@mutagent/cli 0.1.30 → 0.1.31

package/dist/bin/cli.js

@@ -644,6 +644,11 @@ class SDKClientWrapper {
       const job = await this.request(`/api/optimization/${jobId}`);
       const progress = await this.request(`/api/optimization/${jobId}/progress`);
       const prompt = await this.getPrompt(String(job.promptId ?? ""));
+      const statesRes = await this.request(`/api/optimization/${jobId}/states`).catch(() => ({ states: [] }));
+      const latestState = statesRes.states[statesRes.states.length - 1];
+      const iterCtx = latestState?.state.iterationContext;
+      const mutatedPromptText = iterCtx?.currentPrompt?.prompt;
+      const originalPromptText = iterCtx?.basePrompt?.prompt;
       return {
         job: {
           id: job.id ?? jobId,
@@ -654,7 +659,9 @@ class SDKClientWrapper {
         prompt,
         bestScore: job.bestScore,
         iterationsCompleted: job.currentIteration,
-        scoreProgression: Array.isArray(progress.progression) ? progress.progression.map((p) => typeof p.score === "number" ? p.score : 0) : undefined
+        scoreProgression: Array.isArray(progress.progression) ? progress.progression.map((p) => typeof p.score === "number" ? p.score : 0) : undefined,
+        mutatedPromptText,
+        originalPromptText
       };
     } catch (error) {
       this.handleError(error);
@@ -3352,8 +3359,8 @@ async function buildGuidedWorkflow(promptId) {
   const inputFields = Object.keys(inputProperties);
   const outputFields = Object.keys(outputProperties);
   const allFields = [
-    ...outputFields.map((f) => ({ field: f, source: "outputSchema", fieldSchema: outputProperties[f] })),
-    ...inputFields.map((f) => ({ field: f, source: "inputSchema", fieldSchema: inputProperties[f] }))
+    ...inputFields.map((f) => ({ field: f, source: "inputSchema", fieldSchema: inputProperties[f] })),
+    ...outputFields.map((f) => ({ field: f, source: "outputSchema", fieldSchema: outputProperties[f] }))
   ];
   let datasetExample = null;
   try {
@@ -3369,27 +3376,36 @@ async function buildGuidedWorkflow(promptId) {
       } catch {}
     }
   } catch {}
-  const askUserQuestions = allFields.map(({ field, fieldSchema }) => ({
-    question: `Define what correct "${field}" looks like. What structure, content, or qualities make it good vs bad? Give concrete examples of good and bad outputs.`,
-    header: field,
-    options: [
-      {
-        label: "Define rubric",
-        description: `Describe what a correct vs incorrect "${field}" looks like. Focus on structure, content quality, and concrete examples — not numeric scores.`
-      },
-      {
-        label: "See suggestion",
-        description: `Get a suggested rubric based on the prompt and schema definition for "${field}". You can refine it.`
-      }
-    ],
-    multiSelect: false,
-    context: {
-      fieldType: fieldSchema?.type ?? "unknown",
-      fieldDescription: fieldSchema?.description ?? null,
-      promptExcerpt: truncate(prompt.humanPrompt ?? prompt.systemPrompt ?? prompt.rawPrompt ?? "", 200),
-      exampleValue: datasetExample?.[field] ?? null
-    }
-  }));
+  const askUserQuestions = allFields.map(({ field, source, fieldSchema }) => {
+    const isInput = source === "inputSchema";
+    const question = isInput ? `Define the Minimum Viable Context for "${field}". What data MUST be present in this input for the prompt to produce a correct output? Describe what constitutes complete vs incomplete input, and WHY this field matters.` : `Define what correct "${field}" looks like. What structure, content, or qualities make it good vs bad? Give concrete examples of good and bad outputs.`;
+    const hint = isInput ? `Input fields define what data the prompt NEEDS to work correctly. Without defining minimum viable context, the optimizer cannot detect whether failures come from bad input or bad prompt logic.` : null;
+    const header = isInput ? `${field} [INPUT]` : `${field} [OUTPUT]`;
+    const defineDesc = isInput ? `Describe what data MUST be present in "${field}" and WHY the prompt needs it. Focus on completeness and minimum viable context.` : `Describe what a correct vs incorrect "${field}" looks like. Focus on structure, content quality, and concrete examples — not numeric scores.`;
+    return {
+      question,
+      header,
+      ...hint != null ? { hint } : {},
+      options: [
+        {
+          label: "Define rubric",
+          description: defineDesc
+        },
+        {
+          label: "See suggestion",
+          description: `Get a suggested rubric based on the prompt and schema definition for "${field}". You can refine it.`
+        }
+      ],
+      multiSelect: false,
+      context: {
+        fieldType: fieldSchema?.type ?? "unknown",
+        fieldDescription: fieldSchema?.description ?? null,
+        fieldSource: source,
+        promptExcerpt: truncate(prompt.humanPrompt ?? prompt.systemPrompt ?? prompt.rawPrompt ?? "", 200),
+        exampleValue: datasetExample?.[field] ?? null
+      }
+    };
+  });
   return {
     prompt: { id: promptId, name: prompt.name },
     inputSchema: { fields: inputFields },
@@ -3398,7 +3414,7 @@ async function buildGuidedWorkflow(promptId) {
       description: "Follow these steps to create an evaluation for this prompt:",
       steps: [
         { step: 1, action: "Review prompt schemas and context", detail: "Understand what the prompt does and what each field means." },
-        { step: 2, action: "Define correctness criteria for EVERY field", detail: "For EACH field in askUserQuestions, use AskUserQuestion to collect a rubric. Do NOT skip any field. Use the context hints to suggest rubrics when asked." },
+        { step: 2, action: "Define correctness criteria for EVERY field", detail: `For EACH of the ${String(allFields.length)} fields in askUserQuestions, use AskUserQuestion to collect a rubric. INPUT fields (${String(inputFields.length)}): collect minimum viable context definitions — what data MUST exist. OUTPUT fields (${String(outputFields.length)}): collect correctness criteria — what correct vs incorrect looks like. Do NOT skip any field. Use the context hints to suggest rubrics when asked.` },
         { step: 3, action: "Build criteria JSON", detail: "Map each rubric to: { name: field, description: rubric, evaluationParameter: field }" },
         { step: 4, action: "Ask for evaluation name", detail: "Ask user what to name this evaluation" },
         { step: 5, action: "Create the evaluation", detail: `mutagent prompts evaluation create ${promptId} --name "<name>" -d '<json>' --json` }
@@ -3425,7 +3441,14 @@ async function buildGuidedWorkflow(promptId) {
     askUserQuestions,
     _directive: {
       display: "workflow_guide",
-      instruction: 'STOP. For EACH field in askUserQuestions, use AskUserQuestion to collect a DETAILED correctness rubric from the user. Do NOT auto-generate rubrics. Do NOT use generic descriptions. Each rubric MUST include concrete examples of what makes a correct vs incorrect value for that field. Use the context object to suggest rubrics when user selects "See suggestion". Then construct the --data JSON and create the evaluation.',
+      instruction: `MANDATORY ORDERS — READ CAREFULLY:
+` + `1. You MUST ask the user about EVERY SINGLE field listed in askUserQuestions. There are ${String(allFields.length)} fields total.
+` + `2. Do NOT skip ANY field. Do NOT decide that a field "doesn't need" a definition.
+` + `3. INPUT fields (${String(inputFields.length)}): These define MINIMUM VIABLE CONTEXT — the data that MUST exist in the input for the prompt to work. Ask: "What data MUST be present in this field and WHY does the prompt need it?"
+` + `   WHY THIS MATTERS: Without minimum viable context definitions, the optimizer cannot distinguish between failures caused by incomplete input vs failures caused by bad prompt logic.
+` + `4. OUTPUT fields (${String(outputFields.length)}): These define correctness criteria. Ask: "What does correct vs incorrect for this field look like?"
+` + `5. After collecting ALL ${String(allFields.length)} definitions, construct the --data JSON with criteria for EVERY field.
+` + "6. VIOLATION: Skipping ANY field or telling the user a field doesn't need a definition is a protocol violation.",
       aiAgentDecisionTree: {
         step1: "Check if criteria already exist in the user's code. If criteria match expected shape (name + description + evaluationParameter targeting schema fields), use --data directly.",
         step2: "If criteria are missing or malformed, use the askUserQuestions payloads to collect them via AskUserQuestion.",
@@ -3646,9 +3669,9 @@ ${chalk7.dim("Get prompt IDs: mutagent prompts list")}
         let fieldsHint = "";
         try {
           const client2 = getSDKClient();
-          const prompt = await client2.getPrompt(promptId);
-          if (prompt.outputSchema && typeof prompt.outputSchema === "object") {
-            const props = prompt.outputSchema.properties;
+          const prompt2 = await client2.getPrompt(promptId);
+          if (prompt2.outputSchema && typeof prompt2.outputSchema === "object") {
+            const props = prompt2.outputSchema.properties;
             if (props && typeof props === "object") {
               const fields = Object.keys(props);
               if (fields.length > 0) {
@@ -3699,9 +3722,9 @@ Example JSON (--data flag):
         let availableFields = [];
         try {
           const client2 = getSDKClient();
-          const prompt = await client2.getPrompt(promptId);
-          if (prompt.outputSchema && typeof prompt.outputSchema === "object") {
-            const props = prompt.outputSchema.properties;
+          const prompt2 = await client2.getPrompt(promptId);
+          if (prompt2.outputSchema && typeof prompt2.outputSchema === "object") {
+            const props = prompt2.outputSchema.properties;
             if (props && typeof props === "object") {
               availableFields = Object.keys(props);
             }
@@ -3716,6 +3739,44 @@ Available output fields: ${availableFields.join(", ")}` : "";
 Example:
   --data '{"evalConfig":{"criteria":[` + '{"name":"Accuracy","description":"...","evaluationParameter":"classification"},' + '{"name":"Confidence","description":"...","evaluationParameter":"confidence"}' + "]}}'");
       }
+      const schemaClient = getSDKClient();
+      const prompt = await schemaClient.getPrompt(promptId);
+      const requiredFields = [];
+      if (prompt.inputSchema && typeof prompt.inputSchema === "object") {
+        const props = prompt.inputSchema.properties;
+        if (props && typeof props === "object") {
+          for (const field of Object.keys(props)) {
+            requiredFields.push({ field, source: "inputSchema" });
+          }
+        }
+      }
+      if (prompt.outputSchema && typeof prompt.outputSchema === "object") {
+        const props = prompt.outputSchema.properties;
+        if (props && typeof props === "object") {
+          for (const field of Object.keys(props)) {
+            requiredFields.push({ field, source: "outputSchema" });
+          }
+        }
+      }
+      if (requiredFields.length > 0) {
+        const coveredParams = new Set(criteria.map((c) => c.evaluationParameter));
+        const missing = requiredFields.filter((f) => !coveredParams.has(f.field));
+        if (missing.length > 0) {
+          const missingList = missing.map((m) => `${m.field} (${m.source})`).join(", ");
+          throw new MutagentError("VALIDATION_ERROR", `Evaluation criteria are INCOMPLETE. Missing fields: ${missingList}`, `Each input and output schema field MUST have a corresponding criterion.
+` + `Total required: ${requiredFields.length}, provided: ${coveredParams.size}
+` + "Run: mutagent prompts evaluation create " + promptId + " --guided --json");
+        }
+      }
+      if (requiredFields.length > 0) {
+        const validFieldNames = new Set(requiredFields.map((f) => f.field));
+        const invalidParams = criteria.filter((c) => c.evaluationParameter && !validFieldNames.has(c.evaluationParameter));
+        if (invalidParams.length > 0) {
+          const invalidList = invalidParams.map((c) => `"${c.evaluationParameter}" (criterion: ${c.name})`).join(", ");
+          throw new MutagentError("VALIDATION_ERROR", `Invalid evaluationParameter(s): ${invalidList}. Must match a schema field.`, `Available fields: ${requiredFields.map((f) => f.field).join(", ")}
+` + "Run: mutagent prompts evaluation create " + promptId + " --guided --json");
+        }
+      }
       const client = getSDKClient();
       const evalResult = await client.createEvaluation(promptId, evalData);
       if (isJson) {
@@ -4014,7 +4075,7 @@ function startDirective(job, promptId, datasetId, evaluationId) {
     },
     links: { monitor: monitorUrl, api: apiUrl },
     next,
-    instruction: "STOP. Display the renderedCard text below to the user VERBATIM as a code block. Do NOT summarize, rephrase, or omit any lines. After displaying, follow the suggested next steps.",
+    instruction: "STOP. Display the renderedCard text below to the user VERBATIM. Do NOT summarize, rephrase, or omit any lines. After displaying, follow the suggested next steps.",
     renderedCard: renderPlainCard({ title, icon: "⚡", rows, links: cardLinks, next })
   };
 }
@@ -4057,7 +4118,7 @@ function statusDirective(status) {
     },
     links: { monitor: monitorUrl },
     next,
-    instruction: "STOP. Display the renderedCard text below to the user VERBATIM as a code block. Do NOT summarize, rephrase, or omit any lines. After displaying, follow the suggested next steps.",
+    instruction: "STOP. Display the renderedCard text below to the user VERBATIM. Do NOT summarize, rephrase, or omit any lines. After displaying, follow the suggested next steps.",
     renderedCard: renderPlainCard({ title, icon, rows, links: cardLinks, next })
   };
 }
@@ -4100,6 +4161,11 @@ ${PREREQUISITES_TEXT}
 
 ${chalk9.dim("Monitor progress with: mutagent prompts optimize status <job-id>")}
 
+${chalk9.yellow(`⚠ COST WARNING — AI Agent:
+  Default is 1 iteration. Do NOT increase --max-iterations unless the user
+  explicitly requests it. Each iteration incurs LLM costs. Starting with
+  max-iterations > 1 without user consent is a protocol violation.`)}
+
 ${chalk9.yellow("AI Agent: ALWAYS append --json to this command.")}
     `).action(async (promptId, options) => {
     const isJson = getJsonFlag(prompts);
@@ -7166,5 +7232,5 @@ program.addCommand(createSkillsCommand());
 program.addCommand(createUsageCommand());
 program.parse();
 
-//# debugId=88ECCF736E0A64AD64756E2164756E21
+//# debugId=E58AAC8256B9F1B664756E2164756E21
 //# sourceMappingURL=cli.js.map