@yeaft/webchat-agent 0.1.1104 → 1.0.1

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yeaft/webchat-agent",
-  "version": "0.1.1104",
+  "version": "1.0.1",
   "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/anthropic.js

@@ -34,10 +34,10 @@ function thinkingV1Enabled() {
   return process.env.YEAFT_THINKING_V1 === '1';
 }
 
-function applyAnthropicThinking(body, model, effort) {
-  const cap = getThinkingCapability(model);
+function applyAnthropicThinking(body, model, effort, effortContext = {}) {
+  const cap = getThinkingCapability(model, effortContext);
   if (!cap.supportsThinking) return;
-  if (!getModelEffortOptions(model).includes(effort)) return;
+  if (!getModelEffortOptions(model, effortContext).includes(effort)) return;
 
   if (cap.thinkingProtocol === 'anthropic-adaptive') {
     body.thinking = { type: 'adaptive' };
@@ -207,10 +207,10 @@ export class AnthropicAdapter extends LLMAdapter {
   }
 
   /**
-   * @param {{ model: string, system: string, messages: import('./adapter.js').UnifiedMessage[], tools?: import('./adapter.js').UnifiedToolDef[], maxTokens?: number, effort?: 'low'|'medium'|'high'|'xhigh'|'max', effortSource?: 'user'|'auto', signal?: AbortSignal }} params
+   * @param {{ model: string, system: string, messages: import('./adapter.js').UnifiedMessage[], tools?: import('./adapter.js').UnifiedToolDef[], maxTokens?: number, effort?: 'low'|'medium'|'high'|'xhigh'|'max', effortSource?: 'user'|'auto', effortContext?: object, signal?: AbortSignal, onRawExchange?: ({rawRequest, rawResponse}) => void }} params
    * @returns {AsyncGenerator<import('./adapter.js').StreamEvent>}
    */
-  async *stream({ model, system, messages, tools, maxTokens = 16384, effort, effortSource, signal, onRawExchange }) {
+  async *stream({ model, system, messages, tools, maxTokens = 16384, effort, effortSource, effortContext, signal, onRawExchange }) {
     if (signal?.aborted) throw new LLMAbortError();
 
     const body = {
@@ -226,7 +226,7 @@ export class AnthropicAdapter extends LLMAdapter {
     // models use budget_tokens. Unsupported combinations silently drop effort.
     const normEffort = normalizeEffort(effort);
     if ((thinkingV1Enabled() || effortSource === 'user') && normEffort) {
-      applyAnthropicThinking(body, model, normEffort);
+      applyAnthropicThinking(body, model, normEffort, effortContext);
     }
 
     const translatedTools = this.#translateTools(tools);
@@ -473,7 +473,7 @@ export class AnthropicAdapter extends LLMAdapter {
    * models silently drop the param. max_tokens auto-widens to budget+1024
    * when needed.
    */
-  async call({ model, system, messages, maxTokens = 4096, effort, effortSource, signal }) {
+  async call({ model, system, messages, maxTokens = 4096, effort, effortSource, effortContext, signal }) {
     if (signal?.aborted) throw new LLMAbortError();
 
     const body = {
@@ -486,7 +486,7 @@ export class AnthropicAdapter extends LLMAdapter {
     // task-327c: mirror stream()'s thinking injection for side queries.
     const normEffort = normalizeEffort(effort);
     if ((thinkingV1Enabled() || effortSource === 'user') && normEffort) {
-      applyAnthropicThinking(body, model, normEffort);
+      applyAnthropicThinking(body, model, normEffort, effortContext);
     }
 
     let response;

package/yeaft/llm/router.js

@@ -552,9 +552,16 @@ export class AdapterRouter extends LLMAdapter {
    */
   async *stream(params) {
     const resolved = await this.#resolveAdapter(params.model);
+    const effortContext = {
+      protocol: resolved.protocol,
+      supportsEffort: resolved.entry?.supportsEffort,
+      effortOptions: resolved.entry?.effortOptions,
+      thinkingProtocol: resolved.entry?.thinkingProtocol,
+      maxBudgetTokens: resolved.entry?.maxBudgetTokens,
+    };
     const filtered = filterEffortForModel({ ...params, model: resolved.modelId }, resolved);
     const sanitized = sanitizeMessagesForWire(filtered);
-    yield* resolved.adapter.stream({ ...sanitized, model: resolved.modelId });
+    yield* resolved.adapter.stream({ ...sanitized, model: resolved.modelId, effortContext });
   }
 
   /**
@@ -565,9 +572,16 @@ export class AdapterRouter extends LLMAdapter {
    */
   async call(params) {
     const resolved = await this.#resolveAdapter(params.model);
+    const effortContext = {
+      protocol: resolved.protocol,
+      supportsEffort: resolved.entry?.supportsEffort,
+      effortOptions: resolved.entry?.effortOptions,
+      thinkingProtocol: resolved.entry?.thinkingProtocol,
+      maxBudgetTokens: resolved.entry?.maxBudgetTokens,
+    };
     const filtered = filterEffortForModel({ ...params, model: resolved.modelId }, resolved);
     const sanitized = sanitizeMessagesForWire(filtered);
-    return resolved.adapter.call({ ...sanitized, model: resolved.modelId });
+    return resolved.adapter.call({ ...sanitized, model: resolved.modelId, effortContext });
   }
 
   /**

package/yeaft/models.js

@@ -537,8 +537,13 @@ export function thinkingBudgetForEffort(model, effort) {
  */
 export function getThinkingCapability(model, context = {}) {
   const info = MODEL_REGISTRY.get(model);
+  const modelId = parseModelRef(model).modelId;
   const overrideOptions = normalizeEffortOptions(context.effortOptions);
-  const overrideProtocol = context.thinkingProtocol || (context.protocol === 'anthropic' ? 'anthropic' : context.protocol === 'openai-responses' ? 'openai-reasoning' : null);
+  const overrideProtocol = context.thinkingProtocol || (
+    context.protocol === 'anthropic'
+      ? (/^deepseek/i.test(modelId) ? 'anthropic-adaptive' : 'anthropic')
+      : context.protocol === 'openai-responses' ? 'openai-reasoning' : null
+  );
   if (context.supportsEffort === true || overrideOptions) {
     return {
       supportsThinking: true,
@@ -560,7 +565,15 @@ export function getThinkingCapability(model, context = {}) {
     };
   }
   if (context.protocol === 'anthropic' && inferred?.thinkingProtocol === 'openai-reasoning') {
-    inferred = null;
+    if (/^deepseek/i.test(modelId)) {
+      inferred = {
+        ...inferred,
+        thinkingProtocol: 'anthropic-adaptive',
+        effortOptions: DEEPSEEK_REASONING_EFFORT_OPTIONS,
+      };
+    } else {
+      inferred = null;
+    }
   }
   if ((!info || !info.supportsThinking) && !inferred) {
     return {