@nexus-cortex/core 4.88.0 → 4.88.2

package/dist/orchestrator/CortexOrchestrator.js

@@ -37,7 +37,7 @@ import { slashCommandRegistry } from '../commands/SlashCommandRegistry.js';
 import { prefixMcpToolName, parseMcpToolName } from '../mcp/mcpToolNamespacing.js';
 import { DecisionStore } from '../training/DecisionStore.js';
 import { formatPriorReminder, formatFamilyReminder, formatApproachReminder } from '../training/DecisionPriorInjector.js';
-import { resolveConsultRung, resolveMentorConfig, bounceMessage, rateLimitedMessage, } from '../training/mentorConsult.js';
+import { resolveConsultRung, resolveMentorConfig, bounceMessage, TEMPLATE_REFRAME, TEMPLATE_INTERVIEW, rateLimitedMessage, } from '../training/mentorConsult.js';
 import { resolveThrashState, resolveThrashConfig } from '../training/thrashDetector.js';
 import { classifyErrorFamily } from '../training/errorFamily.js';
 import { classifyToolOutcome } from '../training/toolOutcome.js';
@@ -2385,7 +2385,7 @@ export class CortexOrchestrator {
                         temperature: options.parameters?.temperature,
                         maxTokens: options.parameters?.maxTokens,
                         topP: options.parameters?.topP,
-                        reasoningEffort: options.parameters?.reasoningEffort, // GPT-5.1 reasoning level
+                        reasoningEffort: this.consumeEffortPulse() ?? options.parameters?.reasoningEffort, // effort pulse > request param > card
                         stream: options.streaming,
                         staticSystemPrompt: this.currentStaticSystemPrompt, // R28
                         conversationId: this.currentConversationId, // R28b
@@ -4009,7 +4009,7 @@ export class CortexOrchestrator {
                     temperature: options.parameters?.temperature,
                     maxTokens: options.parameters?.maxTokens,
                     topP: options.parameters?.topP,
-                    reasoningEffort: options.parameters?.reasoningEffort, // GPT-5.1 reasoning level
+                    reasoningEffort: this.consumeEffortPulse() ?? options.parameters?.reasoningEffort, // effort pulse > request param > card
                     stream: true, // STREAMING continuation
                     staticSystemPrompt: this.currentStaticSystemPrompt, // R28
                     conversationId: this.currentConversationId, // R28b
@@ -6887,6 +6887,17 @@ export class CortexOrchestrator {
             return undefined; // fail-open: never let the thrash read break a turn
         }
     }
+    /** EFFORT PULSE (ablation 'think harder' arm, def-efdbb67fd8): continuations
+     *  remaining at elevated reasoning effort after a cumulative-thrash trigger. */
+    effortPulseRemaining = 0;
+    /** While the pulse is armed, continuation requests run at the elevated level
+     *  (CORTEX_EFFORT_PULSE_LEVEL, default 'high'); decrements per consumed request. */
+    consumeEffortPulse() {
+        if (this.effortPulseRemaining <= 0)
+            return undefined;
+        this.effortPulseRemaining -= 1;
+        return (process.env.CORTEX_EFFORT_PULSE_LEVEL || 'high');
+    }
     /** Honored AskForAdvice consults per session (rate-limit + rung state; persists across turns). */
     mentorConsultCounts = new Map();
     /**
@@ -6919,21 +6930,30 @@ export class CortexOrchestrator {
             return { success: true, llmContent: rateLimitedMessage(resolveMentorConfig().maxConsults), metadata: { source: 'mentor-consult', rung } };
         }
         let hint;
-        try {
-            hint = await this.helperMiddleware.generateMentorHint({
-                rung,
-                task: this.lastRealUserText(),
-                failed,
-                question: input?.question,
-                helperModelId: this.config.reactiveMentorship?.helperModelId,
-            });
+        if (process.env.CORTEX_MENTOR_TEMPLATE === 'true') {
+            // Ablation-ladder rung 1 (no-LLM mentor): the FIXED self-interrogation
+            // replaces the helper call at the SAME rung — identical trigger, bounce
+            // ladder, rate limit and timing as the LLM mentor; only the hint CONTENT
+            // differs (the isolation the ladder needs). Zero cost, zero latency.
+            hint = rung === 'interview' ? TEMPLATE_INTERVIEW : TEMPLATE_REFRAME;
         }
-        catch (err) {
-            return {
-                success: true,
-                llmContent: 'Advice is unavailable right now — keep working the problem: re-read the task and try a distinct approach.',
-                metadata: { source: 'mentor-consult', rung, error: String(err).slice(0, 120) },
-            };
+        else {
+            try {
+                hint = await this.helperMiddleware.generateMentorHint({
+                    rung,
+                    task: this.lastRealUserText(),
+                    failed,
+                    question: input?.question,
+                    helperModelId: this.config.reactiveMentorship?.helperModelId,
+                });
+            }
+            catch (err) {
+                return {
+                    success: true,
+                    llmContent: 'Advice is unavailable right now — keep working the problem: re-read the task and try a distinct approach.',
+                    metadata: { source: 'mentor-consult', rung, error: String(err).slice(0, 120) },
+                };
+            }
         }
         this.mentorConsultCounts.set(sessionId, honored + 1);
         if (store) {
@@ -6949,6 +6969,7 @@ export class CortexOrchestrator {
                 detail: {
                     tag: 'mentor_episode',
                     rung,
+                    ...(process.env.CORTEX_MENTOR_TEMPLATE === 'true' ? { template: true } : {}),
                     turn: this.turnNumber,
                     helperModel: this.config.reactiveMentorship?.helperModelId ?? 'default',
                     question: input?.question ?? null,
@@ -7060,6 +7081,30 @@ export class CortexOrchestrator {
                         }
                     }
                 }
+                // EFFORT PULSE trigger (def-efdbb67fd8, 'think harder' ablation arm): on
+                // cumulative thrash, escalate reasoning effort for the next N continuation
+                // requests — deep deliberation bought exactly at the stuck point. The
+                // introspective CONTRAST to the mentor's decorrelative consult: same
+                // trigger, same rstan gate (cumulative threshold), INDEPENDENT of
+                // CORTEX_MENTOR_AUTO so each ablation arm isolates cleanly. Ship dark.
+                if (process.env.CORTEX_EFFORT_PULSE === 'true' && outcome.status !== 'ok') {
+                    const cfgP = resolveThrashConfig();
+                    const [pOutcomes, pPrior] = await Promise.all([
+                        store.recentOutcomes(cfgP.window),
+                        store.failureCount(this.currentSessionId ?? 'unknown'),
+                    ]);
+                    pOutcomes.push(false);
+                    const pCum = pPrior + 1;
+                    const pState = resolveThrashState(pOutcomes, pCum, cfgP, pCum);
+                    if (pState.thrashing) {
+                        const turns = Math.max(1, parseInt(process.env.CORTEX_EFFORT_PULSE_TURNS || '2', 10) || 2);
+                        if (this.effortPulseRemaining < turns)
+                            this.effortPulseRemaining = turns;
+                        if (this.config.debug) {
+                            console.log(`[EffortPulse] armed (trigger: ${pState.trigger}, cumFails: ${pCum}) — next ${this.effortPulseRemaining} continuation(s) at elevated effort`);
+                        }
+                    }
+                }
             }
             catch (err) {
                 if (this.config.debug) {