@kognai/orchestrator-core 0.1.3 → 0.2.0

package/dist/lib/engine-agents.js

@@ -45,7 +45,7 @@ class SupervisorAgent {
             : '';
         // Sherlock v2: inject ASMR episodic memory context (AMD-21-03) — fail-open
         const memoryContext = await (0, sherlock_memory_1.getSherlockMemoryContext)(task.context || task.id);
-        const userPrompt = `Review the following code generated for task ${task.id}.\n\n## Task Spec\n${task.context}${memoryContext}\n\n## Generated Files (${files.length})\n${fileContents}${integrityContext}\n\n## Pre-computed Fence Check (authoritative — do NOT infer from display format)\n${fenceCheckLines}\n\n## Instructions\nCRITICAL CHECK: Use the Pre-computed Fence Check above. If any file shows FENCE DETECTED, REJECT. Do NOT infer fence presence from the <file_content> display tags — those are display-only wrappers.\nAlso check: Did the file lose existing functionality? If a file shrank significantly, REJECT.\n\n## Categorical grade (use a discrete letter — no fake precision)\n- A: production-perfect. No improvements possible. Ship.\n- B: good with minor polish needed (rename, comment, formatting).\n- C: works but has noticeable issues (missed edge case, weak abstraction, partial spec coverage). APPROVED with caveats.\n- D: significant problems (broken edge case, regression risk, anti-pattern). REJECT.\n- F: broken, unsafe, doesn't meet spec, or fence/integrity failure. REJECT.\n\nThe deterministic gate already ran (typecheck + structural). If a file got here, syntax is valid — focus your review on substance, not parseability.\n\nRespond with a JSON object:\n{\n  "verdict": "APPROVED" or "REJECTED",\n  "grade": "A" | "B" | "C" | "D" | "F",\n  "score_rationale": "ONE sentence naming the specific factor that determined the grade. Vague 'good code' is NOT acceptable.",\n  "summary": "brief review summary",\n  "issues": [{"severity": "critical|high|medium|low", "file": "path", "description": "..."}],\n  "strengths": ["..."]\n}`;
+        const userPrompt = `Review the following code generated for task ${task.id}.\n\n## Task Spec\n${task.context}${memoryContext}\n\n## Generated Files (${files.length})\n${fileContents}${integrityContext}\n\n## Pre-computed Fence Check (authoritative — do NOT infer from display format)\n${fenceCheckLines}\n\n## YOUR REVIEW LENS — Specification & Integration (this is your ONLY job)\nYou are ONE of two INDEPENDENT reviewers. Judge ONLY: (1) SPEC COVERAGE — every required export/function/behavior is present and matches the task spec; REJECT if partial, stubbed, or TODO. (2) INTEGRATION — imports resolve to real files/symbols, types and contracts match the files this depends on, and referenced files exist. Do NOT base your grade on security or runtime concerns — the other reviewer owns those. This file passes your lens only if it is spec-complete AND integrates cleanly.\n\n## Instructions\nCRITICAL CHECK: Use the Pre-computed Fence Check above. If any file shows FENCE DETECTED, REJECT. Do NOT infer fence presence from the <file_content> display tags — those are display-only wrappers.\nAlso check: Did the file lose existing functionality? If a file shrank significantly, REJECT.\n\n## Categorical grade (use a discrete letter — no fake precision)\n- A: production-perfect. No improvements possible. Ship.\n- B: good with minor polish needed (rename, comment, formatting).\n- C: works but has noticeable issues (missed edge case, weak abstraction, partial spec coverage). APPROVED with caveats.\n- D: significant problems (broken edge case, regression risk, anti-pattern). REJECT.\n- F: broken, unsafe, doesn't meet spec, or fence/integrity failure. REJECT.\n\nThe deterministic gate already ran (typecheck + structural). If a file got here, syntax is valid — focus your review on substance, not parseability.\n\nRespond with a JSON object:\n{\n  "verdict": "APPROVED" or "REJECTED",\n  "grade": "A" | "B" | "C" | "D" | "F",\n  "score_rationale": "ONE sentence naming the specific factor that determined the grade. Vague 'good code' is NOT acceptable.",\n  "summary": "brief review summary",\n  "issues": [{"severity": "critical|high|medium|low", "file": "path", "description": "..."}],\n  "strengths": ["..."]\n}`;
         const startTime = Date.now();
         // B.15: DeepSeek via ClawRouter for standard tasks (~$0.02/task vs $0.07 dual-supervisor)
         // Retain Claude Sonnet only for audit/refactor-complex (high-stakes)
@@ -115,7 +115,7 @@ class Supervisor2Agent {
         const integrityContext2 = task._integrityFailed
             ? `\n\n## ⚠️ INTEGRITY ALERT\n${task._integrityDetails}\nThis file was flagged for destructive rewrite. The original was preserved. REJECT this task.\n`
             : '';
-        const userPrompt = `Review the following code generated for task ${task.id}.\n\n## Task Spec\n${task.context}\n\n## Generated Files (${files.length})\n${fileContents}${integrityContext2}\n\n## Pre-computed Fence Check (authoritative — do NOT infer from display format)\n${fenceCheckLines2}\n\n## Instructions\nCRITICAL CHECK: Use the Pre-computed Fence Check above. If any file shows FENCE DETECTED, REJECT. Do NOT infer fence presence from the <file_content> display tags — those are display-only wrappers.\nAlso check: Did the file lose existing functionality? If a file shrank significantly, REJECT.\n\n## Categorical grade (use a discrete letter — no fake precision)\n- A: production-perfect. No improvements possible. Ship.\n- B: good with minor polish needed (rename, comment, formatting).\n- C: works but has noticeable issues (missed edge case, weak abstraction, partial spec coverage). APPROVED with caveats.\n- D: significant problems (broken edge case, regression risk, anti-pattern). REJECT.\n- F: broken, unsafe, doesn't meet spec, or fence/integrity failure. REJECT.\n\nThe deterministic gate already ran (typecheck + structural). If a file got here, syntax is valid — focus your review on substance, not parseability.\n\nRespond with a JSON object:\n{\n  "verdict": "APPROVED" or "REJECTED",\n  "grade": "A" | "B" | "C" | "D" | "F",\n  "score_rationale": "ONE sentence naming the specific factor that determined the grade. Vague 'good code' is NOT acceptable.",\n  "summary": "brief review summary",\n  "issues": [{"severity": "critical|high|medium|low", "file": "path", "description": "..."}],\n  "strengths": ["..."]\n}`;
+        const userPrompt = `Review the following code generated for task ${task.id}.\n\n## Task Spec\n${task.context}\n\n## Generated Files (${files.length})\n${fileContents}${integrityContext2}\n\n## Pre-computed Fence Check (authoritative — do NOT infer from display format)\n${fenceCheckLines2}\n\n## YOUR REVIEW LENS — Security & Runtime (this is your ONLY job)\nYou are ONE of two INDEPENDENT reviewers. Judge ONLY: (1) SECURITY — injection, secret/credential leakage, unsafe eval/exec/shell, unsanitized input or output (e.g. innerHTML / unescaped HTML), missing authorization or input validation. (2) RUNTIME ROBUSTNESS — error handling, unhandled rejections, resource leaks, missing timeouts/cancellation, crash-on-bad-input. Do NOT re-judge spec completeness — the other reviewer owns that. This file passes your lens only if it is secure AND runtime-robust.\n\n## Instructions\nCRITICAL CHECK: Use the Pre-computed Fence Check above. If any file shows FENCE DETECTED, REJECT. Do NOT infer fence presence from the <file_content> display tags — those are display-only wrappers.\nAlso check: Did the file lose existing functionality? If a file shrank significantly, REJECT.\n\n## Categorical grade (use a discrete letter — no fake precision)\n- A: production-perfect. No improvements possible. Ship.\n- B: good with minor polish needed (rename, comment, formatting).\n- C: works but has noticeable issues (missed edge case, weak abstraction, partial spec coverage). APPROVED with caveats.\n- D: significant problems (broken edge case, regression risk, anti-pattern). REJECT.\n- F: broken, unsafe, doesn't meet spec, or fence/integrity failure. REJECT.\n\nThe deterministic gate already ran (typecheck + structural). If a file got here, syntax is valid — focus your review on substance, not parseability.\n\nRespond with a JSON object:\n{\n  "verdict": "APPROVED" or "REJECTED",\n  "grade": "A" | "B" | "C" | "D" | "F",\n  "score_rationale": "ONE sentence naming the specific factor that determined the grade. Vague 'good code' is NOT acceptable.",\n  "summary": "brief review summary",\n  "issues": [{"severity": "critical|high|medium|low", "file": "path", "description": "..."}],\n  "strengths": ["..."]\n}`;
         const startTime = Date.now();
         // B.15: Use Haiku for second-pass review — 10x cheaper than Sonnet.
         // Founder directive 2026-05-25: if Anthropic depletes, fall back to ClawRouter/DeepSeek
@@ -199,68 +199,53 @@ async function reconcileSupervisorReviews(review1, review2, task, ceo) {
         return { finalReview: review1, review1, review2, consensus: false, escalatedToCEO: false };
     }
     const bothApproved = review1.verdict === 'APPROVED' && review2.verdict === 'APPROVED';
-    const bothRejected = review1.verdict !== 'APPROVED' && review2.verdict !== 'APPROVED';
-    const consensus = bothApproved || bothRejected;
     if (bothApproved) {
-        // Both approve — take the average score, merge strengths
+        // Both lenses passed — average score, merge strengths.
         const avgScore = Math.round((review1.score + review2.score) / 2);
-        (0, orchestrate_engine_1.log)(orchestrate_engine_1.c.green, `  ✓ DUAL CONSENSUS: Both supervisors APPROVED (Sup1: ${review1.score}, Sup2: ${review2.score}, avg: ${avgScore})`);
+        (0, orchestrate_engine_1.log)(orchestrate_engine_1.c.green, `  ✓ DUAL PASS: both lenses APPROVED — Spec/Integration (${review1.score}) + Security/Runtime (${review2.score}), avg ${avgScore}`);
         return {
             finalReview: {
                 verdict: 'APPROVED',
                 score: avgScore,
-                summary: `Dual-approved: Sup1 (${review1.score}/100) + Sup2 (${review2.score}/100)`,
+                summary: `Both lenses passed: Spec/Integration ${review1.score}/100 + Security/Runtime ${review2.score}/100`,
                 issues: [...review1.issues, ...review2.issues],
                 strengths: Array.from(new Set([...review1.strengths, ...review2.strengths])),
             },
             review1, review2, consensus: true, escalatedToCEO: false,
         };
     }
-    if (bothRejected) {
-        // Both reject — merge issues, take lower score
-        const minScore = Math.min(review1.score, review2.score);
-        (0, orchestrate_engine_1.log)(orchestrate_engine_1.c.red, `  ✗ DUAL CONSENSUS: Both supervisors REJECTED (Sup1: ${review1.score}, Sup2: ${review2.score})`);
-        return {
-            finalReview: {
-                verdict: 'REJECTED',
-                score: minScore,
-                summary: `Dual-rejected: Sup1 (${review1.score}/100) + Sup2 (${review2.score}/100). ${review1.summary} | ${review2.summary}`,
-                issues: [...review1.issues, ...review2.issues],
-                strengths: [],
-            },
-            review1, review2, consensus: true, escalatedToCEO: false,
-        };
-    }
-    // CONFLICT — one approved, one rejected → escalate to CEO
-    const approver = review1.verdict === 'APPROVED' ? 'Sup1' : 'Sup2';
-    const rejecter = review1.verdict === 'APPROVED' ? 'Sup2' : 'Sup1';
-    const approvalReview = review1.verdict === 'APPROVED' ? review1 : review2;
-    const rejectionReview = review1.verdict === 'APPROVED' ? review2 : review1;
-    (0, orchestrate_engine_1.log)(orchestrate_engine_1.c.yellow, `  ⚡ SUPERVISOR CONFLICT on ${task.id}: ${approver} APPROVED (${approvalReview.score}), ${rejecter} REJECTED (${rejectionReview.score})`);
-    (0, orchestrate_engine_1.log)(orchestrate_engine_1.c.magenta, `  → Escalating to CEO for final decision...`);
-    try {
-        const ceoDecision = await ceo.resolveReviewConflict(task, approvalReview, rejectionReview, approver, rejecter);
-        const ceoApproves = ceoDecision.toLowerCase().includes('approve');
-        (0, orchestrate_engine_1.log)(ceoApproves ? orchestrate_engine_1.c.green : orchestrate_engine_1.c.red, `  CEO DECISION: ${ceoApproves ? 'APPROVED' : 'REJECTED'} — ${ceoDecision.substring(0, 200)}`);
-        return {
-            finalReview: {
-                verdict: ceoApproves ? 'APPROVED' : 'REJECTED',
-                score: ceoApproves ? approvalReview.score : rejectionReview.score,
-                summary: `CEO resolved conflict (${approver} approved, ${rejecter} rejected): ${ceoDecision.substring(0, 300)}`,
-                issues: rejectionReview.issues,
-                strengths: approvalReview.strengths,
-            },
-            review1, review2, consensus: false, escalatedToCEO: true, ceoDecision,
-        };
-    }
-    catch (error) {
-        // CEO unavailable — default to rejection (safer)
-        (0, orchestrate_engine_1.log)(orchestrate_engine_1.c.yellow, `  CEO unavailable for conflict resolution: ${error.message}. Defaulting to REJECTED.`);
-        return {
-            finalReview: rejectionReview,
-            review1, review2, consensus: false, escalatedToCEO: false,
-        };
-    }
+    // BOTH-MUST-PASS (gap 3). The two reviewers cover DIFFERENT dimensions
+    // (spec/integration vs security/runtime), so passing one does not excuse
+    // failing the other. Any rejection ⇒ REJECTED. No CEO rescue — a real
+    // security or spec failure must be FIXED, not voted away by a third opinion
+    // that never looked at that dimension. This replaces the old
+    // conflict→CEO-decides path that let a single approval override a rejection.
+    const failedLenses = [];
+    if (review1.verdict !== 'APPROVED')
+        failedLenses.push('Spec/Integration');
+    if (review2.verdict !== 'APPROVED')
+        failedLenses.push('Security/Runtime');
+    const rejectedIssues = [
+        ...(review1.verdict !== 'APPROVED' ? review1.issues : []),
+        ...(review2.verdict !== 'APPROVED' ? review2.issues : []),
+    ];
+    const rejectionSummaries = [review1, review2]
+        .filter((r) => r.verdict !== 'APPROVED')
+        .map((r) => r.summary)
+        .join(' | ');
+    (0, orchestrate_engine_1.log)(orchestrate_engine_1.c.red, `  ✗ REJECTED — failed required lens: ${failedLenses.join(' + ')} (both must pass). Spec/Int ${review1.score}, Sec/RT ${review2.score}`);
+    return {
+        finalReview: {
+            verdict: 'REJECTED',
+            score: Math.min(review1.score, review2.score),
+            summary: `Failed required lens (${failedLenses.join(' + ')}): ${rejectionSummaries}`,
+            issues: rejectedIssues,
+            strengths: [],
+        },
+        review1, review2,
+        consensus: review1.verdict === review2.verdict,
+        escalatedToCEO: false,
+    };
 }
 // ===== CEO Agent (Claude via Anthropic API) =====
 class CEOAgent {
@@ -750,6 +735,7 @@ class AgentCreator {
         // supplied a SpawnGate (Kognai wires SAF here), consult it BEFORE creating
         // anything on disk. Approval/rejection only; the citizenship logic below is
         // unchanged (its extraction is tracked separately as TICKET-226).
+        let spawnOwner;
         if (this.spawnGate) {
             const decision = this.spawnGate(spec);
             if (!decision.approved) {
@@ -764,6 +750,9 @@ class AgentCreator {
             }
             if (decision.audit)
                 (0, orchestrate_engine_1.log)(orchestrate_engine_1.c.gray, `  ✓ ${decision.audit}`);
+            // The gate (SAF) resolves the lineage from its requester_did — this is the
+            // running company's context, plumbed in rather than hardcoded here.
+            spawnOwner = decision.owner;
         }
         const agentDir = `./agents/${spec.name}`;
         (0, fs_1.mkdirSync)(agentDir, { recursive: true });
@@ -771,10 +760,13 @@ class AgentCreator {
         // citizen — not a bare agent. Mint citizenship (citizen_id + roll
         // number + Kōpus avatar + ACP baseline) BEFORE writing the agent
         // files so the citizen record can be referenced in the prompt.
+        // Owner-scoped when the gate supplied a lineage (e.g. invoica/voxight);
+        // legacy kognai-internal path otherwise (back-compat for gate-less templates).
         const citizen = (0, citizenship_1.mintCitizen)(spec.name, {
             founding_agent: 'ceo',
             proposing_agent: 'cto',
             citizen_type: 'spawned',
+            owner: spawnOwner,
         });
         // Write agent.yaml
         const yaml = `name: ${spec.name}

package/dist/lib/engine-orchestrator.d.ts

@@ -7,6 +7,8 @@ export declare class Orchestrator {
     private supervisor2;
     private agents;
     private tasks;
+    private buildPath;
+    private triage;
     private stats;
     private taskRuns;
     /**