@warmdrift/kgauto-compiler 2.0.0-alpha.21 → 2.0.0-alpha.23

package/dist/index.js

@@ -43,6 +43,7 @@ __export(index_exports, {
   getArchetypePerfScore: () => getArchetypePerfScore,
   getDefaultFallbackChain: () => getDefaultFallbackChain,
   getDefaultFallbackChainWithGrounding: () => getDefaultFallbackChainWithGrounding,
+  getPerAxisMetrics: () => getPerAxisMetrics,
   getProfile: () => getProfile,
   getReachabilityDiagnostic: () => getReachabilityDiagnostic,
   getSequentialStarterChain: () => getSequentialStarterChain,
@@ -51,6 +52,7 @@ __export(index_exports, {
   getStarterChainWithGrounding: () => getStarterChainWithGrounding,
   hashShape: () => hashShape,
   isArchetype: () => isArchetype,
+  isBrainQueryActiveFor: () => isBrainQueryActiveFor,
   isModelReachable: () => isModelReachable,
   isProviderReachable: () => isProviderReachable,
   learningKey: () => learningKey,
@@ -1944,14 +1946,250 @@ function profilesByProvider(provider) {
   return PROFILES_RAW.filter((p) => p.provider === provider);
 }
 
+// src/brain-query.ts
+var FRESH_SNAPSHOT = {
+  data: null,
+  expiresAt: 0,
+  refreshing: false,
+  warned: false
+};
+var snapshot = { ...FRESH_SNAPSHOT };
+var runtime;
+function configureBrainQuery(rt) {
+  runtime = rt;
+  snapshot = { ...FRESH_SNAPSHOT };
+}
+function createBrainQueryCache(opts) {
+  return () => {
+    const rt = runtime;
+    if (!rt || !rt.enabledTables.has(opts.table)) {
+      return opts.bundledFallback();
+    }
+    const now = Date.now();
+    const stale = snapshot.expiresAt <= now;
+    if (stale && !snapshot.refreshing) {
+      snapshot.refreshing = true;
+      void asyncRefresh(rt);
+    }
+    if (snapshot.data) {
+      const rows = snapshot.data[opts.table];
+      if (Array.isArray(rows) && rows.length > 0) {
+        try {
+          return opts.mapRows(rows);
+        } catch {
+          return opts.bundledFallback();
+        }
+      }
+    }
+    return opts.bundledFallback();
+  };
+}
+var pendingRefresh;
+async function asyncRefresh(rt) {
+  const promise = doRefresh(rt);
+  pendingRefresh = promise;
+  try {
+    await promise;
+  } finally {
+    if (pendingRefresh === promise) pendingRefresh = void 0;
+  }
+}
+var DEFAULT_CONFIG_URL = "https://kgauto-dashboard.vercel.app/api/kgauto-v2/config";
+async function doRefresh(rt) {
+  const url = rt.configEndpoint ?? DEFAULT_CONFIG_URL;
+  try {
+    const res = await rt.fetchImpl(url, { method: "GET" });
+    if (!res.ok) {
+      throw new Error(`brain-query ${res.status}: ${res.statusText}`);
+    }
+    const body = await res.json();
+    if (runtime !== rt) return;
+    snapshot = {
+      data: body,
+      expiresAt: Date.now() + rt.ttlMs,
+      refreshing: false,
+      warned: snapshot.warned
+    };
+  } catch (err) {
+    if (runtime !== rt) return;
+    snapshot.refreshing = false;
+    snapshot.expiresAt = Date.now() + rt.ttlMs;
+    if (!snapshot.warned) {
+      snapshot.warned = true;
+      (rt.onError ?? defaultOnError)(err);
+    }
+  }
+}
+function defaultOnError(err) {
+  console.warn("[kgauto] brain-query failed (using bundled fallback):", err);
+}
+function isBrainQueryActiveFor(table) {
+  return runtime !== void 0 && runtime.enabledTables.has(table);
+}
+async function getPerAxisMetrics(opts) {
+  const fetchFn = opts.fetch ?? fetch;
+  const endpoint = opts.endpoint ?? runtime?.endpoint;
+  if (!endpoint) return null;
+  const windowDays = opts.windowDays ?? 30;
+  const body = {
+    p_app_id: opts.appId,
+    p_archetype: opts.archetype,
+    p_model: opts.model,
+    p_window_days: windowDays,
+    p_quality_floor: opts.qualityFloor ?? null
+  };
+  const headers = {
+    Accept: "application/json",
+    "Content-Type": "application/json",
+    ...opts.apiKey ? { Authorization: `Bearer ${opts.apiKey}` } : {}
+  };
+  try {
+    const res = await fetchFn(`${endpoint}/rpc/get_per_axis_metrics`, {
+      method: "POST",
+      headers,
+      body: JSON.stringify(body)
+    });
+    if (!res.ok) return null;
+    const raw = await res.json();
+    return mapPerAxisMetrics(raw, opts.appId, opts.archetype, opts.model, windowDays);
+  } catch {
+    return null;
+  }
+}
+function mapPerAxisMetrics(raw, fallbackAppId, fallbackArchetype, fallbackModel, fallbackWindowDays) {
+  if (raw === null || raw === void 0) return null;
+  if (typeof raw !== "object") return null;
+  const r = raw;
+  if (Array.isArray(raw)) {
+    if (raw.length === 0) return null;
+    return mapPerAxisMetrics(raw[0], fallbackAppId, fallbackArchetype, fallbackModel, fallbackWindowDays);
+  }
+  const num = (v) => {
+    if (v === null || v === void 0) return null;
+    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;
+  };
+  const int = (v) => {
+    const n = num(v);
+    return n === null ? 0 : Math.trunc(n);
+  };
+  const bool = (v) => {
+    if (v === null || v === void 0) return null;
+    if (typeof v === "boolean") return v;
+    return null;
+  };
+  const str = (v, fallback) => typeof v === "string" ? v : fallback;
+  const cost = r.cost_efficiency ?? {};
+  const time = r.time_efficiency ?? {};
+  const rel = r.reliability ?? {};
+  return {
+    appId: str(r.app_id, fallbackAppId),
+    archetype: str(r.archetype, fallbackArchetype),
+    model: str(r.model, fallbackModel),
+    windowDays: num(r.window_days) ?? fallbackWindowDays,
+    nRows: int(r.n_rows),
+    nRowsClean: int(r.n_rows_clean),
+    nQualityOutcomes: int(r.n_quality_outcomes),
+    magicRate: num(r.magic_rate),
+    qualityFloorMet: bool(r.quality_floor_met),
+    costEfficiency: {
+      avgCostUsd: num(cost.avg_cost_usd),
+      avgCostUsdClean: num(cost.avg_cost_usd_clean),
+      avgInputTokens: num(cost.avg_input_tokens),
+      avgOutputTokens: num(cost.avg_output_tokens),
+      inputTokenRatio: num(cost.input_token_ratio)
+    },
+    timeEfficiency: {
+      avgLatencyMs: num(time.avg_latency_ms),
+      avgTtftMs: num(time.avg_ttft_ms)
+    },
+    reliability: {
+      successRate: num(rel.success_rate),
+      successRateClean: num(rel.success_rate_clean),
+      emptyRate: num(rel.empty_rate),
+      emptyRateClean: num(rel.empty_rate_clean)
+    },
+    evidenceFreshnessDays: num(r.evidence_freshness_days)
+  };
+}
+
+// src/archetype-perf-brain.ts
+function isPerfRow(x) {
+  if (!x || typeof x !== "object") return false;
+  const r = x;
+  return typeof r.model_id === "string" && typeof r.archetype === "string" && typeof r.perf_score === "number";
+}
+function mapRowsToPerfMap(rows) {
+  const out = /* @__PURE__ */ new Map();
+  for (const row of rows) {
+    if (!isPerfRow(row)) continue;
+    const existing = out.get(row.model_id) ?? {};
+    existing[row.archetype] = row.perf_score;
+    out.set(row.model_id, existing);
+  }
+  return out;
+}
+function mapRowsToNMap(rows) {
+  const out = /* @__PURE__ */ new Map();
+  for (const row of rows) {
+    if (!isPerfRow(row)) continue;
+    if (typeof row.n !== "number") continue;
+    const existing = out.get(row.model_id) ?? {};
+    existing[row.archetype] = row.n;
+    out.set(row.model_id, existing);
+  }
+  return out;
+}
+function bundledArchetypePerf() {
+  const out = /* @__PURE__ */ new Map();
+  for (const profile of allProfiles()) {
+    if (profile.archetypePerf) out.set(profile.id, profile.archetypePerf);
+  }
+  return out;
+}
+function bundledArchetypePerfN() {
+  return /* @__PURE__ */ new Map();
+}
+var loadArchetypePerfFromBrain = createBrainQueryCache({
+  table: "kgauto_archetype_perf",
+  mapRows: mapRowsToPerfMap,
+  bundledFallback: bundledArchetypePerf
+});
+var loadArchetypePerfNFromBrain = createBrainQueryCache(
+  {
+    table: "kgauto_archetype_perf",
+    mapRows: mapRowsToNMap,
+    bundledFallback: bundledArchetypePerfN
+  }
+);
+var MEASURED_GROUNDING_MIN_N = 10;
+function getArchetypePerfScore(modelId, archetype) {
+  const score = loadArchetypePerfFromBrain().get(modelId)?.[archetype] ?? 5;
+  const n = loadArchetypePerfNFromBrain().get(modelId)?.[archetype] ?? 0;
+  const grounding = n >= MEASURED_GROUNDING_MIN_N ? "measured" : "judgment";
+  return { score, n, grounding };
+}
+
 // src/advisor.ts
-function runAdvisor(ir, result, profile, policy) {
+var QUALITY_FLOOR_FOR_RECOMMENDATION = 6;
+var TIER_DOWN_COST_RATIO = 0.5;
+var COST_MISMATCHED_CHOSEN_SCORE_CEILING = 7;
+function runAdvisor(ir, result, profile, policy, phase2) {
   const out = [];
   out.push(...detectCachingOff(ir, profile));
   out.push(...detectSingleChunkSystem(ir, profile));
   out.push(...detectToolBloat(ir, result));
   out.push(...detectHistoryUncached(ir, profile));
   out.push(...detectSingleModelArray(ir, policy));
+  if (policy?.posture !== "locked") {
+    out.push(...detectCostMismatchedArchetype(ir, profile, phase2));
+    out.push(...detectModelStaleEvidence(ir, profile));
+    out.push(...detectTierDown(ir, profile, phase2));
+  }
   return out;
 }
 function detectCachingOff(ir, profile) {
@@ -2037,6 +2275,91 @@ function detectSingleModelArray(ir, policy) {
     }
   ];
 }
+function detectCostMismatchedArchetype(ir, profile, phase2) {
+  if (!phase2 || phase2.fallbackChain.length === 0) return [];
+  if (!phase2.profileResolver) return [];
+  const archetype = ir.intent.archetype;
+  const chosenScore = getArchetypePerfScore(profile.id, archetype);
+  const chosenHasRoomToGrow = chosenScore.grounding === "judgment" || chosenScore.score < COST_MISMATCHED_CHOSEN_SCORE_CEILING;
+  if (!chosenHasRoomToGrow) return [];
+  let bestAlt = null;
+  for (const altId of phase2.fallbackChain) {
+    const altProfile = phase2.profileResolver(altId);
+    if (!altProfile) continue;
+    if (altProfile.id === profile.id) continue;
+    const altScore = getArchetypePerfScore(altProfile.id, archetype);
+    if (altScore.score < QUALITY_FLOOR_FOR_RECOMMENDATION) continue;
+    if (altScore.score < chosenScore.score) continue;
+    if (altProfile.costInputPer1m >= profile.costInputPer1m) continue;
+    if (!bestAlt || altScore.score > bestAlt.score.score || altScore.score === bestAlt.score.score && altProfile.costInputPer1m < bestAlt.profile.costInputPer1m) {
+      bestAlt = { id: altId, profile: altProfile, score: altScore };
+    }
+  }
+  if (!bestAlt) return [];
+  const tierDownWouldFire = bestAlt.score.grounding === "measured" && bestAlt.profile.costInputPer1m <= profile.costInputPer1m * TIER_DOWN_COST_RATIO;
+  if (tierDownWouldFire) return [];
+  const chosenGrounding = chosenScore.grounding === "judgment" ? `archetypePerf.${archetype}=judgment` : `archetypePerf.${archetype}=${chosenScore.score}`;
+  const altGrounding = bestAlt.score.grounding === "measured" ? `archetypePerf.${archetype}=${bestAlt.score.score}, measured, n=${bestAlt.score.n}` : `archetypePerf.${archetype}=${bestAlt.score.score}, judgment`;
+  return [
+    {
+      level: "warn",
+      code: "cost-mismatched-archetype",
+      message: `Cost-mismatched-archetype: target=${profile.id} (${chosenGrounding}) selected for ${archetype}. Alternative ${bestAlt.id} (${altGrounding}) is cheaper ($${bestAlt.profile.costInputPer1m}/$${bestAlt.profile.costOutputPer1m} vs $${profile.costInputPer1m}/$${profile.costOutputPer1m} per 1M) at equal-or-better quality.`,
+      suggestion: `Consider declaring \`${bestAlt.id}\` as the primary model for this archetype, or relax to posture='open' to let kgauto select among the chain. If the chosen model is required for compliance/brand reasons, set \`policy.posture = 'locked'\` to silence this rule.`,
+      recommendationType: profile.provider === bestAlt.profile.provider ? "tier-down" : "model-swap",
+      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
+    }
+  ];
+}
+function detectModelStaleEvidence(ir, profile) {
+  if (!isBrainQueryActiveFor("kgauto_archetype_perf")) return [];
+  const archetype = ir.intent.archetype;
+  const chosen = getArchetypePerfScore(profile.id, archetype);
+  if (chosen.grounding !== "judgment") return [];
+  return [
+    {
+      level: "info",
+      code: "model-stale-evidence",
+      message: `Model-stale-evidence: target=${profile.id} archetype=${archetype} is judgment-grounded (n=${chosen.n}) despite brain-query mode being active. Measurement substrate is wired but the brain hasn't accumulated >=10 outcomes for this (model, archetype) tuple yet \u2014 routing decisions remain pre-measured for this slot.`,
+      suggestion: "Verify that `record()` is being called on every call() outcome with the appropriate `actualModel` and `mutationsApplied` fields. Once the brain accumulates n>=10 rows on this tuple, the score promotes from judgment to measured automatically (5-min SWR cache). No code change required from your side \u2014 this is the substrate signaling the gap.",
+      recommendationType: "prompt-fix",
+      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
+    }
+  ];
+}
+function detectTierDown(ir, profile, phase2) {
+  if (!phase2 || phase2.fallbackChain.length === 0) return [];
+  if (!phase2.profileResolver) return [];
+  const archetype = ir.intent.archetype;
+  const chosenScore = getArchetypePerfScore(profile.id, archetype);
+  const chosenCost = profile.costInputPer1m;
+  let bestAlt = null;
+  for (const altId of phase2.fallbackChain) {
+    const altProfile = phase2.profileResolver(altId);
+    if (!altProfile) continue;
+    if (altProfile.id === profile.id) continue;
+    const altScore = getArchetypePerfScore(altProfile.id, archetype);
+    if (altScore.grounding !== "measured") continue;
+    if (altScore.score < QUALITY_FLOOR_FOR_RECOMMENDATION) continue;
+    if (altScore.score < chosenScore.score) continue;
+    if (altProfile.costInputPer1m > chosenCost * TIER_DOWN_COST_RATIO) continue;
+    if (!bestAlt || altProfile.costInputPer1m < bestAlt.profile.costInputPer1m || altProfile.costInputPer1m === bestAlt.profile.costInputPer1m && altScore.score > bestAlt.score.score) {
+      bestAlt = { id: altId, profile: altProfile, score: altScore };
+    }
+  }
+  if (!bestAlt) return [];
+  const chosenDesc = chosenScore.grounding === "measured" ? `archetypePerf.${archetype}=${chosenScore.score} (measured, n=${chosenScore.n})` : `archetypePerf.${archetype}=${chosenScore.score} (${chosenScore.grounding})`;
+  return [
+    {
+      level: "warn",
+      code: "tier-down",
+      message: `Tier-down: target=${profile.id} (${chosenDesc}) selected for ${archetype}. Brain shows ${bestAlt.id} delivers equal-or-better quality (archetypePerf.${archetype}=${bestAlt.score.score}, measured, n=${bestAlt.score.n}) at $${bestAlt.profile.costInputPer1m}/$${bestAlt.profile.costOutputPer1m} per 1M vs $${profile.costInputPer1m}/$${profile.costOutputPer1m} \u2014 a measured tier-down opportunity.`,
+      suggestion: `Move \`${bestAlt.id}\` to primary for this archetype. The brain has n=${bestAlt.score.n} measured outcomes backing the recommendation; this is data, not opinion. If posture='locked' is required (compliance/brand promise), set it explicitly to silence this rule.`,
+      recommendationType: "tier-down",
+      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
+    }
+  ];
+}
 
 // src/compile.ts
 var counter = 0;
@@ -2112,6 +2435,13 @@ function compile(ir, opts = {}) {
       description: "ir.constraints.toolOrchestration='sequential' selected the DeepSeek-tier-0 hunt chain overlay (L-040 parallel-tool cliff doesn't apply at single-step granularity)."
     });
   }
+  const phase2ProfileResolver = opts.profileResolver ? (id) => {
+    try {
+      return opts.profileResolver(id);
+    } catch {
+      return void 0;
+    }
+  } : tryGetProfile;
   const advisories = runAdvisor(
     ir,
     {
@@ -2121,7 +2451,11 @@ function compile(ir, opts = {}) {
       diagnostics
     },
     profile,
-    opts.policy
+    opts.policy,
+    {
+      fallbackChain,
+      profileResolver: phase2ProfileResolver
+    }
   );
   return {
     handle,
@@ -2174,84 +2508,6 @@ function validateFinalFit(ir, profile, tokens) {
   }
 }
 
-// src/brain-query.ts
-var FRESH_SNAPSHOT = {
-  data: null,
-  expiresAt: 0,
-  refreshing: false,
-  warned: false
-};
-var snapshot = { ...FRESH_SNAPSHOT };
-var runtime;
-function configureBrainQuery(rt) {
-  runtime = rt;
-  snapshot = { ...FRESH_SNAPSHOT };
-}
-function createBrainQueryCache(opts) {
-  return () => {
-    const rt = runtime;
-    if (!rt || !rt.enabledTables.has(opts.table)) {
-      return opts.bundledFallback();
-    }
-    const now = Date.now();
-    const stale = snapshot.expiresAt <= now;
-    if (stale && !snapshot.refreshing) {
-      snapshot.refreshing = true;
-      void asyncRefresh(rt);
-    }
-    if (snapshot.data) {
-      const rows = snapshot.data[opts.table];
-      if (Array.isArray(rows) && rows.length > 0) {
-        try {
-          return opts.mapRows(rows);
-        } catch {
-          return opts.bundledFallback();
-        }
-      }
-    }
-    return opts.bundledFallback();
-  };
-}
-var pendingRefresh;
-async function asyncRefresh(rt) {
-  const promise = doRefresh(rt);
-  pendingRefresh = promise;
-  try {
-    await promise;
-  } finally {
-    if (pendingRefresh === promise) pendingRefresh = void 0;
-  }
-}
-var DEFAULT_CONFIG_URL = "https://kgauto-dashboard.vercel.app/api/kgauto-v2/config";
-async function doRefresh(rt) {
-  const url = rt.configEndpoint ?? DEFAULT_CONFIG_URL;
-  try {
-    const res = await rt.fetchImpl(url, { method: "GET" });
-    if (!res.ok) {
-      throw new Error(`brain-query ${res.status}: ${res.statusText}`);
-    }
-    const body = await res.json();
-    if (runtime !== rt) return;
-    snapshot = {
-      data: body,
-      expiresAt: Date.now() + rt.ttlMs,
-      refreshing: false,
-      warned: snapshot.warned
-    };
-  } catch (err) {
-    if (runtime !== rt) return;
-    snapshot.refreshing = false;
-    snapshot.expiresAt = Date.now() + rt.ttlMs;
-    if (!snapshot.warned) {
-      snapshot.warned = true;
-      (rt.onError ?? defaultOnError)(err);
-    }
-  }
-}
-function defaultOnError(err) {
-  console.warn("[kgauto] brain-query failed (using bundled fallback):", err);
-}
-
 // src/pricing-brain.ts
 function isPricingRow(x) {
   if (!x || typeof x !== "object") return false;
@@ -3923,63 +4179,6 @@ function clamp(n) {
   return Math.max(0, Math.min(1, n));
 }
 
-// src/archetype-perf-brain.ts
-function isPerfRow(x) {
-  if (!x || typeof x !== "object") return false;
-  const r = x;
-  return typeof r.model_id === "string" && typeof r.archetype === "string" && typeof r.perf_score === "number";
-}
-function mapRowsToPerfMap(rows) {
-  const out = /* @__PURE__ */ new Map();
-  for (const row of rows) {
-    if (!isPerfRow(row)) continue;
-    const existing = out.get(row.model_id) ?? {};
-    existing[row.archetype] = row.perf_score;
-    out.set(row.model_id, existing);
-  }
-  return out;
-}
-function mapRowsToNMap(rows) {
-  const out = /* @__PURE__ */ new Map();
-  for (const row of rows) {
-    if (!isPerfRow(row)) continue;
-    if (typeof row.n !== "number") continue;
-    const existing = out.get(row.model_id) ?? {};
-    existing[row.archetype] = row.n;
-    out.set(row.model_id, existing);
-  }
-  return out;
-}
-function bundledArchetypePerf() {
-  const out = /* @__PURE__ */ new Map();
-  for (const profile of allProfiles()) {
-    if (profile.archetypePerf) out.set(profile.id, profile.archetypePerf);
-  }
-  return out;
-}
-function bundledArchetypePerfN() {
-  return /* @__PURE__ */ new Map();
-}
-var loadArchetypePerfFromBrain = createBrainQueryCache({
-  table: "kgauto_archetype_perf",
-  mapRows: mapRowsToPerfMap,
-  bundledFallback: bundledArchetypePerf
-});
-var loadArchetypePerfNFromBrain = createBrainQueryCache(
-  {
-    table: "kgauto_archetype_perf",
-    mapRows: mapRowsToNMap,
-    bundledFallback: bundledArchetypePerfN
-  }
-);
-var MEASURED_GROUNDING_MIN_N = 10;
-function getArchetypePerfScore(modelId, archetype) {
-  const score = loadArchetypePerfFromBrain().get(modelId)?.[archetype] ?? 5;
-  const n = loadArchetypePerfNFromBrain().get(modelId)?.[archetype] ?? 0;
-  const grounding = n >= MEASURED_GROUNDING_MIN_N ? "measured" : "judgment";
-  return { score, n, grounding };
-}
-
 // src/models-brain.ts
 function isModelRow(x) {
   if (!x || typeof x !== "object") return false;
@@ -4130,6 +4329,7 @@ function compile2(ir, opts) {
   getArchetypePerfScore,
   getDefaultFallbackChain,
   getDefaultFallbackChainWithGrounding,
+  getPerAxisMetrics,
   getProfile,
   getReachabilityDiagnostic,
   getSequentialStarterChain,
@@ -4138,6 +4338,7 @@ function compile2(ir, opts) {
   getStarterChainWithGrounding,
   hashShape,
   isArchetype,
+  isBrainQueryActiveFor,
   isModelReachable,
   isProviderReachable,
   learningKey,