dsh-router-core 0.4.5 → 0.4.7

package/src/shared.test.ts

@@ -0,0 +1,35 @@
+/**
+ * 外部 OpenAI 客户端归因标签（clientIdentityOf）的闸门。
+ *
+ * 2026-10-15 事故反查暴露的盲区：`/v1` 没有宿主内部头的请求全落成匿名
+ * usage 记录（296 条对不上任何会话）。这里锁住三级兜底的优先级与边界。
+ * 用 node --test 跑（Node 原生 TS 剥离，零新依赖）。
+ */
+import { test } from 'node:test'
+import assert from 'node:assert/strict'
+import { clientIdentityOf } from './shared.ts'
+
+test('归因优先级：user > prompt_cache_key > User-Agent', () => {
+  assert.equal(clientIdentityOf({ user: 'zed-thread-1', prompt_cache_key: 'k' }, 'Zed/0.2'), 'ext/user:zed-thread-1')
+  assert.equal(clientIdentityOf({ prompt_cache_key: 'k1' }, 'Zed/0.2'), 'ext/cache:k1')
+  assert.equal(clientIdentityOf({}, 'Zed/0.2'), 'ext/ua:Zed/0.2')
+  assert.equal(clientIdentityOf({}, undefined), undefined, '什么都没有 → 保持匿名')
+})
+
+test('空串/非字符串的 user 与 prompt_cache_key 当没有，落 UA 兜底', () => {
+  assert.equal(clientIdentityOf({ user: '' }, 'Zed/0.2'), 'ext/ua:Zed/0.2')
+  assert.equal(clientIdentityOf({ user: 123 }, 'Zed/0.2'), 'ext/ua:Zed/0.2')
+  assert.equal(clientIdentityOf({ user: '', prompt_cache_key: '' }, ''), undefined)
+})
+
+test('值截断到 64 字符（防超长标识灌进 usage 记录）', () => {
+  const long = 'x'.repeat(100)
+  assert.equal(clientIdentityOf({ user: long }, undefined), `ext/user:${'x'.repeat(64)}`)
+  assert.equal(clientIdentityOf({}, long), `ext/ua:${'x'.repeat(64)}`)
+})
+
+test('宿主会话 id 形态与外部标签形态不撞（lastHit 过滤互不误命中）', () => {
+  const host = 'session-72847baa-e309-42ad-aa6b-15107cdc6518'
+  const ext = clientIdentityOf({ user: '72847baa-e309-42ad-aa6b-15107cdc6518' }, undefined)
+  assert.notEqual(ext, host, '前缀区分宿主/外部，精确匹配过滤不会串')
+})

package/src/shared.ts

@@ -192,6 +192,17 @@ export interface ExtControlItem {
   body?: string
   /** 文本可编辑（缺省 = 只读；见 `ExtControl.editable`）。 */
   editable?: boolean
+  /**
+   * 分组名（可选；缺省 = 无组）—— 详情页按它分节渲染，组间插组标题行；
+   * 无组照旧平铺（见 `ExtControl.group`）。
+   */
+  group?: string
+  /**
+   * 参与拖动排序（可选；缺省 = 参与）—— `false` 的条目不给拖把手
+   * （见 `ExtControl.sortable`）。⚠️ 服务端**带值透传**，`false` 也会收到；
+   * 判断必须写 `sortable !== false`，`!sortable` 会把缺省（参与）当不参与。
+   */
+  sortable?: boolean
   /**
    * 自建条目（可删）还是内置（不可删、只有开关与还原）。
    *
@@ -321,3 +332,25 @@ export interface RouterChartResponse {
   error?: string
   chart?: RouterChartBucket[]
 }
+
+/**
+ * 外部 OpenAI 兼容客户端的归因标签（`/v1` 没有宿主内部头 `x-dsh-router-session`
+ * 时的兜底）——否则 usage 记录里全是匿名流量，出事反查不出是谁在打
+ * （2026-10-15 实测：296 条匿名记录对不上任何会话）。
+ *
+ * 取值优先级：`user`（OpenAI 标准终端用户字段）→ `prompt_cache_key`
+ * （Zed 等客户端的缓存键，常带线程粒度）→ `User-Agent`（保底到客户端粒度）。
+ * 值截断到 64 字符；三者皆无返回 undefined（保持匿名）。
+ */
+export function clientIdentityOf(
+  body: { user?: unknown; prompt_cache_key?: unknown },
+  userAgent: string | undefined,
+): string | undefined {
+  const tag = (kind: string, v: unknown): string | undefined =>
+    typeof v === 'string' && v !== '' ? `ext/${kind}:${v.slice(0, 64)}` : undefined
+  return (
+    tag('user', body.user)
+    ?? tag('cache', body.prompt_cache_key)
+    ?? (userAgent !== undefined && userAgent !== '' ? `ext/ua:${userAgent.slice(0, 64)}` : undefined)
+  )
+}

package/src/suppliers/contract.ts

@@ -13,6 +13,7 @@ import type { Context } from '@deepseek-ai/cordis'
 import type { ChatRequest, ModelInfo, SupplierAccount } from '../router/types.ts'
 import type { SupplierConfigStore } from '../supplier-config.ts'
 import type { CredentialStore } from '../credential-store.ts'
+import type { ProtocolId } from '../protocol/types.ts'
 
 /** 供应商配置：loader 注入的运行时环境。 */
 export interface SupplierEnv {
@@ -47,9 +48,16 @@ export interface SupplierEnv {
  * `source` 是**可选标签**，声明这个供应商是从哪儿来的。核心用它把 `/health` 的
  * `source` 报准（面板按它分「内置 / 插件」两组，面板和插件页都看这个值）。
  * 不声明 = `external`（独立安装的供应商插件的现状，一个字段都不用改）。
+ *
+ * `wire` 是**协议标签**（可选）：这个供应商的上游说哪种协议。缺省
+ * `'chat'`（= OpenAI Chat Completions，今天所有供应商的现状，一个都不用改）。
+ * 声明成别种时，核心在把 `req` 交给插件前把规范请求体翻成它的协议，
+ * 并把它的响应流翻回规范形态 —— 插件看到的仍是「OpenAI SSE」，
+ * 于是**契约不变、已有插件全都不动**。详见 `protocol/`。
  */
 export type SupplierFactory = ((env: SupplierEnv) => SupplierModule) & {
   source?: 'builtin'
+  wire?: ProtocolId
 }
 
 /**
@@ -117,8 +125,13 @@ export type ChatOnceResult =
   | { ok: true; stream: ReadableStream<Uint8Array> }
   /** 非流式：核心写 JSON。 */
   | { ok: true; status: number; body: string }
-  /** 失败：核心据此做冷却/换号。 */
-  | { ok: false; state: AccountState; message: string }
+  /**
+   * 失败：核心据此做冷却/换号。
+   * `resetAt` = 上游自己声明的重置时刻（epoch ms，可选）：限流类错误带它时
+   * 核心冷却到该点，而不是指数退避——频率配额的重置是小时级的，5min 上限的
+   * 退避只会到点回来再撞（2026-10-15 的 6004 事故：11 个号轮流被打冷）。
+   */
+  | { ok: false; state: AccountState; message: string; resetAt?: number }
 
 /** 供应商模块 —— 契约（核心必须，差异化可选）。 */
 export interface SupplierModule {

package/src/suppliers/opencode/plugin.test.ts

@@ -1,17 +1,24 @@
 /**
- * opencode(Zen 免费档)握手契约测试 —— 锁住 2026-09-22 实测的那组硬门。
+ * opencode(OpenCode Zen 免费档)握手契约测试 —— 锁住 2026-10-15 实测的那组硬门。
  *
- * 为什么值得固化：这四条**都不是可选优化，缺一即 403**
+ * 为什么值得固化：这些**都不是可选优化，缺一即 403**
  * `FreeTierError: OpenCode's free tier can only be used from within OpenCode`。
- * 它们全是「看不出来」的那种约定——代码删掉一行 headers、或把 UA 改回
- * `opencode`，单测本该立刻变红，否则下次只能在真机对话里发现整个供应商死了。
+ * 它们全是「看不出来」的那种约定——代码删掉一行、或把端点/工具名改回去，
+ * 单测本该立刻变红，否则下次只能在真机对话里发现整个供应商死了。
  *
- * 实测依据（2026-09-22，逐项消融验证）：
- *   - UA 必须带版本且 ≥1.18：裸 `opencode`→403；`opencode/1.17.x`→426；1.18.0 起 200
- *   - 必须同时有 `x-opencode-session`/`x-opencode-request`，且形状是
- *     `ses_<12hex><14base62>` / `msg_<同形状>`——格式错也 403
- *   - tools 里必须**同时**有 bash 和 read，只给一个也 403
- *   - 非流式（`stream:false`）即使三样齐全也 403 → 必须恒发 stream:true
+ * 实测依据（2026-10-15，逐项消融 + 单因子隔离）：
+ *   - 端点必须是 `/inference/v1/chat/completions`：`/zen/v1` 已废弃，
+ *     对任何头部/工具组合一律 403（旧实现死在这里）
+ *   - `tools` 里必须存在**小写精确名为 `read`** 的工具：
+ *     `read` 单独→200；`bash`/`glob`/`write`/`shell` 单独→403；
+ *     `read` 改名 `READ`→403；但 `read` 的 `parameters` 完全不看
+ *     （`properties:{}`、无 `parameters`、`parameters:null` 全 200）
+ *   - session 的 12 位 hex 前缀必须是倒序时间戳且低 12 位 ∈ [0,63]：
+ *     纯随机 hex→403、sha256 派生→403、counter=64/300→403
+ *   - 非流式（`stream:false`）即使其余齐全也 403 → 必须恒发 stream:true
+ *
+ * 已实测**不敏感**（故不作为断言，改回旧值也不算错）：UA 具体版本、
+ * `x-opencode-client` 的 cli/desktop、工具 description。
  *
  * 用 node --test 跑（Node 原生剥 TS 类型）：
  *   node --test src/suppliers/opencode/plugin.test.ts
@@ -28,6 +35,19 @@ import type { ChatRequest } from '../../router/types.ts'
 const SES_RE = /^ses_[0-9a-f]{12}[0-9A-Za-z]{14}$/
 const MSG_RE = /^msg_[0-9a-f]{12}[0-9A-Za-z]{14}$/
 
+/** 上游免费档的推理端点。`/zen/v1` 已废弃，出现即回归。 */
+const EXPECTED_CHAT_URL = 'https://opencode.ai/inference/v1/chat/completions'
+
+/**
+ * 从 session id 还原出 counter：12 位 hex 是 `~(ts * 0x1000 + counter)`，
+ * 故先取反还原 `ts * 0x1000 + counter`，再取低 12 位即 counter
+ * （`ts * 0x1000` 的低 12 位恒为 0）。
+ */
+function invertedCounter(id: string): bigint {
+  const time = BigInt('0x' + id.slice(4, 16))
+  return ~time & 0xffffffffffffn & 0xfffn
+}
+
 interface Captured {
   url: string
   headers: Record<string, string>
@@ -70,25 +90,38 @@ function request(over: Partial<ChatRequest> = {}): ChatRequest {
   }
 }
 
-test('chatOnce 必须带版本化 UA + 合法形状的 session/request 头', async () => {
+test('chat 必须打 /inference/v1（/zen/v1 已废弃，全组合 403）', async () => {
   const stub = stubFetch()
   try {
     const r = await makePlugin().chatOnce('', 'auto', request())
     assert.equal(r.ok, true, '握手齐全时应当成功')
-    const h = stub.captured[0]!.headers
+    assert.equal(stub.captured[0]!.url, EXPECTED_CHAT_URL, '/zen/v1 是本次故障根因，不得回退')
+  } finally {
+    stub.restore()
+  }
+})
 
-    // UA 必须带版本号（裸 'opencode' 会被上游按非 CLI 拒掉）
-    assert.match(h['User-Agent'] ?? '', /^opencode\/\d+\.\d+/, 'UA 必须带 opencode/ 版本号')
+test('session 的 12 位 hex 必须是倒序时间戳，低 12 位 counter ∈ [0,63]', async () => {
+  const stub = stubFetch()
+  try {
+    await makePlugin().chatOnce('', 'auto', request())
+    const h = stub.captured[0]!.headers
 
-    // session/request 形状不对上游是**拒绝**而不是忽略 —— 形状必须严格
+    // 形状不对上游是**拒绝**而不是忽略 —— 形状必须严格
     assert.match(h['x-opencode-session'] ?? '', SES_RE, 'x-opencode-session 形状不对会被 403')
     assert.match(h['x-opencode-request'] ?? '', MSG_RE, 'x-opencode-request 形状不对会被 403')
+
+    // 纯随机 hex 与 sha256 派生 hex 都落不进 [0,63] 窗口 → 必然 403
+    for (const key of ['x-opencode-session', 'x-opencode-request'] as const) {
+      const counter = invertedCounter(h[key]!)
+      assert.ok(counter <= 63n, `${key} 的 counter 必须 ∈ [0,63]（实测 64/300 会 403），得到 ${counter}`)
+    }
   } finally {
     stub.restore()
   }
 })
 
-test('请求体恒为流式，且 tools 同时含 bash 与 read', async () => {
+test('tools 必须含小写 read；bash 不是门，且不得被注入（旧实现的回归点）', async () => {
   const stub = stubFetch()
   try {
     // 调用方明确要非流式 —— 上游免费档对非流式一律 403，插件必须翻成流式
@@ -96,32 +129,53 @@ test('请求体恒为流式，且 tools 同时含 bash 与 read', async () => {
     const body = stub.captured[0]!.body
     assert.equal(body.stream, true, '免费档只支持流式，非流式必须翻成 stream:true')
 
-    const names = (body.tools as Array<{ function?: { name?: string } }>).map((t) => t.function?.name)
-    assert.ok(names.includes('bash'), 'tools 缺 bash 会被 403')
+    const tools = body.tools as Array<{ function?: { name?: string } }>
+    const names = tools.map((t) => t.function?.name)
     assert.ok(names.includes('read'), 'tools 缺 read 会被 403')
+    // 实测：bash 不是门（bash 单独 → 403），注入它只会污染调用方的工具表
+    assert.equal(names.includes('bash'), false, 'bash 不是免费档的门，不该被注入')
+    assert.equal(tools.length, 1, '无调用方工具时只应补 read 一个占位')
+  } finally {
+    stub.restore()
+  }
+})
+
+test('read 的 parameters 不被检查（空 schema 即可通过）', async () => {
+  const stub = stubFetch()
+  try {
+    await makePlugin().chatOnce('', 'auto', request())
+    const tools = stub.captured[0]!.body.tools as Array<{ function: { name: string; parameters?: unknown } }>
+    const read = tools.find((t) => t.function.name === 'read')!
+    // 上游只看名字，故占位用空 properties 是对的（别"补"成真 schema 增加噪音）
+    assert.deepEqual(read.function.parameters, { type: 'object', properties: {} })
   } finally {
     stub.restore()
   }
 })
 
-test('调用方自带的同名工具不被占位工具覆盖（bash/read 只补缺）', async () => {
+test('调用方自带的同名工具不被占位工具覆盖（read 只补缺）', async () => {
   const stub = stubFetch()
   try {
+    const realRead = {
+      type: 'function',
+      function: { name: 'read', description: 'REAL READ', parameters: { type: 'object', properties: { filePath: { type: 'string' } } } },
+    }
     const realBash = {
       type: 'function',
       function: { name: 'bash', description: 'REAL BASH', parameters: { type: 'object', properties: { command: { type: 'string' } } } },
     }
     await makePlugin().chatOnce('', 'auto', request({
-      rawBody: JSON.stringify({ model: 'big-pickle', messages: [], stream: true, tools: [realBash] }),
+      rawBody: JSON.stringify({ model: 'big-pickle', messages: [], stream: true, tools: [realRead, realBash] }),
     }))
     const tools = stub.captured[0]!.body.tools as Array<{ function: { name: string; description?: string } }>
 
-    const bash = tools.filter((t) => t.function.name === 'bash')
-    assert.equal(bash.length, 1, 'bash 不能被注入成两条（否则真实工具签名被挤掉）')
-    assert.equal(bash[0]!.function.description, 'REAL BASH', '调用方自己的定义必须原样保留')
+    const read = tools.filter((t) => t.function.name === 'read')
+    assert.equal(read.length, 1, 'read 不能被注入成两条（否则真实工具签名被挤掉）')
+    assert.equal(read[0]!.function.description, 'REAL READ', '调用方自己的定义必须原样保留')
+    assert.equal(tools.length, 2, '已有 read 时不该再补任何占位')
 
-    // read 缺失 → 补上；bash 已有 → 不重复
-    assert.ok(tools.some((t) => t.function.name === 'read'), '缺的 read 要补上')
+    // 调用方的 bash 必须原样留着（它不是门，但也绝不能被删）
+    assert.equal(tools.filter((t) => t.function.name === 'bash').length, 1, '调用方的 bash 不得被动')
   } finally {
     stub.restore()
   }

package/src/suppliers/opencode/plugin.ts

@@ -1,14 +1,33 @@
 /**
- * opencode 供应商插件 —— 参考 9Router(open-sse) 的 opencode free 实现。
+ * opencode 供应商插件 —— OpenCode Zen 免费档（匿名 `Bearer public` 通道）。
  *
- * 上游：https://opencode.ai/zen/v1（OpenAI 兼容，免费档）
- *   - chat:  POST /zen/v1/chat/completions（**只能流式**）
- *   - models: GET  /zen/v1/models（透传，只取免费模型）
+ * 上游（2026-10-15 实测）：
+ *   - chat:   POST https://opencode.ai/inference/v1/chat/completions（**只能流式**）
+ *   - models: GET  https://opencode.ai/inference/v1/models（两个端点都通，取免费模型）
  *
- * 「免费」不等于「无鉴权」：Zen 免费档仍要求**模拟真 CLI 的握手**，否则一律
- * 403 `FreeTierError: OpenCode's free tier can only be used from within OpenCode`
- * （2026-09-22 实测）。三个条件缺一即 403，见下方 `UA` / `sessionId` / `decoyTools`。
- * 免费档也没有账号可填 —— 所以无凭证/连接池/签到，面板只显示模型卡片
+ * ⚠️ `/zen/v1/chat/completions` **已废弃**：2026-10-15 全矩阵实测该端点对任何
+ * 头部/工具组合一律 403，迁到 `/inference/v1` 后同样请求立刻 200。旧的
+ * `zen/v1` 常量是本次故障的根因之一，别再改回去。
+ * 注：2026-10-16 复核，`/zen/v1/chat/completions` 在握手完全正确时也能 200
+ * （即"一律 403"仅对当时的随机 hex session 成立）；两个端点的差异对正确握手
+ * 不可见。`/inference/v1` 仍是**实测更稳**的那个，保留不回退。
+ *
+ * 「免费」不等于「无鉴权」：仍要求**模拟真 CLI 的握手**，否则一律
+ * 403 `FreeTierError: OpenCode's free tier can only be used from within OpenCode`。
+ * 2026-10-15 逐项消融后，真正的门只剩两条（其余全不敏感）：
+ *
+ *   1. `tools` 里必须存在一个**小写精确名为 `read`** 的工具。
+ *      —— `parameters` 完全不看（实测 `properties:{}`、无 `parameters`、
+ *      `parameters:null` 全部 200）；但名字大小写敏感（`READ` → 403），且
+ *      `bash`/`glob`/`write` 等**不能替代** `read`（各自单独都给 403）。
+ *   2. `x-opencode-session` 的 12 位 hex 前缀必须是**倒序时间戳**，且其低 12 位
+ *      （counter）落在 `[0,63]`（见 `newSessionId`）。
+ *
+ * 已实测**不敏感**（改了照样 200，故不写死）：UA 具体版本（`opencode/1.18.31`
+ * 与 `opencode/latest/2.0.29/cli` 等价）、`x-opencode-client` 的 `cli`/`desktop`、
+ * `x-opencode-request` 有无、`x-opencode-project` 是否 40-hex、工具 `description`。
+ *
+ * 免费档没有账号可填 —— 所以无凭证/连接池/签到，面板只显示模型卡片
  * （连接池、签到由能力检测自动隐藏）。
  */
 import { randomBytes } from 'node:crypto'
@@ -23,32 +42,38 @@ export const priority = 0
 /** 面板图标。 */
 export const icon = 'data:image/png;base64,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'
 
-const BASE = 'https://opencode.ai/zen/v1'
+/** 上游免费档的推理端点。`/zen/v1` 已于 2026-10-15 废弃（全组合 403）。 */
+const BASE = 'https://opencode.ai/inference/v1'
 const MODELS_URL = `${BASE}/models`
 const CHAT_URL = `${BASE}/chat/completions`
 
 /**
- * User-Agent 必须带**版本号**且 ≥ 1.18。
- *
- * 裸 `opencode` → 403 FreeTierError；`opencode/1.17.x` → **426 UpgradeRequired**
- * （上游明确要求升级客户端）；`opencode/1.18.0` 起通过。别去掉版本号，
- * 也别写死成能过的最高版本——取一个已实测通过的稳定值即可。
+ * User-Agent。**不是**门禁（实测 `opencode/1.18.31` 与真 CLI 的
+ * `opencode/1.19.0 ai-sdk/...` 等价，裸 `opencode` 也不影响本门）。
+ * 保留一个带版本号的值只是别让上游按非 CLI 流量处理。
  */
 const UA = 'opencode/1.18.31'
 
 /**
- * 免费档按 session 记账（每次生成一个新的即可——见 `chatOnce` 的实测说明），
- * `x-opencode-session` / `x-opencode-request` **必须**是上游认的形状，格式不对
- * 同样 403（不是忽略，是拒绝）：
- *   ses_<12 位小写 hex><14 位 base62>、msg_<同形状>
+ * 造上游认的 session/request id。
+ *
+ * ⚠️ 12 位 hex 前缀**不是任意 hex，也不是摘要派生**，它是真客户端的
+ * `~(Date.now() * 0x1000 + counter)`：倒序时间戳。2026-10-15 实测（逐项消融）：
+ *   - 倒序时间戳 counter=0/1/62/63 → 200；counter=64/65/100 → 403
+ *   - 纯随机 hex → 403；sha256 派生 hex → 403（9router #4780 的成因）
+ * 即上游会把低 12 位当每毫秒计数器读，只接受 `[0,63]`；随机/派生值落进该窗口
+ * 只有约 1.6% 概率，所以旧实现几乎必然 403。（2026-10-16 每个边界值各打 3 次
+ * 复核：63 三次全 200、64 三次全 403，边界就是 63/64 之间。）
+ *
+ * counter 固定取 1：本插件是自己 session 的唯一写入方，做每毫秒计数器只增加状态。
  */
 function newSessionId(prefix: 'ses_' | 'msg_'): string {
   const B62 = '0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz'
-  const bytes = randomBytes(20)
-  let time = ''
-  for (let i = 0; i < 6; i++) time += bytes[i]!.toString(16).padStart(2, '0')
+  const current = BigInt(Date.now()) * 0x1000n + 1n
+  const time = (~current & 0xffffffffffffn).toString(16).padStart(12, '0')
+  const bytes = randomBytes(14)
   let rand = ''
-  for (let i = 6; i < 20; i++) rand += B62[bytes[i]! % 62] ?? ''
+  for (let i = 0; i < 14; i++) rand += B62[bytes[i]! % 62] ?? ''
   return `${prefix}${time}${rand}`
 }
 
@@ -67,15 +92,21 @@ function sessionHeaders(): Record<string, string> {
 }
 
 /**
- * 免费档要求 tools 里**同时**存在 `bash` 和 `read`，否则 403。
+ * 免费档要求的占位工具：**只需要 `read` 一个**，小写精确。
+ *
+ * 2026-10-15 逐项消融实测：
+ *   - `read` 单独 → 200；`bash` 单独 → 403；`glob`/`write`/`shell` 单独 → 403
+ *   - `read` + 任意其他（含空 schema 的 bash/glob/grep…）→ 200
+ *   - 把 `read` 改名成 `READ` → 403（大小写敏感）
+ *   - `read` 的 `parameters` 完全不看：`properties:{}`、无 `parameters`、
+ *     `parameters:null`、`parameters:{}` 全部 200
  *
- * 为什么是「注入」而不是「覆盖」：上游只是**校验这两个名字在不在**，真正的
- * 工具照常工作（实测真实工具在前、占位在后，混合下发 200）。所以调用方已经
- * 带了同名工具时保持原样不动，缺哪个补哪个——绝不覆盖调用方自己的定义，
- * 否则会把真实工具签名糊掉。
+ * 所以旧版「必须同时有 bash 和 read」是错的——**`bash` 不是门，`read` 才是**。
+ * 为什么是「注入」而不是「覆盖」：上游只校验名字在不在，真实工具照常工作，
+ * 所以调用方已带同名工具时保持原样不动，缺哪个补哪个——绝不覆盖调用方自己的
+ * 定义，否则会把真实工具签名糊掉。
  */
 const DECOY_TOOLS: ReadonlyArray<{ name: string; description: string }> = [
-  { name: 'bash', description: 'This tool is currently unavailable and must not be used.' },
   { name: 'read', description: 'This tool is currently unavailable and must not be used.' },
 ]
 
@@ -178,7 +209,7 @@ export default function factory(env: SupplierEnv): SupplierModule {
         // 免费档**只支持流式**：`stream:false` 即使握手齐全也 403（实测）。
         // 客户端要非流式时代理层会聚合 SSE，所以这里恒为 true。
         obj.stream = true
-        // 上游要求 tools 里同时有 bash/read，否则 403（缺哪个补哪个）
+        // 上游要求 tools 里存在小写 `read`，否则 403（缺就补，已有则原样保留）
         ensureDecoyTools(obj)
         // effort 白名单：'off'/'none' 必须**删字段**（上游没有这两个值，
         // 传 'off' 直接 400 invalid_request_error）；low/high/max 实测可过。