@warmdrift/kgauto-compiler 2.0.0-alpha.65 → 2.0.0-alpha.66

package/dist/{chunk-XTZ2ZQJQ.mjs → chunk-CMB6PZWS.mjs}

@@ -1,5 +1,5 @@
 // src/version.ts
-var LIBRARY_VERSION = "2.0.0-alpha.65";
+var LIBRARY_VERSION = "2.0.0-alpha.66";
 
 // src/key-health.ts
 var JSON_HEADERS = { "Content-Type": "application/json" };

package/dist/{chunk-DJVGJEJU.mjs → chunk-SBFSYCQG.mjs}

@@ -333,6 +333,10 @@ function readBrainReadEnv(envSource = defaultEnv()) {
   if (!anonKey) missingEnv.push("KGAUTO_V2_BRAIN_ANON_KEY");
   return { endpoint, jwt, anonKey, missingEnv };
 }
+function isSameModelRetryEnabledFromEnv(envSource = defaultEnv()) {
+  const raw = (envSource.KGAUTO_SAME_MODEL_RETRY ?? "").trim().toLowerCase();
+  return raw === "1" || raw === "true";
+}
 
 // src/fallback.ts
 var STARTER_CHAINS_GROUNDED = {
@@ -701,6 +705,7 @@ export {
   isModelReachable,
   getReachabilityDiagnostic,
   readBrainReadEnv,
+  isSameModelRetryEnabledFromEnv,
   loadChainsFromBrain,
   getDefaultFallbackChain,
   getStarterChain,

package/dist/glassbox/index.d.mts

@@ -1,6 +1,6 @@
-import { G as GlassboxEvent } from '../types-D-G7KXeI.mjs';
-export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-D-G7KXeI.mjs';
-import '../ir-9rrfI3yB.mjs';
+import { G as GlassboxEvent } from '../types-8pWRDbUg.mjs';
+export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-8pWRDbUg.mjs';
+import '../ir-BTvyl8-w.mjs';
 import '../dialect.mjs';
 
 /**

package/dist/glassbox/index.d.ts

@@ -1,6 +1,6 @@
-import { G as GlassboxEvent } from '../types-DSknEy2B.js';
-export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-DSknEy2B.js';
-import '../ir-Dx_UXPfP.js';
+import { G as GlassboxEvent } from '../types-BdYlJz0Z.js';
+export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-BdYlJz0Z.js';
+import '../ir-Qnw6L265.js';
 import '../dialect.js';
 
 /**

package/dist/glassbox-routes/format.d.mts

@@ -1,5 +1,5 @@
-import { T as TraceHealth } from '../types-B4zlO58J.mjs';
-import '../ir-9rrfI3yB.mjs';
+import { T as TraceHealth } from '../types-Cya0jPAF.mjs';
+import '../ir-BTvyl8-w.mjs';
 import '../dialect.mjs';
 
 /**

package/dist/glassbox-routes/format.d.ts

@@ -1,5 +1,5 @@
-import { T as TraceHealth } from '../types-iwyFQHt3.js';
-import '../ir-Dx_UXPfP.js';
+import { T as TraceHealth } from '../types-DxYXuFxj.js';
+import '../ir-Qnw6L265.js';
 import '../dialect.js';
 
 /**

package/dist/glassbox-routes/index.d.mts

@@ -1,7 +1,7 @@
-import { G as GlassboxEvent } from '../types-D-G7KXeI.mjs';
-import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-B4zlO58J.mjs';
-export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-B4zlO58J.mjs';
-import '../ir-9rrfI3yB.mjs';
+import { G as GlassboxEvent } from '../types-8pWRDbUg.mjs';
+import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-Cya0jPAF.mjs';
+export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-Cya0jPAF.mjs';
+import '../ir-BTvyl8-w.mjs';
 import '../dialect.mjs';
 
 /**

package/dist/glassbox-routes/index.d.ts

@@ -1,7 +1,7 @@
-import { G as GlassboxEvent } from '../types-DSknEy2B.js';
-import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-iwyFQHt3.js';
-export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-iwyFQHt3.js';
-import '../ir-Dx_UXPfP.js';
+import { G as GlassboxEvent } from '../types-BdYlJz0Z.js';
+import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-DxYXuFxj.js';
+export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-DxYXuFxj.js';
+import '../ir-Qnw6L265.js';
 import '../dialect.js';
 
 /**

package/dist/glassbox-routes/index.mjs

@@ -1,7 +1,7 @@
 import {
   ARCHETYPE_FLOOR_DEFAULT,
   getDefaultFallbackChain
-} from "../chunk-DJVGJEJU.mjs";
+} from "../chunk-SBFSYCQG.mjs";
 import {
   tryGetProfile
 } from "../chunk-4UO4CCSP.mjs";

package/dist/glassbox-routes/react/index.d.mts

@@ -1,6 +1,6 @@
 import * as react_jsx_runtime from 'react/jsx-runtime';
-import { a as TraceDetail } from '../../types-B4zlO58J.mjs';
-import '../../ir-9rrfI3yB.mjs';
+import { a as TraceDetail } from '../../types-Cya0jPAF.mjs';
+import '../../ir-BTvyl8-w.mjs';
 import '../../dialect.mjs';
 
 /**

package/dist/glassbox-routes/react/index.d.ts

@@ -1,6 +1,6 @@
 import * as react_jsx_runtime from 'react/jsx-runtime';
-import { a as TraceDetail } from '../../types-iwyFQHt3.js';
-import '../../ir-Dx_UXPfP.js';
+import { a as TraceDetail } from '../../types-DxYXuFxj.js';
+import '../../ir-Qnw6L265.js';
 import '../../dialect.js';
 
 /**

package/dist/index.d.mts

@@ -1,5 +1,5 @@
-import { C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, S as SystemModelMessage, e as CompileResult, B as BestPracticeAdvisory, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OracleScore, i as Adapter, j as PerAxisMetrics, k as Provider, l as ChainEntry, G as Grounding } from './ir-9rrfI3yB.mjs';
-export { m as CallAttempt, n as CallError, o as ChainModelEntry, p as ChainWithGrounding, q as Constraints, E as EffortLevel, F as FallbackReason, r as GoldenCaptureOptions, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, s as MutationApplied, t as NormalizedTokens, u as OutcomeKind, v as PerAxisMetricsByModel, w as PromptSection, x as SectionKind, y as ShadowProbeConfig, T as ToolCall, z as ToolDefinition, D as captureGoldenIr, J as parseGoldenCaptureRate, K as resolveGoldenCaptureRate, L as shouldCaptureGolden } from './ir-9rrfI3yB.mjs';
+import { C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, S as SystemModelMessage, e as CompileResult, B as BestPracticeAdvisory, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OracleScore, i as Adapter, j as PerAxisMetrics, k as Provider, l as ChainEntry, G as Grounding } from './ir-BTvyl8-w.mjs';
+export { m as CallAttempt, n as CallError, o as ChainModelEntry, p as ChainWithGrounding, q as Constraints, E as EffortLevel, F as FallbackReason, r as GoldenCaptureOptions, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, s as MutationApplied, t as NormalizedTokens, u as OutcomeKind, v as PerAxisMetricsByModel, w as PromptSection, x as SectionKind, y as ShadowProbeConfig, T as ToolCall, z as ToolDefinition, D as captureGoldenIr, J as parseGoldenCaptureRate, K as resolveGoldenCaptureRate, L as shouldCaptureGolden } from './ir-BTvyl8-w.mjs';
 import { ModelProfile, ArchetypeConvention } from './profiles.mjs';
 export { ALIASES, CacheStrategy, CliffRule, LATENCY_TIER_MS, LatencyTier, LoweringSpec, RecoveryRule, StructuredOutputCapability, SystemPromptMode, allProfiles, getProfile, latencyTierOf, profilesByProvider, tryGetProfile } from './profiles.mjs';
 export { BrainForwardConfig, BrainForwardRoutes, createBrainForwardRoutes } from './brain-proxy.mjs';
@@ -78,6 +78,19 @@ interface ExecuteErr {
     errorCode: string;
     message: string;
     raw: unknown;
+    /**
+     * alpha.66 — real token usage when the failure is a SYNTHESIZED
+     * structured-output contract violation (the provider returned 200 with
+     * usage; the response just violated the contract). Absent on transport-
+     * level failures (4xx/5xx/network) where no usage exists. Lets the
+     * brain's failure rows carry true `tokens_in` for the cliff class —
+     * PB's 2026-05-18 `chain-walk-on-structured-output-parse-failed` ask 4
+     * (tokens_in=0 on a 72s parse-fail made diagnosis ambiguous).
+     */
+    tokens?: {
+        input: number;
+        output: number;
+    };
 }
 type ExecuteResult = ExecuteOk | ExecuteErr;
 declare function execute(request: CompiledRequest, opts?: ExecuteOptions): Promise<ExecuteResult>;
@@ -597,6 +610,12 @@ interface OutcomePayload {
     system_prompt_chars?: number;
     fell_over_from?: string;
     fallback_reason?: 'rate_limit' | 'provider_auth_failed' | 'provider_error' | 'cliff' | 'cost_cap' | 'contract_violation';
+    /**
+     * True when the same-model retry (CallOptions.sameModelRetry) fired
+     * during this call. Key OMITTED (not false) when no retry fired — safe
+     * against brains that haven't applied migration 039 yet.
+     */
+    retried_same_model?: boolean;
     /**
      * Per-call SectionRewrite[] captured at compile time. Omitted (sent as
      * undefined → stored NULL) when no rewrites fired. Powers cross-app
@@ -923,7 +942,7 @@ declare function runGoldenEval(opts: GoldenEvalOptions): Promise<GoldenEvalRunRe
  * guard in `tests/version.test.ts` fails the suite (and therefore
  * `prepublishOnly`) when they diverge — a stale constant cannot reach npm.
  */
-declare const LIBRARY_VERSION = "2.0.0-alpha.65";
+declare const LIBRARY_VERSION = "2.0.0-alpha.66";
 
 /**
  * Oracle contract — how an app tells the brain whether a response was good.

package/dist/index.d.ts

@@ -1,5 +1,5 @@
-import { C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, S as SystemModelMessage, e as CompileResult, B as BestPracticeAdvisory, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OracleScore, i as Adapter, j as PerAxisMetrics, k as Provider, l as ChainEntry, G as Grounding } from './ir-Dx_UXPfP.js';
-export { m as CallAttempt, n as CallError, o as ChainModelEntry, p as ChainWithGrounding, q as Constraints, E as EffortLevel, F as FallbackReason, r as GoldenCaptureOptions, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, s as MutationApplied, t as NormalizedTokens, u as OutcomeKind, v as PerAxisMetricsByModel, w as PromptSection, x as SectionKind, y as ShadowProbeConfig, T as ToolCall, z as ToolDefinition, D as captureGoldenIr, J as parseGoldenCaptureRate, K as resolveGoldenCaptureRate, L as shouldCaptureGolden } from './ir-Dx_UXPfP.js';
+import { C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, S as SystemModelMessage, e as CompileResult, B as BestPracticeAdvisory, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OracleScore, i as Adapter, j as PerAxisMetrics, k as Provider, l as ChainEntry, G as Grounding } from './ir-Qnw6L265.js';
+export { m as CallAttempt, n as CallError, o as ChainModelEntry, p as ChainWithGrounding, q as Constraints, E as EffortLevel, F as FallbackReason, r as GoldenCaptureOptions, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, s as MutationApplied, t as NormalizedTokens, u as OutcomeKind, v as PerAxisMetricsByModel, w as PromptSection, x as SectionKind, y as ShadowProbeConfig, T as ToolCall, z as ToolDefinition, D as captureGoldenIr, J as parseGoldenCaptureRate, K as resolveGoldenCaptureRate, L as shouldCaptureGolden } from './ir-Qnw6L265.js';
 import { ModelProfile, ArchetypeConvention } from './profiles.js';
 export { ALIASES, CacheStrategy, CliffRule, LATENCY_TIER_MS, LatencyTier, LoweringSpec, RecoveryRule, StructuredOutputCapability, SystemPromptMode, allProfiles, getProfile, latencyTierOf, profilesByProvider, tryGetProfile } from './profiles.js';
 export { BrainForwardConfig, BrainForwardRoutes, createBrainForwardRoutes } from './brain-proxy.js';
@@ -78,6 +78,19 @@ interface ExecuteErr {
     errorCode: string;
     message: string;
     raw: unknown;
+    /**
+     * alpha.66 — real token usage when the failure is a SYNTHESIZED
+     * structured-output contract violation (the provider returned 200 with
+     * usage; the response just violated the contract). Absent on transport-
+     * level failures (4xx/5xx/network) where no usage exists. Lets the
+     * brain's failure rows carry true `tokens_in` for the cliff class —
+     * PB's 2026-05-18 `chain-walk-on-structured-output-parse-failed` ask 4
+     * (tokens_in=0 on a 72s parse-fail made diagnosis ambiguous).
+     */
+    tokens?: {
+        input: number;
+        output: number;
+    };
 }
 type ExecuteResult = ExecuteOk | ExecuteErr;
 declare function execute(request: CompiledRequest, opts?: ExecuteOptions): Promise<ExecuteResult>;
@@ -597,6 +610,12 @@ interface OutcomePayload {
     system_prompt_chars?: number;
     fell_over_from?: string;
     fallback_reason?: 'rate_limit' | 'provider_auth_failed' | 'provider_error' | 'cliff' | 'cost_cap' | 'contract_violation';
+    /**
+     * True when the same-model retry (CallOptions.sameModelRetry) fired
+     * during this call. Key OMITTED (not false) when no retry fired — safe
+     * against brains that haven't applied migration 039 yet.
+     */
+    retried_same_model?: boolean;
     /**
      * Per-call SectionRewrite[] captured at compile time. Omitted (sent as
      * undefined → stored NULL) when no rewrites fired. Powers cross-app
@@ -923,7 +942,7 @@ declare function runGoldenEval(opts: GoldenEvalOptions): Promise<GoldenEvalRunRe
  * guard in `tests/version.test.ts` fails the suite (and therefore
  * `prepublishOnly`) when they diverge — a stale constant cannot reach npm.
  */
-declare const LIBRARY_VERSION = "2.0.0-alpha.65";
+declare const LIBRARY_VERSION = "2.0.0-alpha.66";
 
 /**
  * Oracle contract — how an app tells the brain whether a response was good.

package/dist/index.js

@@ -5251,6 +5251,9 @@ function buildPayload(input, reg) {
     system_prompt_chars: input.systemPromptChars ?? reg?.systemPromptChars,
     fell_over_from: fellOverFrom,
     fallback_reason: fallbackReason,
+    // alpha.66 — omitted-undefined keeps the key out of the JSON body
+    // entirely for non-retry rows (pre-migration-039 brain safety).
+    retried_same_model: input.retriedSameModel,
     // alpha.29 — translator activity (migration 019). Send NULL when no
     // rewrites fired so the brain's "did the translator do anything?"
     // queries can use `IS NOT NULL` cleanly.
@@ -5607,6 +5610,10 @@ function readBrainReadEnv(envSource = defaultEnv()) {
   if (!anonKey) missingEnv.push("KGAUTO_V2_BRAIN_ANON_KEY");
   return { endpoint, jwt, anonKey, missingEnv };
 }
+function isSameModelRetryEnabledFromEnv(envSource = defaultEnv()) {
+  const raw = (envSource.KGAUTO_SAME_MODEL_RETRY ?? "").trim().toLowerCase();
+  return raw === "1" || raw === "true";
+}
 
 // src/streaming.ts
 var ANTHROPIC_URL = "https://api.anthropic.com/v1/messages";
@@ -6958,6 +6965,8 @@ async function call(ir, opts = {}) {
   let activeCompile = initial;
   let lastErr;
   const failedProviders = /* @__PURE__ */ new Set();
+  const sameModelRetryEnabled = opts.sameModelRetry ?? isSameModelRetryEnabledFromEnv();
+  let retriedSameModel = false;
   for (let i = 0; i < targetsToTry.length; i++) {
     const targetModel = targetsToTry[i];
     const targetProfile = tryGetProfile(targetModel);
@@ -6995,15 +7004,37 @@ async function call(ir, opts = {}) {
     );
     const targetSupportsStreaming = targetProfile?.streaming === true;
     const streamingOnChunk = opts.onChunk && !opts.noStream && targetSupportsStreaming ? opts.onChunk : void 0;
-    const exec = await execute(activeCompile.request, {
+    const execOpts = {
       apiKeys: opts.apiKeys,
       fetchImpl: opts.fetchImpl,
       providerOverrides: opts.providerOverrides,
       onChunk: streamingOnChunk
-    });
-    const validated = exec.ok ? validateStructuredContract(exec, ir) : exec;
+    };
+    const exec = await execute(activeCompile.request, execOpts);
+    let validated = exec.ok ? validateStructuredContract(exec, ir) : exec;
+    let servedByRetry = false;
+    if (!validated.ok && isStructuredContractViolation(validated.errorCode) && sameModelRetryEnabled && !retriedSameModel) {
+      retriedSameModel = true;
+      attempts.push({
+        model: targetModel,
+        status: validated.errorType,
+        errorCode: validated.errorCode,
+        message: validated.message
+      });
+      const retryRequest = validated.errorCode === "max_tokens_on_structured_output" ? raiseOutputBudget(activeCompile.request, targetProfile) : activeCompile.request;
+      safeEmit(
+        () => emitExecuteAttempt(traceId, ir.appId, { model: targetModel, attemptIndex: i })
+      );
+      const retryExec = await execute(retryRequest, execOpts);
+      validated = retryExec.ok ? validateStructuredContract(retryExec, ir) : retryExec;
+      servedByRetry = true;
+    }
     if (validated.ok) {
-      attempts.push({ model: targetModel, status: "success" });
+      attempts.push({
+        model: targetModel,
+        status: "success",
+        ...servedByRetry ? { sameModelRetry: true } : {}
+      });
       const latencyMs2 = Date.now() - start;
       safeEmit(
         () => emitExecuteSuccess(traceId, ir.appId, {
@@ -7040,7 +7071,10 @@ async function call(ir, opts = {}) {
         finishReason: validated.response.finishReason,
         totalMs: latencyMs2,
         fellOverFrom: fellOver ? initial.target : void 0,
-        fallbackReason
+        fallbackReason,
+        // alpha.66 — migration 039. Omitted (not false) when no retry fired
+        // so pre-039 brains never see the key.
+        retriedSameModel: retriedSameModel || void 0
       });
       if (fellOver) {
         const firstFailed = attempts.find((a) => a.status !== "success");
@@ -7098,6 +7132,7 @@ async function call(ir, opts = {}) {
         servedBy: targetModel,
         fellOverFrom: fellOver ? initial.target : void 0,
         fallbackReason,
+        retriedSameModel: retriedSameModel || void 0,
         unreachableFiltered,
         policyBlockedFiltered,
         traceId,
@@ -7112,7 +7147,8 @@ async function call(ir, opts = {}) {
       model: targetModel,
       status: validated.errorType,
       errorCode: validated.errorCode,
-      message: validated.message
+      message: validated.message,
+      ...servedByRetry ? { sameModelRetry: true } : {}
     });
     lastErr = validated;
     if (validated.errorType === "terminal" || opts.noFallback) {
@@ -7123,15 +7159,20 @@ async function call(ir, opts = {}) {
       break;
     }
   }
+  const lastAttempted = [...attempts].reverse().find(
+    (a) => a.status !== "success" && a.errorCode !== "auth_inferred" && a.errorCode !== "compile_error"
+  );
   const latencyMs = Date.now() - start;
   await record({
     handle: initial.handle,
-    tokensIn: 0,
-    tokensOut: 0,
+    tokensIn: lastErr?.tokens?.input ?? 0,
+    tokensOut: lastErr?.tokens?.output ?? 0,
     latencyMs,
     success: false,
     errorType: lastErr?.errorCode,
-    promptPreview: extractPromptPreview(ir)
+    promptPreview: extractPromptPreview(ir),
+    actualModel: lastAttempted?.model,
+    retriedSameModel: retriedSameModel || void 0
   });
   const filteredNote = unreachableFiltered && unreachableFiltered.length > 0 ? ` (also auto-filtered: [${unreachableFiltered.join(", ")}] \u2014 no API key)` : "";
   const blockedNote = policyBlockedFiltered && policyBlockedFiltered.length > 0 ? ` (also policy-blocked: [${policyBlockedFiltered.join(", ")}])` : "";
@@ -7376,7 +7417,10 @@ function validateStructuredContract(exec, ir) {
       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
+      raw: exec.response.raw,
+      // alpha.66 — the provider DID serve this call; carry the real usage so
+      // exhausted-chain failure rows aren't stamped tokens_in=0 (PB 05-18 ask 4).
+      tokens: { input: exec.response.tokens.input, output: exec.response.tokens.output }
     };
   }
   if (!exec.response.text) {
@@ -7392,10 +7436,35 @@ function validateStructuredContract(exec, ir) {
       errorType: "retryable",
       errorCode: "structured_output_parse_failed",
       message: err instanceof Error ? err.message : String(err),
-      raw: exec.response.raw
+      raw: exec.response.raw,
+      tokens: { input: exec.response.tokens.input, output: exec.response.tokens.output }
     };
   }
 }
+function isStructuredContractViolation(code) {
+  return code === "max_tokens_on_structured_output" || code === "structured_output_parse_failed";
+}
+function raiseOutputBudget(request, profile) {
+  if (!profile) return request;
+  if (request.provider === "anthropic") {
+    if (typeof request.max_tokens === "number" && request.max_tokens < profile.maxOutputTokens) {
+      return { ...request, max_tokens: profile.maxOutputTokens };
+    }
+    return request;
+  }
+  if (request.provider === "google") {
+    const gc = request.generationConfig;
+    const cap = gc?.maxOutputTokens;
+    if (typeof cap === "number" && cap < profile.maxOutputTokens) {
+      return {
+        ...request,
+        generationConfig: { ...gc, maxOutputTokens: profile.maxOutputTokens }
+      };
+    }
+    return request;
+  }
+  return request;
+}
 function normalizeFallbackReason(attempts) {
   const first = attempts.find((a) => a.status !== "success");
   if (!first) return void 0;
@@ -7428,6 +7497,8 @@ function safeEmit(fn) {
 var _internal = {
   compileAndRegister,
   validateStructuredContract,
+  isStructuredContractViolation,
+  raiseOutputBudget,
   extractPromptPreview,
   normalizeFallbackReason,
   generateTraceId,
@@ -8216,7 +8287,7 @@ function createBrainForwardRoutes(config) {
 }
 
 // src/version.ts
-var LIBRARY_VERSION = "2.0.0-alpha.65";
+var LIBRARY_VERSION = "2.0.0-alpha.66";
 
 // src/key-health.ts
 var JSON_HEADERS2 = { "Content-Type": "application/json" };

package/dist/index.mjs

@@ -15,7 +15,7 @@ import {
 import {
   LIBRARY_VERSION,
   createKeyHealthRoute
-} from "./chunk-XTZ2ZQJQ.mjs";
+} from "./chunk-CMB6PZWS.mjs";
 import {
   ABSOLUTE_FLOOR,
   ARCHETYPE_FLOOR_DEFAULT,
@@ -37,10 +37,11 @@ import {
   isBrainQueryActiveFor,
   isModelReachable,
   isProviderReachable,
+  isSameModelRetryEnabledFromEnv,
   loadChainsFromBrain,
   readBrainReadEnv,
   resolveProviderKey
-} from "./chunk-DJVGJEJU.mjs";
+} from "./chunk-SBFSYCQG.mjs";
 import {
   ALIASES,
   LATENCY_TIER_MS,
@@ -3231,6 +3232,9 @@ function buildPayload(input, reg) {
     system_prompt_chars: input.systemPromptChars ?? reg?.systemPromptChars,
     fell_over_from: fellOverFrom,
     fallback_reason: fallbackReason,
+    // alpha.66 — omitted-undefined keeps the key out of the JSON body
+    // entirely for non-retry rows (pre-migration-039 brain safety).
+    retried_same_model: input.retriedSameModel,
     // alpha.29 — translator activity (migration 019). Send NULL when no
     // rewrites fired so the brain's "did the translator do anything?"
     // queries can use `IS NOT NULL` cleanly.
@@ -4203,6 +4207,8 @@ async function call(ir, opts = {}) {
   let activeCompile = initial;
   let lastErr;
   const failedProviders = /* @__PURE__ */ new Set();
+  const sameModelRetryEnabled = opts.sameModelRetry ?? isSameModelRetryEnabledFromEnv();
+  let retriedSameModel = false;
   for (let i = 0; i < targetsToTry.length; i++) {
     const targetModel = targetsToTry[i];
     const targetProfile = tryGetProfile(targetModel);
@@ -4240,15 +4246,37 @@ async function call(ir, opts = {}) {
     );
     const targetSupportsStreaming = targetProfile?.streaming === true;
     const streamingOnChunk = opts.onChunk && !opts.noStream && targetSupportsStreaming ? opts.onChunk : void 0;
-    const exec = await execute(activeCompile.request, {
+    const execOpts = {
       apiKeys: opts.apiKeys,
       fetchImpl: opts.fetchImpl,
       providerOverrides: opts.providerOverrides,
       onChunk: streamingOnChunk
-    });
-    const validated = exec.ok ? validateStructuredContract(exec, ir) : exec;
+    };
+    const exec = await execute(activeCompile.request, execOpts);
+    let validated = exec.ok ? validateStructuredContract(exec, ir) : exec;
+    let servedByRetry = false;
+    if (!validated.ok && isStructuredContractViolation(validated.errorCode) && sameModelRetryEnabled && !retriedSameModel) {
+      retriedSameModel = true;
+      attempts.push({
+        model: targetModel,
+        status: validated.errorType,
+        errorCode: validated.errorCode,
+        message: validated.message
+      });
+      const retryRequest = validated.errorCode === "max_tokens_on_structured_output" ? raiseOutputBudget(activeCompile.request, targetProfile) : activeCompile.request;
+      safeEmit(
+        () => emitExecuteAttempt(traceId, ir.appId, { model: targetModel, attemptIndex: i })
+      );
+      const retryExec = await execute(retryRequest, execOpts);
+      validated = retryExec.ok ? validateStructuredContract(retryExec, ir) : retryExec;
+      servedByRetry = true;
+    }
     if (validated.ok) {
-      attempts.push({ model: targetModel, status: "success" });
+      attempts.push({
+        model: targetModel,
+        status: "success",
+        ...servedByRetry ? { sameModelRetry: true } : {}
+      });
       const latencyMs2 = Date.now() - start;
       safeEmit(
         () => emitExecuteSuccess(traceId, ir.appId, {
@@ -4285,7 +4313,10 @@ async function call(ir, opts = {}) {
         finishReason: validated.response.finishReason,
         totalMs: latencyMs2,
         fellOverFrom: fellOver ? initial.target : void 0,
-        fallbackReason
+        fallbackReason,
+        // alpha.66 — migration 039. Omitted (not false) when no retry fired
+        // so pre-039 brains never see the key.
+        retriedSameModel: retriedSameModel || void 0
       });
       if (fellOver) {
         const firstFailed = attempts.find((a) => a.status !== "success");
@@ -4343,6 +4374,7 @@ async function call(ir, opts = {}) {
         servedBy: targetModel,
         fellOverFrom: fellOver ? initial.target : void 0,
         fallbackReason,
+        retriedSameModel: retriedSameModel || void 0,
         unreachableFiltered,
         policyBlockedFiltered,
         traceId,
@@ -4357,7 +4389,8 @@ async function call(ir, opts = {}) {
       model: targetModel,
       status: validated.errorType,
       errorCode: validated.errorCode,
-      message: validated.message
+      message: validated.message,
+      ...servedByRetry ? { sameModelRetry: true } : {}
     });
     lastErr = validated;
     if (validated.errorType === "terminal" || opts.noFallback) {
@@ -4368,15 +4401,20 @@ async function call(ir, opts = {}) {
       break;
     }
   }
+  const lastAttempted = [...attempts].reverse().find(
+    (a) => a.status !== "success" && a.errorCode !== "auth_inferred" && a.errorCode !== "compile_error"
+  );
   const latencyMs = Date.now() - start;
   await record({
     handle: initial.handle,
-    tokensIn: 0,
-    tokensOut: 0,
+    tokensIn: lastErr?.tokens?.input ?? 0,
+    tokensOut: lastErr?.tokens?.output ?? 0,
     latencyMs,
     success: false,
     errorType: lastErr?.errorCode,
-    promptPreview: extractPromptPreview(ir)
+    promptPreview: extractPromptPreview(ir),
+    actualModel: lastAttempted?.model,
+    retriedSameModel: retriedSameModel || void 0
   });
   const filteredNote = unreachableFiltered && unreachableFiltered.length > 0 ? ` (also auto-filtered: [${unreachableFiltered.join(", ")}] \u2014 no API key)` : "";
   const blockedNote = policyBlockedFiltered && policyBlockedFiltered.length > 0 ? ` (also policy-blocked: [${policyBlockedFiltered.join(", ")}])` : "";
@@ -4621,7 +4659,10 @@ function validateStructuredContract(exec, ir) {
       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
+      raw: exec.response.raw,
+      // alpha.66 — the provider DID serve this call; carry the real usage so
+      // exhausted-chain failure rows aren't stamped tokens_in=0 (PB 05-18 ask 4).
+      tokens: { input: exec.response.tokens.input, output: exec.response.tokens.output }
     };
   }
   if (!exec.response.text) {
@@ -4637,10 +4678,35 @@ function validateStructuredContract(exec, ir) {
       errorType: "retryable",
       errorCode: "structured_output_parse_failed",
       message: err instanceof Error ? err.message : String(err),
-      raw: exec.response.raw
+      raw: exec.response.raw,
+      tokens: { input: exec.response.tokens.input, output: exec.response.tokens.output }
     };
   }
 }
+function isStructuredContractViolation(code) {
+  return code === "max_tokens_on_structured_output" || code === "structured_output_parse_failed";
+}
+function raiseOutputBudget(request, profile) {
+  if (!profile) return request;
+  if (request.provider === "anthropic") {
+    if (typeof request.max_tokens === "number" && request.max_tokens < profile.maxOutputTokens) {
+      return { ...request, max_tokens: profile.maxOutputTokens };
+    }
+    return request;
+  }
+  if (request.provider === "google") {
+    const gc = request.generationConfig;
+    const cap = gc?.maxOutputTokens;
+    if (typeof cap === "number" && cap < profile.maxOutputTokens) {
+      return {
+        ...request,
+        generationConfig: { ...gc, maxOutputTokens: profile.maxOutputTokens }
+      };
+    }
+    return request;
+  }
+  return request;
+}
 function normalizeFallbackReason(attempts) {
   const first = attempts.find((a) => a.status !== "success");
   if (!first) return void 0;
@@ -4673,6 +4739,8 @@ function safeEmit(fn) {
 var _internal = {
   compileAndRegister,
   validateStructuredContract,
+  isStructuredContractViolation,
+  raiseOutputBudget,
   extractPromptPreview,
   normalizeFallbackReason,
   generateTraceId,

package/dist/{ir-9rrfI3yB.d.mts → ir-BTvyl8-w.d.mts}

@@ -1065,12 +1065,40 @@ interface CallOptions {
      * capture never fires.
      */
     goldenCapture?: GoldenCaptureOptions;
+    /**
+     * alpha.66 — one same-model retry on a structured-output contract violation
+     * (`max_tokens_on_structured_output` / `structured_output_parse_failed`)
+     * BEFORE chain-walking to a lower-ranked tier. Filed by playbacksam
+     * 2026-07-18 (`structured-output-truncation-retry-same-model`): on PB's
+     * summarize surface the walk destination is measured quality-inferior
+     * (golden eval run #3), so one more attempt at the leader beats walking.
+     *
+     * Consent posture (matches KGAUTO_AUTO_PROMOTE / KGAUTO_GOLDEN_CAPTURE):
+     * default OFF; `true` wins, explicit `false` beats the env, undefined
+     * falls through to the `KGAUTO_SAME_MODEL_RETRY` env var ('1'/'true').
+     * Default-off because the retry adds one full model round-trip of latency
+     * in the failure class — the consumer decides whether their fallback tier
+     * is bad enough to pay that.
+     *
+     * Bounds: at most ONE retry per call(), on whichever target hits the
+     * contract violation first. Truncation-class retries additionally raise
+     * any explicit wire output cap sitting below the profile's
+     * `maxOutputTokens` (e.g. a terse-clamp or providerOverrides cap);
+     * requests already at the profile ceiling re-roll unchanged.
+     */
+    sameModelRetry?: boolean;
 }
 interface CallAttempt {
     model: string;
     status: 'success' | 'retryable' | 'terminal';
     errorCode?: string;
     message?: string;
+    /**
+     * alpha.66 — true on the attempt row that was the one-shot same-model
+     * retry (CallOptions.sameModelRetry). The preceding row for the same model
+     * carries the contract violation that triggered it.
+     */
+    sameModelRetry?: boolean;
 }
 /**
  * Why fallback fired. Normalized for `CallResult.fallbackReason` (alpha.9).
@@ -1126,6 +1154,14 @@ interface CallResult {
     fellOverFrom?: string;
     /** Set only when fallback fired. Normalized cause. */
     fallbackReason?: FallbackReason;
+    /**
+     * alpha.66 — true when the one-shot same-model retry
+     * (CallOptions.sameModelRetry) fired during this call, regardless of
+     * whether the retry rescued the call or the chain walked afterwards.
+     * Cross-check `attempts[]` (the `sameModelRetry: true` row) and
+     * `fellOverFrom` to distinguish rescued-by-retry from retried-then-walked.
+     */
+    retriedSameModel?: boolean;
     /**
      * alpha.10. Models that auto-filter dropped from the fallback walk because
      * their provider had no reachable API key. Empty when nothing was filtered
@@ -1323,6 +1359,13 @@ interface RecordInput {
      * keep in sync with the wire-contract enum (TraceDetail.fallbackReason).
      */
     fallbackReason?: 'rate_limit' | 'provider_auth_failed' | 'provider_error' | 'cliff' | 'cost_cap' | 'contract_violation';
+    /**
+     * alpha.66 — true when the one-shot same-model retry fired during this
+     * call (migration 039 `retried_same_model`). Powers the offline rollup:
+     * retry rate on the cliff class, and — joined with `fell_over_from` —
+     * how often the retry rescued the call vs the chain walking anyway.
+     */
+    retriedSameModel?: boolean;
 }
 /**
  * alpha.20 Entry 4: kinds of consumer-declared outcomes feeding the quality

package/dist/{ir-Dx_UXPfP.d.ts → ir-Qnw6L265.d.ts}

@@ -1065,12 +1065,40 @@ interface CallOptions {
      * capture never fires.
      */
     goldenCapture?: GoldenCaptureOptions;
+    /**
+     * alpha.66 — one same-model retry on a structured-output contract violation
+     * (`max_tokens_on_structured_output` / `structured_output_parse_failed`)
+     * BEFORE chain-walking to a lower-ranked tier. Filed by playbacksam
+     * 2026-07-18 (`structured-output-truncation-retry-same-model`): on PB's
+     * summarize surface the walk destination is measured quality-inferior
+     * (golden eval run #3), so one more attempt at the leader beats walking.
+     *
+     * Consent posture (matches KGAUTO_AUTO_PROMOTE / KGAUTO_GOLDEN_CAPTURE):
+     * default OFF; `true` wins, explicit `false` beats the env, undefined
+     * falls through to the `KGAUTO_SAME_MODEL_RETRY` env var ('1'/'true').
+     * Default-off because the retry adds one full model round-trip of latency
+     * in the failure class — the consumer decides whether their fallback tier
+     * is bad enough to pay that.
+     *
+     * Bounds: at most ONE retry per call(), on whichever target hits the
+     * contract violation first. Truncation-class retries additionally raise
+     * any explicit wire output cap sitting below the profile's
+     * `maxOutputTokens` (e.g. a terse-clamp or providerOverrides cap);
+     * requests already at the profile ceiling re-roll unchanged.
+     */
+    sameModelRetry?: boolean;
 }
 interface CallAttempt {
     model: string;
     status: 'success' | 'retryable' | 'terminal';
     errorCode?: string;
     message?: string;
+    /**
+     * alpha.66 — true on the attempt row that was the one-shot same-model
+     * retry (CallOptions.sameModelRetry). The preceding row for the same model
+     * carries the contract violation that triggered it.
+     */
+    sameModelRetry?: boolean;
 }
 /**
  * Why fallback fired. Normalized for `CallResult.fallbackReason` (alpha.9).
@@ -1126,6 +1154,14 @@ interface CallResult {
     fellOverFrom?: string;
     /** Set only when fallback fired. Normalized cause. */
     fallbackReason?: FallbackReason;
+    /**
+     * alpha.66 — true when the one-shot same-model retry
+     * (CallOptions.sameModelRetry) fired during this call, regardless of
+     * whether the retry rescued the call or the chain walked afterwards.
+     * Cross-check `attempts[]` (the `sameModelRetry: true` row) and
+     * `fellOverFrom` to distinguish rescued-by-retry from retried-then-walked.
+     */
+    retriedSameModel?: boolean;
     /**
      * alpha.10. Models that auto-filter dropped from the fallback walk because
      * their provider had no reachable API key. Empty when nothing was filtered
@@ -1323,6 +1359,13 @@ interface RecordInput {
      * keep in sync with the wire-contract enum (TraceDetail.fallbackReason).
      */
     fallbackReason?: 'rate_limit' | 'provider_auth_failed' | 'provider_error' | 'cliff' | 'cost_cap' | 'contract_violation';
+    /**
+     * alpha.66 — true when the one-shot same-model retry fired during this
+     * call (migration 039 `retried_same_model`). Powers the offline rollup:
+     * retry rate on the cliff class, and — joined with `fell_over_from` —
+     * how often the retry rescued the call vs the chain walking anyway.
+     */
+    retriedSameModel?: boolean;
 }
 /**
  * alpha.20 Entry 4: kinds of consumer-declared outcomes feeding the quality

package/dist/key-health.js

@@ -25,7 +25,7 @@ __export(key_health_exports, {
 module.exports = __toCommonJS(key_health_exports);
 
 // src/version.ts
-var LIBRARY_VERSION = "2.0.0-alpha.65";
+var LIBRARY_VERSION = "2.0.0-alpha.66";
 
 // src/key-health.ts
 var JSON_HEADERS = { "Content-Type": "application/json" };

package/dist/key-health.mjs

@@ -1,6 +1,6 @@
 import {
   createKeyHealthRoute
-} from "./chunk-XTZ2ZQJQ.mjs";
+} from "./chunk-CMB6PZWS.mjs";
 export {
   createKeyHealthRoute
 };

package/dist/profiles.d.mts

@@ -1,4 +1,4 @@
-import { k as Provider } from './ir-9rrfI3yB.mjs';
+import { k as Provider } from './ir-BTvyl8-w.mjs';
 import { IntentArchetypeName } from './dialect.mjs';
 
 /**

package/dist/profiles.d.ts

@@ -1,4 +1,4 @@
-import { k as Provider } from './ir-Dx_UXPfP.js';
+import { k as Provider } from './ir-Qnw6L265.js';
 import { IntentArchetypeName } from './dialect.js';
 
 /**

package/dist/{types-D-G7KXeI.d.mts → types-8pWRDbUg.d.mts}

@@ -1,4 +1,4 @@
-import { s as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, m as CallAttempt } from './ir-9rrfI3yB.mjs';
+import { s as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, m as CallAttempt } from './ir-BTvyl8-w.mjs';
 
 /**
  * Glass-Box observability types (alpha.17).

package/dist/{types-DSknEy2B.d.ts → types-BdYlJz0Z.d.ts}

@@ -1,4 +1,4 @@
-import { s as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, m as CallAttempt } from './ir-Dx_UXPfP.js';
+import { s as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, m as CallAttempt } from './ir-Qnw6L265.js';
 
 /**
  * Glass-Box observability types (alpha.17).

package/dist/{types-B4zlO58J.d.mts → types-Cya0jPAF.d.mts}

@@ -1,4 +1,4 @@
-import { i as Adapter, x as SectionKind } from './ir-9rrfI3yB.mjs';
+import { i as Adapter, x as SectionKind } from './ir-BTvyl8-w.mjs';
 
 /**
  * Internal config + hook types for createGlassboxRoutes().

package/dist/{types-iwyFQHt3.d.ts → types-DxYXuFxj.d.ts}

@@ -1,4 +1,4 @@
-import { i as Adapter, x as SectionKind } from './ir-Dx_UXPfP.js';
+import { i as Adapter, x as SectionKind } from './ir-Qnw6L265.js';
 
 /**
  * Internal config + hook types for createGlassboxRoutes().

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@warmdrift/kgauto-compiler",
-  "version": "2.0.0-alpha.65",
+  "version": "2.0.0-alpha.66",
   "description": "Prompt compiler with executable provider knowledge for multi-model AI apps: normalized multi-provider transport with fallback chains, compile-time cliff guards, a curated model registry, and a telemetry flight recorder. Swap models without rewriting prompts.",
   "main": "./dist/index.js",
   "module": "./dist/index.mjs",