...CLIENT_HEADERS,
+        Authorization: `Bearer ${apiKey}`,
+        'Content-Type': 'application/json',
+      },
+      body: JSON.stringify(body),
+    })
+  } catch (err) {
+    return { error: `network: ${err.message}` }
+  }
+  if (!res.ok) {
+    return { error: `HTTP ${res.status} ${(await res.text().catch(() => '')).slice(0, 120)}` }
+  }
+  let buf = ''
+  let reasonChars = 0
+  let contentChars = 0
+  let streamErr = null
+  let finish = null
+  try {
+    const reader = res.body.getReader()
+    const dec = new TextDecoder()
+    for (;;) {
+      const { done, value } = await reader.read()
+      if (done) break
+      buf += dec.decode(value, { stream: true })
+      let nl
+      while ((nl = buf.indexOf('\n')) >= 0) {
+        const line = buf.slice(0, nl).trim()
+        buf = buf.slice(nl + 1)
+        if (!line.startsWith('data:')) continue
+        const payload = line.slice(5).trim()
+        if (payload === '[DONE]') continue
+        try {
+          const j = JSON.parse(payload)
+          if (j.error || (j.code != null && j.code !== 0)) streamErr = payload.slice(0, 120)
+          const ch = j.choices?.[0]
+          if (ch?.delta?.reasoning_content) reasonChars += ch.delta.reasoning_content.length
+          if (ch?.delta?.content) contentChars += ch.delta.content.length
+          if (ch?.finish_reason) finish = ch.finish_reason
+        } catch {
+          // incomplete JSON inside a complete line — skip it
+        }
+      }
+      if (reasonChars + contentChars > 6000) break
+    }
+  } catch {
+    // partial stream is still evidence
+  }
+  return { reasonChars, contentChars, finish, streamErr }
+}
+
+async function runEfforts(models, apiKey) {
+  for (const m of models) {
+    console.log(`\n${m.id}`)
+    for (const e of EFFORT_LEVELS) {
+      const label = (e ?? '(omit)').padEnd(8)
+      process.stdout.write(`  ${label}`)
+      const r = await probeEffort(m.id, apiKey, e)
+      if (r.error || r.streamErr) {
+        console.log(` ✗  ${r.error ?? r.streamErr}`)
+      } else {
+        console.log(
+          ` reason=${String(r.reasonChars).padStart(5)}  content=${String(r.contentChars).padStart(4)}  finish=${r.finish ?? '-'}`,
+        )
+      }
+    }
+  }
+  return 0
+}
+
+const models = parseModels(readFileSync(PATCH_FILE, 'utf8'))
+if (models.length === 0) {
+  console.error(`Failed to parse any model from ${PATCH_FILE} — patch layout changed?`)
+  process.exit(1)
+}
+
+if (process.argv.includes('--list')) {
+  for (const m of models) {
+    const input = m.input.length ? m.input.join('+') : 'text'
+    console.log(
+      `${m.id.padEnd(18)} contextWindow=${String(m.contextWindow ?? '?').padEnd(7)} maxTokens=${String(m.maxTokens ?? '?').padEnd(6)} input=${input}`,
+    )
+  }
+  console.log(`\n${models.length} models configured`)
+  process.exit(0)
+}
+
+const apiKey = loadApiKey()
+if (!apiKey) {
+  console.error('CODEBUDDY_API_KEY not found (env or ~/.dsh/.credentials.yaml)')
+  process.exit(1)
+}
+
+if (process.argv.includes('--sync')) {
+  try {
+    process.exit(await runSync(models, apiKey))
+  } catch (err) {
+    console.error(`catalog fetch failed: ${err.message}`)
+    process.exit(1)
+  }
+}
+
+const effortIdx = process.argv.indexOf('--efforts')
+if (effortIdx !== -1) {
+  const requested = process.argv.slice(effortIdx + 1).filter((a) => !a.startsWith('--'))
+  const targets = requested.length
+    ? models.filter((m) => requested.includes(m.id))
+    : models.filter((m) => !m.reasoningEfforts)
+  if (targets.length === 0) {
+    console.error('no matching models to probe')
+    process.exit(1)
+  }
+  await runEfforts(targets, apiKey)
+  process.exit(0)
+}
+
+let failed = 0
+for (const m of models) {
+  process.stdout.write(`probing ${m.id.padEnd(18)} ... `)
+  const r = await probeModel(m.id, apiKey)
+  console.log(r.ok ? 'OK' : `FAIL (${r.detail})`)
+  if (!r.ok) failed++
+}
+console.log(`\n${models.length - failed}/${models.length} models available`)
+process.exit(failed > 0 ? 1 : 0)