@linxiraos/pi-ai 1.0.0

package/src/auth-gateway/server.ts

@@ -0,0 +1,836 @@
+/**
+ * omp auth-gateway HTTP server.
+ *
+ * Accepts any provider-format request (OpenAI chat-completions, Anthropic
+ * messages, OpenAI Responses) and dispatches through pi-ai's `streamSimple()`
+ * — which handles credential injection, anthropic-beta headers, codex
+ * websocket transport, and all the per-provider intricacies. The gateway is
+ * pure protocol translation: foreign wire → omp Context → pi-ai stream() →
+ * omp events → foreign wire.
+ *
+ * Endpoints:
+ *   GET  /healthz                          → unauth; ok + version
+ *   GET  /v1/usage                         → aggregated provider usage (5-min per-credential cache via AuthStorage)
+ *   GET  /v1/credentials/check             → per-credential auth probe (diagnose 401s in a multi-account pool)
+ *   GET  /v1/models                        → list known models from the registry
+ *   POST /v1/chat/completions              → OpenAI chat-completions in/out
+ *   POST /v1/messages                      → Anthropic messages in/out
+ *   POST /v1/responses                     → OpenAI Responses in/out
+ */
+
+import { Effort } from "@linxiraos/pi-catalog/effort";
+import { extractHttpStatusFromError, extractRetryHint, logger } from "@linxiraos/pi-utils";
+import type { ApiKeyResolver } from "../auth-retry";
+import type { AuthStorage } from "../auth-storage";
+import * as AIError from "../error";
+import { classifyGatewayError } from "../error/gateway";
+import { isUsageLimitOutcome } from "../error/rate-limit";
+import * as anthropicMessages from "../providers/anthropic-messages-server";
+import * as openaiChat from "../providers/openai-chat-server";
+import * as openaiResponses from "../providers/openai-responses-server";
+import * as piNative from "../providers/pi-native-server";
+import { completeSimple, streamSimple } from "../stream";
+import type { Api, AssistantMessageEventStream, Context, Model, SimpleStreamOptions } from "../types";
+import { deterministicUuid } from "../utils/deterministic-id";
+import { parseBind } from "../utils/parse-bind";
+import {
+	captureRequestHeaders,
+	corsHeaders,
+	gatewayResponseHeaders,
+	isAuthorized,
+	json,
+	resolvePeer,
+	withCors,
+} from "./http";
+import type {
+	AuthGatewayServerHandle,
+	AuthGatewayServerOptions,
+	AuthGatewayFormatModule as FormatModule,
+	AuthGatewayParsedRequest as ParsedFormatRequest,
+} from "./types";
+import { DEFAULT_AUTH_GATEWAY_BIND } from "./types";
+
+// ParsedFormatRequest / ParsedFormatOptions / FormatModule come from ./types.
+
+export type ModelResolver = (modelId: string) => Model<Api> | undefined;
+
+export interface AuthGatewayBootOptions extends AuthGatewayServerOptions {
+	/** Source of credentials. Caller wires this to a broker-backed AuthStorage. */
+	storage: AuthStorage;
+	/**
+	 * Resolve a client-requested model id to a pi-ai Model. Caller supplies
+	 * this from a ModelRegistry (lives in `coding-agent` to avoid an inverse
+	 * dependency in `pi-ai`).
+	 */
+	resolveModel: ModelResolver;
+	/** Optional supplier for `/v1/models` listing. Returns the full model array. */
+	listModels?: () => Iterable<Model<Api>>;
+}
+
+// `parseBind` lives in ../utils/parse-bind so the gateway and broker can't
+// drift on accepted inputs (e.g. empty hostname, IPv6 brackets).
+
+const FORMAT_ROUTES: Record<string, { module: FormatModule; label: string }> = {
+	"/v1/chat/completions": { module: openaiChat, label: "openai-chat" },
+	"/v1/messages": { module: anthropicMessages, label: "anthropic-messages" },
+	"/v1/responses": { module: openaiResponses, label: "openai-responses" },
+};
+
+// (passthrough fast-path removed — it bypassed pi-ai provider logic, in
+// particular the Anthropic Claude-Code OAuth system-prompt prefix injection.
+// Every request now takes the translate path so credential-specific request
+// shaping always applies.)
+
+// Options the caller's wire format may carry but the resolved provider can't
+// honour are dropped silently in `buildStreamOptions`. We used to 400 here
+// (`Unsupported option: temperature for openai-codex-responses`), but every
+// realistic client (llm-git, openai SDK, anthropic SDK) bakes some of these
+// defaults in without knowing which model they'll resolve to. Failing loudly
+// just turned that into per-call config hell. Silent strip is what the
+// upstream provider would do anyway when it ignores extra fields.
+
+/**
+ * Derive a stable cache identity from the parts of the request that don't
+ * change turn-to-turn within a logical conversation: model id, system prompt,
+ * tool definitions, and the first message (the conversation seed). Codex-class
+ * backends only cache prefixes when an explicit `prompt_cache_key` is set;
+ * without one, two requests with the same prefix but different trailing
+ * messages don't coalesce. This bridges Anthropic-style clients (which signal
+ * caching via `cache_control` markers rather than an opaque key) to Codex's
+ * keyed model so cross-protocol caching "just works".
+ *
+ * Including the first message scopes the key to one logical conversation:
+ * two different chats with the same system prompt no longer share a cache
+ * bucket and can't trample each other's prefix-tree entries.
+ *
+ * Anthropic-backed requests ignore `sessionId`; the key is harmless there.
+ */
+function deriveSessionId(modelId: string, context: Context): string {
+	const parts: string[] = [modelId];
+	if (context.systemPrompt && context.systemPrompt.length > 0) {
+		parts.push(context.systemPrompt.join("\n\n"));
+	}
+	if (context.tools && context.tools.length > 0) {
+		parts.push(JSON.stringify(context.tools));
+	}
+	const first = context.messages?.[0];
+	if (first) {
+		// Strip timestamp / provider metadata so the hash is stable across turns
+		// of the same conversation (omp re-stamps every parsed Message). role +
+		// content is what's actually on the wire.
+		parts.push(JSON.stringify({ role: first.role, content: first.content }));
+	}
+	const seed = parts.join("\u0000");
+	// The 36-char UUID flows through unchanged:
+	// `normalizeOpenAIPromptCacheKey` accepts ≤64 chars verbatim.
+	return deterministicUuid(seed);
+}
+
+function buildStreamOptions(parsed: ParsedFormatRequest, api: Api, signal: AbortSignal): SimpleStreamOptions {
+	const opts: SimpleStreamOptions = { signal };
+	const { options } = parsed;
+	// Codex backend rejects every sampling control with
+	// `Unsupported parameter: …` (#3117). Strip the full set for that one
+	// provider; everything else is harmless to forward — `streamSimple` ignores
+	// what the underlying provider doesn't honour.
+	const isCodex = api === "openai-codex-responses";
+	if (options.maxOutputTokens !== undefined) opts.maxTokens = options.maxOutputTokens;
+	if (options.temperature !== undefined && !isCodex) opts.temperature = options.temperature;
+	if (options.topP !== undefined && !isCodex) opts.topP = options.topP;
+	if (options.topK !== undefined && !isCodex) opts.topK = options.topK;
+	if (options.minP !== undefined && !isCodex) opts.minP = options.minP;
+	if (options.stopSequences !== undefined && !isCodex) opts.stopSequences = options.stopSequences;
+	if (options.presencePenalty !== undefined && !isCodex) opts.presencePenalty = options.presencePenalty;
+	if (options.frequencyPenalty !== undefined && !isCodex) opts.frequencyPenalty = options.frequencyPenalty;
+	if (options.repetitionPenalty !== undefined && !isCodex) opts.repetitionPenalty = options.repetitionPenalty;
+	if (options.metadata !== undefined) opts.metadata = options.metadata;
+	if (options.headers !== undefined) opts.headers = { ...(opts.headers ?? {}), ...options.headers };
+	if (options.toolChoice !== undefined) {
+		opts.toolChoice =
+			typeof options.toolChoice !== "object"
+				? options.toolChoice
+				: "type" in options.toolChoice
+					? options.toolChoice
+					: { type: "tool", name: options.toolChoice.name };
+	}
+	if (options.reasoning !== undefined) opts.reasoning = options.reasoning;
+	if (options.disableReasoning !== undefined) opts.disableReasoning = options.disableReasoning;
+	if (options.hideThinkingSummary !== undefined) opts.hideThinkingSummary = options.hideThinkingSummary;
+	if (options.taskBudget !== undefined) opts.taskBudget = options.taskBudget;
+	if (options.serviceTier !== undefined) opts.serviceTier = options.serviceTier;
+	if (options.cacheRetention !== undefined) opts.cacheRetention = options.cacheRetention;
+	if (options.include !== undefined) opts.include = options.include;
+	// Client-supplied `prompt_cache_key` wins; otherwise derive a stable
+	// key from the model + system + tools so prefix caching engages on
+	// Codex-class backends across turns of the same logical conversation.
+	const promptCacheKey = options.promptCacheKey ?? deriveSessionId(parsed.modelId, parsed.context);
+	opts.promptCacheKey = promptCacheKey;
+	opts.sessionId = promptCacheKey;
+	if (options.thinkingBudgets) {
+		opts.thinkingBudgets = { ...(opts.thinkingBudgets ?? {}), ...options.thinkingBudgets };
+	}
+	if (options.explicitThinkingBudgetTokens !== undefined) {
+		// Mirror Rust's `resolve_thinking_budget`: explicit budget pins onto
+		// whichever effort the client requested (or High when unspecified) and
+		// ALSO sets the effort so providers that gate on `reasoning` actually
+		// surface the budget.
+		const effort = options.reasoning ?? Effort.High;
+		opts.thinkingBudgets = {
+			...(opts.thinkingBudgets ?? {}),
+			[effort]: options.explicitThinkingBudgetTokens,
+		};
+		opts.reasoning ??= effort;
+	}
+	// Fields that don't yet have a matching pi-ai `SimpleStreamOptions` slot.
+	// Surfaced once in debug logs so they show up when wiring a new provider,
+	// but NEVER widened into `options.extra` — every consumer would have to
+	// re-implement the typed parse to read them back out.
+	// TODO(pi-ai): land first-class fields and replace these blocks.
+	if (
+		options.parallelToolCalls !== undefined ||
+		options.previousResponseId !== undefined ||
+		options.seed !== undefined ||
+		options.logitBias !== undefined ||
+		options.user !== undefined ||
+		options.responseFormat !== undefined
+	) {
+		logger.debug("auth-gateway dropped unsupported typed options", {
+			api,
+			parallelToolCalls: options.parallelToolCalls,
+			previousResponseId: options.previousResponseId,
+			seed: options.seed,
+			hasLogitBias: options.logitBias !== undefined,
+			user: options.user,
+			hasResponseFormat: options.responseFormat !== undefined,
+		});
+	}
+	return opts;
+}
+
+/**
+ * Hook fired by {@link streamSimple} when the upstream request fails in a
+ * way that's rotatable — today that's HTTP 401 (credential is bad) and
+ * usage-limit phrasing matched by {@link isUsageLimitError} (Codex's
+ * `usage_limit_reached`, Anthropic's `usage_limit_reached`, Google's
+ * `resource_exhausted`, …). The two cases need different storage actions:
+ *
+ * - **usage-limit** → {@link AuthStorage.markUsageLimitReached}. Marks just
+ *   the current session's credential as temporarily blocked (honouring
+ *   `retry-after` / `resets_at` hints when present) and returns `true` only
+ *   when a sibling credential is still available. Burning the credential
+ *   with `invalidateCredentialMatching` here would orphan accounts whose
+ *   reset window is several hours away — exactly the bug this helper exists
+ *   to avoid.
+ * - **auth-failure** → {@link AuthStorage.invalidateCredentialMatching}.
+ *   Suspect/delete the row so it doesn't get re-picked next request.
+ *
+ * In both branches we return the next `getApiKey` result (sticky on the
+ * same `sessionId`) so streamSimple can transparently retry the pre-emit
+ * failure with a fresh credential. Returning `undefined` aborts the retry
+ * and surfaces the original error to the caller.
+ */
+async function refreshGatewayApiKeyAfterAuthError(
+	storage: AuthStorage,
+	model: Model<Api>,
+	sessionId: string,
+	provider: string,
+	oldKey: string,
+	error: unknown,
+	signal: AbortSignal,
+	format: string,
+	peer: string,
+): Promise<string | undefined> {
+	const message = error instanceof Error ? error.message : String(error);
+	const status = extractHttpStatusFromError(error);
+	if (AIError.isUsageLimit(error) || isUsageLimitOutcome(status, message)) {
+		const retryAfterMs = extractRetryHint(undefined, message);
+		const { switched, retryAtMs } = await storage.markUsageLimitReached(provider, sessionId, {
+			retryAfterMs,
+			baseUrl: model.baseUrl,
+			modelId: model.id,
+			apiKey: oldKey,
+			signal,
+		});
+		logger.debug("auth-gateway retrying provider request after usage-limit block", {
+			format,
+			provider,
+			peer,
+			switched,
+			retryAfterMs,
+			retryAtMs,
+			error: message,
+		});
+		if (!switched) return undefined;
+		return storage.getApiKey(provider, sessionId, { modelId: model.id, signal });
+	}
+	await storage.invalidateCredentialMatching(provider, oldKey, { sessionId, signal });
+	logger.debug("auth-gateway retrying provider request after credential invalidation", {
+		format,
+		provider,
+		peer,
+		error: message,
+	});
+	return storage.getApiKey(provider, sessionId, { modelId: model.id, signal });
+}
+
+/**
+ * Build the {@link ApiKeyResolver} handed to `streamSimple` for a gateway
+ * request. Drives the central a/b/c auth-retry policy server-side:
+ *
+ * - initial resolve → the credential already resolved for this request.
+ * - step (b) `!lastChance` → force-refresh the SAME session-sticky credential
+ *   (a peer/broker may have rotated its token out from under our cached copy).
+ * - step (c) `lastChance` → {@link refreshGatewayApiKeyAfterAuthError} switches
+ *   to a sibling (usage-limit block vs credential invalidation by error class).
+ *
+ * `lastKey` tracks the most recent bearer so the switch step invalidates the
+ * credential that actually failed.
+ */
+function buildGatewayApiKeyResolver(
+	storage: AuthStorage,
+	model: Model<Api>,
+	sessionId: string,
+	initialKey: string,
+	requestSignal: AbortSignal,
+	format: string,
+	peer: string,
+): ApiKeyResolver {
+	let lastKey = initialKey;
+	return async ({ lastChance, error, signal }) => {
+		const sig = signal ?? requestSignal;
+		if (error === undefined) {
+			lastKey = initialKey;
+			return initialKey;
+		}
+		if (!lastChance) {
+			const refreshed = await storage.getApiKey(model.provider, sessionId, {
+				modelId: model.id,
+				signal: sig,
+				forceRefresh: true,
+			});
+			lastKey = refreshed ?? lastKey;
+			return refreshed;
+		}
+		const next = await refreshGatewayApiKeyAfterAuthError(
+			storage,
+			model,
+			sessionId,
+			model.provider,
+			lastKey,
+			error,
+			sig,
+			format,
+			peer,
+		);
+		lastKey = next ?? lastKey;
+		return next;
+	};
+}
+
+function clientClosedResponse(route: { module: FormatModule }): Response {
+	return route.module.formatError(499, "request_aborted", "client closed request");
+}
+
+function mirrorRequestAbort(req: Request): AbortController {
+	const controller = new AbortController();
+	if (req.signal.aborted) {
+		controller.abort(req.signal.reason);
+	} else {
+		req.signal.addEventListener("abort", () => controller.abort(req.signal.reason), { once: true });
+	}
+	return controller;
+}
+
+// (handlePassthrough removed — see note above.)
+
+async function handleFormatEndpoint(
+	route: { module: FormatModule; label: string },
+	bootOpts: AuthGatewayBootOptions,
+	req: Request,
+	peer: string,
+): Promise<Response> {
+	const startedAt = performance.now();
+	const requestId = crypto.randomUUID();
+	const controller = mirrorRequestAbort(req);
+	if (controller.signal.aborted) return clientClosedResponse(route);
+
+	let body: unknown;
+	try {
+		body = await req.json();
+	} catch (error) {
+		if (controller.signal.aborted) return clientClosedResponse(route);
+		return route.module.formatError(400, "invalid_request_error", `Invalid JSON body: ${String(error)}`);
+	}
+	if (controller.signal.aborted) return clientClosedResponse(route);
+
+	// All three supported wire formats put the model id on a top-level `model`
+	// field. Read it without running the full strict schema so the route can
+	// produce a coherent error envelope when the model id is missing.
+	const modelId =
+		typeof body === "object" && body !== null && typeof (body as { model?: unknown }).model === "string"
+			? (body as { model: string }).model
+			: undefined;
+	if (!modelId) {
+		return route.module.formatError(400, "invalid_request_error", "Missing top-level `model` field");
+	}
+
+	const model = bootOpts.resolveModel(modelId);
+	if (!model) {
+		return route.module.formatError(404, "invalid_request_error", `Unknown model: ${modelId}`);
+	}
+
+	// Parse the wire-format request BEFORE resolving the credential so we
+	// have a stable per-conversation `sessionId` to thread into AuthStorage.
+	// Sticky-credential tracking and `markUsageLimitReached` both key off
+	// this id; without it `getApiKey` would re-roundrobin every request
+	// and `markUsageLimitReached` would no-op (it can only mark the
+	// credential it last handed out to that session).
+	let parsed: ParsedFormatRequest;
+	try {
+		parsed = route.module.parseRequest(body, req.headers);
+	} catch (error) {
+		if (controller.signal.aborted) return clientClosedResponse(route);
+		const message = error instanceof Error ? error.message : String(error);
+		return route.module.formatError(400, "invalid_request_error", message);
+	}
+	// Merge gateway-captured passthrough headers under the parser's own
+	// captures. Parsers that set `options.headers` themselves win (they may
+	// have stripped or normalized values); the gateway's allow-list fills in
+	// anything they didn't touch.
+	{
+		const captured = captureRequestHeaders(req.headers);
+		parsed.options.headers = { ...captured, ...(parsed.options.headers ?? {}) };
+	}
+	if (controller.signal.aborted) return clientClosedResponse(route);
+
+	// Sticky credential id: honour the client's `prompt_cache_key` when
+	// supplied (so external session ids align), otherwise derive from
+	// modelId + system + tools + first message. Mirrored into
+	// streamOpts.sessionId / promptCacheKey by `buildStreamOptions`.
+	const sessionId = parsed.options.promptCacheKey ?? deriveSessionId(parsed.modelId, parsed.context);
+	parsed.options.promptCacheKey ??= sessionId;
+
+	// pi-ai's stream() does NOT consult AuthStorage — the caller (us) is
+	// expected to resolve the credential and pass it as `options.apiKey`.
+	// For OAuth providers this returns the access token (refreshed via the
+	// broker override on AuthStorage when needed).
+	let apiKey: string | undefined;
+	try {
+		apiKey = await bootOpts.storage.getApiKey(model.provider, sessionId, {
+			modelId: model.id,
+			signal: controller.signal,
+		});
+	} catch (error) {
+		if (controller.signal.aborted) return clientClosedResponse(route);
+		const classified = classifyGatewayError(error);
+		logger.warn("auth-gateway getApiKey threw", { provider: model.provider, peer, error: classified.message });
+		return route.module.formatError(classified.status, classified.type, classified.message);
+	}
+	if (controller.signal.aborted) return clientClosedResponse(route);
+	if (!apiKey) {
+		return route.module.formatError(
+			401,
+			"authentication_error",
+			`No credential available for provider ${model.provider}`,
+		);
+	}
+
+	const streamOpts = buildStreamOptions(parsed, model.api, controller.signal);
+	streamOpts.apiKey = buildGatewayApiKeyResolver(
+		bootOpts.storage,
+		model,
+		sessionId,
+		apiKey,
+		controller.signal,
+		route.label,
+		peer,
+	);
+
+	logger.info("auth-gateway request", {
+		requestId,
+		format: route.label,
+		model: parsed.modelId,
+		resolvedProvider: model.provider,
+		resolvedModel: model.id,
+		stream: parsed.stream,
+		peer,
+	});
+
+	if (!parsed.stream) {
+		try {
+			if (controller.signal.aborted) return clientClosedResponse(route);
+			const message = await completeSimple(model, parsed.context, streamOpts);
+			if (message.stopReason === "aborted" || message.stopReason === "error") {
+				const errorMessage =
+					message.errorMessage ??
+					(message.stopReason === "aborted" ? "Request was aborted" : "Upstream request failed");
+				logger.warn("auth-gateway non-streaming failed", {
+					format: route.label,
+					reason: message.stopReason,
+					error: errorMessage,
+					peer,
+				});
+				if (message.stopReason === "aborted") {
+					return route.module.formatError(499, "request_aborted", errorMessage);
+				}
+				const classified = classifyGatewayError(errorMessage);
+				return route.module.formatError(classified.status, classified.type, errorMessage);
+			}
+			return json(
+				200,
+				route.module.encodeResponse(message, parsed.modelId),
+				gatewayResponseHeaders(model, { requestId, message, startedAt }),
+			);
+		} catch (error) {
+			if (controller.signal.aborted) return clientClosedResponse(route);
+			const classified = classifyGatewayError(error);
+			logger.warn("auth-gateway non-streaming aborted", {
+				format: route.label,
+				error: classified.message,
+				peer,
+			});
+			return route.module.formatError(classified.status, classified.type, classified.message);
+		}
+	}
+
+	let events: AssistantMessageEventStream;
+	try {
+		if (controller.signal.aborted) return clientClosedResponse(route);
+		events = streamSimple(model, parsed.context, streamOpts);
+	} catch (error) {
+		const classified = classifyGatewayError(error);
+		logger.warn("auth-gateway streamSimple threw", { format: route.label, error: classified.message, peer });
+		return route.module.formatError(classified.status, classified.type, classified.message);
+	}
+	if (controller.signal.aborted) return clientClosedResponse(route);
+
+	const sseStream = route.module.encodeStream(events, parsed.modelId, parsed.options, {
+		signal: controller.signal,
+		onCancel: reason => {
+			if (!controller.signal.aborted) {
+				controller.abort(reason instanceof Error ? reason : new Error("client closed request"));
+			}
+		},
+	});
+	return new Response(sseStream, {
+		status: 200,
+		headers: {
+			...gatewayResponseHeaders(model, { requestId }),
+			"Content-Type": "text/event-stream; charset=utf-8",
+			"Cache-Control": "no-cache",
+			Connection: "keep-alive",
+			// Disable proxy buffering (nginx and ingress controllers honor this).
+			// Without it the SSE stream gets held until the buffer flushes, which
+			// stalls the long-thinking-budget calls we exist to support.
+			"X-Accel-Buffering": "no",
+		},
+	});
+}
+
+/**
+ * Pi-native fast path: `POST /v1/pi/stream`. Accepts the canonical pi-ai
+ * `Context` directly (no wire-format round-trip) and emits a bandwidth-shrunk
+ * event stream matching `pi-agent`'s `streamProxy`. Skips the OpenAI /
+ * Anthropic / Responses translation layers — those exist to bridge foreign
+ * SDKs (llm-git, anthropic-sdk, openai-sdk), and bridging back to pi-native
+ * just to bridge forward again is wasted work.
+ *
+ * Every other gateway concern (bearer auth, model resolve, credential fetch,
+ * abort mirroring, codex temperature/topP strip, prefix-cache key derivation,
+ * Claude-Code OAuth shaping inside `streamSimple`) still applies — only
+ * `parseRequest`/`encodeResponse`/`encodeStream` differ from the format-endpoint
+ * path.
+ */
+async function handlePiNative(bootOpts: AuthGatewayBootOptions, req: Request, peer: string): Promise<Response> {
+	const startedAt = performance.now();
+	const requestId = crypto.randomUUID();
+	const controller = mirrorRequestAbort(req);
+	const aborted = (): Response => piNative.formatError(499, "request_aborted", "client closed request");
+	if (controller.signal.aborted) return aborted();
+
+	let body: unknown;
+	try {
+		body = await req.json();
+	} catch (error) {
+		if (controller.signal.aborted) return aborted();
+		return piNative.formatError(400, "invalid_request_error", `Invalid JSON body: ${String(error)}`);
+	}
+	if (controller.signal.aborted) return aborted();
+
+	let parsed: piNative.PiNativeParsedRequest;
+	try {
+		parsed = piNative.parseRequest(body, req.headers);
+	} catch (error) {
+		if (controller.signal.aborted) return aborted();
+		const message = error instanceof Error ? error.message : String(error);
+		return piNative.formatError(400, "invalid_request_error", message);
+	}
+
+	const model = bootOpts.resolveModel(parsed.modelId);
+	if (!model) {
+		return piNative.formatError(404, "invalid_request_error", `Unknown model: ${parsed.modelId}`);
+	}
+	// Pi-native already parsed `streamOpts.sessionId` (when set by the
+	// client); fall back to the derived key so credential-stickiness lines
+	// up with cache-prefix stickiness — same identity used for both means
+	// the next turn of this conversation reuses the same credential until
+	// it hits a usage cap, then markUsageLimitReached can hand off.
+	const sessionId = parsed.options.sessionId ?? deriveSessionId(parsed.modelId, parsed.context);
+	parsed.options.sessionId ??= sessionId;
+
+	let apiKey: string | undefined;
+	try {
+		apiKey = await bootOpts.storage.getApiKey(model.provider, sessionId, {
+			modelId: model.id,
+			signal: controller.signal,
+		});
+	} catch (error) {
+		if (controller.signal.aborted) return aborted();
+		const classified = classifyGatewayError(error);
+		logger.warn("auth-gateway getApiKey threw", { provider: model.provider, peer, error: classified.message });
+		return piNative.formatError(classified.status, classified.type, classified.message);
+	}
+	if (controller.signal.aborted) return aborted();
+	if (!apiKey) {
+		return piNative.formatError(
+			401,
+			"authentication_error",
+			`No credential available for provider ${model.provider}`,
+		);
+	}
+
+	// Build the SimpleStreamOptions actually handed to `streamSimple`. We
+	// trust the client's options (already allow-listed by `parseRequest`) and
+	// only inject server-controlled fields. The codex sampling strip mirrors
+	// `buildStreamOptions` — Codex rejects every one with a 400 (#3117).
+	const streamOpts: SimpleStreamOptions = { ...parsed.options, apiKey, signal: controller.signal };
+	streamOpts.apiKey = buildGatewayApiKeyResolver(
+		bootOpts.storage,
+		model,
+		sessionId,
+		apiKey,
+		controller.signal,
+		"pi-native",
+		peer,
+	);
+	if (model.api === "openai-codex-responses") {
+		delete streamOpts.temperature;
+		delete streamOpts.topP;
+		delete streamOpts.topK;
+		delete streamOpts.minP;
+		delete streamOpts.stopSequences;
+		delete streamOpts.presencePenalty;
+		delete streamOpts.frequencyPenalty;
+		delete streamOpts.repetitionPenalty;
+	}
+	// Merge gateway-captured passthrough headers under the client's own
+	// headers — the client's values win when they collide.
+	const captured = captureRequestHeaders(req.headers);
+	streamOpts.headers = { ...captured, ...(streamOpts.headers ?? {}) };
+	streamOpts.sessionId ??= sessionId;
+
+	logger.info("auth-gateway request", {
+		requestId,
+		format: "pi-native",
+		model: parsed.modelId,
+		resolvedProvider: model.provider,
+		resolvedModel: model.id,
+		stream: parsed.stream,
+		peer,
+	});
+
+	if (!parsed.stream) {
+		try {
+			if (controller.signal.aborted) return aborted();
+			const message = await completeSimple(model, parsed.context, streamOpts);
+			if (message.stopReason === "aborted" || message.stopReason === "error") {
+				const errorMessage =
+					message.errorMessage ??
+					(message.stopReason === "aborted" ? "Request was aborted" : "Upstream request failed");
+				logger.warn("auth-gateway non-streaming failed", {
+					format: "pi-native",
+					reason: message.stopReason,
+					error: errorMessage,
+					peer,
+				});
+				if (message.stopReason === "aborted") {
+					return piNative.formatError(499, "request_aborted", errorMessage);
+				}
+				const classified = classifyGatewayError(errorMessage);
+				return piNative.formatError(classified.status, classified.type, errorMessage);
+			}
+			return json(200, { message }, gatewayResponseHeaders(model, { requestId, message, startedAt }));
+		} catch (error) {
+			if (controller.signal.aborted) return aborted();
+			const classified = classifyGatewayError(error);
+			logger.warn("auth-gateway non-streaming aborted", { format: "pi-native", error: classified.message, peer });
+			return piNative.formatError(classified.status, classified.type, classified.message);
+		}
+	}
+
+	let events: AssistantMessageEventStream;
+	try {
+		if (controller.signal.aborted) return aborted();
+		events = streamSimple(model, parsed.context, streamOpts);
+	} catch (error) {
+		const classified = classifyGatewayError(error);
+		logger.warn("auth-gateway streamSimple threw", { format: "pi-native", error: classified.message, peer });
+		return piNative.formatError(classified.status, classified.type, classified.message);
+	}
+	if (controller.signal.aborted) return aborted();
+
+	const sseStream = piNative.encodeStream(events, parsed.modelId, parsed.options, {
+		signal: controller.signal,
+		onCancel: reason => {
+			if (!controller.signal.aborted) {
+				controller.abort(reason instanceof Error ? reason : new Error("client closed request"));
+			}
+		},
+	});
+	return new Response(sseStream, {
+		status: 200,
+		headers: {
+			...gatewayResponseHeaders(model, { requestId }),
+			"Content-Type": "text/event-stream; charset=utf-8",
+			"Cache-Control": "no-cache",
+			Connection: "keep-alive",
+			"X-Accel-Buffering": "no",
+		},
+	});
+}
+
+/**
+ * Snapshot of `GET /v1/usage` — `fetchUsageReports` already caches reports at
+ * a 5-minute per-credential TTL (with jitter, plus last-good fallback on
+ * failure) inside `AuthStorage`, so this handler is a thin wrapper that
+ * surfaces the same data to HTTP callers (notably the macOS usage widget).
+ */
+async function handleUsage(storage: AuthStorage, signal: AbortSignal): Promise<Response> {
+	const reports = (await storage.fetchUsageReports?.({ signal })) ?? [];
+	// Drop the heavy provider-specific `raw` payload — UI consumers only need
+	// `limits` + `metadata`. Match the broker's `/v1/usage` shape so a single
+	// client struct (Swift widget, llm-git, ...) works against either endpoint.
+	const trimmed = reports.map(({ raw: _raw, ...rest }) => rest);
+	return json(200, { generatedAt: Date.now(), reports: trimmed });
+}
+
+/**
+ * Per-credential health probe surfaced on `GET /v1/credentials/check`. Tells
+ * the caller exactly which row in their broker is producing 401s — the
+ * aggregate `/v1/usage` endpoint silently drops failed credentials, which is
+ * the wrong shape when you're diagnosing auth.
+ *
+ * The probe is sequential (one credential at a time) to avoid synchronized
+ * N-account fan-out tripping per-IP rate limits on provider `/usage`
+ * endpoints. For multi-account pools that's the difference between getting
+ * a clean diagnosis and getting a 429 storm.
+ */
+async function handleCredentialsCheck(storage: AuthStorage, signal: AbortSignal): Promise<Response> {
+	const credentials = await storage.checkCredentials({ signal });
+	return json(200, { generatedAt: Date.now(), credentials });
+}
+
+function handleModelsList(opts: AuthGatewayBootOptions): Response {
+	const seen = new Set<string>();
+	const data: Array<{ id: string; object: "model"; owned_by: string; api: Api }> = [];
+	for (const model of opts.listModels?.() ?? []) {
+		const id = `${model.provider}/${model.id}`;
+		if (seen.has(id)) continue;
+		seen.add(id);
+		data.push({
+			id,
+			object: "model",
+			owned_by: model.provider,
+			api: model.api,
+		});
+	}
+	return json(200, { object: "list", data });
+}
+
+export function startAuthGateway(opts: AuthGatewayBootOptions): AuthGatewayServerHandle {
+	const bind = parseBind(opts.bind ?? DEFAULT_AUTH_GATEWAY_BIND);
+	const tokens = new Set<string>(opts.bearerTokens);
+	const version = opts.version;
+
+	const server = Bun.serve({
+		hostname: bind.hostname,
+		port: bind.port,
+		fetch: async (req): Promise<Response> => {
+			const url = new URL(req.url);
+			const pathname = url.pathname;
+			const peer = resolvePeer(req);
+			// CORS preflight is always answered without auth — browsers send
+			// preflights pre-authentication and a 401 here breaks the actual
+			// request before the bearer is ever attached.
+			if (req.method === "OPTIONS") {
+				return new Response(null, { status: 204, headers: corsHeaders(req) });
+			}
+			try {
+				if (req.method === "GET" && pathname === "/healthz") {
+					return withCors(json(200, { ok: true, version }), req);
+				}
+				if (!isAuthorized(req, tokens)) {
+					logger.info("auth-gateway request unauthorized", { method: req.method, path: pathname, peer });
+					return withCors(json(401, { error: "unauthorized" }), req);
+				}
+
+				// Aggregated usage — backed by AuthStorage's 5-min per-credential cache.
+				// Same shape as the broker's `/v1/usage`, so widget/llm-git speak to either with the
+				// same client struct.
+				if (req.method === "GET" && pathname === "/v1/usage") {
+					return withCors(await handleUsage(opts.storage, req.signal), req);
+				}
+
+				// Per-credential auth probe — diagnoses which row in a multi-account
+				// pool is producing 401s. Aggregated `/v1/usage` silently drops failed
+				// credentials, so we need a separate endpoint that captures errors.
+				if (req.method === "GET" && pathname === "/v1/credentials/check") {
+					return withCors(await handleCredentialsCheck(opts.storage, req.signal), req);
+				}
+
+				// Provider-format dispatch.
+				const formatRoute = FORMAT_ROUTES[pathname];
+				if (formatRoute && req.method === "POST") {
+					return withCors(await handleFormatEndpoint(formatRoute, opts, req, peer), req);
+				}
+
+				// Pi-native fast path. Same auth + provider plumbing as the
+				// foreign-wire routes, just without the wire-format translation.
+				if (req.method === "POST" && pathname === "/v1/pi/stream") {
+					return withCors(await handlePiNative(opts, req, peer), req);
+				}
+
+				// Model catalog.
+				if (req.method === "GET" && pathname === "/v1/models") {
+					return withCors(handleModelsList(opts), req);
+				}
+
+				// Route-table miss: no format module to defer to, so we emit a
+				// plain JSON 404 rather than guessing at a protocol-specific envelope.
+				return withCors(json(404, { error: `No route: ${req.method} ${pathname}` }), req);
+			} catch (error) {
+				logger.error("auth-gateway handler crashed", {
+					method: req.method,
+					path: pathname,
+					peer,
+					error: String(error),
+				});
+				return withCors(json(500, { error: "internal error" }), req);
+			}
+		},
+		// Max-out Bun's idle timeout. Long thinking-budget calls can sit idle
+		// for minutes before the first token arrives; the default kills them.
+		idleTimeout: 255,
+	});
+
+	const boundHost = server.hostname ?? bind.hostname;
+	const boundPort = server.port ?? bind.port;
+	return {
+		url: `http://${boundHost}:${boundPort}`,
+		port: boundPort,
+		hostname: boundHost,
+		close: async () => {
+			server.stop(true);
+		},
+	};
+}