mohdel 0.111.0 → 0.113.0

package/README.md

@@ -1,6 +1,6 @@
 # Mohdel
 
-Self-hosted LLM gateway and SDK for Node — think LiteLLM, for the JS world. One `answer()` call for 11 providers; swap models by changing one string; get real per-call USD cost back on every result, with OpenTelemetry built in and process isolation when you need it. Your keys, your infra, no SaaS proxy in the path.
+Self-hosted LLM gateway and SDK for Node — think LiteLLM, for the JS world. One `answer()` call for 13 providers; swap models by changing one string; get real per-call USD cost back on every result, with OpenTelemetry built in and process isolation when you need it. Your keys, your infra, no SaaS proxy in the path.
 
 ```bash
 npm install -g mohdel
@@ -8,7 +8,7 @@ mo                                     # interactive setup — pick a provider,
 mo ask gemini/gemini-3-flash-preview "why is the sky blue"
 ```
 
-Providers: Anthropic, OpenAI, Gemini, Mistral, Groq, xAI, Cerebras, Fireworks, DeepSeek, OpenRouter, Novita. Node 22+, ES modules.
+Providers: Anthropic, OpenAI, Gemini, Mistral, Groq, xAI, Cerebras, Fireworks, DeepSeek, Qwen Cloud, Xiaomi, OpenRouter, Novita. Node 22+, ES modules.
 
 ## Why mohdel
 
@@ -306,6 +306,8 @@ What each provider supports through mohdel's unified interface:
 | DeepSeek | No | Yes | Yes | No | No | DSML tool-call fallback when model emits tags in content |
 | Fireworks | Yes | Yes | Yes | No | Yes (`reasoning_effort`) | OpenAI SDK + `baseURL`; model id auto-prefixed |
 | Mistral | No | Yes | Yes | No | No | `tool_choice: "any"` = required |
+| Qwen Cloud | No | Yes | No | No | Yes (`enable_thinking` + `thinking_budget`) | Alibaba DashScope intl; hybrid models think by default — effort `none` sends explicit off |
+| Xiaomi | No | Yes | Yes | No | Auto | MiMo; shared chat-completions path, `reasoning_content` captured |
 | OpenRouter | Yes | Yes | Yes | No | Varies | Meta-provider; `providerOptions.openrouter` for routing prefs |
 | Novita | No | No | No | No | No | Image generation only |
 

package/js/client/call_transcription.js

@@ -0,0 +1,85 @@
+/**
+ * Send a TranscriptionEnvelope to thin-gate's `POST /v1/transcription`.
+ *
+ * Transcription is one-shot: single JSON response body, no streaming,
+ * no cooldown/rate-limit. `audio.fileUri` must be a `file://` or
+ * `data:` URI — `file://` requires that the gate's sessions share a
+ * filesystem with the caller; `data:` carries the bytes inline subject
+ * to the gate's body-size cap.
+ *
+ * @module client/call_transcription
+ */
+
+import { requestUnix } from './transport.js'
+import { MohdelTypedError } from '#core'
+
+/**
+ * @param {import('#core/transcription.js').TranscriptionEnvelope} envelope
+ * @param {object} options
+ * @param {string} options.socketPath
+ * @param {AbortSignal} [options.signal]
+ * @param {string} [options.path]  HTTP path; defaults to '/v1/transcription'
+ * @returns {Promise<import('#core/transcription.js').TranscriptionResult>}
+ */
+export async function callTranscription (envelope, { socketPath, signal, path = '/v1/transcription' }) {
+  const res = await requestUnix({
+    socketPath,
+    path,
+    method: 'POST',
+    body: envelope,
+    signal
+  })
+
+  const body = await readAll(res)
+
+  if (res.statusCode !== 200) {
+    throw MohdelTypedError.fromJSON(parseErrorBody(body, res.statusCode ?? 0))
+  }
+
+  let parsed
+  try {
+    parsed = JSON.parse(body)
+  } catch (e) {
+    throw new MohdelTypedError(
+      'thin-gate returned non-JSON transcription response',
+      { type: 'PROTOCOL_INVALID_EVENT', retryable: false }
+    )
+  }
+
+  if (!parsed || typeof parsed !== 'object' || parsed.status !== 'completed' || typeof parsed.text !== 'string') {
+    throw new MohdelTypedError(
+      'thin-gate returned malformed TranscriptionResult',
+      { type: 'PROTOCOL_INVALID_EVENT', retryable: false }
+    )
+  }
+  return parsed
+}
+
+/**
+ * @param {AsyncIterable<Buffer|string>} stream
+ * @returns {Promise<string>}
+ */
+async function readAll (stream) {
+  let s = ''
+  for await (const c of stream) s += typeof c === 'string' ? c : c.toString('utf8')
+  return s
+}
+
+/**
+ * @param {string} body
+ * @param {number} status
+ * @returns {import('#core/errors.js').TypedError}
+ */
+function parseErrorBody (body, status) {
+  try {
+    const parsed = JSON.parse(body)
+    if (parsed && typeof parsed === 'object' && typeof parsed.type === 'string') {
+      return parsed
+    }
+  } catch {}
+  return {
+    type: 'PROTOCOL_HTTP_ERROR',
+    message: `thin-gate returned HTTP ${status}`,
+    retryable: status >= 500
+  }
+}

package/js/client/index.js

@@ -4,6 +4,7 @@
  * Public surface (0.90):
  *   - call(envelope, { socketPath, signal }): AsyncGenerator<Event>
  *   - callImage(envelope, { socketPath, signal }): Promise<ImageResult>
+ *   - callTranscription(envelope, { socketPath, signal }): Promise<TranscriptionResult>
  *
  * No provider SDKs are imported transitively. This module can be
  * consumed by callers that must not pull openai-node, anthropic-sdk,
@@ -14,3 +15,4 @@
 
 export { call } from './call.js'
 export { callImage } from './call_image.js'
+export { callTranscription } from './call_transcription.js'

package/js/session/adapters/_chat_completions.js

@@ -55,10 +55,14 @@ const DSML_PARAM_RE = /<\uFF5CDSML\uFF5Cparameter\s+name="([^"]+)"(?:\s+string="
  * @property {'openai'|'mistral'|'cerebras'} [toolChoiceFlavor]
  * @property {'user'|'safety_identifier'} [identifierField]
  *   Defaults to 'user'.
- * @property {'reasoning_effort'|'cerebras_zai'} [reasoningField]
+ * @property {'reasoning_effort'|'cerebras_zai'|'qwen'} [reasoningField]
  *   How to wire outputEffort into the request. `reasoning_effort`
  *   sets `args.reasoning_effort = effort`. `cerebras_zai` flips
  *   `args.disable_reasoning = false` instead (zai-family only).
+ *   `qwen` sets `args.enable_thinking` plus a numeric
+ *   `args.thinking_budget` from the spec's effort level — Qwen hybrid
+ *   models think by default, so `enable_thinking: false` must be sent
+ *   explicitly to switch thinking off.
  * @property {boolean} [parseDsml]
  *   Extract DeepSeek DSML function-call blocks from message content
  *   when native `tool_calls` is absent.
@@ -351,6 +355,11 @@ function buildRequest (envelope, spec, config) {
       if (effort !== 'none') delete args.temperature
       if (config.reasoningField === 'cerebras_zai' && /zai/i.test(bareOf(envelope.model))) {
         args.disable_reasoning = (effort === 'none')
+      } else if (config.reasoningField === 'qwen') {
+        args.enable_thinking = (effort !== 'none')
+        if (effort !== 'none' && typeof headroom === 'number') {
+          args.thinking_budget = headroom
+        }
       } else {
         args.reasoning_effort = effort
       }

package/js/session/adapters/_errors.js

@@ -243,6 +243,10 @@ const providerOverrides = {
     if (code === 'rate_limit_exceeded') return tierResult(detail)
     return undefined
   },
+  qwen (_err, code, detail) {
+    if (code === 'rate_limit_exceeded') return tierResult(detail)
+    return undefined
+  },
   novita (_err, code, detail) {
     if (code === 'rate_limit_exceeded') return tierResult(detail)
     return undefined

package/js/session/adapters/index.js

@@ -23,6 +23,7 @@ import { mistral } from './mistral.js'
 import { novita } from './novita.js'
 import { openai } from './openai.js'
 import { openrouter } from './openrouter.js'
+import { qwen } from './qwen.js'
 import { xai } from './xai.js'
 import { xiaomi } from './xiaomi.js'
 
@@ -39,6 +40,7 @@ export const adapters = Object.freeze({
   novita,
   openai,
   openrouter,
+  qwen,
   xai,
   xiaomi
 })

package/js/session/adapters/qwen.js

@@ -0,0 +1,37 @@
+/**
+ * Qwen Cloud adapter — OpenAI-compatible chat completions against
+ * Alibaba's international DashScope endpoint. Reasoning arrives as
+ * the standard `reasoning_content` field handled by the shared core;
+ * thinking is wired via `reasoningField: 'qwen'` (`enable_thinking`
+ * + `thinking_budget`) because Qwen hybrid models think by default.
+ *
+ * @module session/adapters/qwen
+ */
+
+import OpenAI from 'openai'
+
+import { runChatCompletions } from './_chat_completions.js'
+import { streamingDispatcher } from './_dispatcher.js'
+
+const BASE_URL = 'https://dashscope-intl.aliyuncs.com/compatible-mode/v1'
+
+/**
+ * @param {import('#core/envelope.js').CallEnvelope} envelope
+ * @param {{client?: any, signal?: AbortSignal, log?: any, span?: any}} [deps]
+ * @returns {AsyncGenerator<import('#core/events.js').Event>}
+ */
+export async function * qwen (envelope, deps = {}) {
+  const client = deps.client ?? new OpenAI({
+    apiKey: envelope.auth.key,
+    baseURL: envelope.auth.baseURL || BASE_URL,
+    fetchOptions: { dispatcher: streamingDispatcher() }
+  })
+  yield * runChatCompletions(envelope, client, {
+    provider: 'qwen',
+    reasoningField: 'qwen'
+  }, {
+    signal: deps.signal,
+    log: deps.log,
+    span: deps.span
+  })
+}

package/js/session/driver.js

@@ -16,6 +16,7 @@ import readline from 'node:readline'
 
 import { run } from './run.js'
 import { runImage } from './run_image.js'
+import { runTranscription } from './run_transcription.js'
 import { setCatalog } from './adapters/_catalog.js'
 
 // Bounded memory for pre-dequeue cancels. Hostile/buggy supervisors
@@ -148,6 +149,16 @@ export async function drive (stdin, stdout) {
         } else {
           stdout.write(JSON.stringify({ type: 'error', error: out.error }) + '\n')
         }
+      } else if (envelope.op === 'transcription') {
+        // Same one-shot contract as the image path; shape matches
+        // `js/core/transcription.js` after the tag strip.
+        const { op: _op, ...trEnv } = envelope
+        const out = await runTranscription(trEnv)
+        if (out.ok) {
+          stdout.write(JSON.stringify({ type: 'transcription_done', result: out.result }) + '\n')
+        } else {
+          stdout.write(JSON.stringify({ type: 'error', error: out.error }) + '\n')
+        }
       } else {
         for await (const ev of run(envelope, { signal: controller.signal })) {
           stdout.write(JSON.stringify(ev) + '\n')

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mohdel",
-  "version": "0.111.0",
+  "version": "0.113.0",
   "license": "MIT",
   "author": {
     "name": "Christophe Le Bars",
@@ -84,10 +84,10 @@
   },
   "optionalDependencies": {
     "@clack/prompts": "^1.5.1",
-    "@opentelemetry/exporter-trace-otlp-grpc": "^0.218.0",
-    "@opentelemetry/sdk-node": "^0.218.0",
+    "@opentelemetry/exporter-trace-otlp-grpc": "^0.219.0",
+    "@opentelemetry/sdk-node": "^0.219.0",
     "chalk": "^5.4.0",
-    "mohdel-thin-gate-linux-x64-gnu": "0.111.0"
+    "mohdel-thin-gate-linux-x64-gnu": "0.113.0"
   },
   "dependencies": {
     "@anthropic-ai/sdk": "^0.104.1",

package/src/cli/onboard.js

@@ -83,6 +83,13 @@ const PROVIDER_INFO = {
     url: 'https://novita.ai/dashboard/key',
     hint: 'Create an API key at novita.ai → Dashboard → API Key',
     free: false
+  },
+  qwen: {
+    label: 'Qwen Cloud',
+    description: 'Qwen 3.7 Max/Plus, 3.6 Flash — reasoning, coding, long context. Free quota for new users.',
+    url: 'https://home.qwencloud.com/api-keys',
+    hint: 'Create an API key at home.qwencloud.com → API Keys',
+    free: false
   }
 }
 

package/src/lib/providers.js

@@ -88,6 +88,15 @@ const providers = {
     },
     creators: []
   },
+  qwen: {
+    sdk: 'openai',
+    api: 'chatCompletions',
+    apiKeyEnv: 'QWEN_API_SK',
+    createConfiguration: apiKey => ({ baseURL: 'https://dashscope-intl.aliyuncs.com/compatible-mode/v1', apiKey }),
+    creators: ['alibaba'],
+    contextSemantics: 'shared',
+    outputCapStrategy: 'accept'
+  },
   xai: {
     sdk: 'openai',
     apiKeyEnv: 'XAI_API_SK',