promptpilot 0.1.8 → 0.2.2

package/dist/index.js

@@ -1420,6 +1420,16 @@ Mode: Ultra compression. Minimize tokens aggressively.` : getOptimizationSystemP
         };
       }
       const claudeTiersOnly = isClaudeTiersOnlyTargetSet(availableTargets);
+      const routingCandidates = claudeTiersOnly ? filterClaudeTierCandidates(availableTargets, options.input, options.routingPriority) : availableTargets;
+      if (routingCandidates.length === 1) {
+        return {
+          selectedTarget: stripInternalTargetFields(routingCandidates[0]),
+          rankedTargets: [{ ...stripInternalTargetFields(routingCandidates[0]), rank: 1, reason: "Selected by Claude tier pre-filter based on prompt signals." }],
+          routingReason: "Selected by Claude tier pre-filter based on prompt signals.",
+          routingWarnings: [],
+          routingProvider: "heuristic"
+        };
+      }
       const response = await this.client.generateJson({
         model: routerModel,
         timeoutMs: options.input.timeoutMs ?? this.config.timeoutMs,
@@ -1437,7 +1447,7 @@ Mode: Ultra compression. Minimize tokens aggressively.` : getOptimizationSystemP
             targetHints: options.input.targetHints ?? [],
             workloadBias: options.workloadBias,
             routingPriority: options.routingPriority,
-            candidateTargets: availableTargets.map((target) => ({
+            candidateTargets: routingCandidates.map((target) => ({
               id: target.id,
               provider: target.provider,
               model: target.model,
@@ -1456,7 +1466,7 @@ Mode: Ultra compression. Minimize tokens aggressively.` : getOptimizationSystemP
         new Set((response.rankedTargetIds ?? []).map((value) => value.trim()).filter(Boolean))
       ).slice(0, Math.max(1, options.routingTopK));
       const rankedTargets = rankedTargetIds.map((id, index) => {
-        const target = availableTargets.find((candidate) => candidate.id === id);
+        const target = routingCandidates.find((candidate) => candidate.id === id);
         if (!target) {
           return null;
         }
@@ -1467,7 +1477,7 @@ Mode: Ultra compression. Minimize tokens aggressively.` : getOptimizationSystemP
         };
       }).filter((value) => value !== null);
       const selectedTargetId = response.selectedTargetId?.trim();
-      const selectedTargetCandidate = (selectedTargetId && availableTargets.find((candidate) => candidate.id === selectedTargetId)) ?? (rankedTargets[0] ? availableTargets.find(
+      const selectedTargetCandidate = (selectedTargetId && routingCandidates.find((candidate) => candidate.id === selectedTargetId)) ?? (rankedTargets[0] ? routingCandidates.find(
         (candidate) => candidate.provider === rankedTargets[0].provider && candidate.model === rankedTargets[0].model && candidate.label === rankedTargets[0].label
       ) ?? null : null);
       if (!selectedTargetCandidate || rankedTargets.length === 0) {
@@ -1665,12 +1675,13 @@ function buildDownstreamRoutingSystemPrompt(priority, workloadBias, claudeTiersO
   if (claudeTiersOnly) {
     lines.push(
       "You are choosing between Claude model tiers (Haiku, Sonnet, Opus).",
-      "Haiku: fastest and cheapest. Best for email, chat, support, summarization, and simple rewrites. Avoid for deep coding or multi-step reasoning.",
-      "Sonnet: balanced cost and capability. Best for coding, debugging, refactoring, writing, and general-purpose tasks. The default for most prompts.",
-      "Opus: most capable and most expensive. Reserve for complex architecture decisions, multi-constraint agentic planning, long-horizon reasoning, or prompts that clearly require the strongest model.",
-      "When routing priority is cheapest_adequate: prefer Haiku for lightweight tasks, Sonnet for most code and writing tasks, and Opus only when clearly necessary.",
-      "When routing priority is best_quality: prefer Opus for code and reasoning, Sonnet for writing and simple code.",
-      "When routing priority is fastest_adequate: prefer Haiku unless the task clearly needs Sonnet-level capability."
+      "Haiku: fastest and cheapest. ONLY suitable for email, chat, support, summarization, and trivial one-sentence rewrites. Do NOT use Haiku for any coding, debugging, refactoring, or technical tasks.",
+      "Sonnet: balanced cost and capability. The DEFAULT for all coding, debugging, refactoring, writing, and general-purpose tasks. If the prompt mentions code, a file, a module, a bug, or any technical work, choose Sonnet at minimum.",
+      "Opus: most capable and most expensive. Use for complex architecture decisions, multi-constraint agentic planning, system design, long-horizon reasoning, or when the prompt explicitly requires the strongest model.",
+      "When routing priority is cheapest_adequate: Haiku for non-technical lightweight tasks only, Sonnet for anything involving code or technical content, Opus only when clearly necessary.",
+      "When routing priority is best_quality: Opus for all code and reasoning tasks, Sonnet for writing and non-technical tasks.",
+      "When routing priority is fastest_adequate: Haiku only for lightweight non-technical tasks, Sonnet otherwise.",
+      "IMPORTANT: refactor, debug, fix, auth, module, CI, test, and TypeScript are all coding signals \u2014 always choose Sonnet or Opus for these, never Haiku."
     );
   }
   return lines.join("\n");
@@ -1706,6 +1717,39 @@ function isClaudeTiersOnlyTargetSet(targets) {
     (t) => t.provider === "anthropic" && /haiku|sonnet|opus/i.test(t.model)
   );
 }
+function isCodeSignal(input) {
+  const task = (input.task ?? "").toLowerCase();
+  const preset = (input.preset ?? "").toLowerCase();
+  const hints = input.targetHints ?? [];
+  return task === "code" || preset === "code" || hints.some((h) => ["coding", "agentic", "tool_use", "refactor", "debugging", "architecture"].includes(h)) || /\b(refactor|debug|fix|auth|module|ci|test|typescript|javascript|function|class|api|endpoint|build|deploy|lint|migration)\b/i.test(input.prompt);
+}
+function isArchitectureSignal(input) {
+  const hints = input.targetHints ?? [];
+  return hints.includes("architecture") || /\b(architect|architecture|design system|migration plan|multi.?step|long.?horizon|agentic.*plan|system design|microservice|monolith)\b/i.test(input.prompt);
+}
+function filterClaudeTierCandidates(targets, input, priority) {
+  if (priority === "best_quality") {
+    const filtered = targets.filter((t) => /opus|sonnet/i.test(t.model));
+    return filtered.length > 0 ? filtered : targets;
+  }
+  if (priority === "cheapest_adequate") {
+    if (isArchitectureSignal(input)) {
+      const filtered2 = targets.filter((t) => /opus|sonnet/i.test(t.model));
+      return filtered2.length > 0 ? filtered2 : targets;
+    }
+    if (isCodeSignal(input)) {
+      const sonnet = targets.find((t) => /sonnet/i.test(t.model));
+      return sonnet ? [sonnet] : targets.filter((t) => !/haiku/i.test(t.model));
+    }
+    const filtered = targets.filter((t) => /haiku|sonnet/i.test(t.model));
+    return filtered.length > 0 ? filtered : targets;
+  }
+  if (priority === "fastest_adequate") {
+    const filtered = targets.filter((t) => !/opus/i.test(t.model));
+    return filtered.length > 0 ? filtered : targets;
+  }
+  return targets;
+}
 function selectClaudeTierHeuristic(input, priority, targets) {
   const haiku = targets.find((t) => /haiku/i.test(t.model)) ?? null;
   const sonnet = targets.find((t) => /sonnet/i.test(t.model)) ?? null;
@@ -1716,15 +1760,18 @@ function selectClaudeTierHeuristic(input, priority, targets) {
   const prompt = input.prompt;
   const isLightweight = ["email", "chat", "support", "summarization"].includes(task) || ["email", "chat", "support", "summarization"].includes(preset) || hints.some((h) => ["email", "support", "chat", "summarization"].includes(h));
   const needsOpus = /\b(architect|architecture|design system|migration plan|multi.?step|complex.*refactor|long.?horizon|agentic.*plan)\b/i.test(prompt) || hints.includes("architecture") || priority === "best_quality";
+  const isCodeTask = ["code"].includes(task) || ["code"].includes(preset) || hints.some((h) => ["coding", "agentic", "tool_use", "refactor", "debugging", "architecture"].includes(h)) || /\b(refactor|debug|fix|auth|module|ci|test|typescript|javascript|function|class|api|endpoint)\b/i.test(prompt);
   if (priority === "fastest_adequate") {
-    return isLightweight || !needsOpus ? haiku ?? sonnet : sonnet ?? haiku;
+    if (needsOpus) return opus ?? sonnet;
+    if (isCodeTask) return sonnet ?? opus;
+    return haiku ?? sonnet;
   }
-  if (needsOpus) {
+  if (priority === "best_quality") {
     return opus ?? sonnet ?? haiku;
   }
-  if (isLightweight && priority === "cheapest_adequate") {
-    return haiku ?? sonnet;
-  }
+  if (needsOpus) return opus ?? sonnet;
+  if (isCodeTask) return sonnet ?? opus;
+  if (isLightweight) return haiku ?? sonnet;
   return sonnet ?? haiku ?? opus;
 }
 function inferCapabilities(target) {