fix(providers): only send Codex prompt cache params to native OpenAI endpoints (#1342)

buildRequestBody emitted prompt_cache_key and prompt_cache_retention for
every Codex request. The ChatGPT subscription OAuth backend
(chatgpt.com/backend-api) rejects prompt_cache_retention with HTTP 400
"Unsupported parameter", so every LLM call from an agent using a ChatGPT
OAuth Codex provider failed at the first think step.

Gate both params on isOpenAINativeEndpoint(apiBase), the same native-only
policy already documented and applied by CacheMiddleware. This covers both
CodexProvider and CodexAdapter, since the adapter delegates to the same
builder. Server-side prefix caching still works on the OAuth backend
without these params.

Update the two prompt-cache tests to assert both cases: native endpoints
include the params, the OAuth backend omits them.

internal/providers/adapter_codex_test.go

@@ -388,18 +388,20 @@ func TestCodexAdapter_FromResponse_PromptTokensDetailsAlias(t *testing.T) {
 }
 
 func TestCodexAdapter_ToRequestPromptCacheControls(t *testing.T) {
-	a, _ := NewCodexAdapter(ProviderConfig{})
-	body, _, err := a.ToRequest(ChatRequest{
+	req := ChatRequest{
 		Messages: []Message{{Role: "user", Content: "hi"}},
 		Options: map[string]any{
 			OptPromptCacheKey:       "tenant/agent/session",
 			OptPromptCacheRetention: "1h",
 		},
-	})
+	}
+
+	// Native OpenAI endpoint accepts prompt cache params.
+	nativeAdapter, _ := NewCodexAdapter(ProviderConfig{BaseURL: "https://api.openai.com/v1"})
+	body, _, err := nativeAdapter.ToRequest(req)
 	if err != nil {
 		t.Fatalf("ToRequest: %v", err)
 	}
-
 	var payload map[string]any
 	if err := json.Unmarshal(body, &payload); err != nil {
 		t.Fatalf("decode body: %v", err)
@@ -410,6 +412,23 @@ func TestCodexAdapter_ToRequestPromptCacheControls(t *testing.T) {
 	if got := payload["prompt_cache_retention"]; got != "1h" {
 		t.Fatalf("prompt_cache_retention = %v, want 1h", got)
 	}
+
+	// Default ChatGPT OAuth backend rejects these params with HTTP 400 — must omit.
+	oauthAdapter, _ := NewCodexAdapter(ProviderConfig{})
+	oauthBody, _, err := oauthAdapter.ToRequest(req)
+	if err != nil {
+		t.Fatalf("ToRequest (oauth): %v", err)
+	}
+	var oauthPayload map[string]any
+	if err := json.Unmarshal(oauthBody, &oauthPayload); err != nil {
+		t.Fatalf("decode oauth body: %v", err)
+	}
+	if _, ok := oauthPayload["prompt_cache_key"]; ok {
+		t.Fatal("prompt_cache_key must not be sent to the ChatGPT OAuth backend")
+	}
+	if _, ok := oauthPayload["prompt_cache_retention"]; ok {
+		t.Fatal("prompt_cache_retention must not be sent to the ChatGPT OAuth backend")
+	}
 }
 
 func TestCodexAdapter_CapabilitiesCacheControl(t *testing.T) {

internal/providers/codex_build.go

@@ -140,11 +140,18 @@ func (p *CodexProvider) buildRequestBody(req ChatRequest, stream bool) map[strin
 		body["reasoning"] = map[string]any{"effort": level}
 	}
 
-	if cacheKey, ok := req.Options[OptPromptCacheKey]; ok {
-		body["prompt_cache_key"] = cacheKey
-	}
-	if retention, ok := req.Options[OptPromptCacheRetention]; ok {
-		body["prompt_cache_retention"] = retention
+	// Prompt caching params (prompt_cache_key, prompt_cache_retention) are accepted
+	// only by native OpenAI endpoints. The ChatGPT subscription OAuth backend
+	// (chatgpt.com/backend-api) rejects them with HTTP 400, so gate on the endpoint —
+	// the same native-only policy CacheMiddleware already applies. Server-side prefix
+	// caching still works on the OAuth backend without these params.
+	if isOpenAINativeEndpoint(p.apiBase) {
+		if cacheKey, ok := req.Options[OptPromptCacheKey]; ok {
+			body["prompt_cache_key"] = cacheKey
+		}
+		if retention, ok := req.Options[OptPromptCacheRetention]; ok {
+			body["prompt_cache_retention"] = retention
+		}
 	}
 
 	return body

internal/providers/codex_test.go

@@ -1131,21 +1131,33 @@ func TestCodexBuildRequestBody_NilFunction_HandlesGracefully(t *testing.T) {
 }
 
 func TestCodexBuildRequestBodyPromptCacheControls(t *testing.T) {
-	p := NewCodexProvider("test", &staticTokenSource{token: "tok"}, "", "gpt-4o")
-
-	body := p.buildRequestBody(ChatRequest{
+	req := ChatRequest{
 		Messages: []Message{{Role: "user", Content: "Hello"}},
 		Options: map[string]any{
 			OptPromptCacheKey:       "agent/session/provider",
 			OptPromptCacheRetention: "24h",
 		},
-	}, true)
+	}
 
+	// Native OpenAI endpoint accepts prompt cache params.
+	native := NewCodexProvider("test", &staticTokenSource{token: "tok"}, "https://api.openai.com/v1", "gpt-4o")
+	body := native.buildRequestBody(req, true)
 	if got := body["prompt_cache_key"]; got != "agent/session/provider" {
-		t.Fatalf("prompt_cache_key = %v, want agent/session/provider", got)
+		t.Fatalf("native prompt_cache_key = %v, want agent/session/provider", got)
 	}
 	if got := body["prompt_cache_retention"]; got != "24h" {
-		t.Fatalf("prompt_cache_retention = %v, want 24h", got)
+		t.Fatalf("native prompt_cache_retention = %v, want 24h", got)
+	}
+
+	// ChatGPT subscription OAuth backend (default apiBase) rejects these params
+	// with HTTP 400, so they must be omitted.
+	oauth := NewCodexProvider("test", &staticTokenSource{token: "tok"}, "", "gpt-4o")
+	oauthBody := oauth.buildRequestBody(req, true)
+	if _, ok := oauthBody["prompt_cache_key"]; ok {
+		t.Fatal("prompt_cache_key must not be sent to the ChatGPT OAuth backend")
+	}
+	if _, ok := oauthBody["prompt_cache_retention"]; ok {
+		t.Fatal("prompt_cache_retention must not be sent to the ChatGPT OAuth backend")
 	}
 }