diff --git a/src/shared/utils/classify429.ts b/src/shared/utils/classify429.ts
index 6c6f4eee2e..531a5bae67 100644
--- a/src/shared/utils/classify429.ts
+++ b/src/shared/utils/classify429.ts
@@ -66,6 +66,23 @@ const QUOTA_PATTERNS: ReadonlyArray<RegExp> = [
   // the 429 is misclassified as transient rate_limit and retried every
   // ~60s against a budget that only resets at UTC midnight.
   /daily free allocation/i,
+
+  // Modal-hosted OpenAI-compatible endpoints (e.g. self-hosted Kimi K3).
+  // Body: {"error":"usage limit reached"}, no nested "message"/"quota"/
+  // "daily" wording. Without this pattern the 429 falls through to
+  // "rate_limit" (short cooldown), so combo round-robin's per-conversation
+  // session stickiness (#3825) keeps re-targeting the same exhausted
+  // connection every turn instead of a long lockout that lets the sticky
+  // target fail over to another account.
+  //
+  // Matches the "error" JSON key with "usage limit reached" as its value.
+  // Extra sibling fields (e.g. {"error":"usage limit reached", "code":"..."})
+  // still match. A different key like {"detail":"..."} or a qualified value
+  // like {"error":"Per-minute usage limit reached"} does NOT match. Bare
+  // string bodies without a JSON wrapper also do NOT match.
+  // Trailing punctuation/whitespace before the closing quote is tolerated
+  // because real API responses may include a period or trailing space.
+  /"error"\s*:\s*"usage limit reached[.\s]*"/i,
 ];
 
 /**
diff --git a/tests/unit/classify429.test.ts b/tests/unit/classify429.test.ts
index 234c42ee73..08ba94de12 100644
--- a/tests/unit/classify429.test.ts
+++ b/tests/unit/classify429.test.ts
@@ -164,6 +164,63 @@ test("looksLikeQuotaExhausted: rejects empty / null / non-quota text", () => {
   assert.equal(looksLikeQuotaExhausted("server error 500"), false);
 });
 
+test("classify429: Modal-hosted endpoint 'usage limit reached' body returns 'quota_exhausted'", () => {
+  // Real body observed from a self-hosted Modal OpenAI-compatible endpoint:
+  // {"error":"usage limit reached"} - a bare string value, no "message"/
+  // "daily"/"quota" wording, so none of the prior patterns matched and the
+  // 429 fell through to a 60s rate_limit cooldown. Combo round-robin's
+  // per-conversation session stickiness (#3825) then kept re-targeting the
+  // same exhausted connection on every turn of a long-running session.
+  const body = { error: "usage limit reached" };
+  assert.equal(looksLikeQuotaExhausted(body), true);
+  assert.equal(classify429({ status: 429, body }), "quota_exhausted");
+  assert.equal(classify429({ status: 429, body: JSON.stringify(body) }), "quota_exhausted");
+  // Case variation must also match.
+  assert.equal(
+    classify429({ status: 429, body: { error: "USAGE LIMIT REACHED" } }),
+    "quota_exhausted"
+  );
+  // Whitespace around JSON object must also match (bodyToText does not trim).
+  assert.equal(
+    classify429({ status: 429, body: ' { "error" : "usage limit reached" } ' }),
+    "quota_exhausted"
+  );
+  // Extra sibling fields must still match.
+  assert.equal(
+    classify429({
+      status: 429,
+      body: { error: "usage limit reached", code: "RESOURCE_EXHAUSTED" },
+    }),
+    "quota_exhausted"
+  );
+  // Trailing punctuation/whitespace must still match.
+  assert.equal(classify429({ status: 429, body: { error: "usage limit reached." } }), "quota_exhausted");
+  assert.equal(classify429({ status: 429, body: { error: "usage limit reached " } }), "quota_exhausted");
+});
+
+test("classify429: qualified transient 'usage limit reached' messages stay rate_limit", () => {
+  // The Modal pattern requires the "error" JSON key with exactly "usage
+  // limit reached" as its value - anything else is a transient rate limit
+  // and must NOT be locked out long-term.
+  assert.equal(
+    classify429({ status: 429, body: "Per-minute usage limit reached, retry in 60s." }),
+    "rate_limit"
+  );
+  assert.equal(
+    classify429({ status: 429, body: { error: { message: "RPM usage limit reached" } } }),
+    "rate_limit"
+  );
+  // Bare string body (no JSON "error" key) must NOT match.
+  assert.equal(classify429({ status: 429, body: "usage limit reached" }), "rate_limit");
+  // Different JSON key (not "error") must NOT match.
+  assert.equal(classify429({ status: 429, body: { detail: "usage limit reached" } }), "rate_limit");
+  // Qualified value under the "error" key must NOT match.
+  assert.equal(
+    classify429({ status: 429, body: { error: "Per-minute usage limit reached" } }),
+    "rate_limit"
+  );
+});
+
 test("ambiguous 'daily rate limit' messages classify as quota_exhausted (intentional)", () => {
   // Codex audit LOW: messages combining 'daily' or 'monthly' with 'limit'
   // match the quota regex even when paired with 'rate'. This is intentional