diff --git a/docs-site/src/content/docs/guides/codex-app-models.md b/docs-site/src/content/docs/guides/codex-app-models.md
index c84f8f6407..053fba1669 100644
--- a/docs-site/src/content/docs/guides/codex-app-models.md
+++ b/docs-site/src/content/docs/guides/codex-app-models.md
@@ -105,6 +105,11 @@ them by ignoring `visibility`. See [Codex Desktop native-allowlist compatibility
 for the command, disable-key semantics, and safety constraints.
 
 ## Integration path
+
+`ocx init`, `ocx start`, and `ocx sync` wire the shared Codex config and catalog into the proxy; see
+[Codex Integration](/guides/codex-integration/) for config injection, catalog sync, shims, WebSocket
+fallback, and restore mechanics.
+
 ## Native quota fallback limitation
 
 When the Codex app exhausts its native five-hour quota it can switch to a reserve
@@ -137,11 +142,6 @@ mode is active; if the client rewrites or refuses it before the request leaves,
 setting changes that. Treat the explicit-selection route as worth trying rather than a confirmed
 workaround.
 
-
-`ocx init`, `ocx start`, and `ocx sync` wire the shared Codex config and catalog into the proxy; see
-[Codex Integration](/guides/codex-integration/) for config injection, catalog sync, shims, WebSocket
-fallback, and restore mechanics.
-
 ## Why routed models show up
 
 Codex's model picker expects Codex-shaped catalog entries. opencodex builds routed entries by cloning
diff --git a/src/server/responses/core.ts b/src/server/responses/core.ts
index 96bb1f8ce5..1f56ed979a 100644
--- a/src/server/responses/core.ts
+++ b/src/server/responses/core.ts
@@ -6144,7 +6144,15 @@ async function handleResponsesInner(
           {
             abortSignal: upstream.signal,
             label: safeHostLabel(builtContinuationRequest.url),
-            ...(continuationTransientPolicy ? { attempts: continuationTransientPolicy.attempts } : {}),
+            // Same request-scoped budget as the initial send and the 429/rotation refetches:
+            // a terminal-guard continuation is another leg of ONE request, so handing it a
+            // fresh `attempts` would let one request exceed the configured total-send ceiling.
+            ...(continuationTransientPolicy
+              ? {
+                attempts: remainingTransientSendBudget(continuationTransientPolicy.attempts),
+                onSendsConsumed: noteTransientSends,
+              }
+              : {}),
           },
           );
       } finally {
diff --git a/tests/transient-budget-scope-source.test.ts b/tests/transient-budget-scope-source.test.ts
new file mode 100644
index 0000000000..437973ba32
--- /dev/null
+++ b/tests/transient-budget-scope-source.test.ts
@@ -0,0 +1,52 @@
+import { describe, expect, test } from "bun:test";
+import { readFileSync } from "node:fs";
+import { join } from "node:path";
+
+const source = (relative: string): string =>
+  readFileSync(join(import.meta.dir, "..", "src", ...relative.split("/")), "utf8");
+
+/**
+ * `transientRetryOn5xx.attempts` is ONE request-wide total-send budget, not a per-leg
+ * allowance. A Responses request can reach upstream on several legs — the initial send, a
+ * 429/account-rotation refetch, and the terminal-guard continuation — and each leg calls
+ * `fetchWithTransientRetry` separately. The budget only holds if every leg draws from the
+ * shared request-scoped counter.
+ *
+ * The continuation leg shipped on the raw policy value instead, so a request that reached it
+ * received a fresh full `attempts` allowance: with `attempts: 3` an initial send that had
+ * already spent its budget could still emit three more upstream sends. Runtime coverage in
+ * `tests/upstream-transient-retry.test.ts` proves the helper reports and honors a remainder;
+ * it cannot prove that every call site asks for one, because a site that forgets simply
+ * passes a larger number. This asserts the wiring at the source, which is the only place the
+ * omission is visible.
+ */
+describe("transient send budget stays request-scoped", () => {
+  test("every transient-retry call site draws from the shared counter", () => {
+    const core = source("server/responses/core.ts");
+
+    // One owner per request, declared before any leg can send.
+    expect(core.match(/let transientSendsUsed = 0;/g)).toHaveLength(1);
+    expect(core.match(/const remainingTransientSendBudget = \(budget: number\): number =>/g)).toHaveLength(1);
+
+    // Initial send, 429/rotation refetch, and terminal-guard continuation: three legs, three
+    // reports into the same counter.
+    expect(core.match(/onSendsConsumed: noteTransientSends/g)).toHaveLength(3);
+
+    // The refetch and continuation legs must ask for the REMAINDER. Only the initial send may
+    // pass a policy value directly, because nothing has been spent yet.
+    expect(core.match(/attempts: remainingTransientSendBudget\(/g)).toHaveLength(2);
+    expect(core).toContain("attempts: remainingTransientSendBudget(refetchTransientPolicy.attempts)");
+    expect(core).toContain("attempts: remainingTransientSendBudget(continuationTransientPolicy.attempts)");
+
+    // The regressed shape: a leg handing itself a fresh full budget.
+    expect(core).not.toContain("attempts: continuationTransientPolicy.attempts }");
+    expect(core).not.toContain("attempts: refetchTransientPolicy.attempts }");
+  });
+
+  test("the helper still exposes the seam those call sites depend on", () => {
+    const retry = source("lib/upstream-retry.ts");
+    expect(retry).toContain("onSendsConsumed?: (sends: number) => void;");
+    // Reported in `finally` so every exit path — return, throw, abort — feeds the counter.
+    expect(retry).toMatch(/} finally \{\n\s*opts\.onSendsConsumed\?\.\(sent\);/);
+  });
+});