cartethyia 2.5.9 → 2.6.0

package/src/console/inferhub/types.ts

@@ -0,0 +1,183 @@
+/**
+ * InferHub market types.
+ *
+ * Two layers, deliberately separate. The raw shapes mirror the wire exactly,
+ * including the split the InferHub API makes between **decimal strings**
+ * (`/api/catalog`, `/api/budgets` — `$/Mtok` as `"3.5"`) and **numbers**
+ * (`/api/market` — `3.5`). Collapsing them into one type is how a string price
+ * silently becomes `NaN` or a number price becomes `"3.5" + 1`, so the two
+ * sources have two types and never share a field.
+ *
+ * Everything here is untrusted remote data: labels and `systemPromptNote` arrive
+ * from the marketplace and must be rendered as text, never interpolated into a
+ * prompt or a URL.
+ */
+
+/** One price tier: a distinct ask and how many live providers offer it. */
+export interface Tier {
+  readonly price: number;
+  /** Percentage below the official price; null when the official price is unknown. */
+  readonly discountPct: number | null;
+  /** Live providers offering this price. */
+  readonly avail: number;
+}
+
+/** Upstream family health, as InferHub reports it. */
+export type FamilyState = "operational" | "degraded" | "major_outage" | "low_traffic";
+
+export interface ModelHealth {
+  /** Null below InferHub's volume floor (10 requests). Null means unknown, not bad. */
+  readonly successRatePct: number | null;
+  /** Null below the sample floor (3 samples). */
+  readonly avgTtftMs: number | null;
+  readonly avgTps: number | null;
+  readonly familyState: FamilyState | null;
+}
+
+export interface ModelCapabilities {
+  readonly contextWindow: number | null;
+  readonly maxOutputTokens: number | null;
+  readonly modalities: readonly string[];
+  readonly outputModality: "text" | "image" | "video";
+  readonly reasoningLevels: readonly string[];
+  readonly supportsCache: boolean;
+}
+
+/** One normalized marketplace model. */
+export interface ModelMarket {
+  /** `"cc/claude-opus-5-5"` — the prefix is load-bearing and never stripped. */
+  readonly id: string;
+  readonly prefix: string;
+  readonly family: string;
+  readonly modelId: string;
+  /** Cross-prefix comparison key; see `normalize.ts`. */
+  readonly aliasKey: string;
+  readonly officialIn: number | null;
+  readonly officialOut: number | null;
+  /** Ascending by price. */
+  readonly tiersIn: readonly Tier[];
+  readonly tiersOut: readonly Tier[];
+  readonly minAskIn: number | null;
+  readonly minAskOut: number | null;
+  readonly maxAskIn: number | null;
+  readonly maxAskOut: number | null;
+  /** Zero means no live provider — listed, but never routed to. */
+  readonly providersLive: number;
+  readonly lastRate: number | null;
+  readonly cacheHitRate: number | null;
+  readonly health: ModelHealth;
+  readonly capabilities: ModelCapabilities;
+  /** Untrusted remote prose describing how the upstream rewrites the system prompt. */
+  readonly systemPromptNote: string | null;
+  /** Whether this model may be routed to right now. */
+  readonly routable: boolean;
+}
+
+/** Per-upstream-family summary for the providers view. */
+export interface ProviderMarket {
+  readonly prefix: string;
+  readonly family: string;
+  readonly activeProviders: number;
+  readonly familyState: FamilyState | null;
+  readonly availability24hPct: number | null;
+  readonly availability7dPct: number | null;
+  readonly p50DurationMs: number | null;
+  readonly p95DurationMs: number | null;
+  readonly avgTtftMs1h: number | null;
+  readonly avgTps1h: number | null;
+}
+
+/** Platform pricing configuration. */
+export interface PricingConfig {
+  readonly maxAskPctOfOfficial: number | null;
+  readonly platformFeePct: number | null;
+  readonly publisherSharePct: number | null;
+}
+
+/** Which upstream a snapshot came from, and when. */
+export type InferhubSource = "market" | "status" | "catalog" | "pricingConfig";
+
+/** A snapshot plus its freshness, as the API reports it. */
+export interface SnapshotMeta {
+  readonly fetchedAt: number;
+  readonly ageSeconds: number;
+  /** True when the snapshot is older than its freshness window. */
+  readonly stale: boolean;
+}
+
+// ── Raw wire shapes ────────────────────────────────────────────────────────
+
+/** `GET /api/market` — money is numeric here. */
+export interface RawMarketModel {
+  readonly slug?: unknown;
+  readonly family?: unknown;
+  readonly minAskIn?: unknown;
+  readonly minAskOut?: unknown;
+  readonly maxAskIn?: unknown;
+  readonly maxAskOut?: unknown;
+  readonly lastRate?: unknown;
+}
+
+/** `GET /api/catalog` — money is a decimal string here. */
+export interface RawCatalogModel {
+  readonly id?: unknown;
+  /** The routable slug, e.g. `claude-haiku-4-5`. `id` is a UUID. */
+  readonly upstreamModelId?: unknown;
+  readonly label?: unknown;
+  readonly officialIn?: unknown;
+  readonly officialOut?: unknown;
+  /**
+   * `[price, providerCount]` tuples — InferHub pre-aggregates tiers, so one
+   * entry is one distinct price and its live-provider count. Money is numeric
+   * here even though `officialIn`/`officialOut` are decimal strings.
+   */
+  readonly pricePointsIn?: unknown;
+  readonly pricePointsOut?: unknown;
+  readonly asksIn?: unknown;
+  readonly asksOut?: unknown;
+  readonly enabled?: unknown;
+  readonly modelDisabled?: unknown;
+  readonly supportsCache?: unknown;
+  readonly cacheHitRate?: unknown;
+  readonly contextWindow?: unknown;
+  readonly maxOutputTokens?: unknown;
+  readonly reasoningLevels?: unknown;
+  readonly modalities?: unknown;
+  readonly outputModality?: unknown;
+}
+
+export interface RawUpstreamCard {
+  readonly id?: unknown;
+  readonly slug?: unknown;
+  readonly prefix?: unknown;
+  readonly label?: unknown;
+  readonly status?: unknown;
+  readonly activeProviders?: unknown;
+  readonly systemPromptNote?: unknown;
+  readonly models?: unknown;
+}
+
+/** `GET /api/status` — per-model health and per-family state. */
+export interface RawStatusModel {
+  /** The full `prefix/model` key. `slug` is the fallback for older payloads. */
+  readonly id?: unknown;
+  readonly slug?: unknown;
+  readonly successRatePct?: unknown;
+  readonly avgTtftMs?: unknown;
+  readonly avgTps?: unknown;
+}
+
+export interface RawStatusFamily {
+  readonly slug?: unknown;
+  readonly prefix?: unknown;
+  /** Display name, e.g. `Antigravity`. `slug` is the lowercase key. */
+  readonly family?: unknown;
+  readonly currentAvailabilityPct?: unknown;
+  readonly state?: unknown;
+  readonly availability24hPct?: unknown;
+  readonly availability7dPct?: unknown;
+  readonly p50DurationMs?: unknown;
+  readonly p95DurationMs?: unknown;
+  readonly avgTtftMs1h?: unknown;
+  readonly avgTps1h?: unknown;
+}

package/src/console/observability/contracts.ts

@@ -1,12 +1,10 @@
 import type { ConsoleAccessResolver } from "../auth/access";
 // Telemetry domain: system health, usage, and telemetry-event read API.
 
-import { readdir } from "node:fs/promises";
 import { ConsoleDomainError, requireTenantScope } from "../shared/errors";
 import { conditionalResponse, type CachedResult } from "../shared/read-through-cache";
 import { parseQueryLimit } from "../shared/query";
 import type { AccessDecision } from "../../security/access-control";
-import { resolveTelemetryCompactAfterDays } from "../../config/telemetry";
 import { USAGE_DIMENSIONS, type UsageDimension } from "./usage-dimensions";
 
 /**
@@ -23,31 +21,6 @@ export function isSupportedUsagePeriod(period: string): boolean {
   return Number.isSafeInteger(amount) && amount >= 1 && amount <= cap;
 }
 
-/**
- * Why a request detail cannot be shown.
- *
- * A compacted request is genuinely gone: past the compaction age its row was
- * folded into a per-day summary, and the individual request no longer exists
- * anywhere. That is by design, not data loss — but "Request not found" reads
- * like the latter, and an operator who thinks history was lost will go looking
- * for a bug that is not there.
- *
- * The age is checked against the summary files rather than assumed: a request
- * from before retention is aged out, not compacted, and the two deserve
- * different words.
- */
-async function detailNotFoundMessage(requestId: string, noun = "Request"): Promise<string> {
-  const compacted = await readdir("./data/telemetry-events").catch(() => [] as string[]);
-  const hasSummaries = compacted.some((file) => file.endsWith(".index.json"));
-  if (!hasSummaries) return `${noun} ${requestId} not found`;
-  return (
-    `${noun} ${requestId} is no longer stored individually: requests older than ` +
-    `${resolveTelemetryCompactAfterDays()} day(s) are compacted into daily summaries. ` +
-    `Its totals still count in Usage — raise CARTETHYIA_TELEMETRY_COMPACT_AFTER_DAYS ` +
-    `to keep individual requests longer.`
-  );
-}
-
 export { USAGE_PERIODS, type UsagePeriod } from "./usage-periods";
 
 export interface SystemHealthResponse {
@@ -413,13 +386,7 @@ export function createObservabilityOperations(config: ObservabilityConfig) {
           includePayload,
           createdAt,
         );
-        if (!detail) {
-          throw new ConsoleDomainError(
-            "event_not_found",
-            404,
-            await detailNotFoundMessage(eventId, "Event"),
-          );
-        }
+        if (!detail) throw new ConsoleDomainError("event_not_found", 404, `Event ${eventId} not found`);
         return detail;
       },
     requirePeriod(period: string): string {
@@ -509,13 +476,7 @@ export function createObservabilityOperations(config: ObservabilityConfig) {
       ): Promise<UsageRequestDetail> {
         const tenantId = operations.requireTenant(access);
         const detail = await config.store.usageRequestDetail(tenantId, requestId);
-        if (!detail) {
-          throw new ConsoleDomainError(
-            "event_not_found",
-            404,
-            await detailNotFoundMessage(requestId),
-          );
-        }
+        if (!detail) throw new ConsoleDomainError("event_not_found", 404, `Request ${requestId} not found`);
         const payloads = detail.payloads ?? null;
         return { ...detail, payloads };
       },