From e3a8f0e0a1abd8fed7264741bb500bacd015f4e5 Mon Sep 17 00:00:00 2001
From: 134 <a134@134deMacBook-Pro.local>
Date: Wed, 21 Oct 2026 10:00:23 +0800
Subject: [PATCH] fix(agent): consider the full fallback pool within the
 request deadline

---
 lib/ai/runtime/agent-model-failover.ts              | 2 --
 tests/integration/agent-multimodal-failover.test.ts | 9 ++++++---
 2 files changed, 6 insertions(+), 5 deletions(-)

diff --git a/lib/ai/runtime/agent-model-failover.ts b/lib/ai/runtime/agent-model-failover.ts
index 8a609e42d..9180ec5fb 100644
--- a/lib/ai/runtime/agent-model-failover.ts
+++ b/lib/ai/runtime/agent-model-failover.ts
@@ -67,11 +67,9 @@ export function createAgentModelFailover(entries: AgentModelEntry[], circuit: Ag
     const totalSignal = signal ? AbortSignal.any([signal, deadline]) : deadline;
     const ordered = [...entries].sort((a, b) => Number(b.model.id === preferred) - Number(a.model.id === preferred));
     let last: unknown;
-    let attempts = 0;
     for (const entry of ordered) {
       totalSignal.throwIfAborted();
       if (await circuit.isPaused(entry)) continue;
-      if (attempts++ >= 4) break;
       const attemptSignal = AbortSignal.any([totalSignal, AbortSignal.timeout(AGENT_MODEL_ATTEMPT_MS)]);
       try {
         const result = await execute(entry, attemptSignal);
diff --git a/tests/integration/agent-multimodal-failover.test.ts b/tests/integration/agent-multimodal-failover.test.ts
index 87f1e83ea..c21530d2b 100644
--- a/tests/integration/agent-multimodal-failover.test.ts
+++ b/tests/integration/agent-multimodal-failover.test.ts
@@ -54,11 +54,14 @@ test("invalid input, content policy and cancellation neither quarantine nor retr
   assert.equal(s.faults.size, 0);
 });
 
-test("multiple failed backups are skipped; no model is tried twice and attempts are bounded", async () => {
+test("every eligible backup is considered once; a fixed small retry count cannot strand a healthy later model", async () => {
   const s = store(); const calls: string[] = [];
   await assert.rejects(createAgentModelFailover(["a", "b", "c", "d", "e"].map(id => entry(id)), s.circuit)(async e => { calls.push(e.model.id); throw Object.assign(new Error("busy"), { status: 503 }); }));
-  assert.deepEqual(calls, ["a", "b", "c", "d"]);
-  assert.equal(s.faults.size, 4);
+  assert.deepEqual(calls, ["a", "b", "c", "d", "e"]);
+  assert.equal(s.faults.size, 5);
+  calls.length = 0;
+  const result = await createAgentModelFailover(["a", "b", "c", "d", "e", "healthy"].map(id => entry(id)), s.circuit)(async e => { calls.push(e.model.id); return "ok"; });
+  assert.equal(result, "ok"); assert.deepEqual(calls, ["healthy"]);
 });
 
 test("late success cannot clear a fault recorded by another in-flight request", async () => {