fest-gateway 0.1.0

package/dist/server/pipeline/dispatch.js

@@ -0,0 +1,138 @@
+/**
+ * Choose a pipeline for one request, and hand off.
+ *
+ * The fork lives here rather than inside either pipeline so that neither can
+ * reach into the other. `passthrough.ts` has no idea routing exists;
+ * `substitute.ts` has no idea a subscription exists. That separation is what
+ * keeps "bytes are forwarded verbatim on the subscription path" true by
+ * construction rather than by discipline.
+ *
+ * Order of operations matters and is deliberate:
+ *
+ *   1. identity        — who is asking (both paths need this)
+ *   2. body            — read as opaque bytes, inspected only via a copy
+ *   3. route           — on the requested model
+ *   4. credential      — resolved for the chosen route, refusing rather than
+ *                        falling back
+ *   5. dispatch
+ *
+ * Routing happens AFTER the body is read because the model id lives in the
+ * body, not the URL. That is also why a request whose model cannot be read
+ * stays on the pass-through path: routing a request we could not identify is
+ * how traffic ends up somewhere nobody intended.
+ */
+import { randomUUID } from "node:crypto";
+import { detectInbound } from "../auth/posture.js";
+import { isSubscriptionCredential } from "../secret/fingerprint.js";
+import { readBodyBytes, isTooLarge, peekRequest } from "../http/body.js";
+import { anthropicError, statusForErrorType } from "../http/errors.js";
+import { resolveRoute, resolveWithoutCallerCredential, defaultAnthropicUpstream, } from "../routes/resolve.js";
+import { resolveCredential } from "../credentials/resolve.js";
+import { handleMessages } from "./passthrough.js";
+import { handleSubstitute, refuse } from "./substitute.js";
+import { fireworksAdapter } from "../adapters/fireworks.js";
+import { anthropicAdapter } from "../adapters/anthropic.js";
+import { log } from "../log.js";
+const MAX_BODY_BYTES = 32 * 1024 * 1024;
+const ADAPTERS = {
+    fireworks: fireworksAdapter,
+    anthropic: anthropicAdapter,
+};
+export async function dispatchMessages(req, res, ctx) {
+    // The common case by far, and the one that must stay cheapest: no routing
+    // config at all means there is nothing to decide, so do not read the body
+    // twice or construct a decision just to discard it.
+    //
+    // An `anthropic` upstream counts as routing config even with no routes
+    // pointing at it — it is the default destination for Anthropic models in the
+    // key posture, so there IS a decision to make.
+    if (ctx.routes.routes.length === 0 && defaultAnthropicUpstream(ctx.routes) === null) {
+        await handleMessages(req, res, ctx);
+        return;
+    }
+    const inbound = detectInbound(req.url ?? "/", req.headers);
+    const body = await readBodyBytes(req, MAX_BODY_BYTES);
+    if (isTooLarge(body)) {
+        const type = "invalid_request_error";
+        res.writeHead(statusForErrorType(type), { "content-type": "application/json" });
+        res.end(anthropicError(type, `Fest: request body exceeds ${MAX_BODY_BYTES} bytes.`));
+        return;
+    }
+    const peek = peekRequest(body);
+    /**
+     * With no caller credential there is nothing to pass through, so an Anthropic
+     * model no route claims falls back to the `anthropic` upstream rather than to
+     * a first-message failure. With one, routing is posture-blind as before —
+     * the developer's own credential serves anything unrouted, and nothing
+     * diverts it onto org spend.
+     */
+    const decision = inbound.upstreamCredential === null
+        ? resolveWithoutCallerCredential(ctx.routes, peek.model)
+        : resolveRoute(ctx.routes, peek.model);
+    if (decision.pipeline === "passthrough") {
+        // Hand the already-read bytes on rather than re-reading a consumed stream.
+        await handleMessages(req, res, { ...ctx, routeId: decision.route?.id ?? null }, body);
+        return;
+    }
+    // ── substitute ──────────────────────────────────────────────────────────────
+    const resolved = ctx.resolveIdentity(inbound.identityToken);
+    if (resolved !== null)
+        ctx.touchToken?.(resolved.tokenId);
+    const isSubscription = inbound.upstreamCredential !== null && isSubscriptionCredential(inbound.upstreamCredential);
+    const credential = resolveCredential(decision, inbound.upstreamCredential === null
+        ? null
+        : { fingerprint: inbound.upstreamCredential.fingerprint, isSubscription }, ctx.secrets);
+    const base = {
+        id: randomUUID(),
+        startedAt: Date.now(),
+        posture: inbound.posture,
+        identityCarrier: inbound.identity.carrier,
+        callerFingerprint: inbound.identity.tokenFingerprint,
+        userId: resolved?.userId ?? null,
+        tokenId: resolved?.tokenId ?? null,
+        // The credential that SERVED the request is the server's, so the inbound
+        // fingerprint is not the right thing to record here. It is null rather
+        // than the caller's, because recording the caller's would read as though
+        // theirs was used.
+        credentialFingerprint: null,
+        // A server-held credential is org spend, always. This is the field the
+        // posture screen keys off.
+        credentialOrigin: "fallback_server",
+        sessionId: headerValue(req, "x-claude-code-session-id"),
+        clientVersion: headerValue(req, "user-agent"),
+    };
+    if (!credential.ok) {
+        log.warn("substitute refused: no server credential", {
+            route: decision.route?.id ?? null,
+            upstream: decision.upstream?.id ?? null,
+            model: decision.requestedModel,
+        });
+        refuse(res, ctx.sink, base, decision, credential.considered, credential.message);
+        return;
+    }
+    const adapter = ADAPTERS[decision.upstream?.adapter ?? ""];
+    if (adapter === undefined || credential.secret === null) {
+        // Unreachable through validated config; refusing beats guessing.
+        refuse(res, ctx.sink, base, decision, credential.considered, `Fest: no adapter for upstream ${JSON.stringify(decision.upstream?.id ?? "?")}.`);
+        return;
+    }
+    await handleSubstitute(req, res, {
+        base,
+        decision,
+        adapter,
+        secret: credential.secret,
+        considered: credential.considered,
+        body,
+        path: inbound.effectivePath,
+        betas: headerValue(req, "anthropic-beta"),
+        stream: peek.stream,
+        sink: ctx.sink,
+    });
+}
+function headerValue(req, name) {
+    const v = req.headers[name];
+    if (v === undefined)
+        return null;
+    return Array.isArray(v) ? (v[0] ?? null) : v;
+}
+//# sourceMappingURL=dispatch.js.map

package/dist/server/pipeline/dispatch.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"dispatch.js","sourceRoot":"","sources":["../../../server/pipeline/dispatch.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;GAsBG;AAEH,OAAO,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AAGzC,OAAO,EAAE,aAAa,EAAE,MAAM,oBAAoB,CAAC;AACnD,OAAO,EAAE,wBAAwB,EAAE,MAAM,0BAA0B,CAAC;AACpE,OAAO,EAAE,aAAa,EAAE,UAAU,EAAE,WAAW,EAAE,MAAM,iBAAiB,CAAC;AACzE,OAAO,EAAE,cAAc,EAAE,kBAAkB,EAAE,MAAM,mBAAmB,CAAC;AACvE,OAAO,EACL,YAAY,EACZ,8BAA8B,EAC9B,wBAAwB,GACzB,MAAM,sBAAsB,CAAC;AAE9B,OAAO,EAAE,iBAAiB,EAAE,MAAM,2BAA2B,CAAC;AAE9D,OAAO,EAAE,cAAc,EAAE,MAAM,kBAAkB,CAAC;AAElD,OAAO,EAAE,gBAAgB,EAAE,MAAM,EAAE,MAAM,iBAAiB,CAAC;AAC3D,OAAO,EAAE,gBAAgB,EAAE,MAAM,0BAA0B,CAAC;AAC5D,OAAO,EAAE,gBAAgB,EAAE,MAAM,0BAA0B,CAAC;AAG5D,OAAO,EAAE,GAAG,EAAE,MAAM,WAAW,CAAC;AAEhC,MAAM,cAAc,GAAG,EAAE,GAAG,IAAI,GAAG,IAAI,CAAC;AAExC,MAAM,QAAQ,GAAsC;IAClD,SAAS,EAAE,gBAAgB;IAC3B,SAAS,EAAE,gBAAgB;CAC5B,CAAC;AAQF,MAAM,CAAC,KAAK,UAAU,gBAAgB,CACpC,GAAoB,EACpB,GAAmB,EACnB,GAAoB;IAEpB,0EAA0E;IAC1E,0EAA0E;IAC1E,oDAAoD;IACpD,EAAE;IACF,uEAAuE;IACvE,6EAA6E;IAC7E,+CAA+C;IAC/C,IAAI,GAAG,CAAC,MAAM,CAAC,MAAM,CAAC,MAAM,KAAK,CAAC,IAAI,wBAAwB,CAAC,GAAG,CAAC,MAAM,CAAC,KAAK,IAAI,EAAE,CAAC;QACpF,MAAM,cAAc,CAAC,GAAG,EAAE,GAAG,EAAE,GAAG,CAAC,CAAC;QACpC,OAAO;IACT,CAAC;IAED,MAAM,OAAO,GAAG,aAAa,CAAC,GAAG,CAAC,GAAG,IAAI,GAAG,EAAE,GAAG,CAAC,OAAO,CAAC,CAAC;IAC3D,MAAM,IAAI,GAAG,MAAM,aAAa,CAAC,GAAG,EAAE,cAAc,CAAC,CAAC;IACtD,IAAI,UAAU,CAAC,IAAI,CAAC,EAAE,CAAC;QACrB,MAAM,IAAI,GAAG,uBAAuB,CAAC;QACrC,GAAG,CAAC,SAAS,CAAC,kBAAkB,CAAC,IAAI,CAAC,EAAE,EAAE,cAAc,EAAE,kBAAkB,EAAE,CAAC,CAAC;QAChF,GAAG,CAAC,GAAG,CAAC,cAAc,CAAC,IAAI,EAAE,8BAA8B,cAAc,SAAS,CAAC,CAAC,CAAC;QACrF,OAAO;IACT,CAAC;IAED,MAAM,IAAI,GAAG,WAAW,CAAC,IAAI,CAAC,CAAC;IAC/B;;;;;;OAMG;IACH,MAAM,QAAQ,GACZ,OAAO,CAAC,kBAAkB,KAAK,IAAI;QACjC,CAAC,CAAC,8BAA8B,CAAC,GAAG,CAAC,MAAM,EAAE,IAAI,CAAC,KAAK,CAAC;QACxD,CAAC,CAAC,YAAY,CAAC,GAAG,CAAC,MAAM,EAAE,IAAI,CAAC,KAAK,CAAC,CAAC;IAE3C,IAAI,QAAQ,CAAC,QAAQ,KAAK,aAAa,EAAE,CAAC;QACxC,2EAA2E;QAC3E,MAAM,cAAc,CAAC,GAAG,EAAE,GAAG,EAAE,EAAE,GAAG,GAAG,EAAE,OAAO,EAAE,QAAQ,CAAC,KAAK,EAAE,EAAE,IAAI,IAAI,EAAE,EAAE,IAAI,CAAC,CAAC;QACtF,OAAO;IACT,CAAC;IAED,+EAA+E;IAC/E,MAAM,QAAQ,GAAG,GAAG,CAAC,eAAe,CAAC,OAAO,CAAC,aAAa,CAAC,CAAC;IAC5D,IAAI,QAAQ,KAAK,IAAI;QAAE,GAAG,CAAC,UAAU,EAAE,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC;IAE1D,MAAM,cAAc,GAClB,OAAO,CAAC,kBAAkB,KAAK,IAAI,IAAI,wBAAwB,CAAC,OAAO,CAAC,kBAAkB,CAAC,CAAC;IAE9F,MAAM,UAAU,GAAG,iBAAiB,CAClC,QAAQ,EACR,OAAO,CAAC,kBAAkB,KAAK,IAAI;QACjC,CAAC,CAAC,IAAI;QACN,CAAC,CAAC,EAAE,WAAW,EAAE,OAAO,CAAC,kBAAkB,CAAC,WAAW,EAAE,cAAc,EAAE,EAC3E,GAAG,CAAC,OAAO,CACZ,CAAC;IAEF,MAAM,IAAI,GAAG;QACX,EAAE,EAAE,UAAU,EAAE;QAChB,SAAS,EAAE,IAAI,CAAC,GAAG,EAAE;QACrB,OAAO,EAAE,OAAO,CAAC,OAAO;QACxB,eAAe,EAAE,OAAO,CAAC,QAAQ,CAAC,OAAO;QACzC,iBAAiB,EAAE,OAAO,CAAC,QAAQ,CAAC,gBAAgB;QACpD,MAAM,EAAE,QAAQ,EAAE,MAAM,IAAI,IAAI;QAChC,OAAO,EAAE,QAAQ,EAAE,OAAO,IAAI,IAAI;QAClC,yEAAyE;QACzE,uEAAuE;QACvE,yEAAyE;QACzE,mBAAmB;QACnB,qBAAqB,EAAE,IAAI;QAC3B,uEAAuE;QACvE,2BAA2B;QAC3B,gBAAgB,EAAE,iBAAqC;QACvD,SAAS,EAAE,WAAW,CAAC,GAAG,EAAE,0BAA0B,CAAC;QACvD,aAAa,EAAE,WAAW,CAAC,GAAG,EAAE,YAAY,CAAC;KAC9C,CAAC;IAEF,IAAI,CAAC,UAAU,CAAC,EAAE,EAAE,CAAC;QACnB,GAAG,CAAC,IAAI,CAAC,0CAA0C,EAAE;YACnD,KAAK,EAAE,QAAQ,CAAC,KAAK,EAAE,EAAE,IAAI,IAAI;YACjC,QAAQ,EAAE,QAAQ,CAAC,QAAQ,EAAE,EAAE,IAAI,IAAI;YACvC,KAAK,EAAE,QAAQ,CAAC,cAAc;SAC/B,CAAC,CAAC;QACH,MAAM,CAAC,GAAG,EAAE,GAAG,CAAC,IAAI,EAAE,IAAI,EAAE,QAAQ,EAAE,UAAU,CAAC,UAAU,EAAE,UAAU,CAAC,OAAO,CAAC,CAAC;QACjF,OAAO;IACT,CAAC;IAED,MAAM,OAAO,GAAG,QAAQ,CAAC,QAAQ,CAAC,QAAQ,EAAE,OAAO,IAAI,EAAE,CAAC,CAAC;IAC3D,IAAI,OAAO,KAAK,SAAS,IAAI,UAAU,CAAC,MAAM,KAAK,IAAI,EAAE,CAAC;QACxD,iEAAiE;QACjE,MAAM,CACJ,GAAG,EACH,GAAG,CAAC,IAAI,EACR,IAAI,EACJ,QAAQ,EACR,UAAU,CAAC,UAAU,EACrB,iCAAiC,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,QAAQ,EAAE,EAAE,IAAI,GAAG,CAAC,GAAG,CACjF,CAAC;QACF,OAAO;IACT,CAAC;IAED,MAAM,gBAAgB,CAAC,GAAG,EAAE,GAAG,EAAE;QAC/B,IAAI;QACJ,QAAQ;QACR,OAAO;QACP,MAAM,EAAE,UAAU,CAAC,MAAM;QACzB,UAAU,EAAE,UAAU,CAAC,UAAU;QACjC,IAAI;QACJ,IAAI,EAAE,OAAO,CAAC,aAAa;QAC3B,KAAK,EAAE,WAAW,CAAC,GAAG,EAAE,gBAAgB,CAAC;QACzC,MAAM,EAAE,IAAI,CAAC,MAAM;QACnB,IAAI,EAAE,GAAG,CAAC,IAAI;KACf,CAAC,CAAC;AACL,CAAC;AAED,SAAS,WAAW,CAAC,GAAoB,EAAE,IAAY;IACrD,MAAM,CAAC,GAAG,GAAG,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC;IAC5B,IAAI,CAAC,KAAK,SAAS;QAAE,OAAO,IAAI,CAAC;IACjC,OAAO,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;AAC/C,CAAC"}

package/dist/server/pipeline/passthrough.js

@@ -0,0 +1,392 @@
+/**
+ * The pass-through pipeline: relay a request to Anthropic unchanged, and
+ * observe it on the way past.
+ *
+ * This pipeline deliberately does NOT use a provider-adapter interface.
+ * Adapters exist to transform requests; this path exists precisely not to. On
+ * the subscription posture the developer's own OAuth bearer is forwarded and
+ * Anthropic validates the request shape, so every byte of the body and every
+ * inbound header is sacred. Keeping this as its own pipeline — rather than a
+ * "no-op adapter" — is what makes that rule enforceable instead of aspirational,
+ * and leaves a future routing/rewriting path free to grow without endangering
+ * it.
+ *
+ * What Fest is allowed to do here: authenticate the caller, relay bytes, tee
+ * the stream for metering, and record what happened.
+ */
+import { randomUUID } from "node:crypto";
+import { EMPTY_USAGE } from "../../shared/types.js";
+import { detectInbound } from "../auth/posture.js";
+import { buildUpstreamHeaders, buildDownstreamHeaders, parseRateLimit } from "../http/headers.js";
+import { anthropicError, statusForErrorType, sseErrorEvent, describeUpstreamFailure, } from "../http/errors.js";
+import { readBodyBytes, isTooLarge, peekRequest } from "../http/body.js";
+import { createSseParser } from "../http/sse.js";
+import { createUsageAccumulator, usageFromJson } from "../usage/accumulator.js";
+import { priceUsage } from "../usage/pricing.js";
+import { pipeWithTee, beginStream } from "../http/pipe.js";
+import { isSubscriptionCredential } from "../secret/fingerprint.js";
+import { badIdentity, noIdentity, noUpstreamCredential } from "../auth/gateway-401.js";
+import { isWarmupPreflightRefusal } from "./preflight.js";
+import { log } from "../log.js";
+const MAX_BODY_BYTES = 32 * 1024 * 1024;
+function credentialOrigin(hasSubscription, hasKey) {
+    if (hasSubscription)
+        return "inbound_subscription";
+    if (hasKey)
+        return "inbound_key";
+    return "none";
+}
+export async function handleMessages(req, res, ctx, 
+/**
+ * Bytes already read from the request, when the dispatcher had to read them
+ * to see the model id. The stream can only be consumed once, so re-reading
+ * here would hang. Still the SAME bytes — the dispatcher inspects a copy and
+ * never rewrites on this path.
+ */
+preRead) {
+    const startedAt = Date.now();
+    const id = randomUUID();
+    const inbound = detectInbound(req.url ?? "/", req.headers);
+    const sessionId = headerValue(req, "x-claude-code-session-id");
+    const clientVersion = headerValue(req, "user-agent");
+    const credentialsConsidered = inbound.upstreamCredential === null
+        ? [{ source: "inbound", result: "missing", reason: "no credential on the request" }]
+        : [
+            {
+                source: isSubscriptionCredential(inbound.upstreamCredential)
+                    ? "inbound_subscription"
+                    : "inbound_key",
+                result: "used",
+            },
+        ];
+    // Resolve identity before anything else, but note the two failure modes are
+    // NOT the same:
+    //   * a token was presented and did not resolve  -> always an error, because
+    //     the developer believes they are attributed and they are not;
+    //   * no token at all -> an error only when the deployment requires one.
+    // Conflating them would let a revoked token silently degrade to
+    // "unattributed" instead of telling the developer their token is dead.
+    const presentedToken = inbound.identityToken;
+    const resolved = ctx.resolveIdentity(presentedToken);
+    if (resolved !== null)
+        ctx.touchToken?.(resolved.tokenId);
+    const finish = (partial) => {
+        const rec = {
+            id,
+            startedAt,
+            endedAt: Date.now(),
+            posture: inbound.posture,
+            identityCarrier: inbound.identity.carrier,
+            callerFingerprint: inbound.identity.tokenFingerprint,
+            userId: resolved?.userId ?? null,
+            tokenId: resolved?.tokenId ?? null,
+            credentialFingerprint: inbound.upstreamCredential?.fingerprint ?? null,
+            credentialOrigin: credentialOrigin(inbound.upstreamCredential !== null && isSubscriptionCredential(inbound.upstreamCredential), inbound.upstreamCredential !== null && !isSubscriptionCredential(inbound.upstreamCredential)),
+            sessionId,
+            requestedModel: null,
+            servedModel: null,
+            upstream: ctx.upstreamBaseUrl,
+            stream: false,
+            httpStatus: null,
+            partial: false,
+            usage: EMPTY_USAGE,
+            costUsd: null,
+            costBasis: "none",
+            notionalCostUsd: null,
+            ttfbMs: null,
+            durationMs: Date.now() - startedAt,
+            bytesIn: 0,
+            bytesOut: 0,
+            upstreamRequestId: null,
+            rateLimit: null,
+            clientVersion,
+            pipeline: "passthrough",
+            routeId: ctx.routeId ?? null,
+            // Recorded even though this path never chooses: a record that says
+            // "the caller's own subscription served this, nothing else was tried"
+            // is what makes the absence of a substitution provable rather than
+            // assumed.
+            credentialsConsidered,
+            ...partial,
+        };
+        // Fire and forget: metering must never delay or fail a request.
+        ctx.sink.record(rec);
+    };
+    const denyIdentity = (message) => {
+        const type = "authentication_error";
+        finish({ status: "identity_denied", httpStatus: statusForErrorType(type) });
+        res.writeHead(statusForErrorType(type), {
+            "content-type": "application/json",
+            // Nothing a retry can fix; the client honours this ahead of its own
+            // status-code rules, so say so rather than letting it back off ten times.
+            "x-should-retry": "false",
+        });
+        // Claude Code renders error.message to the developer, so this string is
+        // the user-facing UI. Write it for a human who is mid-task.
+        res.end(anthropicError(type, message));
+    };
+    if (presentedToken !== null && resolved === null) {
+        denyIdentity(badIdentity().message);
+        return;
+    }
+    // Identity is separate from the upstream credential and is never forwarded.
+    // Without it usage cannot be attributed, which for a team gateway defeats
+    // the point — so a team deployment sets FEST_REQUIRE_IDENTITY=1.
+    if (ctx.requireIdentity && resolved === null) {
+        denyIdentity(noIdentity(ctx.publicBaseUrl ?? "https://<fest>").message);
+        return;
+    }
+    const bodyOrTooLarge = preRead ?? (await readBodyBytes(req, MAX_BODY_BYTES));
+    if (isTooLarge(bodyOrTooLarge)) {
+        const type = "invalid_request_error";
+        finish({ status: "bad_request", httpStatus: statusForErrorType(type) });
+        res.writeHead(statusForErrorType(type), { "content-type": "application/json" });
+        res.end(anthropicError(type, `Fest: request body exceeds ${MAX_BODY_BYTES} bytes.`));
+        return;
+    }
+    const body = bodyOrTooLarge;
+    const peek = peekRequest(body);
+    const isSubscription = inbound.upstreamCredential !== null && isSubscriptionCredential(inbound.upstreamCredential);
+    // Key posture with a Fest token and nothing to pay with: the credential
+    // position holds our OWN token, which we will not forward, and no route
+    // claimed this model. Anthropic's 401 would be accurate but useless — it
+    // would describe a credential the developer never sent. Checked here, after
+    // the body, so the message can name the model that has no home.
+    if (inbound.upstreamCredential === null && inbound.identity.carrier === "auth_header") {
+        const failure = noUpstreamCredential(peek.model);
+        finish({
+            status: "bad_request",
+            httpStatus: failure.status,
+            requestedModel: peek.model,
+            bytesIn: body.byteLength,
+            errorType: failure.type,
+        });
+        res.writeHead(failure.status, {
+            "content-type": "application/json",
+            "x-should-retry": String(failure.retryable),
+        });
+        res.end(anthropicError(failure.type, failure.message));
+        return;
+    }
+    const target = new URL(inbound.effectivePath, ctx.upstreamBaseUrl);
+    const headers = buildUpstreamHeaders(req.headers, {
+        posture: inbound.posture,
+        upstreamHost: target.host,
+    });
+    const abort = new AbortController();
+    let upstream;
+    try {
+        upstream = await fetch(target, {
+            method: req.method ?? "POST",
+            headers,
+            // The exact bytes we received. No parse, no re-serialise.
+            body,
+            signal: abort.signal,
+            redirect: "error",
+        });
+    }
+    catch (err) {
+        // Failed before any bytes were sent, so we can still choose a status.
+        const type = "api_error";
+        log.warn("upstream request failed", { id, error: String(err).slice(0, 200) });
+        finish({
+            status: "upstream_error",
+            httpStatus: statusForErrorType(type),
+            requestedModel: peek.model,
+            bytesIn: body.byteLength,
+            errorType: type,
+            errorMessage: String(err).slice(0, 200),
+        });
+        res.writeHead(statusForErrorType(type), { "content-type": "application/json" });
+        res.end(anthropicError(type, "Fest: upstream request failed."));
+        return;
+    }
+    const rateLimit = parseRateLimit(upstream.headers);
+    const upstreamRequestId = upstream.headers.get("request-id") ?? upstream.headers.get("anthropic-request-id");
+    const downstream = buildDownstreamHeaders(upstream.headers, { stream: peek.stream });
+    const base = {
+        requestedModel: peek.model,
+        servedModel: peek.model,
+        stream: peek.stream,
+        httpStatus: upstream.status,
+        bytesIn: body.byteLength,
+        upstreamRequestId,
+        rateLimit,
+    };
+    // ── upstream refused ───────────────────────────────────────────────────────
+    /**
+     * Handled before the stream/non-stream split, because a refusal is the same
+     * shape either way: a short JSON body instead of an SSE stream, even when the
+     * client asked for one. Buffering it costs nothing and is the only chance to
+     * read it — a teed stream is relayed and discarded.
+     *
+     * This is the response whose CONTENT matters most and the one Fest used to
+     * record least. It recorded the status and dropped the reason, so a row said
+     * "400" and nothing about which field the provider objected to; the answer
+     * reached the developer's terminal and lived only until it scrolled.
+     */
+    if (!upstream.ok) {
+        const failure = describeUpstreamFailure(upstream.status, await upstream.text());
+        /**
+         * Anthropic refusing the session-start warmup ping is expected and benign —
+         * see `preflight.ts`. It still gets a row, with the provider's own reason
+         * on it, because a thing that happens is worth recording; it just is not a
+         * failure and is not warned about, or every session would open with an
+         * error in the dashboard and a warning in the log.
+         */
+        const preflight = isWarmupPreflightRefusal({
+            httpStatus: upstream.status,
+            stream: peek.stream,
+            maxTokens: peek.maxTokens,
+            rateLimit,
+            subscription: isSubscription,
+        });
+        /**
+         * Logged as well as recorded, because the database is the wrong place to
+         * find out something is broken NOW. Fest logs nothing per request by
+         * design — a gateway that narrates every relay is unreadable — but a
+         * refusal is not routine traffic, and a developer watching an error appear
+         * in their client had no way to see the reason without querying SQLite.
+         *
+         * `rateLimit` rides along: a 429 that carries no quota headers is a
+         * different fact from one that does, and that difference is invisible in
+         * the status alone.
+         */
+        if (!preflight) {
+            log.warn("upstream refused", {
+                id,
+                status: upstream.status,
+                type: failure.type,
+                message: failure.message,
+                model: peek.model,
+                stream: peek.stream,
+                bytesIn: body.byteLength,
+                pipeline: "passthrough",
+                upstreamRequestId,
+                rateLimit,
+            });
+        }
+        // `stream: false`: whatever the client asked for, this is a JSON error, and
+        // framing it as a stream would tell the client to expect SSE it won't get.
+        res.writeHead(upstream.status, buildDownstreamHeaders(upstream.headers, { stream: false }));
+        res.end(failure.text);
+        finish({
+            ...base,
+            status: preflight ? "preflight_refused" : "upstream_error",
+            bytesOut: Buffer.byteLength(failure.text),
+            errorType: failure.type,
+            errorMessage: failure.message,
+        });
+        return;
+    }
+    // ── non-streaming ──────────────────────────────────────────────────────────
+    if (!peek.stream || !upstream.body) {
+        const text = await upstream.text();
+        let usage = EMPTY_USAGE;
+        try {
+            usage = usageFromJson(JSON.parse(text)) ?? EMPTY_USAGE;
+        }
+        catch {
+            // A body we cannot read is still relayed; we just report less.
+        }
+        const priced = priceUsage(peek.model, usage, isSubscription);
+        res.writeHead(upstream.status, downstream);
+        res.end(text);
+        finish({
+            ...base,
+            status: "ok",
+            usage,
+            costUsd: priced.cost,
+            costBasis: priced.basis,
+            notionalCostUsd: priced.notionalCost,
+            bytesOut: Buffer.byteLength(text),
+        });
+        return;
+    }
+    // ── streaming ──────────────────────────────────────────────────────────────
+    const parser = createSseParser();
+    const acc = createUsageAccumulator();
+    beginStream(res, upstream.status, downstream);
+    try {
+        const result = await pipeWithTee(upstream.body, res, {
+            abort,
+            startedAt,
+            observe: (chunk) => {
+                for (const event of parser.push(chunk))
+                    acc.apply(event);
+            },
+        });
+        for (const event of parser.flush())
+            acc.apply(event);
+        const usage = acc.snapshot();
+        const priced = priceUsage(peek.model, usage, isSubscription);
+        const streamErr = acc.streamError();
+        /**
+         * Only a 2xx reaches here — a non-2xx was recorded and relayed above, which
+         * is what keeps an upstream that refuses everything from reading as a zero
+         * error rate. (It once did: status was derived purely from client aborts
+         * and in-band SSE error events, so a provider refusing before emitting any
+         * event produced a tidy, successful-looking record.)
+         *
+         * What remains are failures that happen once the stream is under way, and
+         * a client abort wins over them: the developer walking away is the more
+         * specific fact about what happened.
+         */
+        let status = "ok";
+        if (result.clientAborted)
+            status = "client_abort";
+        else if (streamErr)
+            status = "stream_error";
+        if (!result.clientAborted)
+            res.end();
+        finish({
+            ...base,
+            status,
+            // Input tokens were already consumed upstream even on an abort, so the
+            // partial usage is real and worth recording — just flagged as incomplete.
+            partial: result.clientAborted || !result.completed,
+            usage,
+            costUsd: priced.cost,
+            costBasis: priced.basis,
+            notionalCostUsd: priced.notionalCost,
+            ttfbMs: result.ttfbMs,
+            bytesOut: result.bytesOut,
+            ...(streamErr ? { errorType: streamErr.type, errorMessage: streamErr.message } : {}),
+        });
+        if (result.observerErrors > 0) {
+            log.warn("metering observer errored mid-stream", { id, errors: result.observerErrors });
+        }
+    }
+    catch (err) {
+        // Headers are already sent, so the status cannot be changed. Synthesise a
+        // terminal SSE error frame so the client sees a clean end rather than a
+        // truncated stream. Never retry: partial text is already on screen.
+        for (const event of parser.flush())
+            acc.apply(event);
+        const usage = acc.snapshot();
+        const priced = priceUsage(peek.model, usage, isSubscription);
+        log.warn("upstream stream failed mid-flight", { id, error: String(err).slice(0, 200) });
+        if (!res.writableEnded) {
+            res.write(sseErrorEvent("api_error", "Fest: upstream connection closed mid-stream."));
+            res.end();
+        }
+        finish({
+            ...base,
+            status: "stream_error",
+            partial: true,
+            usage,
+            costUsd: priced.cost,
+            costBasis: priced.basis,
+            notionalCostUsd: priced.notionalCost,
+            errorType: "api_error",
+            errorMessage: String(err).slice(0, 200),
+        });
+    }
+}
+function headerValue(req, name) {
+    const v = req.headers[name];
+    if (v === undefined)
+        return null;
+    return Array.isArray(v) ? (v[0] ?? null) : v;
+}
+//# sourceMappingURL=passthrough.js.map

package/dist/server/pipeline/passthrough.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"passthrough.js","sourceRoot":"","sources":["../../../server/pipeline/passthrough.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAEH,OAAO,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AAQzC,OAAO,EAAE,WAAW,EAAE,MAAM,uBAAuB,CAAC;AACpD,OAAO,EAAE,aAAa,EAAE,MAAM,oBAAoB,CAAC;AACnD,OAAO,EAAE,oBAAoB,EAAE,sBAAsB,EAAE,cAAc,EAAE,MAAM,oBAAoB,CAAC;AAClG,OAAO,EACL,cAAc,EACd,kBAAkB,EAClB,aAAa,EACb,uBAAuB,GACxB,MAAM,mBAAmB,CAAC;AAC3B,OAAO,EAAE,aAAa,EAAE,UAAU,EAAE,WAAW,EAAE,MAAM,iBAAiB,CAAC;AACzE,OAAO,EAAE,eAAe,EAAE,MAAM,gBAAgB,CAAC;AACjD,OAAO,EAAE,sBAAsB,EAAE,aAAa,EAAE,MAAM,yBAAyB,CAAC;AAChF,OAAO,EAAE,UAAU,EAAE,MAAM,qBAAqB,CAAC;AACjD,OAAO,EAAE,WAAW,EAAE,WAAW,EAAE,MAAM,iBAAiB,CAAC;AAC3D,OAAO,EAAE,wBAAwB,EAAE,MAAM,0BAA0B,CAAC;AAEpE,OAAO,EAAE,WAAW,EAAE,UAAU,EAAE,oBAAoB,EAAE,MAAM,wBAAwB,CAAC;AACvF,OAAO,EAAE,wBAAwB,EAAE,MAAM,gBAAgB,CAAC;AAC1D,OAAO,EAAE,GAAG,EAAE,MAAM,WAAW,CAAC;AAEhC,MAAM,cAAc,GAAG,EAAE,GAAG,IAAI,GAAG,IAAI,CAAC;AA0BxC,SAAS,gBAAgB,CAAC,eAAwB,EAAE,MAAe;IACjE,IAAI,eAAe;QAAE,OAAO,sBAAsB,CAAC;IACnD,IAAI,MAAM;QAAE,OAAO,aAAa,CAAC;IACjC,OAAO,MAAM,CAAC;AAChB,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,cAAc,CAClC,GAAoB,EACpB,GAAmB,EACnB,GAAuB;AACvB;;;;;GAKG;AACH,OAAiC;IAEjC,MAAM,SAAS,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;IAC7B,MAAM,EAAE,GAAG,UAAU,EAAE,CAAC;IACxB,MAAM,OAAO,GAAG,aAAa,CAAC,GAAG,CAAC,GAAG,IAAI,GAAG,EAAE,GAAG,CAAC,OAAO,CAAC,CAAC;IAE3D,MAAM,SAAS,GAAG,WAAW,CAAC,GAAG,EAAE,0BAA0B,CAAC,CAAC;IAC/D,MAAM,aAAa,GAAG,WAAW,CAAC,GAAG,EAAE,YAAY,CAAC,CAAC;IAErD,MAAM,qBAAqB,GACzB,OAAO,CAAC,kBAAkB,KAAK,IAAI;QACjC,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,SAAS,EAAE,MAAM,EAAE,SAAS,EAAE,MAAM,EAAE,8BAA8B,EAAE,CAAC;QACpF,CAAC,CAAC;YACE;gBACE,MAAM,EAAE,wBAAwB,CAAC,OAAO,CAAC,kBAAkB,CAAC;oBAC1D,CAAC,CAAC,sBAAsB;oBACxB,CAAC,CAAC,aAAa;gBACjB,MAAM,EAAE,MAAM;aACf;SACF,CAAC;IAER,4EAA4E;IAC5E,gBAAgB;IAChB,6EAA6E;IAC7E,mEAAmE;IACnE,yEAAyE;IACzE,gEAAgE;IAChE,uEAAuE;IACvE,MAAM,cAAc,GAAG,OAAO,CAAC,aAAa,CAAC;IAC7C,MAAM,QAAQ,GAAG,GAAG,CAAC,eAAe,CAAC,cAAc,CAAC,CAAC;IACrD,IAAI,QAAQ,KAAK,IAAI;QAAE,GAAG,CAAC,UAAU,EAAE,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC;IAE1D,MAAM,MAAM,GAAG,CAAC,OAAyD,EAAQ,EAAE;QACjF,MAAM,GAAG,GAAgB;YACvB,EAAE;YACF,SAAS;YACT,OAAO,EAAE,IAAI,CAAC,GAAG,EAAE;YACnB,OAAO,EAAE,OAAO,CAAC,OAAO;YACxB,eAAe,EAAE,OAAO,CAAC,QAAQ,CAAC,OAAO;YACzC,iBAAiB,EAAE,OAAO,CAAC,QAAQ,CAAC,gBAAgB;YACpD,MAAM,EAAE,QAAQ,EAAE,MAAM,IAAI,IAAI;YAChC,OAAO,EAAE,QAAQ,EAAE,OAAO,IAAI,IAAI;YAClC,qBAAqB,EAAE,OAAO,CAAC,kBAAkB,EAAE,WAAW,IAAI,IAAI;YACtE,gBAAgB,EAAE,gBAAgB,CAChC,OAAO,CAAC,kBAAkB,KAAK,IAAI,IAAI,wBAAwB,CAAC,OAAO,CAAC,kBAAkB,CAAC,EAC3F,OAAO,CAAC,kBAAkB,KAAK,IAAI,IAAI,CAAC,wBAAwB,CAAC,OAAO,CAAC,kBAAkB,CAAC,CAC7F;YACD,SAAS;YACT,cAAc,EAAE,IAAI;YACpB,WAAW,EAAE,IAAI;YACjB,QAAQ,EAAE,GAAG,CAAC,eAAe;YAC7B,MAAM,EAAE,KAAK;YACb,UAAU,EAAE,IAAI;YAChB,OAAO,EAAE,KAAK;YACd,KAAK,EAAE,WAAW;YAClB,OAAO,EAAE,IAAI;YACb,SAAS,EAAE,MAAM;YACjB,eAAe,EAAE,IAAI;YACrB,MAAM,EAAE,IAAI;YACZ,UAAU,EAAE,IAAI,CAAC,GAAG,EAAE,GAAG,SAAS;YAClC,OAAO,EAAE,CAAC;YACV,QAAQ,EAAE,CAAC;YACX,iBAAiB,EAAE,IAAI;YACvB,SAAS,EAAE,IAAI;YACf,aAAa;YACb,QAAQ,EAAE,aAAa;YACvB,OAAO,EAAE,GAAG,CAAC,OAAO,IAAI,IAAI;YAC5B,mEAAmE;YACnE,sEAAsE;YACtE,mEAAmE;YACnE,WAAW;YACX,qBAAqB;YACrB,GAAG,OAAO;SACX,CAAC;QACF,gEAAgE;QAChE,GAAG,CAAC,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;IACvB,CAAC,CAAC;IAEF,MAAM,YAAY,GAAG,CAAC,OAAe,EAAQ,EAAE;QAC7C,MAAM,IAAI,GAAG,sBAAsB,CAAC;QACpC,MAAM,CAAC,EAAE,MAAM,EAAE,iBAAiB,EAAE,UAAU,EAAE,kBAAkB,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QAC5E,GAAG,CAAC,SAAS,CAAC,kBAAkB,CAAC,IAAI,CAAC,EAAE;YACtC,cAAc,EAAE,kBAAkB;YAClC,oEAAoE;YACpE,0EAA0E;YAC1E,gBAAgB,EAAE,OAAO;SAC1B,CAAC,CAAC;QACH,wEAAwE;QACxE,4DAA4D;QAC5D,GAAG,CAAC,GAAG,CAAC,cAAc,CAAC,IAAI,EAAE,OAAO,CAAC,CAAC,CAAC;IACzC,CAAC,CAAC;IAEF,IAAI,cAAc,KAAK,IAAI,IAAI,QAAQ,KAAK,IAAI,EAAE,CAAC;QACjD,YAAY,CAAC,WAAW,EAAE,CAAC,OAAO,CAAC,CAAC;QACpC,OAAO;IACT,CAAC;IAED,4EAA4E;IAC5E,0EAA0E;IAC1E,iEAAiE;IACjE,IAAI,GAAG,CAAC,eAAe,IAAI,QAAQ,KAAK,IAAI,EAAE,CAAC;QAC7C,YAAY,CAAC,UAAU,CAAC,GAAG,CAAC,aAAa,IAAI,gBAAgB,CAAC,CAAC,OAAO,CAAC,CAAC;QACxE,OAAO;IACT,CAAC;IAED,MAAM,cAAc,GAAG,OAAO,IAAI,CAAC,MAAM,aAAa,CAAC,GAAG,EAAE,cAAc,CAAC,CAAC,CAAC;IAC7E,IAAI,UAAU,CAAC,cAAc,CAAC,EAAE,CAAC;QAC/B,MAAM,IAAI,GAAG,uBAAuB,CAAC;QACrC,MAAM,CAAC,EAAE,MAAM,EAAE,aAAa,EAAE,UAAU,EAAE,kBAAkB,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QACxE,GAAG,CAAC,SAAS,CAAC,kBAAkB,CAAC,IAAI,CAAC,EAAE,EAAE,cAAc,EAAE,kBAAkB,EAAE,CAAC,CAAC;QAChF,GAAG,CAAC,GAAG,CAAC,cAAc,CAAC,IAAI,EAAE,8BAA8B,cAAc,SAAS,CAAC,CAAC,CAAC;QACrF,OAAO;IACT,CAAC;IAED,MAAM,IAAI,GAAG,cAAc,CAAC;IAC5B,MAAM,IAAI,GAAG,WAAW,CAAC,IAAI,CAAC,CAAC;IAC/B,MAAM,cAAc,GAAG,OAAO,CAAC,kBAAkB,KAAK,IAAI,IAAI,wBAAwB,CAAC,OAAO,CAAC,kBAAkB,CAAC,CAAC;IAEnH,wEAAwE;IACxE,wEAAwE;IACxE,yEAAyE;IACzE,4EAA4E;IAC5E,gEAAgE;IAChE,IAAI,OAAO,CAAC,kBAAkB,KAAK,IAAI,IAAI,OAAO,CAAC,QAAQ,CAAC,OAAO,KAAK,aAAa,EAAE,CAAC;QACtF,MAAM,OAAO,GAAG,oBAAoB,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;QACjD,MAAM,CAAC;YACL,MAAM,EAAE,aAAa;YACrB,UAAU,EAAE,OAAO,CAAC,MAAM;YAC1B,cAAc,EAAE,IAAI,CAAC,KAAK;YAC1B,OAAO,EAAE,IAAI,CAAC,UAAU;YACxB,SAAS,EAAE,OAAO,CAAC,IAAI;SACxB,CAAC,CAAC;QACH,GAAG,CAAC,SAAS,CAAC,OAAO,CAAC,MAAM,EAAE;YAC5B,cAAc,EAAE,kBAAkB;YAClC,gBAAgB,EAAE,MAAM,CAAC,OAAO,CAAC,SAAS,CAAC;SAC5C,CAAC,CAAC;QACH,GAAG,CAAC,GAAG,CAAC,cAAc,CAAC,OAAO,CAAC,IAAI,EAAE,OAAO,CAAC,OAAO,CAAC,CAAC,CAAC;QACvD,OAAO;IACT,CAAC;IAED,MAAM,MAAM,GAAG,IAAI,GAAG,CAAC,OAAO,CAAC,aAAa,EAAE,GAAG,CAAC,eAAe,CAAC,CAAC;IACnE,MAAM,OAAO,GAAG,oBAAoB,CAAC,GAAG,CAAC,OAAO,EAAE;QAChD,OAAO,EAAE,OAAO,CAAC,OAAO;QACxB,YAAY,EAAE,MAAM,CAAC,IAAI;KAC1B,CAAC,CAAC;IAEH,MAAM,KAAK,GAAG,IAAI,eAAe,EAAE,CAAC;IACpC,IAAI,QAAkB,CAAC;IACvB,IAAI,CAAC;QACH,QAAQ,GAAG,MAAM,KAAK,CAAC,MAAM,EAAE;YAC7B,MAAM,EAAE,GAAG,CAAC,MAAM,IAAI,MAAM;YAC5B,OAAO;YACP,0DAA0D;YAC1D,IAAI;YACJ,MAAM,EAAE,KAAK,CAAC,MAAM;YACpB,QAAQ,EAAE,OAAO;SAClB,CAAC,CAAC;IACL,CAAC;IAAC,OAAO,GAAG,EAAE,CAAC;QACb,sEAAsE;QACtE,MAAM,IAAI,GAAG,WAAW,CAAC;QACzB,GAAG,CAAC,IAAI,CAAC,yBAAyB,EAAE,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,CAAC,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,CAAC,CAAC;QAC9E,MAAM,CAAC;YACL,MAAM,EAAE,gBAAgB;YACxB,UAAU,EAAE,kBAAkB,CAAC,IAAI,CAAC;YACpC,cAAc,EAAE,IAAI,CAAC,KAAK;YAC1B,OAAO,EAAE,IAAI,CAAC,UAAU;YACxB,SAAS,EAAE,IAAI;YACf,YAAY,EAAE,MAAM,CAAC,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC;SACxC,CAAC,CAAC;QACH,GAAG,CAAC,SAAS,CAAC,kBAAkB,CAAC,IAAI,CAAC,EAAE,EAAE,cAAc,EAAE,kBAAkB,EAAE,CAAC,CAAC;QAChF,GAAG,CAAC,GAAG,CAAC,cAAc,CAAC,IAAI,EAAE,gCAAgC,CAAC,CAAC,CAAC;QAChE,OAAO;IACT,CAAC;IAED,MAAM,SAAS,GAAG,cAAc,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC;IACnD,MAAM,iBAAiB,GACrB,QAAQ,CAAC,OAAO,CAAC,GAAG,CAAC,YAAY,CAAC,IAAI,QAAQ,CAAC,OAAO,CAAC,GAAG,CAAC,sBAAsB,CAAC,CAAC;IACrF,MAAM,UAAU,GAAG,sBAAsB,CAAC,QAAQ,CAAC,OAAO,EAAE,EAAE,MAAM,EAAE,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC;IAErF,MAAM,IAAI,GAAG;QACX,cAAc,EAAE,IAAI,CAAC,KAAK;QAC1B,WAAW,EAAE,IAAI,CAAC,KAAK;QACvB,MAAM,EAAE,IAAI,CAAC,MAAM;QACnB,UAAU,EAAE,QAAQ,CAAC,MAAM;QAC3B,OAAO,EAAE,IAAI,CAAC,UAAU;QACxB,iBAAiB;QACjB,SAAS;KACV,CAAC;IAEF,8EAA8E;IAC9E;;;;;;;;;;OAUG;IACH,IAAI,CAAC,QAAQ,CAAC,EAAE,EAAE,CAAC;QACjB,MAAM,OAAO,GAAG,uBAAuB,CAAC,QAAQ,CAAC,MAAM,EAAE,MAAM,QAAQ,CAAC,IAAI,EAAE,CAAC,CAAC;QAEhF;;;;;;WAMG;QACH,MAAM,SAAS,GAAG,wBAAwB,CAAC;YACzC,UAAU,EAAE,QAAQ,CAAC,MAAM;YAC3B,MAAM,EAAE,IAAI,CAAC,MAAM;YACnB,SAAS,EAAE,IAAI,CAAC,SAAS;YACzB,SAAS;YACT,YAAY,EAAE,cAAc;SAC7B,CAAC,CAAC;QAEH;;;;;;;;;;WAUG;QACH,IAAI,CAAC,SAAS,EAAE,CAAC;YACf,GAAG,CAAC,IAAI,CAAC,kBAAkB,EAAE;gBAC3B,EAAE;gBACF,MAAM,EAAE,QAAQ,CAAC,MAAM;gBACvB,IAAI,EAAE,OAAO,CAAC,IAAI;gBAClB,OAAO,EAAE,OAAO,CAAC,OAAO;gBACxB,KAAK,EAAE,IAAI,CAAC,KAAK;gBACjB,MAAM,EAAE,IAAI,CAAC,MAAM;gBACnB,OAAO,EAAE,IAAI,CAAC,UAAU;gBACxB,QAAQ,EAAE,aAAa;gBACvB,iBAAiB;gBACjB,SAAS;aACV,CAAC,CAAC;QACL,CAAC;QACD,4EAA4E;QAC5E,2EAA2E;QAC3E,GAAG,CAAC,SAAS,CAAC,QAAQ,CAAC,MAAM,EAAE,sBAAsB,CAAC,QAAQ,CAAC,OAAO,EAAE,EAAE,MAAM,EAAE,KAAK,EAAE,CAAC,CAAC,CAAC;QAC5F,GAAG,CAAC,GAAG,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC;QACtB,MAAM,CAAC;YACL,GAAG,IAAI;YACP,MAAM,EAAE,SAAS,CAAC,CAAC,CAAC,mBAAmB,CAAC,CAAC,CAAC,gBAAgB;YAC1D,QAAQ,EAAE,MAAM,CAAC,UAAU,CAAC,OAAO,CAAC,IAAI,CAAC;YACzC,SAAS,EAAE,OAAO,CAAC,IAAI;YACvB,YAAY,EAAE,OAAO,CAAC,OAAO;SAC9B,CAAC,CAAC;QACH,OAAO;IACT,CAAC;IAED,8EAA8E;IAC9E,IAAI,CAAC,IAAI,CAAC,MAAM,IAAI,CAAC,QAAQ,CAAC,IAAI,EAAE,CAAC;QACnC,MAAM,IAAI,GAAG,MAAM,QAAQ,CAAC,IAAI,EAAE,CAAC;QACnC,IAAI,KAAK,GAAG,WAAW,CAAC;QACxB,IAAI,CAAC;YACH,KAAK,GAAG,aAAa,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,IAAI,WAAW,CAAC;QACzD,CAAC;QAAC,MAAM,CAAC;YACP,+DAA+D;QACjE,CAAC;QACD,MAAM,MAAM,GAAG,UAAU,CAAC,IAAI,CAAC,KAAK,EAAE,KAAK,EAAE,cAAc,CAAC,CAAC;QAC7D,GAAG,CAAC,SAAS,CAAC,QAAQ,CAAC,MAAM,EAAE,UAAU,CAAC,CAAC;QAC3C,GAAG,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;QACd,MAAM,CAAC;YACL,GAAG,IAAI;YACP,MAAM,EAAE,IAAI;YACZ,KAAK;YACL,OAAO,EAAE,MAAM,CAAC,IAAI;YACpB,SAAS,EAAE,MAAM,CAAC,KAAK;YACvB,eAAe,EAAE,MAAM,CAAC,YAAY;YACpC,QAAQ,EAAE,MAAM,CAAC,UAAU,CAAC,IAAI,CAAC;SAClC,CAAC,CAAC;QACH,OAAO;IACT,CAAC;IAED,8EAA8E;IAC9E,MAAM,MAAM,GAAG,eAAe,EAAE,CAAC;IACjC,MAAM,GAAG,GAAG,sBAAsB,EAAE,CAAC;IAErC,WAAW,CAAC,GAAG,EAAE,QAAQ,CAAC,MAAM,EAAE,UAAU,CAAC,CAAC;IAE9C,IAAI,CAAC;QACH,MAAM,MAAM,GAAG,MAAM,WAAW,CAAC,QAAQ,CAAC,IAAI,EAAE,GAAG,EAAE;YACnD,KAAK;YACL,SAAS;YACT,OAAO,EAAE,CAAC,KAAK,EAAE,EAAE;gBACjB,KAAK,MAAM,KAAK,IAAI,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC;oBAAE,GAAG,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;YAC3D,CAAC;SACF,CAAC,CAAC;QACH,KAAK,MAAM,KAAK,IAAI,MAAM,CAAC,KAAK,EAAE;YAAE,GAAG,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;QAErD,MAAM,KAAK,GAAG,GAAG,CAAC,QAAQ,EAAE,CAAC;QAC7B,MAAM,MAAM,GAAG,UAAU,CAAC,IAAI,CAAC,KAAK,EAAE,KAAK,EAAE,cAAc,CAAC,CAAC;QAC7D,MAAM,SAAS,GAAG,GAAG,CAAC,WAAW,EAAE,CAAC;QAEpC;;;;;;;;;;WAUG;QACH,IAAI,MAAM,GAAkB,IAAI,CAAC;QACjC,IAAI,MAAM,CAAC,aAAa;YAAE,MAAM,GAAG,cAAc,CAAC;aAC7C,IAAI,SAAS;YAAE,MAAM,GAAG,cAAc,CAAC;QAE5C,IAAI,CAAC,MAAM,CAAC,aAAa;YAAE,GAAG,CAAC,GAAG,EAAE,CAAC;QAErC,MAAM,CAAC;YACL,GAAG,IAAI;YACP,MAAM;YACN,uEAAuE;YACvE,0EAA0E;YAC1E,OAAO,EAAE,MAAM,CAAC,aAAa,IAAI,CAAC,MAAM,CAAC,SAAS;YAClD,KAAK;YACL,OAAO,EAAE,MAAM,CAAC,IAAI;YACpB,SAAS,EAAE,MAAM,CAAC,KAAK;YACvB,eAAe,EAAE,MAAM,CAAC,YAAY;YACpC,MAAM,EAAE,MAAM,CAAC,MAAM;YACrB,QAAQ,EAAE,MAAM,CAAC,QAAQ;YACzB,GAAG,CAAC,SAAS,CAAC,CAAC,CAAC,EAAE,SAAS,EAAE,SAAS,CAAC,IAAI,EAAE,YAAY,EAAE,SAAS,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;SACrF,CAAC,CAAC;QAEH,IAAI,MAAM,CAAC,cAAc,GAAG,CAAC,EAAE,CAAC;YAC9B,GAAG,CAAC,IAAI,CAAC,sCAAsC,EAAE,EAAE,EAAE,EAAE,MAAM,EAAE,MAAM,CAAC,cAAc,EAAE,CAAC,CAAC;QAC1F,CAAC;IACH,CAAC;IAAC,OAAO,GAAG,EAAE,CAAC;QACb,0EAA0E;QAC1E,wEAAwE;QACxE,oEAAoE;QACpE,KAAK,MAAM,KAAK,IAAI,MAAM,CAAC,KAAK,EAAE;YAAE,GAAG,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;QACrD,MAAM,KAAK,GAAG,GAAG,CAAC,QAAQ,EAAE,CAAC;QAC7B,MAAM,MAAM,GAAG,UAAU,CAAC,IAAI,CAAC,KAAK,EAAE,KAAK,EAAE,cAAc,CAAC,CAAC;QAC7D,GAAG,CAAC,IAAI,CAAC,mCAAmC,EAAE,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,CAAC,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,CAAC,CAAC;QACxF,IAAI,CAAC,GAAG,CAAC,aAAa,EAAE,CAAC;YACvB,GAAG,CAAC,KAAK,CAAC,aAAa,CAAC,WAAW,EAAE,8CAA8C,CAAC,CAAC,CAAC;YACtF,GAAG,CAAC,GAAG,EAAE,CAAC;QACZ,CAAC;QACD,MAAM,CAAC;YACL,GAAG,IAAI;YACP,MAAM,EAAE,cAAc;YACtB,OAAO,EAAE,IAAI;YACb,KAAK;YACL,OAAO,EAAE,MAAM,CAAC,IAAI;YACpB,SAAS,EAAE,MAAM,CAAC,KAAK;YACvB,eAAe,EAAE,MAAM,CAAC,YAAY;YACpC,SAAS,EAAE,WAAW;YACtB,YAAY,EAAE,MAAM,CAAC,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC;SACxC,CAAC,CAAC;IACL,CAAC;AACH,CAAC;AAED,SAAS,WAAW,CAAC,GAAoB,EAAE,IAAY;IACrD,MAAM,CAAC,GAAG,GAAG,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC;IAC5B,IAAI,CAAC,KAAK,SAAS;QAAE,OAAO,IAAI,CAAC;IACjC,OAAO,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;AAC/C,CAAC"}

package/dist/server/pipeline/preflight.js

@@ -0,0 +1,66 @@
+/**
+ * Claude Code's session-start warmup ping, and Anthropic's refusal of it.
+ *
+ * On every new session Claude Code sends exactly one non-streaming request
+ * before anything else: a ~311-byte, `max_tokens: 1`, no-system, no-tools ping.
+ * On a Claude subscription bearer Anthropic refuses it. The client shrugs this
+ * off and the session proceeds normally on the streaming requests that follow,
+ * so nothing is broken — but Fest recorded it as `upstream_error`, which put a
+ * failure in the dashboard at the start of every single session and warned
+ * about it in the gateway log.
+ *
+ * Captured directly against `api.anthropic.com` with Fest out of the path
+ * (`tools/capture-server.ts`, MODE=forward), so this is Anthropic's behaviour
+ * and not something the gateway causes:
+ *
+ *     POST /v1/messages?beta=true   311 bytes
+ *       model: claude-opus-5   stream: false   max_tokens: 1
+ *       system_present: false  tool_count: 0   message_count: 1
+ *     -> 429  x-should-retry: true   NO anthropic-ratelimit-* headers
+ *        {"type":"error","error":{"type":"rate_limit_error","message":"Error"}}
+ *
+ * The same ping on a server-held org API key returns 200, and the streaming
+ * turn 1.7s later on the same subscription bearer returns 200 with the full
+ * quota header set reading `allowed` at 0.37 utilization. So it is not a quota
+ * rejection; it is a shape rejection wearing a quota status code.
+ *
+ * WHY THIS IS NOT A BLANKET "IGNORE 429s"
+ *
+ * The discriminator is the headers. A genuine rate limit arrives with the
+ * `anthropic-ratelimit-unified-*` set, which is precisely what Claude Code's
+ * own `/status` reads; this refusal arrives with none of them. A developer who
+ * really is out of quota still gets a loud, correctly-counted error — which
+ * matters, because an error rate that reads as zero during an outage is the
+ * failure mode this codebase has already been bitten by once (see
+ * `test/stream-status.test.ts`).
+ *
+ * WHAT IS DELIBERATELY NOT DONE HERE
+ *
+ * The 429 is still relayed to Claude Code unchanged. Fest reclassifies its own
+ * record of what happened; it does not edit what Anthropic said. Rewriting a
+ * provider's response into a friendlier one would make the gateway a source of
+ * fiction about the provider, which is worth more than a tidy client UI.
+ */
+/**
+ * Four conditions, all required. Each one narrows the predicate past something
+ * a real failure could look like:
+ *
+ *  - `subscription` — the refusal is specific to the OAuth bearer; the same
+ *    request on an org key succeeds, so an org key's 429 is always real.
+ *  - `!stream` — Claude Code sends exactly one non-streaming request per
+ *    session, this one. Every real turn streams.
+ *  - `maxTokens === 1` — a ping, not work. No genuine request asks for a token.
+ *  - `rateLimit === null` — a real quota 429 carries its headers.
+ *
+ * Narrow on purpose. If Anthropic changes the ping or starts answering it
+ * properly, this stops matching and the row goes back to being a plain
+ * `upstream_error` — which is the right direction to fail: visible, not silent.
+ */
+export function isWarmupPreflightRefusal(c) {
+    return (c.httpStatus === 429 &&
+        c.subscription &&
+        !c.stream &&
+        c.maxTokens === 1 &&
+        c.rateLimit === null);
+}
+//# sourceMappingURL=preflight.js.map

package/dist/server/pipeline/preflight.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"preflight.js","sourceRoot":"","sources":["../../../server/pipeline/preflight.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA0CG;AAgBH;;;;;;;;;;;;;;GAcG;AACH,MAAM,UAAU,wBAAwB,CAAC,CAAqB;IAC5D,OAAO,CACL,CAAC,CAAC,UAAU,KAAK,GAAG;QACpB,CAAC,CAAC,YAAY;QACd,CAAC,CAAC,CAAC,MAAM;QACT,CAAC,CAAC,SAAS,KAAK,CAAC;QACjB,CAAC,CAAC,SAAS,KAAK,IAAI,CACrB,CAAC;AACJ,CAAC"}