@yeaft/webchat-agent 1.0.156 → 1.0.157

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yeaft/webchat-agent",
-  "version": "1.0.156",
+  "version": "1.0.157",
   "description": "Remote worker agent for Yeaft Web Code Agent — connects the native Yeaft engine, CLI providers, and workbench tools",
   "main": "index.js",
   "type": "module",

package/yeaft/llm/openai-responses.js

@@ -41,6 +41,7 @@ import {
 } from './adapter.js';
 import {
   normalizeEffort,
+  getModelEffortOptions,
   getThinkingCapability,
   mapEffortToOpenAIReasoning,
 } from '../models.js';
@@ -57,10 +58,9 @@ function thinkingV1Enabled() {
 }
 
 /**
- * Translate a normalised effort ('minimal'|'low'|'medium'|'high'|'xhigh'|'max') into the value
- * accepted by the OpenAI Responses `reasoning.effort` field. Responses today
- * accepts 'minimal'|'low'|'medium'|'high'|'xhigh'. Unsupported Anthropic-only
- * adaptive efforts must be dropped, not downgraded.
+ * Translate a normalised effort into the OpenAI Responses `reasoning.effort`
+ * field. Model capability filtering happens at the call site because `max` is
+ * supported by GPT-5.6 but not by older OpenAI reasoning models.
  */
 function effortForResponses(effort) {
   return mapEffortToOpenAIReasoning(effort);
@@ -248,7 +248,7 @@ export class OpenAIResponsesAdapter extends LLMAdapter {
   // ─── Streaming ──────────────────────────────────────────
 
   /**
-   * @param {{ model: string, system: string, messages: import('./adapter.js').UnifiedMessage[], tools?: import('./adapter.js').UnifiedToolDef[], maxTokens?: number, effort?: 'minimal'|'low'|'medium'|'high'|'xhigh'|'max', effortSource?: 'user'|'auto', extraBody?: object, signal?: AbortSignal, onRawExchange?: ({rawRequest, rawResponse}) => void }} params
+   * @param {{ model: string, system: string, messages: import('./adapter.js').UnifiedMessage[], tools?: import('./adapter.js').UnifiedToolDef[], maxTokens?: number, effort?: 'minimal'|'low'|'medium'|'high'|'xhigh'|'max', effortSource?: 'user'|'auto', effortContext?: object, extraBody?: object, signal?: AbortSignal, onRawExchange?: ({rawRequest, rawResponse}) => void, onRequestStart?: () => void }} params
    *
    * NOTE on `extraBody`: any keys you spread here are merged verbatim into
    * the wire body and — because the verbatim debug feature is intentionally
@@ -257,7 +257,7 @@ export class OpenAIResponsesAdapter extends LLMAdapter {
    * `api-key` headers are auto-redacted (see `redactRawRequest` in
    * `adapter.js`); request-body fields are caller-controlled.
    */
-  async *stream({ model, system, messages, tools, maxTokens = 16384, effort, effortSource, extraBody, signal, onRawExchange, onRequestStart }) {
+  async *stream({ model, system, messages, tools, maxTokens = 16384, effort, effortSource, effortContext = {}, extraBody, signal, onRawExchange, onRequestStart }) {
     if (signal?.aborted) throw new LLMAbortError();
 
     const body = {
@@ -276,8 +276,9 @@ export class OpenAIResponsesAdapter extends LLMAdapter {
     // Unknown / unsupported models silently drop the field.
     const normEffort = normalizeEffort(effort);
     if ((thinkingV1Enabled() || effortSource === 'user') && normEffort) {
-      const cap = getThinkingCapability(model);
-      if (cap.supportsThinking && cap.thinkingProtocol === 'openai-reasoning') {
+      const cap = getThinkingCapability(model, effortContext);
+      const supportedEfforts = getModelEffortOptions(model, effortContext);
+      if (cap.supportsThinking && cap.thinkingProtocol === 'openai-reasoning' && supportedEfforts.includes(normEffort)) {
         const wireEffort = effortForResponses(normEffort);
         if (wireEffort) body.reasoning = { effort: wireEffort };
       }
@@ -517,7 +518,7 @@ export class OpenAIResponsesAdapter extends LLMAdapter {
    * expose them, mirror the stream() instrumentation. Parity with
    * anthropic.js's `call()`.
    */
-  async call({ model, system, messages, maxTokens = 4096, effort, effortSource, extraBody, signal, onRequestStart }) {
+  async call({ model, system, messages, maxTokens = 4096, effort, effortSource, effortContext = {}, extraBody, signal, onRequestStart }) {
     if (signal?.aborted) throw new LLMAbortError();
 
     const body = {
@@ -530,8 +531,9 @@ export class OpenAIResponsesAdapter extends LLMAdapter {
     // Mirror stream()'s thinking injection for non-streaming side queries.
     const normEffort = normalizeEffort(effort);
     if ((thinkingV1Enabled() || effortSource === 'user') && normEffort) {
-      const cap = getThinkingCapability(model);
-      if (cap.supportsThinking && cap.thinkingProtocol === 'openai-reasoning') {
+      const cap = getThinkingCapability(model, effortContext);
+      const supportedEfforts = getModelEffortOptions(model, effortContext);
+      if (cap.supportsThinking && cap.thinkingProtocol === 'openai-reasoning' && supportedEfforts.includes(normEffort)) {
         const wireEffort = effortForResponses(normEffort);
         if (wireEffort) body.reasoning = { effort: wireEffort };
       }

package/yeaft/models.js

@@ -34,7 +34,7 @@ import { lookupModelLimitSync } from './llm/models-dev.js';
  * @property {'anthropic' | 'anthropic-adaptive' | 'openai-reasoning' | 'none'} [thinkingProtocol] — task-327a:
  *   'anthropic' → thinking:{type:'enabled', budget_tokens:N}
  *   'anthropic-adaptive' → thinking:{type:'adaptive'} + output_config:{effort}
- *   'openai-reasoning' → reasoning:{effort:'minimal'|'low'|'medium'|'high'|'xhigh'}
+ *   'openai-reasoning' → reasoning:{effort:'minimal'|'low'|'medium'|'high'|'xhigh'|'max'}
  *   'none' (default) → parameter silently dropped by router
  * @property {'minimal' | 'low' | 'medium' | 'high' | 'xhigh' | 'max' | null} [defaultEffort] — adapter-level default
  *   when caller doesn't specify effort (null = no default / decision-tree decides).
@@ -405,12 +405,12 @@ export const ANTHROPIC_THINKING_BUDGETS = {
 };
 
 /**
- * Map a Yeaft effort level to the OpenAI reasoning.effort enum. Current
- * OpenAI APIs expose xhigh, but not Anthropic's 'max'. Unsupported values are
- * dropped by returning null; the adapter MUST NOT error.
+ * Map a Yeaft effort level to the OpenAI reasoning.effort enum. Supported
+ * values are model-dependent; capability filtering happens before this wire
+ * translation. Unsupported values are dropped; the adapter MUST NOT error.
  *
  * @param {Effort} effort
- * @returns {'minimal' | 'low' | 'medium' | 'high' | 'xhigh' | null}
+ * @returns {'minimal' | 'low' | 'medium' | 'high' | 'xhigh' | 'max' | null}
  */
 export function mapEffortToOpenAIReasoning(effort) {
   if (!effort) return null;
@@ -420,11 +420,13 @@ export function mapEffortToOpenAIReasoning(effort) {
     case 'medium': return 'medium';
     case 'high': return 'high';
     case 'xhigh': return 'xhigh';
+    case 'max': return 'max';
     default: return null;
   }
 }
 
 export const OPENAI_REASONING_EFFORT_OPTIONS = ['minimal', 'low', 'medium', 'high', 'xhigh'];
+export const OPENAI_MAX_REASONING_EFFORT_OPTIONS = ['low', 'medium', 'high', 'xhigh', 'max'];
 export const ANTHROPIC_MANUAL_EFFORT_OPTIONS = ['low', 'medium', 'high'];
 export const ANTHROPIC_ADAPTIVE_EFFORT_OPTIONS = ['low', 'medium', 'high', 'xhigh', 'max'];
 export const ANTHROPIC_ADAPTIVE_MAX_EFFORT_OPTIONS = ['low', 'medium', 'high', 'max'];
@@ -454,6 +456,19 @@ function inferThinkingCapability(model) {
   const id = parseModelRef(model).modelId.toLowerCase();
   if (!id) return null;
 
+  // GPT-5.6 adds the model-specific `max` tier. Keep this ahead of the generic
+  // GPT inference so older OpenAI models do not advertise a value they reject.
+  // Provider suffixes such as gpt-5.6-sol are variants of the same model family.
+  if (/^gpt-5\.6($|[-.])/.test(id)) {
+    return {
+      supportsThinking: true,
+      thinkingProtocol: 'openai-reasoning',
+      defaultEffort: null,
+      maxBudgetTokens: null,
+      effortOptions: OPENAI_MAX_REASONING_EFFORT_OPTIONS,
+    };
+  }
+
   if (/^(gpt-5|o1|o3|o4|chatgpt-|codex-)/.test(id)) {
     return { supportsThinking: true, thinkingProtocol: 'openai-reasoning', defaultEffort: null, maxBudgetTokens: null };
   }