@rely-ai/caliber 1.20.0-dev.1773695140 → 1.20.0-dev.1773696270

package/dist/bin.js

@@ -2857,7 +2857,7 @@ Return the complete updated AgentSetup JSON incorporating the user's changes. Re
         system: REFINE_SYSTEM_PROMPT,
         prompt,
         messages: conversationHistory,
-        maxTokens: 64e3
+        maxTokens: 16e3
       },
       {
         onText: (text) => {
@@ -6576,18 +6576,12 @@ async function scoreAndRefine(setup, dir, sessionHistory, callbacks) {
       const issueNames = issues.map((i) => i.check).join(", ");
       callbacks.onStatus(`Fixing ${issues.length} scoring issue${issues.length === 1 ? "" : "s"}: ${issueNames}...`);
     }
-    const feedbackMessage = buildFeedbackMessage(issues);
-    const refined = await refineSetup(currentSetup, feedbackMessage, sessionHistory);
-    if (!refined) {
+    const patched = await applyTargetedFixes(currentSetup, issues, sessionHistory);
+    if (!patched) {
       if (callbacks?.onStatus) callbacks.onStatus("Refinement failed, keeping current setup");
       return bestSetup;
     }
-    sessionHistory.push({ role: "user", content: feedbackMessage });
-    sessionHistory.push({
-      role: "assistant",
-      content: `Applied scoring fixes for: ${issues.map((i) => i.check).join(", ")}`
-    });
-    currentSetup = refined;
+    currentSetup = patched;
   }
   const finalIssues = validateSetup(currentSetup, dir, cachedExists);
   const finalLostPoints = countIssuePoints(finalIssues);
@@ -6596,6 +6590,46 @@ async function scoreAndRefine(setup, dir, sessionHistory, callbacks) {
   }
   return bestSetup;
 }
+async function applyTargetedFixes(setup, issues, sessionHistory) {
+  const { claudeMd, agentsMd } = extractConfigContent(setup);
+  const content = claudeMd ?? agentsMd;
+  if (!content) return null;
+  const feedbackMessage = buildFeedbackMessage(issues);
+  const contentLabel = claudeMd ? "CLAUDE.md" : "AGENTS.md";
+  const prompt = [
+    `Here is the current ${contentLabel} content:
+`,
+    "```markdown",
+    content,
+    "```\n",
+    feedbackMessage,
+    `
+Return ONLY the fixed ${contentLabel} content as raw markdown (no JSON, no code fences). Apply the fixes above and nothing else \u2014 do not rewrite, restructure, or make cosmetic changes.`
+  ].join("\n");
+  try {
+    const raw = await llmCall({
+      system: `You fix scoring issues in AI agent configuration files. Return only the fixed markdown content \u2014 no explanations, no JSON wrappers, no code fences.`,
+      prompt,
+      maxTokens: 8e3
+    });
+    const fixed = stripMarkdownFences(raw).trim();
+    if (!fixed || fixed.length < 50) return null;
+    const patched = JSON.parse(JSON.stringify(setup));
+    if (claudeMd) {
+      patched.claude.claudeMd = fixed;
+    } else {
+      patched.codex.agentsMd = fixed;
+    }
+    sessionHistory.push({ role: "user", content: feedbackMessage });
+    sessionHistory.push({
+      role: "assistant",
+      content: `Applied scoring fixes for: ${issues.map((i) => i.check).join(", ")}`
+    });
+    return patched;
+  } catch {
+    return null;
+  }
+}
 async function runScoreRefineWithSpinner(setup, dir, sessionHistory) {
   const spinner = ora2("Validating setup against scoring criteria...").start();
   try {
@@ -7123,6 +7157,7 @@ async function initCommand(options) {
   const TASK_CONFIG = display.add("Generating configs");
   const TASK_SKILLS_GEN = display.add("Generating skills");
   const TASK_SKILLS_SEARCH = wantsSkills ? display.add("Searching community skills") : -1;
+  const TASK_SCORE_REFINE = display.add("Validating & refining setup");
   display.start();
   display.enableWaitingContent();
   try {
@@ -7231,6 +7266,29 @@ async function initCommand(options) {
       }
     })() : Promise.resolve();
     await Promise.all([generatePromise, searchPromise]);
+    if (generatedSetup) {
+      display.update(TASK_SCORE_REFINE, "running");
+      const sessionHistory2 = [];
+      sessionHistory2.push({
+        role: "assistant",
+        content: summarizeSetup("Initial generation", generatedSetup)
+      });
+      try {
+        const refined = await scoreAndRefine(generatedSetup, process.cwd(), sessionHistory2, {
+          onStatus: (msg) => display.update(TASK_SCORE_REFINE, "running", msg)
+        });
+        if (refined !== generatedSetup) {
+          display.update(TASK_SCORE_REFINE, "done", "Refined");
+          generatedSetup = refined;
+        } else {
+          display.update(TASK_SCORE_REFINE, "done", "Passed");
+        }
+      } catch {
+        display.update(TASK_SCORE_REFINE, "done", "Skipped");
+      }
+    } else {
+      display.update(TASK_SCORE_REFINE, "failed", "No setup to validate");
+    }
   } catch (err) {
     display.stop();
     const msg = err instanceof Error ? err.message : "Unknown error";
@@ -7268,7 +7326,6 @@ async function initCommand(options) {
     role: "assistant",
     content: summarizeSetup("Initial generation", generatedSetup)
   });
-  generatedSetup = await runScoreRefineWithSpinner(generatedSetup, process.cwd(), sessionHistory);
   console.log(title.bold("  Step 3/4 \u2014 Review\n"));
   const setupFiles = collectSetupFiles(generatedSetup, targetAgent);
   const staged = stageFiles(setupFiles, process.cwd());

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@rely-ai/caliber",
-  "version": "1.20.0-dev.1773695140",
+  "version": "1.20.0-dev.1773696270",
   "description": "Analyze your codebase and generate optimized AI agent configs (CLAUDE.md, .cursorrules, skills) — no API key needed",
   "type": "module",
   "bin": {