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

package/dist/profiles.js

@@ -553,9 +553,19 @@ 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,
@@ -1280,9 +1290,12 @@ 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,
@@ -1374,6 +1387,122 @@ 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

package/dist/profiles.mjs

@@ -13,7 +13,7 @@ import {
   registerProfiles,
   resolveModelAlias,
   tryGetProfile
-} from "./chunk-OK2TMFRR.mjs";
+} from "./chunk-T53ISC2F.mjs";
 export {
   ALIASES,
   LATENCY_TIER_MS,

package/dist/{types-Cp9ot1HV.d.ts → types-BCHv34P7.d.ts}

@@ -1,4 +1,4 @@
-import { j as Adapter, F as FallbackReason, y as SectionKind } from './ir-CTx026t0.js';
+import { j as Adapter, F as FallbackReason, y as SectionKind } from './ir-BWnE6LaB.js';
 
 /**
  * Wire contract for the Glass-Box Chrome extension's brain-poll endpoint.

package/dist/{types-cBzinzUR.d.mts → types-BgvLmT3s.d.mts}

@@ -1,4 +1,4 @@
-import { t as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, n as CallAttempt } from './ir-DeYMLWge.mjs';
+import { t as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, n as CallAttempt } from './ir-BPYh68mv.mjs';
 
 /**
  * Glass-Box observability types (alpha.17).

package/dist/{types-BKbRtmUb.d.ts → types-BhxC4hdx.d.ts}

@@ -1,4 +1,4 @@
-import { t as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, n as CallAttempt } from './ir-CTx026t0.js';
+import { t as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, n as CallAttempt } from './ir-BWnE6LaB.js';
 
 /**
  * Glass-Box observability types (alpha.17).

package/dist/{types-DD36cCbZ.d.mts → types-eLelJBj-.d.mts}

@@ -1,4 +1,4 @@
-import { j as Adapter, F as FallbackReason, y as SectionKind } from './ir-DeYMLWge.mjs';
+import { j as Adapter, F as FallbackReason, y as SectionKind } from './ir-BPYh68mv.mjs';
 
 /**
  * Wire contract for the Glass-Box Chrome extension's brain-poll endpoint.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@warmdrift/kgauto-compiler",
-  "version": "2.0.0-alpha.90",
+  "version": "2.0.0-alpha.92",
   "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",