sentinelayer-cli 0.1.0

package/src/prompt/generator.js

@@ -0,0 +1,106 @@
+export const SUPPORTED_PROMPT_TARGETS = Object.freeze([
+  "claude",
+  "cursor",
+  "copilot",
+  "codex",
+  "generic",
+]);
+
+const TARGET_GUIDANCE = Object.freeze({
+  claude: [
+    "Use strict PR-scoped execution with explicit plan -> implement -> verify loop.",
+    "Prioritize deterministic checks before optional AI-dependent steps.",
+    "Record evidence for every claim in the final report.",
+  ],
+  cursor: [
+    "Apply edits in small, reviewable commits.",
+    "Keep changes aligned with existing project conventions and tests.",
+    "Run local verification before proposing follow-up iterations.",
+  ],
+  copilot: [
+    "Generate code with strong type and error-path handling.",
+    "Avoid introducing implicit behavior changes in existing modules.",
+    "Include targeted tests for all new logic paths.",
+  ],
+  codex: [
+    "Operate autonomously but keep one PR scope at a time.",
+    "Use deterministic ingest/spec context as primary source of truth.",
+    "Fail closed when requirements are ambiguous or unsafe.",
+  ],
+  generic: [
+    "Follow the provided spec exactly.",
+    "Implement in incremental, verifiable steps.",
+    "Document assumptions and unresolved risks.",
+  ],
+});
+
+function normalizeTarget(target) {
+  const normalized = String(target || "generic").trim().toLowerCase();
+  if (!SUPPORTED_PROMPT_TARGETS.includes(normalized)) {
+    throw new Error(
+      `Unsupported prompt target '${target}'. Use one of: ${SUPPORTED_PROMPT_TARGETS.join(", ")}`
+    );
+  }
+  return normalized;
+}
+
+function buildAgentHeader(target) {
+  const headers = {
+    claude: "Claude Code execution prompt",
+    cursor: "Cursor execution prompt",
+    copilot: "GitHub Copilot execution prompt",
+    codex: "Codex execution prompt",
+    generic: "Generic execution prompt",
+  };
+  return headers[target] || headers.generic;
+}
+
+export function resolvePromptTarget(target) {
+  return normalizeTarget(target);
+}
+
+export function defaultPromptFileName(target) {
+  const normalized = normalizeTarget(target);
+  if (normalized === "generic") {
+    return "PROMPT.md";
+  }
+  return `PROMPT_${normalized}.md`;
+}
+
+export function generateExecutionPrompt({
+  specMarkdown,
+  target = "generic",
+  projectPath,
+  generatedAt = new Date().toISOString(),
+} = {}) {
+  const resolvedTarget = normalizeTarget(target);
+  const guidance = TARGET_GUIDANCE[resolvedTarget] || TARGET_GUIDANCE.generic;
+
+  const specText = String(specMarkdown || "").trim();
+  if (!specText) {
+    throw new Error("Spec content is empty. Generate or provide a spec before creating a prompt.");
+  }
+
+  const guidanceMarkdown = guidance.map((item, index) => `${index + 1}. ${item}`).join("\n");
+
+  return `# ${buildAgentHeader(resolvedTarget)}
+
+Generated: ${generatedAt}
+Agent target: ${resolvedTarget}
+Workspace: ${projectPath || "(not provided)"}
+
+## Operating Rules
+${guidanceMarkdown}
+
+## Execution Workflow
+1. Read the full spec and derive a concrete PR-scoped implementation plan.
+2. Implement only the current scope with deterministic checkpoints.
+3. Run verification commands and capture outcomes.
+4. Report findings first (bugs/regressions), then summarize changes.
+
+## Source Spec (Authoritative)
+\n\`\`\`markdown
+${specText}
+\`\`\`
+`;
+}