@rely-ai/caliber 1.11.1 → 1.12.0

package/dist/bin.js

@@ -1554,7 +1554,7 @@ SCORING CRITERIA \u2014 your output is scored deterministically. Optimize for 10
 Existence (25 pts):
 - CLAUDE.md exists (6 pts) \u2014 always generate for claude/both targets
 - AGENTS.md exists (6 pts) \u2014 always generate for codex target (serves as primary instructions file)
-- Skills configured (8 pts) \u2014 generate exactly 3 focused skills for full points (6 base + 1 per extra, cap 2). Two skills = 7 pts, three = 8 pts.
+- Skills configured (8 pts) \u2014 generate at least 3 skills for full points. Generate more if the project has multiple distinct tools, frameworks, or workflows that benefit from dedicated skills.
 - MCP servers mentioned (3 pts) \u2014 reference detected MCP integrations
 - For "both" target: .cursorrules/.cursor/rules/ exist (3+3 pts), cross-platform parity (2 pts)
 
@@ -1588,7 +1588,7 @@ Bonus (5 pts):
 
 OUTPUT SIZE CONSTRAINTS \u2014 these are critical:
 - CLAUDE.md / AGENTS.md: MUST be under 100 lines for maximum score. Aim for 70-90 lines. Be extremely concise \u2014 only commands, architecture overview, and key conventions. Use bullet points and tables, not prose.
-- Skills: generate exactly 3 skills per target platform. Only go above 3 for large multi-framework projects.
+- Skills: generate 3-6 skills per target platform based on project complexity. Each skill should cover a distinct tool, workflow, or domain \u2014 don't pad with generic skills.
 - Each skill content: max 150 lines. Focus on patterns and examples, not exhaustive docs.
 - Cursor rules: max 5 .mdc files.
 - If the project is large, prioritize depth on the 3-4 most critical tools over breadth across everything.`;
@@ -1646,7 +1646,7 @@ CoreSetup schema:
 }
 
 IMPORTANT: Do NOT generate full skill content. Only output skill topic names and descriptions.
-Skills will be generated separately. Generate exactly 3 skill topics per target platform.
+Skills will be generated separately. Generate 3-6 skill topics per target platform based on project complexity. Each topic should cover a distinct tool, workflow, or domain.
 
 Skill topic description MUST include WHAT it does + WHEN to use it with specific trigger phrases.
 Example: "Manages database migrations. Use when user says 'run migration', 'create migration', 'db schema change', or modifies files in db/migrations/."
@@ -1687,7 +1687,7 @@ Bonus (5 pts): Hooks (2 pts), AGENTS.md (1 pt), OpenSkills format (2 pts) \u2014
 OUTPUT SIZE CONSTRAINTS:
 - CLAUDE.md / AGENTS.md: MUST be under 100 lines. Aim for 70-90 lines.
 - Cursor rules: max 5 .mdc files.
-- Skill topics: exactly 3 per platform (name + description only, no content).`;
+- Skill topics: 3-6 per platform based on project complexity (name + description only, no content).`;
 var SKILL_GENERATION_PROMPT = `You generate a single skill file for a coding agent (Claude Code, Cursor, or Codex).
 
 Given project context and a skill topic, produce a focused SKILL.md body.
@@ -6398,21 +6398,32 @@ async function evaluateDismissals(failingChecks, fingerprint) {
     name: c.name,
     suggestion: c.suggestion
   }));
+  const hasBuildFiles = fingerprint.fileTree.some(
+    (f) => /^(package\.json|Makefile|Cargo\.toml|go\.mod|pyproject\.toml|requirements\.txt|build\.gradle|pom\.xml)$/i.test(f.split("/").pop() || "")
+  );
+  const topFiles = fingerprint.fileTree.slice(0, 30).join(", ");
   try {
     const result = await llmJsonCall({
       system: `You evaluate whether scoring checks are applicable to a project.
-Given the project's languages/frameworks and a list of failing checks, return which checks are NOT applicable.
+Given the project context and a list of failing checks, return which checks are NOT applicable.
+
+Only dismiss checks that truly don't apply. Examples:
+- "Build/test/lint commands" for a GitOps/Helm/Terraform/config repo with no build system
+- "Build/test/lint commands" for a repo with only YAML, HCL, or config files and no package.json/Makefile
+- "Dependency coverage" for a repo with no package manager
 
-Only dismiss checks that truly don't apply \u2014 e.g. "Build/test/lint commands" for a pure Terraform/HCL repo with no build system.
 Do NOT dismiss checks that could reasonably apply even if the project doesn't use them yet.
 
 Return {"dismissed": [{"id": "check_id", "reason": "brief reason"}]} or {"dismissed": []} if all apply.`,
       prompt: `Languages: ${fingerprint.languages.join(", ") || "none"}
 Frameworks: ${fingerprint.frameworks.join(", ") || "none"}
+Tools: ${fingerprint.tools.join(", ") || "none"}
+Has build files (package.json, Makefile, etc.): ${hasBuildFiles ? "yes" : "no"}
+Top files: ${topFiles}
 
 Failing checks:
 ${JSON.stringify(checkList, null, 2)}`,
-      maxTokens: 200,
+      maxTokens: 300,
       ...fastModel ? { model: fastModel } : {}
     });
     if (!Array.isArray(result.dismissed)) return [];

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@rely-ai/caliber",
-  "version": "1.11.1",
+  "version": "1.12.0",
   "description": "Analyze your codebase and generate optimized AI agent configs (CLAUDE.md, .cursorrules, skills) — no API key needed",
   "type": "module",
   "bin": {