[claude] Agent debate on borderline eval requests (#21) (#72)

artifacts/api-server/src/routes/events.ts

@@ -214,6 +214,39 @@ function translateEvent(ev: BusEvent): object | null {
       }
       return null;
 
+    // ── Debate events (#21) ────────────────────────────────────────────────
+    case "debate:argument": {
+      void logWorldEvent(
+        "debate:argument",
+        `${ev.agent} argues to ${ev.position}: ${ev.argument.slice(0, 80)}`,
+        "beta",
+        ev.jobId,
+      );
+      return {
+        type: "agent_debate",
+        jobId: ev.jobId,
+        agent: ev.agent,
+        position: ev.position,
+        argument: ev.argument,
+      };
+    }
+    case "debate:verdict": {
+      void logWorldEvent(
+        "debate:verdict",
+        `Verdict: ${ev.accepted ? "accepted" : "rejected"} — ${ev.reason.slice(0, 80)}`,
+        "beta",
+        ev.jobId,
+      );
+      return {
+        type: "agent_debate",
+        jobId: ev.jobId,
+        agent: "Beta",
+        position: "verdict",
+        argument: `Final verdict: ${ev.accepted ? "ACCEPTED" : "REJECTED"} — ${ev.reason}`,
+        accepted: ev.accepted,
+      };
+    }
+
     default:
       return null;
   }

artifacts/api-server/src/routes/jobs.ts

@@ -1,6 +1,6 @@
 import { Router, type Request, type Response } from "express";
 import { randomUUID, createHash } from "crypto";
-import { db, jobs, invoices, type Job } from "@workspace/db";
+import { db, jobs, invoices, jobDebates, type Job } from "@workspace/db";
 import { eq, and } from "drizzle-orm";
 import { CreateJobBody, GetJobParams } from "@workspace/api-zod";
 import { lnbitsService } from "../lib/lnbits.js";
@@ -41,17 +41,74 @@ async function runEvalInBackground(
 ): Promise<void> {
   const evalStart = Date.now();
   try {
-    const evalResult = await agentService.evaluateRequest(request);
+    let evalResult = await agentService.evaluateRequest(request);
     latencyHistogram.record("eval_phase", Date.now() - evalStart);
 
     logger.info("eval result", {
       jobId,
       accepted:     evalResult.accepted,
       reason:       evalResult.reason,
+      confidence:   evalResult.confidence,
       inputTokens:  evalResult.inputTokens,
       outputTokens: evalResult.outputTokens,
     });
 
+    // ── Borderline debate (#21) ─────────────────────────────────────────
+    // When the eval model reports low confidence, run a mini debate to
+    // produce a more defensible accept/reject decision.
+    if (evalResult.confidence === "low") {
+      logger.info("borderline eval — starting debate", { jobId });
+      eventBus.publish({ type: "job:state", jobId, state: "evaluating" }); // keep beta thinking
+
+      const initialPosition = evalResult.accepted ? "accept" : "reject";
+      const debateResult = await agentService.runDebate(
+        request,
+        initialPosition as "accept" | "reject",
+        evalResult.reason,
+        (agent, position, argument) => {
+          eventBus.publish({ type: "debate:argument", jobId, agent, position, argument });
+        },
+      );
+
+      // Publish the final verdict event
+      eventBus.publish({
+        type: "debate:verdict",
+        jobId,
+        accepted: debateResult.verdict.accepted,
+        reason: debateResult.verdict.reason,
+      });
+
+      // Store debate transcript
+      try {
+        await db.insert(jobDebates).values({
+          id: randomUUID(),
+          jobId,
+          argFor: debateResult.argFor,
+          argAgainst: debateResult.argAgainst,
+          verdict: JSON.stringify(debateResult.verdict),
+          verdictAccepted: String(debateResult.verdict.accepted),
+          verdictReason: debateResult.verdict.reason,
+        });
+      } catch (dbErr) {
+        logger.warn("failed to store debate transcript", { jobId, err: String(dbErr) });
+      }
+
+      // Override the eval result with the debate's verdict
+      evalResult = {
+        accepted: debateResult.verdict.accepted,
+        reason: debateResult.verdict.reason,
+        confidence: "high", // post-debate, confidence is resolved
+        inputTokens: evalResult.inputTokens + debateResult.inputTokens,
+        outputTokens: evalResult.outputTokens + debateResult.outputTokens,
+      };
+
+      logger.info("debate concluded", {
+        jobId,
+        accepted: evalResult.accepted,
+        reason: evalResult.reason,
+      });
+    }
+
     if (evalResult.accepted) {
       const { estimatedInputTokens, estimatedOutputTokens } = pricingService.estimateRequestCost(request, agentService.workModel);
       const breakdown = await pricingService.calculateWorkFeeSats(