@warmdrift/kgauto-compiler 2.0.0-alpha.85 → 2.0.0-alpha.87

package/dist/profiles.js

@@ -23,11 +23,15 @@ __export(profiles_exports, {
   ALIASES: () => ALIASES,
   LATENCY_TIER_MS: () => LATENCY_TIER_MS,
   _setProfileBrainHook: () => _setProfileBrainHook,
+  _testClearConsumerProfiles: () => _testClearConsumerProfiles,
   allProfiles: () => allProfiles,
   allProfilesRaw: () => allProfilesRaw,
+  bestEffortProfile: () => bestEffortProfile,
   getProfile: () => getProfile,
+  inferProviderFromId: () => inferProviderFromId,
   latencyTierOf: () => latencyTierOf,
   profilesByProvider: () => profilesByProvider,
+  registerProfiles: () => registerProfiles,
   resolveModelAlias: () => resolveModelAlias,
   tryGetProfile: () => tryGetProfile
 });
@@ -244,7 +248,7 @@ var PROFILES_RAW = [
   },
   {
     id: "claude-opus-4-7",
-    verifiedAgainstDocs: "2026-05-08",
+    verifiedAgainstDocs: "2026-08-14",
     provider: "anthropic",
     status: "legacy",
     // s76 2026-07-25: docs list as LEGACY (Opus 5 is current). Bundle said current while the brain already said legacy — the two disagreed.
@@ -294,7 +298,7 @@ var PROFILES_RAW = [
   },
   {
     id: "claude-opus-4-6",
-    verifiedAgainstDocs: "2026-05-08",
+    verifiedAgainstDocs: "2026-08-14",
     provider: "anthropic",
     status: "legacy",
     maxContextTokens: 1e6,
@@ -339,7 +343,7 @@ var PROFILES_RAW = [
   },
   {
     id: "claude-sonnet-4-6",
-    verifiedAgainstDocs: "2026-05-08",
+    verifiedAgainstDocs: "2026-08-14",
     provider: "anthropic",
     status: "legacy",
     // s76 2026-07-25: docs list as LEGACY (Sonnet 5 is current).
@@ -380,7 +384,7 @@ var PROFILES_RAW = [
   },
   {
     id: "claude-haiku-4-5",
-    verifiedAgainstDocs: "2026-05-08",
+    verifiedAgainstDocs: "2026-08-14",
     provider: "anthropic",
     status: "current",
     maxContextTokens: 2e5,
@@ -443,7 +447,7 @@ var PROFILES_RAW = [
   // ── Google ──
   {
     id: "gemini-2.5-flash",
-    verifiedAgainstDocs: "2026-05-08",
+    verifiedAgainstDocs: "2026-08-14",
     provider: "google",
     status: "current",
     maxContextTokens: 1048576,
@@ -549,7 +553,7 @@ 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: "2026-05-13",
+    verifiedAgainstDocs: "2026-08-14",
     provider: "google",
     status: "current",
     maxContextTokens: 1048576,
@@ -644,9 +648,21 @@ var PROFILES_RAW = [
   },
   {
     id: "gemini-2.5-pro",
-    verifiedAgainstDocs: "2026-05-08",
+    verifiedAgainstDocs: "2026-08-08",
     provider: "google",
-    status: "current",
+    // alpha.86 — AVAILABILITY demotion, not a provider-lifecycle mirror.
+    // Google's docs still list 2.5-pro as Stable, but the serving layer
+    // disagrees per account cohort: tt-intel measured 4/4 refusals
+    // ("no longer available to new users", 2026-08-02, their bench), and
+    // kgauto's own key 429s on a FreeTier per-day quota (probed
+    // 2026-08-08). A chain anchor that hard-fails for whole account
+    // cohorts is not an anchor; `legacy` removes it from cost-ordered
+    // chains and new selection while keeping the profile resolvable for
+    // consumers who pass it explicitly. Re-promotion is one status flip
+    // if Google restores availability. The registry cannot yet express
+    // account-cohort availability as data — see interfaces/kgauto.md
+    // ## Requested 2026-08-02.
+    status: "legacy",
     maxContextTokens: 1048576,
     maxOutputTokens: 65535,
     maxTools: 128,
@@ -676,7 +692,7 @@ var PROFILES_RAW = [
       }
     ],
     strengths: ["reasoning", "1m_context", "structured_output", "tool_use"],
-    weaknesses: ["pricing_above_200k"],
+    weaknesses: ["pricing_above_200k", "availability_restricted_by_account_cohort"],
     // Master plan §3.3 anchor: tier-2 cross-provider in almost every chain.
     // Sits on the frontier at perf-9 — close to Sonnet but cheaper input.
     archetypePerf: {
@@ -694,7 +710,7 @@ var PROFILES_RAW = [
   },
   {
     id: "gemini-3.1-pro-preview",
-    verifiedAgainstDocs: "2026-05-08",
+    verifiedAgainstDocs: "2026-08-14",
     provider: "google",
     status: "preview",
     maxContextTokens: 1048576,
@@ -754,7 +770,7 @@ var PROFILES_RAW = [
   // deepseek-reasoner registered as aliases (see ALIASES below).
   {
     id: "deepseek-v4-flash",
-    verifiedAgainstDocs: "2026-05-08",
+    verifiedAgainstDocs: "2026-08-14",
     provider: "deepseek",
     status: "current",
     maxContextTokens: 1e6,
@@ -792,7 +808,7 @@ var PROFILES_RAW = [
     // baseQuality (it carries 'reasoning') with no latency counterweight,
     // leapfrogged gemini-2.5-flash as PB's summarize leader once reachable.
     latencyTier: "slow",
-    notes: "Cheap workhorse. 1M context, 384k max output. Cache-hit input $0.0028/M (1/50\xD7 of miss). Aliased as `deepseek-chat` (non-thinking) and `deepseek-reasoner` (thinking) \u2014 see ALIASES.",
+    notes: "Cheap workhorse. 1M context, 384k max output. Cache-hit input $0.0028/M (1/50\xD7 of miss). Aliased as `deepseek-chat` (non-thinking) and `deepseek-reasoner` (thinking) \u2014 see ALIASES. PRICING MODEL CHANGE effective 2026-08-16: DeepSeek moves to peak/off-peak billing, off-peak at HALF the peak rate (docs, fetched 2026-08-14). Encoded prices are peak; cost math for off-peak traffic will overstate by up to 2x until time-of-day pricing is modeled.",
     // Master plan §6.2 anchor. Brain-validated tier 1 cross-provider for
     // classify (169 rows, 0% empty). Tier 0 for summarize-with-no-tools.
     // Falls off on hunt (sequential tools — L-040) and reasoning depth.
@@ -827,7 +843,7 @@ var PROFILES_RAW = [
   },
   {
     id: "deepseek-v4-pro",
-    verifiedAgainstDocs: "2026-06-22",
+    verifiedAgainstDocs: "2026-08-14",
     provider: "deepseek",
     status: "current",
     maxContextTokens: 1e6,
@@ -868,7 +884,7 @@ 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. $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.",
+    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. PRICING MODEL CHANGE effective 2026-08-16: DeepSeek moves to peak/off-peak billing, off-peak at HALF the peak rate (docs, fetched 2026-08-14). Encoded prices are peak; cost math for off-peak traffic will overstate by up to 2x until time-of-day pricing is modeled.",
     // Master plan §3.3: tier 3 cross-provider for plan chain. Reasoning
     // bumped one notch over V4-Flash; same parallel-tool ceiling.
     archetypePerf: {
@@ -1476,13 +1492,109 @@ var PROFILES_RAW = [
       // +1 vs 2.5-flash — but still below Sonnet/Opus reasoning floor
     }
   },
+  // ─────────────────────────────────────────────────────────────────────────
+  // gemini-3.6-flash — hand-onboarded s86 (2026-08-14), step-zero against
+  // Google docs (models/gemini-3.6-flash page + pricing via catalog-sync's
+  // two-agreeing-extraction the same day). Docs list it as Stable and call
+  // it "previous-generation Flash" (3.5-flash is now "legacy" in Google's
+  // wording) — but kgauto status is 'preview' per the standing entry rule:
+  // a new model enters on capability data verified only at the docs level,
+  // preview is unreachable by starter chains, and promotion to routing
+  // happens through brain evidence, not through Google's marketing tier.
+  // This also means NO family-primary retarget rides this release.
+  // ─────────────────────────────────────────────────────────────────────────
+  {
+    id: "gemini-3.6-flash",
+    verifiedAgainstDocs: "2026-08-14",
+    versionAdded: "2.0.0-alpha.87",
+    provider: "google",
+    status: "preview",
+    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 Flash-family guard (2.5\u21923.5 lineage); no 3.6 brain evidence yet"
+      },
+      {
+        metric: "tool_count",
+        threshold: 20,
+        action: "drop_to_top_relevant",
+        reason: "Inherited from Flash family: tool reliability drops above ~20 (despite 128 hard limit). Validate per (archetype, model) after n\u226520."
+      },
+      {
+        metric: "thinking_with_short_output",
+        threshold: 1,
+        action: "force_thinking_budget_zero",
+        reason: "Thinking supported per Google docs; same drain risk as 2.5/3.5-flash \u2014 thinking tokens consume maxOutputTokens"
+      },
+      {
+        metric: "tool_count",
+        threshold: 1,
+        whenIntent: "summarize",
+        action: "strip_tools",
+        reason: "Inherited from 2.5-flash s11 cliff: summarize+tools \u2192 empty response. Preemptive family guard until brain evidence on 3.6-flash specifically."
+      }
+    ],
+    costInputPer1m: 1.5,
+    costOutputPer1m: 7.5,
+    lowering: {
+      ...GOOGLE_LOWERING_BASE,
+      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: "Inherited Flash-family pattern: 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: ["agentic_loops", "coding", "1m_context", "parallel_tools", "thinking_mode"],
+    weaknesses: ["no_brain_evidence_yet"],
+    notes: "Hand-onboarded s86 (2026-08-14), step-zero verified: 1,048,576 ctx / 65,536 out (Google model page), $1.50/$7.50 per 1M (catalog-sync two-agreeing-extraction same day \u2014 CHEAPER output than 3.5-flash at $9.00). Docs status Stable; enters kgauto as preview per standing rule (availability != lifecycle, promotion via brain evidence). Function calling + structured outputs + thinking per docs; parallel tools assumed per Flash family (docs silent). First flagged by the release watcher s73; sat in the onboarding queue 18 days \u2014 drained by the s86 catalog-sync run.",
+    archetypePerf: {
+      hunt: 9,
+      // Flash-family parallel-tool tier (L-040); inherited from 3.5-flash
+      classify: 7,
+      summarize: 7,
+      transform: 7,
+      ask: 8,
+      generate: 8,
+      plan: 7,
+      extract: 7,
+      critique: 5
+    }
+  },
   {
     // Auto-onboarded 2026-07-01 from `claude-sonnet-4-6`; VERIFIED 2026-07-02
     // against the claude-api reference (cc-portfolio ratification pass). The
-    // clone got context right (1M) and pricing right at sticker ($3/$15 —
-    // NOTE an introductory $2/$10 per MTok runs through 2026-08-31; sticker
-    // encoded here per the time-bounded-pricing convention, intro belongs in
-    // brain kgauto_pricing if worth capturing). The clone got max output WRONG:
+    // clone got context right (1M). PRICING RE-BASED 2026-08-14 (s86): the
+    // introductory $2/$10 was made PERMANENT — Anthropic's pricing page now
+    // states the scheduled Sept-1 increase to $3/$15 "will not occur", so
+    // $2/$10 IS the sticker. (The $3/$15 encoded 07-02 was correct under the
+    // time-bounded-pricing convention at the time; the bound was cancelled.) The clone got max output WRONG:
     // Sonnet 5 is 128k, not 4-6's 64k — corrected. New tokenizer (~30% more
     // tokens for the same text vs 4-6): byte-budget consumers should re-baseline.
     // API quirks (claude-api ref): (a) NON-DEFAULT temperature/top_p/top_k
@@ -1495,7 +1607,7 @@ var PROFILES_RAW = [
     // 'preview' per the Fable precedent — no brain evidence yet; promotion to
     // 'current' is an explicit call. (L-049/L-081.)
     id: "claude-sonnet-5",
-    verifiedAgainstDocs: "2026-07-02",
+    verifiedAgainstDocs: "2026-08-14",
     provider: "anthropic",
     status: "current",
     // s76 2026-07-25: GA per docs; was preview.
@@ -1507,8 +1619,9 @@ var PROFILES_RAW = [
     systemPromptMode: "inline",
     streaming: true,
     cliffs: [],
-    costInputPer1m: 3,
-    costOutputPer1m: 15,
+    costInputPer1m: 2,
+    // 2026-08-14: intro price made PERMANENT (docs: "the previously scheduled increase to $3/$15 on September 1, 2026 will not occur")
+    costOutputPer1m: 10,
     lowering: ANTHROPIC_LOWERING_BASE,
     recovery: [
       { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" },
@@ -1699,6 +1812,55 @@ function _setProfileBrainHook(hook) {
 function canonicalId(id) {
   return brainHook.resolveAlias?.(id) ?? ALIASES[id] ?? id;
 }
+var CONSUMER_INDEX = /* @__PURE__ */ new Map();
+function registerProfiles(profiles) {
+  for (const p of profiles) CONSUMER_INDEX.set(p.id, p);
+}
+function _testClearConsumerProfiles() {
+  CONSUMER_INDEX.clear();
+}
+function inferProviderFromId(id) {
+  const lower = id.toLowerCase();
+  if (lower.startsWith("claude")) return "anthropic";
+  if (lower.startsWith("gemini") || lower.startsWith("palm")) return "google";
+  if (lower.startsWith("gpt") || /^o[0-9]/.test(lower) || lower.startsWith("chatgpt")) return "openai";
+  if (lower.startsWith("deepseek")) return "deepseek";
+  if (lower.startsWith("glm")) return "zai";
+  if (lower.startsWith("kimi") || lower.startsWith("moonshot")) return "moonshot";
+  return void 0;
+}
+function bestEffortProfile(id) {
+  const canonical = canonicalId(id);
+  const existing = tryGetProfile(canonical);
+  if (existing) return existing;
+  const provider = inferProviderFromId(canonical);
+  if (!provider) return void 0;
+  const donor = PROFILES_RAW.find((p) => p.provider === provider && p.status === "current") ?? PROFILES_RAW.find((p) => p.provider === provider);
+  if (!donor) return void 0;
+  const synth = {
+    id: canonical,
+    provider,
+    bestEffort: true,
+    status: "preview",
+    maxContextTokens: 128e3,
+    maxOutputTokens: 4096,
+    maxTools: donor.maxTools,
+    parallelToolCalls: false,
+    structuredOutput: "none",
+    systemPromptMode: donor.systemPromptMode,
+    streaming: donor.streaming,
+    cliffs: [],
+    costInputPer1m: 0,
+    costOutputPer1m: 0,
+    lowering: donor.lowering,
+    recovery: [],
+    strengths: [],
+    weaknesses: ["best-effort synthesized profile \u2014 no verified capability data"],
+    notes: `Synthesized by onUnprofiledModel:'best-effort' (provider ${provider} inferred from id; wire mechanics borrowed from ${donor.id}). No cliffs, no measured knowledge, cost UNKNOWN (recorded as 0). Replace via registerProfiles() for real guards.`
+  };
+  CONSUMER_INDEX.set(canonical, synth);
+  return synth;
+}
 function resolveModelAlias(id) {
   return canonicalId(id);
 }
@@ -1707,18 +1869,22 @@ var PROFILE_INDEX = new Map(
 );
 function getProfile(id) {
   const canonical = canonicalId(id);
+  const fromConsumer = CONSUMER_INDEX.get(canonical);
+  if (fromConsumer) return fromConsumer;
   const fromBrain = brainHook.getProfile?.(canonical);
   if (fromBrain) return fromBrain;
   const p = PROFILE_INDEX.get(canonical);
   if (!p) {
     const known = [...PROFILE_INDEX.keys(), ...Object.keys(ALIASES)].join(", ");
-    throw new Error(`Unknown model id: "${id}". Known: ${known}`);
+    throw new Error(
+      `Unknown model id: "${id}". Known: ${known}. For a model kgauto has not onboarded: registerProfiles([...]) supplies a full profile, or policy.onUnprofiledModel:'best-effort' compiles with guards skipped (alpha.87).`
+    );
   }
   return p;
 }
 function tryGetProfile(id) {
   const canonical = canonicalId(id);
-  return brainHook.getProfile?.(canonical) ?? PROFILE_INDEX.get(canonical);
+  return CONSUMER_INDEX.get(canonical) ?? brainHook.getProfile?.(canonical) ?? PROFILE_INDEX.get(canonical);
 }
 function allProfiles() {
   return PROFILES_RAW;
@@ -1734,11 +1900,15 @@ function profilesByProvider(provider) {
   ALIASES,
   LATENCY_TIER_MS,
   _setProfileBrainHook,
+  _testClearConsumerProfiles,
   allProfiles,
   allProfilesRaw,
+  bestEffortProfile,
   getProfile,
+  inferProviderFromId,
   latencyTierOf,
   profilesByProvider,
+  registerProfiles,
   resolveModelAlias,
   tryGetProfile
 });

package/dist/profiles.mjs

@@ -2,23 +2,31 @@ import {
   ALIASES,
   LATENCY_TIER_MS,
   _setProfileBrainHook,
+  _testClearConsumerProfiles,
   allProfiles,
   allProfilesRaw,
+  bestEffortProfile,
   getProfile,
+  inferProviderFromId,
   latencyTierOf,
   profilesByProvider,
+  registerProfiles,
   resolveModelAlias,
   tryGetProfile
-} from "./chunk-FD3NFXDC.mjs";
+} from "./chunk-YVKTTTTB.mjs";
 export {
   ALIASES,
   LATENCY_TIER_MS,
   _setProfileBrainHook,
+  _testClearConsumerProfiles,
   allProfiles,
   allProfilesRaw,
+  bestEffortProfile,
   getProfile,
+  inferProviderFromId,
   latencyTierOf,
   profilesByProvider,
+  registerProfiles,
   resolveModelAlias,
   tryGetProfile
 };

package/dist/{types-y42mwdvM.d.mts → types-7sF3adLO.d.mts}

@@ -1,4 +1,4 @@
-import { s as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, m as CallAttempt } from './ir-JUoybVM-.mjs';
+import { t as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, n as CallAttempt } from './ir-P3NEU8z4.mjs';
 
 /**
  * Glass-Box observability types (alpha.17).

package/dist/{types-BqTKRECa.d.ts → types-B_MjEB0h.d.ts}

@@ -1,4 +1,4 @@
-import { s as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, m as CallAttempt } from './ir-DfrM71pA.js';
+import { t as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, n as CallAttempt } from './ir-ZgN2KkuF.js';
 
 /**
  * Glass-Box observability types (alpha.17).

package/dist/{types-DLDJbCha.d.ts → types-BgfuVIVI.d.ts}

@@ -1,4 +1,4 @@
-import { i as Adapter, F as FallbackReason, x as SectionKind } from './ir-DfrM71pA.js';
+import { j as Adapter, F as FallbackReason, y as SectionKind } from './ir-ZgN2KkuF.js';
 
 /**
  * Wire contract for the Glass-Box Chrome extension's brain-poll endpoint.

package/dist/{types-BGH_KAHu.d.mts → types-CK0zUlVE.d.mts}

@@ -1,4 +1,4 @@
-import { i as Adapter, F as FallbackReason, x as SectionKind } from './ir-JUoybVM-.mjs';
+import { j as Adapter, F as FallbackReason, y as SectionKind } from './ir-P3NEU8z4.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.85",
+  "version": "2.0.0-alpha.87",
   "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",
@@ -77,6 +77,12 @@
     "model-fallback"
   ],
   "license": "MIT",
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/stue/kgauto.git",
+    "directory": "v2"
+  },
+  "homepage": "https://github.com/stue/kgauto/tree/main/v2#readme",
   "peerDependencies": {
     "js-tiktoken": ">=1.0.0",
     "react": ">=18.0.0",
@@ -103,4 +109,4 @@
     "typescript": "^5.7.0",
     "vitest": "^3.1.0"
   }
-}
+}