@warmdrift/kgauto-compiler 2.0.0-alpha.7 → 2.0.0-alpha.8

package/dist/index.js

@@ -524,7 +524,11 @@ function lowerAnthropic(ir, profile, hints) {
       system: systemBlocks,
       messages,
       tools,
-      max_tokens: hints.forceTerseOutput ? 200 : Math.min(profile.maxOutputTokens, 4096)
+      // alpha.8: trust profile.maxOutputTokens. The historical Math.min(_, 4096)
+      // floor surprised every consumer once (PB-Cairn contract-gaps brief, Gap 3).
+      // Profile is the single source of truth; consumers wanting a tighter
+      // budget can pass providerOverrides.anthropic.max_tokens explicitly.
+      max_tokens: hints.forceTerseOutput ? 200 : profile.maxOutputTokens
     },
     diagnostics: {
       cacheableTokens,
@@ -1806,31 +1810,31 @@ async function call(ir, opts = {}) {
       fetchImpl: opts.fetchImpl,
       providerOverrides: opts.providerOverrides
     });
-    if (exec.ok) {
+    const validated = exec.ok ? validateStructuredContract(exec, ir) : exec;
+    if (validated.ok) {
       attempts.push({ model: targetModel, status: "success" });
       const latencyMs2 = Date.now() - start;
-      const responseWithStructured = withStructuredOutput(exec.response, ir);
       await record({
         handle: initial.handle,
-        tokensIn: responseWithStructured.tokens.input,
-        tokensOut: responseWithStructured.tokens.output,
+        tokensIn: validated.response.tokens.input,
+        tokensOut: validated.response.tokens.output,
         latencyMs: latencyMs2,
         success: true,
-        emptyResponse: responseWithStructured.tokens.output === 0,
-        toolsCalled: responseWithStructured.toolCalls.map((tc) => tc.name),
+        emptyResponse: validated.response.tokens.output === 0,
+        toolsCalled: validated.response.toolCalls.map((tc) => tc.name),
         actualModel: targetModel !== initial.target ? targetModel : void 0,
         mutationsApplied: targetModel !== initial.target ? activeCompile.mutationsApplied.map((m) => m.id) : void 0,
         promptPreview: extractPromptPreview(ir),
-        responsePreview: responseWithStructured.text.slice(0, 200),
-        cacheReadInputTokens: responseWithStructured.tokens.cached,
-        cacheCreationInputTokens: responseWithStructured.tokens.cacheCreated
+        responsePreview: validated.response.text.slice(0, 200),
+        cacheReadInputTokens: validated.response.tokens.cached,
+        cacheCreationInputTokens: validated.response.tokens.cacheCreated
       });
       return {
         handle: initial.handle,
         actualModel: targetModel,
         requestedModel: initial.target,
         provider: activeCompile.provider,
-        response: responseWithStructured,
+        response: validated.response,
         latencyMs: latencyMs2,
         mutationsApplied: activeCompile.mutationsApplied,
         attempts
@@ -1838,12 +1842,12 @@ async function call(ir, opts = {}) {
     }
     attempts.push({
       model: targetModel,
-      status: exec.errorType,
-      errorCode: exec.errorCode,
-      message: exec.message
+      status: validated.errorType,
+      errorCode: validated.errorCode,
+      message: validated.message
     });
-    lastErr = exec;
-    if (exec.errorType === "terminal" || opts.noFallback) {
+    lastErr = validated;
+    if (validated.errorType === "terminal" || opts.noFallback) {
       break;
     }
   }
@@ -1880,17 +1884,35 @@ function extractPromptPreview(ir) {
   if (lastHist) return lastHist.slice(0, 200);
   return void 0;
 }
-function withStructuredOutput(response, ir) {
-  if (!ir.constraints?.structuredOutput) return response;
-  if (!response.text) return response;
+function validateStructuredContract(exec, ir) {
+  if (!ir.constraints?.structuredOutput) {
+    return { ok: true, response: exec.response };
+  }
+  const finish = (exec.response.finishReason ?? "").toLowerCase();
+  if (finish === "max_tokens" || finish === "length") {
+    return {
+      ok: false,
+      status: exec.status,
+      errorType: "retryable",
+      errorCode: "max_tokens_on_structured_output",
+      message: `Provider returned finishReason="${exec.response.finishReason}" on a structured-output call \u2014 output truncated mid-token, JSON cannot be valid`,
+      raw: exec.response.raw
+    };
+  }
+  if (!exec.response.text) {
+    return { ok: true, response: exec.response };
+  }
   try {
-    const parsed = JSON.parse(response.text);
-    return { ...response, structuredOutput: parsed };
+    const parsed = JSON.parse(exec.response.text);
+    return { ok: true, response: { ...exec.response, structuredOutput: parsed } };
   } catch (err) {
     return {
-      ...response,
-      structuredOutput: null,
-      parseError: err instanceof Error ? err.message : String(err)
+      ok: false,
+      status: exec.status,
+      errorType: "retryable",
+      errorCode: "structured_output_parse_failed",
+      message: err instanceof Error ? err.message : String(err),
+      raw: exec.response.raw
     };
   }
 }

package/dist/index.mjs

@@ -408,7 +408,11 @@ function lowerAnthropic(ir, profile, hints) {
       system: systemBlocks,
       messages,
       tools,
-      max_tokens: hints.forceTerseOutput ? 200 : Math.min(profile.maxOutputTokens, 4096)
+      // alpha.8: trust profile.maxOutputTokens. The historical Math.min(_, 4096)
+      // floor surprised every consumer once (PB-Cairn contract-gaps brief, Gap 3).
+      // Profile is the single source of truth; consumers wanting a tighter
+      // budget can pass providerOverrides.anthropic.max_tokens explicitly.
+      max_tokens: hints.forceTerseOutput ? 200 : profile.maxOutputTokens
     },
     diagnostics: {
       cacheableTokens,
@@ -1288,31 +1292,31 @@ async function call(ir, opts = {}) {
       fetchImpl: opts.fetchImpl,
       providerOverrides: opts.providerOverrides
     });
-    if (exec.ok) {
+    const validated = exec.ok ? validateStructuredContract(exec, ir) : exec;
+    if (validated.ok) {
       attempts.push({ model: targetModel, status: "success" });
       const latencyMs2 = Date.now() - start;
-      const responseWithStructured = withStructuredOutput(exec.response, ir);
       await record({
         handle: initial.handle,
-        tokensIn: responseWithStructured.tokens.input,
-        tokensOut: responseWithStructured.tokens.output,
+        tokensIn: validated.response.tokens.input,
+        tokensOut: validated.response.tokens.output,
         latencyMs: latencyMs2,
         success: true,
-        emptyResponse: responseWithStructured.tokens.output === 0,
-        toolsCalled: responseWithStructured.toolCalls.map((tc) => tc.name),
+        emptyResponse: validated.response.tokens.output === 0,
+        toolsCalled: validated.response.toolCalls.map((tc) => tc.name),
         actualModel: targetModel !== initial.target ? targetModel : void 0,
         mutationsApplied: targetModel !== initial.target ? activeCompile.mutationsApplied.map((m) => m.id) : void 0,
         promptPreview: extractPromptPreview(ir),
-        responsePreview: responseWithStructured.text.slice(0, 200),
-        cacheReadInputTokens: responseWithStructured.tokens.cached,
-        cacheCreationInputTokens: responseWithStructured.tokens.cacheCreated
+        responsePreview: validated.response.text.slice(0, 200),
+        cacheReadInputTokens: validated.response.tokens.cached,
+        cacheCreationInputTokens: validated.response.tokens.cacheCreated
       });
       return {
         handle: initial.handle,
         actualModel: targetModel,
         requestedModel: initial.target,
         provider: activeCompile.provider,
-        response: responseWithStructured,
+        response: validated.response,
         latencyMs: latencyMs2,
         mutationsApplied: activeCompile.mutationsApplied,
         attempts
@@ -1320,12 +1324,12 @@ async function call(ir, opts = {}) {
     }
     attempts.push({
       model: targetModel,
-      status: exec.errorType,
-      errorCode: exec.errorCode,
-      message: exec.message
+      status: validated.errorType,
+      errorCode: validated.errorCode,
+      message: validated.message
     });
-    lastErr = exec;
-    if (exec.errorType === "terminal" || opts.noFallback) {
+    lastErr = validated;
+    if (validated.errorType === "terminal" || opts.noFallback) {
       break;
     }
   }
@@ -1362,17 +1366,35 @@ function extractPromptPreview(ir) {
   if (lastHist) return lastHist.slice(0, 200);
   return void 0;
 }
-function withStructuredOutput(response, ir) {
-  if (!ir.constraints?.structuredOutput) return response;
-  if (!response.text) return response;
+function validateStructuredContract(exec, ir) {
+  if (!ir.constraints?.structuredOutput) {
+    return { ok: true, response: exec.response };
+  }
+  const finish = (exec.response.finishReason ?? "").toLowerCase();
+  if (finish === "max_tokens" || finish === "length") {
+    return {
+      ok: false,
+      status: exec.status,
+      errorType: "retryable",
+      errorCode: "max_tokens_on_structured_output",
+      message: `Provider returned finishReason="${exec.response.finishReason}" on a structured-output call \u2014 output truncated mid-token, JSON cannot be valid`,
+      raw: exec.response.raw
+    };
+  }
+  if (!exec.response.text) {
+    return { ok: true, response: exec.response };
+  }
   try {
-    const parsed = JSON.parse(response.text);
-    return { ...response, structuredOutput: parsed };
+    const parsed = JSON.parse(exec.response.text);
+    return { ok: true, response: { ...exec.response, structuredOutput: parsed } };
   } catch (err) {
     return {
-      ...response,
-      structuredOutput: null,
-      parseError: err instanceof Error ? err.message : String(err)
+      ok: false,
+      status: exec.status,
+      errorType: "retryable",
+      errorCode: "structured_output_parse_failed",
+      message: err instanceof Error ? err.message : String(err),
+      raw: exec.response.raw
     };
   }
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@warmdrift/kgauto-compiler",
-  "version": "2.0.0-alpha.7",
+  "version": "2.0.0-alpha.8",
   "description": "Prompt compiler + central learning brain for multi-model AI apps. Swap models without rewriting prompts.",
   "main": "./dist/index.js",
   "module": "./dist/index.mjs",