@yeaft/webchat-agent 0.1.508 → 0.1.509

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yeaft/webchat-agent",
-  "version": "0.1.508",
+  "version": "0.1.509",
   "description": "Remote agent for Yeaft WebChat — connects worker machines to the central server",
   "main": "index.js",
   "type": "module",

package/unify/engine.js

@@ -26,6 +26,7 @@ import { buildMemoryInjection } from './memory/layout.js';
 import { runStopHooks } from './stop-hooks.js';
 import { getThreadStore, MAIN_THREAD_ID } from './threads/store.js';
 import { pickEffort, parseEffortPrefix } from './effort.js';
+import { normalizeEffort } from './models.js';
 
 /**
  * task-324 — Turn cap removed.
@@ -446,9 +447,12 @@ export class Engine {
 
     // task-327b: `/max` / `/high` / `/medium` / `/low` prefix override.
     // Explicit caller-supplied userEffort wins over the prefix.
+    // task-327c nit: defensively normalize caller-supplied userEffort BEFORE
+    // the merge, so an invalid caller value (e.g. 'ULTRA') does not shadow a
+    // valid prompt prefix.
     const parsed = parseEffortPrefix(prompt);
     const effectivePrompt = parsed.cleanedPrompt;
-    const effectiveUserEffort = userEffort || parsed.effort || null;
+    const effectiveUserEffort = normalizeEffort(userEffort) || parsed.effort || null;
 
     // ─── task-325a: engine-owned AbortController ─────────────
     // We create our own controller for this query run so `engine.abort()`

package/unify/llm/anthropic.js

@@ -297,8 +297,13 @@ export class AnthropicAdapter extends LLMAdapter {
 
   /**
    * Non-streaming call for side queries.
+   *
+   * task-327c: accepts `effort` for internal scenario-tagged calls
+   * (consolidate/dream/recall/light). Guards mirror stream() — unsupported
+   * models silently drop the param. max_tokens auto-widens to budget+1024
+   * when needed.
    */
-  async call({ model, system, messages, maxTokens = 4096, signal }) {
+  async call({ model, system, messages, maxTokens = 4096, effort, signal }) {
     if (signal?.aborted) throw new LLMAbortError();
 
     const body = {
@@ -308,6 +313,20 @@ export class AnthropicAdapter extends LLMAdapter {
       messages: this.#translateMessages(messages),
     };
 
+    // task-327c: mirror stream()'s thinking injection for side queries.
+    const normEffort = normalizeEffort(effort);
+    if (thinkingV1Enabled() && normEffort) {
+      const cap = getThinkingCapability(model);
+      if (cap.supportsThinking && cap.thinkingProtocol === 'anthropic') {
+        const budget = thinkingBudgetForEffort(model, normEffort);
+        if (budget && budget > 0) {
+          const minMax = budget + 1024;
+          if (body.max_tokens < minMax) body.max_tokens = minMax;
+          body.thinking = { type: 'enabled', budget_tokens: budget };
+        }
+      }
+    }
+
     const response = await fetch(`${this.#baseUrl}/v1/messages`, {
       method: 'POST',
       headers: {

package/unify/llm/chat-completions.js

@@ -350,8 +350,12 @@ export class ChatCompletionsAdapter extends LLMAdapter {
 
   /**
    * Non-streaming call for side queries.
+   *
+   * task-327c: accepts `effort` for internal scenario-tagged calls
+   * (consolidate/dream/recall/light). Feature-flag + capability guards
+   * mirror stream() exactly; unsupported models silently drop the param.
    */
-  async call({ model, system, messages, maxTokens = 4096, extraBody, signal }) {
+  async call({ model, system, messages, maxTokens = 4096, effort, extraBody, signal }) {
     if (signal?.aborted) throw new LLMAbortError();
 
     const body = {
@@ -360,6 +364,18 @@ export class ChatCompletionsAdapter extends LLMAdapter {
       ...this.#maxTokensBody(model, maxTokens),
     };
 
+    // task-327c: mirror stream()'s reasoning.effort injection for side queries.
+    const normEffort = normalizeEffort(effort);
+    if (thinkingV1Enabled() && normEffort) {
+      const cap = getThinkingCapability(model);
+      if (cap.supportsThinking && cap.thinkingProtocol === 'openai-reasoning') {
+        const reasoningEffort = mapEffortToOpenAIReasoning(normEffort);
+        if (reasoningEffort) {
+          body.reasoning = { effort: reasoningEffort };
+        }
+      }
+    }
+
     // extraBody allows callers to pass through any additional/override parameters
     if (extraBody) Object.assign(body, extraBody);
 

package/unify/memory/consolidate.js

@@ -15,6 +15,7 @@
  */
 
 import { extractMemories } from './extract.js';
+import { pickEffort } from '../effort.js';
 
 // ─── Constants ──────────────────────────────────────────────────
 
@@ -103,6 +104,11 @@ async function generateSummary(messages, adapter, config) {
       system,
       messages: [{ role: 'user', content: `Summarize this conversation:\n\n${conversation}` }],
       maxTokens: 1024,
+      // task-327c: consolidate is a high-complexity side-query; flag as
+      // 'max' effort so supported models use extended thinking / reasoning.
+      // Router/adapter silently drops the param for models that don't
+      // support thinking, or when UNIFY_THINKING_V1 is off.
+      effort: pickEffort({ scenario: 'consolidate' }),
     });
     return result.text.trim();
   } catch {

package/unify/memory/dream.js

@@ -16,6 +16,7 @@ import { existsSync, readFileSync, writeFileSync, mkdirSync, readdirSync, unlink
 import { join } from 'path';
 import { scanEntries, findStaleEntries, findDuplicateGroups, summarizeScan } from './scan.js';
 import { MAX_ENTRIES } from './store.js';
+import { pickEffort } from '../effort.js';
 import {
   ensureLayout,
   renderIndex,
@@ -399,6 +400,9 @@ async function llmCall(adapter, config, system, prompt) {
       system,
       messages: [{ role: 'user', content: prompt }],
       maxTokens: 4096,
+      // task-327c: dream is self-reflective memory maintenance — flag 'max'
+      // so supported models use the full thinking budget.
+      effort: pickEffort({ scenario: 'dream' }),
     });
 
     const text = result.text.trim();
@@ -761,6 +765,9 @@ Write the narrative as Markdown with:
       system,
       messages: [{ role: 'user', content: prompt }],
       maxTokens: 2048,
+      // task-327c: dream narrative synthesis — same 'max' tier as the
+      // dream phase above; both pass through dream's self-reflection loop.
+      effort: pickEffort({ scenario: 'dream' }),
     });
     const text = (result?.text || '').trim();
     if (!text) return null;

package/unify/memory/extract.js

@@ -9,6 +9,7 @@
  */
 
 import { MEMORY_KINDS } from './store.js';
+import { pickEffort } from '../effort.js';
 
 /**
  * Build the extraction prompt.
@@ -69,6 +70,9 @@ export async function extractMemories({ messages, adapter, config }) {
       system,
       messages: [{ role: 'user', content: extractionPrompt }],
       maxTokens: 2048,
+      // task-327c: extract runs inside the consolidate pipeline — the
+      // JSON-structured output benefits from the same 'max' thinking tier.
+      effort: pickEffort({ scenario: 'consolidate' }),
     });
 
     const text = result.text.trim();

package/unify/memory/recall.js

@@ -14,6 +14,7 @@
  */
 
 import { createHash } from 'crypto';
+import { pickEffort } from '../effort.js';
 
 // ─── Constants ──────────────────────────────────────────────────
 
@@ -148,6 +149,9 @@ Select the ${MAX_RECALL_RESULTS} most relevant entries. Return a JSON array of e
       system,
       messages,
       maxTokens: 512,
+      // task-327c: recall step-3 is a cheap classifier pass (pick N out of
+      // 15 candidates). Flag 'low' so supported models skip deep reasoning.
+      effort: pickEffort({ scenario: 'recall' }),
     });
 
     // Parse the JSON array from the response