fest-gateway 0.1.0

package/dist/server/usage/prices/catalog.js

@@ -0,0 +1,136 @@
+/**
+ * litellm's price file, normalised into Fest's rate shape.
+ *
+ * This module is the ONLY place that knows litellm's JSON schema. Both the
+ * offline sync tool (`tools/prices-sync.ts`) and the runtime refresh
+ * (`refresh.ts`) go through `normaliseCatalog`, so the vendored snapshot and a
+ * freshly fetched file can never be two different formats.
+ *
+ * Why normalise at all rather than ship litellm's file verbatim:
+ *
+ *  - litellm prices PER TOKEN in floats like `5e-7`. Fest prices per million,
+ *    because that is how vendors publish and how a human checks the number.
+ *    Converting once at build time keeps float noise out of the hot path.
+ *  - The raw file is 2.9MB and mostly fields Fest will never read (image,
+ *    audio, video, batch, OCR, embedding rates). Filtering to chat models and
+ *    the five token buckets Fest actually meters takes it to ~700KB, and ~48KB
+ *    gzipped — small enough to vendor.
+ *  - Tier suffixes are encoded in KEY NAMES (`input_cost_per_token_above_200k_tokens`),
+ *    with thresholds that vary by model: 128k, 200k, 256k, 272k, 512k all
+ *    appear. Discovering them by parsing the suffix is the only thing that
+ *    doesn't rot; a hardcoded 200k silently understates every other model.
+ */
+export const LITELLM_URL = "https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json";
+const PER_MILLION = 1_000_000;
+function numeric(v) {
+    return typeof v === "number" && Number.isFinite(v) && v >= 0 ? v : null;
+}
+/**
+ * Build one tier from a suffixed family of litellm keys.
+ *
+ * Returns null unless BOTH input and output are published — a tier with only
+ * half its rates would price every call short, which is exactly the silent
+ * undercount this codebase treats as the worst failure mode.
+ *
+ * The cache fallbacks are deliberate and worth stating plainly: when a provider
+ * publishes no cache rate, cached tokens are billed as ordinary input, so the
+ * input rate is the correct fallback. Zero would claim cache reads are free,
+ * and Anthropic's ×0.1 / ×1.25 / ×2.0 multipliers are an *Anthropic* policy —
+ * applying them to Fireworks would be inventing a discount nobody offers.
+ */
+function tierFrom(entry, suffix) {
+    const input = numeric(entry[`input_cost_per_token${suffix}`]);
+    const output = numeric(entry[`output_cost_per_token${suffix}`]);
+    if (input === null || output === null)
+        return null;
+    const cacheRead = numeric(entry[`cache_read_input_token_cost${suffix}`]);
+    const write5m = numeric(entry[`cache_creation_input_token_cost${suffix}`]);
+    // Anthropic spells the 1h variant with `_above_1hr` BEFORE the tier suffix.
+    const write1h = numeric(entry[`cache_creation_input_token_cost_above_1hr${suffix}`]) ??
+        numeric(entry[`cache_creation_input_token_cost_above_1hr`]);
+    return {
+        cachePublished: cacheRead !== null || write5m !== null,
+        rates: {
+            inputPerMillion: input * PER_MILLION,
+            outputPerMillion: output * PER_MILLION,
+            cacheReadPerMillion: (cacheRead ?? input) * PER_MILLION,
+            cacheWrite5mPerMillion: (write5m ?? input) * PER_MILLION,
+            cacheWrite1hPerMillion: (write1h ?? write5m ?? input) * PER_MILLION,
+        },
+    };
+}
+/** `input_cost_per_token_above_200k_tokens` → 200_000. Ignores flex/priority tiers. */
+const TIER_SUFFIX = /^input_cost_per_token_above_(\d+)k_tokens$/;
+function tierSuffixes(entry) {
+    const found = [];
+    for (const key of Object.keys(entry)) {
+        const m = TIER_SUFFIX.exec(key);
+        if (m?.[1] === undefined)
+            continue;
+        found.push({ above: Number(m[1]) * 1000, suffix: `_above_${m[1]}k_tokens` });
+    }
+    return found.sort((a, b) => a.above - b.above);
+}
+/**
+ * Web search, normalised to USD per thousand queries to match how Fest counts
+ * `usage.webSearches`. litellm publishes it per query, keyed by context size;
+ * Fest cannot know which size was requested, so it takes the medium rate as the
+ * representative one and falls back to whichever is present.
+ */
+function webSearchPerThousand(entry) {
+    const raw = entry["search_context_cost_per_query"];
+    if (raw === null || typeof raw !== "object")
+        return null;
+    const sizes = raw;
+    const perQuery = numeric(sizes["search_context_size_medium"]) ??
+        numeric(sizes["search_context_size_low"]) ??
+        numeric(sizes["search_context_size_high"]);
+    return perQuery === null ? null : perQuery * 1000;
+}
+/**
+ * Filter litellm's raw file down to priceable chat models.
+ *
+ * Non-chat modes (embedding, image, audio, rerank) are dropped: Fest meters
+ * `/v1/messages` and nothing else, so an image model's per-pixel rate is weight
+ * with no reader. A chat model with no usable base tier is dropped too —
+ * carrying an entry we cannot price would turn an honest "unknown model" into a
+ * confident zero.
+ */
+export function normaliseCatalog(raw, opts) {
+    if (raw === null || typeof raw !== "object") {
+        throw new Error("price catalog is not an object");
+    }
+    const models = {};
+    for (const [key, value] of Object.entries(raw)) {
+        // litellm carries a `sample_spec` pseudo-entry documenting its own schema.
+        if (key === "sample_spec" || value === null || typeof value !== "object")
+            continue;
+        const entry = value;
+        if (entry["mode"] !== "chat")
+            continue;
+        const base = tierFrom(entry, "");
+        if (base === null)
+            continue;
+        const provider = typeof entry["litellm_provider"] === "string" ? entry["litellm_provider"] : "";
+        const tiers = [];
+        for (const { above, suffix } of tierSuffixes(entry)) {
+            const t = tierFrom(entry, suffix);
+            if (t !== null)
+                tiers.push({ aboveTokens: above, rates: t.rates });
+        }
+        models[key] = {
+            label: key,
+            provider,
+            base: base.rates,
+            tiers,
+            priority: tierFrom(entry, "_priority")?.rates ?? null,
+            webSearchPerThousand: webSearchPerThousand(entry),
+            cacheRatesPublished: base.cachePublished,
+        };
+    }
+    if (Object.keys(models).length === 0) {
+        throw new Error("price catalog contained no priceable chat models");
+    }
+    return { source: opts.source, fetchedAt: opts.fetchedAt, models };
+}
+//# sourceMappingURL=catalog.js.map

package/dist/server/usage/prices/catalog.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"catalog.js","sourceRoot":"","sources":["../../../../server/usage/prices/catalog.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;GAqBG;AAyCH,MAAM,CAAC,MAAM,WAAW,GACtB,6FAA6F,CAAC;AAEhG,MAAM,WAAW,GAAG,SAAS,CAAC;AAE9B,SAAS,OAAO,CAAC,CAAU;IACzB,OAAO,OAAO,CAAC,KAAK,QAAQ,IAAI,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC;AAC1E,CAAC;AAED;;;;;;;;;;;;GAYG;AACH,SAAS,QAAQ,CACf,KAA8B,EAC9B,MAAc;IAEd,MAAM,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,uBAAuB,MAAM,EAAE,CAAC,CAAC,CAAC;IAC9D,MAAM,MAAM,GAAG,OAAO,CAAC,KAAK,CAAC,wBAAwB,MAAM,EAAE,CAAC,CAAC,CAAC;IAChE,IAAI,KAAK,KAAK,IAAI,IAAI,MAAM,KAAK,IAAI;QAAE,OAAO,IAAI,CAAC;IAEnD,MAAM,SAAS,GAAG,OAAO,CAAC,KAAK,CAAC,8BAA8B,MAAM,EAAE,CAAC,CAAC,CAAC;IACzE,MAAM,OAAO,GAAG,OAAO,CAAC,KAAK,CAAC,kCAAkC,MAAM,EAAE,CAAC,CAAC,CAAC;IAC3E,4EAA4E;IAC5E,MAAM,OAAO,GACX,OAAO,CAAC,KAAK,CAAC,4CAA4C,MAAM,EAAE,CAAC,CAAC;QACpE,OAAO,CAAC,KAAK,CAAC,2CAA2C,CAAC,CAAC,CAAC;IAE9D,OAAO;QACL,cAAc,EAAE,SAAS,KAAK,IAAI,IAAI,OAAO,KAAK,IAAI;QACtD,KAAK,EAAE;YACL,eAAe,EAAE,KAAK,GAAG,WAAW;YACpC,gBAAgB,EAAE,MAAM,GAAG,WAAW;YACtC,mBAAmB,EAAE,CAAC,SAAS,IAAI,KAAK,CAAC,GAAG,WAAW;YACvD,sBAAsB,EAAE,CAAC,OAAO,IAAI,KAAK,CAAC,GAAG,WAAW;YACxD,sBAAsB,EAAE,CAAC,OAAO,IAAI,OAAO,IAAI,KAAK,CAAC,GAAG,WAAW;SACpE;KACF,CAAC;AACJ,CAAC;AAED,uFAAuF;AACvF,MAAM,WAAW,GAAG,4CAA4C,CAAC;AAEjE,SAAS,YAAY,CAAC,KAA8B;IAClD,MAAM,KAAK,GAA6C,EAAE,CAAC;IAC3D,KAAK,MAAM,GAAG,IAAI,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC;QACrC,MAAM,CAAC,GAAG,WAAW,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;QAChC,IAAI,CAAC,EAAE,CAAC,CAAC,CAAC,KAAK,SAAS;YAAE,SAAS;QACnC,KAAK,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,IAAI,EAAE,MAAM,EAAE,UAAU,CAAC,CAAC,CAAC,CAAC,UAAU,EAAE,CAAC,CAAC;IAC/E,CAAC;IACD,OAAO,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC;AACjD,CAAC;AAED;;;;;GAKG;AACH,SAAS,oBAAoB,CAAC,KAA8B;IAC1D,MAAM,GAAG,GAAG,KAAK,CAAC,+BAA+B,CAAC,CAAC;IACnD,IAAI,GAAG,KAAK,IAAI,IAAI,OAAO,GAAG,KAAK,QAAQ;QAAE,OAAO,IAAI,CAAC;IACzD,MAAM,KAAK,GAAG,GAA8B,CAAC;IAC7C,MAAM,QAAQ,GACZ,OAAO,CAAC,KAAK,CAAC,4BAA4B,CAAC,CAAC;QAC5C,OAAO,CAAC,KAAK,CAAC,yBAAyB,CAAC,CAAC;QACzC,OAAO,CAAC,KAAK,CAAC,0BAA0B,CAAC,CAAC,CAAC;IAC7C,OAAO,QAAQ,KAAK,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,QAAQ,GAAG,IAAI,CAAC;AACpD,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,gBAAgB,CAAC,GAAY,EAAE,IAA2C;IACxF,IAAI,GAAG,KAAK,IAAI,IAAI,OAAO,GAAG,KAAK,QAAQ,EAAE,CAAC;QAC5C,MAAM,IAAI,KAAK,CAAC,gCAAgC,CAAC,CAAC;IACpD,CAAC;IAED,MAAM,MAAM,GAA8B,EAAE,CAAC;IAE7C,KAAK,MAAM,CAAC,GAAG,EAAE,KAAK,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,GAA8B,CAAC,EAAE,CAAC;QAC1E,2EAA2E;QAC3E,IAAI,GAAG,KAAK,aAAa,IAAI,KAAK,KAAK,IAAI,IAAI,OAAO,KAAK,KAAK,QAAQ;YAAE,SAAS;QACnF,MAAM,KAAK,GAAG,KAAgC,CAAC;QAC/C,IAAI,KAAK,CAAC,MAAM,CAAC,KAAK,MAAM;YAAE,SAAS;QAEvC,MAAM,IAAI,GAAG,QAAQ,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC;QACjC,IAAI,IAAI,KAAK,IAAI;YAAE,SAAS;QAE5B,MAAM,QAAQ,GAAG,OAAO,KAAK,CAAC,kBAAkB,CAAC,KAAK,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,kBAAkB,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;QAEhG,MAAM,KAAK,GAAqD,EAAE,CAAC;QACnE,KAAK,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,IAAI,YAAY,CAAC,KAAK,CAAC,EAAE,CAAC;YACpD,MAAM,CAAC,GAAG,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC;YAClC,IAAI,CAAC,KAAK,IAAI;gBAAE,KAAK,CAAC,IAAI,CAAC,EAAE,WAAW,EAAE,KAAK,EAAE,KAAK,EAAE,CAAC,CAAC,KAAK,EAAE,CAAC,CAAC;QACrE,CAAC;QAED,MAAM,CAAC,GAAG,CAAC,GAAG;YACZ,KAAK,EAAE,GAAG;YACV,QAAQ;YACR,IAAI,EAAE,IAAI,CAAC,KAAK;YAChB,KAAK;YACL,QAAQ,EAAE,QAAQ,CAAC,KAAK,EAAE,WAAW,CAAC,EAAE,KAAK,IAAI,IAAI;YACrD,oBAAoB,EAAE,oBAAoB,CAAC,KAAK,CAAC;YACjD,mBAAmB,EAAE,IAAI,CAAC,cAAc;SACzC,CAAC;IACJ,CAAC;IAED,IAAI,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACrC,MAAM,IAAI,KAAK,CAAC,kDAAkD,CAAC,CAAC;IACtE,CAAC;IAED,OAAO,EAAE,MAAM,EAAE,IAAI,CAAC,MAAM,EAAE,SAAS,EAAE,IAAI,CAAC,SAAS,EAAE,MAAM,EAAE,CAAC;AACpE,CAAC"}

package/dist/server/usage/prices/refresh.js

@@ -0,0 +1,125 @@
+/**
+ * Keeping prices current without ever making the gateway depend on the network.
+ *
+ * The contract, in order of importance:
+ *
+ *  1. Boot always succeeds and always has prices. The vendored snapshot is
+ *     enough; a clone with no network, a CI run and an airgapped container all
+ *     price correctly.
+ *  2. A refresh that fails changes NOTHING. No half-written cache, no cleared
+ *     table, no thrown error — a warn line and the previous rates.
+ *  3. No request ever waits on a fetch. The refresh runs on an unref'd timer
+ *     off the hot path, and swaps a whole catalog in one assignment.
+ *
+ * Prices drift slowly (a vendor repricing is a quarterly event, not a daily
+ * one), so a 24h cadence is generous. The point of refreshing at all is that a
+ * *new model* appears the week it ships, and an unpriced model is a row the
+ * dashboard has to report as "n/a".
+ */
+import { readFileSync, writeFileSync, renameSync, mkdirSync } from "node:fs";
+import { dirname } from "node:path";
+import { log } from "../../log.js";
+import { LITELLM_URL, normaliseCatalog } from "./catalog.js";
+import { loadSnapshot, setPriceCatalog } from "./table.js";
+const DAY_MS = 24 * 60 * 60 * 1000;
+const FETCH_TIMEOUT_MS = 10_000;
+export function pricingOptionsFromEnv(dbPath) {
+    const raw = (process.env.FEST_PRICING_REFRESH ?? "").trim().toLowerCase();
+    return {
+        // Next to the database, so the one writable volume an operator already
+        // mounts carries the cache too.
+        cachePath: `${dirname(dbPath)}/prices.json`,
+        refreshEnabled: !["0", "false", "no", "off"].includes(raw),
+        url: (process.env.FEST_PRICING_URL ?? "").trim() || LITELLM_URL,
+    };
+}
+/**
+ * Load the best catalog available, synchronously, at boot.
+ *
+ * The cache wins only if it is genuinely NEWER than the snapshot. That
+ * comparison is on `fetchedAt` rather than file mtime because the failure it
+ * prevents is real: upgrading Fest ships a newer snapshot, and a stale
+ * `data/prices.json` left in the volume would otherwise shadow it forever.
+ */
+export function loadPrices(opts) {
+    const snapshot = loadSnapshot();
+    let cached = null;
+    try {
+        cached = JSON.parse(readFileSync(opts.cachePath, "utf8"));
+    }
+    catch {
+        // No cache, or one this build cannot read. The snapshot is always enough.
+    }
+    const chosen = cached !== null && typeof cached.fetchedAt === "number" && cached.fetchedAt > snapshot.fetchedAt
+        ? cached
+        : snapshot;
+    setPriceCatalog(chosen);
+    log.info("prices loaded", {
+        models: Object.keys(chosen.models).length,
+        fetchedAt: new Date(chosen.fetchedAt).toISOString(),
+        from: chosen === snapshot ? "snapshot" : "cache",
+    });
+    return chosen;
+}
+async function fetchCatalog(url) {
+    const res = await fetch(url, {
+        signal: AbortSignal.timeout(FETCH_TIMEOUT_MS),
+        headers: { accept: "application/json" },
+    });
+    if (!res.ok)
+        throw new Error(`HTTP ${res.status}`);
+    return normaliseCatalog(await res.json(), { source: url, fetchedAt: Date.now() });
+}
+/**
+ * One refresh attempt. Never throws, never partially applies.
+ *
+ * The catalog is validated by `normaliseCatalog` (which throws on an empty or
+ * malformed file) BEFORE anything is written or swapped, so a truncated
+ * response cannot leave Fest pricing from a table with three models in it.
+ */
+export async function refreshPricesOnce(opts) {
+    try {
+        const catalog = await fetchCatalog(opts.url);
+        // Atomic: a crash mid-write leaves the old cache, never a torn one.
+        mkdirSync(dirname(opts.cachePath), { recursive: true });
+        const tmp = `${opts.cachePath}.${process.pid}.tmp`;
+        writeFileSync(tmp, JSON.stringify(catalog));
+        renameSync(tmp, opts.cachePath);
+        setPriceCatalog(catalog);
+        log.info("prices refreshed", { models: Object.keys(catalog.models).length, url: opts.url });
+        return true;
+    }
+    catch (err) {
+        log.warn("price refresh failed, keeping current rates", {
+            url: opts.url,
+            error: err instanceof Error ? err.message : String(err),
+        });
+        return false;
+    }
+}
+/**
+ * Start the background refresh loop. Returns a handle whose `stop` is safe to
+ * call more than once, so shutdown ordering never has to be careful.
+ *
+ * Both timers are `unref`'d: prices must never be the reason the process won't
+ * exit, and a short-lived CLI command that happens to load the table should not
+ * hang for 24 hours waiting for a tick.
+ */
+export function startPriceRefresh(opts) {
+    if (!opts.refreshEnabled) {
+        log.info("price refresh disabled", { reason: "FEST_PRICING_REFRESH" });
+        return { stop: () => { } };
+    }
+    const timers = [];
+    // Deferred rather than awaited: boot must not wait on GitHub.
+    timers.push(setTimeout(() => void refreshPricesOnce(opts), 0).unref());
+    timers.push(setInterval(() => void refreshPricesOnce(opts), DAY_MS).unref());
+    return {
+        stop: () => {
+            for (const t of timers)
+                clearTimeout(t);
+            timers.length = 0;
+        },
+    };
+}
+//# sourceMappingURL=refresh.js.map

package/dist/server/usage/prices/refresh.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"refresh.js","sourceRoot":"","sources":["../../../../server/usage/prices/refresh.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAEH,OAAO,EAAE,YAAY,EAAE,aAAa,EAAE,UAAU,EAAE,SAAS,EAAE,MAAM,SAAS,CAAC;AAC7E,OAAO,EAAE,OAAO,EAAE,MAAM,WAAW,CAAC;AACpC,OAAO,EAAE,GAAG,EAAE,MAAM,cAAc,CAAC;AACnC,OAAO,EAAE,WAAW,EAAE,gBAAgB,EAAE,MAAM,cAAc,CAAC;AAE7D,OAAO,EAAE,YAAY,EAAE,eAAe,EAAE,MAAM,YAAY,CAAC;AAE3D,MAAM,MAAM,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,IAAI,CAAC;AACnC,MAAM,gBAAgB,GAAG,MAAM,CAAC;AAShC,MAAM,UAAU,qBAAqB,CAAC,MAAc;IAClD,MAAM,GAAG,GAAG,CAAC,OAAO,CAAC,GAAG,CAAC,oBAAoB,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;IAC1E,OAAO;QACL,uEAAuE;QACvE,gCAAgC;QAChC,SAAS,EAAE,GAAG,OAAO,CAAC,MAAM,CAAC,cAAc;QAC3C,cAAc,EAAE,CAAC,CAAC,GAAG,EAAE,OAAO,EAAE,IAAI,EAAE,KAAK,CAAC,CAAC,QAAQ,CAAC,GAAG,CAAC;QAC1D,GAAG,EAAE,CAAC,OAAO,CAAC,GAAG,CAAC,gBAAgB,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,IAAI,WAAW;KAChE,CAAC;AACJ,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,UAAU,UAAU,CAAC,IAAuC;IAChE,MAAM,QAAQ,GAAG,YAAY,EAAE,CAAC;IAEhC,IAAI,MAAM,GAAwB,IAAI,CAAC;IACvC,IAAI,CAAC;QACH,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,YAAY,CAAC,IAAI,CAAC,SAAS,EAAE,MAAM,CAAC,CAAiB,CAAC;IAC5E,CAAC;IAAC,MAAM,CAAC;QACP,0EAA0E;IAC5E,CAAC;IAED,MAAM,MAAM,GACV,MAAM,KAAK,IAAI,IAAI,OAAO,MAAM,CAAC,SAAS,KAAK,QAAQ,IAAI,MAAM,CAAC,SAAS,GAAG,QAAQ,CAAC,SAAS;QAC9F,CAAC,CAAC,MAAM;QACR,CAAC,CAAC,QAAQ,CAAC;IAEf,eAAe,CAAC,MAAM,CAAC,CAAC;IACxB,GAAG,CAAC,IAAI,CAAC,eAAe,EAAE;QACxB,MAAM,EAAE,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,MAAM;QACzC,SAAS,EAAE,IAAI,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,CAAC,WAAW,EAAE;QACnD,IAAI,EAAE,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,OAAO;KACjD,CAAC,CAAC;IACH,OAAO,MAAM,CAAC;AAChB,CAAC;AAED,KAAK,UAAU,YAAY,CAAC,GAAW;IACrC,MAAM,GAAG,GAAG,MAAM,KAAK,CAAC,GAAG,EAAE;QAC3B,MAAM,EAAE,WAAW,CAAC,OAAO,CAAC,gBAAgB,CAAC;QAC7C,OAAO,EAAE,EAAE,MAAM,EAAE,kBAAkB,EAAE;KACxC,CAAC,CAAC;IACH,IAAI,CAAC,GAAG,CAAC,EAAE;QAAE,MAAM,IAAI,KAAK,CAAC,QAAQ,GAAG,CAAC,MAAM,EAAE,CAAC,CAAC;IACnD,OAAO,gBAAgB,CAAC,MAAM,GAAG,CAAC,IAAI,EAAE,EAAE,EAAE,MAAM,EAAE,GAAG,EAAE,SAAS,EAAE,IAAI,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC;AACpF,CAAC;AAED;;;;;;GAMG;AACH,MAAM,CAAC,KAAK,UAAU,iBAAiB,CAAC,IAAoB;IAC1D,IAAI,CAAC;QACH,MAAM,OAAO,GAAG,MAAM,YAAY,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;QAE7C,oEAAoE;QACpE,SAAS,CAAC,OAAO,CAAC,IAAI,CAAC,SAAS,CAAC,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;QACxD,MAAM,GAAG,GAAG,GAAG,IAAI,CAAC,SAAS,IAAI,OAAO,CAAC,GAAG,MAAM,CAAC;QACnD,aAAa,CAAC,GAAG,EAAE,IAAI,CAAC,SAAS,CAAC,OAAO,CAAC,CAAC,CAAC;QAC5C,UAAU,CAAC,GAAG,EAAE,IAAI,CAAC,SAAS,CAAC,CAAC;QAEhC,eAAe,CAAC,OAAO,CAAC,CAAC;QACzB,GAAG,CAAC,IAAI,CAAC,kBAAkB,EAAE,EAAE,MAAM,EAAE,MAAM,CAAC,IAAI,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,MAAM,EAAE,GAAG,EAAE,IAAI,CAAC,GAAG,EAAE,CAAC,CAAC;QAC5F,OAAO,IAAI,CAAC;IACd,CAAC;IAAC,OAAO,GAAG,EAAE,CAAC;QACb,GAAG,CAAC,IAAI,CAAC,6CAA6C,EAAE;YACtD,GAAG,EAAE,IAAI,CAAC,GAAG;YACb,KAAK,EAAE,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC;SACxD,CAAC,CAAC;QACH,OAAO,KAAK,CAAC;IACf,CAAC;AACH,CAAC;AAMD;;;;;;;GAOG;AACH,MAAM,UAAU,iBAAiB,CAAC,IAAoB;IACpD,IAAI,CAAC,IAAI,CAAC,cAAc,EAAE,CAAC;QACzB,GAAG,CAAC,IAAI,CAAC,wBAAwB,EAAE,EAAE,MAAM,EAAE,sBAAsB,EAAE,CAAC,CAAC;QACvE,OAAO,EAAE,IAAI,EAAE,GAAG,EAAE,GAAE,CAAC,EAAE,CAAC;IAC5B,CAAC;IAED,MAAM,MAAM,GAAqB,EAAE,CAAC;IAEpC,8DAA8D;IAC9D,MAAM,CAAC,IAAI,CAAC,UAAU,CAAC,GAAG,EAAE,CAAC,KAAK,iBAAiB,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC,CAAC,KAAK,EAAE,CAAC,CAAC;IACvE,MAAM,CAAC,IAAI,CAAC,WAAW,CAAC,GAAG,EAAE,CAAC,KAAK,iBAAiB,CAAC,IAAI,CAAC,EAAE,MAAM,CAAC,CAAC,KAAK,EAAE,CAAC,CAAC;IAE7E,OAAO;QACL,IAAI,EAAE,GAAG,EAAE;YACT,KAAK,MAAM,CAAC,IAAI,MAAM;gBAAE,YAAY,CAAC,CAAC,CAAC,CAAC;YACxC,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC;QACpB,CAAC;KACF,CAAC;AACJ,CAAC"}

package/dist/server/usage/prices/table.js

@@ -0,0 +1,102 @@
+/**
+ * The in-memory price table: load once at boot, look up per request.
+ *
+ * Two hard rules, both inherited from the hand-maintained table this replaced:
+ *
+ *  1. **An unknown model yields `null`, never an estimate.** A guessed rate
+ *     that looks plausible is worse than an honest "n/a", because nobody
+ *     investigates a number that looks fine.
+ *  2. **A gateway must not depend on an external price service to serve a
+ *     request.** The vendored snapshot is loaded synchronously at boot and is
+ *     always sufficient. `refresh.ts` only ever swaps in something newer.
+ *
+ * Note the deliberate absence of prefix matching. The old table had 13 entries
+ * and matched `claude-opus-4-1-20260101` by longest prefix out of necessity.
+ * litellm carries every dated snapshot id explicitly alongside its alias, so
+ * prefix matching would now be guessing where an exact answer exists — and it
+ * guesses badly: `claude-opus-4-5-20251101` would match a bare `claude-opus`
+ * entry at a completely different rate.
+ */
+import { gunzipSync } from "node:zlib";
+import { readFileSync } from "node:fs";
+import { fileURLToPath } from "node:url";
+const SNAPSHOT_URL = new URL("./snapshot.json.gz", import.meta.url);
+/** The vendored snapshot. Always present, always loadable, never the network. */
+export function loadSnapshot() {
+    const gz = readFileSync(fileURLToPath(SNAPSHOT_URL));
+    return JSON.parse(gunzipSync(gz).toString("utf8"));
+}
+let current = null;
+export function priceCatalog() {
+    current ??= loadSnapshot();
+    return current;
+}
+/**
+ * Swap in a newer catalog. Called only by `refresh.ts` after a successful
+ * fetch; a failed refresh leaves the previous table exactly as it was.
+ */
+export function setPriceCatalog(catalog) {
+    current = catalog;
+}
+/** Test seam: forget the loaded table so the next lookup re-reads the snapshot. */
+export function resetPriceCatalog() {
+    current = null;
+}
+/** Strip Claude Code's client-side context-window tag before matching. */
+function normalise(model) {
+    return model.trim().toLowerCase().replace(/\[1m\]$/, "");
+}
+/**
+ * Candidate catalog keys for a model id, most specific first.
+ *
+ * The provider-qualified forms matter for the substitute path: a route sends
+ * `accounts/fireworks/models/deepseek-v4-pro`, which litellm files under
+ * `fireworks_ai/accounts/fireworks/models/deepseek-v4-pro`. It also files many
+ * Fireworks models under the bare leaf (`fireworks_ai/deepseek-v4-pro-0813`),
+ * hence the basename form.
+ */
+function candidates(id, provider) {
+    const out = [id];
+    if (provider !== undefined && provider !== "") {
+        out.push(`${provider}/${id}`);
+        const leaf = id.slice(id.lastIndexOf("/") + 1);
+        if (leaf !== id)
+            out.push(`${provider}/${leaf}`);
+    }
+    return out;
+}
+export function lookupRate(model, provider) {
+    if (model === null || model === undefined || model.trim() === "")
+        return null;
+    const { models } = priceCatalog();
+    // Raw first, then normalised: an id that matches verbatim should never be
+    // reinterpreted by the `[1m]` strip.
+    for (const id of [model.trim(), normalise(model)]) {
+        for (const key of candidates(id, provider)) {
+            const hit = models[key];
+            if (hit !== undefined)
+                return hit;
+        }
+    }
+    return null;
+}
+/**
+ * Pick the tier a call is billed at.
+ *
+ * Long-context tiers are thresholds on the WHOLE request: once context crosses
+ * the line the entire call prices at the higher rate, which is how Anthropic
+ * and Google both bill it. Priority overrides only the base tier — no vendor
+ * currently publishes a priority rate for a long-context tier, so combining
+ * them would be inventing a price.
+ */
+export function tierFor(rate, contextTokens, serviceTier) {
+    for (let i = rate.tiers.length - 1; i >= 0; i -= 1) {
+        const tier = rate.tiers[i];
+        if (tier !== undefined && contextTokens > tier.aboveTokens)
+            return tier.rates;
+    }
+    if (serviceTier === "priority" && rate.priority !== null)
+        return rate.priority;
+    return rate.base;
+}
+//# sourceMappingURL=table.js.map

package/dist/server/usage/prices/table.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"table.js","sourceRoot":"","sources":["../../../../server/usage/prices/table.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;GAkBG;AAEH,OAAO,EAAE,UAAU,EAAE,MAAM,WAAW,CAAC;AACvC,OAAO,EAAE,YAAY,EAAE,MAAM,SAAS,CAAC;AACvC,OAAO,EAAE,aAAa,EAAE,MAAM,UAAU,CAAC;AAKzC,MAAM,YAAY,GAAG,IAAI,GAAG,CAAC,oBAAoB,EAAE,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;AAEpE,iFAAiF;AACjF,MAAM,UAAU,YAAY;IAC1B,MAAM,EAAE,GAAG,YAAY,CAAC,aAAa,CAAC,YAAY,CAAC,CAAC,CAAC;IACrD,OAAO,IAAI,CAAC,KAAK,CAAC,UAAU,CAAC,EAAE,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAiB,CAAC;AACrE,CAAC;AAED,IAAI,OAAO,GAAwB,IAAI,CAAC;AAExC,MAAM,UAAU,YAAY;IAC1B,OAAO,KAAK,YAAY,EAAE,CAAC;IAC3B,OAAO,OAAO,CAAC;AACjB,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,eAAe,CAAC,OAAqB;IACnD,OAAO,GAAG,OAAO,CAAC;AACpB,CAAC;AAED,mFAAmF;AACnF,MAAM,UAAU,iBAAiB;IAC/B,OAAO,GAAG,IAAI,CAAC;AACjB,CAAC;AAED,0EAA0E;AAC1E,SAAS,SAAS,CAAC,KAAa;IAC9B,OAAO,KAAK,CAAC,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC,OAAO,CAAC,SAAS,EAAE,EAAE,CAAC,CAAC;AAC3D,CAAC;AAED;;;;;;;;GAQG;AACH,SAAS,UAAU,CAAC,EAAU,EAAE,QAA4B;IAC1D,MAAM,GAAG,GAAG,CAAC,EAAE,CAAC,CAAC;IACjB,IAAI,QAAQ,KAAK,SAAS,IAAI,QAAQ,KAAK,EAAE,EAAE,CAAC;QAC9C,GAAG,CAAC,IAAI,CAAC,GAAG,QAAQ,IAAI,EAAE,EAAE,CAAC,CAAC;QAC9B,MAAM,IAAI,GAAG,EAAE,CAAC,KAAK,CAAC,EAAE,CAAC,WAAW,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC;QAC/C,IAAI,IAAI,KAAK,EAAE;YAAE,GAAG,CAAC,IAAI,CAAC,GAAG,QAAQ,IAAI,IAAI,EAAE,CAAC,CAAC;IACnD,CAAC;IACD,OAAO,GAAG,CAAC;AACb,CAAC;AAED,MAAM,UAAU,UAAU,CACxB,KAAgC,EAChC,QAA6B;IAE7B,IAAI,KAAK,KAAK,IAAI,IAAI,KAAK,KAAK,SAAS,IAAI,KAAK,CAAC,IAAI,EAAE,KAAK,EAAE;QAAE,OAAO,IAAI,CAAC;IAC9E,MAAM,EAAE,MAAM,EAAE,GAAG,YAAY,EAAE,CAAC;IAElC,0EAA0E;IAC1E,qCAAqC;IACrC,KAAK,MAAM,EAAE,IAAI,CAAC,KAAK,CAAC,IAAI,EAAE,EAAE,SAAS,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC;QAClD,KAAK,MAAM,GAAG,IAAI,UAAU,CAAC,EAAE,EAAE,QAAQ,CAAC,EAAE,CAAC;YAC3C,MAAM,GAAG,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC;YACxB,IAAI,GAAG,KAAK,SAAS;gBAAE,OAAO,GAAG,CAAC;QACpC,CAAC;IACH,CAAC;IACD,OAAO,IAAI,CAAC;AACd,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,OAAO,CACrB,IAAe,EACf,aAAqB,EACrB,WAA+B;IAE/B,KAAK,IAAI,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC;QACnD,MAAM,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;QAC3B,IAAI,IAAI,KAAK,SAAS,IAAI,aAAa,GAAG,IAAI,CAAC,WAAW;YAAE,OAAO,IAAI,CAAC,KAAK,CAAC;IAChF,CAAC;IACD,IAAI,WAAW,KAAK,UAAU,IAAI,IAAI,CAAC,QAAQ,KAAK,IAAI;QAAE,OAAO,IAAI,CAAC,QAAQ,CAAC;IAC/E,OAAO,IAAI,CAAC,IAAI,CAAC;AACnB,CAAC"}

package/dist/server/usage/pricing.js

@@ -0,0 +1,103 @@
+/**
+ * Per-model rate lookup and call pricing.
+ *
+ * Two rules carried over from the cost engine in fw-ai/fireconnect
+ * (packages/setup-cli/lib/harnesses/claude/usage/pricing.mjs), both of which are
+ * easy to get wrong and expensive to get wrong:
+ *
+ *  1. The token buckets are DISJOINT and priced at different rates. Cache reads
+ *     are ~10% of the input rate; a 5-minute cache write is ~125% and a 1-hour
+ *     write ~200%. Folding cache reads into input inflates a bill by an order of
+ *     magnitude on a cache-heavy agent workload, which Claude Code is.
+ *  2. An unknown model yields `null`, never an estimate. A guessed rate that
+ *     looks plausible is worse than an honest "n/a", because nobody
+ *     investigates a number that looks fine.
+ *
+ * Rates come from the vendored litellm catalog in `prices/` — see that
+ * directory for why they are a snapshot with a background refresh rather than a
+ * live vendor call. This file replaced a hand-maintained 13-entry Claude table
+ * that had drifted 3x on Opus 5; the drift is the argument for not hand-
+ * maintaining rates.
+ *
+ * ── Two dollar figures, and why they are not interchangeable ─────────────────
+ *
+ * `cost` is ORG SPEND: what the organisation will be invoiced. It is null for a
+ * subscription request, because the developer's own plan absorbed it and there
+ * is no invoice.
+ *
+ * `notionalCost` is VALUE: what the same call would have cost at published API
+ * rates, computed on every path including subscription. It exists because the
+ * largest slice of Fest's traffic is otherwise invisible in money terms — a
+ * team lead cannot see what their Max seats are actually delivering.
+ *
+ * They are separate fields, separate columns and separate UI stats precisely so
+ * that "add them up" is never the path of least resistance. A notional figure
+ * summed into spend is a fabricated invoice.
+ */
+import { contextTokens } from "../../shared/types.js";
+import { lookupRate, tierFor } from "./prices/table.js";
+export { lookupRate } from "./prices/table.js";
+/**
+ * Price one call at list rates.
+ *
+ * The tier is chosen from the call's own context size and service tier: a
+ * long-context request bills entirely at the higher tier, which is how both
+ * Anthropic and Google publish it, and is where Claude Code's 1M-context
+ * sessions were previously understated by half.
+ */
+function listPrice(rate, usage) {
+    const t = tierFor(rate, contextTokens(usage), usage.serviceTier);
+    const perMillion = usage.inputTokens * t.inputPerMillion +
+        usage.cacheReadTokens * t.cacheReadPerMillion +
+        usage.cacheWrite5mTokens * t.cacheWrite5mPerMillion +
+        usage.cacheWrite1hTokens * t.cacheWrite1hPerMillion +
+        usage.outputTokens * t.outputPerMillion;
+    // A model with no published search rate contributes nothing for searches
+    // rather than borrowing another vendor's price.
+    const search = rate.webSearchPerThousand === null
+        ? 0
+        : (usage.webSearches * rate.webSearchPerThousand) / 1_000;
+    return perMillion / 1_000_000 + search;
+}
+/**
+ * Price a call.
+ *
+ * `postureIsSubscription` still yields `basis: "subscription"` with a null
+ * `cost` — that invariant is unchanged and load-bearing. What is new is that
+ * `notionalCost` is populated anyway, so the value of subscription work is
+ * recorded without ever entering a spend total.
+ *
+ * `provider` is litellm's provider key, needed on the substitute path where the
+ * served model is a provider-native id. See `ADAPTER_PRICE_PROVIDERS`.
+ */
+export function priceUsage(model, usage, postureIsSubscription, provider) {
+    const rate = lookupRate(model, provider);
+    if (rate === null) {
+        // Unpriced, not free — and unpriced on both figures. We do not know what
+        // this model charges, so we do not know what it would have been worth.
+        return {
+            cost: null,
+            basis: postureIsSubscription ? "subscription" : "none",
+            notionalCost: null,
+            rateLabel: null,
+        };
+    }
+    const value = listPrice(rate, usage);
+    if (postureIsSubscription) {
+        return { cost: null, basis: "subscription", notionalCost: value, rateLabel: rate.label };
+    }
+    return { cost: value, basis: "list", notionalCost: value, rateLabel: rate.label };
+}
+/**
+ * Fraction of context that came from cache. The single most actionable number
+ * on a gateway dashboard: a collapsed cache hit rate is where both spend and
+ * latency go, and it usually means something upstream is mangling
+ * `cache_control`.
+ */
+export function cacheHitRatio(usage) {
+    const context = contextTokens(usage);
+    if (context <= 0)
+        return null;
+    return usage.cacheReadTokens / context;
+}
+//# sourceMappingURL=pricing.js.map

package/dist/server/usage/pricing.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"pricing.js","sourceRoot":"","sources":["../../../server/usage/pricing.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAmCG;AAEH,OAAO,EAAE,aAAa,EAAE,MAAM,uBAAuB,CAAC;AAGtD,OAAO,EAAE,UAAU,EAAE,OAAO,EAAE,MAAM,mBAAmB,CAAC;AAGxD,OAAO,EAAE,UAAU,EAAE,MAAM,mBAAmB,CAAC;AAe/C;;;;;;;GAOG;AACH,SAAS,SAAS,CAAC,IAAe,EAAE,KAAmB;IACrD,MAAM,CAAC,GAAG,OAAO,CAAC,IAAI,EAAE,aAAa,CAAC,KAAK,CAAC,EAAE,KAAK,CAAC,WAAW,CAAC,CAAC;IAEjE,MAAM,UAAU,GACd,KAAK,CAAC,WAAW,GAAG,CAAC,CAAC,eAAe;QACrC,KAAK,CAAC,eAAe,GAAG,CAAC,CAAC,mBAAmB;QAC7C,KAAK,CAAC,kBAAkB,GAAG,CAAC,CAAC,sBAAsB;QACnD,KAAK,CAAC,kBAAkB,GAAG,CAAC,CAAC,sBAAsB;QACnD,KAAK,CAAC,YAAY,GAAG,CAAC,CAAC,gBAAgB,CAAC;IAE1C,yEAAyE;IACzE,gDAAgD;IAChD,MAAM,MAAM,GAAG,IAAI,CAAC,oBAAoB,KAAK,IAAI;QAC/C,CAAC,CAAC,CAAC;QACH,CAAC,CAAC,CAAC,KAAK,CAAC,WAAW,GAAG,IAAI,CAAC,oBAAoB,CAAC,GAAG,KAAK,CAAC;IAE5D,OAAO,UAAU,GAAG,SAAS,GAAG,MAAM,CAAC;AACzC,CAAC;AAED;;;;;;;;;;GAUG;AACH,MAAM,UAAU,UAAU,CACxB,KAAoB,EACpB,KAAmB,EACnB,qBAA8B,EAC9B,QAA6B;IAE7B,MAAM,IAAI,GAAG,UAAU,CAAC,KAAK,EAAE,QAAQ,CAAC,CAAC;IACzC,IAAI,IAAI,KAAK,IAAI,EAAE,CAAC;QAClB,yEAAyE;QACzE,uEAAuE;QACvE,OAAO;YACL,IAAI,EAAE,IAAI;YACV,KAAK,EAAE,qBAAqB,CAAC,CAAC,CAAC,cAAc,CAAC,CAAC,CAAC,MAAM;YACtD,YAAY,EAAE,IAAI;YAClB,SAAS,EAAE,IAAI;SAChB,CAAC;IACJ,CAAC;IAED,MAAM,KAAK,GAAG,SAAS,CAAC,IAAI,EAAE,KAAK,CAAC,CAAC;IAErC,IAAI,qBAAqB,EAAE,CAAC;QAC1B,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE,cAAc,EAAE,YAAY,EAAE,KAAK,EAAE,SAAS,EAAE,IAAI,CAAC,KAAK,EAAE,CAAC;IAC3F,CAAC;IACD,OAAO,EAAE,IAAI,EAAE,KAAK,EAAE,KAAK,EAAE,MAAM,EAAE,YAAY,EAAE,KAAK,EAAE,SAAS,EAAE,IAAI,CAAC,KAAK,EAAE,CAAC;AACpF,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,aAAa,CAAC,KAAmB;IAC/C,MAAM,OAAO,GAAG,aAAa,CAAC,KAAK,CAAC,CAAC;IACrC,IAAI,OAAO,IAAI,CAAC;QAAE,OAAO,IAAI,CAAC;IAC9B,OAAO,KAAK,CAAC,eAAe,GAAG,OAAO,CAAC;AACzC,CAAC"}

package/dist/shared/api.js

@@ -0,0 +1,19 @@
+/**
+ * The dashboard wire contract.
+ *
+ * Declared here rather than imported from `server/store/queries.ts` because the
+ * browser bundle must not reach into server modules — that file's import graph
+ * reaches `node:sqlite`. Declaring the shapes independently would normally
+ * invite drift, so `server/api/routes.ts` asserts every response with
+ * `satisfies`: if a query's return type stops matching what the dashboard is
+ * promised, `tsc --noEmit` fails at the route rather than the browser failing at
+ * runtime. Fields added server-side are invisible to the UI, which is harmless;
+ * fields removed or renamed break the build, which is the direction that
+ * matters.
+ *
+ * Accounting rules the UI must honour are documented on the fields themselves,
+ * because a field named `pricedCostUsd` is otherwise indistinguishable from a
+ * total.
+ */
+export {};
+//# sourceMappingURL=api.js.map

package/dist/shared/api.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"api.js","sourceRoot":"","sources":["../../shared/api.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG"}

package/dist/shared/base-models.js

@@ -0,0 +1,28 @@
+/**
+ * Anthropic's own current base models, shared between the gateway menu, the
+ * `fest claude` preflight and the Routing screen's classification. Keeping one
+ * list prevents the CLI from warning about a model the server no longer
+ * considers a base, and prevents the server from offering a model the CLI does
+ * not know to check.
+ *
+ * A model missing from here is not a cosmetic gap. Three things go quietly
+ * wrong: the gateway menu never offers it, `fest claude` never warns that this
+ * gateway cannot serve it, and the dashboard files a rule that redirects it
+ * under "models this gateway adds" instead of "substitutions" — which is the
+ * heading an operator reads as harmless.
+ *
+ * Checked against the lineup table at
+ * platform.claude.com/docs/en/docs/about-claude/models/overview on 2026-09-22.
+ * Ids are the Claude API ids; from the 4.6 generation on, a dateless id is
+ * itself a pinned snapshot, which is why only Haiku carries a date. Legacy
+ * models that are still servable (Fable 5, Opus 4.8 and earlier, Sonnet 4.6
+ * and earlier) are deliberately absent: this is the list Fest offers and
+ * checks, not an archive.
+ */
+export const BASE_MODELS = [
+    { id: "claude-fable-5-1", name: "Fable 5.1" },
+    { id: "claude-opus-5", name: "Opus 5" },
+    { id: "claude-sonnet-5", name: "Sonnet 5" },
+    { id: "claude-haiku-4-5-20251001", name: "Haiku 4.5" },
+];
+//# sourceMappingURL=base-models.js.map

package/dist/shared/base-models.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"base-models.js","sourceRoot":"","sources":["../../shared/base-models.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;GAoBG;AAOH,MAAM,CAAC,MAAM,WAAW,GAA6B;IACnD,EAAE,EAAE,EAAE,kBAAkB,EAAE,IAAI,EAAE,WAAW,EAAE;IAC7C,EAAE,EAAE,EAAE,eAAe,EAAE,IAAI,EAAE,QAAQ,EAAE;IACvC,EAAE,EAAE,EAAE,iBAAiB,EAAE,IAAI,EAAE,UAAU,EAAE;IAC3C,EAAE,EAAE,EAAE,2BAA2B,EAAE,IAAI,EAAE,WAAW,EAAE;CAC9C,CAAC"}

package/dist/shared/demotion-vars.js

@@ -0,0 +1,17 @@
+/**
+ * Env vars that demote Claude Code off subscription auth onto key auth.
+ *
+ * Setting any of these makes `isAnthropicAuthEnabled()` return false, silently
+ * moving a developer off their own Max/Team subscription onto whatever key the
+ * gateway holds. `tools/canary.ts` clears these before spawning a *test*
+ * client; `cli/claude.ts` clears them before spawning a *developer's* client,
+ * for the identical reason. One list, so they cannot drift apart.
+ */
+export const DEMOTION_VARS = [
+    "ANTHROPIC_API_KEY",
+    "ANTHROPIC_AUTH_TOKEN",
+    "ANTHROPIC_BASE_URL",
+    "CLAUDE_CODE_USE_BEDROCK",
+    "CLAUDE_CODE_USE_VERTEX",
+];
+//# sourceMappingURL=demotion-vars.js.map

package/dist/shared/demotion-vars.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"demotion-vars.js","sourceRoot":"","sources":["../../shared/demotion-vars.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AACH,MAAM,CAAC,MAAM,aAAa,GAAG;IAC3B,mBAAmB;IACnB,sBAAsB;IACtB,oBAAoB;IACpB,yBAAyB;IACzB,wBAAwB;CAChB,CAAC"}

package/dist/shared/series.js

@@ -0,0 +1,50 @@
+/**
+ * Zero-filling a gapped hourly series.
+ *
+ * The API returns gaps AS GAPS: an hour with no traffic is absent, not zero,
+ * because the server cannot tell "nobody worked" from "outside retention",
+ * whereas the caller knows exactly which range it asked for. Filling is
+ * therefore the caller's job, and this is that job done once.
+ *
+ * It lives in `shared/` rather than in the chart component for one reason: the
+ * bug it guards against is invisible in a component test and obvious in a unit
+ * test. The hour KEY must be the epoch-ms hour start, never a formatted clock
+ * label — a label repeats every 24 hours, so on any range wider than a day two
+ * different hours collide on one categorical band and a stacked layout rejects
+ * the data outright. Format at draw time; key on time.
+ */
+export const HOUR_MS = 3_600_000;
+/**
+ * Hours drawn at most. Beyond this the bars are sub-pixel on any real screen,
+ * so the extra columns cost layout time and communicate nothing.
+ */
+export const MAX_HOURS = 168;
+/** Epoch-ms hour starts covering `[fromMs, toMs]`, oldest first. */
+export function hourStarts(fromMs, toMs, maxHours = MAX_HOURS) {
+    if (!Number.isFinite(fromMs) || !Number.isFinite(toMs) || toMs < fromMs)
+        return [];
+    const lastHour = Math.floor(toMs / HOUR_MS) * HOUR_MS;
+    const firstHour = Math.floor(fromMs / HOUR_MS) * HOUR_MS;
+    const available = Math.floor((lastHour - firstHour) / HOUR_MS) + 1;
+    const count = Math.min(available, maxHours);
+    // When the range is longer than the cap, keep the RECENT end. A truncated
+    // chart that drops the last hour would be actively misleading on a dashboard
+    // whose whole job is "what is happening now".
+    const start = lastHour - (count - 1) * HOUR_MS;
+    const hours = [];
+    for (let t = start; t <= lastHour; t += HOUR_MS)
+        hours.push(t);
+    return hours;
+}
+/**
+ * Join a gapped series onto a dense hour axis.
+ *
+ * `fill` supplies the value for an hour with no row, so callers never have to
+ * decide whether a missing hour is zero — here it genuinely is zero traffic,
+ * which is different from an unpriced or unavailable figure elsewhere.
+ */
+export function fillSeries(rows, fromMs, toMs, project, maxHours = MAX_HOURS) {
+    const byHour = new Map(rows.map((r) => [r.hourStart, r]));
+    return hourStarts(fromMs, toMs, maxHours).map((t) => project(t, byHour.get(t)));
+}
+//# sourceMappingURL=series.js.map

package/dist/shared/series.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"series.js","sourceRoot":"","sources":["../../shared/series.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAEH,MAAM,CAAC,MAAM,OAAO,GAAG,SAAS,CAAC;AAEjC;;;GAGG;AACH,MAAM,CAAC,MAAM,SAAS,GAAG,GAAG,CAAC;AAE7B,oEAAoE;AACpE,MAAM,UAAU,UAAU,CAAC,MAAc,EAAE,IAAY,EAAE,QAAQ,GAAG,SAAS;IAC3E,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,IAAI,IAAI,GAAG,MAAM;QAAE,OAAO,EAAE,CAAC;IAEnF,MAAM,QAAQ,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,GAAG,OAAO,CAAC,GAAG,OAAO,CAAC;IACtD,MAAM,SAAS,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,GAAG,OAAO,CAAC,GAAG,OAAO,CAAC;IACzD,MAAM,SAAS,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,QAAQ,GAAG,SAAS,CAAC,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC;IACnE,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,SAAS,EAAE,QAAQ,CAAC,CAAC;IAE5C,0EAA0E;IAC1E,6EAA6E;IAC7E,8CAA8C;IAC9C,MAAM,KAAK,GAAG,QAAQ,GAAG,CAAC,KAAK,GAAG,CAAC,CAAC,GAAG,OAAO,CAAC;IAE/C,MAAM,KAAK,GAAa,EAAE,CAAC;IAC3B,KAAK,IAAI,CAAC,GAAG,KAAK,EAAE,CAAC,IAAI,QAAQ,EAAE,CAAC,IAAI,OAAO;QAAE,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;IAC/D,OAAO,KAAK,CAAC;AACf,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,UAAU,CACxB,IAAqB,EACrB,MAAc,EACd,IAAY,EACZ,OAA2D,EAC3D,QAAQ,GAAG,SAAS;IAEpB,MAAM,MAAM,GAAG,IAAI,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,SAAS,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;IAC1D,OAAO,UAAU,CAAC,MAAM,EAAE,IAAI,EAAE,QAAQ,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,CAAC,EAAE,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;AAClF,CAAC"}