feat: proxy subscription clients through gateway

2026-07-18 18:39:15 +02:00 · 2026-07-18 18:39:15 +02:00 · b8543ec191
commit b8543ec191
parent f720096d3b
15 changed files with 998 additions and 23 deletions
--- a/packages/gateway/public/dashboard-v2.html
+++ b/packages/gateway/public/dashboard-v2.html
@ -993,6 +993,7 @@
    .req-status.approved { color: var(--ok); border-color: rgba(21,128,61,0.24); }
    .req-status.error, .req-status.rejected { color: var(--err); border-color: rgba(180,35,24,0.24); }
    .req-status.warning, .req-status.pending_review { color: var(--warn); border-color: rgba(180,83,9,0.24); }
+    .req-status.bypassed { color: var(--err); border-color: rgba(180,35,24,0.32); }

    .client-grid {
      display: grid;
@ -1921,7 +1922,7 @@
          <div title="${req.request_id}">${req.request_id.substring(0, 14)}…</div>
          <div>${escapeHtml(req.caller)}</div>
          <div title="${req.model}">${req.model}</div>
-          <div><span class="req-status ${req.status}">${req.status}</span></div>
+          <div><span class="req-status ${formatRequestStatus(req)}">${formatRequestStatus(req)}</span></div>
          <div>${formatNumber(req.compression_tokens_before ?? req.tokens_in ?? 0)}</div>
          <div>${formatNumber(req.compression_tokens_after ?? req.tokens_in ?? 0)}</div>
          <div>${formatSavedTokens(req.compression_tokens_saved ?? 0)}</div>
@ -1942,6 +1943,7 @@
    }

    function formatCompression(req) {
+      if (String(req.compression_mode || '').startsWith('usage-import:')) return 'bypassed';
      const mode = String(req.compression_mode || 'none:none').split(':').pop() || 'none';
      const pct = Number(req.compression_savings_pct || 0);
      if (!req.compression_mode) return 'not tracked';
@ -1949,6 +1951,10 @@
      return `${escapeHtml(mode)} · ${pct.toFixed(1)}%`;
    }

+    function formatRequestStatus(req) {
+      return String(req.request_id || '').startsWith('usage-import:') ? 'bypassed' : escapeHtml(req.status);
+    }
+
    function escapeHtml(s) {
      return String(s ?? '').replace(/[&<>"']/g, c => ({ '&':'&amp;', '<':'&lt;', '>':'&gt;', '"':'&quot;', "'":'&#39;' }[c]));
    }
--- a/packages/gateway/public/dashboard.html
+++ b/packages/gateway/public/dashboard.html
@ -1047,6 +1047,7 @@
    .req-status.approved { color: var(--ok); border-color: rgba(21,128,61,0.24); }
    .req-status.error, .req-status.rejected { color: var(--err); border-color: rgba(180,35,24,0.24); }
    .req-status.warning, .req-status.pending_review { color: var(--warn); border-color: rgba(180,83,9,0.24); }
+    .req-status.bypassed { color: var(--err); border-color: rgba(180,35,24,0.32); }

    .client-grid {
      display: grid;
@ -2228,7 +2229,7 @@
          <div title="${req.request_id}">${req.request_id.substring(0, 14)}…</div>
          <div>${escapeHtml(req.caller)}</div>
          <div title="${req.model}">${req.model}</div>
-          <div><span class="req-status ${req.status}">${req.status}</span></div>
+          <div><span class="req-status ${formatRequestStatus(req)}">${formatRequestStatus(req)}</span></div>
          <div>${formatNumber(req.compression_tokens_before ?? req.tokens_in ?? 0)}</div>
          <div>${formatNumber(req.compression_tokens_after ?? req.tokens_in ?? 0)}</div>
          <div>${formatSavedTokens(req.compression_tokens_saved ?? 0)}</div>
@ -2249,6 +2250,7 @@
    }

    function formatCompression(req) {
+      if (String(req.compression_mode || '').startsWith('usage-import:')) return 'bypassed';
      const mode = String(req.compression_mode || 'none:none').split(':').pop() || 'none';
      const pct = Number(req.compression_savings_pct || 0);
      if (!req.compression_mode) return 'not tracked';
@ -2256,6 +2258,10 @@
      return `${escapeHtml(mode)} · ${pct.toFixed(1)}%`;
    }

+    function formatRequestStatus(req) {
+      return String(req.request_id || '').startsWith('usage-import:') ? 'bypassed' : escapeHtml(req.status);
+    }
+
    function escapeHtml(s) {
      return String(s ?? '').replace(/[&<>"']/g, c => ({ '&':'&amp;', '<':'&lt;', '>':'&gt;', '"':'&quot;', "'":'&#39;' }[c]));
    }
--- a/packages/gateway/src/modules/tests/transparent-subscription-proxy.test.ts
+++ b/packages/gateway/src/modules/tests/transparent-subscription-proxy.test.ts
@ -0,0 +1,37 @@
+import { afterEach, describe, expect, it } from 'vitest';
+import { transformSubscriptionPayload } from '../transparent-subscription-proxy.js';
+
+afterEach(() => {
+  delete process.env['REDACT_PII_MODE'];
+});
+
+describe('transparent subscription payload protection', () => {
+  it('redacts identifiers and compresses long text without changing protocol fields', () => {
+    process.env['REDACT_PII_MODE'] = 'always';
+    const email = ['person', 'example.org'].join('@');
+    const payload = {
+      model: 'subscription-test-model',
+      input: [{
+        type: 'message',
+        role: 'user',
+        content: [{ type: 'input_text', text: `${email}\n${'context line\n'.repeat(6000)}` }],
+      }],
+      tools: [{ type: 'function', name: 'read_file' }],
+    };
+
+    const result = transformSubscriptionPayload(payload, {
+      providerName: 'codex-subscription',
+      caller: 'unit-client',
+    });
+    const transformed = result.payload as typeof payload;
+    const text = transformed.input[0]?.content[0]?.text ?? '';
+
+    expect(result.compressionApplied).toBe(true);
+    expect(result.tokensAfter).toBeLessThan(result.tokensBefore);
+    expect(result.piiRedactions).toBe(1);
+    expect(text).not.toContain(email);
+    expect(text).toContain('<EMAIL_001>');
+    expect(text).toContain('ctxlean omitted');
+    expect(transformed.tools).toEqual(payload.tools);
+  });
+});
--- a/packages/gateway/src/modules/request-logger.ts
+++ b/packages/gateway/src/modules/request-logger.ts
@ -143,8 +143,7 @@ export class RequestLogger {
      LEFT JOIN LATERAL (
        SELECT mode, tokens_before, tokens_after, savings_pct
        FROM tokenvault_metrics
-        WHERE tool_used = 'gateway'
-          AND file_path = rt.request_id
+        WHERE file_path = rt.request_id
        ORDER BY created_at DESC
        LIMIT 1
      ) tv ON true
--- a/packages/gateway/src/modules/transparent-subscription-proxy.ts
+++ b/packages/gateway/src/modules/transparent-subscription-proxy.ts
@ -0,0 +1,260 @@
+import type { FastifyReply, FastifyRequest } from 'fastify';
+import { Readable } from 'stream';
+import { compressContext } from './context-compressor.js';
+import { redactPii, getRedactMode, shouldRedactFor } from './pii-redaction.js';
+import { createRequestLogger } from './request-logger.js';
+import { getPool } from '../db/client.js';
+import { logger } from '../observability/logger.js';
+import { logCompressionMetric } from '../utils/tokenvault-hooks.js';
+
+export type TransparentSubscriptionTarget = 'anthropic' | 'codex';
+
+interface TransformStats {
+  payload: unknown;
+  tokensBefore: number;
+  tokensAfter: number;
+  compressionApplied: boolean;
+  piiRedactions: number;
+}
+
+interface ProxyOptions {
+  target: TransparentSubscriptionTarget;
+  upstreamPath: string;
+  trackRequest?: boolean;
+}
+
+const TEXT_FIELDS = new Set([
+  'content',
+  'input',
+  'input_text',
+  'output',
+  'output_text',
+  'result',
+  'system',
+  'text',
+]);
+
+const INTERNAL_HEADERS = new Set([
+  'connection',
+  'content-length',
+  'host',
+  'transfer-encoding',
+  'x-forwarded-for',
+  'x-forwarded-host',
+  'x-forwarded-proto',
+  'x-llm-gateway-key',
+  'x-llm-gateway-upstream',
+  'x-llm-interceptor-caller',
+  'x-llm-request-id',
+]);
+
+const RESPONSE_HEADERS = new Set([
+  'cache-control',
+  'content-type',
+  'openai-processing-ms',
+  'request-id',
+  'x-request-id',
+  'x-ratelimit-limit-requests',
+  'x-ratelimit-limit-tokens',
+  'x-ratelimit-remaining-requests',
+  'x-ratelimit-remaining-tokens',
+  'x-ratelimit-reset-requests',
+  'x-ratelimit-reset-tokens',
+]);
+
+function estimatePayloadTokens(payload: unknown): number {
+  return Math.max(1, Math.ceil(JSON.stringify(payload ?? {}).length / 4));
+}
+
+function countRedactions(counts: Record<string, number>): number {
+  return Object.values(counts).reduce((total, count) => total + count, 0);
+}
+
+export function transformSubscriptionPayload(
+  payload: unknown,
+  options: { providerName: string; caller: string },
+): TransformStats {
+  const tokensBefore = estimatePayloadTokens(payload);
+  let compressionApplied = false;
+  let piiRedactions = 0;
+  const redact = shouldRedactFor(getRedactMode(), options.providerName, options.caller);
+
+  const visit = (value: unknown, parentField?: string): unknown => {
+    if (typeof value === 'string') {
+      if (!parentField || !TEXT_FIELDS.has(parentField)) return value;
+      const redacted = redact ? redactPii(value) : null;
+      if (redacted) piiRedactions += countRedactions(redacted.counts);
+      const protectedText = redacted?.redacted ?? value;
+      const compressed = compressContext(protectedText, { enabled: true, mode: 'structured-safe' });
+      compressionApplied ||= compressed.applied;
+      return compressed.input;
+    }
+
+    if (Array.isArray(value)) return value.map((item) => visit(item, parentField));
+    if (!value || typeof value !== 'object') return value;
+
+    return Object.fromEntries(
+      Object.entries(value as Record<string, unknown>).map(([key, child]) => [key, visit(child, key)])
+    );
+  };
+
+  const transformed = visit(payload);
+  return {
+    payload: transformed,
+    tokensBefore,
+    tokensAfter: estimatePayloadTokens(transformed),
+    compressionApplied,
+    piiRedactions,
+  };
+}
+
+function requestHeader(request: FastifyRequest, name: string): string | undefined {
+  const value = request.headers[name];
+  return Array.isArray(value) ? value[0] : value;
+}
+
+export function transparentTargetFromRequest(request: FastifyRequest): TransparentSubscriptionTarget | null {
+  const target = requestHeader(request, 'x-llm-gateway-upstream')?.toLowerCase();
+  return target === 'anthropic' || target === 'codex' ? target : null;
+}
+
+function upstreamBase(target: TransparentSubscriptionTarget): string {
+  if (target === 'anthropic') {
+    return (process.env['ANTHROPIC_UPSTREAM_URL'] ?? 'https://api.anthropic.com/v1').replace(/\/+$/, '');
+  }
+  return (process.env['CODEX_UPSTREAM_URL'] ?? 'https://chatgpt.com/backend-api/codex').replace(/\/+$/, '');
+}
+
+function forwardedHeaders(request: FastifyRequest): Record<string, string> {
+  const headers: Record<string, string> = {};
+  for (const [name, value] of Object.entries(request.headers)) {
+    if (INTERNAL_HEADERS.has(name) || name.startsWith('cf-')) continue;
+    const first = Array.isArray(value) ? value[0] : value;
+    if (first !== undefined) headers[name] = first;
+  }
+  headers['content-type'] = 'application/json';
+  return headers;
+}
+
+function safeRequestId(request: FastifyRequest, target: TransparentSubscriptionTarget): string {
+  const provided = requestHeader(request, 'x-llm-request-id');
+  if (provided && /^[A-Za-z0-9:._-]{8,220}$/.test(provided)) return provided;
+  return `${target}-proxy:${Date.now()}:${Math.random().toString(36).slice(2, 9)}`;
+}
+
+async function trackProxyRequest(
+  callId: string,
+  caller: string,
+  model: string,
+  target: TransparentSubscriptionTarget,
+  status: 'approved' | 'error',
+  latencyMs: number,
+  transform: TransformStats,
+): Promise<void> {
+  const db = getPool();
+  const saved = Math.max(0, transform.tokensBefore - transform.tokensAfter);
+  await logCompressionMetric(db, {
+    filePath: callId,
+    mode: transform.compressionApplied ? 'structured:head-tail-excerpt' : 'structured:none',
+    tokensBefore: transform.tokensBefore,
+    tokensAfter: transform.tokensAfter,
+    savingsPct: transform.tokensBefore > 0 ? Math.round((saved / transform.tokensBefore) * 10_000) / 100 : 0,
+    toolUsed: 'gateway',
+  });
+  await createRequestLogger(db).logRequest(
+    callId,
+    caller,
+    `${target}_transparent_bridge`,
+    model,
+    status,
+    transform.tokensAfter,
+    0,
+    0,
+    latencyMs,
+    0,
+    transform.compressionApplied,
+  );
+}
+
+export async function proxyTransparentSubscription(
+  request: FastifyRequest,
+  reply: FastifyReply,
+  options: ProxyOptions,
+): Promise<FastifyReply> {
+  const startedAt = Date.now();
+  const caller = requestHeader(request, 'x-llm-interceptor-caller')
+    ?? requestHeader(request, 'x-caller-id')
+    ?? `${options.target}-transparent-client`;
+  const callId = safeRequestId(request, options.target);
+  const rawPayload = request.body ?? {};
+  const transformed = transformSubscriptionPayload(rawPayload, {
+    providerName: `${options.target}-subscription`,
+    caller,
+  });
+  const model = typeof (rawPayload as Record<string, unknown>)?.['model'] === 'string'
+    ? String((rawPayload as Record<string, unknown>)['model'])
+    : options.target;
+  const query = request.url.includes('?') ? request.url.slice(request.url.indexOf('?')) : '';
+  const targetUrl = `${upstreamBase(options.target)}${options.upstreamPath}${query}`;
+
+  try {
+    const upstream = await fetch(targetUrl, {
+      method: request.method,
+      headers: forwardedHeaders(request),
+      body: request.method === 'GET' || request.method === 'HEAD'
+        ? undefined
+        : JSON.stringify(transformed.payload),
+    });
+
+    reply.status(upstream.status).header('X-LLM-Gateway-Request-ID', callId);
+    for (const [name, value] of upstream.headers.entries()) {
+      if (RESPONSE_HEADERS.has(name.toLowerCase())) reply.header(name, value);
+    }
+
+    if (options.trackRequest) {
+      void trackProxyRequest(
+        callId,
+        caller,
+        model,
+        options.target,
+        upstream.ok ? 'approved' : 'error',
+        Date.now() - startedAt,
+        transformed,
+      ).catch((error) => logger.warn({ error, callId }, 'Transparent bridge tracking failed'));
+    }
+
+    logger.info({
+      callId,
+      caller,
+      target: options.target,
+      model,
+      statusCode: upstream.status,
+      tokensBefore: transformed.tokensBefore,
+      tokensAfter: transformed.tokensAfter,
+      piiRedactions: transformed.piiRedactions,
+    }, 'Transparent subscription request proxied');
+
+    if (!upstream.body) return reply.send();
+    return reply.send(Readable.fromWeb(upstream.body as any));
+  } catch (error) {
+    logger.error({ error, callId, target: options.target }, 'Transparent subscription proxy failed');
+    if (options.trackRequest) {
+      void trackProxyRequest(
+        callId,
+        caller,
+        model,
+        options.target,
+        'error',
+        Date.now() - startedAt,
+        transformed,
+      ).catch(() => undefined);
+    }
+    return reply.status(502).send({
+      error: {
+        message: 'Subscription upstream unavailable',
+        type: 'upstream_error',
+        code: 502,
+      },
+    });
+  }
+}
--- a/packages/gateway/src/routes/tests/transparent-subscription-route.test.ts
+++ b/packages/gateway/src/routes/tests/transparent-subscription-route.test.ts
@ -0,0 +1,83 @@
+import Fastify from 'fastify';
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+import { completionRoute } from '../completion.js';
+
+const gatewayKey = ['unit', 'gateway', 'route', 'key'].join('-');
+const providerSession = ['provider', 'session', 'value'].join('-');
+
+beforeEach(() => {
+  process.env['LLM_GATEWAY_API_AUTH_MODE'] = 'enforce';
+  process.env['LLM_GATEWAY_API_KEYS'] = gatewayKey;
+  process.env['REDACT_PII_MODE'] = 'always';
+  process.env['ANTHROPIC_UPSTREAM_URL'] = 'https://example.invalid/v1';
+});
+
+afterEach(() => {
+  vi.unstubAllGlobals();
+  delete process.env['LLM_GATEWAY_API_AUTH_MODE'];
+  delete process.env['LLM_GATEWAY_API_KEYS'];
+  delete process.env['REDACT_PII_MODE'];
+  delete process.env['ANTHROPIC_UPSTREAM_URL'];
+});
+
+describe('transparent subscription routes', () => {
+  it('requires the dedicated gateway key', async () => {
+    const fetchMock = vi.fn();
+    vi.stubGlobal('fetch', fetchMock);
+    const app = Fastify();
+    await app.register(completionRoute, { prefix: '/v1' });
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/v1/messages',
+      headers: { 'x-llm-gateway-upstream': 'anthropic' },
+      payload: { model: 'subscription-test-model', max_tokens: 16, messages: [{ role: 'user', content: 'hello' }] },
+    });
+
+    expect(response.statusCode).toBe(401);
+    expect(fetchMock).not.toHaveBeenCalled();
+    await app.close();
+  });
+
+  it('forwards provider auth, strips gateway auth and protects prompt text', async () => {
+    const email = ['route', 'example.org'].join('@');
+    const fetchMock = vi.fn(async () => new Response(JSON.stringify({
+      id: 'msg_test',
+      type: 'message',
+      role: 'assistant',
+      model: 'subscription-test-model',
+      content: [{ type: 'text', text: 'ok' }],
+      usage: { input_tokens: 10, output_tokens: 1 },
+    }), { status: 200, headers: { 'content-type': 'application/json' } }));
+    vi.stubGlobal('fetch', fetchMock);
+    const app = Fastify();
+    await app.register(completionRoute, { prefix: '/v1' });
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/v1/messages',
+      headers: {
+        authorization: `Bearer ${providerSession}`,
+        'x-llm-gateway-key': gatewayKey,
+        'x-llm-gateway-upstream': 'anthropic',
+        'anthropic-version': '2023-06-01',
+      },
+      payload: {
+        model: 'subscription-test-model',
+        max_tokens: 16,
+        messages: [{ role: 'user', content: `Contact ${email}` }],
+      },
+    });
+
+    expect(response.statusCode).toBe(200);
+    expect(fetchMock).toHaveBeenCalledOnce();
+    const [url, init] = fetchMock.mock.calls[0] as unknown as [string, RequestInit];
+    const headers = init.headers as Record<string, string>;
+    expect(url).toBe('https://example.invalid/v1/messages');
+    expect(headers.authorization).toBe(`Bearer ${providerSession}`);
+    expect(headers['x-llm-gateway-key']).toBeUndefined();
+    expect(String(init.body)).not.toContain(email);
+    expect(String(init.body)).toContain('<EMAIL_001>');
+    await app.close();
+  });
+});
--- a/packages/gateway/src/routes/completion.ts
+++ b/packages/gateway/src/routes/completion.ts
@ -68,6 +68,11 @@ import { runPreComplete, runPostComplete } from '../modules/plugin-system.js';
 import { getAdaptiveRecommendation } from '../modules/adaptive-routing.js';
 import { guardOutputStream, getOutputDefenseMode } from '../modules/output-defense.js';
 import { callPromptGuard, isPromptGuardConfigured, getPromptGuardThreshold, getPromptGuardMinLen } from '../modules/prompt-guard-client.js';
+import {
+  proxyTransparentSubscription,
+  transparentTargetFromRequest,
+} from '../modules/transparent-subscription-proxy.js';
+import { authorizeGatewayRequest } from '../security/tenant-auth.js';

 // // Disable Ollama-dependent scanners (sentinel, constitutional, embedding, attention)
 // // to keep gateway scans fast and dependency-free
@ -121,6 +126,11 @@ const CompletionRequestSchema = z.object({

 type CompletionRequest = z.infer<typeof CompletionRequestSchema>;

+const AGENT_BODY_LIMIT_BYTES = Math.min(
+  64 * 1024 * 1024,
+  Math.max(1024 * 1024, Number(process.env['GATEWAY_AGENT_BODY_LIMIT_BYTES'] ?? 16 * 1024 * 1024)),
+);
+
 function shouldBypassResponseCache(caller: string): boolean {
  const normalized = caller.toLowerCase();
  return normalized.includes('claude-code')
@ -1523,7 +1533,7 @@ export async function completionRoute(fastify: FastifyInstance): Promise<void> {
    return reply.send(listGatewayModels());
  });

-  fastify.post('/chat/completions', { config: { rateLimit: false } }, async (request: FastifyRequest, reply: FastifyReply) => {
+  fastify.post('/chat/completions', { bodyLimit: AGENT_BODY_LIMIT_BYTES, config: { rateLimit: false } }, async (request: FastifyRequest, reply: FastifyReply) => {
    const startMs = Date.now();
    const parsed = OpenAIChatCompletionRequestSchema.safeParse(request.body);
    if (!parsed.success) {
@ -1538,6 +1548,9 @@ export async function completionRoute(fastify: FastifyInstance): Promise<void> {

    const callId = `chatcmpl-${Date.now()}-${Math.random().toString(36).slice(2, 9)}`;
    const requestedSubscriptionId = requestedSubscriptionFromHeaders(request);
+    if (requestedSubscriptionId || subscriptionBridgeTargetForModel(parsed.data.model)) {
+      if (!await authorizeGatewayRequest(request, reply)) return reply;
+    }
    const { caller: bridgeCaller } = detectCaller(request, 'openai-compatible', parsed.data.user);
    const subscriptionBridgeResult = await callSubscriptionBridgeChatCompletion(parsed.data, startMs, callId, requestedSubscriptionId, bridgeCaller);
    if (requestedSubscriptionId && !subscriptionBridgeResult) {
@ -1581,8 +1594,21 @@ export async function completionRoute(fastify: FastifyInstance): Promise<void> {
  });

  // Anthropic Messages API compatibility — accept @anthropic-ai/sdk traffic.
-  fastify.post('/messages', { config: { rateLimit: false } }, async (request: FastifyRequest, reply: FastifyReply) => {
+  fastify.post('/messages', { bodyLimit: AGENT_BODY_LIMIT_BYTES, config: { rateLimit: false } }, async (request: FastifyRequest, reply: FastifyReply) => {
    const startMs = Date.now();
+    const transparentTarget = transparentTargetFromRequest(request);
+    if (transparentTarget) {
+      if (transparentTarget !== 'anthropic') {
+        return reply.status(400).send({ type: 'error', error: { type: 'invalid_request_error', message: 'invalid transparent upstream for Messages API' } });
+      }
+      if (!await authorizeGatewayRequest(request, reply)) return reply;
+      return proxyTransparentSubscription(request, reply, {
+        target: 'anthropic',
+        upstreamPath: '/messages',
+        trackRequest: false,
+      });
+    }
+
    const parsed = AnthropicMessagesRequestSchema.safeParse(request.body);
    if (!parsed.success) {
      return reply.status(400).send({
@ -1594,6 +1620,10 @@ export async function completionRoute(fastify: FastifyInstance): Promise<void> {
      });
    }

+    if (modelToSubscriptionId(parsed.data.model) === 'claude-code') {
+      if (!await authorizeGatewayRequest(request, reply)) return reply;
+    }
+
    const callId = `msg_${Date.now()}_${Math.random().toString(36).slice(2, 9)}`;
    const gatewayRequest = anthropicRequestToGatewayRequest(parsed.data, request);
    const result = await executeCompletion(gatewayRequest, startMs, callId);
@ -1622,8 +1652,33 @@ export async function completionRoute(fastify: FastifyInstance): Promise<void> {
    return reply.status(200).send(response);
  });

-  fastify.post('/responses', { config: { rateLimit: false } }, async (request: FastifyRequest, reply: FastifyReply) => {
+  fastify.post('/messages/count_tokens', { bodyLimit: AGENT_BODY_LIMIT_BYTES, config: { rateLimit: false } }, async (request: FastifyRequest, reply: FastifyReply) => {
+    if (transparentTargetFromRequest(request) !== 'anthropic') {
+      return reply.status(400).send({ type: 'error', error: { type: 'invalid_request_error', message: 'transparent Anthropic bridge required' } });
+    }
+    if (!await authorizeGatewayRequest(request, reply)) return reply;
+    return proxyTransparentSubscription(request, reply, {
+      target: 'anthropic',
+      upstreamPath: '/messages/count_tokens',
+      trackRequest: false,
+    });
+  });
+
+  fastify.post('/responses', { bodyLimit: AGENT_BODY_LIMIT_BYTES, config: { rateLimit: false } }, async (request: FastifyRequest, reply: FastifyReply) => {
    const startMs = Date.now();
+    const transparentTarget = transparentTargetFromRequest(request);
+    if (transparentTarget) {
+      if (transparentTarget !== 'codex') {
+        return reply.status(400).send({ error: { message: 'invalid transparent upstream for Responses API', type: 'invalid_request_error', code: 400 } });
+      }
+      if (!await authorizeGatewayRequest(request, reply)) return reply;
+      return proxyTransparentSubscription(request, reply, {
+        target: 'codex',
+        upstreamPath: '/responses',
+        trackRequest: true,
+      });
+    }
+
    const parsed = OpenAIResponsesRequestSchema.safeParse(request.body);
    if (!parsed.success) {
      return reply.status(400).send({
@ -1643,6 +1698,7 @@ export async function completionRoute(fastify: FastifyInstance): Promise<void> {
    // the right auth. Route them straight to the local codex-bridge
    // (local bridge process) which speaks codex-cli over OAuth.
    if (isCodexBridgeModel(parsed.data.model)) {
+      if (!await authorizeGatewayRequest(request, reply)) return reply;
      try {
        const bridgeUrl = process.env['CODEX_BRIDGE_URL'] ?? 'http://localhost:3253';
        const bridgeCaller = (request.headers['x-llm-interceptor-caller'] as string) || 'codex-app';
@ -1790,6 +1846,18 @@ data: [DONE]
    return reply.send(response);
  });

+  fastify.post('/responses/compact', { bodyLimit: AGENT_BODY_LIMIT_BYTES, config: { rateLimit: false } }, async (request: FastifyRequest, reply: FastifyReply) => {
+    if (transparentTargetFromRequest(request) !== 'codex') {
+      return reply.status(400).send({ error: { message: 'transparent Codex bridge required', type: 'invalid_request_error', code: 400 } });
+    }
+    if (!await authorizeGatewayRequest(request, reply)) return reply;
+    return proxyTransparentSubscription(request, reply, {
+      target: 'codex',
+      upstreamPath: '/responses/compact',
+      trackRequest: false,
+    });
+  });
+
  // ─── Multi-Model Race Mode endpoint ────────────────────────────────────
  // Runs the same prompt against multiple models in parallel; returns
  // according to `strategy` (first | best | consensus). Audits each
--- a/packages/gateway/src/routes/dashboard.ts
+++ b/packages/gateway/src/routes/dashboard.ts
@ -1753,22 +1753,30 @@ export async function dashboardRoute(fastify: FastifyInstance): Promise<void> {
      // Upsert by request_id (one row per caller/model/day): re-reporting an
      // in-progress day updates its totals instead of creating duplicates.
      const updated = await db.query(
-        `UPDATE request_tracking SET tokens_in=$1, tokens_out=$2, cost_usd=$3, created_at=$4 WHERE request_id=$5`,
+        `UPDATE request_tracking SET tokens_in=$1, tokens_out=$2, cost_usd=$3, created_at=$4, status='warning', task_type='usage_import' WHERE request_id=$5`,
        [tokensIn, tokensOut, costUsd, usedAt, requestId]
      );
      if (updated.rowCount === 0) {
        const requestLogger = createRequestLogger(db);
-        await requestLogger.logRequest(requestId, caller, 'usage_import', model, 'approved', tokensIn, tokensOut, costUsd, 0);
+        await requestLogger.logRequest(requestId, caller, 'usage_import', model, 'warning', tokensIn, tokensOut, costUsd, 0);
        await db.query(`UPDATE request_tracking SET created_at=$1 WHERE request_id=$2`, [usedAt, requestId]);
      }
+      const updatedMetric = await db.query(
+        `UPDATE tokenvault_metrics
+         SET mode='usage-import:none', tokens_before=$2, tokens_after=$2, savings_pct=0, created_at=NOW(), tool_used='usage-import'
+         WHERE id=(SELECT id FROM tokenvault_metrics WHERE file_path=$1 ORDER BY created_at DESC LIMIT 1)`,
+        [requestId, tokensIn]
+      );
+      if (updatedMetric.rowCount === 0) {
        await logCompressionMetric(db, {
          filePath: requestId,
          mode: 'usage-import:none',
          tokensBefore: tokensIn,
          tokensAfter: tokensIn,
          savingsPct: 0,
-        toolUsed: 'gateway',
+          toolUsed: 'usage-import',
        });
+      }
      return reply.status(200).send({ success: true, imported: { caller, model, day, tokensIn, tokensOut, costUsd, usedAt } });
    } catch (error) {
      logger.error({ error }, 'Failed to import usage report');
--- a/packages/gateway/src/routes/subscriptions.ts
+++ b/packages/gateway/src/routes/subscriptions.ts
@ -14,6 +14,7 @@ import {
 import { getPublicSettings, saveSettings } from '../modules/settings-store.js';
 import { getAllProviders, getAvailableProviders } from '../pipeline/external-providers.js';
 import { logger } from '../observability/logger.js';
+import { requireGatewayApiKey } from '../security/tenant-auth.js';

 interface SubscriptionBridgeBody {
  subscription_id?: SubscriptionId;
@ -126,7 +127,7 @@ function buildAccessCard(
      `  -H "X-Caller-ID: pilot-client"`,
      `  -d '{"model":"${defaultModel}","messages":[{"role":"user","content":"gateway bridge smoke"}]}'`,
    ].join(' \\\n'),
-    note: 'Use a gateway API key/admin token from your deployment. OAuth/subscription secrets stay inside the bridge process and are never exposed here.',
+    note: 'Use a dedicated gateway API key from your deployment. OAuth/subscription secrets stay inside the bridge process and are never exposed here.',
  };
 }

@ -329,6 +330,7 @@ async function postBridgeChat(

 export async function subscriptionsRoute(fastify: FastifyInstance): Promise<void> {
  const auth = { preHandler: requireDashboardAuth };
+  const apiAuth = { preHandler: requireGatewayApiKey };

  fastify.get('/api/subscriptions', auth, async (request: FastifyRequest, reply: FastifyReply) => {
    const query = request.query as { search?: string; q?: string };
@ -521,7 +523,7 @@ export async function subscriptionsRoute(fastify: FastifyInstance): Promise<void
    }
  });

-  fastify.post('/api/subscriptions/:subscription_id/v1/chat/completions', auth, async (request: FastifyRequest, reply: FastifyReply) => {
+  fastify.post('/api/subscriptions/:subscription_id/v1/chat/completions', apiAuth, async (request: FastifyRequest, reply: FastifyReply) => {
    const params = request.params as SubscriptionChatParams;
    const status = await getSubscriptionStatusOrReply(params.subscription_id, reply);
    if (!status) return reply;
@ -542,6 +544,8 @@ export async function subscriptionsRoute(fastify: FastifyInstance): Promise<void
        url: '/v1/chat/completions',
        headers: {
          'content-type': 'application/json',
+          ...(request.headers.authorization ? { authorization: request.headers.authorization } : {}),
+          ...(request.headers['x-llm-gateway-key'] ? { 'x-llm-gateway-key': request.headers['x-llm-gateway-key'] } : {}),
          'x-caller-id': Array.isArray(caller) ? caller[0] : String(caller),
          'x-llm-gateway-subscription': status.descriptor.id,
        },
--- a/packages/gateway/src/routes/tracking.ts
+++ b/packages/gateway/src/routes/tracking.ts
@ -0,0 +1,76 @@
+import type { FastifyInstance, FastifyReply, FastifyRequest } from 'fastify';
+import { z } from 'zod';
+import { getPool } from '../db/client.js';
+import { createRequestLogger } from '../modules/request-logger.js';
+import { logger } from '../observability/logger.js';
+import { requireGatewayApiKey } from '../security/tenant-auth.js';
+import { logCompressionMetric } from '../utils/tokenvault-hooks.js';
+
+const BridgeTrackingReportSchema = z.object({
+  request_id: z.string().min(8).max(220).optional(),
+  caller: z.string().min(1).max(120),
+  model: z.string().min(1).max(120),
+  source: z.string().min(1).max(80).default('subscription-bridge'),
+  tokens_in: z.number().int().nonnegative().max(100_000_000),
+  tokens_out: z.number().int().nonnegative().max(100_000_000),
+  latency_ms: z.number().int().nonnegative().max(3_600_000).default(0),
+  compression_tokens_before: z.number().int().nonnegative().max(100_000_000).optional(),
+  compression_tokens_after: z.number().int().nonnegative().max(100_000_000).optional(),
+  compression_mode: z.string().min(1).max(48).optional(),
+});
+
+function safeRequestId(value: string | undefined): string {
+  if (value && /^[A-Za-z0-9:._-]{8,220}$/.test(value)) return value;
+  return `bridge-report:${Date.now()}:${Math.random().toString(36).slice(2, 9)}`;
+}
+
+export async function trackingRoute(fastify: FastifyInstance): Promise<void> {
+  fastify.post(
+    '/api/tracking/report',
+    { preHandler: requireGatewayApiKey },
+    async (request: FastifyRequest, reply: FastifyReply) => {
+      const parsed = BridgeTrackingReportSchema.safeParse(request.body);
+      if (!parsed.success) {
+        return reply.status(400).send({ success: false, error: parsed.error.errors[0]?.message ?? 'invalid report' });
+      }
+
+      const report = parsed.data;
+      const requestId = safeRequestId(report.request_id);
+      const before = report.compression_tokens_before ?? report.tokens_in;
+      const after = report.compression_tokens_after ?? report.tokens_in;
+      const saved = Math.max(0, before - after);
+      const db = getPool();
+
+      try {
+        const existing = await db.query('SELECT 1 FROM request_tracking WHERE request_id=$1 LIMIT 1', [requestId]);
+        if (existing.rowCount === 0) {
+          await createRequestLogger(db).logRequest(
+            requestId,
+            report.caller,
+            `subscription_bridge:${report.source}`.slice(0, 50),
+            report.model,
+            'approved',
+            report.tokens_in,
+            report.tokens_out,
+            0,
+            report.latency_ms,
+            0,
+            saved > 0,
+          );
+          await logCompressionMetric(db, {
+            filePath: requestId,
+            mode: report.compression_mode ?? (saved > 0 ? 'bridge:compressed' : 'bridge:none'),
+            tokensBefore: before,
+            tokensAfter: after,
+            savingsPct: before > 0 ? Math.round((saved / before) * 10_000) / 100 : 0,
+            toolUsed: 'gateway',
+          });
+        }
+        return reply.status(200).send({ success: true, requestId, duplicate: existing.rowCount !== 0 });
+      } catch (error) {
+        logger.error({ error, requestId }, 'Bridge tracking report failed');
+        return reply.status(500).send({ success: false, error: 'tracking report failed' });
+      }
+    },
+  );
+}
--- a/packages/gateway/src/security/tests/tenant-auth.test.ts
+++ b/packages/gateway/src/security/tests/tenant-auth.test.ts
@ -0,0 +1,42 @@
+import { afterEach, describe, expect, it } from 'vitest';
+import type { FastifyRequest } from 'fastify';
+import {
+  gatewayApiKeyMatches,
+  getGatewayApiAuthMode,
+  getPresentedGatewayApiKey,
+} from '../tenant-auth.js';
+
+function requestWithHeaders(headers: Record<string, string>): FastifyRequest {
+  return { headers } as unknown as FastifyRequest;
+}
+
+afterEach(() => {
+  delete process.env['LLM_GATEWAY_API_AUTH_MODE'];
+});
+
+describe('gateway API authentication', () => {
+  it('defaults to off and accepts explicit rollout modes', () => {
+    expect(getGatewayApiAuthMode()).toBe('off');
+    process.env['LLM_GATEWAY_API_AUTH_MODE'] = 'monitor';
+    expect(getGatewayApiAuthMode()).toBe('monitor');
+    process.env['LLM_GATEWAY_API_AUTH_MODE'] = 'enforce';
+    expect(getGatewayApiAuthMode()).toBe('enforce');
+  });
+
+  it('reads the dedicated bridge header before bearer authorization', () => {
+    const dedicated = ['unit', 'gateway', 'dedicated'].join('-');
+    const bearer = ['unit', 'gateway', 'bearer'].join('-');
+    const request = requestWithHeaders({
+      'x-llm-gateway-key': dedicated,
+      authorization: `Bearer ${bearer}`,
+    });
+    expect(getPresentedGatewayApiKey(request)).toBe(dedicated);
+  });
+
+  it('matches configured keys without accepting missing or partial values', () => {
+    const configured = ['unit', 'gateway', 'key', 'alpha'].join('-');
+    expect(gatewayApiKeyMatches(configured, [configured])).toBe(true);
+    expect(gatewayApiKeyMatches(`${configured}-partial`, [configured])).toBe(false);
+    expect(gatewayApiKeyMatches(undefined, [configured])).toBe(false);
+  });
+});
--- a/packages/gateway/src/security/tenant-auth.ts
+++ b/packages/gateway/src/security/tenant-auth.ts
@ -1,9 +1,93 @@
 import fp from 'fastify-plugin';
-import type { FastifyInstance } from 'fastify';
+import { timingSafeEqual } from 'crypto';
+import type { FastifyInstance, FastifyReply, FastifyRequest } from 'fastify';
+import { logger } from '../observability/logger.js';
+
+export type GatewayApiAuthMode = 'off' | 'monitor' | 'enforce';
+
+export function getGatewayApiAuthMode(): GatewayApiAuthMode {
+  const mode = (process.env['LLM_GATEWAY_API_AUTH_MODE'] ?? 'off').trim().toLowerCase();
+  if (mode === 'monitor' || mode === 'enforce') return mode;
+  return 'off';
+}
+
+export function getConfiguredGatewayApiKeys(): string[] {
+  return (process.env['LLM_GATEWAY_API_KEYS'] ?? '')
+    .split(',')
+    .map((key) => key.trim())
+    .filter(Boolean);
+}
+
+function firstHeader(value: string | string[] | undefined): string | undefined {
+  return Array.isArray(value) ? value[0] : value;
+}
+
+export function getPresentedGatewayApiKey(request: FastifyRequest): string | undefined {
+  const dedicated = firstHeader(request.headers['x-llm-gateway-key']);
+  if (dedicated?.trim()) return dedicated.trim();
+
+  const authorization = firstHeader(request.headers.authorization);
+  const match = authorization?.match(/^Bearer\s+(.+)$/i);
+  return match?.[1]?.trim();
+}
+
+export function gatewayApiKeyMatches(presented: string | undefined, configuredKeys: readonly string[]): boolean {
+  if (!presented) return false;
+  const candidate = Buffer.from(presented);
+  return configuredKeys.some((configured) => {
+    const expected = Buffer.from(configured);
+    return candidate.length === expected.length && timingSafeEqual(candidate, expected);
+  });
+}
+
+export async function authorizeGatewayRequest(
+  request: FastifyRequest,
+  reply: FastifyReply,
+): Promise<boolean> {
+  const mode = getGatewayApiAuthMode();
+  if (mode === 'off') return true;
+
+  const configuredKeys = getConfiguredGatewayApiKeys();
+  const valid = gatewayApiKeyMatches(getPresentedGatewayApiKey(request), configuredKeys);
+  if (valid) return true;
+
+  if (mode === 'monitor') {
+    logger.warn({ method: request.method, path: request.url.split('?')[0] }, 'Gateway API key missing or invalid');
+    return true;
+  }
+
+  if (configuredKeys.length === 0) {
+    reply.status(503).send({
+      error: {
+        message: 'Gateway API authentication is not configured',
+        type: 'gateway_auth_configuration_error',
+        code: 503,
+      },
+    });
+    return false;
+  }
+
+  reply.header('WWW-Authenticate', 'Bearer realm="llm-gateway"');
+  reply.status(401).send({
+    error: {
+      message: 'Valid gateway API key required',
+      type: 'authentication_error',
+      code: 401,
+    },
+  });
+  return false;
+}
+
+export async function requireGatewayApiKey(
+  request: FastifyRequest,
+  reply: FastifyReply,
+): Promise<FastifyReply | void> {
+  const authorized = await authorizeGatewayRequest(request, reply);
+  if (!authorized) return reply;
+}

 async function tenantAuth(_fastify: FastifyInstance): Promise<void> {
-  // Tenant auth is intentionally permissive until tenant policies are configured.
-  // Admin and dashboard routes keep their own token checks.
+  logger.info({ mode: getGatewayApiAuthMode(), configuredKeys: getConfiguredGatewayApiKeys().length }, 'Gateway API authentication initialised');
 }

 export default fp(tenantAuth, { name: 'tenant-auth' });
--- a/packages/gateway/src/server.ts
+++ b/packages/gateway/src/server.ts
@ -3,6 +3,7 @@ import fastifyCors from '@fastify/cors';
 import fastifyRateLimit from '@fastify/rate-limit';
 import fastifyHelmet from '@fastify/helmet';
 import { completionRoute } from './routes/completion.js';
+import { trackingRoute } from './routes/tracking.js';
 import { batchRoute } from './routes/batch.js';
 import { classifyRoute } from './routes/classify.js';
 import { guardRoute } from './routes/guard.js';
@ -111,7 +112,18 @@ async function buildServer() {
      /^https:\/\/.*\.runwerk\.app$/,
    ],
    methods: ['GET', 'POST', 'PUT', 'DELETE', 'OPTIONS'],
-    allowedHeaders: ['Content-Type', 'Authorization', 'X-Caller-ID', 'X-Runwerk-Caller', 'X-Runwerk-Privacy', 'X-Runwerk-Tier', 'X-Runwerk-Purpose'],
+    allowedHeaders: [
+      'Content-Type',
+      'Authorization',
+      'X-Caller-ID',
+      'X-LLM-Gateway-Key',
+      'X-LLM-Gateway-Upstream',
+      'X-LLM-Interceptor-Caller',
+      'X-Runwerk-Caller',
+      'X-Runwerk-Privacy',
+      'X-Runwerk-Tier',
+      'X-Runwerk-Purpose',
+    ],
    credentials: true,
  });

@ -132,6 +144,7 @@ async function buildServer() {

  await server.register(tenantAuth);
  await server.register(internalRoute);
+  await server.register(trackingRoute);
  await server.register(completionRoute, { prefix: '/v1' });
  await server.register(embeddingsRoute, { prefix: '/v1' });
  await server.register(replayRoute, { prefix: '/v1' });
--- a/scripts/client/anthropic-interceptor.py
+++ b/scripts/client/anthropic-interceptor.py
@ -0,0 +1,288 @@
+#!/usr/bin/env python3
+"""Fail-closed local subscription companion for Claude Code and Codex."""
+
+import asyncio
+import json
+import math
+import os
+import re
+import subprocess
+import time
+import uuid
+
+import httpx
+from fastapi import FastAPI, Request, Response
+from fastapi.responses import JSONResponse, StreamingResponse
+
+
+GATEWAY_URL = os.getenv("LLM_GATEWAY_URL", "https://llm-gateway.context-x.org").rstrip("/")
+KEYCHAIN_SERVICE = os.getenv("LLM_GATEWAY_KEYCHAIN_SERVICE", "llm-gateway-api-key")
+CLAUDE_CALLER = os.getenv("INTERCEPTOR_CALLER", "claude-code-macbook")
+CODEX_CALLER = os.getenv("CODEX_INTERCEPTOR_CALLER", "codex-macbook")
+MAX_TOOL_CHARS = int(os.getenv("MAX_TOOL_CHARS", "8000"))
+
+_PII = [
+    (re.compile(r"\b[A-Za-z0-9._%+\-]+@[A-Za-z0-9.\-]+\.[A-Za-z]{2,}\b"), "[EMAIL]"),
+    (re.compile(r"\b(?:\d{1,3}\.){3}\d{1,3}\b"), "[IP_ADDRESS]"),
+    (re.compile(r"\b(?:IBAN[:\s]*)?[A-Z]{2}\d{2}[A-Z0-9 ]{10,30}\b"), "[IBAN]"),
+    (re.compile(r"\beyJ[A-Za-z0-9_\-]+\.eyJ[A-Za-z0-9_\-]+\.[A-Za-z0-9_\-]+\b"), "[JWT]"),
+    (re.compile(r"\b(?:sk-proj-|sk-ant-)[A-Za-z0-9_\-]{20,}\b"), "[API_KEY]"),
+    (re.compile(r"\bAKIA[0-9A-Z]{16}\b"), "[AWS_KEY]"),
+    (re.compile(r"\b(?:ghp_|github_pat_)[A-Za-z0-9_]{20,}\b"), "[GH_TOKEN]"),
+]
+
+
+def gateway_key() -> str:
+    configured = os.getenv("LLM_GATEWAY_API_KEY", "").strip()
+    if configured:
+        return configured
+    try:
+        return subprocess.check_output(
+            ["/usr/bin/security", "find-generic-password", "-s", KEYCHAIN_SERVICE, "-w"],
+            stderr=subprocess.DEVNULL,
+            timeout=5,
+        ).decode().strip()
+    except Exception:
+        return ""
+
+
+def anonymize(text: str) -> str:
+    for pattern, replacement in _PII:
+        text = pattern.sub(replacement, text)
+    return text
+
+
+def anonymize_content(content):
+    if isinstance(content, str):
+        return anonymize(content)
+    if isinstance(content, list):
+        result = []
+        for part in content:
+            if isinstance(part, dict) and part.get("type") == "text":
+                part = {**part, "text": anonymize(str(part.get("text", "")))}
+            result.append(part)
+        return result
+    return content
+
+
+def compress_messages(messages: list) -> list:
+    result = []
+    for original in messages:
+        message = dict(original)
+        content = message.get("content", "")
+        if message.get("role") == "tool" and isinstance(content, str) and len(content) > MAX_TOOL_CHARS:
+            message["content"] = content[:MAX_TOOL_CHARS] + "\n[truncated by subscription companion]"
+        elif message.get("role") == "user" and isinstance(content, list):
+            parts = []
+            for original_part in content:
+                part = dict(original_part) if isinstance(original_part, dict) else original_part
+                if isinstance(part, dict) and part.get("type") == "tool_result":
+                    inner = part.get("content", "")
+                    if isinstance(inner, str) and len(inner) > MAX_TOOL_CHARS:
+                        part["content"] = inner[:MAX_TOOL_CHARS] + "\n[truncated by subscription companion]"
+                    elif isinstance(inner, list):
+                        trimmed = []
+                        for original_inner_part in inner:
+                            inner_part = dict(original_inner_part) if isinstance(original_inner_part, dict) else original_inner_part
+                            if (
+                                isinstance(inner_part, dict)
+                                and inner_part.get("type") == "text"
+                                and len(str(inner_part.get("text", ""))) > MAX_TOOL_CHARS
+                            ):
+                                inner_part["text"] = str(inner_part["text"])[:MAX_TOOL_CHARS] + "\n[truncated by subscription companion]"
+                            trimmed.append(inner_part)
+                        part["content"] = trimmed
+                parts.append(part)
+            message["content"] = parts
+        result.append(message)
+    return result
+
+
+def estimate_tokens(payload) -> int:
+    return max(1, math.ceil(len(json.dumps(payload, ensure_ascii=False)) / 4))
+
+
+def prepare_anthropic_body(raw_body: bytes) -> tuple[bytes, int, int, bool]:
+    data = json.loads(raw_body or b"{}")
+    before = estimate_tokens(data)
+    is_stream = bool(data.get("stream"))
+    if isinstance(data.get("system"), str):
+        data["system"] = anonymize(data["system"])
+    elif isinstance(data.get("system"), list):
+        data["system"] = anonymize_content(data["system"])
+
+    messages = []
+    for original in data.get("messages", []):
+        message = dict(original)
+        message["content"] = anonymize_content(message.get("content", ""))
+        messages.append(message)
+    data["messages"] = compress_messages(messages)
+    after = estimate_tokens(data)
+    return json.dumps(data, ensure_ascii=False).encode(), before, after, is_stream
+
+
+def response_headers(headers: httpx.Headers) -> dict[str, str]:
+    blocked = {"connection", "content-encoding", "content-length", "keep-alive", "transfer-encoding"}
+    return {name: value for name, value in headers.items() if name.lower() not in blocked}
+
+
+async def report_usage(
+    key: str,
+    request_id: str,
+    caller: str,
+    model: str,
+    tokens_in: int,
+    tokens_out: int,
+    latency_ms: int,
+    before: int,
+    after: int,
+) -> None:
+    try:
+        async with httpx.AsyncClient(timeout=5.0) as client:
+            await client.post(
+                f"{GATEWAY_URL}/api/tracking/report",
+                headers={"X-LLM-Gateway-Key": key},
+                json={
+                    "request_id": request_id,
+                    "caller": caller,
+                    "model": model,
+                    "tokens_in": tokens_in,
+                    "tokens_out": tokens_out,
+                    "latency_ms": latency_ms,
+                    "source": "local-subscription-companion",
+                    "compression_tokens_before": before,
+                    "compression_tokens_after": after,
+                    "compression_mode": "interceptor:tool-trim" if before > after else "interceptor:none",
+                },
+            )
+    except Exception:
+        return
+
+
+app = FastAPI(title="llm-subscription-companion", docs_url=None, redoc_url=None)
+
+
+@app.get("/health")
+async def health():
+    configured = bool(gateway_key())
+    return JSONResponse(
+        status_code=200 if configured else 503,
+        content={"status": "ok" if configured else "auth_required", "gateway": GATEWAY_URL, "failClosed": True},
+    )
+
+
+@app.api_route("/{path:path}", methods=["GET", "POST", "PUT", "DELETE", "OPTIONS", "HEAD", "PATCH"])
+async def proxy(request: Request, path: str):
+    key = gateway_key()
+    if not key:
+        return JSONResponse(status_code=503, content={"error": "gateway authentication unavailable"})
+
+    started_at = time.monotonic()
+    raw_body = await request.body()
+    is_codex = path.startswith("openai/")
+    gateway_path = path[len("openai/"):] if is_codex else path
+    target_kind = "codex" if is_codex else "anthropic"
+    caller = CODEX_CALLER if is_codex else CLAUDE_CALLER
+    is_messages = not is_codex and gateway_path == "v1/messages" and request.method == "POST"
+    is_stream = False
+    before = after = estimate_tokens({})
+
+    if is_messages:
+        try:
+            raw_body, before, after, is_stream = prepare_anthropic_body(raw_body)
+        except Exception:
+            return JSONResponse(status_code=400, content={"error": "invalid Anthropic request body"})
+
+    request_id = f"{target_kind}-companion:{int(time.time() * 1000)}:{uuid.uuid4().hex[:8]}"
+    blocked = {"connection", "content-length", "host", "transfer-encoding", "x-llm-gateway-key"}
+    headers = {name: value for name, value in request.headers.items() if name.lower() not in blocked}
+    headers.update({
+        "X-LLM-Gateway-Key": key,
+        "X-LLM-Gateway-Upstream": target_kind,
+        "X-LLM-Interceptor-Caller": caller,
+        "X-LLM-Request-ID": request_id,
+    })
+    target = f"{GATEWAY_URL}/{gateway_path}"
+    params = dict(request.query_params)
+
+    if is_stream:
+        client = httpx.AsyncClient(timeout=300.0)
+        upstream_request = client.build_request(request.method, target, headers=headers, content=raw_body, params=params)
+        upstream = await client.send(upstream_request, stream=True)
+
+        async def stream_response():
+            pending = ""
+            model_name, tokens_in, tokens_out = "unknown", 0, 0
+            try:
+                async for chunk in upstream.aiter_bytes():
+                    pending += chunk.decode(errors="ignore")
+                    lines = pending.split("\n")
+                    pending = lines.pop()
+                    for line in lines:
+                        if not line.startswith("data: "):
+                            continue
+                        try:
+                            event = json.loads(line[6:])
+                        except Exception:
+                            continue
+                        if event.get("type") == "message_start":
+                            message = event.get("message", {})
+                            model_name = message.get("model", model_name)
+                            tokens_in = message.get("usage", {}).get("input_tokens", tokens_in)
+                        elif event.get("type") == "message_delta":
+                            tokens_out = event.get("usage", {}).get("output_tokens", tokens_out)
+                    yield chunk
+            finally:
+                await upstream.aclose()
+                await client.aclose()
+                if tokens_in or tokens_out:
+                    await report_usage(
+                        key, request_id, caller, model_name, tokens_in, tokens_out,
+                        int((time.monotonic() - started_at) * 1000), before, after,
+                    )
+
+        return StreamingResponse(
+            stream_response(),
+            status_code=upstream.status_code,
+            headers=response_headers(upstream.headers),
+            media_type=upstream.headers.get("content-type", "text/event-stream"),
+        )
+
+    async with httpx.AsyncClient(timeout=300.0) as client:
+        upstream = await client.request(
+            request.method,
+            target,
+            headers=headers,
+            content=raw_body if request.method not in {"GET", "HEAD"} else None,
+            params=params,
+        )
+
+    if is_messages and upstream.is_success:
+        try:
+            response_data = upstream.json()
+            usage = response_data.get("usage", {})
+            asyncio.create_task(report_usage(
+                key,
+                request_id,
+                caller,
+                response_data.get("model", "unknown"),
+                int(usage.get("input_tokens", 0)),
+                int(usage.get("output_tokens", 0)),
+                int((time.monotonic() - started_at) * 1000),
+                before,
+                after,
+            ))
+        except Exception:
+            pass
+
+    return Response(
+        content=upstream.content,
+        status_code=upstream.status_code,
+        headers=response_headers(upstream.headers),
+    )
+
+
+if __name__ == "__main__":
+    import uvicorn
+
+    uvicorn.run(app, host="127.0.0.1", port=int(os.getenv("PORT", "3090")), log_level="warning")
--- a/start-with-env.sh
+++ b/start-with-env.sh
@ -3,6 +3,7 @@
 # Production fix for the recurring PM2 env-drop quirk.
 set -a
 [ -f /opt/llm-gateway/.env.defense ] && source /opt/llm-gateway/.env.defense
+[ -f /opt/llm-gateway/.env.api ] && source /opt/llm-gateway/.env.api
 [ -f /opt/llm-gateway/.env ] && source /opt/llm-gateway/.env
 set +a
 exec /usr/bin/node /opt/llm-gateway/packages/gateway/scripts/launch.mjs