fix(routing): strip prompt_cache_key for NVIDIA NIM (#7617) (#7709)

Codex CLI injects prompt_cache_key natively for its own prompt caching.
injectPromptCacheKey() only guards against the router injecting a NEW
key for nvidia/codex/xai — it never strips a key that arrived already
present in the inbound body. NVIDIA NIM's OpenAI-compatible wrapper
rejects the field with a 400, and NIM has no documented support for
prompt caching (providerSupportsCaching already treats nvidia as
non-cache-capable).

Adds a provider-wide STRIP_RULES entry in paramSupport.ts (match-all,
since prompt_cache_key rejection isn't model-specific) so
stripUnsupportedParams() drops it for every nvidia target before the
request reaches DefaultExecutor.

changelog.d/fixes/7617-nvidia-strip-cache-key.md

@@ -0,0 +1 @@
+- fix(routing): strip `prompt_cache_key` for NVIDIA NIM — Codex CLI injects it, NIM's OpenAI-compatible wrapper 400s on it (#7617)

open-sse/translator/paramSupport.ts

@@ -50,6 +50,11 @@ const STRIP_RULES: StripRule[] = [
   // (format:"openai") does not accept the Claude-style `thinking` body field
   // and returns 400 "Unsupported parameter(s): thinking". Upstream #2268.
   { provider: "nvidia", match: /minimax-m2\.7/i, drop: ["thinking"] },
+  // NVIDIA NIM: OpenAI-compatible wrapper 400s on `prompt_cache_key` (Codex CLI
+  // injects it natively for its own prompt caching). NIM has no documented
+  // support for this field (providerSupportsCaching already treats nvidia as
+  // non-cache-capable) — safe to drop provider-wide, not model-specific. #7617.
+  { provider: "nvidia", match: /.*/, drop: ["prompt_cache_key"] },
   // VolcEngine Ark caps the Kimi coding-plan endpoint at max_tokens <= 32768
   // server-side ("integer above maximum value, expected a value <= 32768"),
   // independent of the model's own catalog ceiling. Confirmed against two

tests/unit/nvidia-strip-prompt-cache-key.test.ts

@@ -0,0 +1,41 @@
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { stripUnsupportedParams } from "../../open-sse/translator/paramSupport.ts";
+
+test("#7617: stripUnsupportedParams strips prompt_cache_key for nvidia when present", () => {
+  const body: Record<string, unknown> = {
+    model: "some-nvidia-model",
+    prompt_cache_key: "codex-cli-session-abc123",
+    max_tokens: 512,
+  };
+  stripUnsupportedParams("nvidia", "some-nvidia-model", body);
+  assert.equal(
+    body.prompt_cache_key,
+    undefined,
+    "prompt_cache_key must be stripped for nvidia"
+  );
+  assert.equal(body.max_tokens, 512, "unrelated params must be preserved");
+});
+
+test("#7617: stripUnsupportedParams preserves prompt_cache_key for non-nvidia providers (e.g. openai)", () => {
+  const body: Record<string, unknown> = {
+    model: "gpt-5.4",
+    prompt_cache_key: "some-cache-key",
+  };
+  stripUnsupportedParams("openai", "gpt-5.4", body);
+  assert.equal(
+    body.prompt_cache_key,
+    "some-cache-key",
+    "prompt_cache_key must be preserved for non-nvidia providers"
+  );
+});
+
+test("#7617: stripUnsupportedParams is a no-op for nvidia when prompt_cache_key is absent", () => {
+  const body: Record<string, unknown> = {
+    model: "some-nvidia-model",
+    max_tokens: 256,
+  };
+  stripUnsupportedParams("nvidia", "some-nvidia-model", body);
+  assert.equal("prompt_cache_key" in body, false);
+  assert.equal(body.max_tokens, 256);
+});