scripts/test-hermes-evolution.ts

import axios from "axios";
import "dotenv/config";

const API_BASE = "http://localhost:3000/api/v1";
const API_KEY = process.env.API_KEYS?.split(",")[0] || "";

async function runHermesStressTest() {
  console.log("🚀 Starting HERMES Closed-Loop Evolution Stress Test...");

  try {
    // Phase 0: Get current policy
    console.log("\n--- Phase 0: Fetching Current Policy ---");
    const authRes = await axios.get(`${API_BASE}/auth/session`, {
        headers: { Authorization: `Bearer ${API_KEY}` }
    });
    console.log("Authenticated as:", authRes.data.tenant.name);

    // Phase 1: Generate Successful Baseline Outcomes
    console.log("\n--- Phase 1: Generating Successful Baseline Outcomes ---");
    for (let i = 0; i < 5; i++) {
        const traceId = `trace_success_${i}_${Date.now()}`;
        // Record a mock decision audit first so Hermes can correlate it
        await axios.post(`${API_BASE}/optimize`, {
            logs: [{ 
                step: "test", 
                agent: "TestAgent",
                model: "gpt-4",
                input: "test", 
                output: "test output",
                prompt_tokens: 100, 
                completion_tokens: 50 
            }],
            mode: "fast",
            decision: { mode: "recommend" }
        }, {
            headers: { Authorization: `Bearer ${API_KEY}`, "X-TCG-Trace-Id": traceId }
        });

        await axios.post(`${API_BASE}/feedback`, {
            trace_id: traceId,
            success: true,
            score: 0.95,
            feedback: "Perfect optimization"
        }, {
            headers: { Authorization: `Bearer ${API_KEY}` }
        });
        process.stdout.write(".");
    }
    console.log("\nDone.");

    // Phase 2: Generate Critical Failure Outcomes (Simulation)
    console.log("\n--- Phase 2: Generating Critical Failure Outcomes (Simulation) ---");
    console.log("Simulating a high failure rate...");
    for (let i = 0; i < 20; i++) {
        const traceId = `trace_fail_${i}_${Date.now()}`;
        // Record a mock decision audit
        await axios.post(`${API_BASE}/optimize`, {
            logs: [{ 
                step: "test", 
                agent: "TestAgent",
                model: "gpt-4",
                input: "test", 
                output: "test output",
                prompt_tokens: 100, 
                completion_tokens: 50 
            }],
            mode: "fast",
            decision: { mode: "recommend" }
        }, {
            headers: { Authorization: `Bearer ${API_KEY}`, "X-TCG-Trace-Id": traceId }
        });

        await axios.post(`${API_BASE}/feedback`, {
            trace_id: traceId,
            success: false,
            score: 0.2,
            feedback: "Context lost during pruning"
        }, {
            headers: { Authorization: `Bearer ${API_KEY}` }
        });
        process.stdout.write("X");
    }
    console.log("\nDone.");

    // Phase 3: Trigger Hermes Evolution
    console.log("\n--- Phase 3: Triggering Hermes Evolution ---");
    const evolveRes = await axios.post(`${API_BASE}/hermes/evolve`, {}, {
        headers: { Authorization: `Bearer ${API_KEY}` }
    });

    console.log("Status:", evolveRes.status);
    console.log("Rationale:", evolveRes.data.proposal.rationale);
    
    const newWeights = evolveRes.data.proposal.newWeights;
    const newThresholds = evolveRes.data.proposal.newThresholds;
    
    console.log("\nNew Weights:", JSON.stringify(newWeights, null, 2));
    console.log("New Thresholds:", JSON.stringify(newThresholds, null, 2));

    // Assertions
    if (evolveRes.data.proposal.rationale.includes("CRITICAL: Low success rate")) {
        console.log("\n✅ SUCCESS: Hermes detected the degradation and pivoted to Quality Restoration.");
    } else {
        console.log("\n❌ FAILURE: Hermes did not detect the degradation correctly.");
    }

  } catch (error: any) {
    console.error("\n❌ Test Failed:", error.response?.data || error.message);
  }
}

runHermesStressTest();