mohdel 0.97.1 → 0.98.1

package/js/factory/bridge.js

@@ -71,7 +71,7 @@ import { createRealtimeDeltaBuffer } from '../../src/lib/utils.js'
  * @returns {Promise<any>}            AnswerResult (matches the factory's return shape).
  */
 export async function runAnswer ({ provider, model, modelKey, configuration, prompt, options = {} }, deps = {}) {
-  const envelope = toEnvelope({ provider, model, configuration, prompt, options })
+  const envelope = toEnvelope({ modelKey, configuration, prompt, options })
 
   // If the caller passed a `realtimeHandler`, feed every `delta`
   // event into a buffer that invokes the handler on batches matching
@@ -150,20 +150,23 @@ export async function runAnswerImage ({ provider, model, configuration, prompt,
 
 /**
  * @param {object} args
- * @param {string} args.provider
- * @param {string} args.model
+ * @param {string} args.modelKey      Mohdel catalog key `<provider>/<bare>`. The
+ *                                    envelope carries the mohdel key, not the
+ *                                    upstream provider model name — adapters
+ *                                    look up `spec.model` via `catalogKey()`
+ *                                    and use that for the actual API call.
  * @param {any} args.configuration
  * @param {string | any[] | {system?: any, messages: any[]}} args.prompt
  * @param {any} args.options
  * @returns {import('#core/envelope.js').CallEnvelope}
  */
-function toEnvelope ({ provider, model, configuration, prompt, options }) {
+function toEnvelope ({ modelKey, configuration, prompt, options }) {
   /** @type {import('#core/envelope.js').CallEnvelope} */
   const envelope = {
     callId: options.callId || newCallId(),
     authId: options.authId || 'local',
     auth: configToAuth(configuration),
-    model: /** @type {import('#core/model-id.js').ModelId} */ (`${provider}/${model}`),
+    model: /** @type {import('#core/model-id.js').ModelId} */ (modelKey),
     prompt: toEnvelopePrompt(prompt)
   }
 

package/js/session/adapters/_chat_completions.js

@@ -107,6 +107,7 @@ export async function * runChatCompletions (envelope, client, config, deps = {})
 
   let content = message.content || ''
   let toolCalls = message.tool_calls
+  const reasoning = message.reasoning_content || null
 
   if (config.parseDsml && content && (!toolCalls || !toolCalls.length)) {
     const dsml = parseDsmlToolCalls(content)
@@ -121,7 +122,7 @@ export async function * runChatCompletions (envelope, client, config, deps = {})
   }
 
   yield finalize({
-    envelope, content, toolCalls, usage, finishReason, start, first
+    envelope, content, toolCalls, usage, finishReason, start, first, reasoning
   })
 }
 
@@ -140,6 +141,7 @@ async function * runStreaming (envelope, client, args, config, start, deps) {
   // F53: accumulate via array + join to avoid per-delta V8 cons-string
   // churn on long streams.
   const contentParts = []
+  const reasoningParts = []
   let first = null
   let finishReason = null
   let usage = {}
@@ -163,6 +165,15 @@ async function * runStreaming (envelope, client, args, config, start, deps) {
   try {
     for await (const chunk of stream) {
       const choice = chunk.choices?.[0]
+      // DeepSeek V4 / deepseek-reasoner / Cerebras reasoning models emit
+      // `delta.reasoning_content` chunks before visible content. Capture
+      // them so multi-turn callers can roundtrip reasoning back to the
+      // API (DeepSeek V4 hard-rejects assistant messages without it).
+      // Token count comes from `usage.completion_tokens_details.reasoning_tokens`.
+      if (choice?.delta?.reasoning_content) {
+        if (first === null) first = String(process.hrtime.bigint())
+        reasoningParts.push(choice.delta.reasoning_content)
+      }
       if (choice?.delta?.content) {
         if (first === null) first = String(process.hrtime.bigint())
         contentParts.push(choice.delta.content)
@@ -227,7 +238,8 @@ async function * runStreaming (envelope, client, args, config, start, deps) {
     usage,
     finishReason,
     start,
-    first
+    first,
+    reasoning: reasoningParts.length ? reasoningParts.join('') : null
   })
 }
 
@@ -239,11 +251,12 @@ async function * runStreaming (envelope, client, args, config, start, deps) {
  *   usage: any,
  *   finishReason: string | null,
  *   start: string,
- *   first: string | null
+ *   first: string | null,
+ *   reasoning?: string | null
  * }} p
  * @returns {import('#core/events.js').DoneEvent}
  */
-function finalize ({ envelope, content, toolCalls, usage, finishReason, start, first }) {
+function finalize ({ envelope, content, toolCalls, usage, finishReason, start, first, reasoning = null }) {
   const end = String(process.hrtime.bigint())
   const inputTokens = usage.prompt_tokens || 0
   const totalOutputTokens = usage.completion_tokens || 0
@@ -274,6 +287,7 @@ function finalize ({ envelope, content, toolCalls, usage, finishReason, start, f
   if (toolCalls && toolCalls.length > 0) {
     done.result.toolCalls = fromCerebrasToolCalls(toolCalls)
   }
+  if (reasoning) done.result.reasoning = reasoning
   return done
 }
 
@@ -349,11 +363,12 @@ function toChatMessages (prompt) {
         content: flattenText(m.content)
       }
     }
+    const reasoning = m.role === 'assistant' ? extractReasoning(m.content) : null
     if (m.role === 'assistant' && m.toolCalls?.length) {
       // Chat Completions assistant turn: optional `content` + the
       // `tool_calls` array. `arguments` must be a JSON string on
       // the wire.
-      return {
+      const msg = {
         role: 'assistant',
         content: flattenText(m.content) || '',
         tool_calls: m.toolCalls.map(tc => ({
@@ -365,8 +380,12 @@ function toChatMessages (prompt) {
           }
         }))
       }
+      if (reasoning) msg.reasoning_content = reasoning
+      return msg
     }
-    return { role: m.role, content: flattenText(m.content) }
+    const msg = { role: m.role, content: flattenText(m.content) }
+    if (reasoning) msg.reasoning_content = reasoning
+    return msg
   })
 }
 
@@ -383,6 +402,13 @@ function flattenText (content) {
   return content.filter(p => p.type === 'text' && p.text).map(p => p.text).join('\n')
 }
 
+/** @param {string | import('#core/envelope.js').MessagePart[]} content */
+function extractReasoning (content) {
+  if (typeof content === 'string' || !Array.isArray(content)) return null
+  const parts = content.filter(p => p.type === 'reasoning' && p.text).map(p => p.text)
+  return parts.length ? parts.join('\n') : null
+}
+
 /**
  * @param {any} args
  * @param {import('#core/envelope.js').MediaRef[]} images

package/js/session/adapters/fireworks.js

@@ -2,9 +2,9 @@
  * Fireworks adapter — OpenAI-compatible chat completions with
  * streaming, over api.fireworks.ai/inference/v1.
  *
- * Fireworks model IDs carry an `accounts/fireworks/models/` prefix;
- * envelopes can supply either form. The `mutateArgs` hook normalizes
- * `args.model` before the request leaves the adapter.
+ * Catalog `spec.model` carries the full upstream id (with the
+ * `accounts/fireworks/models/` prefix). The adapter forwards it
+ * verbatim — no normalization needed.
  *
  * Implementation uses the OpenAI SDK with a custom baseURL — the
  * wire shape is identical and the SDK's streaming iterator matches
@@ -18,7 +18,6 @@ import OpenAI from 'openai'
 import { runChatCompletions } from './_chat_completions.js'
 
 const BASE_URL = 'https://api.fireworks.ai/inference/v1'
-const MODEL_PREFIX = 'accounts/fireworks/models/'
 
 /**
  * @param {import('#core/envelope.js').CallEnvelope} envelope
@@ -32,12 +31,7 @@ export async function * fireworks (envelope, deps = {}) {
   })
   yield * runChatCompletions(envelope, client, {
     provider: 'fireworks',
-    stream: true,
-    mutateArgs: (env, args) => {
-      if (!args.model.includes('/')) {
-        args.model = `${MODEL_PREFIX}${args.model}`
-      }
-    }
+    stream: true
   }, {
     signal: deps.signal,
     log: deps.log,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mohdel",
-  "version": "0.97.1",
+  "version": "0.98.1",
   "license": "MIT",
   "author": {
     "name": "Christophe Le Bars",
@@ -87,7 +87,7 @@
     "@opentelemetry/exporter-trace-otlp-grpc": "^0.215.0",
     "@opentelemetry/sdk-node": "^0.215.0",
     "chalk": "^5.4.0",
-    "mohdel-thin-gate-linux-x64-gnu": "0.97.1"
+    "mohdel-thin-gate-linux-x64-gnu": "0.98.1"
   },
   "dependencies": {
     "@anthropic-ai/sdk": "^0.91.1",
@@ -96,12 +96,13 @@
     "@opentelemetry/api": "^1.9.1",
     "env-paths": "^4.0.0",
     "groq-sdk": "^1.1.2",
-    "openai": "^6.34.0"
+    "openai": "^6.35.0"
   },
   "lint-staged": {
     "*.{js,cjs}": "standard"
   },
   "devDependencies": {
+    "gpt-tokenizer": "^3.4.0",
     "lint-staged": "^16.4.0",
     "release-it": "^20.0.1",
     "standard": "^17.1.2",

package/src/lib/catalog/fireworks.js

@@ -1,7 +1,7 @@
 // Fireworks' /inference/v1/models is openai-compatible for listing.
-// Upstream IDs carry a `accounts/fireworks/models/` prefix that the
-// runtime adapter adds back automatically — strip it here so curated
-// entries stay short.
+// Upstream IDs carry an `accounts/fireworks/models/` prefix. The mohdel
+// catalog key uses the bare suffix (`fireworks/<bare>`), but `spec.model`
+// preserves the full upstream id so the adapter sends it verbatim.
 const DEFAULT_BASE_URL = 'https://api.fireworks.ai/inference/v1'
 const FW_PREFIX = 'accounts/fireworks/models/'
 
@@ -29,7 +29,7 @@ export default (sdkConfig) => {
     getModelInfo: async (id) => {
       const m = (await load()).find(x => shortId(x.id) === id)
       if (!m) return null
-      const info = { model: shortId(m.id) }
+      const info = { model: m.id }
       if (typeof m.context_length === 'number') info.contextTokenLimit = m.context_length
       if (typeof m.created === 'number') info.created = m.created
       return info

package/src/lib/index.js

@@ -723,4 +723,5 @@ const createModelProxy = (resolvedModelId, modelSpec, handlers, aliasOutputEffor
 
 export { silent } from './logger.js'
 export { loadCuratedCache, getCuratedCacheSnapshot } from './curated-cache.js'
+export { effectiveContextLimit } from './utils.js'
 export default mohdel

package/src/lib/providers.js

@@ -3,26 +3,34 @@ const providers = {
     sdk: 'anthropic',
     apiKeyEnv: 'ANTHROPIC_API_SK',
     createConfiguration: apiKey => ({ apiKey }),
-    creators: ['anthropic']
+    creators: ['anthropic'],
+    contextSemantics: 'shared',
+    outputCapStrategy: 'error'
   },
   cerebras: {
     sdk: 'cerebras',
     apiKeyEnv: 'CEREBRAS_API_SK',
     createConfiguration: apiKey => ({ apiKey }),
-    creators: ['openai', 'zai']
+    creators: ['openai', 'zai'],
+    contextSemantics: 'shared',
+    outputCapStrategy: 'accept'
   },
   deepseek: {
     sdk: 'openai',
     api: 'chatCompletions',
     apiKeyEnv: 'DEEPSEEK_API_SK',
     createConfiguration: apiKey => ({ baseURL: 'https://api.deepseek.com', apiKey }),
-    creators: ['deepseek']
+    creators: ['deepseek'],
+    contextSemantics: 'shared',
+    outputCapStrategy: 'accept'
   },
   gemini: {
     sdk: 'gemini',
     apiKeyEnv: 'GEMINI_API_SK',
     createConfiguration: apiKey => ({ apiKey }),
-    creators: ['google']
+    creators: ['google'],
+    contextSemantics: 'separate',
+    outputCapStrategy: 'accept'
   },
   groq: {
     sdk: 'groq',
@@ -41,7 +49,9 @@ const providers = {
     sdk: 'fireworks',
     apiKeyEnv: 'FIREWORKS_API_SK',
     createConfiguration: apiKey => ({ apiKey, baseURL: 'https://api.fireworks.ai/inference/v1' }),
-    creators: ['meta', 'alibaba']
+    creators: ['meta', 'alibaba'],
+    contextSemantics: 'shared',
+    outputCapStrategy: 'accept'
   },
   novita: {
     sdk: 'openai',
@@ -49,13 +59,17 @@ const providers = {
     imageHandler: 'novita',
     apiKeyEnv: 'NOVITA_API_SK',
     createConfiguration: apiKey => ({ apiKey, baseURL: 'https://api.novita.ai/openai' }),
-    creators: ['deepseek', 'openai', 'bfl']
+    creators: ['deepseek', 'openai', 'bfl'],
+    contextSemantics: 'shared',
+    outputCapStrategy: 'error'
   },
   openai: {
     sdk: 'openai',
     apiKeyEnv: 'OPENAI_API_SK',
     createConfiguration: apiKey => ({ apiKey }),
-    creators: ['openai']
+    creators: ['openai'],
+    contextSemantics: 'shared',
+    outputCapStrategy: 'accept'
   },
   openrouter: {
     sdk: 'openrouter',
@@ -78,7 +92,9 @@ const providers = {
     sdk: 'openai',
     apiKeyEnv: 'XAI_API_SK',
     createConfiguration: apiKey => ({ baseURL: 'https://api.x.ai/v1', apiKey }),
-    creators: ['xai']
+    creators: ['xai'],
+    contextSemantics: 'shared',
+    outputCapStrategy: 'accept'
   }
 }
 

package/src/lib/utils.js

@@ -3,6 +3,17 @@ export const sanitizeOutput = str => {
   return str.replace(/\0/g, '\uFFFD').trim()
 }
 
+// Practical input ceiling for a model spec: catalog `contextTokenLimit`
+// minus any empirically-derived `inputCeilingMargin` reserve the API
+// silently steals (reasoning floor, structural overhead, pricing-tier
+// cliffs). Reduces to `contextTokenLimit` when the margin field is
+// unset. Consumers computing safe input budgets should call this in
+// place of `spec.contextTokenLimit`.
+export const effectiveContextLimit = spec => {
+  if (!spec || spec.contextTokenLimit == null) return 0
+  return Math.max(0, spec.contextTokenLimit - (spec.inputCeilingMargin ?? 0))
+}
+
 export const translateModelInfo = (model, infoTranslate = {}) => {
   if (!model || typeof model !== 'object') return model