@warmdrift/kgauto-compiler 2.0.0-alpha.90 → 2.0.0-alpha.92

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 FallbackReason, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OutcomeSource, i as OracleScore, j as Adapter, k as PerAxisMetrics, l as Provider, m as ChainEntry, G as Grounding } from './ir-DeYMLWge.mjs';
-export { n as CallAttempt, o as CallError, p as ChainModelEntry, q as ChainWithGrounding, r as Constraints, E as EffortLevel, s as GoldenCaptureOptions, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, t as MutationApplied, u as NormalizedTokens, v as OutcomeKind, w as PerAxisMetricsByModel, x as PromptSection, y as SectionKind, z as ShadowProbeConfig, T as ToolCall, D as ToolDefinition, J as captureGoldenIr, K as hasMutation, L as mutationId, Q as parseGoldenCaptureRate, U as resolveGoldenCaptureRate, V as shouldCaptureGolden } from './ir-DeYMLWge.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 FallbackReason, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OutcomeSource, i as OracleScore, j as Adapter, k as PerAxisMetrics, l as Provider, m as ChainEntry, G as Grounding } from './ir-BPYh68mv.mjs';
+export { n as CallAttempt, o as CallError, p as ChainModelEntry, q as ChainWithGrounding, r as Constraints, E as EffortLevel, s as GoldenCaptureOptions, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, t as MutationApplied, u as NormalizedTokens, v as OutcomeKind, w as PerAxisMetricsByModel, x as PromptSection, y as SectionKind, z as ShadowProbeConfig, T as ToolCall, D as ToolDefinition, J as captureGoldenIr, K as hasMutation, L as mutationId, Q as parseGoldenCaptureRate, U as resolveGoldenCaptureRate, V as shouldCaptureGolden } from './ir-BPYh68mv.mjs';
 import { ModelProfile, ArchetypeConvention } from './profiles.mjs';
 export { ALIASES, CacheStrategy, CliffRule, LATENCY_TIER_MS, LatencyTier, LoweringSpec, RecoveryRule, StructuredOutputCapability, SystemPromptMode, _testClearConsumerProfiles, allProfiles, bestEffortProfile, getProfile, inferProviderFromId, latencyTierOf, profilesByProvider, registerProfiles, resolveModelAlias, tryGetProfile } from './profiles.mjs';
 export { BrainForwardConfig, BrainForwardRoutes, createBrainForwardRoutes } from './brain-proxy.mjs';
@@ -47,6 +47,13 @@ interface CompileOptions {
      * time. `registerCompile` reads it to derive the linkage. Absent ⇒ root call.
      */
     parentHandle?: string;
+    /**
+     * alpha.92 — lower for a gateway route (execute-leg contract). Threaded
+     * from `CallOptions.route`. The canonical profile drives every pass
+     * (cliffs, scoring, budgets) unchanged; only the wire destination and the
+     * learning-key encoding change. Absent ⇒ direct vendor wire.
+     */
+    route?: 'direct' | 'openrouter';
 }
 
 /**
@@ -666,6 +673,10 @@ interface OutcomePayload {
     provider?: string;
     shape_key?: string;
     learning_key?: string;
+    /** alpha.92 (migration 067) — route identity: 'direct' | 'openrouter'. */
+    route?: string;
+    /** alpha.92 — gateway-reported serving host (OpenRouter `provider` field). */
+    route_host?: string;
     mutations_applied: string[];
     tokens_in: number;
     tokens_out: number;
@@ -1270,7 +1281,7 @@ declare function runStrategyEvalWithAttribution(opts: Omit<GoldenEvalOptions, 'a
  * 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.90";
+declare const LIBRARY_VERSION = "2.0.0-alpha.92";
 
 /**
  * Oracle contract — how an app tells the brain whether a response was good.
@@ -2668,7 +2679,7 @@ declare function getPerAxisMetrics(opts: GetPerAxisMetricsOpts): Promise<PerAxis
  * no execute() handler, no env-var convention). Narrowing here keeps the
  * reachability check structurally honest.
  */
-type SupportedProvider = 'anthropic' | 'google' | 'openai' | 'deepseek' | 'zai' | 'moonshot';
+type SupportedProvider = 'anthropic' | 'google' | 'openai' | 'deepseek' | 'zai' | 'moonshot' | 'openrouter';
 /**
  * Per-provider env var names kgauto recognizes. Order doesn't matter —
  * first-present wins. Multiple names per provider supported because Google

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 FallbackReason, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OutcomeSource, i as OracleScore, j as Adapter, k as PerAxisMetrics, l as Provider, m as ChainEntry, G as Grounding } from './ir-CTx026t0.js';
-export { n as CallAttempt, o as CallError, p as ChainModelEntry, q as ChainWithGrounding, r as Constraints, E as EffortLevel, s as GoldenCaptureOptions, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, t as MutationApplied, u as NormalizedTokens, v as OutcomeKind, w as PerAxisMetricsByModel, x as PromptSection, y as SectionKind, z as ShadowProbeConfig, T as ToolCall, D as ToolDefinition, J as captureGoldenIr, K as hasMutation, L as mutationId, Q as parseGoldenCaptureRate, U as resolveGoldenCaptureRate, V as shouldCaptureGolden } from './ir-CTx026t0.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 FallbackReason, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OutcomeSource, i as OracleScore, j as Adapter, k as PerAxisMetrics, l as Provider, m as ChainEntry, G as Grounding } from './ir-BWnE6LaB.js';
+export { n as CallAttempt, o as CallError, p as ChainModelEntry, q as ChainWithGrounding, r as Constraints, E as EffortLevel, s as GoldenCaptureOptions, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, t as MutationApplied, u as NormalizedTokens, v as OutcomeKind, w as PerAxisMetricsByModel, x as PromptSection, y as SectionKind, z as ShadowProbeConfig, T as ToolCall, D as ToolDefinition, J as captureGoldenIr, K as hasMutation, L as mutationId, Q as parseGoldenCaptureRate, U as resolveGoldenCaptureRate, V as shouldCaptureGolden } from './ir-BWnE6LaB.js';
 import { ModelProfile, ArchetypeConvention } from './profiles.js';
 export { ALIASES, CacheStrategy, CliffRule, LATENCY_TIER_MS, LatencyTier, LoweringSpec, RecoveryRule, StructuredOutputCapability, SystemPromptMode, _testClearConsumerProfiles, allProfiles, bestEffortProfile, getProfile, inferProviderFromId, latencyTierOf, profilesByProvider, registerProfiles, resolveModelAlias, tryGetProfile } from './profiles.js';
 export { BrainForwardConfig, BrainForwardRoutes, createBrainForwardRoutes } from './brain-proxy.js';
@@ -47,6 +47,13 @@ interface CompileOptions {
      * time. `registerCompile` reads it to derive the linkage. Absent ⇒ root call.
      */
     parentHandle?: string;
+    /**
+     * alpha.92 — lower for a gateway route (execute-leg contract). Threaded
+     * from `CallOptions.route`. The canonical profile drives every pass
+     * (cliffs, scoring, budgets) unchanged; only the wire destination and the
+     * learning-key encoding change. Absent ⇒ direct vendor wire.
+     */
+    route?: 'direct' | 'openrouter';
 }
 
 /**
@@ -666,6 +673,10 @@ interface OutcomePayload {
     provider?: string;
     shape_key?: string;
     learning_key?: string;
+    /** alpha.92 (migration 067) — route identity: 'direct' | 'openrouter'. */
+    route?: string;
+    /** alpha.92 — gateway-reported serving host (OpenRouter `provider` field). */
+    route_host?: string;
     mutations_applied: string[];
     tokens_in: number;
     tokens_out: number;
@@ -1270,7 +1281,7 @@ declare function runStrategyEvalWithAttribution(opts: Omit<GoldenEvalOptions, 'a
  * 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.90";
+declare const LIBRARY_VERSION = "2.0.0-alpha.92";
 
 /**
  * Oracle contract — how an app tells the brain whether a response was good.
@@ -2668,7 +2679,7 @@ declare function getPerAxisMetrics(opts: GetPerAxisMetricsOpts): Promise<PerAxis
  * no execute() handler, no env-var convention). Narrowing here keeps the
  * reachability check structurally honest.
  */
-type SupportedProvider = 'anthropic' | 'google' | 'openai' | 'deepseek' | 'zai' | 'moonshot';
+type SupportedProvider = 'anthropic' | 'google' | 'openai' | 'deepseek' | 'zai' | 'moonshot' | 'openrouter';
 /**
  * Per-provider env var names kgauto recognizes. Order doesn't matter —
  * first-present wins. Multiple names per provider supported because Google

package/dist/index.js

@@ -889,9 +889,19 @@ var PROFILES_RAW = [
     // thresholds. The brain will validate/relax these as evidence accumulates
     // per (archetype, model) tuple. Currently ZERO brain rows for this model.
     id: "gemini-2.5-flash-lite",
-    verifiedAgainstDocs: "2026-08-14",
+    verifiedAgainstDocs: "2026-08-27",
     provider: "google",
-    status: "current",
+    // alpha.91 — AVAILABILITY demotion, same class as gemini-2.5-pro below.
+    // Google refuses the model for newer account cohorts: "This model is no
+    // longer available to new users. Please update your code to use
+    // models/gemini-3.5-flash-lite" (tt-intel TT-49 runtime evidence,
+    // 2026-08-26, interfaces/kgauto.md ## Requested). It ranked FIRST in
+    // getFallbackChain('gemini-2.5-pro') — a dead-for-new-accounts model as
+    // the cheapest fallback pick. `legacy` removes it from chains and new
+    // selection; explicit-id consumers still resolve it. Successor is the
+    // one Google's own error names: gemini-3.5-flash-lite (onboarded this
+    // release, same price point).
+    status: "legacy",
     maxContextTokens: 1048576,
     maxOutputTokens: 65536,
     maxTools: 128,
@@ -1616,9 +1626,12 @@ var PROFILES_RAW = [
     //
     // Cliffs are HYPOTHESIZED from 2.5-flash family; brain evidence pending.
     id: "gemini-3.1-flash-lite",
-    verifiedAgainstDocs: "2026-05-21",
+    verifiedAgainstDocs: "2026-08-27",
     provider: "google",
-    status: "preview",
+    // alpha.91 — docs page now labels it "Stable" and the brain row has been
+    // `current` since catalog-sync moved it; the bundle lagged (the L-073
+    // layer-parity shape, caught during the TT-49 delisting release).
+    status: "current",
     maxContextTokens: 1048576,
     maxOutputTokens: 65536,
     maxTools: 128,
@@ -1710,6 +1723,122 @@ var PROFILES_RAW = [
     }
   },
   // ─────────────────────────────────────────────────────────────────────────
+  // Gemini 3.5 Flash-Lite — onboarded s91 (2026-08-27)
+  //
+  // The successor Google's own delisting error names: "This model
+  // [gemini-2.5-flash-lite] is no longer available to new users. Please
+  // update your code to use models/gemini-3.5-flash-lite" (tt-intel TT-49,
+  // 2026-08-26). Same price point as the 2.5 Lite rung it replaces in the
+  // summarize/classify starter chains ($0.30/$2.50 vs the old $0.10/$0.40
+  // era pricing), same 1M/65K limits, same 10× cache discount ($0.03 vs
+  // $0.30 input). Positioning: "fastest, most cost-effective 3.5 model for
+  // high-throughput execution … high-volume agentic tasks, translation, and
+  // simple data processing."
+  //
+  // Cliffs + archetypePerf inherited from the Flash-Lite family as starter
+  // hypotheses — ZERO brain rows on this id yet; first 50 rows per archetype
+  // validate or relax.
+  //
+  // Specs verified 2026-08-27 against:
+  //   ai.google.dev/gemini-api/docs/models/gemini-3.5-flash-lite
+  //     (Stable; input limit 1,048,576; output limit 65,536; function
+  //     calling / structured outputs / thinking / caching all Supported)
+  //   ai.google.dev/gemini-api/docs/pricing ($0.30 in / $2.50 out incl.
+  //     thinking tokens; context caching $0.03)
+  // ─────────────────────────────────────────────────────────────────────────
+  {
+    id: "gemini-3.5-flash-lite",
+    verifiedAgainstDocs: "2026-08-27",
+    provider: "google",
+    status: "current",
+    maxContextTokens: 1048576,
+    maxOutputTokens: 65536,
+    maxTools: 128,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "separate",
+    streaming: true,
+    cliffs: [
+      {
+        metric: "input_tokens",
+        threshold: 8e3,
+        action: "downgrade_quality_warning",
+        reason: "Inherited from Flash family: quality degrades above ~8K. Lite tier \u2014 assume it degrades at least as fast. Re-tune from brain after n\u226520."
+      },
+      {
+        metric: "tool_count",
+        threshold: 10,
+        action: "drop_to_top_relevant",
+        reason: "Conservative: Flash drops at 20, Lite tier is smaller \u2014 assume tighter ceiling until brain proves otherwise."
+      },
+      {
+        metric: "thinking_with_short_output",
+        threshold: 1,
+        action: "force_thinking_budget_zero",
+        reason: "Thinking Supported per model page \u2014 same drain risk as Flash; thinking tokens consume maxOutputTokens."
+      },
+      {
+        // Strong prior: Flash hit 5/5 empty rate on summarize+tools (s11
+        // trust artifact). Flash-Lite siblings carry the guard preemptively;
+        // brain telemetry confirms or relaxes per id.
+        metric: "tool_count",
+        threshold: 1,
+        whenIntent: "summarize",
+        action: "strip_tools",
+        reason: "Inherited from Flash s11 cliff: summarize+tools \u2192 empty response. Preemptive guard until brain evidence on 3.5 Flash-Lite specifically."
+      }
+    ],
+    costInputPer1m: 0.3,
+    costOutputPer1m: 2.5,
+    lowering: {
+      ...GOOGLE_LOWERING_BASE,
+      // Cache discount 10× — pricing page: $0.03/M cached vs $0.30/M input.
+      // Material for repeat-prompt workloads (classify shape).
+      cache: { ...GOOGLE_LOWERING_BASE.cache, discount: 0.1 },
+      thinking: { field: "generationConfig.thinkingConfig.thinkingBudget", default: "auto" }
+    },
+    recovery: [
+      {
+        signal: "empty_response_after_tool",
+        action: "retry_with_params",
+        retryParams: { "generationConfig.thinkingConfig.thinkingBudget": 0 },
+        maxRetries: 1,
+        reason: "Known on Flash family: empty after tool result \u2014 retry with thinking off."
+      },
+      {
+        signal: "empty_response",
+        action: "retry_with_params",
+        retryParams: { "generationConfig.thinkingConfig.thinkingBudget": 0 },
+        maxRetries: 1,
+        reason: "Empty response \u2014 try with thinking off."
+      },
+      {
+        signal: "malformed_function_call",
+        action: "escalate",
+        reason: "MALFORMED_FUNCTION_CALL maps to stop \u2014 escalate to next target."
+      }
+    ],
+    strengths: ["low_cost", "speed", "volume", "classification", "summarize", "1m_context", "cache_friendly"],
+    weaknesses: ["complex_reasoning", "large_tool_sets", "complex_schemas", "structured_output_unproven", "long_context_quality"],
+    notes: 'Onboarded s91 (2026-08-27) as the Google-named successor to availability-restricted gemini-2.5-flash-lite. $0.30/$2.50 per 1M, 1M context, 65K max output, cache $0.03/M (10\xD7 discount). "Fastest, most cost-effective 3.5 model for high-throughput execution." Cliffs hypothesized from the Flash family \u2014 brain evidence pending.',
+    // Lite-tier floor for summarize/classify chains. ZERO brain rows — all
+    // values are starter hypotheses anchored to the Flash-Lite family shape.
+    archetypePerf: {
+      classify: 6,
+      // starter hypothesis — verify (Flash is 7, lite likely ≤)
+      summarize: 6,
+      // starter hypothesis — verify; cliff strips tools
+      transform: 6,
+      // starter hypothesis — verify
+      ask: 5,
+      hunt: 5,
+      generate: 4,
+      extract: 4,
+      plan: 3,
+      critique: 3
+    }
+  },
+  // ─────────────────────────────────────────────────────────────────────────
   // Gemini 3.5 Flash — hand-onboarded s37 (2026-05-21)
   //
   // Google positioning ("Most intelligent for sustained frontier performance
@@ -2848,6 +2977,9 @@ function hashShape(s) {
 function learningKey(archetype, model, shape) {
   return `${DIALECT_VERSION}::${archetype}::${model}::${hashShape(shape)}`;
 }
+function routeScopedModel(model, route) {
+  return route === "openrouter" ? `${model}@openrouter` : model;
+}
 
 // src/passes.ts
 function passSlice(ir) {
@@ -3505,7 +3637,26 @@ function passApplyConventions(ir, profile) {
 }
 
 // src/lower.ts
+var OPENROUTER_VENDOR_SLUG = {
+  anthropic: "anthropic",
+  google: "google",
+  openai: "openai",
+  deepseek: "deepseek",
+  zai: "z-ai",
+  moonshot: "moonshotai"
+};
+var OPENROUTER_ID_OVERRIDES = Object.freeze({});
+function openRouterModelId(profile) {
+  const override = OPENROUTER_ID_OVERRIDES[profile.id];
+  if (override) return override;
+  const slug = OPENROUTER_VENDOR_SLUG[profile.provider];
+  if (!slug) throw new Error(`No OpenRouter vendor slug for provider ${profile.provider} (model ${profile.id})`);
+  return `${slug}/${profile.id}`;
+}
 function lower(ir, profile, hints = {}) {
+  if (hints.route === "openrouter") {
+    return lowerOpenRouter(ir, profile);
+  }
   switch (profile.provider) {
     case "anthropic":
       return lowerAnthropic(ir, profile, hints);
@@ -3858,6 +4009,23 @@ function lowerZai(ir, profile, hints) {
     }
   };
 }
+function lowerOpenRouter(ir, profile) {
+  const parts = buildOpenAICompatibleParts(ir);
+  return {
+    request: {
+      provider: "openrouter",
+      model: openRouterModelId(profile),
+      messages: parts.messages,
+      tools: parts.tools,
+      response_format: parts.response_format
+    },
+    diagnostics: {
+      cacheableTokens: 0,
+      historyCacheableTokens: parts.historyCacheableTokens,
+      estimatedCacheSavingsUsd: 0
+    }
+  };
+}
 function lowerMoonshot(ir, profile) {
   const parts = buildOpenAICompatibleParts(ir);
   return {
@@ -3959,7 +4127,8 @@ var SUPPORTED_PROVIDERS = Object.freeze([
   "openai",
   "deepseek",
   "zai",
-  "moonshot"
+  "moonshot",
+  "openrouter"
 ]);
 function isSupportedProvider(p) {
   return SUPPORTED_PROVIDERS.includes(p);
@@ -3980,7 +4149,11 @@ var PROVIDER_ENV_KEYS = Object.freeze({
   // alpha.65 — Moonshot AI (Kimi family). MOONSHOT_API_KEY is canonical
   // (api.moonshot.ai); KIMI_API_KEY covers the platform.kimi.ai rebrand
   // surface so consumers keyed under either name resolve.
-  moonshot: Object.freeze(["MOONSHOT_API_KEY", "KIMI_API_KEY"])
+  moonshot: Object.freeze(["MOONSHOT_API_KEY", "KIMI_API_KEY"]),
+  // alpha.92 — the OpenRouter gateway route (execute-leg contract). This is
+  // a ROUTE key, not a vendor key: it never makes a vendor "reachable" for
+  // chain composition — only explicit `route: 'openrouter'` calls use it.
+  openrouter: Object.freeze(["OPENROUTER_API_KEY"])
 });
 function defaultEnv() {
   return typeof process !== "undefined" && process.env ? process.env : {};
@@ -4186,7 +4359,7 @@ var STARTER_CHAINS_GROUNDED = {
     { id: "gemini-2.5-flash", grounding: "judgment", reason: "Cost-sensitive primary \u2014 engineer pick" },
     { id: "deepseek-v4-flash", grounding: "measured", reason: "Brain-validated tier 1 for cost-sensitive summarize workloads", n: 169 },
     { id: "claude-haiku-4-5", grounding: "judgment", reason: "Quality safety net" },
-    { id: "gemini-2.5-flash-lite", grounding: "judgment", reason: "Emergency floor \u2014 onboarded s22, no brain evidence yet" },
+    { id: "gemini-3.5-flash-lite", grounding: "judgment", reason: "Emergency floor. alpha.91: retargeted from availability-restricted gemini-2.5-flash-lite (Google refuses it for new accounts, TT-49 2026-08-26); 3.5-flash-lite is the successor Google\u2019s own error names" },
     { id: "gpt-5.4-nano", grounding: "judgment", reason: "alpha.33: third-provider tail (archetypePerf=6) \u2014 cheapest OpenAI, matches summarize cost-band" }
   ],
   // Brain-validated DeepSeek tier 1 (169 rows, 0% empty rate).
@@ -4194,7 +4367,7 @@ var STARTER_CHAINS_GROUNDED = {
     { id: "gemini-2.5-flash", grounding: "judgment", reason: "Cost-sensitive primary \u2014 engineer pick" },
     { id: "deepseek-v4-flash", grounding: "measured", reason: "Brain-validated tier 1 (169 rows, 0% empty rate)", n: 169 },
     { id: "claude-haiku-4-5", grounding: "judgment", reason: "Quality safety net" },
-    { id: "gemini-2.5-flash-lite", grounding: "judgment", reason: "Cache-discount 10\xD7 floor for repeat-prompt workloads" },
+    { id: "gemini-3.5-flash-lite", grounding: "judgment", reason: "Cache-discount 10\xD7 floor for repeat-prompt workloads ($0.03/M cached). alpha.91: retargeted from availability-restricted gemini-2.5-flash-lite" },
     { id: "gpt-5.4-nano", grounding: "judgment", reason: "alpha.33: third-provider tail (archetypePerf=6) \u2014 cheapest OpenAI for classify" }
   ]
 };
@@ -6058,12 +6231,13 @@ function compile(ir, opts = {}) {
   const lowered = lower(workingIR, profile, {
     forceThinkingZero: cliffs.value.loweringHints.forceThinkingZero,
     forceTerseOutput: cliffs.value.loweringHints.forceTerseOutput,
-    wireOverrides
+    wireOverrides,
+    route: opts.route
   });
   validateFinalFit(workingIR, profile, inputTokens);
   const handle = makeHandle();
   const finalShape = computeShape(workingIR, inputTokens);
-  const _learningKey = learningKey(ir.intent.archetype, profile.id, finalShape);
+  const _learningKey = learningKey(ir.intent.archetype, routeScopedModel(profile.id, opts.route), finalShape);
   const historyCacheMarkIndex = computeHistoryCacheMarkIndex(workingIR);
   const systemMessages = buildSystemMessages(workingIR, profile.provider);
   const systemCacheMarkIndex = lastCacheableSystemIndex(systemMessages);
@@ -6290,7 +6464,7 @@ function validateFinalFit(ir, profile, tokens) {
 }
 
 // src/version.ts
-var LIBRARY_VERSION = "2.0.0-alpha.90";
+var LIBRARY_VERSION = "2.0.0-alpha.92";
 
 // src/pricing-brain.ts
 function isPricingRow(x) {
@@ -6565,7 +6739,7 @@ async function flushBrainDeadLetter() {
 }
 var compileRegistry = /* @__PURE__ */ new Map();
 var REGISTRY_MAX_ENTRIES = 1e4;
-function registerCompile(appId, archetype, ir, result, parentHandle) {
+function registerCompile(appId, archetype, ir, result, parentHandle, route) {
   if (compileRegistry.size >= REGISTRY_MAX_ENTRIES) {
     const cutoff = Math.floor(REGISTRY_MAX_ENTRIES * 0.25);
     let evicted = 0;
@@ -6597,7 +6771,10 @@ function registerCompile(appId, archetype, ir, result, parentHandle) {
     model: result.target,
     provider: result.provider,
     shapeKey,
-    learningKey: learningKey(archetype, result.target, shape),
+    route: route ?? "direct",
+    // alpha.92 — route-scoped for non-direct routes: the same model id on a
+    // gateway is a different serving stack (Venice/StreamLake, 2026-08-26).
+    learningKey: learningKey(archetype, routeScopedModel(result.target, route), shape),
     estimatedTokensIn: tokens,
     mutationsApplied: result.mutationsApplied.map((m) => m.id),
     // alpha.30: cache the in-memory advisories so record() can auto-persist
@@ -6791,6 +6968,11 @@ function buildPayload(input, reg) {
     provider: reg?.provider,
     shape_key: reg?.shapeKey,
     learning_key: reg?.learningKey,
+    // alpha.92 (migration 067) — route identity on every row. Input wins
+    // (a manual record() replaying through a gateway states its own route);
+    // registry fills the call() path; column default covers legacy writers.
+    route: input.route ?? reg?.route ?? "direct",
+    route_host: input.routeHost,
     mutations_applied: mutationsApplied,
     tokens_in: input.tokensIn,
     tokens_out: input.tokensOut,
@@ -7585,6 +7767,7 @@ var OPENAI_URL = "https://api.openai.com/v1/chat/completions";
 var DEEPSEEK_URL = "https://api.deepseek.com/chat/completions";
 var ZAI_URL = "https://api.z.ai/api/paas/v4/chat/completions";
 var MOONSHOT_URL = "https://api.moonshot.ai/v1/chat/completions";
+var OPENROUTER_URL = "https://openrouter.ai/api/v1/chat/completions";
 async function execute(request, opts = {}) {
   const merged = applyOverrides(request, opts.providerOverrides);
   switch (merged.provider) {
@@ -7608,6 +7791,12 @@ async function execute(request, opts = {}) {
         opts,
         { provider: "moonshot", url: MOONSHOT_URL, missingKeyMessage: "MOONSHOT_API_KEY missing" }
       );
+    case "openrouter":
+      return executeOpenAICompatible(
+        merged,
+        opts,
+        { provider: "openrouter", url: OPENROUTER_URL, missingKeyMessage: "OPENROUTER_API_KEY missing" }
+      );
     default: {
       const _exhaustive = merged;
       throw new Error(`execute(): no executor for provider: ${JSON.stringify(_exhaustive)}`);
@@ -7807,6 +7996,7 @@ function normalizeOpenAILike(raw) {
   const r = raw;
   const choice = r.choices?.[0];
   const text = choice?.message?.content ?? "";
+  const routeHost = typeof raw?.provider === "string" ? raw.provider : void 0;
   const toolCalls = (choice?.message?.tool_calls ?? []).filter((tc) => tc.function?.name).map((tc, i) => ({
     id: tc.id ?? `tc-${i}`,
     name: tc.function.name,
@@ -7819,7 +8009,7 @@ function normalizeOpenAILike(raw) {
     total: u.total_tokens ?? (u.prompt_tokens ?? 0) + (u.completion_tokens ?? 0),
     cached: u.prompt_tokens_details?.cached_tokens
   };
-  return { text, structuredOutput: null, toolCalls, tokens, finishReason: choice?.finish_reason, raw };
+  return { text, structuredOutput: null, toolCalls, tokens, finishReason: choice?.finish_reason, routeHost, raw };
 }
 function applyOverrides(request, overrides) {
   if (!overrides) return request;
@@ -7844,6 +8034,9 @@ function classifyHttpError2(status, body) {
   if (status === 404) {
     return { ok: false, status, errorType: "retryable", errorCode: "model_not_found", message, raw: body };
   }
+  if (status >= 400 && status < 500 && isModelUnavailableMessage(message)) {
+    return { ok: false, status, errorType: "retryable", errorCode: "model_not_found", message, raw: body };
+  }
   if (status === 401 || status === 403) {
     return { ok: false, status, errorType: "terminal", errorCode: "auth", message, raw: body };
   }
@@ -7855,6 +8048,12 @@ function classifyHttpError2(status, body) {
   }
   return { ok: false, status, errorType: "terminal", errorCode: "unknown", message, raw: body };
 }
+function isModelUnavailableMessage(message) {
+  if (!message) return false;
+  return /no longer (available|supported)|not available to new users|has been (deprecated|retired|discontinued|shut ?down)/i.test(
+    message
+  );
+}
 function extractErrorMessage2(body) {
   if (!body || typeof body !== "object") return void 0;
   const b = body;
@@ -8249,6 +8448,7 @@ async function call(ir, opts = {}) {
         handle: initial.handle,
         source: opts.source,
         endUserId: opts.endUserId,
+        route: opts.route,
         tokensIn: 0,
         tokensOut: 0,
         latencyMs: latencyMs2,
@@ -8303,6 +8503,7 @@ async function call(ir, opts = {}) {
         handle: initial.handle,
         source: opts.source,
         endUserId: opts.endUserId,
+        route: opts.route,
         tokensIn: 0,
         tokensOut: 0,
         latencyMs: latencyMs2,
@@ -8446,6 +8647,11 @@ async function call(ir, opts = {}) {
         handle: initial.handle,
         source: opts.source,
         endUserId: opts.endUserId,
+        // alpha.92 — route identity on the row (registry also carries it;
+        // input wins in buildPayload, so state it explicitly here with the
+        // host the gateway reported for the WINNING attempt).
+        route: opts.route,
+        routeHost: validated.response.routeHost,
         tokensIn: validated.response.tokens.input,
         tokensOut: validated.response.tokens.output,
         latencyMs: latencyMs2,
@@ -8526,6 +8732,12 @@ async function call(ir, opts = {}) {
         provider: activeCompile.provider,
         response: validated.response,
         latencyMs: latencyMs2,
+        // alpha.92 — route identity + the served profile's cliffs flagged
+        // route-unverified on a non-direct route (guards still fired; their
+        // evidence is vendor-route evidence — see the execute-leg contract).
+        route: opts.route ?? "direct",
+        routeHost: validated.response.routeHost,
+        routeUnverifiedCliffs: opts.route === "openrouter" ? (tryGetProfile(targetModel)?.cliffs ?? []).map((c) => c.metric) : void 0,
         mutationsApplied: activeCompile.mutationsApplied,
         attempts,
         servedBy: targetModel,
@@ -8570,6 +8782,7 @@ async function call(ir, opts = {}) {
     handle: initial.handle,
     source: opts.source,
     endUserId: opts.endUserId,
+    route: opts.route,
     tokensIn: lastErr?.tokens?.input ?? 0,
     tokensOut: lastErr?.tokens?.output ?? 0,
     latencyMs,
@@ -8613,9 +8826,12 @@ function compileAndRegister(ir, opts) {
     toolRelevanceThreshold: opts.toolRelevanceThreshold,
     compressHistoryAfter: opts.compressHistoryAfter,
     // alpha.68 / Release A — carry the fan-out parent handle through to compile.
-    parentHandle: opts.parentHandle
+    parentHandle: opts.parentHandle,
+    // alpha.92 — gateway route (execute-leg contract): changes the wire +
+    // learning-key encoding, never the pass pipeline.
+    route: opts.route
   });
-  registerCompile(ir.appId, ir.intent.archetype, ir, result, opts.parentHandle);
+  registerCompile(ir.appId, ir.intent.archetype, ir, result, opts.parentHandle, opts.route);
   return result;
 }
 function extractPromptPreview(ir) {
@@ -11054,7 +11270,7 @@ function planDecomposition(args) {
 // src/index.ts
 function compile2(ir, opts) {
   const result = compile(ir, opts);
-  registerCompile(ir.appId, ir.intent.archetype, ir, result, opts?.parentHandle);
+  registerCompile(ir.appId, ir.intent.archetype, ir, result, opts?.parentHandle, opts?.route);
   return result;
 }
 // Annotate the CommonJS export names for ESM import in node: