@nexus-cortex/core 4.108.4 → 4.108.7

package/dist/orchestrator/CortexOrchestrator.js

@@ -38,8 +38,9 @@ import { toolFactory } from '../tools/ToolFactory.js';
 import { resolveToolProfile, resolveToolAnchor, resolveFrameProfile, resolveLiftNudge, resolveLiftPlan, resolveEndTurnResolver, resolveHeadlessDropAskUser, resolveDeferredLoading, isNarrowProfile, isToolAllowedByProfile, applyToolProfile, webToolBlocked, resolveVisionHelperModel, resolveVisionHandoffMax, resolveSliceNudge, resolveSliceBlock, resolveSliceBlockAt, resolveSliceBlockMax } from '../tools/ToolProfile.js';
 import { sliceReadFile, decideSliceBlock } from './sliceBlock.js';
 import { TurnEvidence, evaluateEndTurnGates, buildMissingEndTurnReminder } from './endTurnGates.js';
-import { resolveCompactionResume, pickDropped, buildCompactionResumeReminder, isCompactionResumeMessage, resolveTaskText, resumeMemoryTargetTokens, coversAll, buildRollingSeedMessage, memoryIndexLine, upsertMemoryIndex } from './compactionResume.js';
+import { resolveCompactionResume, pickDropped, buildCompactionResumeReminder, isCompactionResumeMessage, resolveTaskText, resumeMemoryTargetTokens, coversAll, buildRollingSeedMessage, memoryIndexLine, upsertMemoryIndex, scaleEstimateToRequestView, approxCharsOf } from './compactionResume.js';
 import { renderResumeMemoryPrompt, buildRebuildInstructions, resolveCheckpointBand, resolveRearmBand } from './compactionResumeTemplate.js';
+import { resolveCortexStateDir, cortexStatePath } from '../utils/stateDir.js';
 import { collectWorkspaceDelta, detectCheckCommand, runCheck, resolveJudgeGroundingConfig } from '../training/judgeEvidence.js';
 import { resolveMentorRoleConfig, describeMentorWire, describeMentorDelivery, mentorSurfaceTimeoutMs } from '../training/mentorRole.js';
 import { resolveOuterToolDeadlineMs } from './outerToolTimeout.js';
@@ -2245,7 +2246,7 @@ export class CortexOrchestrator {
                             // orchestrator (config.projectPath is unset there → was landing
                             // at the server's cwd).
                             const recRoot = process.env.PROJECT_ROOT || this.config.projectPath || this.config.workingDirectory || process.cwd();
-                            const recPath = pathJoin(recRoot, '.cortex', 'training', 'cortex-samples.jsonl');
+                            const recPath = cortexStatePath(recRoot, 'training', 'cortex-samples.jsonl');
                             appendJsonlRotating(recPath, JSON.stringify(sample));
                         }
                         catch (recErr) {
@@ -5692,7 +5693,7 @@ export class CortexOrchestrator {
     /** HB-COMPACTION-RESUME: best-effort write of the resume memory to .cortex/memory/resume-<session>.md (synchronous fs). */
     writeResumeMemoryFile(text, why) {
         try {
-            const cortexDir = pathJoin(this.config.projectPath || process.cwd(), '.cortex');
+            const cortexDir = resolveCortexStateDir(this.config.projectPath || process.cwd()).dir;
             const dir = pathJoin(cortexDir, 'memory');
             mkdirSync(dir, { recursive: true });
             const name = `resume-${this.currentSessionId}`;
@@ -5742,7 +5743,26 @@ export class CortexOrchestrator {
         // Thinking blocks are NEVER stripped — stripping mutates messageHistory,
         // breaks prompt-cache prefix continuity, and causes model confusion.
         // Context overflow is handled by compaction (whole-message removal).
-        const currentTokens = this.contextBudgetManager.estimateTotalTokens(this.messageHistory);
+        const rawEstimate = this.contextBudgetManager.estimateTotalTokens(this.messageHistory);
+        // 4.108.5 (R129, TB4.0 cad-model): the stored history can weigh 1M+ "tokens" of raw tool output while the REQUEST carries the
+        // aged-pruned view (68K real tokens). Compaction must judge what the request sends, or the checkpoint rung and the proactive drop
+        // fire on a false reading and re-fire every iteration (tokensAfter stays above the threshold). Scale the estimate by the
+        // pruned-view/raw char ratio; below the pruner's 50%-utilization gate the ratio is 1 and nothing changes.
+        let requestRatio = 1;
+        let currentTokens = rawEstimate;
+        try {
+            const rawChars = approxCharsOf(this.messageHistory);
+            const prunedChars = approxCharsOf(this.convertToCanonicalMessages(this.messageHistory));
+            currentTokens = scaleEstimateToRequestView(rawEstimate, rawChars, prunedChars);
+            requestRatio = rawEstimate > 0 ? currentTokens / rawEstimate : 1;
+            if (this.config.debug && requestRatio < 1) {
+                console.log(`[Orchestrator Context] request-view estimate: ${currentTokens} of ${rawEstimate} stored tokens (ratio ${requestRatio.toFixed(3)})`);
+            }
+        }
+        catch {
+            currentTokens = rawEstimate;
+            requestRatio = 1;
+        }
         // Budget overrides: use actual token counts instead of hardcoded estimates.
         // - actualToolTokens: real tool schema size (tools are NOT in messageHistory)
         // - actualSystemTokens: 0 because system messages are embedded as <system-reminder>
@@ -5807,7 +5827,7 @@ export class CortexOrchestrator {
             // Select messages that fit within budget (standard path — no thinking stripping)
             const selectedMessages = this.contextBudgetManager.selectMessages(this.messageHistory, budget.availableForHistory, { strategy, preserveToolCalls: true }, model);
             const removedCount = this.messageHistory.length - selectedMessages.length;
-            const newTokenCount = this.contextBudgetManager.estimateTotalTokens(selectedMessages);
+            const newTokenCount = Math.max(1, Math.round(this.contextBudgetManager.estimateTotalTokens(selectedMessages) * requestRatio));
             if (this.config.debug) {
                 console.log(`[OK] Context managed: ${this.messageHistory.length} -> ${selectedMessages.length} messages`);
                 console.log(` Token reduction: ${currentTokens} -> ${newTokenCount} tokens (removed ${currentTokens - newTokenCount})`);
@@ -7704,7 +7724,8 @@ export class CortexOrchestrator {
         }
         if (!this.decisionStore) {
             const root = this.config.projectPath || process.cwd();
-            const storePath = `${root}/.cortex/decisions.jsonl`;
+            // 4.108.6 (R131): state root resolves to a WRITABLE location (read-only workdirs fall back; never fatal)
+            const storePath = cortexStatePath(root, 'decisions.jsonl');
             this.decisionStore = new DecisionStore(storePath);
         }
         return this.decisionStore;