chore: snapshot main sync

apps/control/src/routes/gateway.ts

@@ -1,13 +1,13 @@
 import type { FastifyInstance, FastifyRequest, FastifyReply } from 'fastify';
 import type { Sql } from '../db.js';
-import type { FleetState } from '../services/fleet-state.js';
-import type { DeltaEmitter } from '../index.js';
 import {
   VIRTUAL_MODELS,
   resolveCandidates,
   splitComposite,
 } from '../services/gateway.js';
 import { resolveProviderBaseUrl } from '../services/llama-providers.js';
+import { recordFailure, recordSuccess } from '../services/circuit-breaker.js';
+import type { AppContext } from '../app-context.js';
 
 /**
  * P7.1: OpenAI-compatible auto:* gateway.
@@ -25,11 +25,9 @@ import { resolveProviderBaseUrl } from '../services/llama-providers.js';
  */
 export function registerGatewayRoutes(
   app: FastifyInstance,
-  sql: Sql,
-  fleet: FleetState,
-  _emitter: DeltaEmitter,
+  ctx: AppContext,
 ): void {
-  // ─── model catalog ───────────────────────────────────────────────────────
+  const { sql, fleet } = ctx;
 
   app.get('/v1/models', async (_req: FastifyRequest, reply: FastifyReply) => {
     return reply.send({
@@ -43,10 +41,6 @@ export function registerGatewayRoutes(
     });
   });
 
-  // ─── props (for getModelContext) ─────────────────────────────────────────
-  // Resolve candidates and proxy the first healthy candidate's props so the
-  // caller can read default_generation_settings.n_ctx.
-
   app.get('/upstream/:model/props', async (req: FastifyRequest, reply: FastifyReply) => {
     const { model } = req.params as { model: string };
     const { candidates } = await resolveCandidates(sql, fleet, model);
@@ -69,8 +63,6 @@ export function registerGatewayRoutes(
     return reply.status(503).send({ error: 'no healthy candidate for virtual model', model });
   });
 
-  // ─── chat completions (dispatch with failover) ───────────────────────────
-
   app.post('/v1/chat/completions', async (req: FastifyRequest, reply: FastifyReply) => {
     const body = req.body as Record<string, unknown>;
     const requestedModel = body?.model as string | undefined;
@@ -113,11 +105,20 @@ export function registerGatewayRoutes(
         });
 
         if (!res.ok) {
-          // HTTP error before body — eligible for failover to the next candidate.
+          recordFailure(compositeId);
+          continue;
+        }
+
+        // A null body on an OK response is a broken upstream; fail over to the
+        // next candidate (nothing has been committed to the client yet).
+        const reader = stream ? res.body?.getReader() : null;
+        if (stream && !reader) {
+          recordFailure(compositeId);
           continue;
         }
 
         // Success: dispatch chosen. Log and stream/return through.
+        recordSuccess(compositeId);
         await logDispatch(sql, {
           virtualModel,
           chosen: compositeId,
@@ -128,16 +129,11 @@ export function registerGatewayRoutes(
           durationMs: Date.now() - startedAt,
         });
 
-        if (stream) {
+        if (stream && reader) {
           reply.header('Content-Type', 'text/event-stream');
           reply.header('Cache-Control', 'no-cache');
           reply.header('Connection', 'keep-alive');
           reply.raw.writeHead(200);
-          const reader = res.body?.getReader();
-          if (!reader) {
-            reply.raw.end();
-            return;
-          }
           const decoder = new TextDecoder();
           try {
             while (true) {
@@ -155,7 +151,7 @@ export function registerGatewayRoutes(
         const json = await res.json();
         return reply.send(json);
       } catch {
-        // Connection error — failover to the next candidate.
+        recordFailure(compositeId);
         continue;
       }
     }