@warmdrift/kgauto-compiler 2.0.0-alpha.49 → 2.0.0-alpha.51

package/dist/profiles.js

@@ -64,6 +64,115 @@ var GOOGLE_LOWERING_BASE = {
 };
 var PROFILES_RAW = [
   // ── Anthropic ──
+  {
+    // VERIFIED 2026-06-21 against the claude-api reference. Claude Fable 5 is
+    // Anthropic's most capable model — a NEW tier ABOVE the opus/sonnet/haiku
+    // line (the id has no MAJOR-MINOR version, just `-5`) at 2× Opus pricing
+    // ($10/$50 vs $5/$25). Hand-authored, not cloned: the onboarder regex can't
+    // match this shape (new family + single-segment version), AND a clone would
+    // carry the wrong (opus) pricing. status:'preview' — no brain evidence yet;
+    // it earns per-archetype placement through the machinery like any model, and
+    // at 2× Opus the quality-floor + cost gates do real work before it leads.
+    // API quirk (claude-api ref): an explicit `thinking:{type:"disabled"}` 400s
+    // on Fable 5 (omit instead) — MOOT here: ANTHROPIC_LOWERING_BASE carries no
+    // `thinking` field, so kgauto never sends the param. (L-049/L-081.)
+    id: "claude-fable-5",
+    verifiedAgainstDocs: "2026-06-21",
+    provider: "anthropic",
+    status: "preview",
+    maxContextTokens: 1e6,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "grammar",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 10,
+    costOutputPer1m: 50,
+    lowering: ANTHROPIC_LOWERING_BASE,
+    recovery: [
+      {
+        signal: "rate_limit",
+        action: "escalate",
+        reason: "429 from Anthropic \u2014 escalate to fallback chain"
+      },
+      {
+        signal: "model_not_found",
+        action: "escalate",
+        reason: "Model deprecated/renamed \u2014 escalate (L-061)"
+      }
+    ],
+    strengths: ["reasoning", "agentic_coding", "long_context", "reliable_tool_use", "structured_output"],
+    weaknesses: ["cost", "latency"],
+    notes: "Most capable Anthropic model (2026-06); a new tier ABOVE Opus at 2\xD7 pricing ($10/$50). 1M ctx, 128k out. Verified 2026-06-21 against the claude-api reference. status:preview \u2014 no brain evidence yet; earns placement via the machinery (cost + quality-floor gates govern). latencyTier derives `slow` (weaknesses includes latency \u2014 minutes-long turns on hard tasks). The thinking-disabled-400 API quirk is moot for kgauto (Anthropic lowering emits no thinking param).",
+    // archetypePerf cloned from claude-opus-4-8 as a conservative estimate
+    // (Fable >= Opus in capability, but unmeasured — no fabricated superiority).
+    // Re-rank once brain rows exist.
+    archetypePerf: {
+      critique: 10,
+      plan: 10,
+      generate: 9,
+      ask: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 8,
+      summarize: 8,
+      classify: 8
+    }
+  },
+  {
+    // Auto-onboarded 2026-06-04 from `claude-opus-4-7`; VERIFIED 2026-06-21
+    // against the claude-api reference. Opus 4.8 shares 4.7's surface exactly —
+    // 1M ctx, 128k out, $5/$25, no new breaking changes — so the clone's
+    // capability data was accurate (confirmed field-by-field). archetypePerf is
+    // still cloned from 4.7 (a sound estimate: 4.8 >= 4.7) — re-rank once brain
+    // rows exist. Promoted to status:'current' 2026-06-21 (Sacha: "include Claude
+    // Opus 4.8") — now live-selectable. Ties with the still-current claude-opus-4-7
+    // (identical $5/$25 + cloned perf) resolve to 4-8 by array order. (L-049/L-081.)
+    id: "claude-opus-4-8",
+    verifiedAgainstDocs: "2026-06-21",
+    provider: "anthropic",
+    status: "current",
+    maxContextTokens: 1e6,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "grammar",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 5,
+    costOutputPer1m: 25,
+    lowering: ANTHROPIC_LOWERING_BASE,
+    recovery: [
+      {
+        signal: "rate_limit",
+        action: "escalate",
+        reason: "429 from Anthropic \u2014 escalate to fallback chain"
+      },
+      {
+        signal: "model_not_found",
+        action: "escalate",
+        reason: "Model deprecated/renamed \u2014 escalate (L-061)"
+      }
+    ],
+    strengths: ["reasoning", "agentic_coding", "long_context", "reliable_tool_use", "structured_output"],
+    weaknesses: ["cost", "latency"],
+    notes: "Frontier Opus (2026-06), the current recommended Opus-tier default. Verified 2026-06-21 against the claude-api reference: 1M ctx, 128k out, $5/$25 \u2014 identical surface to 4.7, no new breaking changes. Promoted to status:current 2026-06-21 (live-selectable); coexists with claude-opus-4-7 (still current).",
+    // Cloned archetypePerf from claude-opus-4-7 — re-evaluate once brain rows exist.
+    archetypePerf: {
+      critique: 10,
+      plan: 10,
+      generate: 9,
+      ask: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 8,
+      summarize: 8,
+      classify: 8
+    }
+  },
   {
     id: "claude-opus-4-7",
     verifiedAgainstDocs: "2026-05-08",
@@ -631,7 +740,7 @@ var PROFILES_RAW = [
   },
   {
     id: "deepseek-v4-pro",
-    verifiedAgainstDocs: "2026-05-08",
+    verifiedAgainstDocs: "2026-06-22",
     provider: "deepseek",
     status: "current",
     maxContextTokens: 1e6,
@@ -649,10 +758,14 @@ var PROFILES_RAW = [
         reason: "Sequential tool calls only \u2014 L-040"
       }
     ],
-    // Profile carries REGULAR pricing, not the 75%-off promo (ends 2026-05-31).
-    // Under-estimating cost is worse than over-estimating for budget caps.
-    costInputPer1m: 1.74,
-    costOutputPer1m: 3.48,
+    // Verified against the live DeepSeek pricing page 2026-06-22: the 75%-off
+    // launch "promo" ($0.435/$0.87) did NOT revert on 2026-05-31 — it became the
+    // standing price (docs show $0.435 in / $0.87 out cache-miss, $0.003625
+    // cache-hit in; no promo caveat). The old $1.74/$3.48 "regular post-promo"
+    // was a projected revert that never happened (L-073/L-081) — it 4×-overstated
+    // cost and suppressed V4-Pro in cost-aware ranking.
+    costInputPer1m: 0.435,
+    costOutputPer1m: 0.87,
     lowering: {
       system: { mode: "inline" },
       cache: { strategy: "unsupported" },
@@ -668,7 +781,7 @@ var PROFILES_RAW = [
     // of the served set. Tag derivation would say 'medium'; the measurement says
     // otherwise.
     latencyTier: "slow",
-    notes: "Pro tier. 1M context, 384k max output. Regular pricing $1.74/$3.48; 75% promo through 2026-05-31 ($0.435/$0.87). Default mode = thinking.",
+    notes: "Pro tier. 1M context, 384k max output. $0.435/$0.87 per 1M (cache-hit in $0.003625) \u2014 verified live 2026-06-22; the launch promo became the standing price (no revert). Default mode = thinking.",
     // Master plan §3.3: tier 3 cross-provider for plan chain. Reasoning
     // bumped one notch over V4-Flash; same parallel-tool ceiling.
     archetypePerf: {

package/dist/profiles.mjs

@@ -8,7 +8,7 @@ import {
   latencyTierOf,
   profilesByProvider,
   tryGetProfile
-} from "./chunk-PKOFXEB3.mjs";
+} from "./chunk-CXH7KC4D.mjs";
 export {
   ALIASES,
   LATENCY_TIER_MS,

package/dist/{types-xBWYk9Mm.d.mts → types-BTeRoSvM.d.mts}

@@ -1,4 +1,4 @@
-import { i as Adapter, w as SectionKind } from './ir-BBKgsUX7.mjs';
+import { i as Adapter, w as SectionKind } from './ir-dDcG8Pvu.mjs';
 
 /**
  * Internal config + hook types for createGlassboxRoutes().

package/dist/{types-B6z-8phQ.d.ts → types-B_pdPjxm.d.ts}

@@ -1,4 +1,4 @@
-import { i as Adapter, w as SectionKind } from './ir-D3n-pBYI.js';
+import { i as Adapter, w as SectionKind } from './ir-rUUojj0s.js';
 
 /**
  * Internal config + hook types for createGlassboxRoutes().

package/dist/{types-Bzw8tH-7.d.ts → types-UXPxWabQ.d.ts}

@@ -1,4 +1,4 @@
-import { r as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, m as CallAttempt } from './ir-D3n-pBYI.js';
+import { r as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, m as CallAttempt } from './ir-rUUojj0s.js';
 
 /**
  * Glass-Box observability types (alpha.17).

package/dist/{types-DMJqRzDF.d.mts → types-vGo-h0tZ.d.mts}

@@ -1,4 +1,4 @@
-import { r as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, m as CallAttempt } from './ir-BBKgsUX7.mjs';
+import { r as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, m as CallAttempt } from './ir-dDcG8Pvu.mjs';
 
 /**
  * Glass-Box observability types (alpha.17).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@warmdrift/kgauto-compiler",
-  "version": "2.0.0-alpha.49",
+  "version": "2.0.0-alpha.51",
   "description": "Prompt compiler + central learning brain for multi-model AI apps. Swap models without rewriting prompts.",
   "main": "./dist/index.js",
   "module": "./dist/index.mjs",