dsh-tap 0.20.0

package/scripts/probe-routing.mjs

@@ -0,0 +1,226 @@
+#!/usr/bin/env node
+/**
+ * Route-allocation probe (topic 3 of docs/rules/): what dimensions does the
+ * gateway allocate routes by, i.e. what exactly does 14407 "route config not
+ * found" key on?
+ *
+ * Spell under test (AGENTS.md:24): "`/v2/videos/generations`、`/v2/3d/generations`
+ * 路由存在但当前账号一律 14407 route config not found（无可用模型）"。
+ *
+ * Anchor evidence (docs/probes/media-2026-08-17.json): 14407 messages embed
+ * the MODEL name — "Video model [hunyuan-video-t2v] route config not found" —
+ * so 14407 is a per-(endpoint-family, model) lookup failure, not a 404.
+ *
+ * /v3/config (fetched live in this script) lists 24 models for this account
+ * incl. hunyuan-image-v3.0-art and NO video/3d models — but deepseek-v3,
+ * which routes fine on chat, is NOT among them, so "catalog membership" is
+ * not the route table. The route table is its own per-(family, model) map.
+ *
+ * Pre-registered predictions (2026-08-19):
+ *   P1  chat + bogus model → 14407, message echoes the model name verbatim
+ *   P2  images + bogus model → 14407 with "image model" wording
+ *   P3  images + glm-5.2 (a chat-family model that IS in the catalog)
+ *       → 14407 (route configs are family-scoped, catalog membership does
+ *       not cross families)
+ *   P4  chat + hunyuan-image-v3.0-art (image model) → 14407
+ *   P5  chat + glm-5.0 (catalog-listed, NOT in the cli agents list)
+ *       → 200 (route config exists; the cli list gates CLI display, not
+ *       routing)  [costs one max_tokens=1 completion if it passes]
+ *   P6  unknown family /v2/foo/generations → 404 "Route Not Found"
+ *       (path routing is a layer separate from model routing)
+ * Exploratory (no firm prediction): videos + omitted model field.
+ *
+ * Low rate: 2s spacing, single account; all arms are designed to fail at
+ * the route layer (zero generation cost) except P5's tiny completion.
+ *
+ * Usage: node scripts/probe-routing.mjs [--out docs/probes/routing-YYYY-MM-DD.jsonl]
+ */
+
+import { readFileSync, existsSync, appendFileSync, mkdirSync } from 'node:fs'
+import { homedir } from 'node:os'
+import { join, dirname } from 'node:path'
+
+const DSH_HOME = process.env.DSH_HOME ?? join(homedir(), '.dsh')
+const GATEWAY = 'https://copilot.tencent.com'
+const OUT = process.argv[process.argv.indexOf('--out') + 1] && process.argv.includes('--out')
+  ? process.argv[process.argv.indexOf('--out') + 1]
+  : join('docs', 'probes', `routing-${new Date().toISOString().slice(0, 10)}.jsonl`)
+const SPACING_MS = 2000
+
+function resolveKey() {
+  try {
+    const cfg = JSON.parse(readFileSync(join(DSH_HOME, 'codebuddy-plugin.json'), 'utf8'))
+    const active = (cfg.apiKeys ?? []).find((k) => k.name === cfg.activeApiKey)
+    if (active?.key) return active.key
+  } catch { /* fall through */ }
+  if (process.env.CODEBUDDY_API_KEY) return process.env.CODEBUDDY_API_KEY
+  const credFile = join(DSH_HOME, '.credentials.yaml')
+  if (existsSync(credFile)) {
+    const m = readFileSync(credFile, 'utf8').match(/^\s*CODEBUDDY_API_KEY:\s*["']?([^"'\s]+)["']?\s*$/m)
+    if (m) return m[1]
+  }
+  throw new Error('no CodeBuddy credential found')
+}
+
+const sleep = (ms) => new Promise((r) => setTimeout(r, ms))
+const redact = (v) => typeof v === 'string'
+  ? v.replace(/[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}/gi, '<redacted>').replace(/\b\d{6,}\b/g, '<redacted>')
+  : v
+
+const HEADERS = (key) => ({
+  'Content-Type': 'application/json',
+  Accept: 'application/json',
+  Authorization: `Bearer ${key}`,
+  'User-Agent': 'CLI/unknown CodeBuddy/2.136.0',
+})
+
+const ARMS = [
+  { name: 'P1-chat-bogus-model', method: 'POST', path: '/v2/chat/completions',
+    body: { model: 'no-such-model-xyz', stream: true, max_tokens: 1, messages: [{ role: 'user', content: 'hi' }] },
+    expect: { code: 14407 }, note: 'message echoes model name verbatim' },
+  { name: 'P2-images-bogus-model', method: 'POST', path: '/v2/images/generations',
+    body: { model: 'no-such-image-model', prompt: 'x', size: '1024x1024', n: 1 },
+    expect: { code: 14407 }, note: '"image model" wording' },
+  { name: 'P3-images-chat-model', method: 'POST', path: '/v2/images/generations',
+    body: { model: 'glm-5.2', prompt: 'x', size: '1024x1024', n: 1 },
+    expect: { code: 14407 }, note: 'catalog chat model does not cross into the image family' },
+  { name: 'P4-chat-image-model', method: 'POST', path: '/v2/chat/completions',
+    body: { model: 'hunyuan-image-v3.0-art', stream: true, max_tokens: 1, messages: [{ role: 'user', content: 'hi' }] },
+    expect: { code: 14407 }, note: 'image model has no chat route' },
+  { name: 'P5-chat-catalog-noncli', method: 'POST', path: '/v2/chat/completions',
+    body: { model: 'glm-5.0', stream: true, stream_options: { include_usage: true }, max_tokens: 1, messages: [{ role: 'user', content: 'hi' }] },
+    expect: { status: 200 }, note: 'catalog-listed but not cli-enabled still routes' },
+  { name: 'P6-unknown-family', method: 'POST', path: '/v2/foo/generations',
+    body: { model: 'x' },
+    expect: { status: 404 }, note: 'path routing layer is separate from model routing' },
+  { name: 'X1-videos-no-model', method: 'POST', path: '/v2/videos/generations',
+    body: { prompt: 'x' },
+    expect: null, note: 'exploratory: error shape when model field omitted' },
+]
+
+// Round 2 (2026-08-19): the matrix refuted the naive "14407 = (family,model)"
+// picture and revealed a THREE-LAYER structure. Revised rules under test:
+//   R-R2  media-family registry miss → family-specific code + verbatim echo:
+//         image → 14401, video/3d → 14407 ("<Family> model [<echo>] route
+//         config not found"); the miss means "model unknown in that family",
+//         a GLOBALLY known model passes lookup and fails later at dispatch
+//         (P3: images+glm-5.2 → bare 500; P4: chat+image-model → 500/11103)
+//   R-R3  chat-family registry miss → 11102 "model [echo] service info not
+//         found"; the chat registry is independent of /v3/config AND of the
+//         cli-enabled list (kimi-k2.5 routes fine though not cli-listed;
+//         glm-5.0 is catalog-listed yet 11102)
+//   R-R4  model known but backend unsupported in context → HTTP 500, either
+//         bare (image family) or 11103 "Backend [x] is not supported" (chat)
+// Pre-registered confirmations (unobserved cases):
+const ARMS_R2 = [
+  { name: 'C1-images-no-model', method: 'POST', path: '/v2/images/generations',
+    body: { prompt: 'x', size: '1024x1024', n: 1 },
+    expect: { code: 14401 }, note: 'image family echoes empty model like video X1' },
+  { name: 'C2-video-fresh-bogus', method: 'POST', path: '/v2/videos/generations',
+    body: { model: 'no-such-video-model', prompt: 'x' },
+    expect: { code: 14407 }, note: 'verbatim echo of a fresh unobserved name' },
+  { name: 'C3-3d-no-model', method: 'POST', path: '/v2/3d/generations',
+    body: { prompt: 'x' },
+    expect: { code: 14407 }, note: '3d family echoes empty model' },
+  { name: 'C4-chat-fresh-bogus', method: 'POST', path: '/v2/chat/completions',
+    body: { model: 'no-such-model-abc', stream: true, max_tokens: 1, messages: [{ role: 'user', content: 'hi' }] },
+    expect: { code: 11102 }, note: 'chat registry miss wording is stable' },
+  { name: 'C5-chat-notcli-k25', method: 'POST', path: '/v2/chat/completions',
+    body: { model: 'kimi-k2.5', stream: true, stream_options: { include_usage: true }, max_tokens: 1, messages: [{ role: 'user', content: 'hi' }] },
+    expect: { status: 200 }, note: 'cli list does not gate routing (k2.5 not cli-listed)' },
+  { name: 'C6-chat-volc-alias', method: 'POST', path: '/v2/chat/completions',
+    body: { model: 'deepseek-v3-2-volc', stream: true, max_tokens: 1, messages: [{ role: 'user', content: 'hi' }] },
+    expect: { code: 11102 }, note: 'registry keyed by public ids; catalog-internal volc id unknown' },
+  { name: 'C7-chat-hunyuan-chat', method: 'POST', path: '/v2/chat/completions',
+    body: { model: 'hunyuan-chat', stream: true, max_tokens: 1, messages: [{ role: 'user', content: 'hi' }] },
+    expect: { code: 11103 }, note: 'catalog chat-branded model resolves but its backend unsupported' },
+  { name: 'C8-3d-known-image-model', method: 'POST', path: '/v2/3d/generations',
+    body: { model: 'hunyuan-image-v3.0-art', prompt: 'x' },
+    expect: { status: 500 }, note: 'globally-known model passes 3d lookup, dies at dispatch — NOT 14407' },
+]
+
+// Round 3 (2026-08-19): C1–C5 HIT confirmed R-R2/R-R3; C6/C7 MISS showed the
+// chat registry also resolves catalog-internal aliases (deepseek-v3-2-volc
+// → 200) and more backends (hunyuan-chat → 200). C8 refined R-R4: known
+// model + unsupported family → 11103 "Backend [aiart] is not supported for
+// 3d generation" — but with HTTP 400 (3d), not the 500 seen on chat (P4)
+// and bare 500 on images (P3). R-R4 final form under test: the dispatch
+// failure is 11103 with a per-family envelope: 3d/video → HTTP 400 with
+// "...for <family> generation", chat → HTTP 500 bare wording, images →
+// HTTP 500 with no JSON envelope at all.
+const ARMS_R3 = [
+  { name: 'D1-video-known-chat-model', method: 'POST', path: '/v2/videos/generations',
+    body: { model: 'glm-5.2', prompt: 'x' },
+    expect: { status: 400, code: 11103 }, note: '11103 with "for video generation" wording, mirroring C8' },
+  { name: 'D2-images-known-chat-model', method: 'POST', path: '/v2/images/generations',
+    body: { model: 'kimi-k2.6', prompt: 'x', size: '1024x1024', n: 1 },
+    expect: { status: 500, code: null }, note: 'images family envelope is a bare 500 (P3 replication with a fresh model)' },
+  { name: 'D3-chat-image-model-retest', method: 'POST', path: '/v2/chat/completions',
+    body: { model: 'hunyuan-image-v3.0-art', stream: true, max_tokens: 1, messages: [{ role: 'user', content: 'hi' }] },
+    expect: { status: 500, code: 11103 }, note: 'P4 envelope stability retest' },
+]
+
+// Round 4 (2026-08-19): D1 MISS refined the video family — 14407 is its
+// CATCH-ALL code, not only "route config not found": a globally known chat
+// model on /v2/videos got 400/14407 "unsupported video params". D2/D3 HIT
+// confirmed the images bare-500 and chat 500/11103 envelopes. Final R-R4
+// prediction: a SECOND unobserved chat model on video repeats the catch-all.
+const ARMS_R4 = [
+  { name: 'E1-video-chat-model-2', method: 'POST', path: '/v2/videos/generations',
+    body: { model: 'kimi-k2.6', prompt: 'x' },
+    expect: { status: 400, code: 14407 }, note: 'video catch-all: 14407 "unsupported video params" for any known non-video model' },
+]
+
+async function probe(key, arm) {
+  const t0 = Date.now()
+  try {
+    const res = await fetch(`${GATEWAY}${arm.path}`, {
+      method: arm.method,
+      headers: HEADERS(key),
+      body: JSON.stringify(arm.body),
+    })
+    const text = await res.text()
+    let json = null
+    try { json = JSON.parse(text) } catch { /* SSE stream or HTML */ }
+    return {
+      name: arm.name, path: arm.path, model: arm.body?.model ?? null, expect: arm.expect, note: arm.note,
+      status: res.status, code: json?.code ?? null,
+      msg: redact(json?.msg ?? json?.error_msg ?? (json ? JSON.stringify(json).slice(0, 160) : text.slice(0, 160))),
+      ms: Date.now() - t0,
+    }
+  } catch (err) {
+    return { name: arm.name, path: arm.path, model: arm.body?.model ?? null, expect: arm.expect, note: arm.note, error: String(err?.message ?? err), ms: Date.now() - t0 }
+  }
+}
+
+function judge(rec) {
+  if (!rec.expect) return ''
+  const okStatus = rec.expect.status == null || rec.status === rec.expect.status
+  const okCode = rec.expect.code == null || rec.code === rec.expect.code
+  return okStatus && okCode ? ' HIT' : ' MISS'
+}
+
+const key = resolveKey()
+mkdirSync(dirname(OUT), { recursive: true })
+appendFileSync(OUT, JSON.stringify({ type: 'run', at: new Date().toISOString(), gateway: GATEWAY }) + '\n')
+console.log(`evidence → ${OUT}`)
+
+const SET = process.argv.includes('--set') ? process.argv[process.argv.indexOf('--set') + 1] : 'matrix'
+const arms = SET === 'r2' ? ARMS_R2 : SET === 'r3' ? ARMS_R3 : SET === 'r4' ? ARMS_R4 : ARMS
+
+// anchor: the account's catalog (proves which families are entitled)
+const catRes = await fetch(`${GATEWAY}/v3/config`, { headers: { Accept: 'application/json', Authorization: `Bearer ${key}`, 'x-api-key': key, 'User-Agent': 'CLI/unknown CodeBuddy/2.136.0', 'X-Product': 'SaaS' } })
+const cat = await catRes.json().catch(() => null)
+const models = (cat?.data?.models ?? []).map((m) => m.id)
+const cliModels = (cat?.data?.agents ?? []).find((a) => a.name === 'cli')?.models ?? []
+appendFileSync(OUT, JSON.stringify({ type: 'catalog', at: new Date().toISOString(), models, cliModels }) + '\n')
+console.log(`catalog: ${models.length} models, cli-enabled: ${cliModels.length}`)
+await sleep(SPACING_MS)
+
+for (const arm of arms) {
+  const rec = { type: 'probe', ...await probe(key, arm) }
+  appendFileSync(OUT, JSON.stringify(rec) + '\n')
+  console.log(`${arm.name.padEnd(24)} status=${rec.status ?? '-'} code=${rec.code ?? '-'} msg=${(rec.msg ?? rec.error ?? '').slice(0, 110)}${judge(rec)}`)
+  await sleep(SPACING_MS)
+}
+console.log('done.')

package/scripts/probe-trae-3003-diagnosis.mjs

@@ -0,0 +1,139 @@
+#!/usr/bin/node
+// 诊断探测：Trae inline 面 "3003 all models failed" 故障取证（2026-08-24）。
+//
+// 背景：用户经 dsh-codebuddy-plugin 使用 TraeWork CN 模型报
+//   "trae 3003 all models failed" + PI_AI_ERROR。
+// 本脚本做最小对照实验并落证据盘（docs/probes/trae-3003-diagnosis-<ts>.json）：
+//   [0] 双额度池余额（只读 ide_user_ent_usage，req_source 0/1）
+//   [A] inline_chat + 非默认模型（glm-5.3）
+//   [B] inline_chat + 账户默认模型（kimi-k2.6）
+//   [C] function=chat_v3 + glm-5.3（同信封同鉴权对照——健康则证明故障只在
+//       inline_chat 的模型解析层，非凭据/非信封/非配额）
+//   [R] remote 会话创建（含边缘节点漂移重试观测；成功即停，避免占并发额度）
+// 判读要点：错误事件名在 `event:` 行（data JSON 内无 event 字段）；业务级拒绝
+// 恒为 JSON 信封，**裸文本 404/403 是 TLB/WAF 层行为**（docs/diagnosis-trae-3003.md）。
+// 纪律：raw 面请求间隔 ≥20s（4011 限流很紧）；本脚本单轮共 4 个 chat 请求。
+import fs from 'node:fs'
+import { homedir } from 'node:os'
+import { join } from 'node:path'
+
+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 { readJson } = await import('../core/json-store.js')
+const { createTraeOAuth } = await import('../providers/trae/oauth.js')
+const { buildChatRequest, traeOutboundHeaders } = await import('../providers/trae/gateway.js')
+const { createRemoteSession } = await import('../providers/trae/remote.js')
+
+const SETTINGS = { traeAuthBaseURL: 'https://api.trae.cn', traeChatBaseURL: MCHOST, traeLoginHost: 'https://www.trae.cn' }
+const oauth = createTraeOAuth({ readAuth: () => readJson(AUTH_PATH), writeAuth: (v) => fs.writeFileSync(AUTH_PATH, JSON.stringify(v)) })
+const cred = await oauth.resolveTraeCredential(SETTINGS)
+if (!cred) { console.error('未登录（先在设置卡 TraeWork CN 分区登录或跑 probe-trae-live.mjs --login）'); process.exit(1) }
+const authStore = readJson(AUTH_PATH)
+const token = String(cred.authorization).replace(/^Cloud-IDE-JWT\s+/, '')
+const device = authStore.device ?? {}
+const uid = String(authStore.account?.uid ?? '')
+const ideHeaders = (requestId) => ({
+  ...traeOutboundHeaders(device, uid ? Number(uid) : undefined, requestId),
+  Authorization: `Cloud-IDE-JWT ${token}`,
+  'X-Cloudide-Token': token,
+  'x-ide-token': token,
+})
+const sleep = (ms) => new Promise((r) => setTimeout(r, ms))
+
+/** 解析 raw 面 SSE：event 名在 event: 行；裸 {code} 数据也算业务错误。 */
+function parseRawSse(raw) {
+  let lastEvt = null
+  const events = []
+  for (const line of raw.split('\n')) {
+    const l = line.trim()
+    if (!l) { lastEvt = null; continue }
+    if (l.startsWith('event:')) { lastEvt = l.slice(6).trim(); continue }
+    if (!l.startsWith('data:')) continue
+    try {
+      const j = JSON.parse(l.slice(5).trim())
+      const evName = (typeof j.event === 'string' && j.event) ? j.event : lastEvt
+      const isErr = evName === 'error' || (typeof j.code === 'number' && j.response === undefined && j.usage === undefined)
+      events.push({ ev: evName ?? null, code: isErr ? j.code : undefined, message: isErr ? j.message : undefined, providerModel: j.provider_model_name })
+    } catch { /* 非 JSON data 忽略 */ }
+  }
+  return events
+}
+
+async function rawChatProbe(label, fn, model /* undefined=不带 model 字段 */) {
+  const payload = { stream: true, messages: [{ role: 'user', content: '只回复两个字：成功' }] }
+  if (model !== undefined) payload.model = model
+  const { body, requestId } = buildChatRequest(payload, `diag-${label}`)
+  body.function = fn
+  const t0 = Date.now()
+  const resp = await fetch(`${MCHOST}/api/agent/v3/llm_utils_chat`, {
+    method: 'POST', headers: ideHeaders(requestId), body: JSON.stringify(body), signal: AbortSignal.timeout(60_000),
+  })
+  const raw = await resp.text()
+  const events = parseRawSse(raw)
+  return {
+    label, fn, model: model ?? '(omitted)', status: resp.status,
+    errors: events.filter((e) => e.code != null).slice(0, 2),
+    providerModels: [...new Set(events.map((e) => e.providerModel).filter(Boolean))],
+    hasText: events.some((e) => e.ev === 'output'),
+    ms: Date.now() - t0, rawBytes: raw.length,
+  }
+}
+
+const evidence = {
+  at: new Date().toISOString(), note: '诊断脚本见 scripts/probe-trae-3003-diagnosis.mjs；结论见 docs/diagnosis-trae-3003.md',
+  accountUidMasked: uid.slice(0, 3) + '***', steps: {},
+}
+
+console.log('[0] 双额度池余额（只读）')
+for (const src of [0, 1]) {
+  try {
+    const r = await fetch('https://api.trae.cn/trae/api/v2/pay/ide_user_ent_usage', {
+      method: 'POST', headers: ideHeaders(crypto.randomUUID()),
+      body: JSON.stringify({ require_usage: true, req_source: src }), signal: AbortSignal.timeout(15_000),
+    })
+    const j = await r.json().catch(() => null)
+    evidence.steps[`pool_src${src}`] = (j?.user_entitlement_pack_list ?? []).map((p) => ({
+      endpoint: p?.entitlement_base_info?.available_endpoint, limit: p?.entitlement_base_info?.quota?.credits_limit, used: p?.usage?.credits_amount,
+    }))
+    console.log(`   req_source=${src}:`, JSON.stringify(evidence.steps[`pool_src${src}`]))
+  } catch (e) { evidence.steps[`pool_src${src}`] = { error: String(e?.message ?? e).slice(0, 120) } }
+}
+
+await sleep(21_000)
+console.log('[A] inline_chat + glm-5.3（非默认模型）')
+evidence.steps.A_inline_nonDefault = await rawChatProbe('A', 'inline_chat', 'glm-5.3')
+console.log('  ', JSON.stringify(evidence.steps.A_inline_nonDefault))
+await sleep(21_000)
+console.log('[B] inline_chat + kimi-k2.6（账户默认模型）')
+evidence.steps.B_inline_default = await rawChatProbe('B', 'inline_chat', 'kimi-k2.6')
+console.log('  ', JSON.stringify(evidence.steps.B_inline_default))
+await sleep(21_000)
+console.log('[C] chat_v3 + glm-5.3（对照组：信封/鉴权健康性）')
+evidence.steps.C_chatv3 = await rawChatProbe('C', 'chat_v3', 'glm-5.3')
+console.log('  ', JSON.stringify(evidence.steps.C_chatv3))
+await sleep(21_000)
+console.log('[R] remote 会话创建（观测边缘节点漂移；成功即停善后）')
+try {
+  const s = await createRemoteSession(MCHOST + '/api/remote/v1', token, 'glm-5.3', [{ role: 'user', content: '只回复两个字：成功' }])
+  evidence.steps.R_remote_create = { ok: true, sessionIdMasked: s.sessionId.slice(0, 10) + '…' }
+  console.log('   创建成功 sid=' + s.sessionId.slice(0, 10) + '…')
+} catch (e) {
+  evidence.steps.R_remote_create = { ok: false, error: String(e?.message ?? e).slice(0, 240) }
+  console.log('  ', evidence.steps.R_remote_create.error)
+}
+
+const a = evidence.steps.A_inline_nonDefault, b = evidence.steps.B_inline_default, c = evidence.steps.C_chatv3
+evidence.verdict =
+  a.errors.length && b.errors.length && !c.errors.length
+    ? `INLINE-FACE-BROKEN: inline_chat 对默认与非默认模型一律 ${a.errors[0].code}，同信封 chat_v3 正常 → 服务端 inline 面模型解析故障（非凭据/配额/信封）；解法=切 remote 传输`
+    : a.errors.length && !b.errors.length
+      ? 'DEFAULT-ONLY: 仅非默认模型失败 → 与 trae-cloud-api.md §5.1 一致（function 位钉死到账户默认模型）'
+      : '见证据详情'
+console.log('==>', evidence.verdict)
+
+const outPath = join('docs/probes', `trae-3003-diagnosis-${Date.now()}.json`)
+fs.mkdirSync('docs/probes', { recursive: true })
+fs.writeFileSync(outPath, JSON.stringify(evidence, null, 2))
+console.log('证据落盘 →', outPath)