sentinelayer-cli 0.6.2 → 0.8.1

package/src/agents/jules/loop.js

@@ -1,450 +1,460 @@
-import { randomUUID } from "node:crypto";
-import { createMultiProviderApiClient } from "../../ai/client.js";
-import { evaluateBudget } from "../../cost/budget.js";
-import { dispatchTool, createAgentContext, BudgetExhaustedError } from "./tools/dispatch.js";
-import { JULES_DEFINITION } from "./config/definition.js";
-import { shouldSpawnSubAgents, runJulesSwarm } from "./swarm/orchestrator.js";
-import { frontendAnalyze } from "./tools/frontend-analyze.js";
-
-/**
- * Jules Tanaka — Agentic Loop
- *
- * Core state machine: LLM → tool_use → execute → result → LLM → repeat
- * With sub-agent swarm integration for large codebases.
- *
- * This loop is self-contained: it uses the existing ai/client.js for LLM calls,
- * the existing cost/budget.js for budget enforcement, and the Jules tool
- * dispatch for tool execution. No dependency on Batches O-Q.
- */
-
-const DEFAULT_MAX_TURNS = 25;
-const HEARTBEAT_INTERVAL_TURNS = 5;
-
-/**
- * Run Jules' agentic audit loop.
- *
- * @param {object} config
- * @param {string} config.systemPrompt - Jules' full system prompt
- * @param {object} config.scopeMap - { primary, secondary, tertiary } file lists
- * @param {string} config.rootPath - Codebase root
- * @param {object} [config.omarBaseline] - Deterministic baseline findings (if available)
- * @param {object} [config.blackboard] - Shared blackboard for cross-agent findings
- * @param {object} [config.memory] - Memory index for cross-run recall
- * @param {object} [config.budget] - Budget overrides
- * @param {object} [config.provider] - LLM provider overrides
- * @param {string} [config.mode] - "primary" | "secondary" | "tertiary"
- * @param {number} [config.maxTurns] - Max loop iterations
- * @param {AbortController} [config.abortController]
- * @param {function} [config.onEvent] - Streaming event callback
- * @returns {AsyncGenerator<JulesEvent>} Yields events as they occur
- */
-export async function* julesAuditLoop(config) {
-  const {
-    systemPrompt,
-    scopeMap,
-    rootPath,
-    omarBaseline,
-    blackboard,
-    memory,
-    provider,
-    mode = "primary",
-    maxTurns = DEFAULT_MAX_TURNS,
-    abortController,
-    onEvent,
-  } = config;
-
-  const budget = { ...JULES_DEFINITION.budget, ...config.budget };
-  const runId = `jules-${Date.now()}-${randomUUID().slice(0, 8)}`;
-  const startedAt = Date.now();
-  const client = createMultiProviderApiClient(provider || {});
-
-  const ctx = createAgentContext({
-    agentIdentity: { id: JULES_DEFINITION.id, persona: JULES_DEFINITION.persona },
-    budget,
-    runId,
-    onEvent,
-  });
-
-  const emit = (event, payload) => {
-    const evt = {
-      stream: "sl_event",
-      event,
-      agent: { id: JULES_DEFINITION.id, persona: JULES_DEFINITION.persona, color: JULES_DEFINITION.color, avatar: JULES_DEFINITION.avatar },
-      payload,
-      usage: {
-        costUsd: ctx.usage.costUsd,
-        outputTokens: ctx.usage.outputTokens,
-        toolCalls: ctx.usage.toolCalls,
-        durationMs: Date.now() - startedAt,
-      },
-    };
-    if (onEvent) onEvent(evt);
-    return evt;
-  };
-
-  yield emit("agent_start", { mode, runId, maxTurns, budget });
-
-  // ── Phase 0: Prerequisites ────────────────────────────────────────
-
-  yield emit("progress", { phase: "prerequisites", message: "Detecting framework..." });
-
-  let framework = {};
-  try {
-    framework = frontendAnalyze({ operation: "detect_framework", path: rootPath });
-    ctx.usage.toolCalls++;
-    yield emit("tool_result", { tool: "FrontendAnalyze", operation: "detect_framework", result: { framework: framework.framework, componentCount: framework.componentCount } });
-  } catch { /* proceed without */ }
-
-  // ── Phase 1: Swarm or direct? ─────────────────────────────────────
-
-  const spawnDecision = shouldSpawnSubAgents(scopeMap);
-  let swarmFindings = [];
-
-  if (spawnDecision.spawn && blackboard) {
-    yield emit("progress", { phase: "swarm", message: `Large frontend (${spawnDecision.reason}). Spawning sub-agents...` });
-
-    const swarmResult = await runJulesSwarm({
-      scopeMap,
-      rootPath,
-      blackboard,
-      budget: { ...budget, maxCostUsd: budget.maxCostUsd * 0.6 }, // 60% for swarm
-      provider,
-      parentAbort: abortController,
-      onEvent,
-    });
-
-    swarmFindings = swarmResult.agentResults.flatMap(r => r.findings);
-    ctx.usage.costUsd += swarmResult.usage.totalCostUsd;
-    ctx.usage.toolCalls += swarmResult.usage.totalToolCalls;
-
-    yield emit("swarm_complete", {
-      totalFindings: swarmFindings.length,
-      totalAgents: swarmResult.usage.totalAgents,
-      totalCostUsd: swarmResult.usage.totalCostUsd,
-    });
-  }
-
-  // ── Phase 2: Jules primary deep analysis (agentic LLM loop) ──────
-
-  yield emit("progress", { phase: "deep_analysis", message: "Starting deep analysis..." });
-
-  // Build context for LLM — BLIND-FIRST: no Omar baseline or swarm findings
-  // in the initial context. Only codebase metadata and memory recall (past runs,
-  // not current-run findings). Swarm/baseline reconciliation happens AFTER the
-  // independent deep analysis completes.
-  const contextParts = [];
-  contextParts.push(`Framework: ${framework.framework || "unknown"}`);
-  contextParts.push(`Mode: ${mode}`);
-  contextParts.push(`Components: ${framework.componentCount || "unknown"}`);
-  contextParts.push(`Scope: ${(scopeMap.primary || []).length} primary files`);
-
-  if (memory) {
-    try {
-      const recalled = memory.query ? memory.query({
-        files: (scopeMap.primary || []).map(f => f.path || f),
-        limit: 10,
-      }) : [];
-      if (recalled.length > 0) {
-        contextParts.push(`\nPrevious findings recalled from memory (${recalled.length}):`);
-        for (const r of recalled) {
-          contextParts.push(`- ${r.content || r.text || JSON.stringify(r).slice(0, 100)}`);
-        }
-      }
-    } catch { /* memory recall failure is non-blocking */ }
-  }
-
-  const messages = [
-    { role: "user", content: contextParts.join("\n") +
-      "\n\nPerform your deep analysis now. Use FileRead, Grep, Glob, and FrontendAnalyze tools as needed. " +
-      "Return your findings in a ```json code block as an array of { severity, file, line, title, evidence, rootCause, recommendedFix, trafficLight, reproduction, user_impact, confidence }." },
-  ];
-
-  const allFindings = [...swarmFindings];
-  let turnCount = 0;
-
-  while (turnCount < maxTurns) {
-    if (abortController?.signal.aborted) {
-      yield emit("agent_abort", { reason: "user_cancelled" });
-      break;
-    }
-
-    // Budget check before LLM call
-    const preCheck = evaluateBudget({
-      sessionSummary: {
-        costUsd: ctx.usage.costUsd,
-        outputTokens: ctx.usage.outputTokens,
-        durationMs: Date.now() - startedAt,
-        toolCalls: ctx.usage.toolCalls,
-      },
-      ...budget,
-    });
-
-    if (preCheck.blocking) {
-      yield emit("budget_stop", { reasons: preCheck.reasons });
-      break;
-    }
-
-    if (preCheck.warnings.length > 0) {
-      yield emit("budget_warning", { warnings: preCheck.warnings });
-    }
-
-    turnCount++;
-
-    // Heartbeat
-    if (turnCount % HEARTBEAT_INTERVAL_TURNS === 0) {
-      yield emit("heartbeat", {
-        turnsCompleted: turnCount,
-        turnsMax: maxTurns,
-        findingsSoFar: allFindings.length,
-        budgetRemaining: {
-          costUsd: Math.max(0, budget.maxCostUsd - ctx.usage.costUsd),
-          pct: Math.max(0, 100 - (ctx.usage.costUsd / budget.maxCostUsd * 100)),
-        },
-      });
-    }
-
-    // Call LLM — format system prompt + messages into a single prompt
-    // for the MultiProviderApiClient which uses a completions-style API
-    let response;
-    try {
-      response = await client.invoke({
-        prompt: formatPromptForClient(systemPrompt, messages),
-      });
-    } catch (err) {
-      yield emit("llm_error", { error: err.message, turn: turnCount });
-      break;
-    }
-
-    const responseText = response.text || "";
-    ctx.usage.outputTokens += Math.ceil(responseText.length / 4);
-    ctx.usage.costUsd += (Math.ceil(responseText.length / 4) / 1_000_000) * 15;
-
-    yield emit("reasoning", {
-      phase: "deep_analysis",
-      turn: turnCount,
-      summary: responseText.slice(0, 200),
-    });
-
-    // Parse tool_use blocks
-    const toolCalls = parseToolUseBlocks(responseText);
-
-    if (toolCalls.length === 0) {
-      // No tools — extract findings from response
-      const parsed = extractJsonFindings(responseText);
-      for (const finding of parsed) {
-        allFindings.push(finding);
-        yield emit("finding", { ...finding });
-        if (blackboard) {
-          try {
-            await blackboard.appendEntry({
-              agentId: JULES_DEFINITION.id,
-              source: "jules-primary",
-              ...finding,
-            });
-          } catch { /* blackboard write failure non-blocking */ }
-        }
-      }
-      messages.push({ role: "assistant", content: responseText });
-      break; // LLM is done
-    }
-
-    // Execute tool calls
-    const results = [];
-    for (const call of toolCalls) {
-      try {
-        const result = await dispatchTool(call.tool, call.input, ctx);
-        results.push({ tool: call.tool, result });
-        yield emit("tool_call", { tool: call.tool, input: sanitizeForEvent(call.input) });
-      } catch (err) {
-        if (err instanceof BudgetExhaustedError) {
-          yield emit("budget_stop", { reason: err.message });
-          break;
-        }
-        results.push({ tool: call.tool, error: err.message });
-      }
-    }
-
-    // Feed results back
-    messages.push({ role: "assistant", content: responseText });
-    messages.push({
-      role: "user",
-      content: results.map(r =>
-        r.error
-          ? `Tool ${r.tool} failed: ${r.error}`
-          : `Tool ${r.tool} result:\n${JSON.stringify(r.result).slice(0, 3000)}`,
-      ).join("\n\n") + "\n\nContinue your analysis. If done, return findings in a ```json code block.",
-    });
-  }
-
-  // ── Phase 2b: Reconciliation (post-blind-pass) ─────────────────────
-  // Now that the independent analysis is complete, cross-reference with
-  // swarm findings and Omar baseline. This preserves blind-first: the
-  // persona formed its own opinion before seeing prior conclusions.
-
-  const hasSwarmContext = swarmFindings.length > 0;
-  const baselineFindings = omarBaseline
-    ? (omarBaseline.findings || omarBaseline.summary || [])
-    : [];
-  const hasBaselineContext = Array.isArray(baselineFindings) && baselineFindings.length > 0;
-
-  if (hasSwarmContext || hasBaselineContext) {
-    yield emit("progress", { phase: "reconciliation", message: "Cross-referencing with sub-agent and baseline findings..." });
-
-    const reconcileParts = [];
-    reconcileParts.push("Your independent analysis is complete. Now cross-reference with the following prior findings.");
-    reconcileParts.push("For each prior finding: confirm if your analysis agrees, dispute with evidence if you disagree, or flag as missed if you did not cover it.");
-
-    if (hasSwarmContext) {
-      reconcileParts.push(`\nYour sub-agents found ${swarmFindings.length} findings:`);
-      for (const f of swarmFindings.slice(0, 30)) {
-        reconcileParts.push(`- [${f.severity || "P3"}] ${f.file || ""}:${f.line || ""} ${f.title || f.type || ""}`);
-      }
-    }
-
-    if (hasBaselineContext) {
-      reconcileParts.push(`\nOmar baseline reported ${baselineFindings.length} findings:`);
-      for (const f of baselineFindings.slice(0, 20)) {
-        reconcileParts.push(`- [${f.severity || ""}] ${f.file || ""}:${f.line || ""} ${f.message || f.title || ""}`);
-      }
-    }
-
-    reconcileParts.push("\nReturn any additional or revised findings as a JSON array in a ```json code block. If no changes, return an empty array [].");
-
-    messages.push({ role: "user", content: reconcileParts.join("\n") });
-
-    // Budget check before reconciliation turn
-    const reconcilePreCheck = evaluateBudget({
-      sessionSummary: {
-        costUsd: ctx.usage.costUsd,
-        outputTokens: ctx.usage.outputTokens,
-        durationMs: Date.now() - startedAt,
-        toolCalls: ctx.usage.toolCalls,
-      },
-      ...budget,
-    });
-
-    if (!reconcilePreCheck.blocking) {
-      try {
-        const reconcileResponse = await client.invoke({
-          prompt: formatPromptForClient(systemPrompt, messages),
-        });
-
-        const reconcileText = reconcileResponse.text || "";
-        ctx.usage.outputTokens += Math.ceil(reconcileText.length / 4);
-        ctx.usage.costUsd += (Math.ceil(reconcileText.length / 4) / 1_000_000) * 15;
-
-        yield emit("reasoning", { phase: "reconciliation", summary: reconcileText.slice(0, 200) });
-
-        const reconcileFindings = extractJsonFindings(reconcileText);
-        for (const finding of reconcileFindings) {
-          allFindings.push(finding);
-          yield emit("finding", { ...finding, source: "reconciliation" });
-          if (blackboard) {
-            try {
-              await blackboard.appendEntry({
-                agentId: JULES_DEFINITION.id,
-                source: "jules-reconciliation",
-                ...finding,
-              });
-            } catch { /* blackboard write failure non-blocking */ }
-          }
-        }
-
-        messages.push({ role: "assistant", content: reconcileText });
-      } catch (err) {
-        yield emit("llm_error", { error: err.message, phase: "reconciliation" });
-      }
-    } else {
-      yield emit("budget_stop", { reasons: reconcilePreCheck.reasons, phase: "reconciliation" });
-    }
-  }
-
-  // ── Phase 3: Build final report ───────────────────────────────────
-
-  const durationMs = Date.now() - startedAt;
-  const severityCounts = { P0: 0, P1: 0, P2: 0, P3: 0 };
-  for (const f of allFindings) {
-    const sev = (f.severity || "P3").toUpperCase();
-    if (severityCounts[sev] !== undefined) severityCounts[sev]++;
-    else severityCounts.P3++;
-  }
-
-  const report = {
-    runId,
-    persona: JULES_DEFINITION.persona,
-    mode,
-    framework: framework.framework || "unknown",
-    status: "completed",
-    findings: allFindings,
-    summary: {
-      total: allFindings.length,
-      ...severityCounts,
-      blocking: severityCounts.P0 > 0 || severityCounts.P1 > 0,
-    },
-    usage: {
-      turns: turnCount,
-      costUsd: ctx.usage.costUsd,
-      outputTokens: ctx.usage.outputTokens,
-      toolCalls: ctx.usage.toolCalls,
-      durationMs,
-    },
-    signature: JULES_DEFINITION.signature,
-  };
-
-  yield emit("agent_complete", {
-    ...report.summary,
-    costUsd: ctx.usage.costUsd,
-    durationMs,
-    turns: turnCount,
-  });
-
-  return report;
-}
-
-// ── Helpers ──────────────────────────────────────────────────────────
-
-function parseToolUseBlocks(text) {
-  const calls = [];
-  const regex = /```tool_use\s*\n([\s\S]*?)```/g;
-  let match;
-  while ((match = regex.exec(text)) !== null) {
-    try {
-      const parsed = JSON.parse(match[1].trim());
-      if (parsed.tool && parsed.input) calls.push(parsed);
-    } catch { /* skip malformed */ }
-  }
-  return calls;
-}
-
-function extractJsonFindings(text) {
-  const jsonMatch = text.match(/```json\s*\n([\s\S]*?)```/);
-  if (!jsonMatch) return [];
-  try {
-    const parsed = JSON.parse(jsonMatch[1].trim());
-    if (Array.isArray(parsed)) return parsed;
-    if (parsed.findings && Array.isArray(parsed.findings)) return parsed.findings;
-  } catch { /* skip malformed */ }
-  return [];
-}
-
-function sanitizeForEvent(input) {
-  const sanitized = { ...input };
-  if (typeof sanitized.content === "string" && sanitized.content.length > 200) {
-    sanitized.content = `[${sanitized.content.length} chars]`;
-  }
-  return sanitized;
-}
-
-/**
- * Format system prompt + chat messages into a single prompt string
- * for MultiProviderApiClient which uses a completions-style API.
- */
-function formatPromptForClient(systemPrompt, messages) {
-  const parts = [];
-  if (systemPrompt) parts.push(systemPrompt);
-  for (const msg of messages) {
-    const role = msg.role === "assistant" ? "ASSISTANT" : "USER";
-    parts.push(`\n${role}:\n${msg.content}`);
-  }
-  return parts.join("\n");
-}
+import { randomUUID } from "node:crypto";
+import { createMultiProviderApiClient } from "../../ai/client.js";
+import { evaluateBudget } from "../../cost/budget.js";
+import { estimateTokens } from "../../cost/tokenizer.js";
+import { dispatchTool, createAgentContext, BudgetExhaustedError } from "./tools/dispatch.js";
+import { JULES_DEFINITION } from "./config/definition.js";
+import { shouldSpawnSubAgents, runJulesSwarm } from "./swarm/orchestrator.js";
+import { frontendAnalyze } from "./tools/frontend-analyze.js";
+import { createAgentEvent } from "../../events/schema.js";
+
+/**
+ * Jules Tanaka — Agentic Loop
+ *
+ * Core state machine: LLM → tool_use → execute → result → LLM → repeat
+ * With sub-agent swarm integration for large codebases.
+ *
+ * This loop is self-contained: it uses the existing ai/client.js for LLM calls,
+ * the existing cost/budget.js for budget enforcement, and the Jules tool
+ * dispatch for tool execution. No dependency on Batches O-Q.
+ */
+
+const DEFAULT_MAX_TURNS = 25;
+const HEARTBEAT_INTERVAL_TURNS = 5;
+
+/**
+ * Run Jules' agentic audit loop.
+ *
+ * @param {object} config
+ * @param {string} config.systemPrompt - Jules' full system prompt
+ * @param {object} config.scopeMap - { primary, secondary, tertiary } file lists
+ * @param {string} config.rootPath - Codebase root
+ * @param {object} [config.omarBaseline] - Deterministic baseline findings (if available)
+ * @param {object} [config.blackboard] - Shared blackboard for cross-agent findings
+ * @param {object} [config.memory] - Memory index for cross-run recall
+ * @param {object} [config.budget] - Budget overrides
+ * @param {object} [config.provider] - LLM provider overrides
+ * @param {string} [config.mode] - "primary" | "secondary" | "tertiary"
+ * @param {number} [config.maxTurns] - Max loop iterations
+ * @param {AbortController} [config.abortController]
+ * @param {function} [config.onEvent] - Streaming event callback
+ * @returns {AsyncGenerator<JulesEvent>} Yields events as they occur
+ */
+export async function* julesAuditLoop(config) {
+  const {
+    systemPrompt,
+    scopeMap,
+    rootPath,
+    omarBaseline,
+    blackboard,
+    memory,
+    provider,
+    mode = "primary",
+    maxTurns = DEFAULT_MAX_TURNS,
+    abortController,
+    onEvent,
+  } = config;
+
+  const budget = { ...JULES_DEFINITION.budget, ...config.budget };
+  const runId = `jules-${Date.now()}-${randomUUID().slice(0, 8)}`;
+  const startedAt = Date.now();
+  const client = createMultiProviderApiClient(provider || {});
+
+  const ctx = createAgentContext({
+    agentIdentity: { id: JULES_DEFINITION.id, persona: JULES_DEFINITION.persona },
+    budget,
+    runId,
+    onEvent,
+  });
+
+  const emit = (event, payload) => {
+    const evt = createAgentEvent({
+      event,
+      agent: {
+        id: JULES_DEFINITION.id,
+        persona: JULES_DEFINITION.persona,
+        color: JULES_DEFINITION.color,
+        avatar: JULES_DEFINITION.avatar,
+      },
+      payload,
+      usage: {
+        costUsd: ctx.usage.costUsd,
+        outputTokens: ctx.usage.outputTokens,
+        toolCalls: ctx.usage.toolCalls,
+        durationMs: Date.now() - startedAt,
+      },
+      runId,
+      sessionId: ctx.sessionId,
+    });
+    if (onEvent) onEvent(evt);
+    return evt;
+  };
+
+  yield emit("agent_start", { mode, runId, maxTurns, budget });
+
+  // ── Phase 0: Prerequisites ────────────────────────────────────────
+
+  yield emit("progress", { phase: "prerequisites", message: "Detecting framework..." });
+
+  let framework = {};
+  try {
+    framework = frontendAnalyze({ operation: "detect_framework", path: rootPath });
+    ctx.usage.toolCalls++;
+    yield emit("tool_result", { tool: "FrontendAnalyze", operation: "detect_framework", result: { framework: framework.framework, componentCount: framework.componentCount } });
+  } catch { /* proceed without */ }
+
+  // ── Phase 1: Swarm or direct? ─────────────────────────────────────
+
+  const spawnDecision = shouldSpawnSubAgents(scopeMap);
+  let swarmFindings = [];
+
+  if (spawnDecision.spawn && blackboard) {
+    yield emit("progress", { phase: "swarm", message: `Large frontend (${spawnDecision.reason}). Spawning sub-agents...` });
+
+    const swarmResult = await runJulesSwarm({
+      scopeMap,
+      rootPath,
+      blackboard,
+      budget: { ...budget, maxCostUsd: budget.maxCostUsd * 0.6 }, // 60% for swarm
+      provider,
+      parentAbort: abortController,
+      onEvent,
+    });
+
+    swarmFindings = swarmResult.agentResults.flatMap(r => r.findings);
+    ctx.usage.costUsd += swarmResult.usage.totalCostUsd;
+    ctx.usage.toolCalls += swarmResult.usage.totalToolCalls;
+
+    yield emit("swarm_complete", {
+      totalFindings: swarmFindings.length,
+      totalAgents: swarmResult.usage.totalAgents,
+      totalCostUsd: swarmResult.usage.totalCostUsd,
+    });
+  }
+
+  // ── Phase 2: Jules primary deep analysis (agentic LLM loop) ──────
+
+  yield emit("progress", { phase: "deep_analysis", message: "Starting deep analysis..." });
+
+  // Build context for LLM — BLIND-FIRST: no Omar baseline or swarm findings
+  // in the initial context. Only codebase metadata and memory recall (past runs,
+  // not current-run findings). Swarm/baseline reconciliation happens AFTER the
+  // independent deep analysis completes.
+  const contextParts = [];
+  contextParts.push(`Framework: ${framework.framework || "unknown"}`);
+  contextParts.push(`Mode: ${mode}`);
+  contextParts.push(`Components: ${framework.componentCount || "unknown"}`);
+  contextParts.push(`Scope: ${(scopeMap.primary || []).length} primary files`);
+
+  if (memory) {
+    try {
+      const recalled = memory.query ? memory.query({
+        files: (scopeMap.primary || []).map(f => f.path || f),
+        limit: 10,
+      }) : [];
+      if (recalled.length > 0) {
+        contextParts.push(`\nPrevious findings recalled from memory (${recalled.length}):`);
+        for (const r of recalled) {
+          contextParts.push(`- ${r.content || r.text || JSON.stringify(r).slice(0, 100)}`);
+        }
+      }
+    } catch { /* memory recall failure is non-blocking */ }
+  }
+
+  const messages = [
+    { role: "user", content: contextParts.join("\n") +
+      "\n\nPerform your deep analysis now. Use FileRead, Grep, Glob, and FrontendAnalyze tools as needed. " +
+      "Return your findings in a ```json code block as an array of { severity, file, line, title, evidence, rootCause, recommendedFix, trafficLight, reproduction, user_impact, confidence }." },
+  ];
+
+  const allFindings = [...swarmFindings];
+  let turnCount = 0;
+
+  while (turnCount < maxTurns) {
+    if (abortController?.signal.aborted) {
+      yield emit("agent_abort", { reason: "user_cancelled" });
+      break;
+    }
+
+    // Budget check before LLM call
+    const preCheck = evaluateBudget({
+      sessionSummary: {
+        costUsd: ctx.usage.costUsd,
+        outputTokens: ctx.usage.outputTokens,
+        durationMs: Date.now() - startedAt,
+        toolCalls: ctx.usage.toolCalls,
+      },
+      ...budget,
+    });
+
+    if (preCheck.blocking) {
+      yield emit("budget_stop", { reasons: preCheck.reasons });
+      break;
+    }
+
+    if (preCheck.warnings.length > 0) {
+      yield emit("budget_warning", { warnings: preCheck.warnings });
+    }
+
+    turnCount++;
+
+    // Heartbeat
+    if (turnCount % HEARTBEAT_INTERVAL_TURNS === 0) {
+      yield emit("heartbeat", {
+        turnsCompleted: turnCount,
+        turnsMax: maxTurns,
+        findingsSoFar: allFindings.length,
+        budgetRemaining: {
+          costUsd: Math.max(0, budget.maxCostUsd - ctx.usage.costUsd),
+          pct: Math.max(0, 100 - (ctx.usage.costUsd / budget.maxCostUsd * 100)),
+        },
+      });
+    }
+
+    // Call LLM — format system prompt + messages into a single prompt
+    // for the MultiProviderApiClient which uses a completions-style API
+    let response;
+    try {
+      response = await client.invoke({
+        prompt: formatPromptForClient(systemPrompt, messages),
+      });
+    } catch (err) {
+      yield emit("llm_error", { error: err.message, turn: turnCount });
+      break;
+    }
+
+    const responseText = response.text || "";
+    const responseTokens = estimateTokens(responseText, { provider: "anthropic" });
+    ctx.usage.outputTokens += responseTokens;
+    ctx.usage.costUsd += (responseTokens / 1_000_000) * 15;
+
+    yield emit("reasoning", {
+      phase: "deep_analysis",
+      turn: turnCount,
+      summary: responseText.slice(0, 200),
+    });
+
+    // Parse tool_use blocks
+    const toolCalls = parseToolUseBlocks(responseText);
+
+    if (toolCalls.length === 0) {
+      // No tools — extract findings from response
+      const parsed = extractJsonFindings(responseText);
+      for (const finding of parsed) {
+        allFindings.push(finding);
+        yield emit("finding", { ...finding });
+        if (blackboard) {
+          try {
+            await blackboard.appendEntry({
+              agentId: JULES_DEFINITION.id,
+              source: "jules-primary",
+              ...finding,
+            });
+          } catch { /* blackboard write failure non-blocking */ }
+        }
+      }
+      messages.push({ role: "assistant", content: responseText });
+      break; // LLM is done
+    }
+
+    // Execute tool calls
+    const results = [];
+    for (const call of toolCalls) {
+      try {
+        const result = await dispatchTool(call.tool, call.input, ctx);
+        results.push({ tool: call.tool, result });
+        yield emit("tool_call", { tool: call.tool, input: sanitizeForEvent(call.input) });
+      } catch (err) {
+        if (err instanceof BudgetExhaustedError) {
+          yield emit("budget_stop", { reason: err.message });
+          break;
+        }
+        results.push({ tool: call.tool, error: err.message });
+      }
+    }
+
+    // Feed results back
+    messages.push({ role: "assistant", content: responseText });
+    messages.push({
+      role: "user",
+      content: results.map(r =>
+        r.error
+          ? `Tool ${r.tool} failed: ${r.error}`
+          : `Tool ${r.tool} result:\n${JSON.stringify(r.result).slice(0, 3000)}`,
+      ).join("\n\n") + "\n\nContinue your analysis. If done, return findings in a ```json code block.",
+    });
+  }
+
+  // ── Phase 2b: Reconciliation (post-blind-pass) ─────────────────────
+  // Now that the independent analysis is complete, cross-reference with
+  // swarm findings and Omar baseline. This preserves blind-first: the
+  // persona formed its own opinion before seeing prior conclusions.
+
+  const hasSwarmContext = swarmFindings.length > 0;
+  const baselineFindings = omarBaseline
+    ? (omarBaseline.findings || omarBaseline.summary || [])
+    : [];
+  const hasBaselineContext = Array.isArray(baselineFindings) && baselineFindings.length > 0;
+
+  if (hasSwarmContext || hasBaselineContext) {
+    yield emit("progress", { phase: "reconciliation", message: "Cross-referencing with sub-agent and baseline findings..." });
+
+    const reconcileParts = [];
+    reconcileParts.push("Your independent analysis is complete. Now cross-reference with the following prior findings.");
+    reconcileParts.push("For each prior finding: confirm if your analysis agrees, dispute with evidence if you disagree, or flag as missed if you did not cover it.");
+
+    if (hasSwarmContext) {
+      reconcileParts.push(`\nYour sub-agents found ${swarmFindings.length} findings:`);
+      for (const f of swarmFindings.slice(0, 30)) {
+        reconcileParts.push(`- [${f.severity || "P3"}] ${f.file || ""}:${f.line || ""} ${f.title || f.type || ""}`);
+      }
+    }
+
+    if (hasBaselineContext) {
+      reconcileParts.push(`\nOmar baseline reported ${baselineFindings.length} findings:`);
+      for (const f of baselineFindings.slice(0, 20)) {
+        reconcileParts.push(`- [${f.severity || ""}] ${f.file || ""}:${f.line || ""} ${f.message || f.title || ""}`);
+      }
+    }
+
+    reconcileParts.push("\nReturn any additional or revised findings as a JSON array in a ```json code block. If no changes, return an empty array [].");
+
+    messages.push({ role: "user", content: reconcileParts.join("\n") });
+
+    // Budget check before reconciliation turn
+    const reconcilePreCheck = evaluateBudget({
+      sessionSummary: {
+        costUsd: ctx.usage.costUsd,
+        outputTokens: ctx.usage.outputTokens,
+        durationMs: Date.now() - startedAt,
+        toolCalls: ctx.usage.toolCalls,
+      },
+      ...budget,
+    });
+
+    if (!reconcilePreCheck.blocking) {
+      try {
+        const reconcileResponse = await client.invoke({
+          prompt: formatPromptForClient(systemPrompt, messages),
+        });
+
+        const reconcileText = reconcileResponse.text || "";
+        const reconcileTokens = estimateTokens(reconcileText, { provider: "anthropic" });
+        ctx.usage.outputTokens += reconcileTokens;
+        ctx.usage.costUsd += (reconcileTokens / 1_000_000) * 15;
+
+        yield emit("reasoning", { phase: "reconciliation", summary: reconcileText.slice(0, 200) });
+
+        const reconcileFindings = extractJsonFindings(reconcileText);
+        for (const finding of reconcileFindings) {
+          allFindings.push(finding);
+          yield emit("finding", { ...finding, source: "reconciliation" });
+          if (blackboard) {
+            try {
+              await blackboard.appendEntry({
+                agentId: JULES_DEFINITION.id,
+                source: "jules-reconciliation",
+                ...finding,
+              });
+            } catch { /* blackboard write failure non-blocking */ }
+          }
+        }
+
+        messages.push({ role: "assistant", content: reconcileText });
+      } catch (err) {
+        yield emit("llm_error", { error: err.message, phase: "reconciliation" });
+      }
+    } else {
+      yield emit("budget_stop", { reasons: reconcilePreCheck.reasons, phase: "reconciliation" });
+    }
+  }
+
+  // ── Phase 3: Build final report ───────────────────────────────────
+
+  const durationMs = Date.now() - startedAt;
+  const severityCounts = { P0: 0, P1: 0, P2: 0, P3: 0 };
+  for (const f of allFindings) {
+    const sev = (f.severity || "P3").toUpperCase();
+    if (severityCounts[sev] !== undefined) severityCounts[sev]++;
+    else severityCounts.P3++;
+  }
+
+  const report = {
+    runId,
+    persona: JULES_DEFINITION.persona,
+    mode,
+    framework: framework.framework || "unknown",
+    status: "completed",
+    findings: allFindings,
+    summary: {
+      total: allFindings.length,
+      ...severityCounts,
+      blocking: severityCounts.P0 > 0 || severityCounts.P1 > 0,
+    },
+    usage: {
+      turns: turnCount,
+      costUsd: ctx.usage.costUsd,
+      outputTokens: ctx.usage.outputTokens,
+      toolCalls: ctx.usage.toolCalls,
+      durationMs,
+    },
+    signature: JULES_DEFINITION.signature,
+  };
+
+  yield emit("agent_complete", {
+    ...report.summary,
+    costUsd: ctx.usage.costUsd,
+    durationMs,
+    turns: turnCount,
+  });
+
+  return report;
+}
+
+// ── Helpers ──────────────────────────────────────────────────────────
+
+function parseToolUseBlocks(text) {
+  const calls = [];
+  const regex = /```tool_use\s*\n([\s\S]*?)```/g;
+  let match;
+  while ((match = regex.exec(text)) !== null) {
+    try {
+      const parsed = JSON.parse(match[1].trim());
+      if (parsed.tool && parsed.input) calls.push(parsed);
+    } catch { /* skip malformed */ }
+  }
+  return calls;
+}
+
+function extractJsonFindings(text) {
+  const jsonMatch = text.match(/```json\s*\n([\s\S]*?)```/);
+  if (!jsonMatch) return [];
+  try {
+    const parsed = JSON.parse(jsonMatch[1].trim());
+    if (Array.isArray(parsed)) return parsed;
+    if (parsed.findings && Array.isArray(parsed.findings)) return parsed.findings;
+  } catch { /* skip malformed */ }
+  return [];
+}
+
+function sanitizeForEvent(input) {
+  const sanitized = { ...input };
+  if (typeof sanitized.content === "string" && sanitized.content.length > 200) {
+    sanitized.content = `[${sanitized.content.length} chars]`;
+  }
+  return sanitized;
+}
+
+/**
+ * Format system prompt + chat messages into a single prompt string
+ * for MultiProviderApiClient which uses a completions-style API.
+ */
+function formatPromptForClient(systemPrompt, messages) {
+  const parts = [];
+  if (systemPrompt) parts.push(systemPrompt);
+  for (const msg of messages) {
+    const role = msg.role === "assistant" ? "ASSISTANT" : "USER";
+    parts.push(`\n${role}:\n${msg.content}`);
+  }
+  return parts.join("\n");
+}