mohdel 0.100.0 → 0.102.0

package/js/session/adapters/_chat_completions.js

@@ -330,14 +330,16 @@ function buildRequest (envelope, spec, config) {
 
   if (spec.thinkingEffortLevels) {
     const effort = envelope.outputEffort ?? spec.defaultThinkingEffort ?? 'low'
-    if (effort && effort !== 'none') {
+    if (effort && spec.thinkingEffortLevels[effort] != null) {
       const headroom = spec.thinkingEffortLevels[effort]
       if (args.max_tokens && typeof headroom === 'number') {
         args.max_tokens += headroom
       }
-      delete args.temperature
+      // When reasoning is disabled ('none') the model accepts
+      // temperature again — only strip it when reasoning is on.
+      if (effort !== 'none') delete args.temperature
       if (config.reasoningField === 'cerebras_zai' && /zai/i.test(bareOf(envelope.model))) {
-        args.disable_reasoning = false
+        args.disable_reasoning = (effort === 'none')
       } else {
         args.reasoning_effort = effort
       }

package/js/session/adapters/_dispatcher.js

@@ -0,0 +1,29 @@
+/**
+ * Shared undici dispatcher for streaming-LLM adapters.
+ *
+ * Default `globalThis.fetch` (undici on Node 18+) closes a stream when
+ * no body chunk has arrived for `bodyTimeout` ms — 300 000 ms (5 min)
+ * by default. Reasoning models stream zero bytes during their thinking
+ * phase, so any non-trivial task on a thinking-capable provider can
+ * blow that limit and surface as a `NET_ERROR / "terminated"` mid-run.
+ *
+ * We disable the inter-chunk idle timeout. Cancellation comes from
+ * three layers above us:
+ *   1. caller's `AbortSignal` (per-run timeout, user cancel)
+ *   2. SDK request-level timeout (OpenAI/Anthropic/Groq default 600 s)
+ *   3. provider-side stream limits
+ *
+ * Headers timeout stays bounded — connect + first response must be
+ * fast even when the body afterwards may be slow.
+ *
+ * Singleton: undici Agents own a connection pool, so we keep one per
+ * process and let it manage per-origin keep-alive across all adapters.
+ */
+import { Agent } from 'undici'
+
+let _agent = null
+
+export function streamingDispatcher () {
+  if (!_agent) _agent = new Agent({ bodyTimeout: 0, headersTimeout: 60_000 })
+  return _agent
+}

package/js/session/adapters/anthropic.js

@@ -34,6 +34,7 @@ import {
   fromAnthropicToolCalls,
   toToolChoice
 } from './_tools.js'
+import { streamingDispatcher } from './_dispatcher.js'
 
 /**
  * Approximate chars-per-token used to estimate Anthropic thinking
@@ -80,7 +81,10 @@ const ANTHROPIC_DEFAULT_MAX_TOKENS = 4096
  * @returns {AsyncGenerator<import('#core/events.js').Event>}
  */
 export async function * anthropic (envelope, deps = {}) {
-  const client = deps.client ?? new Anthropic({ apiKey: envelope.auth.key })
+  const client = deps.client ?? new Anthropic({
+    apiKey: envelope.auth.key,
+    fetchOptions: { dispatcher: streamingDispatcher() }
+  })
   const signal = deps.signal
   const log = deps.log
   const start = String(process.hrtime.bigint())

package/js/session/adapters/deepseek.js

@@ -11,6 +11,7 @@
 import OpenAI from 'openai'
 
 import { runChatCompletions } from './_chat_completions.js'
+import { streamingDispatcher } from './_dispatcher.js'
 
 const BASE_URL = 'https://api.deepseek.com'
 
@@ -20,7 +21,11 @@ const BASE_URL = 'https://api.deepseek.com'
  * @returns {AsyncGenerator<import('#core/events.js').Event>}
  */
 export async function * deepseek (envelope, deps = {}) {
-  const client = deps.client ?? new OpenAI({ apiKey: envelope.auth.key, baseURL: envelope.auth.baseURL || BASE_URL })
+  const client = deps.client ?? new OpenAI({
+    apiKey: envelope.auth.key,
+    baseURL: envelope.auth.baseURL || BASE_URL,
+    fetchOptions: { dispatcher: streamingDispatcher() }
+  })
   yield * runChatCompletions(envelope, client, {
     provider: 'deepseek',
     parseDsml: true

package/js/session/adapters/fireworks.js

@@ -16,6 +16,7 @@
 import OpenAI from 'openai'
 
 import { runChatCompletions } from './_chat_completions.js'
+import { streamingDispatcher } from './_dispatcher.js'
 
 const BASE_URL = 'https://api.fireworks.ai/inference/v1'
 
@@ -27,7 +28,8 @@ const BASE_URL = 'https://api.fireworks.ai/inference/v1'
 export async function * fireworks (envelope, deps = {}) {
   const client = deps.client ?? new OpenAI({
     apiKey: envelope.auth.key,
-    baseURL: envelope.auth.baseURL || BASE_URL
+    baseURL: envelope.auth.baseURL || BASE_URL,
+    fetchOptions: { dispatcher: streamingDispatcher() }
   })
   yield * runChatCompletions(envelope, client, {
     provider: 'fireworks',

package/js/session/adapters/groq.js

@@ -7,6 +7,7 @@
 import Groq from 'groq-sdk'
 
 import { runChatCompletions } from './_chat_completions.js'
+import { streamingDispatcher } from './_dispatcher.js'
 
 /**
  * @param {import('#core/envelope.js').CallEnvelope} envelope
@@ -14,7 +15,10 @@ import { runChatCompletions } from './_chat_completions.js'
  * @returns {AsyncGenerator<import('#core/events.js').Event>}
  */
 export async function * groq (envelope, deps = {}) {
-  const client = deps.client ?? new Groq({ apiKey: envelope.auth.key })
+  const client = deps.client ?? new Groq({
+    apiKey: envelope.auth.key,
+    fetchOptions: { dispatcher: streamingDispatcher() }
+  })
   yield * runChatCompletions(envelope, client, { provider: 'groq' }, {
     signal: deps.signal,
     log: deps.log,

package/js/session/adapters/mistral.js

@@ -10,6 +10,7 @@
 import OpenAI from 'openai'
 
 import { runChatCompletions } from './_chat_completions.js'
+import { streamingDispatcher } from './_dispatcher.js'
 
 const BASE_URL = 'https://api.mistral.ai/v1'
 
@@ -19,7 +20,11 @@ const BASE_URL = 'https://api.mistral.ai/v1'
  * @returns {AsyncGenerator<import('#core/events.js').Event>}
  */
 export async function * mistral (envelope, deps = {}) {
-  const client = deps.client ?? new OpenAI({ apiKey: envelope.auth.key, baseURL: envelope.auth.baseURL || BASE_URL })
+  const client = deps.client ?? new OpenAI({
+    apiKey: envelope.auth.key,
+    baseURL: envelope.auth.baseURL || BASE_URL,
+    fetchOptions: { dispatcher: streamingDispatcher() }
+  })
   yield * runChatCompletions(envelope, client, {
     provider: 'mistral',
     toolChoiceFlavor: 'mistral'

package/js/session/adapters/openai.js

@@ -35,6 +35,7 @@ import {
   fromOpenAIToolCalls,
   toToolChoice
 } from './_tools.js'
+import { streamingDispatcher } from './_dispatcher.js'
 
 /**
  * @param {import('#core/envelope.js').CallEnvelope} envelope
@@ -44,7 +45,8 @@ import {
 export async function * openai (envelope, deps = {}) {
   const client = deps.client ?? new OpenAI({
     apiKey: envelope.auth.key,
-    ...(envelope.auth.baseURL ? { baseURL: envelope.auth.baseURL } : {})
+    ...(envelope.auth.baseURL ? { baseURL: envelope.auth.baseURL } : {}),
+    fetchOptions: { dispatcher: streamingDispatcher() }
   })
   const signal = deps.signal
   const log = deps.log
@@ -223,19 +225,18 @@ function buildRequest (envelope, input, instructions) {
 
   // Thinking: when the spec has `thinkingEffortLevels`, set
   // `reasoning.effort` and add the thinking-budget headroom on top
-  // of the user's `outputBudget`. `reasoning` is an OpenAI-only
-  // parameter — xAI reasoning is automatic, so add the headroom
-  // but skip the request field on xAI.
+  // of the user's `outputBudget`. Both OpenAI (gpt-5.x) and xAI
+  // (grok-4.3+) accept the same `reasoning: { effort }` shape on
+  // the Responses API, including the literal value 'none' to
+  // disable reasoning entirely.
   if (spec?.thinkingEffortLevels) {
     const effort = envelope.outputEffort ?? spec.defaultThinkingEffort ?? 'low'
-    if (effort && effort !== 'none') {
+    if (effort && spec.thinkingEffortLevels[effort] != null) {
       const headroom = spec.thinkingEffortLevels[effort]
       if (request.max_output_tokens && typeof headroom === 'number') {
         request.max_output_tokens += headroom
       }
-      if (provider === 'openai') {
-        request.reasoning = { effort }
-      }
+      request.reasoning = { effort }
     }
   }
 

package/js/session/adapters/openrouter.js

@@ -12,6 +12,7 @@
 import OpenAI from 'openai'
 
 import { runChatCompletions } from './_chat_completions.js'
+import { streamingDispatcher } from './_dispatcher.js'
 
 const BASE_URL = 'https://openrouter.ai/api/v1'
 
@@ -47,7 +48,8 @@ export async function * openrouter (envelope, deps = {}) {
   const client = deps.client ?? new OpenAI({
     apiKey: envelope.auth.key,
     baseURL: envelope.auth.baseURL || BASE_URL,
-    defaultHeaders
+    defaultHeaders,
+    fetchOptions: { dispatcher: streamingDispatcher() }
   })
 
   yield * runChatCompletions(envelope, client, {

package/js/session/adapters/xai.js

@@ -10,6 +10,7 @@
 import OpenAI from 'openai'
 
 import { openai } from './openai.js'
+import { streamingDispatcher } from './_dispatcher.js'
 
 const BASE_URL = 'https://api.x.ai/v1'
 
@@ -21,7 +22,8 @@ const BASE_URL = 'https://api.x.ai/v1'
 export async function * xai (envelope, deps = {}) {
   const client = deps.client ?? new OpenAI({
     apiKey: envelope.auth.key,
-    baseURL: envelope.auth.baseURL || BASE_URL
+    baseURL: envelope.auth.baseURL || BASE_URL,
+    fetchOptions: { dispatcher: streamingDispatcher() }
   })
   yield * openai(envelope, { ...deps, client })
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mohdel",
-  "version": "0.100.0",
+  "version": "0.102.0",
   "license": "MIT",
   "author": {
     "name": "Christophe Le Bars",
@@ -84,26 +84,27 @@
   },
   "optionalDependencies": {
     "@clack/prompts": "^1.3.0",
-    "@opentelemetry/exporter-trace-otlp-grpc": "^0.216.0",
-    "@opentelemetry/sdk-node": "^0.216.0",
+    "@opentelemetry/exporter-trace-otlp-grpc": "^0.217.0",
+    "@opentelemetry/sdk-node": "^0.217.0",
     "chalk": "^5.4.0",
-    "mohdel-thin-gate-linux-x64-gnu": "0.100.0"
+    "mohdel-thin-gate-linux-x64-gnu": "0.102.0"
   },
   "dependencies": {
-    "@anthropic-ai/sdk": "^0.91.1",
+    "@anthropic-ai/sdk": "^0.95.1",
     "@cerebras/cerebras_cloud_sdk": "^1.61.1",
-    "@google/genai": "^1.51.0",
+    "@google/genai": "^2.0.0",
     "@opentelemetry/api": "^1.9.1",
     "env-paths": "^4.0.0",
     "groq-sdk": "^1.1.2",
-    "openai": "^6.35.0"
+    "openai": "^6.37.0",
+    "undici": "^7.24.5"
   },
   "lint-staged": {
     "*.{js,cjs}": "standard"
   },
   "devDependencies": {
     "gpt-tokenizer": "^3.4.0",
-    "lint-staged": "^16.4.0",
+    "lint-staged": "^17.0.3",
     "release-it": "^20.0.1",
     "standard": "^17.1.2",
     "vitest": "^4.1.5"