@warmdrift/kgauto-compiler 2.0.0-alpha.63 → 2.0.0-alpha.65

package/dist/index.js

@@ -27,6 +27,7 @@ __export(index_exports, {
   ARCHETYPE_FLOOR_DEFAULT: () => ARCHETYPE_FLOOR_DEFAULT,
   CallError: () => CallError,
   DEFAULT_FINDINGS_ENDPOINT: () => DEFAULT_FINDINGS_ENDPOINT,
+  DEFAULT_PROMOTIONS_ENDPOINT: () => DEFAULT_PROMOTIONS_ENDPOINT,
   DIALECT_VERSION: () => DIALECT_VERSION,
   FamilyResolutionError: () => FamilyResolutionError,
   INTENT_ARCHETYPES: () => INTENT_ARCHETYPES,
@@ -38,6 +39,8 @@ __export(index_exports, {
   PROVIDER_ENV_KEYS: () => PROVIDER_ENV_KEYS,
   RULE_SEQUENTIAL_TOOL_CLIFF: () => RULE_SEQUENTIAL_TOOL_CLIFF,
   TRANSLATOR_FLOOR: () => TRANSLATOR_FLOOR,
+  _testResetPromotions: () => _testResetPromotions,
+  _testWaitForPromotionsRefresh: () => _testWaitForPromotionsRefresh,
   allProfiles: () => allProfiles,
   applyArchetypeConvention: () => applyArchetypeConvention,
   applySectionRewrites: () => applySectionRewrites,
@@ -57,6 +60,7 @@ __export(index_exports, {
   compile: () => compile2,
   compileForAISDKv6: () => compileForAISDKv6,
   configureBrain: () => configureBrain,
+  configurePromotionsBrain: () => configurePromotionsBrain,
   countTokens: () => countTokens,
   createBrainForwardRoutes: () => createBrainForwardRoutes,
   createKeyHealthRoute: () => createKeyHealthRoute,
@@ -68,6 +72,7 @@ __export(index_exports, {
   getActionableAdvisories: () => getActionableAdvisories,
   getAllStarterChains: () => getAllStarterChains,
   getAllStarterChainsWithGrounding: () => getAllStarterChainsWithGrounding,
+  getApplicablePromotion: () => getApplicablePromotion,
   getArchetypePerfScore: () => getArchetypePerfScore,
   getDefaultFallbackChain: () => getDefaultFallbackChain,
   getDefaultFallbackChainWithGrounding: () => getDefaultFallbackChainWithGrounding,
@@ -83,10 +88,12 @@ __export(index_exports, {
   getStarterChainWithGrounding: () => getStarterChainWithGrounding,
   hashShape: () => hashShape,
   isArchetype: () => isArchetype,
+  isAutoPromoteEnabledFromEnv: () => isAutoPromoteEnabledFromEnv,
   isBrainQueryActiveFor: () => isBrainQueryActiveFor,
   isBrainSync: () => isBrainSync,
   isExclusionFindingsBrainActive: () => isExclusionFindingsBrainActive,
   isModelReachable: () => isModelReachable,
+  isPromotionsBrainActive: () => isPromotionsBrainActive,
   isProviderReachable: () => isProviderReachable,
   latencyTierOf: () => latencyTierOf,
   learningKey: () => learningKey,
@@ -1711,6 +1718,156 @@ var PROFILES_RAW = [
       classify: 8
       // overkill
     }
+  },
+  // ── Z.ai (GLM) ──
+  // alpha.65 (2026-07-18): first Z.ai profile, provider `zai` added same
+  // release. Step-zero sources (L-049/L-081, all fetched 2026-07-18):
+  // docs.z.ai/guides/llm/glm-5.2 (capabilities), docs.z.ai api-reference
+  // chat-completion (wire params: thinking.type enabled|disabled default
+  // enabled, reasoning_effort max..none, tool_stream, max output 131072),
+  // docs.z.ai/guides/overview/pricing ($1.4 in / $4.4 out, cached input
+  // $0.26, cache storage limited-time free), vLLM recipe (native ctx
+  // 1,048,576). NO AI-trained numbers.
+  {
+    id: "glm-5.2",
+    verifiedAgainstDocs: "2026-07-18",
+    provider: "zai",
+    status: "preview",
+    maxContextTokens: 1048576,
+    maxOutputTokens: 131072,
+    // No documented tool-count cap — conservative default pending measured
+    // evidence; matches the deepseek posture, not a doc number.
+    maxTools: 16,
+    // Hosted-API docs document tools + tool_choice + tool_stream but say
+    // NOTHING about multiple tool_calls per response. Conservative false so
+    // hunt-shaped ranking doesn't over-promise; flip on doc or probe
+    // evidence (L-081: undocumented ≠ true).
+    parallelToolCalls: false,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 1.4,
+    costOutputPer1m: 4.4,
+    lowering: {
+      system: { mode: "inline" },
+      // Automatic prefix caching ("intelligent caching mechanism") — no wire
+      // marker to emit, DeepSeek-style. Cache-hit input $0.26/M (~0.19× of
+      // miss) tracked in notes; storage limited-time free as of 2026-07-18.
+      cache: { strategy: "unsupported" },
+      tools: { format: "openai" },
+      // thinking.type: 'enabled' (server default) | 'disabled'. lowerZai
+      // emits an explicit disable only on forceThinkingZero cliffs.
+      thinking: { field: "thinking.type", default: "auto" }
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" },
+      { signal: "model_not_found", action: "escalate", reason: "Deprecated \u2014 escalate (L-061)" }
+    ],
+    strengths: ["cost", "1m_context", "code", "tool_use", "json_output", "reasoning"],
+    weaknesses: ["parallel_tools"],
+    notes: "GLM-5.2 (Z.ai, 2026-06): agentic-coding flagship. 1,048,576 ctx / 131,072 max out. $1.40/$4.40 per 1M; cached input $0.26/M (automatic caching, no marker; storage limited-time free as of 2026-07-18). Thinking on by default (`thinking.type`), `reasoning_effort` max..none. Parallel tool calls UNDOCUMENTED on the hosted API \u2014 profile says false until doc or probe evidence. status:preview \u2014 no brain evidence yet; earns placement via the machinery.",
+    // Starter hypotheses — verify with telemetry/probes; NO brain evidence
+    // yet. Anchored on the vendor's agentic-coding positioning at a
+    // mid-cost tier, discounted for the unproven parallel-tool story.
+    archetypePerf: {
+      generate: 7,
+      // agentic-coding positioning — starter hypothesis
+      transform: 7,
+      // starter hypothesis
+      plan: 7,
+      // starter hypothesis
+      summarize: 7,
+      // tolerant archetype — starter hypothesis
+      ask: 6,
+      extract: 6,
+      critique: 6,
+      classify: 6,
+      hunt: 5
+      // tools work; parallel unproven → discounted
+    }
+  },
+  // ── Moonshot (Kimi) ──
+  // alpha.65 (2026-07-18): first Moonshot profile, provider `moonshot` added
+  // same release. Step-zero sources (all fetched 2026-07-18):
+  // platform.kimi.ai/docs/pricing/chat-k3 ($3.00 cache-miss in / $0.30
+  // cache-hit in / $15.00 out; 1,048,576 ctx; ToolCalls + JSON Mode +
+  // Partial Mode), kimi-k3-quickstart (max_completion_tokens default
+  // 131072, up to 1,048,576; base https://api.moonshot.ai/v1; vision via
+  // base64 or ms://file-id, NO public image URLs), models-overview
+  // (temperature FIXED 1.0, top_p FIXED 0.95, n FIXED 1, always reasons,
+  // reasoning_effort 'max' only), tool-calls guide ("can choose to call
+  // multiple tools at once … will tend to call them in parallel").
+  {
+    id: "kimi-k3",
+    verifiedAgainstDocs: "2026-07-18",
+    provider: "moonshot",
+    status: "preview",
+    maxContextTokens: 1048576,
+    // Documented ceiling; default is 131,072 when unset.
+    maxOutputTokens: 1048576,
+    // No hard doc cap; docs recommend dynamic tool loading at "dozens or
+    // hundreds" — 64 matches the frontier-agentic posture (sonnet-5 tier).
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 3,
+    costOutputPer1m: 15,
+    lowering: {
+      system: { mode: "inline" },
+      // Automatic context caching — no wire marker. Cache-hit input $0.30/M
+      // (0.1× of miss) tracked in notes.
+      cache: { strategy: "unsupported" },
+      tools: { format: "openai" }
+      // No thinking knob: reasoning is always-on and reasoning_effort
+      // accepts only 'max' (= the server default). Nothing to lower.
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" },
+      { signal: "model_not_found", action: "escalate", reason: "Deprecated \u2014 escalate (L-061)" }
+    ],
+    strengths: ["quality", "reasoning", "1m_context", "parallel_tools", "tool_use", "code", "vision"],
+    // 'latency' → latencyTierOf derives 'slow': always-on max-effort
+    // reasoner with no off switch; no measured p50 yet, the tag is the
+    // honest prior (reasoner family precedent: deepseek-v4-pro ~48s).
+    weaknesses: ["cost", "latency"],
+    notes: "Kimi K3 (Moonshot, API live 2026-07-16): 2.8T-param open-weight flagship, native vision. 1,048,576 ctx; max_completion_tokens default 131,072, ceiling 1,048,576. $3.00/$15.00 per 1M, cache-hit input $0.30/M (automatic caching). Reasoning ALWAYS ON (reasoning_effort max only, cannot disable; responses may carry reasoning_content). Wire quirks: temperature fixed 1.0 / top_p 0.95 / n 1 (kgauto never emits those); vision input must be base64 or ms://file-id \u2014 public image URLs rejected. status:preview \u2014 earns placement via the machinery.",
+    // Starter hypotheses — verify with telemetry/probes; NO brain evidence
+    // yet. Anchored on documented long-horizon-agentic + parallel-tool
+    // strengths; discounted on terse archetypes where always-on reasoning
+    // bills $15/M output for thinking the archetype doesn't need.
+    archetypePerf: {
+      hunt: 8,
+      // parallel tools + long-horizon agentic — starter hypothesis
+      plan: 8,
+      // starter hypothesis
+      generate: 8,
+      // starter hypothesis
+      critique: 7,
+      ask: 7,
+      extract: 7,
+      transform: 7,
+      summarize: 6,
+      // works, but $15/M output reasoning tax on a tolerant archetype
+      classify: 5
+      // always-on reasoning cost + reasoner hedge risk
+    },
+    // Diagnostics-only (no prompt rewrite): reasoner-family hedge risk on
+    // decisive archetypes, carried as a warning until K3-specific probe
+    // evidence exists. The deepseek-reasoner family showed 8/10 judge
+    // rationales citing hedging on classify (exclusion-finding ID 20,
+    // 2026-05-28); K3 is an always-on reasoner and inherits the risk class,
+    // not the conviction.
+    archetypeConventions: [
+      {
+        archetype: "classify",
+        cliffWarning: "kimi-k3 is an always-on reasoner: decisive archetypes (classify) may hedge, and reasoning tokens bill at $15/M output. Family-prior risk (deepseek-reasoner precedent), not K3-measured \u2014 probe before relying on it for classify.",
+        reason: "Reasoner-family prior (exclusion-finding ID 20, 2026-05-28, deepseek-v4-pro on tt-intel/classify). No K3-specific evidence yet \u2014 warning only, no prompt rewrite."
+      }
+    ]
   }
 ];
 var ALIASES = {
@@ -2602,6 +2759,7 @@ function passApplyCliffs(ir, profile, estimatedInputTokens) {
 var LATENCY_OVERAGE_WEIGHT = 0.6;
 var LATENCY_PENALTY_CAP = 1;
 var QUALITY_GATE_PENALTY = 4;
+var PROMOTION_BOOST = 5;
 function effectiveConventions(profile) {
   const own = profile.archetypeConventions ?? [];
   if (own.length > 0) return own;
@@ -2619,6 +2777,9 @@ function passScoreTargets(ir, opts) {
   const preferredSet = new Set(policy.preferredModels ?? []);
   const scores = [];
   const policyMutations = [];
+  const rawPromotion = opts.promotion;
+  const promotionDeferred = rawPromotion !== void 0 && (policy.preferredModels?.length ?? 0) > 0;
+  const promotion = promotionDeferred ? void 0 : rawPromotion;
   const modelIds = ir.models.filter((m) => typeof m === "string");
   for (const modelId of modelIds) {
     let profile;
@@ -2683,7 +2844,12 @@ function passScoreTargets(ir, opts) {
       );
       if (schemaWeak) qualityGatePenalty = QUALITY_GATE_PENALTY;
     }
-    const rank = qualityScore + callerOrderBoost - costPenalty - reasons.length * 10 + preferredBoost - latencyPenalty - qualityGatePenalty;
+    const isPromoted = promotion?.promotedModel === modelId;
+    if (isPromoted && promotion.suppressQualityGate) {
+      qualityGatePenalty = 0;
+    }
+    const promotionBoost = isPromoted ? PROMOTION_BOOST : 0;
+    const rank = qualityScore + callerOrderBoost - costPenalty - reasons.length * 10 + preferredBoost - latencyPenalty - qualityGatePenalty + promotionBoost;
     scores.push({
       modelId,
       estimatedCostUsd,
@@ -2733,6 +2899,22 @@ function passScoreTargets(ir, opts) {
         description: `Model ${modelId} gated below the quality floor for archetype '${ir.intent.archetype}' \u2014 declared structuredOutput + kgauto convention flags it schema-weak for this contract (structuredOutputHint:'avoid'). Down-ranked out of leadership; retained as graceful fallback only.`
       });
     }
+    if (isPromoted && reasons.length === 0) {
+      policyMutations.push({
+        id: `promotion-applied-${modelId}`,
+        source: "surface_promotion",
+        passName: "score_targets",
+        description: `Model ${modelId} boosted to surface leadership by active promotion #${promotion.id} (golden-eval run ${promotion.evalRunId ?? "n/a"}, Stage 2 auto-promote, 7-day rollback guard)` + (promotion.suppressQualityGate ? " \u2014 schema-weak quality gate suppressed for this tuple (eval measured the floor holding on real workload)" : "")
+      });
+    }
+  }
+  if (promotionDeferred && rawPromotion && modelIds.includes(rawPromotion.promotedModel)) {
+    policyMutations.push({
+      id: `promotion-deferred-consumer-preference-${rawPromotion.promotedModel}`,
+      source: "surface_promotion",
+      passName: "score_targets",
+      description: `Active promotion #${rawPromotion.id} for ${rawPromotion.promotedModel} deferred \u2014 CompilePolicy.preferredModels is declared and explicit consumer preference outranks brain evidence.`
+    });
   }
   return { value: scores, mutations: policyMutations };
 }
@@ -2940,6 +3122,10 @@ function lower(ir, profile, hints = {}) {
       return lowerOpenAI(ir, profile, hints);
     case "deepseek":
       return lowerDeepSeek(ir, profile);
+    case "zai":
+      return lowerZai(ir, profile, hints);
+    case "moonshot":
+      return lowerMoonshot(ir, profile);
     default:
       throw new Error(`No lowering implementation for provider: ${profile.provider}`);
   }
@@ -3228,6 +3414,68 @@ function lowerDeepSeek(ir, profile) {
     }
   };
 }
+function buildOpenAICompatibleParts(ir) {
+  const ordered = sortSections(ir.sections);
+  const systemText = ordered.map((s) => s.text).join("\n\n");
+  const messages = systemText ? [{ role: "system", content: systemText }] : [];
+  for (const m of ir.history ?? []) {
+    if (m.role === "system") continue;
+    messages.push({ role: m.role, content: m.parts ?? m.content });
+  }
+  if (ir.currentTurn && ir.currentTurn.role !== "system") {
+    messages.push({
+      role: ir.currentTurn.role,
+      content: ir.currentTurn.parts ?? ir.currentTurn.content
+    });
+  }
+  const history = (ir.history ?? []).filter((m) => m.role !== "system");
+  const histMarkIndex = resolveHistoryMarkIndex(history.length, ir.historyCachePolicy);
+  const historyCacheableTokens = histMarkIndex >= 0 ? sumHistoryTokens(history, histMarkIndex) : 0;
+  return {
+    messages,
+    tools: ir.tools && ir.tools.length > 0 ? toOpenAITools(ir.tools) : void 0,
+    response_format: ir.constraints?.structuredOutput ? { type: "json_object" } : void 0,
+    historyCacheableTokens
+  };
+}
+function lowerZai(ir, profile, hints) {
+  const parts = buildOpenAICompatibleParts(ir);
+  return {
+    request: {
+      provider: "zai",
+      model: profile.id,
+      messages: parts.messages,
+      tools: parts.tools,
+      response_format: parts.response_format,
+      // Z.ai thinking defaults to enabled server-side; emit an explicit
+      // disable only when a cliff forced thinking off (the GLM analogue of
+      // Gemini's thinkingBudget=0).
+      thinking: hints.forceThinkingZero ? { type: "disabled" } : void 0
+    },
+    diagnostics: {
+      cacheableTokens: 0,
+      historyCacheableTokens: parts.historyCacheableTokens,
+      estimatedCacheSavingsUsd: 0
+    }
+  };
+}
+function lowerMoonshot(ir, profile) {
+  const parts = buildOpenAICompatibleParts(ir);
+  return {
+    request: {
+      provider: "moonshot",
+      model: profile.id,
+      messages: parts.messages,
+      tools: parts.tools,
+      response_format: parts.response_format
+    },
+    diagnostics: {
+      cacheableTokens: 0,
+      historyCacheableTokens: parts.historyCacheableTokens,
+      estimatedCacheSavingsUsd: 0
+    }
+  };
+}
 function sortSections(sections) {
   return [...sections].sort((a, b) => {
     const wa = a.weight ?? 100;
@@ -4121,6 +4369,130 @@ ${originalText}`;
   return { rewrittenIR, rewrites };
 }
 
+// src/promotions-brain.ts
+function isRawPromotionRow(x) {
+  if (!x || typeof x !== "object") return false;
+  const r = x;
+  return (typeof r.id === "number" || typeof r.id === "string") && typeof r.intent_archetype === "string" && typeof r.promoted_model === "string" && typeof r.incumbent_model === "string";
+}
+function coerceId(v) {
+  if (typeof v === "number") return Number.isFinite(v) ? v : null;
+  if (typeof v === "string") {
+    const n = Number(v);
+    return Number.isFinite(n) ? n : null;
+  }
+  return null;
+}
+function mapRowsToPromotions(rows) {
+  const out = [];
+  for (const row of rows) {
+    if (!isRawPromotionRow(row)) continue;
+    const id = coerceId(row.id);
+    if (id === null) continue;
+    out.push({
+      id,
+      archetype: row.intent_archetype,
+      promotedModel: row.promoted_model,
+      incumbentModel: row.incumbent_model,
+      evalRunId: coerceId(row.eval_run_id ?? null),
+      suppressQualityGate: row.suppress_quality_gate === true,
+      promotedAt: typeof row.promoted_at === "string" ? row.promoted_at : ""
+    });
+  }
+  return out;
+}
+var snapshots4 = /* @__PURE__ */ new Map();
+var runtime5;
+var warnedOnce4 = false;
+var DEFAULT_PROMOTIONS_ENDPOINT = "https://kgauto-dashboard.vercel.app/api/kgauto-v2/promotions";
+function isAutoPromoteEnabledFromEnv(envSource) {
+  const env = envSource ?? (typeof process !== "undefined" && process.env ? process.env : {});
+  const raw = (env.KGAUTO_AUTO_PROMOTE ?? "").trim().toLowerCase();
+  return raw === "1" || raw === "true";
+}
+function configurePromotionsBrain(rt) {
+  runtime5 = rt;
+  snapshots4.clear();
+  warnedOnce4 = false;
+}
+function isPromotionsBrainActive() {
+  return runtime5 !== void 0;
+}
+function getApplicablePromotion(opts) {
+  const rt = runtime5;
+  if (!rt) return void 0;
+  const appId = opts.appId;
+  if (!appId || !opts.archetype) return void 0;
+  let snap = snapshots4.get(appId);
+  if (!snap) {
+    snap = { data: [], expiresAt: 0, refreshing: false };
+    snapshots4.set(appId, snap);
+  }
+  const now = Date.now();
+  const stale = snap.expiresAt <= now;
+  if (stale && !snap.refreshing) {
+    snap.refreshing = true;
+    void asyncRefresh5(rt, appId);
+  }
+  return snap.data.find((p) => p.archetype === opts.archetype);
+}
+var pendingRefreshes4 = /* @__PURE__ */ new Map();
+async function asyncRefresh5(rt, appId) {
+  const promise = doRefresh5(rt, appId);
+  pendingRefreshes4.set(appId, promise);
+  try {
+    await promise;
+  } finally {
+    if (pendingRefreshes4.get(appId) === promise) {
+      pendingRefreshes4.delete(appId);
+    }
+  }
+}
+async function doRefresh5(rt, appId) {
+  const url = `${rt.endpoint}?app_id=${encodeURIComponent(appId)}`;
+  let snap = snapshots4.get(appId);
+  if (!snap) {
+    snap = { data: [], expiresAt: 0, refreshing: false };
+    snapshots4.set(appId, snap);
+  }
+  try {
+    const res = await rt.fetchImpl(url, { method: "GET" });
+    if (!res.ok) {
+      throw new Error(`promotions ${res.status}: ${res.statusText}`);
+    }
+    const body = await res.json();
+    if (runtime5 !== rt) return;
+    const rows = Array.isArray(body) ? mapRowsToPromotions(body) : [];
+    snap.data = rows;
+    snap.expiresAt = Date.now() + rt.ttlMs;
+    snap.refreshing = false;
+  } catch (err) {
+    if (runtime5 !== rt) return;
+    snap.refreshing = false;
+    snap.expiresAt = Date.now() + rt.ttlMs;
+    if (!warnedOnce4) {
+      warnedOnce4 = true;
+      (rt.onError ?? defaultOnError5)(err);
+    }
+  }
+}
+function defaultOnError5(err) {
+  console.warn(
+    "[kgauto] promotions fetch failed (promotion boost inactive until next refresh):",
+    err
+  );
+}
+function _testResetPromotions() {
+  runtime5 = void 0;
+  snapshots4.clear();
+  pendingRefreshes4 = /* @__PURE__ */ new Map();
+  warnedOnce4 = false;
+}
+async function _testWaitForPromotionsRefresh() {
+  const pending = Array.from(pendingRefreshes4.values());
+  if (pending.length > 0) await Promise.all(pending);
+}
+
 // src/compile.ts
 var counter = 0;
 function makeHandle() {
@@ -4147,11 +4519,22 @@ function compile(ir, opts = {}) {
     ...toolFiltered.mutations,
     ...compressed.mutations
   ];
+  const activePromotion = getApplicablePromotion({
+    appId: ir.appId,
+    archetype: ir.intent.archetype
+  });
+  const promotion = activePromotion ? {
+    id: activePromotion.id,
+    promotedModel: activePromotion.promotedModel,
+    evalRunId: activePromotion.evalRunId,
+    suppressQualityGate: activePromotion.suppressQualityGate
+  } : void 0;
   const inputTokens = estimateInputTokens(workingIR);
   const scores = passScoreTargets(workingIR, {
     estimatedInputTokens: inputTokens,
     profilesById: resolver,
-    policy: opts.policy
+    policy: opts.policy,
+    promotion
   });
   accumulatedMutations.push(...scores.mutations);
   const target = pickTarget(workingIR, scores.value);
@@ -4483,11 +4866,23 @@ function configureBrain(config) {
   } else {
     configureExclusionFindingsBrain(void 0);
   }
+  const promotionsConsent = config.autoPromote ?? isAutoPromoteEnabledFromEnv();
+  if (promotionsConsent && bq.promotions !== false) {
+    configurePromotionsBrain({
+      endpoint: bq.promotionsEndpoint ?? DEFAULT_PROMOTIONS_ENDPOINT,
+      ttlMs: bq.cacheTtlMs ?? 3e5,
+      fetchImpl: config.fetchImpl ?? fetch,
+      onError: config.onError
+    });
+  } else {
+    configurePromotionsBrain(void 0);
+  }
 }
 function clearBrain() {
   activeConfig = void 0;
   configureBrainQuery(void 0);
   configureExclusionFindingsBrain(void 0);
+  configurePromotionsBrain(void 0);
 }
 var DEAD_LETTER_MAX_ENTRIES = 50;
 var DEAD_LETTER_MAX_ATTEMPTS = 3;
@@ -4604,7 +4999,7 @@ async function flushBrainDeadLetter() {
         if (entry.attempts >= DEAD_LETTER_MAX_ATTEMPTS) {
           const idx = deadLetter.indexOf(entry);
           if (idx !== -1) deadLetter.splice(idx, 1);
-          (config.onError ?? defaultOnError5)(
+          (config.onError ?? defaultOnError6)(
             new Error(
               `brain dead-letter: dropping ${entry.route} payload after ${entry.attempts} attempts \u2014 last error: ${entry.lastError}`
             )
@@ -4754,7 +5149,7 @@ async function record(input) {
     } catch (err) {
       noteFailure(err);
       pushDeadLetter("outcomes", config.endpoint, payload, err);
-      (config.onError ?? defaultOnError5)(err);
+      (config.onError ?? defaultOnError6)(err);
       return;
     }
     maybeAutoFlush();
@@ -4782,7 +5177,7 @@ async function record(input) {
     } catch (err) {
       noteFailure(err);
       pushDeadLetter("compile_outcome_advisories", config.endpoint, advisoryPayload, err);
-      (config.onError ?? defaultOnError5)(err);
+      (config.onError ?? defaultOnError6)(err);
     }
   };
   if (config.sync) {
@@ -4791,7 +5186,7 @@ async function record(input) {
     void send();
   }
 }
-function defaultOnError5(err) {
+function defaultOnError6(err) {
   console.warn("[kgauto] brain record failed:", err);
 }
 function describeBrainWriteFailure(status, route, body) {
@@ -4943,7 +5338,7 @@ async function recordOutcome(input) {
     } catch (err) {
       noteFailure(err);
       pushDeadLetter("compile_outcome_quality", config.endpoint, payload, err);
-      (config.onError ?? defaultOnError5)(err);
+      (config.onError ?? defaultOnError6)(err);
       return { ok: false, reason: "persistence_failed" };
     }
   };
@@ -5014,7 +5409,7 @@ async function recordShadowProbe(input) {
     } catch (err) {
       noteFailure(err);
       pushDeadLetter("probe_outcomes", config.endpoint, row, err);
-      (config.onError ?? defaultOnError5)(err);
+      (config.onError ?? defaultOnError6)(err);
     }
   };
   if (config.sync) {
@@ -5069,7 +5464,7 @@ async function recordGoldenIr(input) {
     } catch (err) {
       noteFailure(err);
       pushDeadLetter("golden_irs", config.endpoint, row, err);
-      (config.onError ?? defaultOnError5)(err);
+      (config.onError ?? defaultOnError6)(err);
     }
   };
   if (config.sync) {
@@ -5137,7 +5532,9 @@ var SUPPORTED_PROVIDERS = Object.freeze([
   "anthropic",
   "google",
   "openai",
-  "deepseek"
+  "deepseek",
+  "zai",
+  "moonshot"
 ]);
 function isSupportedProvider(p) {
   return SUPPORTED_PROVIDERS.includes(p);
@@ -5150,7 +5547,15 @@ var PROVIDER_ENV_KEYS = Object.freeze({
     "GOOGLE_GENERATIVE_AI_API_KEY"
   ]),
   openai: Object.freeze(["OPENAI_API_KEY"]),
-  deepseek: Object.freeze(["DEEPSEEK_API_KEY"])
+  deepseek: Object.freeze(["DEEPSEEK_API_KEY"]),
+  // alpha.65 — Z.ai (GLM family). ZAI_API_KEY is the canonical name; the
+  // Z_AI_API_KEY variant covers the underscore convention some tooling uses
+  // for the brand's "Z.ai" spelling.
+  zai: Object.freeze(["ZAI_API_KEY", "Z_AI_API_KEY"]),
+  // 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"])
 });
 function defaultEnv() {
   return typeof process !== "undefined" && process.env ? process.env : {};
@@ -5494,6 +5899,8 @@ function retryableError(status, code, message, raw) {
 var ANTHROPIC_URL2 = "https://api.anthropic.com/v1/messages";
 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";
 async function execute(request, opts = {}) {
   const merged = applyOverrides(request, opts.providerOverrides);
   switch (merged.provider) {
@@ -5505,6 +5912,18 @@ async function execute(request, opts = {}) {
       return executeOpenAI(merged, opts);
     case "deepseek":
       return executeDeepSeek(merged, opts);
+    case "zai":
+      return executeOpenAICompatible(
+        merged,
+        opts,
+        { provider: "zai", url: ZAI_URL, missingKeyMessage: "ZAI_API_KEY missing" }
+      );
+    case "moonshot":
+      return executeOpenAICompatible(
+        merged,
+        opts,
+        { provider: "moonshot", url: MOONSHOT_URL, missingKeyMessage: "MOONSHOT_API_KEY missing" }
+      );
     default: {
       const _exhaustive = merged;
       throw new Error(`execute(): no executor for provider: ${JSON.stringify(_exhaustive)}`);
@@ -5654,6 +6073,34 @@ async function executeDeepSeek(request, opts) {
   if (!res.ok) return classifyHttpError2(res.status, json);
   return { ok: true, status: res.status, response: normalizeOpenAILike(json) };
 }
+async function executeOpenAICompatible(request, opts, spec) {
+  const apiKey = resolveProviderKey(spec.provider, { apiKeys: opts.apiKeys });
+  if (!apiKey) {
+    return terminalError(401, "auth", spec.missingKeyMessage);
+  }
+  if (opts.onChunk) {
+    return streamOpenAILike(spec.url, request, apiKey, spec.provider, {
+      onChunk: opts.onChunk,
+      fetchImpl: opts.fetchImpl
+    });
+  }
+  const { provider: _provider, ...body } = request;
+  const fetchFn = opts.fetchImpl ?? fetch;
+  let res;
+  let json;
+  try {
+    res = await fetchFn(spec.url, {
+      method: "POST",
+      headers: { authorization: `Bearer ${apiKey}`, "content-type": "application/json" },
+      body: JSON.stringify(body)
+    });
+    json = await res.json().catch(() => ({}));
+  } catch (err) {
+    return retryableError2(0, "network_error", String(err), null);
+  }
+  if (!res.ok) return classifyHttpError2(res.status, json);
+  return { ok: true, status: res.status, response: normalizeOpenAILike(json) };
+}
 function normalizeOpenAILike(raw) {
   const r = raw;
   const choice = r.choices?.[0];
@@ -7769,7 +8216,7 @@ function createBrainForwardRoutes(config) {
 }
 
 // src/version.ts
-var LIBRARY_VERSION = "2.0.0-alpha.63";
+var LIBRARY_VERSION = "2.0.0-alpha.65";
 
 // src/key-health.ts
 var JSON_HEADERS2 = { "Content-Type": "application/json" };
@@ -7818,6 +8265,14 @@ var PROBE_SPECS = [
       return Number.isFinite(n) ? n : void 0;
     }
   },
+  {
+    provider: "moonshot",
+    canonicalEnvName: "MOONSHOT_API_KEY",
+    buildRequest: (key) => ({
+      url: "https://api.moonshot.ai/v1/models",
+      headers: { Authorization: `Bearer ${key}` }
+    })
+  },
   {
     provider: "google",
     canonicalEnvName: "GEMINI_API_KEY",
@@ -7849,6 +8304,11 @@ function createKeyHealthRoute(config) {
       const aiSdk = envKey(env, "GOOGLE_GENERATIVE_AI_API_KEY");
       key = gemini ?? google ?? aiSdk;
       envName = gemini ? "GEMINI_API_KEY" : google ? "GOOGLE_API_KEY" : aiSdk ? "GOOGLE_GENERATIVE_AI_API_KEY" : "GEMINI_API_KEY";
+    } else if (spec.provider === "moonshot") {
+      const moonshot = envKey(env, "MOONSHOT_API_KEY");
+      const kimi = envKey(env, "KIMI_API_KEY");
+      key = moonshot ?? kimi;
+      envName = moonshot ? "MOONSHOT_API_KEY" : kimi ? "KIMI_API_KEY" : "MOONSHOT_API_KEY";
     } else {
       key = envKey(env, spec.canonicalEnvName);
     }
@@ -8389,6 +8849,7 @@ function compile2(ir, opts) {
   ARCHETYPE_FLOOR_DEFAULT,
   CallError,
   DEFAULT_FINDINGS_ENDPOINT,
+  DEFAULT_PROMOTIONS_ENDPOINT,
   DIALECT_VERSION,
   FamilyResolutionError,
   INTENT_ARCHETYPES,
@@ -8400,6 +8861,8 @@ function compile2(ir, opts) {
   PROVIDER_ENV_KEYS,
   RULE_SEQUENTIAL_TOOL_CLIFF,
   TRANSLATOR_FLOOR,
+  _testResetPromotions,
+  _testWaitForPromotionsRefresh,
   allProfiles,
   applyArchetypeConvention,
   applySectionRewrites,
@@ -8419,6 +8882,7 @@ function compile2(ir, opts) {
   compile,
   compileForAISDKv6,
   configureBrain,
+  configurePromotionsBrain,
   countTokens,
   createBrainForwardRoutes,
   createKeyHealthRoute,
@@ -8430,6 +8894,7 @@ function compile2(ir, opts) {
   getActionableAdvisories,
   getAllStarterChains,
   getAllStarterChainsWithGrounding,
+  getApplicablePromotion,
   getArchetypePerfScore,
   getDefaultFallbackChain,
   getDefaultFallbackChainWithGrounding,
@@ -8445,10 +8910,12 @@ function compile2(ir, opts) {
   getStarterChainWithGrounding,
   hashShape,
   isArchetype,
+  isAutoPromoteEnabledFromEnv,
   isBrainQueryActiveFor,
   isBrainSync,
   isExclusionFindingsBrainActive,
   isModelReachable,
+  isPromotionsBrainActive,
   isProviderReachable,
   latencyTierOf,
   learningKey,