jiva-core 0.3.42-dev.c44c725 → 0.3.43-dev.00a530a

package/dist/evaluator/evaluator-agent.d.ts

@@ -0,0 +1,35 @@
+/**
+ * EvaluatorAgent — autonomous supervisor for Jiva tasks.
+ *
+ * Architecture mirrors CodeAgent's single tool-calling loop, but instead of
+ * writing code the evaluator reads workspace files, validates completion against
+ * the directive, and uses interact_with_agent to nudge the main agent when gaps
+ * are found.
+ *
+ * The evaluator has completely isolated LLM state — its own ModelOrchestrator,
+ * its own MCPServerManager, and its own message history. It never shares context
+ * with the agent it supervises.
+ */
+import type { EvaluatorConfig, EvaluationContext, EvaluationResult } from './types.js';
+export declare class EvaluatorAgent {
+    private orchestrator;
+    private mcpManager;
+    private workspace;
+    private targetAgent;
+    private maxIterations;
+    private maxCycles;
+    private virtualTools;
+    constructor(config: EvaluatorConfig);
+    /** Return token usage accumulated by the evaluator's isolated orchestrator. */
+    getOrchestratorTokenUsage(): import("../models/token-tracker.js").TokenUsageSnapshot;
+    /**
+     * Evaluate whether the main agent has completed the goal described in userMessage.
+     *
+     * The evaluator runs its own agentic loop: it reads files, assesses completion,
+     * sends nudges to the main agent via interact_with_agent, and re-checks until
+     * satisfied or the cycle limit is reached.
+     */
+    evaluate(ctx: EvaluationContext): Promise<EvaluationResult>;
+    private runEvaluationCycle;
+}
+//# sourceMappingURL=evaluator-agent.d.ts.map

package/dist/evaluator/evaluator-agent.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"evaluator-agent.d.ts","sourceRoot":"","sources":["../../src/evaluator/evaluator-agent.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAUH,OAAO,KAAK,EACV,eAAe,EACf,iBAAiB,EACjB,gBAAgB,EAEjB,MAAM,YAAY,CAAC;AAiHpB,qBAAa,cAAc;IACzB,OAAO,CAAC,YAAY,CAAoB;IACxC,OAAO,CAAC,UAAU,CAAmB;IACrC,OAAO,CAAC,SAAS,CAAmB;IACpC,OAAO,CAAC,WAAW,CAA8C;IACjE,OAAO,CAAC,aAAa,CAAS;IAC9B,OAAO,CAAC,SAAS,CAAS;IAC1B,OAAO,CAAC,YAAY,CAAmB;gBAE3B,MAAM,EAAE,eAAe;IAUnC,+EAA+E;IAC/E,yBAAyB;IAIzB;;;;;;OAMG;IACG,QAAQ,CAAC,GAAG,EAAE,iBAAiB,GAAG,OAAO,CAAC,gBAAgB,CAAC;YAuFnD,kBAAkB;CA0MjC"}

package/dist/evaluator/evaluator-agent.js

@@ -0,0 +1,369 @@
+/**
+ * EvaluatorAgent — autonomous supervisor for Jiva tasks.
+ *
+ * Architecture mirrors CodeAgent's single tool-calling loop, but instead of
+ * writing code the evaluator reads workspace files, validates completion against
+ * the directive, and uses interact_with_agent to nudge the main agent when gaps
+ * are found.
+ *
+ * The evaluator has completely isolated LLM state — its own ModelOrchestrator,
+ * its own MCPServerManager, and its own message history. It never shares context
+ * with the agent it supervises.
+ */
+import { formatToolResult } from '../models/harmony.js';
+import { logger } from '../utils/logger.js';
+import { deriveEvaluatorDirective } from './directive-adapter.js';
+import { EVALUATOR_VIRTUAL_TOOLS } from './tools/agent-tools.js';
+// ─── Constants ────────────────────────────────────────────────────────────────
+const DEFAULT_MAX_ITERATIONS = 30;
+const DEFAULT_MAX_CYCLES = 5;
+const DOOM_LOOP_THRESHOLD = 3;
+// ─── Result parsing ───────────────────────────────────────────────────────────
+/**
+ * Extract a structured EvaluationResult from the evaluator's final LLM response.
+ * Looks for a JSON code block; falls back to a best-effort parse.
+ */
+/** Strip Harmony/XML tool-call markup so it doesn't pollute the summary text. */
+function stripToolCallMarkup(text) {
+    return text
+        .replace(/<tool_call>[\s\S]*?<\/tool_call>/gi, '')
+        .replace(/<\|call\|>[\s\S]*?<\|return\|>/g, '')
+        .replace(/<\|channel\|>\w+/g, '')
+        .trim();
+}
+function parseEvaluationResult(text, nudgesSent, cyclesRan, evidence) {
+    // Strip any tool-call markup before attempting to parse
+    const stripped = stripToolCallMarkup(text);
+    try {
+        // Look for ```json ... ``` block (prefer stripped, fall back to raw)
+        const jsonMatch = stripped.match(/```json\s*([\s\S]+?)\s*```/i) ||
+            text.match(/```json\s*([\s\S]+?)\s*```/i);
+        const raw = jsonMatch ? jsonMatch[1] : stripped;
+        const parsed = JSON.parse(raw.trim());
+        return {
+            passed: Boolean(parsed.passed),
+            gaps: Array.isArray(parsed.gaps) ? parsed.gaps : [],
+            nudgesSent,
+            cyclesRan,
+            evidence,
+            summary: typeof parsed.summary === 'string' ? parsed.summary : stripped.substring(0, 300),
+        };
+    }
+    catch {
+        // Free-form response — determine pass/fail from keywords in the stripped text
+        const lower = stripped.toLowerCase();
+        const passed = lower.includes('evaluation passed') ||
+            lower.includes('all tasks complete') ||
+            lower.includes('work is complete') ||
+            lower.includes('"passed": true');
+        // If the stripped text is empty (was entirely tool calls) or too short to be
+        // a real verdict, treat this as inconclusive rather than a definitive failure.
+        const inconclusive = stripped.length < 20;
+        return {
+            passed: inconclusive ? false : passed,
+            gaps: inconclusive
+                ? ['Evaluation inconclusive — evaluator did not produce a final verdict']
+                : [],
+            nudgesSent,
+            cyclesRan,
+            evidence,
+            summary: inconclusive
+                ? 'The evaluator did not produce a parseable verdict. Review the agent logs for details.'
+                : stripped.substring(0, 300),
+        };
+    }
+}
+// ─── System prompt ────────────────────────────────────────────────────────────
+function buildSystemPrompt(workspaceDir, evaluatorDirective) {
+    return `You are an autonomous evaluation agent operating in evaluation mode.
+You DO NOT perform tasks — you VALIDATE whether tasks have been completed.
+
+WORKSPACE: ${workspaceDir}
+All relative paths are resolved relative to the workspace directory above.
+Use absolute paths for all file operations.
+
+YOUR EVALUATION APPROACH:
+1. Read workspace files to understand what has been produced.
+2. Compare the actual output against the requirements in your directive.
+3. If gaps exist, send targeted instructions to the main agent using interact_with_agent.
+4. Re-read files after each nudge to confirm corrections were applied.
+5. When you are satisfied (or exhausted your options), output your final JSON assessment.
+
+TOOL SELECTION:
+- To READ files → use the filesystem MCP tools (filesystem__read_file, filesystem__search_files, etc.)
+- To SEND instructions to the main agent → use interact_with_agent
+- To LIST available conversations → use list_agent_conversations
+- To INSPECT conversation history → use get_conversation_history
+
+FINAL OUTPUT — when evaluation is complete, respond with ONLY this JSON (no other text):
+\`\`\`json
+{
+  "passed": true | false,
+  "gaps": ["specific gap 1", "specific gap 2"],
+  "summary": "1-3 sentence assessment."
+}
+\`\`\`
+
+${evaluatorDirective}`;
+}
+// ─── EvaluatorAgent ───────────────────────────────────────────────────────────
+export class EvaluatorAgent {
+    orchestrator;
+    mcpManager;
+    workspace;
+    targetAgent;
+    maxIterations;
+    maxCycles;
+    virtualTools;
+    constructor(config) {
+        this.orchestrator = config.orchestrator;
+        this.mcpManager = config.mcpManager;
+        this.workspace = config.workspace;
+        this.targetAgent = config.targetAgent;
+        this.maxIterations = config.maxIterationsPerCycle ?? DEFAULT_MAX_ITERATIONS;
+        this.maxCycles = config.maxEvaluationCycles ?? DEFAULT_MAX_CYCLES;
+        this.virtualTools = EVALUATOR_VIRTUAL_TOOLS;
+    }
+    /** Return token usage accumulated by the evaluator's isolated orchestrator. */
+    getOrchestratorTokenUsage() {
+        return this.orchestrator.getTokenUsage();
+    }
+    /**
+     * Evaluate whether the main agent has completed the goal described in userMessage.
+     *
+     * The evaluator runs its own agentic loop: it reads files, assesses completion,
+     * sends nudges to the main agent via interact_with_agent, and re-checks until
+     * satisfied or the cycle limit is reached.
+     */
+    async evaluate(ctx) {
+        logger.info('[Evaluator] Starting evaluation');
+        // Optionally load a specific conversation into the target agent
+        if (ctx.targetConversationId) {
+            logger.info(`[Evaluator] Loading conversation: ${ctx.targetConversationId}`);
+            await this.targetAgent.loadConversation(ctx.targetConversationId);
+        }
+        let nudgesSent = 0;
+        let cyclesRan = 0;
+        const evidence = [];
+        // The directive for the evaluator — wraps the workspace directive with evaluation framing
+        const rawDirective = this.workspace.getDirectivePrompt() ?? '';
+        const evaluatorDirective = deriveEvaluatorDirective(rawDirective);
+        const workspaceDir = this.workspace.getWorkspaceDir();
+        const systemPrompt = buildSystemPrompt(workspaceDir, evaluatorDirective);
+        // Build the unified tool list: MCP tools + virtual tools
+        const mcpToolDefs = this.mcpManager
+            .getClient()
+            .getAllTools()
+            .map((t) => ({
+            name: t.name,
+            description: t.description,
+            parameters: t.inputSchema ?? { type: 'object', properties: {} },
+        }));
+        const virtualToolDefs = this.virtualTools.map((t) => ({
+            name: t.name,
+            description: t.description,
+            parameters: t.parameters,
+        }));
+        const allToolDefs = [...mcpToolDefs, ...virtualToolDefs];
+        // Evaluation loop — each cycle runs the full agentic loop once.
+        // The evaluator terminates early if it produces a JSON result with "passed": true.
+        let lastResult = null;
+        for (let cycle = 0; cycle < this.maxCycles; cycle++) {
+            cyclesRan = cycle + 1;
+            logger.info(`[Evaluator] Cycle ${cyclesRan}/${this.maxCycles}`);
+            const cycleResult = await this.runEvaluationCycle({
+                systemPrompt,
+                userMessage: ctx.userMessage,
+                allToolDefs,
+                workspaceDir,
+                nudgesSentRef: { value: nudgesSent },
+                evidence,
+            });
+            nudgesSent = cycleResult.nudgesSentAfter;
+            lastResult = parseEvaluationResult(cycleResult.finalText, nudgesSent, cyclesRan, [...evidence]);
+            logger.info(`[Evaluator] Cycle ${cyclesRan} complete — passed: ${lastResult.passed}, gaps: ${lastResult.gaps.length}`);
+            if (lastResult.passed) {
+                break;
+            }
+        }
+        if (!lastResult) {
+            return {
+                passed: false,
+                gaps: ['Evaluation could not be completed — max cycles reached without a result'],
+                nudgesSent,
+                cyclesRan,
+                evidence,
+                summary: 'Evaluation did not produce a result within the allowed cycle budget.',
+            };
+        }
+        return lastResult;
+    }
+    // ─── Private: single cycle ──────────────────────────────────────────────────
+    async runEvaluationCycle(params) {
+        const { systemPrompt, userMessage, allToolDefs, workspaceDir, nudgesSentRef, evidence } = params;
+        // Fresh isolated message history for each cycle
+        const messages = [
+            { role: 'developer', content: systemPrompt },
+            {
+                role: 'user',
+                content: `Evaluate whether the following goal has been fully and correctly completed:\n\n` +
+                    `GOAL: ${userMessage}\n\n` +
+                    `Start by reading the relevant workspace files, then assess completion against the ` +
+                    `original directive. If gaps exist, use interact_with_agent to guide the main agent. ` +
+                    `When done, output your JSON assessment.`,
+            },
+        ];
+        const toolContext = {
+            workspaceDir,
+            targetAgent: this.targetAgent,
+            onNudgeSent: () => {
+                nudgesSentRef.value++;
+            },
+            onEvidenceFound: (path) => {
+                if (!evidence.includes(path))
+                    evidence.push(path);
+            },
+        };
+        // Doom loop detection
+        const recentCalls = [];
+        let consecutiveApiErrors = 0;
+        const MAX_API_ERRORS = 3;
+        let finalText = '';
+        let iterations = 0;
+        for (let i = 0; i < this.maxIterations; i++) {
+            iterations = i + 1;
+            const iterPct = i / this.maxIterations;
+            const isFinalPhase = iterPct >= 0.90;
+            // Near-limit nudge
+            if (iterPct >= 0.80 && iterPct < 0.90 && i > 0) {
+                const alreadyNudged = messages.some((m) => typeof m.content === 'string' && m.content.includes('FINAL PHASE'));
+                if (!alreadyNudged) {
+                    messages.push({
+                        role: 'user',
+                        content: `You are approaching the iteration limit (${iterations}/${this.maxIterations}). ` +
+                            `Finish any remaining checks and produce your final JSON assessment.`,
+                    });
+                }
+            }
+            if (isFinalPhase) {
+                const alreadyStripped = messages.some((m) => typeof m.content === 'string' && m.content.includes('FINAL PHASE'));
+                if (!alreadyStripped) {
+                    messages.push({
+                        role: 'user',
+                        content: 'FINAL PHASE: Tools are disabled. You MUST respond with ONLY the JSON assessment block now.\n\n' +
+                            '```json\n{"passed": true|false, "gaps": [], "summary": "..."}\n```',
+                    });
+                }
+            }
+            let response;
+            try {
+                response = await this.orchestrator.chatWithFallback({
+                    messages,
+                    tools: isFinalPhase ? [] : allToolDefs,
+                    temperature: 0.1,
+                }, false);
+                consecutiveApiErrors = 0;
+            }
+            catch (error) {
+                consecutiveApiErrors++;
+                const msg = error instanceof Error ? error.message : String(error);
+                logger.error(`[Evaluator] Model error (${consecutiveApiErrors}/${MAX_API_ERRORS}): ${msg}`);
+                if (consecutiveApiErrors >= MAX_API_ERRORS) {
+                    finalText = '[Evaluation aborted due to repeated model errors]';
+                    break;
+                }
+                // Add placeholder so conversation structure remains valid
+                const lastMsg = messages[messages.length - 1];
+                if (lastMsg && lastMsg.role !== 'assistant') {
+                    messages.push({ role: 'assistant', content: '' });
+                }
+                messages.push({
+                    role: 'user',
+                    content: 'The previous request failed. Please continue the evaluation and output your JSON assessment.',
+                });
+                continue;
+            }
+            // Record assistant message
+            if (response.toolCalls && response.toolCalls.length > 0) {
+                messages.push({
+                    role: 'assistant',
+                    content: response.content || null,
+                    tool_calls: response.toolCalls,
+                });
+            }
+            else {
+                messages.push({ role: 'assistant', content: response.content });
+            }
+            // No tool calls → evaluator is done
+            if (!response.toolCalls || response.toolCalls.length === 0) {
+                finalText = response.content || '[No evaluation response]';
+                break;
+            }
+            // Execute tool calls
+            for (const toolCall of response.toolCalls) {
+                const toolName = toolCall.function.name;
+                let toolArgs = {};
+                try {
+                    toolArgs = JSON.parse(toolCall.function.arguments);
+                }
+                catch {
+                    // malformed args — continue with empty args
+                }
+                // Doom loop guard
+                const callSig = `${toolName}:${JSON.stringify(toolArgs)}`;
+                recentCalls.push(callSig);
+                if (recentCalls.length > DOOM_LOOP_THRESHOLD)
+                    recentCalls.shift();
+                if (recentCalls.length === DOOM_LOOP_THRESHOLD &&
+                    recentCalls.every((c) => c === recentCalls[0])) {
+                    logger.warn(`[Evaluator] Doom loop detected for tool: ${toolName}`);
+                    messages.push({
+                        role: 'user',
+                        content: `STOP: You are calling \`${toolName}\` with the same arguments repeatedly. ` +
+                            `Move on and output your final JSON assessment based on what you have found so far.`,
+                    });
+                    break;
+                }
+                logger.info(`[Evaluator] Tool: ${toolName}`);
+                let toolResult;
+                // Check virtual tools first, then fall through to MCP
+                const virtualTool = this.virtualTools.find((t) => t.name === toolName);
+                if (virtualTool) {
+                    try {
+                        toolResult = await virtualTool.execute(toolArgs, toolContext);
+                    }
+                    catch (e) {
+                        toolResult = `Error executing ${toolName}: ${e instanceof Error ? e.message : String(e)}`;
+                    }
+                }
+                else {
+                    // MCP tool
+                    try {
+                        const result = await this.mcpManager.getClient().executeTool(toolName, toolArgs);
+                        toolResult = typeof result === 'string' ? result : JSON.stringify(result);
+                        // Track filesystem reads as evidence
+                        if (toolName.includes('read') ||
+                            toolName.includes('search') ||
+                            toolName.includes('list')) {
+                            const pathArg = toolArgs.path ||
+                                toolArgs.query ||
+                                toolArgs.pattern;
+                            if (pathArg)
+                                toolContext.onEvidenceFound(pathArg);
+                        }
+                    }
+                    catch (e) {
+                        toolResult = `Error executing ${toolName}: ${e instanceof Error ? e.message : String(e)}`;
+                    }
+                }
+                const toolMessage = formatToolResult(toolCall.id, toolName, toolResult);
+                messages.push(toolMessage);
+            }
+        }
+        if (!finalText) {
+            finalText = `[Evaluator reached max iterations (${iterations}) without producing a final assessment]`;
+        }
+        return { finalText, nudgesSentAfter: nudgesSentRef.value };
+    }
+}
+//# sourceMappingURL=evaluator-agent.js.map

package/dist/evaluator/evaluator-agent.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"evaluator-agent.js","sourceRoot":"","sources":["../../src/evaluator/evaluator-agent.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAKH,OAAO,EAAE,gBAAgB,EAAE,MAAM,sBAAsB,CAAC;AAExD,OAAO,EAAE,MAAM,EAAE,MAAM,oBAAoB,CAAC;AAC5C,OAAO,EAAE,wBAAwB,EAAE,MAAM,wBAAwB,CAAC;AAClE,OAAO,EAAE,uBAAuB,EAAuB,MAAM,wBAAwB,CAAC;AAQtF,iFAAiF;AAEjF,MAAM,sBAAsB,GAAG,EAAE,CAAC;AAClC,MAAM,kBAAkB,GAAG,CAAC,CAAC;AAC7B,MAAM,mBAAmB,GAAG,CAAC,CAAC;AAE9B,iFAAiF;AAEjF;;;GAGG;AACH,iFAAiF;AACjF,SAAS,mBAAmB,CAAC,IAAY;IACvC,OAAO,IAAI;SACR,OAAO,CAAC,oCAAoC,EAAE,EAAE,CAAC;SACjD,OAAO,CAAC,iCAAiC,EAAE,EAAE,CAAC;SAC9C,OAAO,CAAC,mBAAmB,EAAE,EAAE,CAAC;SAChC,IAAI,EAAE,CAAC;AACZ,CAAC;AAED,SAAS,qBAAqB,CAC5B,IAAY,EACZ,UAAkB,EAClB,SAAiB,EACjB,QAAkB;IAElB,wDAAwD;IACxD,MAAM,QAAQ,GAAG,mBAAmB,CAAC,IAAI,CAAC,CAAC;IAE3C,IAAI,CAAC;QACH,qEAAqE;QACrE,MAAM,SAAS,GACb,QAAQ,CAAC,KAAK,CAAC,6BAA6B,CAAC;YAC7C,IAAI,CAAC,KAAK,CAAC,6BAA6B,CAAC,CAAC;QAC5C,MAAM,GAAG,GAAG,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC;QAChD,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,IAAI,EAAE,CAAC,CAAC;QAEtC,OAAO;YACL,MAAM,EAAE,OAAO,CAAC,MAAM,CAAC,MAAM,CAAC;YAC9B,IAAI,EAAE,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE;YACnD,UAAU;YACV,SAAS;YACT,QAAQ;YACR,OAAO,EAAE,OAAO,MAAM,CAAC,OAAO,KAAK,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,CAAC,QAAQ,CAAC,SAAS,CAAC,CAAC,EAAE,GAAG,CAAC;SAC1F,CAAC;IACJ,CAAC;IAAC,MAAM,CAAC;QACP,8EAA8E;QAC9E,MAAM,KAAK,GAAG,QAAQ,CAAC,WAAW,EAAE,CAAC;QACrC,MAAM,MAAM,GACV,KAAK,CAAC,QAAQ,CAAC,mBAAmB,CAAC;YACnC,KAAK,CAAC,QAAQ,CAAC,oBAAoB,CAAC;YACpC,KAAK,CAAC,QAAQ,CAAC,kBAAkB,CAAC;YAClC,KAAK,CAAC,QAAQ,CAAC,gBAAgB,CAAC,CAAC;QAEnC,6EAA6E;QAC7E,+EAA+E;QAC/E,MAAM,YAAY,GAAG,QAAQ,CAAC,MAAM,GAAG,EAAE,CAAC;QAE1C,OAAO;YACL,MAAM,EAAE,YAAY,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,MAAM;YACrC,IAAI,EAAE,YAAY;gBAChB,CAAC,CAAC,CAAC,qEAAqE,CAAC;gBACzE,CAAC,CAAC,EAAE;YACN,UAAU;YACV,SAAS;YACT,QAAQ;YACR,OAAO,EAAE,YAAY;gBACnB,CAAC,CAAC,uFAAuF;gBACzF,CAAC,CAAC,QAAQ,CAAC,SAAS,CAAC,CAAC,EAAE,GAAG,CAAC;SAC/B,CAAC;IACJ,CAAC;AACH,CAAC;AAED,iFAAiF;AAEjF,SAAS,iBAAiB,CAAC,YAAoB,EAAE,kBAA0B;IACzE,OAAO;;;aAGI,YAAY;;;;;;;;;;;;;;;;;;;;;;;;;;EA0BvB,kBAAkB,EAAE,CAAC;AACvB,CAAC;AAED,iFAAiF;AAEjF,MAAM,OAAO,cAAc;IACjB,YAAY,CAAoB;IAChC,UAAU,CAAmB;IAC7B,SAAS,CAAmB;IAC5B,WAAW,CAA8C;IACzD,aAAa,CAAS;IACtB,SAAS,CAAS;IAClB,YAAY,CAAmB;IAEvC,YAAY,MAAuB;QACjC,IAAI,CAAC,YAAY,GAAG,MAAM,CAAC,YAAY,CAAC;QACxC,IAAI,CAAC,UAAU,GAAG,MAAM,CAAC,UAAU,CAAC;QACpC,IAAI,CAAC,SAAS,GAAG,MAAM,CAAC,SAAS,CAAC;QAClC,IAAI,CAAC,WAAW,GAAG,MAAM,CAAC,WAAW,CAAC;QACtC,IAAI,CAAC,aAAa,GAAG,MAAM,CAAC,qBAAqB,IAAI,sBAAsB,CAAC;QAC5E,IAAI,CAAC,SAAS,GAAG,MAAM,CAAC,mBAAmB,IAAI,kBAAkB,CAAC;QAClE,IAAI,CAAC,YAAY,GAAG,uBAAuB,CAAC;IAC9C,CAAC;IAED,+EAA+E;IAC/E,yBAAyB;QACvB,OAAO,IAAI,CAAC,YAAY,CAAC,aAAa,EAAE,CAAC;IAC3C,CAAC;IAED;;;;;;OAMG;IACH,KAAK,CAAC,QAAQ,CAAC,GAAsB;QACnC,MAAM,CAAC,IAAI,CAAC,iCAAiC,CAAC,CAAC;QAE/C,gEAAgE;QAChE,IAAI,GAAG,CAAC,oBAAoB,EAAE,CAAC;YAC7B,MAAM,CAAC,IAAI,CAAC,qCAAqC,GAAG,CAAC,oBAAoB,EAAE,CAAC,CAAC;YAC7E,MAAM,IAAI,CAAC,WAAW,CAAC,gBAAgB,CAAC,GAAG,CAAC,oBAAoB,CAAC,CAAC;QACpE,CAAC;QAED,IAAI,UAAU,GAAG,CAAC,CAAC;QACnB,IAAI,SAAS,GAAG,CAAC,CAAC;QAClB,MAAM,QAAQ,GAAa,EAAE,CAAC;QAE9B,0FAA0F;QAC1F,MAAM,YAAY,GAAG,IAAI,CAAC,SAAS,CAAC,kBAAkB,EAAE,IAAI,EAAE,CAAC;QAC/D,MAAM,kBAAkB,GAAG,wBAAwB,CAAC,YAAY,CAAC,CAAC;QAClE,MAAM,YAAY,GAAG,IAAI,CAAC,SAAS,CAAC,eAAe,EAAE,CAAC;QACtD,MAAM,YAAY,GAAG,iBAAiB,CAAC,YAAY,EAAE,kBAAkB,CAAC,CAAC;QAEzE,yDAAyD;QACzD,MAAM,WAAW,GAAG,IAAI,CAAC,UAAU;aAChC,SAAS,EAAE;aACX,WAAW,EAAE;aACb,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;YACX,IAAI,EAAE,CAAC,CAAC,IAAI;YACZ,WAAW,EAAE,CAAC,CAAC,WAAW;YAC1B,UAAU,EAAG,CAAS,CAAC,WAAW,IAAI,EAAE,IAAI,EAAE,QAAiB,EAAE,UAAU,EAAE,EAAE,EAAE;SAClF,CAAC,CAAW,CAAC;QAEhB,MAAM,eAAe,GAAW,IAAI,CAAC,YAAY,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;YAC5D,IAAI,EAAE,CAAC,CAAC,IAAI;YACZ,WAAW,EAAE,CAAC,CAAC,WAAW;YAC1B,UAAU,EAAE,CAAC,CAAC,UAAU;SACzB,CAAC,CAAC,CAAC;QAEJ,MAAM,WAAW,GAAW,CAAC,GAAG,WAAW,EAAE,GAAG,eAAe,CAAC,CAAC;QAEjE,gEAAgE;QAChE,mFAAmF;QACnF,IAAI,UAAU,GAA4B,IAAI,CAAC;QAE/C,KAAK,IAAI,KAAK,GAAG,CAAC,EAAE,KAAK,GAAG,IAAI,CAAC,SAAS,EAAE,KAAK,EAAE,EAAE,CAAC;YACpD,SAAS,GAAG,KAAK,GAAG,CAAC,CAAC;YACtB,MAAM,CAAC,IAAI,CAAC,qBAAqB,SAAS,IAAI,IAAI,CAAC,SAAS,EAAE,CAAC,CAAC;YAEhE,MAAM,WAAW,GAAG,MAAM,IAAI,CAAC,kBAAkB,CAAC;gBAChD,YAAY;gBACZ,WAAW,EAAE,GAAG,CAAC,WAAW;gBAC5B,WAAW;gBACX,YAAY;gBACZ,aAAa,EAAE,EAAE,KAAK,EAAE,UAAU,EAAE;gBACpC,QAAQ;aACT,CAAC,CAAC;YAEH,UAAU,GAAG,WAAW,CAAC,eAAe,CAAC;YACzC,UAAU,GAAG,qBAAqB,CAChC,WAAW,CAAC,SAAS,EACrB,UAAU,EACV,SAAS,EACT,CAAC,GAAG,QAAQ,CAAC,CACd,CAAC;YAEF,MAAM,CAAC,IAAI,CACT,qBAAqB,SAAS,uBAAuB,UAAU,CAAC,MAAM,WAAW,UAAU,CAAC,IAAI,CAAC,MAAM,EAAE,CAC1G,CAAC;YAEF,IAAI,UAAU,CAAC,MAAM,EAAE,CAAC;gBACtB,MAAM;YACR,CAAC;QACH,CAAC;QAED,IAAI,CAAC,UAAU,EAAE,CAAC;YAChB,OAAO;gBACL,MAAM,EAAE,KAAK;gBACb,IAAI,EAAE,CAAC,yEAAyE,CAAC;gBACjF,UAAU;gBACV,SAAS;gBACT,QAAQ;gBACR,OAAO,EAAE,sEAAsE;aAChF,CAAC;QACJ,CAAC;QAED,OAAO,UAAU,CAAC;IACpB,CAAC;IAED,+EAA+E;IAEvE,KAAK,CAAC,kBAAkB,CAAC,MAOhC;QACC,MAAM,EAAE,YAAY,EAAE,WAAW,EAAE,WAAW,EAAE,YAAY,EAAE,aAAa,EAAE,QAAQ,EAAE,GACrF,MAAM,CAAC;QAET,gDAAgD;QAChD,MAAM,QAAQ,GAAc;YAC1B,EAAE,IAAI,EAAE,WAAkB,EAAE,OAAO,EAAE,YAAY,EAAE;YACnD;gBACE,IAAI,EAAE,MAAM;gBACZ,OAAO,EACL,iFAAiF;oBACjF,SAAS,WAAW,MAAM;oBAC1B,oFAAoF;oBACpF,sFAAsF;oBACtF,yCAAyC;aAC5C;SACF,CAAC;QAEF,MAAM,WAAW,GAAyB;YACxC,YAAY;YACZ,WAAW,EAAE,IAAI,CAAC,WAAW;YAC7B,WAAW,EAAE,GAAG,EAAE;gBAChB,aAAa,CAAC,KAAK,EAAE,CAAC;YACxB,CAAC;YACD,eAAe,EAAE,CAAC,IAAY,EAAE,EAAE;gBAChC,IAAI,CAAC,QAAQ,CAAC,QAAQ,CAAC,IAAI,CAAC;oBAAE,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;YACpD,CAAC;SACF,CAAC;QAEF,sBAAsB;QACtB,MAAM,WAAW,GAAa,EAAE,CAAC;QACjC,IAAI,oBAAoB,GAAG,CAAC,CAAC;QAC7B,MAAM,cAAc,GAAG,CAAC,CAAC;QACzB,IAAI,SAAS,GAAG,EAAE,CAAC;QACnB,IAAI,UAAU,GAAG,CAAC,CAAC;QAEnB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,aAAa,EAAE,CAAC,EAAE,EAAE,CAAC;YAC5C,UAAU,GAAG,CAAC,GAAG,CAAC,CAAC;YACnB,MAAM,OAAO,GAAG,CAAC,GAAG,IAAI,CAAC,aAAa,CAAC;YACvC,MAAM,YAAY,GAAG,OAAO,IAAI,IAAI,CAAC;YAErC,mBAAmB;YACnB,IAAI,OAAO,IAAI,IAAI,IAAI,OAAO,GAAG,IAAI,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC;gBAC/C,MAAM,aAAa,GAAG,QAAQ,CAAC,IAAI,CACjC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,CAAC,OAAO,KAAK,QAAQ,IAAI,CAAC,CAAC,OAAO,CAAC,QAAQ,CAAC,aAAa,CAAC,CAC1E,CAAC;gBACF,IAAI,CAAC,aAAa,EAAE,CAAC;oBACnB,QAAQ,CAAC,IAAI,CAAC;wBACZ,IAAI,EAAE,MAAM;wBACZ,OAAO,EACL,4CAA4C,UAAU,IAAI,IAAI,CAAC,aAAa,KAAK;4BACjF,qEAAqE;qBACxE,CAAC,CAAC;gBACL,CAAC;YACH,CAAC;YAED,IAAI,YAAY,EAAE,CAAC;gBACjB,MAAM,eAAe,GAAG,QAAQ,CAAC,IAAI,CACnC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,CAAC,OAAO,KAAK,QAAQ,IAAI,CAAC,CAAC,OAAO,CAAC,QAAQ,CAAC,aAAa,CAAC,CAC1E,CAAC;gBACF,IAAI,CAAC,eAAe,EAAE,CAAC;oBACrB,QAAQ,CAAC,IAAI,CAAC;wBACZ,IAAI,EAAE,MAAM;wBACZ,OAAO,EACL,gGAAgG;4BAChG,oEAAoE;qBACvE,CAAC,CAAC;gBACL,CAAC;YACH,CAAC;YAED,IAAI,QAAQ,CAAC;YACb,IAAI,CAAC;gBACH,QAAQ,GAAG,MAAM,IAAI,CAAC,YAAY,CAAC,gBAAgB,CACjD;oBACE,QAAQ;oBACR,KAAK,EAAE,YAAY,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,WAAW;oBACtC,WAAW,EAAE,GAAG;iBACjB,EACD,KAAK,CACN,CAAC;gBACF,oBAAoB,GAAG,CAAC,CAAC;YAC3B,CAAC;YAAC,OAAO,KAAK,EAAE,CAAC;gBACf,oBAAoB,EAAE,CAAC;gBACvB,MAAM,GAAG,GAAG,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC;gBACnE,MAAM,CAAC,KAAK,CAAC,4BAA4B,oBAAoB,IAAI,cAAc,MAAM,GAAG,EAAE,CAAC,CAAC;gBAE5F,IAAI,oBAAoB,IAAI,cAAc,EAAE,CAAC;oBAC3C,SAAS,GAAG,mDAAmD,CAAC;oBAChE,MAAM;gBACR,CAAC;gBAED,0DAA0D;gBAC1D,MAAM,OAAO,GAAG,QAAQ,CAAC,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;gBAC9C,IAAI,OAAO,IAAI,OAAO,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;oBAC5C,QAAQ,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,OAAO,EAAE,EAAE,EAAE,CAAC,CAAC;gBACpD,CAAC;gBACD,QAAQ,CAAC,IAAI,CAAC;oBACZ,IAAI,EAAE,MAAM;oBACZ,OAAO,EAAE,8FAA8F;iBACxG,CAAC,CAAC;gBACH,SAAS;YACX,CAAC;YAED,2BAA2B;YAC3B,IAAI,QAAQ,CAAC,SAAS,IAAI,QAAQ,CAAC,SAAS,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;gBACxD,QAAQ,CAAC,IAAI,CAAC;oBACZ,IAAI,EAAE,WAAW;oBACjB,OAAO,EAAE,QAAQ,CAAC,OAAO,IAAI,IAAI;oBACjC,UAAU,EAAE,QAAQ,CAAC,SAAS;iBACxB,CAAC,CAAC;YACZ,CAAC;iBAAM,CAAC;gBACN,QAAQ,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,OAAO,EAAE,QAAQ,CAAC,OAAO,EAAE,CAAC,CAAC;YAClE,CAAC;YAED,oCAAoC;YACpC,IAAI,CAAC,QAAQ,CAAC,SAAS,IAAI,QAAQ,CAAC,SAAS,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;gBAC3D,SAAS,GAAG,QAAQ,CAAC,OAAO,IAAI,0BAA0B,CAAC;gBAC3D,MAAM;YACR,CAAC;YAED,qBAAqB;YACrB,KAAK,MAAM,QAAQ,IAAI,QAAQ,CAAC,SAAS,EAAE,CAAC;gBAC1C,MAAM,QAAQ,GAAG,QAAQ,CAAC,QAAQ,CAAC,IAAI,CAAC;gBACxC,IAAI,QAAQ,GAA4B,EAAE,CAAC;gBAC3C,IAAI,CAAC;oBACH,QAAQ,GAAG,IAAI,CAAC,KAAK,CAAC,QAAQ,CAAC,QAAQ,CAAC,SAAS,CAAC,CAAC;gBACrD,CAAC;gBAAC,MAAM,CAAC;oBACP,4CAA4C;gBAC9C,CAAC;gBAED,kBAAkB;gBAClB,MAAM,OAAO,GAAG,GAAG,QAAQ,IAAI,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,EAAE,CAAC;gBAC1D,WAAW,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC;gBAC1B,IAAI,WAAW,CAAC,MAAM,GAAG,mBAAmB;oBAAE,WAAW,CAAC,KAAK,EAAE,CAAC;gBAElE,IACE,WAAW,CAAC,MAAM,KAAK,mBAAmB;oBAC1C,WAAW,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,KAAK,WAAW,CAAC,CAAC,CAAC,CAAC,EAC9C,CAAC;oBACD,MAAM,CAAC,IAAI,CAAC,4CAA4C,QAAQ,EAAE,CAAC,CAAC;oBACpE,QAAQ,CAAC,IAAI,CAAC;wBACZ,IAAI,EAAE,MAAM;wBACZ,OAAO,EAAE,2BAA2B,QAAQ,yCAAyC;4BACnF,oFAAoF;qBACvF,CAAC,CAAC;oBACH,MAAM;gBACR,CAAC;gBAED,MAAM,CAAC,IAAI,CAAC,qBAAqB,QAAQ,EAAE,CAAC,CAAC;gBAE7C,IAAI,UAAkB,CAAC;gBAEvB,sDAAsD;gBACtD,MAAM,WAAW,GAAG,IAAI,CAAC,YAAY,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,QAAQ,CAAC,CAAC;gBAEvE,IAAI,WAAW,EAAE,CAAC;oBAChB,IAAI,CAAC;wBACH,UAAU,GAAG,MAAM,WAAW,CAAC,OAAO,CAAC,QAAQ,EAAE,WAAW,CAAC,CAAC;oBAChE,CAAC;oBAAC,OAAO,CAAC,EAAE,CAAC;wBACX,UAAU,GAAG,mBAAmB,QAAQ,KAAK,CAAC,YAAY,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC;oBAC5F,CAAC;gBACH,CAAC;qBAAM,CAAC;oBACN,WAAW;oBACX,IAAI,CAAC;wBACH,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,UAAU,CAAC,SAAS,EAAE,CAAC,WAAW,CAAC,QAAQ,EAAE,QAAQ,CAAC,CAAC;wBACjF,UAAU,GAAG,OAAO,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,CAAC;wBAE1E,qCAAqC;wBACrC,IACE,QAAQ,CAAC,QAAQ,CAAC,MAAM,CAAC;4BACzB,QAAQ,CAAC,QAAQ,CAAC,QAAQ,CAAC;4BAC3B,QAAQ,CAAC,QAAQ,CAAC,MAAM,CAAC,EACzB,CAAC;4BACD,MAAM,OAAO,GACV,QAAQ,CAAC,IAAe;gCACxB,QAAQ,CAAC,KAAgB;gCACzB,QAAQ,CAAC,OAAkB,CAAC;4BAC/B,IAAI,OAAO;gCAAE,WAAW,CAAC,eAAe,CAAC,OAAO,CAAC,CAAC;wBACpD,CAAC;oBACH,CAAC;oBAAC,OAAO,CAAC,EAAE,CAAC;wBACX,UAAU,GAAG,mBAAmB,QAAQ,KAAK,CAAC,YAAY,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC;oBAC5F,CAAC;gBACH,CAAC;gBAED,MAAM,WAAW,GAAG,gBAAgB,CAAC,QAAQ,CAAC,EAAE,EAAE,QAAQ,EAAE,UAAU,CAAC,CAAC;gBACxE,QAAQ,CAAC,IAAI,CAAC,WAAW,CAAC,CAAC;YAC7B,CAAC;QACH,CAAC;QAED,IAAI,CAAC,SAAS,EAAE,CAAC;YACf,SAAS,GAAG,sCAAsC,UAAU,yCAAyC,CAAC;QACxG,CAAC;QAED,OAAO,EAAE,SAAS,EAAE,eAAe,EAAE,aAAa,CAAC,KAAK,EAAE,CAAC;IAC7D,CAAC;CACF"}

package/dist/evaluator/harness.d.ts

@@ -0,0 +1,31 @@
+/**
+ * EvaluatorHarness — coordinates the main Jiva agent and the evaluator.
+ *
+ * Flow per user message:
+ *   1. Run the main agent (mainAgent.chat(userMessage))
+ *   2. Run the evaluator (evaluatorAgent.evaluate({ userMessage }))
+ *      — the evaluator reads workspace files, identifies gaps, and nudges the
+ *        main agent as needed via its interact_with_agent tool
+ *   3. Return a HarnessResult containing both the main response and the
+ *      evaluation outcome
+ */
+import { EvaluatorAgent } from './evaluator-agent.js';
+import type { IAgent } from '../core/agent-interface.js';
+import type { EvaluationContext, HarnessOptions, HarnessResult } from './types.js';
+export declare class EvaluatorHarness {
+    private mainAgent;
+    private evaluatorAgent;
+    private verbose;
+    constructor(mainAgent: IAgent, evaluatorAgent: EvaluatorAgent, options?: HarnessOptions);
+    /**
+     * Process a user message through the harness:
+     *   1. Main agent processes the request.
+     *   2. Evaluator validates and guides completion.
+     */
+    run(userMessage: string, evalCtx?: Partial<EvaluationContext>): Promise<HarnessResult>;
+    /** Stop both agents cooperatively. */
+    stop(): void;
+    /** Cleanup both agents' resources. */
+    cleanup(): Promise<void>;
+}
+//# sourceMappingURL=harness.d.ts.map

package/dist/evaluator/harness.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"harness.d.ts","sourceRoot":"","sources":["../../src/evaluator/harness.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAGH,OAAO,EAAE,cAAc,EAAE,MAAM,sBAAsB,CAAC;AACtD,OAAO,KAAK,EAAE,MAAM,EAAE,MAAM,4BAA4B,CAAC;AACzD,OAAO,KAAK,EAAE,iBAAiB,EAAE,cAAc,EAAE,aAAa,EAAE,MAAM,YAAY,CAAC;AAEnF,qBAAa,gBAAgB;IAC3B,OAAO,CAAC,SAAS,CAAS;IAC1B,OAAO,CAAC,cAAc,CAAiB;IACvC,OAAO,CAAC,OAAO,CAAU;gBAGvB,SAAS,EAAE,MAAM,EACjB,cAAc,EAAE,cAAc,EAC9B,OAAO,GAAE,cAAmB;IAO9B;;;;OAIG;IACG,GAAG,CAAC,WAAW,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,OAAO,CAAC,iBAAiB,CAAC,GAAG,OAAO,CAAC,aAAa,CAAC;IA2C5F,sCAAsC;IACtC,IAAI,IAAI,IAAI;IAIZ,sCAAsC;IAChC,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC;CAI/B"}

package/dist/evaluator/harness.js

@@ -0,0 +1,70 @@
+/**
+ * EvaluatorHarness — coordinates the main Jiva agent and the evaluator.
+ *
+ * Flow per user message:
+ *   1. Run the main agent (mainAgent.chat(userMessage))
+ *   2. Run the evaluator (evaluatorAgent.evaluate({ userMessage }))
+ *      — the evaluator reads workspace files, identifies gaps, and nudges the
+ *        main agent as needed via its interact_with_agent tool
+ *   3. Return a HarnessResult containing both the main response and the
+ *      evaluation outcome
+ */
+import { logger } from '../utils/logger.js';
+export class EvaluatorHarness {
+    mainAgent;
+    evaluatorAgent;
+    verbose;
+    constructor(mainAgent, evaluatorAgent, options = {}) {
+        this.mainAgent = mainAgent;
+        this.evaluatorAgent = evaluatorAgent;
+        this.verbose = options.verbose ?? true;
+    }
+    /**
+     * Process a user message through the harness:
+     *   1. Main agent processes the request.
+     *   2. Evaluator validates and guides completion.
+     */
+    async run(userMessage, evalCtx) {
+        // ── Step 1: Main agent ─────────────────────────────────────────────────
+        if (this.verbose) {
+            logger.info('[Harness] Main agent processing user request');
+        }
+        const mainResponse = await this.mainAgent.chat(userMessage);
+        if (this.verbose) {
+            logger.info(`[Harness] Main agent done — ${mainResponse.iterations} iterations, ` +
+                `${mainResponse.toolsUsed.length} tools used`);
+        }
+        // ── Step 2: Evaluator ──────────────────────────────────────────────────
+        if (this.verbose) {
+            logger.info('[Harness] Evaluator starting validation');
+        }
+        const evaluationContext = {
+            userMessage,
+            targetConversationId: evalCtx?.targetConversationId,
+        };
+        const evaluation = await this.evaluatorAgent.evaluate(evaluationContext);
+        if (this.verbose) {
+            const outcome = evaluation.passed
+                ? `✓ Passed (${evaluation.cyclesRan} cycles, ${evaluation.nudgesSent} nudges)`
+                : `✗ Failed — ${evaluation.gaps.length} gap(s) remain after ${evaluation.cyclesRan} cycles`;
+            logger.info(`[Harness] Evaluation complete — ${outcome}`);
+        }
+        return {
+            mainAgentResponse: mainResponse.content,
+            mainAgentIterations: mainResponse.iterations,
+            evaluation,
+            mainAgentTokenUsage: this.mainAgent.getTokenUsage(),
+            evaluatorTokenUsage: this.evaluatorAgent.getOrchestratorTokenUsage(),
+        };
+    }
+    /** Stop both agents cooperatively. */
+    stop() {
+        this.mainAgent.stop();
+    }
+    /** Cleanup both agents' resources. */
+    async cleanup() {
+        await this.mainAgent.cleanup();
+        await this.evaluatorAgent['mcpManager'].cleanup();
+    }
+}
+//# sourceMappingURL=harness.js.map

package/dist/evaluator/harness.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"harness.js","sourceRoot":"","sources":["../../src/evaluator/harness.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,OAAO,EAAE,MAAM,EAAE,MAAM,oBAAoB,CAAC;AAK5C,MAAM,OAAO,gBAAgB;IACnB,SAAS,CAAS;IAClB,cAAc,CAAiB;IAC/B,OAAO,CAAU;IAEzB,YACE,SAAiB,EACjB,cAA8B,EAC9B,UAA0B,EAAE;QAE5B,IAAI,CAAC,SAAS,GAAG,SAAS,CAAC;QAC3B,IAAI,CAAC,cAAc,GAAG,cAAc,CAAC;QACrC,IAAI,CAAC,OAAO,GAAG,OAAO,CAAC,OAAO,IAAI,IAAI,CAAC;IACzC,CAAC;IAED;;;;OAIG;IACH,KAAK,CAAC,GAAG,CAAC,WAAmB,EAAE,OAAoC;QACjE,0EAA0E;QAC1E,IAAI,IAAI,CAAC,OAAO,EAAE,CAAC;YACjB,MAAM,CAAC,IAAI,CAAC,8CAA8C,CAAC,CAAC;QAC9D,CAAC;QAED,MAAM,YAAY,GAAG,MAAM,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC,WAAW,CAAC,CAAC;QAE5D,IAAI,IAAI,CAAC,OAAO,EAAE,CAAC;YACjB,MAAM,CAAC,IAAI,CACT,+BAA+B,YAAY,CAAC,UAAU,eAAe;gBACrE,GAAG,YAAY,CAAC,SAAS,CAAC,MAAM,aAAa,CAC9C,CAAC;QACJ,CAAC;QAED,0EAA0E;QAC1E,IAAI,IAAI,CAAC,OAAO,EAAE,CAAC;YACjB,MAAM,CAAC,IAAI,CAAC,yCAAyC,CAAC,CAAC;QACzD,CAAC;QAED,MAAM,iBAAiB,GAAsB;YAC3C,WAAW;YACX,oBAAoB,EAAE,OAAO,EAAE,oBAAoB;SACpD,CAAC;QAEF,MAAM,UAAU,GAAG,MAAM,IAAI,CAAC,cAAc,CAAC,QAAQ,CAAC,iBAAiB,CAAC,CAAC;QAEzE,IAAI,IAAI,CAAC,OAAO,EAAE,CAAC;YACjB,MAAM,OAAO,GAAG,UAAU,CAAC,MAAM;gBAC/B,CAAC,CAAC,aAAa,UAAU,CAAC,SAAS,YAAY,UAAU,CAAC,UAAU,UAAU;gBAC9E,CAAC,CAAC,cAAc,UAAU,CAAC,IAAI,CAAC,MAAM,wBAAwB,UAAU,CAAC,SAAS,SAAS,CAAC;YAC9F,MAAM,CAAC,IAAI,CAAC,mCAAmC,OAAO,EAAE,CAAC,CAAC;QAC5D,CAAC;QAED,OAAO;YACL,iBAAiB,EAAE,YAAY,CAAC,OAAO;YACvC,mBAAmB,EAAE,YAAY,CAAC,UAAU;YAC5C,UAAU;YACV,mBAAmB,EAAE,IAAI,CAAC,SAAS,CAAC,aAAa,EAAE;YACnD,mBAAmB,EAAE,IAAI,CAAC,cAAc,CAAC,yBAAyB,EAAE;SACrE,CAAC;IACJ,CAAC;IAED,sCAAsC;IACtC,IAAI;QACF,IAAI,CAAC,SAAS,CAAC,IAAI,EAAE,CAAC;IACxB,CAAC;IAED,sCAAsC;IACtC,KAAK,CAAC,OAAO;QACX,MAAM,IAAI,CAAC,SAAS,CAAC,OAAO,EAAE,CAAC;QAC/B,MAAM,IAAI,CAAC,cAAc,CAAC,YAAY,CAAC,CAAC,OAAO,EAAE,CAAC;IACpD,CAAC;CACF"}

package/dist/evaluator/index.d.ts

@@ -0,0 +1,47 @@
+/**
+ * Evaluator module — public API.
+ *
+ * Usage (CLI):
+ *   const harness = await createEvaluatorHarness(mainAgent, mcpServers, orchestratorConfig);
+ *   const result = await harness.run(userMessage);
+ *
+ * Usage (HTTP):
+ *   const harness = await createEvaluatorHarness(mainAgent, mcpServers, orchestratorConfig);
+ *   app.post('/api/chat/harness', async (req, res) => {
+ *     const result = await harness.run(req.body.message);
+ *     res.json(result);
+ *   });
+ */
+export { EvaluatorAgent } from './evaluator-agent.js';
+export { EvaluatorHarness } from './harness.js';
+export { deriveEvaluatorDirective } from './directive-adapter.js';
+export { InteractWithAgentTool, ListAgentConversationsTool, GetConversationHistoryTool, EVALUATOR_VIRTUAL_TOOLS, } from './tools/agent-tools.js';
+export type { EvaluatorConfig, EvaluationContext, EvaluationResult, HarnessOptions, HarnessResult, EvaluatorToolContext, } from './types.js';
+import { EvaluatorHarness } from './harness.js';
+import type { IAgent } from '../core/agent-interface.js';
+import type { MCPServerConfig } from '../core/config.js';
+import type { HarnessOptions } from './types.js';
+export interface EvaluatorOrchestratorConfig {
+    endpoint: string;
+    apiKey: string;
+    model: string;
+    useHarmonyFormat?: boolean;
+    /** Optional tool-calling model (same env vars as main agent). */
+    toolCallingEndpoint?: string;
+    toolCallingApiKey?: string;
+    toolCallingModel?: string;
+}
+/**
+ * Create an EvaluatorHarness that pairs the given main agent with a fresh,
+ * isolated EvaluatorAgent.
+ *
+ * The evaluator uses the same model config and MCP servers as the main agent
+ * but runs with completely separate LLM state.
+ *
+ * @param mainAgent        The IAgent instance to supervise.
+ * @param mcpServers       MCP server configs (same set as the main agent uses).
+ * @param orchestratorCfg  Model connection details for the evaluator's orchestrator.
+ * @param harnessOptions   Optional harness behaviour tweaks.
+ */
+export declare function createEvaluatorHarness(mainAgent: IAgent, mcpServers: Record<string, MCPServerConfig>, orchestratorCfg: EvaluatorOrchestratorConfig, harnessOptions?: HarnessOptions): Promise<EvaluatorHarness>;
+//# sourceMappingURL=index.d.ts.map

package/dist/evaluator/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/evaluator/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAEH,OAAO,EAAE,cAAc,EAAE,MAAM,sBAAsB,CAAC;AACtD,OAAO,EAAE,gBAAgB,EAAE,MAAM,cAAc,CAAC;AAChD,OAAO,EAAE,wBAAwB,EAAE,MAAM,wBAAwB,CAAC;AAClE,OAAO,EACL,qBAAqB,EACrB,0BAA0B,EAC1B,0BAA0B,EAC1B,uBAAuB,GACxB,MAAM,wBAAwB,CAAC;AAChC,YAAY,EACV,eAAe,EACf,iBAAiB,EACjB,gBAAgB,EAChB,cAAc,EACd,aAAa,EACb,oBAAoB,GACrB,MAAM,YAAY,CAAC;AAQpB,OAAO,EAAE,gBAAgB,EAAE,MAAM,cAAc,CAAC;AAChD,OAAO,KAAK,EAAE,MAAM,EAAE,MAAM,4BAA4B,CAAC;AACzD,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,mBAAmB,CAAC;AACzD,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,YAAY,CAAC;AAEjD,MAAM,WAAW,2BAA2B;IAC1C,QAAQ,EAAE,MAAM,CAAC;IACjB,MAAM,EAAE,MAAM,CAAC;IACf,KAAK,EAAE,MAAM,CAAC;IACd,gBAAgB,CAAC,EAAE,OAAO,CAAC;IAC3B,iEAAiE;IACjE,mBAAmB,CAAC,EAAE,MAAM,CAAC;IAC7B,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAC3B,gBAAgB,CAAC,EAAE,MAAM,CAAC;CAC3B;AAED;;;;;;;;;;;GAWG;AACH,wBAAsB,sBAAsB,CAC1C,SAAS,EAAE,MAAM,EACjB,UAAU,EAAE,MAAM,CAAC,MAAM,EAAE,eAAe,CAAC,EAC3C,eAAe,EAAE,2BAA2B,EAC5C,cAAc,GAAE,cAAmB,GAClC,OAAO,CAAC,gBAAgB,CAAC,CAiD3B"}